Initial import of code.

densecrf/include/Eigen/src/Core/Reverse.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Ricard Marxer <email@ricardmarxer.com>
+// Copyright (C) 2009-2010 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_REVERSE_H
+#define EIGEN_REVERSE_H
+
+namespace Eigen { 
+
+/** \class Reverse
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the reverse of a vector or matrix
+  *
+  * \param MatrixType the type of the object of which we are taking the reverse
+  *
+  * This class represents an expression of the reverse of a vector.
+  * It is the return type of MatrixBase::reverse() and VectorwiseOp::reverse()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::reverse(), VectorwiseOp::reverse()
+  */
+
+namespace internal {
+
+template<typename MatrixType, int Direction>
+struct traits<Reverse<MatrixType, Direction> >
+ : traits<MatrixType>
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename traits<MatrixType>::XprKind XprKind;
+  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+  enum {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+
+    // let's enable LinearAccess only with vectorization because of the product overhead
+    LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) )
+                 ? LinearAccessBit : 0,
+
+    Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess),
+
+    CoeffReadCost = _MatrixTypeNested::CoeffReadCost
+  };
+};
+
+template<typename PacketScalar, bool ReversePacket> struct reverse_packet_cond
+{
+  static inline PacketScalar run(const PacketScalar& x) { return preverse(x); }
+};
+
+template<typename PacketScalar> struct reverse_packet_cond<PacketScalar,false>
+{
+  static inline PacketScalar run(const PacketScalar& x) { return x; }
+};
+
+} // end namespace internal 
+
+template<typename MatrixType, int Direction> class Reverse
+  : public internal::dense_xpr_base< Reverse<MatrixType, Direction> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<Reverse>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Reverse)
+    using Base::IsRowMajor;
+
+    // next line is necessary because otherwise const version of operator()
+    // is hidden by non-const version defined in this file
+    using Base::operator(); 
+
+  protected:
+    enum {
+      PacketSize = internal::packet_traits<Scalar>::size,
+      IsColMajor = !IsRowMajor,
+      ReverseRow = (Direction == Vertical)   || (Direction == BothDirections),
+      ReverseCol = (Direction == Horizontal) || (Direction == BothDirections),
+      OffsetRow  = ReverseRow && IsColMajor ? PacketSize : 1,
+      OffsetCol  = ReverseCol && IsRowMajor ? PacketSize : 1,
+      ReversePacket = (Direction == BothDirections)
+                    || ((Direction == Vertical)   && IsColMajor)
+                    || ((Direction == Horizontal) && IsRowMajor)
+    };
+    typedef internal::reverse_packet_cond<PacketScalar,ReversePacket> reverse_packet;
+  public:
+
+    inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse)
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+
+    inline Index innerStride() const
+    {
+      return -m_matrix.innerStride();
+    }
+
+    inline Scalar& operator()(Index row, Index col)
+    {
+      eigen_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
+      return coeffRef(row, col);
+    }
+
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      return m_matrix.const_cast_derived().coeffRef(ReverseRow ? m_matrix.rows() - row - 1 : row,
+                                                    ReverseCol ? m_matrix.cols() - col - 1 : col);
+    }
+
+    inline CoeffReturnType coeff(Index row, Index col) const
+    {
+      return m_matrix.coeff(ReverseRow ? m_matrix.rows() - row - 1 : row,
+                            ReverseCol ? m_matrix.cols() - col - 1 : col);
+    }
+
+    inline CoeffReturnType coeff(Index index) const
+    {
+      return m_matrix.coeff(m_matrix.size() - index - 1);
+    }
+
+    inline Scalar& coeffRef(Index index)
+    {
+      return m_matrix.const_cast_derived().coeffRef(m_matrix.size() - index - 1);
+    }
+
+    inline Scalar& operator()(Index index)
+    {
+      eigen_assert(index >= 0 && index < m_matrix.size());
+      return coeffRef(index);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index row, Index col) const
+    {
+      return reverse_packet::run(m_matrix.template packet<LoadMode>(
+                                    ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
+                                    ReverseCol ? m_matrix.cols() - col - OffsetCol : col));
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index row, Index col, const PacketScalar& x)
+    {
+      m_matrix.const_cast_derived().template writePacket<LoadMode>(
+                                      ReverseRow ? m_matrix.rows() - row - OffsetRow : row,
+                                      ReverseCol ? m_matrix.cols() - col - OffsetCol : col,
+                                      reverse_packet::run(x));
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index index) const
+    {
+      return internal::preverse(m_matrix.template packet<LoadMode>( m_matrix.size() - index - PacketSize ));
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index index, const PacketScalar& x)
+    {
+      m_matrix.const_cast_derived().template writePacket<LoadMode>(m_matrix.size() - index - PacketSize, internal::preverse(x));
+    }
+
+    const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    nestedExpression() const 
+    {
+      return m_matrix;
+    }
+
+  protected:
+    typename MatrixType::Nested m_matrix;
+};
+
+/** \returns an expression of the reverse of *this.
+  *
+  * Example: \include MatrixBase_reverse.cpp
+  * Output: \verbinclude MatrixBase_reverse.out
+  *
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::ReverseReturnType
+DenseBase<Derived>::reverse()
+{
+  return derived();
+}
+
+/** This is the const version of reverse(). */
+template<typename Derived>
+inline const typename DenseBase<Derived>::ConstReverseReturnType
+DenseBase<Derived>::reverse() const
+{
+  return derived();
+}
+
+/** This is the "in place" version of reverse: it reverses \c *this.
+  *
+  * In most cases it is probably better to simply use the reversed expression
+  * of a matrix. However, when reversing the matrix data itself is really needed,
+  * then this "in-place" version is probably the right choice because it provides
+  * the following additional features:
+  *  - less error prone: doing the same operation with .reverse() requires special care:
+  *    \code m = m.reverse().eval(); \endcode
+  *  - this API allows to avoid creating a temporary (the current implementation creates a temporary, but that could be avoided using swap)
+  *  - it allows future optimizations (cache friendliness, etc.)
+  *
+  * \sa reverse() */
+template<typename Derived>
+inline void DenseBase<Derived>::reverseInPlace()
+{
+  derived() = derived().reverse().eval();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_REVERSE_H

densecrf/include/Eigen/src/Core/Select.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 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_SELECT_H
+#define EIGEN_SELECT_H
+
+namespace Eigen { 
+
+/** \class Select
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a coefficient wise version of the C++ ternary operator ?:
+  *
+  * \param ConditionMatrixType the type of the \em condition expression which must be a boolean matrix
+  * \param ThenMatrixType the type of the \em then expression
+  * \param ElseMatrixType the type of the \em else expression
+  *
+  * This class represents an expression of a coefficient wise version of the C++ ternary operator ?:.
+  * It is the return type of DenseBase::select() and most of the time this is the only way it is used.
+  *
+  * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const
+  */
+
+namespace internal {
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
+struct traits<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
+ : traits<ThenMatrixType>
+{
+  typedef typename traits<ThenMatrixType>::Scalar Scalar;
+  typedef Dense StorageKind;
+  typedef typename traits<ThenMatrixType>::XprKind XprKind;
+  typedef typename ConditionMatrixType::Nested ConditionMatrixNested;
+  typedef typename ThenMatrixType::Nested ThenMatrixNested;
+  typedef typename ElseMatrixType::Nested ElseMatrixNested;
+  enum {
+    RowsAtCompileTime = ConditionMatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime,
+    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits,
+    CoeffReadCost = traits<typename remove_all<ConditionMatrixNested>::type>::CoeffReadCost
+                  + EIGEN_SIZE_MAX(traits<typename remove_all<ThenMatrixNested>::type>::CoeffReadCost,
+                                   traits<typename remove_all<ElseMatrixNested>::type>::CoeffReadCost)
+  };
+};
+}
+
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
+class Select : internal::no_assignment_operator,
+  public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<Select>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Select)
+
+    Select(const ConditionMatrixType& conditionMatrix,
+           const ThenMatrixType& thenMatrix,
+           const ElseMatrixType& elseMatrix)
+      : m_condition(conditionMatrix), m_then(thenMatrix), m_else(elseMatrix)
+    {
+      eigen_assert(m_condition.rows() == m_then.rows() && m_condition.rows() == m_else.rows());
+      eigen_assert(m_condition.cols() == m_then.cols() && m_condition.cols() == m_else.cols());
+    }
+
+    Index rows() const { return m_condition.rows(); }
+    Index cols() const { return m_condition.cols(); }
+
+    const Scalar coeff(Index i, Index j) const
+    {
+      if (m_condition.coeff(i,j))
+        return m_then.coeff(i,j);
+      else
+        return m_else.coeff(i,j);
+    }
+
+    const Scalar coeff(Index i) const
+    {
+      if (m_condition.coeff(i))
+        return m_then.coeff(i);
+      else
+        return m_else.coeff(i);
+    }
+
+    const ConditionMatrixType& conditionMatrix() const
+    {
+      return m_condition;
+    }
+
+    const ThenMatrixType& thenMatrix() const
+    {
+      return m_then;
+    }
+
+    const ElseMatrixType& elseMatrix() const
+    {
+      return m_else;
+    }
+
+  protected:
+    typename ConditionMatrixType::Nested m_condition;
+    typename ThenMatrixType::Nested m_then;
+    typename ElseMatrixType::Nested m_else;
+};
+
+
+/** \returns a matrix where each coefficient (i,j) is equal to \a thenMatrix(i,j)
+  * if \c *this(i,j), and \a elseMatrix(i,j) otherwise.
+  *
+  * Example: \include MatrixBase_select.cpp
+  * Output: \verbinclude MatrixBase_select.out
+  *
+  * \sa class Select
+  */
+template<typename Derived>
+template<typename ThenDerived,typename ElseDerived>
+inline const Select<Derived,ThenDerived,ElseDerived>
+DenseBase<Derived>::select(const DenseBase<ThenDerived>& thenMatrix,
+                            const DenseBase<ElseDerived>& elseMatrix) const
+{
+  return Select<Derived,ThenDerived,ElseDerived>(derived(), thenMatrix.derived(), elseMatrix.derived());
+}
+
+/** Version of DenseBase::select(const DenseBase&, const DenseBase&) with
+  * the \em else expression being a scalar value.
+  *
+  * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const, class Select
+  */
+template<typename Derived>
+template<typename ThenDerived>
+inline const Select<Derived,ThenDerived, typename ThenDerived::ConstantReturnType>
+DenseBase<Derived>::select(const DenseBase<ThenDerived>& thenMatrix,
+                            typename ThenDerived::Scalar elseScalar) const
+{
+  return Select<Derived,ThenDerived,typename ThenDerived::ConstantReturnType>(
+    derived(), thenMatrix.derived(), ThenDerived::Constant(rows(),cols(),elseScalar));
+}
+
+/** Version of DenseBase::select(const DenseBase&, const DenseBase&) with
+  * the \em then expression being a scalar value.
+  *
+  * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const, class Select
+  */
+template<typename Derived>
+template<typename ElseDerived>
+inline const Select<Derived, typename ElseDerived::ConstantReturnType, ElseDerived >
+DenseBase<Derived>::select(typename ElseDerived::Scalar thenScalar,
+                            const DenseBase<ElseDerived>& elseMatrix) const
+{
+  return Select<Derived,typename ElseDerived::ConstantReturnType,ElseDerived>(
+    derived(), ElseDerived::Constant(rows(),cols(),thenScalar), elseMatrix.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELECT_H

densecrf/include/Eigen/src/Core/SelfAdjointView.h

+// 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_SELFADJOINTMATRIX_H
+#define EIGEN_SELFADJOINTMATRIX_H
+
+namespace Eigen { 
+
+/** \class SelfAdjointView
+  * \ingroup Core_Module
+  *
+  *
+  * \brief Expression of a selfadjoint matrix from a triangular part of a dense matrix
+  *
+  * \param MatrixType the type of the dense matrix storing the coefficients
+  * \param TriangularPart can be either \c #Lower or \c #Upper
+  *
+  * This class is an expression of a sefladjoint matrix from a triangular part of a matrix
+  * with given dense storage of the coefficients. It is the return type of MatrixBase::selfadjointView()
+  * and most of the time this is the only way that it is used.
+  *
+  * \sa class TriangularBase, MatrixBase::selfadjointView()
+  */
+
+namespace internal {
+template<typename MatrixType, unsigned int UpLo>
+struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType>
+{
+  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
+  typedef MatrixType ExpressionType;
+  typedef typename MatrixType::PlainObject DenseMatrixType;
+  enum {
+    Mode = UpLo | SelfAdjoint,
+    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits)
+           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)), // FIXME these flags should be preserved
+    CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost
+  };
+};
+}
+
+template <typename Lhs, int LhsMode, bool LhsIsVector,
+          typename Rhs, int RhsMode, bool RhsIsVector>
+struct SelfadjointProductMatrix;
+
+// FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ??
+template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
+  : public TriangularBase<SelfAdjointView<MatrixType, UpLo> >
+{
+  public:
+
+    typedef TriangularBase<SelfAdjointView> Base;
+    typedef typename internal::traits<SelfAdjointView>::MatrixTypeNested MatrixTypeNested;
+    typedef typename internal::traits<SelfAdjointView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
+
+    /** \brief The type of coefficients in this matrix */
+    typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; 
+
+    typedef typename MatrixType::Index Index;
+
+    enum {
+      Mode = internal::traits<SelfAdjointView>::Mode
+    };
+    typedef typename MatrixType::PlainObject PlainObject;
+
+    inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
+    {}
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+    inline Index outerStride() const { return m_matrix.outerStride(); }
+    inline Index innerStride() const { return m_matrix.innerStride(); }
+
+    /** \sa MatrixBase::coeff()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    inline Scalar coeff(Index row, Index col) const
+    {
+      Base::check_coordinates_internal(row, col);
+      return m_matrix.coeff(row, col);
+    }
+
+    /** \sa MatrixBase::coeffRef()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      Base::check_coordinates_internal(row, col);
+      return m_matrix.const_cast_derived().coeffRef(row, col);
+    }
+
+    /** \internal */
+    const MatrixTypeNestedCleaned& _expression() const { return m_matrix; }
+
+    const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; }
+    MatrixTypeNestedCleaned& nestedExpression() { return *const_cast<MatrixTypeNestedCleaned*>(&m_matrix); }
+
+    /** Efficient self-adjoint matrix times vector/matrix product */
+    template<typename OtherDerived>
+    SelfadjointProductMatrix<MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
+    operator*(const MatrixBase<OtherDerived>& rhs) const
+    {
+      return SelfadjointProductMatrix
+              <MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
+              (m_matrix, rhs.derived());
+    }
+
+    /** Efficient vector/matrix times self-adjoint matrix product */
+    template<typename OtherDerived> friend
+    SelfadjointProductMatrix<OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
+    operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs)
+    {
+      return SelfadjointProductMatrix
+              <OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
+              (lhs.derived(),rhs.m_matrix);
+    }
+
+    /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this:
+      * \f$ this = this + \alpha u v^* + conj(\alpha) v u^* \f$
+      * \returns a reference to \c *this
+      *
+      * The vectors \a u and \c v \b must be column vectors, however they can be
+      * a adjoint expression without any overhead. Only the meaningful triangular
+      * part of the matrix is updated, the rest is left unchanged.
+      *
+      * \sa rankUpdate(const MatrixBase<DerivedU>&, Scalar)
+      */
+    template<typename DerivedU, typename DerivedV>
+    SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, Scalar alpha = Scalar(1));
+
+    /** Perform a symmetric rank K update of the selfadjoint matrix \c *this:
+      * \f$ this = this + \alpha ( u u^* ) \f$ where \a u is a vector or matrix.
+      *
+      * \returns a reference to \c *this
+      *
+      * Note that to perform \f$ this = this + \alpha ( u^* u ) \f$ you can simply
+      * call this function with u.adjoint().
+      *
+      * \sa rankUpdate(const MatrixBase<DerivedU>&, const MatrixBase<DerivedV>&, Scalar)
+      */
+    template<typename DerivedU>
+    SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, Scalar alpha = Scalar(1));
+
+/////////// Cholesky module ///////////
+
+    const LLT<PlainObject, UpLo> llt() const;
+    const LDLT<PlainObject, UpLo> ldlt() const;
+
+/////////// Eigenvalue module ///////////
+
+    /** Real part of #Scalar */
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    /** Return type of eigenvalues() */
+    typedef Matrix<RealScalar, internal::traits<MatrixType>::ColsAtCompileTime, 1> EigenvaluesReturnType;
+
+    EigenvaluesReturnType eigenvalues() const;
+    RealScalar operatorNorm() const;
+    
+    #ifdef EIGEN2_SUPPORT
+    template<typename OtherDerived>
+    SelfAdjointView& operator=(const MatrixBase<OtherDerived>& other)
+    {
+      enum {
+        OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper
+      };
+      m_matrix.const_cast_derived().template triangularView<UpLo>() = other;
+      m_matrix.const_cast_derived().template triangularView<OtherPart>() = other.adjoint();
+      return *this;
+    }
+    template<typename OtherMatrixType, unsigned int OtherMode>
+    SelfAdjointView& operator=(const TriangularView<OtherMatrixType, OtherMode>& other)
+    {
+      enum {
+        OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper
+      };
+      m_matrix.const_cast_derived().template triangularView<UpLo>() = other.toDenseMatrix();
+      m_matrix.const_cast_derived().template triangularView<OtherPart>() = other.toDenseMatrix().adjoint();
+      return *this;
+    }
+    #endif
+
+  protected:
+    MatrixTypeNested m_matrix;
+};
+
+
+// template<typename OtherDerived, typename MatrixType, unsigned int UpLo>
+// internal::selfadjoint_matrix_product_returntype<OtherDerived,SelfAdjointView<MatrixType,UpLo> >
+// operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView<MatrixType,UpLo>& rhs)
+// {
+//   return internal::matrix_selfadjoint_product_returntype<OtherDerived,SelfAdjointView<MatrixType,UpLo> >(lhs.derived(),rhs);
+// }
+
+// selfadjoint to dense matrix
+
+namespace internal {
+
+template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount, ClearOpposite>
+{
+  enum {
+    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
+    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
+  };
+
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount-1, ClearOpposite>::run(dst, src);
+
+    if(row == col)
+      dst.coeffRef(row, col) = real(src.coeff(row, col));
+    else if(row < col)
+      dst.coeffRef(col, row) = conj(dst.coeffRef(row, col) = src.coeff(row, col));
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, 0, ClearOpposite>
+{
+  static inline void run(Derived1 &, const Derived2 &) {}
+};
+
+template<typename Derived1, typename Derived2, int UnrollCount, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount, ClearOpposite>
+{
+  enum {
+    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
+    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
+  };
+
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount-1, ClearOpposite>::run(dst, src);
+
+    if(row == col)
+      dst.coeffRef(row, col) = real(src.coeff(row, col));
+    else if(row > col)
+      dst.coeffRef(col, row) = conj(dst.coeffRef(row, col) = src.coeff(row, col));
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, 0, ClearOpposite>
+{
+  static inline void run(Derived1 &, const Derived2 &) {}
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      for(Index i = 0; i < j; ++i)
+      {
+        dst.copyCoeff(i, j, src);
+        dst.coeffRef(j,i) = conj(dst.coeff(i,j));
+      }
+      dst.copyCoeff(j, j, src);
+    }
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dynamic, ClearOpposite>
+{
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+  typedef typename Derived1::Index Index;
+    for(Index i = 0; i < dst.rows(); ++i)
+    {
+      for(Index j = 0; j < i; ++j)
+      {
+        dst.copyCoeff(i, j, src);
+        dst.coeffRef(j,i) = conj(dst.coeff(i,j));
+      }
+      dst.copyCoeff(i, i, src);
+    }
+  }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* Implementation of MatrixBase methods
+***************************************************************************/
+
+template<typename Derived>
+template<unsigned int UpLo>
+typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
+MatrixBase<Derived>::selfadjointView() const
+{
+  return derived();
+}
+
+template<typename Derived>
+template<unsigned int UpLo>
+typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
+MatrixBase<Derived>::selfadjointView()
+{
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINTMATRIX_H

densecrf/include/Eigen/src/Core/SelfCwiseBinaryOp.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 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_SELFCWISEBINARYOP_H
+#define EIGEN_SELFCWISEBINARYOP_H
+
+namespace Eigen { 
+
+/** \class SelfCwiseBinaryOp
+  * \ingroup Core_Module
+  *
+  * \internal
+  *
+  * \brief Internal helper class for optimizing operators like +=, -=
+  *
+  * This is a pseudo expression class re-implementing the copyCoeff/copyPacket
+  * method to directly performs a +=/-= operations in an optimal way. In particular,
+  * this allows to make sure that the input/output data are loaded only once using
+  * aligned packet loads.
+  *
+  * \sa class SwapWrapper for a similar trick.
+  */
+
+namespace internal {
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct traits<SelfCwiseBinaryOp<BinaryOp,Lhs,Rhs> >
+  : traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >
+{
+  enum {
+    // Note that it is still a good idea to preserve the DirectAccessBit
+    // so that assign can correctly align the data.
+    Flags = traits<CwiseBinaryOp<BinaryOp,Lhs,Rhs> >::Flags | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit),
+    OuterStrideAtCompileTime = Lhs::OuterStrideAtCompileTime,
+    InnerStrideAtCompileTime = Lhs::InnerStrideAtCompileTime
+  };
+};
+}
+
+template<typename BinaryOp, typename Lhs, typename Rhs> class SelfCwiseBinaryOp
+  : public internal::dense_xpr_base< SelfCwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<SelfCwiseBinaryOp>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(SelfCwiseBinaryOp)
+
+    typedef typename internal::packet_traits<Scalar>::type Packet;
+
+    inline SelfCwiseBinaryOp(Lhs& xpr, const BinaryOp& func = BinaryOp()) : m_matrix(xpr), m_functor(func) {}
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+    inline Index outerStride() const { return m_matrix.outerStride(); }
+    inline Index innerStride() const { return m_matrix.innerStride(); }
+    inline const Scalar* data() const { return m_matrix.data(); }
+
+    // note that this function is needed by assign to correctly align loads/stores
+    // TODO make Assign use .data()
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
+      return m_matrix.const_cast_derived().coeffRef(row, col);
+    }
+    inline const Scalar& coeffRef(Index row, Index col) const
+    {
+      return m_matrix.coeffRef(row, col);
+    }
+
+    // note that this function is needed by assign to correctly align loads/stores
+    // TODO make Assign use .data()
+    inline Scalar& coeffRef(Index index)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(Lhs)
+      return m_matrix.const_cast_derived().coeffRef(index);
+    }
+    inline const Scalar& coeffRef(Index index) const
+    {
+      return m_matrix.const_cast_derived().coeffRef(index);
+    }
+
+    template<typename OtherDerived>
+    void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(row >= 0 && row < rows()
+                         && col >= 0 && col < cols());
+      Scalar& tmp = m_matrix.coeffRef(row,col);
+      tmp = m_functor(tmp, _other.coeff(row,col));
+    }
+
+    template<typename OtherDerived>
+    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(index >= 0 && index < m_matrix.size());
+      Scalar& tmp = m_matrix.coeffRef(index);
+      tmp = m_functor(tmp, _other.coeff(index));
+    }
+
+    template<typename OtherDerived, int StoreMode, int LoadMode>
+    void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(row >= 0 && row < rows()
+                        && col >= 0 && col < cols());
+      m_matrix.template writePacket<StoreMode>(row, col,
+        m_functor.packetOp(m_matrix.template packet<StoreMode>(row, col),_other.template packet<LoadMode>(row, col)) );
+    }
+
+    template<typename OtherDerived, int StoreMode, int LoadMode>
+    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(index >= 0 && index < m_matrix.size());
+      m_matrix.template writePacket<StoreMode>(index,
+        m_functor.packetOp(m_matrix.template packet<StoreMode>(index),_other.template packet<LoadMode>(index)) );
+    }
+
+    // reimplement lazyAssign to handle complex *= real
+    // see CwiseBinaryOp ctor for details
+    template<typename RhsDerived>
+    EIGEN_STRONG_INLINE SelfCwiseBinaryOp& lazyAssign(const DenseBase<RhsDerived>& rhs)
+    {
+      EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs,RhsDerived)
+      EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename RhsDerived::Scalar);
+      
+    #ifdef EIGEN_DEBUG_ASSIGN
+      internal::assign_traits<SelfCwiseBinaryOp, RhsDerived>::debug();
+    #endif
+      eigen_assert(rows() == rhs.rows() && cols() == rhs.cols());
+      internal::assign_impl<SelfCwiseBinaryOp, RhsDerived>::run(*this,rhs.derived());
+    #ifndef EIGEN_NO_DEBUG
+      this->checkTransposeAliasing(rhs.derived());
+    #endif
+      return *this;
+    }
+    
+    // overloaded to honor evaluation of special matrices
+    // maybe another solution would be to not use SelfCwiseBinaryOp
+    // at first...
+    SelfCwiseBinaryOp& operator=(const Rhs& _rhs)
+    {
+      typename internal::nested<Rhs>::type rhs(_rhs);
+      return Base::operator=(rhs);
+    }
+
+    Lhs& expression() const 
+    { 
+      return m_matrix;
+    }
+
+    const BinaryOp& functor() const 
+    { 
+      return m_functor;
+    }
+
+  protected:
+    Lhs& m_matrix;
+    const BinaryOp& m_functor;
+
+  private:
+    SelfCwiseBinaryOp& operator=(const SelfCwiseBinaryOp&);
+};
+
+template<typename Derived>
+inline Derived& DenseBase<Derived>::operator*=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  SelfCwiseBinaryOp<internal::scalar_product_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
+  tmp = PlainObject::Constant(rows(),cols(),other);
+  return derived();
+}
+
+template<typename Derived>
+inline Derived& DenseBase<Derived>::operator/=(const Scalar& other)
+{
+  typedef typename internal::conditional<NumTraits<Scalar>::IsInteger,
+                                        internal::scalar_quotient_op<Scalar>,
+                                        internal::scalar_product_op<Scalar> >::type BinOp;
+  typedef typename Derived::PlainObject PlainObject;
+  SelfCwiseBinaryOp<BinOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
+  tmp = PlainObject::Constant(rows(),cols(), NumTraits<Scalar>::IsInteger ? other : Scalar(1)/other);
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFCWISEBINARYOP_H

densecrf/include/Eigen/src/Core/SolveTriangular.h

+// 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

densecrf/include/Eigen/src/Core/StableNorm.h

+// 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)
+{
+  Scalar max = bl.cwiseAbs().maxCoeff();
+  if (max>scale)
+  {
+    ssq = ssq * 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 internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::stableNorm() const
+{
+  using std::min;
+  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 * internal::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
+{
+  using std::pow;
+  using std::min;
+  using std::max;
+  static bool initialized = false;
+  static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
+  if(!initialized)
+  {
+    int ibeta, it, iemin, iemax, iexp;
+    RealScalar abig, eps;
+    // This program calculates the machine-dependent constants
+    // bl, b2, slm, s2m, relerr overfl
+    // from the "basic" machine-dependent numbers
+    // 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  = internal::sqrt(eps);         // tolerance for neglecting asml
+    abig    = RealScalar(1.0/eps - 1.0);
+    initialized = true;
+  }
+  Index n = size();
+  RealScalar ab2 = b2 / RealScalar(n);
+  RealScalar asml = RealScalar(0);
+  RealScalar amed = RealScalar(0);
+  RealScalar abig = RealScalar(0);
+  for(Index j=0; j<n; ++j)
+  {
+    RealScalar ax = internal::abs(coeff(j));
+    if(ax > ab2)     abig += internal::abs2(ax*s2m);
+    else if(ax < b1) asml += internal::abs2(ax*s1m);
+    else             amed += internal::abs2(ax);
+  }
+  if(abig > RealScalar(0))
+  {
+    abig = internal::sqrt(abig);
+    if(abig > overfl)
+    {
+      return rbig;
+    }
+    if(amed > RealScalar(0))
+    {
+      abig = abig/s2m;
+      amed = internal::sqrt(amed);
+    }
+    else
+      return abig/s2m;
+  }
+  else if(asml > RealScalar(0))
+  {
+    if (amed > RealScalar(0))
+    {
+      abig = internal::sqrt(amed);
+      amed = internal::sqrt(asml) / s1m;
+    }
+    else
+      return internal::sqrt(asml)/s1m;
+  }
+  else
+    return internal::sqrt(amed);
+  asml = (min)(abig, amed);
+  abig = (max)(abig, amed);
+  if(asml <= abig*relerr)
+    return abig;
+  else
+    return abig * internal::sqrt(RealScalar(1) + internal::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 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

densecrf/include/Eigen/src/Core/Stride.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// 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_STRIDE_H
+#define EIGEN_STRIDE_H
+
+namespace Eigen { 
+
+/** \class Stride
+  * \ingroup Core_Module
+  *
+  * \brief Holds strides information for Map
+  *
+  * This class holds the strides information for mapping arrays with strides with class Map.
+  *
+  * It holds two values: the inner stride and the outer stride.
+  *
+  * The inner stride is the pointer increment between two consecutive entries within a given row of a
+  * row-major matrix or within a given column of a column-major matrix.
+  *
+  * The outer stride is the pointer increment between two consecutive rows of a row-major matrix or
+  * between two consecutive columns of a column-major matrix.
+  *
+  * These two values can be passed either at compile-time as template parameters, or at runtime as
+  * arguments to the constructor.
+  *
+  * Indeed, this class takes two template parameters:
+  *  \param _OuterStrideAtCompileTime the outer stride, or Dynamic if you want to specify it at runtime.
+  *  \param _InnerStrideAtCompileTime the inner stride, or Dynamic if you want to specify it at runtime.
+  *
+  * Here is an example:
+  * \include Map_general_stride.cpp
+  * Output: \verbinclude Map_general_stride.out
+  *
+  * \sa class InnerStride, class OuterStride, \ref TopicStorageOrders
+  */
+template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime>
+class Stride
+{
+  public:
+    typedef DenseIndex Index;
+    enum {
+      InnerStrideAtCompileTime = _InnerStrideAtCompileTime,
+      OuterStrideAtCompileTime = _OuterStrideAtCompileTime
+    };
+
+    /** Default constructor, for use when strides are fixed at compile time */
+    Stride()
+      : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime)
+    {
+      eigen_assert(InnerStrideAtCompileTime != Dynamic && OuterStrideAtCompileTime != Dynamic);
+    }
+
+    /** Constructor allowing to pass the strides at runtime */
+    Stride(Index outerStride, Index innerStride)
+      : m_outer(outerStride), m_inner(innerStride)
+    {
+      eigen_assert(innerStride>=0 && outerStride>=0);
+    }
+
+    /** Copy constructor */
+    Stride(const Stride& other)
+      : m_outer(other.outer()), m_inner(other.inner())
+    {}
+
+    /** \returns the outer stride */
+    inline Index outer() const { return m_outer.value(); }
+    /** \returns the inner stride */
+    inline Index inner() const { return m_inner.value(); }
+
+  protected:
+    internal::variable_if_dynamic<Index, OuterStrideAtCompileTime> m_outer;
+    internal::variable_if_dynamic<Index, InnerStrideAtCompileTime> m_inner;
+};
+
+/** \brief Convenience specialization of Stride to specify only an inner stride
+  * See class Map for some examples */
+template<int Value = Dynamic>
+class InnerStride : public Stride<0, Value>
+{
+    typedef Stride<0, Value> Base;
+  public:
+    typedef DenseIndex Index;
+    InnerStride() : Base() {}
+    InnerStride(Index v) : Base(0, v) {}
+};
+
+/** \brief Convenience specialization of Stride to specify only an outer stride
+  * See class Map for some examples */
+template<int Value = Dynamic>
+class OuterStride : public Stride<Value, 0>
+{
+    typedef Stride<Value, 0> Base;
+  public:
+    typedef DenseIndex Index;
+    OuterStride() : Base() {}
+    OuterStride(Index v) : Base(v,0) {}
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_STRIDE_H

densecrf/include/Eigen/src/Core/Swap.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// 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_SWAP_H
+#define EIGEN_SWAP_H
+
+namespace Eigen { 
+
+/** \class SwapWrapper
+  * \ingroup Core_Module
+  *
+  * \internal
+  *
+  * \brief Internal helper class for swapping two expressions
+  */
+namespace internal {
+template<typename ExpressionType>
+struct traits<SwapWrapper<ExpressionType> > : traits<ExpressionType> {};
+}
+
+template<typename ExpressionType> class SwapWrapper
+  : public internal::dense_xpr_base<SwapWrapper<ExpressionType> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<SwapWrapper>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper)
+    typedef typename internal::packet_traits<Scalar>::type Packet;
+
+    inline SwapWrapper(ExpressionType& xpr) : m_expression(xpr) {}
+
+    inline Index rows() const { return m_expression.rows(); }
+    inline Index cols() const { return m_expression.cols(); }
+    inline Index outerStride() const { return m_expression.outerStride(); }
+    inline Index innerStride() const { return m_expression.innerStride(); }
+    
+    typedef typename internal::conditional<
+                       internal::is_lvalue<ExpressionType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+                     
+    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
+    inline const Scalar* data() const { return m_expression.data(); }
+
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      return m_expression.const_cast_derived().coeffRef(row, col);
+    }
+
+    inline Scalar& coeffRef(Index index)
+    {
+      return m_expression.const_cast_derived().coeffRef(index);
+    }
+
+    inline Scalar& coeffRef(Index row, Index col) const
+    {
+      return m_expression.coeffRef(row, col);
+    }
+
+    inline Scalar& coeffRef(Index index) const
+    {
+      return m_expression.coeffRef(index);
+    }
+
+    template<typename OtherDerived>
+    void copyCoeff(Index row, Index col, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(row >= 0 && row < rows()
+                         && col >= 0 && col < cols());
+      Scalar tmp = m_expression.coeff(row, col);
+      m_expression.coeffRef(row, col) = _other.coeff(row, col);
+      _other.coeffRef(row, col) = tmp;
+    }
+
+    template<typename OtherDerived>
+    void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(index >= 0 && index < m_expression.size());
+      Scalar tmp = m_expression.coeff(index);
+      m_expression.coeffRef(index) = _other.coeff(index);
+      _other.coeffRef(index) = tmp;
+    }
+
+    template<typename OtherDerived, int StoreMode, int LoadMode>
+    void copyPacket(Index row, Index col, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(row >= 0 && row < rows()
+                        && col >= 0 && col < cols());
+      Packet tmp = m_expression.template packet<StoreMode>(row, col);
+      m_expression.template writePacket<StoreMode>(row, col,
+        _other.template packet<LoadMode>(row, col)
+      );
+      _other.template writePacket<LoadMode>(row, col, tmp);
+    }
+
+    template<typename OtherDerived, int StoreMode, int LoadMode>
+    void copyPacket(Index index, const DenseBase<OtherDerived>& other)
+    {
+      OtherDerived& _other = other.const_cast_derived();
+      eigen_internal_assert(index >= 0 && index < m_expression.size());
+      Packet tmp = m_expression.template packet<StoreMode>(index);
+      m_expression.template writePacket<StoreMode>(index,
+        _other.template packet<LoadMode>(index)
+      );
+      _other.template writePacket<LoadMode>(index, tmp);
+    }
+
+    ExpressionType& expression() const { return m_expression; }
+
+  protected:
+    ExpressionType& m_expression;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SWAP_H

densecrf/include/Eigen/src/Core/Transpose.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009-2010 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_TRANSPOSE_H
+#define EIGEN_TRANSPOSE_H
+
+namespace Eigen { 
+
+/** \class Transpose
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the transpose of a matrix
+  *
+  * \param MatrixType the type of the object of which we are taking the transpose
+  *
+  * This class represents an expression of the transpose of a matrix.
+  * It is the return type of MatrixBase::transpose() and MatrixBase::adjoint()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::transpose(), MatrixBase::adjoint()
+  */
+
+namespace internal {
+template<typename MatrixType>
+struct traits<Transpose<MatrixType> > : traits<MatrixType>
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain;
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename traits<MatrixType>::XprKind XprKind;
+  enum {
+    RowsAtCompileTime = MatrixType::ColsAtCompileTime,
+    ColsAtCompileTime = MatrixType::RowsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit),
+    Flags1 = Flags0 | FlagsLvalueBit,
+    Flags = Flags1 ^ RowMajorBit,
+    CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost,
+    InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret,
+    OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret
+  };
+};
+}
+
+template<typename MatrixType, typename StorageKind> class TransposeImpl;
+
+template<typename MatrixType> class Transpose
+  : public TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>
+{
+  public:
+
+    typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose)
+
+    inline Transpose(MatrixType& matrix) : m_matrix(matrix) {}
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose)
+
+    inline Index rows() const { return m_matrix.cols(); }
+    inline Index cols() const { return m_matrix.rows(); }
+
+    /** \returns the nested expression */
+    const typename internal::remove_all<typename MatrixType::Nested>::type&
+    nestedExpression() const { return m_matrix; }
+
+    /** \returns the nested expression */
+    typename internal::remove_all<typename MatrixType::Nested>::type&
+    nestedExpression() { return m_matrix.const_cast_derived(); }
+
+  protected:
+    typename MatrixType::Nested m_matrix;
+};
+
+namespace internal {
+
+template<typename MatrixType, bool HasDirectAccess = has_direct_access<MatrixType>::ret>
+struct TransposeImpl_base
+{
+  typedef typename dense_xpr_base<Transpose<MatrixType> >::type type;
+};
+
+template<typename MatrixType>
+struct TransposeImpl_base<MatrixType, false>
+{
+  typedef typename dense_xpr_base<Transpose<MatrixType> >::type type;
+};
+
+} // end namespace internal
+
+template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
+  : public internal::TransposeImpl_base<MatrixType>::type
+{
+  public:
+
+    typedef typename internal::TransposeImpl_base<MatrixType>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>)
+
+    inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
+    inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+
+    typedef typename internal::conditional<
+                       internal::is_lvalue<MatrixType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+
+    inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
+    inline const Scalar* data() const { return derived().nestedExpression().data(); }
+
+    inline ScalarWithConstIfNotLvalue& coeffRef(Index row, Index col)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      return derived().nestedExpression().const_cast_derived().coeffRef(col, row);
+    }
+
+    inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      return derived().nestedExpression().const_cast_derived().coeffRef(index);
+    }
+
+    inline const Scalar& coeffRef(Index row, Index col) const
+    {
+      return derived().nestedExpression().coeffRef(col, row);
+    }
+
+    inline const Scalar& coeffRef(Index index) const
+    {
+      return derived().nestedExpression().coeffRef(index);
+    }
+
+    inline CoeffReturnType coeff(Index row, Index col) const
+    {
+      return derived().nestedExpression().coeff(col, row);
+    }
+
+    inline CoeffReturnType coeff(Index index) const
+    {
+      return derived().nestedExpression().coeff(index);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index row, Index col) const
+    {
+      return derived().nestedExpression().template packet<LoadMode>(col, row);
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index row, Index col, const PacketScalar& x)
+    {
+      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(col, row, x);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index index) const
+    {
+      return derived().nestedExpression().template packet<LoadMode>(index);
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index index, const PacketScalar& x)
+    {
+      derived().nestedExpression().const_cast_derived().template writePacket<LoadMode>(index, x);
+    }
+};
+
+/** \returns an expression of the transpose of *this.
+  *
+  * Example: \include MatrixBase_transpose.cpp
+  * Output: \verbinclude MatrixBase_transpose.out
+  *
+  * \warning If you want to replace a matrix by its own transpose, do \b NOT do this:
+  * \code
+  * m = m.transpose(); // bug!!! caused by aliasing effect
+  * \endcode
+  * Instead, use the transposeInPlace() method:
+  * \code
+  * m.transposeInPlace();
+  * \endcode
+  * which gives Eigen good opportunities for optimization, or alternatively you can also do:
+  * \code
+  * m = m.transpose().eval();
+  * \endcode
+  *
+  * \sa transposeInPlace(), adjoint() */
+template<typename Derived>
+inline Transpose<Derived>
+DenseBase<Derived>::transpose()
+{
+  return derived();
+}
+
+/** This is the const version of transpose().
+  *
+  * Make sure you read the warning for transpose() !
+  *
+  * \sa transposeInPlace(), adjoint() */
+template<typename Derived>
+inline const typename DenseBase<Derived>::ConstTransposeReturnType
+DenseBase<Derived>::transpose() const
+{
+  return ConstTransposeReturnType(derived());
+}
+
+/** \returns an expression of the adjoint (i.e. conjugate transpose) of *this.
+  *
+  * Example: \include MatrixBase_adjoint.cpp
+  * Output: \verbinclude MatrixBase_adjoint.out
+  *
+  * \warning If you want to replace a matrix by its own adjoint, do \b NOT do this:
+  * \code
+  * m = m.adjoint(); // bug!!! caused by aliasing effect
+  * \endcode
+  * Instead, use the adjointInPlace() method:
+  * \code
+  * m.adjointInPlace();
+  * \endcode
+  * which gives Eigen good opportunities for optimization, or alternatively you can also do:
+  * \code
+  * m = m.adjoint().eval();
+  * \endcode
+  *
+  * \sa adjointInPlace(), transpose(), conjugate(), class Transpose, class internal::scalar_conjugate_op */
+template<typename Derived>
+inline const typename MatrixBase<Derived>::AdjointReturnType
+MatrixBase<Derived>::adjoint() const
+{
+  return this->transpose(); // in the complex case, the .conjugate() is be implicit here
+                            // due to implicit conversion to return type
+}
+
+/***************************************************************************
+* "in place" transpose implementation
+***************************************************************************/
+
+namespace internal {
+
+template<typename MatrixType,
+  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic>
+struct inplace_transpose_selector;
+
+template<typename MatrixType>
+struct inplace_transpose_selector<MatrixType,true> { // square matrix
+  static void run(MatrixType& m) {
+    m.template triangularView<StrictlyUpper>().swap(m.transpose());
+  }
+};
+
+template<typename MatrixType>
+struct inplace_transpose_selector<MatrixType,false> { // non square matrix
+  static void run(MatrixType& m) {
+    if (m.rows()==m.cols())
+      m.template triangularView<StrictlyUpper>().swap(m.transpose());
+    else
+      m = m.transpose().eval();
+  }
+};
+
+} // end namespace internal
+
+/** This is the "in place" version of transpose(): it replaces \c *this by its own transpose.
+  * Thus, doing
+  * \code
+  * m.transposeInPlace();
+  * \endcode
+  * has the same effect on m as doing
+  * \code
+  * m = m.transpose().eval();
+  * \endcode
+  * and is faster and also safer because in the latter line of code, forgetting the eval() results
+  * in a bug caused by aliasing.
+  *
+  * Notice however that this method is only useful if you want to replace a matrix by its own transpose.
+  * If you just need the transpose of a matrix, use transpose().
+  *
+  * \note if the matrix is not square, then \c *this must be a resizable matrix.
+  *
+  * \sa transpose(), adjoint(), adjointInPlace() */
+template<typename Derived>
+inline void DenseBase<Derived>::transposeInPlace()
+{
+  internal::inplace_transpose_selector<Derived>::run(derived());
+}
+
+/***************************************************************************
+* "in place" adjoint implementation
+***************************************************************************/
+
+/** This is the "in place" version of adjoint(): it replaces \c *this by its own transpose.
+  * Thus, doing
+  * \code
+  * m.adjointInPlace();
+  * \endcode
+  * has the same effect on m as doing
+  * \code
+  * m = m.adjoint().eval();
+  * \endcode
+  * and is faster and also safer because in the latter line of code, forgetting the eval() results
+  * in a bug caused by aliasing.
+  *
+  * Notice however that this method is only useful if you want to replace a matrix by its own adjoint.
+  * If you just need the adjoint of a matrix, use adjoint().
+  *
+  * \note if the matrix is not square, then \c *this must be a resizable matrix.
+  *
+  * \sa transpose(), adjoint(), transposeInPlace() */
+template<typename Derived>
+inline void MatrixBase<Derived>::adjointInPlace()
+{
+  derived() = adjoint().eval();
+}
+
+#ifndef EIGEN_NO_DEBUG
+
+// The following is to detect aliasing problems in most common cases.
+
+namespace internal {
+
+template<typename BinOp,typename NestedXpr,typename Rhs>
+struct blas_traits<SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> >
+ : blas_traits<NestedXpr>
+{
+  typedef SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> XprType;
+  static inline const XprType extract(const XprType& x) { return x; }
+};
+
+template<bool DestIsTransposed, typename OtherDerived>
+struct check_transpose_aliasing_compile_time_selector
+{
+  enum { ret = bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed };
+};
+
+template<bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB>
+struct check_transpose_aliasing_compile_time_selector<DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> >
+{
+  enum { ret =    bool(blas_traits<DerivedA>::IsTransposed) != DestIsTransposed
+               || bool(blas_traits<DerivedB>::IsTransposed) != DestIsTransposed
+  };
+};
+
+template<typename Scalar, bool DestIsTransposed, typename OtherDerived>
+struct check_transpose_aliasing_run_time_selector
+{
+  static bool run(const Scalar* dest, const OtherDerived& src)
+  {
+    return (bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src));
+  }
+};
+
+template<typename Scalar, bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB>
+struct check_transpose_aliasing_run_time_selector<Scalar,DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> >
+{
+  static bool run(const Scalar* dest, const CwiseBinaryOp<BinOp,DerivedA,DerivedB>& src)
+  {
+    return ((blas_traits<DerivedA>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.lhs())))
+        || ((blas_traits<DerivedB>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.rhs())));
+  }
+};
+
+// the following selector, checkTransposeAliasing_impl, based on MightHaveTransposeAliasing,
+// is because when the condition controlling the assert is known at compile time, ICC emits a warning.
+// This is actually a good warning: in expressions that don't have any transposing, the condition is
+// known at compile time to be false, and using that, we can avoid generating the code of the assert again
+// and again for all these expressions that don't need it.
+
+template<typename Derived, typename OtherDerived,
+         bool MightHaveTransposeAliasing
+                 = check_transpose_aliasing_compile_time_selector
+                     <blas_traits<Derived>::IsTransposed,OtherDerived>::ret
+        >
+struct checkTransposeAliasing_impl
+{
+    static void run(const Derived& dst, const OtherDerived& other)
+    {
+        eigen_assert((!check_transpose_aliasing_run_time_selector
+                      <typename Derived::Scalar,blas_traits<Derived>::IsTransposed,OtherDerived>
+                      ::run(extract_data(dst), other))
+          && "aliasing detected during tranposition, use transposeInPlace() "
+             "or evaluate the rhs into a temporary using .eval()");
+
+    }
+};
+
+template<typename Derived, typename OtherDerived>
+struct checkTransposeAliasing_impl<Derived, OtherDerived, false>
+{
+    static void run(const Derived&, const OtherDerived&)
+    {
+    }
+};
+
+} // end namespace internal
+
+template<typename Derived>
+template<typename OtherDerived>
+void DenseBase<Derived>::checkTransposeAliasing(const OtherDerived& other) const
+{
+    internal::checkTransposeAliasing_impl<Derived, OtherDerived>::run(derived(), other);
+}
+#endif
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRANSPOSE_H

densecrf/include/Eigen/src/Core/Transpositions.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010-2011 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_TRANSPOSITIONS_H
+#define EIGEN_TRANSPOSITIONS_H
+
+namespace Eigen { 
+
+/** \class Transpositions
+  * \ingroup Core_Module
+  *
+  * \brief Represents a sequence of transpositions (row/column interchange)
+  *
+  * \param SizeAtCompileTime the number of transpositions, or Dynamic
+  * \param MaxSizeAtCompileTime the maximum number of transpositions, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
+  *
+  * This class represents a permutation transformation as a sequence of \em n transpositions
+  * \f$[T_{n-1} \ldots T_{i} \ldots T_{0}]\f$. It is internally stored as a vector of integers \c indices.
+  * Each transposition \f$ T_{i} \f$ applied on the left of a matrix (\f$ T_{i} M\f$) interchanges
+  * the rows \c i and \c indices[i] of the matrix \c M.
+  * A transposition applied on the right (e.g., \f$ M T_{i}\f$) yields a column interchange.
+  *
+  * Compared to the class PermutationMatrix, such a sequence of transpositions is what is
+  * computed during a decomposition with pivoting, and it is faster when applying the permutation in-place.
+  * 
+  * To apply a sequence of transpositions to a matrix, simply use the operator * as in the following example:
+  * \code
+  * Transpositions tr;
+  * MatrixXf mat;
+  * mat = tr * mat;
+  * \endcode
+  * In this example, we detect that the matrix appears on both side, and so the transpositions
+  * are applied in-place without any temporary or extra copy.
+  *
+  * \sa class PermutationMatrix
+  */
+
+namespace internal {
+template<typename TranspositionType, typename MatrixType, int Side, bool Transposed=false> struct transposition_matrix_product_retval;
+}
+
+template<typename Derived>
+class TranspositionsBase
+{
+    typedef internal::traits<Derived> Traits;
+    
+  public:
+
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar Index;
+
+    Derived& derived() { return *static_cast<Derived*>(this); }
+    const Derived& derived() const { return *static_cast<const Derived*>(this); }
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    Derived& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      indices() = other.indices();
+      return derived();
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Derived& operator=(const TranspositionsBase& other)
+    {
+      indices() = other.indices();
+      return derived();
+    }
+    #endif
+
+    /** \returns the number of transpositions */
+    inline Index size() const { return indices().size(); }
+
+    /** Direct access to the underlying index vector */
+    inline const Index& coeff(Index i) const { return indices().coeff(i); }
+    /** Direct access to the underlying index vector */
+    inline Index& coeffRef(Index i) { return indices().coeffRef(i); }
+    /** Direct access to the underlying index vector */
+    inline const Index& operator()(Index i) const { return indices()(i); }
+    /** Direct access to the underlying index vector */
+    inline Index& operator()(Index i) { return indices()(i); }
+    /** Direct access to the underlying index vector */
+    inline const Index& operator[](Index i) const { return indices()(i); }
+    /** Direct access to the underlying index vector */
+    inline Index& operator[](Index i) { return indices()(i); }
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return derived().indices(); }
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return derived().indices(); }
+
+    /** Resizes to given size. */
+    inline void resize(int size)
+    {
+      indices().resize(size);
+    }
+
+    /** Sets \c *this to represents an identity transformation */
+    void setIdentity()
+    {
+      for(int i = 0; i < indices().size(); ++i)
+        coeffRef(i) = i;
+    }
+
+    // FIXME: do we want such methods ?
+    // might be usefull when the target matrix expression is complex, e.g.:
+    // object.matrix().block(..,..,..,..) = trans * object.matrix().block(..,..,..,..);
+    /*
+    template<typename MatrixType>
+    void applyForwardToRows(MatrixType& mat) const
+    {
+      for(Index k=0 ; k<size() ; ++k)
+        if(m_indices(k)!=k)
+          mat.row(k).swap(mat.row(m_indices(k)));
+    }
+
+    template<typename MatrixType>
+    void applyBackwardToRows(MatrixType& mat) const
+    {
+      for(Index k=size()-1 ; k>=0 ; --k)
+        if(m_indices(k)!=k)
+          mat.row(k).swap(mat.row(m_indices(k)));
+    }
+    */
+
+    /** \returns the inverse transformation */
+    inline Transpose<TranspositionsBase> inverse() const
+    { return Transpose<TranspositionsBase>(derived()); }
+
+    /** \returns the tranpose transformation */
+    inline Transpose<TranspositionsBase> transpose() const
+    { return Transpose<TranspositionsBase>(derived()); }
+
+  protected:
+};
+
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
+struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType> >
+{
+  typedef IndexType Index;
+  typedef Matrix<Index, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+};
+}
+
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType>
+class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType> >
+{
+    typedef internal::traits<Transpositions> Traits;
+  public:
+
+    typedef TranspositionsBase<Transpositions> Base;
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar Index;
+
+    inline Transpositions() {}
+
+    /** Copy constructor. */
+    template<typename OtherDerived>
+    inline Transpositions(const TranspositionsBase<OtherDerived>& other)
+      : m_indices(other.indices()) {}
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** Standard copy constructor. Defined only to prevent a default copy constructor
+      * from hiding the other templated constructor */
+    inline Transpositions(const Transpositions& other) : m_indices(other.indices()) {}
+    #endif
+
+    /** Generic constructor from expression of the transposition indices. */
+    template<typename Other>
+    explicit inline Transpositions(const MatrixBase<Other>& indices) : m_indices(indices)
+    {}
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    Transpositions& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      return Base::operator=(other);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Transpositions& operator=(const Transpositions& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** Constructs an uninitialized permutation matrix of given size.
+      */
+    inline Transpositions(Index size) : m_indices(size)
+    {}
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    IndicesType m_indices;
+};
+
+
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int _PacketAccess>
+struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,_PacketAccess> >
+{
+  typedef IndexType Index;
+  typedef Map<const Matrix<Index,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+};
+}
+
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename IndexType, int PacketAccess>
+class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,PacketAccess>
+ : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,IndexType>,PacketAccess> >
+{
+    typedef internal::traits<Map> Traits;
+  public:
+
+    typedef TranspositionsBase<Map> Base;
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar Index;
+
+    inline Map(const Index* indices)
+      : m_indices(indices)
+    {}
+
+    inline Map(const Index* indices, Index size)
+      : m_indices(indices,size)
+    {}
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    Map& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      return Base::operator=(other);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Map& operator=(const Map& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+    
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    IndicesType m_indices;
+};
+
+namespace internal {
+template<typename _IndicesType>
+struct traits<TranspositionsWrapper<_IndicesType> >
+{
+  typedef typename _IndicesType::Scalar Index;
+  typedef _IndicesType IndicesType;
+};
+}
+
+template<typename _IndicesType>
+class TranspositionsWrapper
+ : public TranspositionsBase<TranspositionsWrapper<_IndicesType> >
+{
+    typedef internal::traits<TranspositionsWrapper> Traits;
+  public:
+
+    typedef TranspositionsBase<TranspositionsWrapper> Base;
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar Index;
+
+    inline TranspositionsWrapper(IndicesType& indices)
+      : m_indices(indices)
+    {}
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    TranspositionsWrapper& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      return Base::operator=(other);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    TranspositionsWrapper& operator=(const TranspositionsWrapper& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    const typename IndicesType::Nested m_indices;
+};
+
+/** \returns the \a matrix with the \a transpositions applied to the columns.
+  */
+template<typename Derived, typename TranspositionsDerived>
+inline const internal::transposition_matrix_product_retval<TranspositionsDerived, Derived, OnTheRight>
+operator*(const MatrixBase<Derived>& matrix,
+          const TranspositionsBase<TranspositionsDerived> &transpositions)
+{
+  return internal::transposition_matrix_product_retval
+           <TranspositionsDerived, Derived, OnTheRight>
+           (transpositions.derived(), matrix.derived());
+}
+
+/** \returns the \a matrix with the \a transpositions applied to the rows.
+  */
+template<typename Derived, typename TranspositionDerived>
+inline const internal::transposition_matrix_product_retval
+               <TranspositionDerived, Derived, OnTheLeft>
+operator*(const TranspositionsBase<TranspositionDerived> &transpositions,
+          const MatrixBase<Derived>& matrix)
+{
+  return internal::transposition_matrix_product_retval
+           <TranspositionDerived, Derived, OnTheLeft>
+           (transpositions.derived(), matrix.derived());
+}
+
+namespace internal {
+
+template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
+struct traits<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
+{
+  typedef typename MatrixType::PlainObject ReturnType;
+};
+
+template<typename TranspositionType, typename MatrixType, int Side, bool Transposed>
+struct transposition_matrix_product_retval
+ : public ReturnByValue<transposition_matrix_product_retval<TranspositionType, MatrixType, Side, Transposed> >
+{
+    typedef typename remove_all<typename MatrixType::Nested>::type MatrixTypeNestedCleaned;
+    typedef typename TranspositionType::Index Index;
+
+    transposition_matrix_product_retval(const TranspositionType& tr, const MatrixType& matrix)
+      : m_transpositions(tr), m_matrix(matrix)
+    {}
+
+    inline int rows() const { return m_matrix.rows(); }
+    inline int cols() const { return m_matrix.cols(); }
+
+    template<typename Dest> inline void evalTo(Dest& dst) const
+    {
+      const int size = m_transpositions.size();
+      Index j = 0;
+
+      if(!(is_same<MatrixTypeNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_matrix)))
+        dst = m_matrix;
+
+      for(int k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)
+        if((j=m_transpositions.coeff(k))!=k)
+        {
+          if(Side==OnTheLeft)
+            dst.row(k).swap(dst.row(j));
+          else if(Side==OnTheRight)
+            dst.col(k).swap(dst.col(j));
+        }
+    }
+
+  protected:
+    const TranspositionType& m_transpositions;
+    typename MatrixType::Nested m_matrix;
+};
+
+} // end namespace internal
+
+/* Template partial specialization for transposed/inverse transpositions */
+
+template<typename TranspositionsDerived>
+class Transpose<TranspositionsBase<TranspositionsDerived> >
+{
+    typedef TranspositionsDerived TranspositionType;
+    typedef typename TranspositionType::IndicesType IndicesType;
+  public:
+
+    Transpose(const TranspositionType& t) : m_transpositions(t) {}
+
+    inline int size() const { return m_transpositions.size(); }
+
+    /** \returns the \a matrix with the inverse transpositions applied to the columns.
+      */
+    template<typename Derived> friend
+    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>
+    operator*(const MatrixBase<Derived>& matrix, const Transpose& trt)
+    {
+      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheRight, true>(trt.m_transpositions, matrix.derived());
+    }
+
+    /** \returns the \a matrix with the inverse transpositions applied to the rows.
+      */
+    template<typename Derived>
+    inline const internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>
+    operator*(const MatrixBase<Derived>& matrix) const
+    {
+      return internal::transposition_matrix_product_retval<TranspositionType, Derived, OnTheLeft, true>(m_transpositions, matrix.derived());
+    }
+
+  protected:
+    const TranspositionType& m_transpositions;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRANSPOSITIONS_H

densecrf/include/Eigen/src/Core/TriangularMatrix.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// 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_TRIANGULARMATRIX_H
+#define EIGEN_TRIANGULARMATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+  
+template<int Side, typename TriangularType, typename Rhs> struct triangular_solve_retval;
+  
+}
+
+/** \internal
+  *
+  * \class TriangularBase
+  * \ingroup Core_Module
+  *
+  * \brief Base class for triangular part in a matrix
+  */
+template<typename Derived> class TriangularBase : public EigenBase<Derived>
+{
+  public:
+
+    enum {
+      Mode = internal::traits<Derived>::Mode,
+      CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime
+    };
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::Index Index;
+    typedef typename internal::traits<Derived>::DenseMatrixType DenseMatrixType;
+    typedef DenseMatrixType DenseType;
+
+    inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); }
+
+    inline Index rows() const { return derived().rows(); }
+    inline Index cols() const { return derived().cols(); }
+    inline Index outerStride() const { return derived().outerStride(); }
+    inline Index innerStride() const { return derived().innerStride(); }
+
+    inline Scalar coeff(Index row, Index col) const  { return derived().coeff(row,col); }
+    inline Scalar& coeffRef(Index row, Index col) { return derived().coeffRef(row,col); }
+
+    /** \see MatrixBase::copyCoeff(row,col)
+      */
+    template<typename Other>
+    EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, Other& other)
+    {
+      derived().coeffRef(row, col) = other.coeff(row, col);
+    }
+
+    inline Scalar operator()(Index row, Index col) const
+    {
+      check_coordinates(row, col);
+      return coeff(row,col);
+    }
+    inline Scalar& operator()(Index row, Index col)
+    {
+      check_coordinates(row, col);
+      return coeffRef(row,col);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    inline Derived& derived() { return *static_cast<Derived*>(this); }
+    #endif // not EIGEN_PARSED_BY_DOXYGEN
+
+    template<typename DenseDerived>
+    void evalTo(MatrixBase<DenseDerived> &other) const;
+    template<typename DenseDerived>
+    void evalToLazy(MatrixBase<DenseDerived> &other) const;
+
+    DenseMatrixType toDenseMatrix() const
+    {
+      DenseMatrixType res(rows(), cols());
+      evalToLazy(res);
+      return res;
+    }
+
+  protected:
+
+    void check_coordinates(Index row, Index col) const
+    {
+      EIGEN_ONLY_USED_FOR_DEBUG(row);
+      EIGEN_ONLY_USED_FOR_DEBUG(col);
+      eigen_assert(col>=0 && col<cols() && row>=0 && row<rows());
+      const int mode = int(Mode) & ~SelfAdjoint;
+      EIGEN_ONLY_USED_FOR_DEBUG(mode);
+      eigen_assert((mode==Upper && col>=row)
+                || (mode==Lower && col<=row)
+                || ((mode==StrictlyUpper || mode==UnitUpper) && col>row)
+                || ((mode==StrictlyLower || mode==UnitLower) && col<row));
+    }
+
+    #ifdef EIGEN_INTERNAL_DEBUGGING
+    void check_coordinates_internal(Index row, Index col) const
+    {
+      check_coordinates(row, col);
+    }
+    #else
+    void check_coordinates_internal(Index , Index ) const {}
+    #endif
+
+};
+
+/** \class TriangularView
+  * \ingroup Core_Module
+  *
+  * \brief Base class for triangular part in a matrix
+  *
+  * \param MatrixType the type of the object in which we are taking the triangular part
+  * \param Mode the kind of triangular matrix expression to construct. Can be #Upper,
+  *             #Lower, #UnitUpper, #UnitLower, #StrictlyUpper, or #StrictlyLower.
+  *             This is in fact a bit field; it must have either #Upper or #Lower, 
+  *             and additionnaly it may have #UnitDiag or #ZeroDiag or neither.
+  *
+  * This class represents a triangular part of a matrix, not necessarily square. Strictly speaking, for rectangular
+  * matrices one should speak of "trapezoid" parts. This class is the return type
+  * of MatrixBase::triangularView() and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::triangularView()
+  */
+namespace internal {
+template<typename MatrixType, unsigned int _Mode>
+struct traits<TriangularView<MatrixType, _Mode> > : traits<MatrixType>
+{
+  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedNonRef;
+  typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
+  typedef MatrixType ExpressionType;
+  typedef typename MatrixType::PlainObject DenseMatrixType;
+  enum {
+    Mode = _Mode,
+    Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode,
+    CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost
+  };
+};
+}
+
+template<int Mode, bool LhsIsTriangular,
+         typename Lhs, bool LhsIsVector,
+         typename Rhs, bool RhsIsVector>
+struct TriangularProduct;
+
+template<typename _MatrixType, unsigned int _Mode> class TriangularView
+  : public TriangularBase<TriangularView<_MatrixType, _Mode> >
+{
+  public:
+
+    typedef TriangularBase<TriangularView> Base;
+    typedef typename internal::traits<TriangularView>::Scalar Scalar;
+
+    typedef _MatrixType MatrixType;
+    typedef typename internal::traits<TriangularView>::DenseMatrixType DenseMatrixType;
+    typedef DenseMatrixType PlainObject;
+
+  protected:
+    typedef typename internal::traits<TriangularView>::MatrixTypeNested MatrixTypeNested;
+    typedef typename internal::traits<TriangularView>::MatrixTypeNestedNonRef MatrixTypeNestedNonRef;
+    typedef typename internal::traits<TriangularView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
+
+    typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType;
+    
+  public:
+    using Base::evalToLazy;
+  
+
+    typedef typename internal::traits<TriangularView>::StorageKind StorageKind;
+    typedef typename internal::traits<TriangularView>::Index Index;
+
+    enum {
+      Mode = _Mode,
+      TransposeMode = (Mode & Upper ? Lower : 0)
+                    | (Mode & Lower ? Upper : 0)
+                    | (Mode & (UnitDiag))
+                    | (Mode & (ZeroDiag))
+    };
+
+    inline TriangularView(const MatrixType& matrix) : m_matrix(matrix)
+    {}
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+    inline Index outerStride() const { return m_matrix.outerStride(); }
+    inline Index innerStride() const { return m_matrix.innerStride(); }
+
+    /** \sa MatrixBase::operator+=() */
+    template<typename Other> TriangularView&  operator+=(const DenseBase<Other>& other) { return *this = m_matrix + other.derived(); }
+    /** \sa MatrixBase::operator-=() */
+    template<typename Other> TriangularView&  operator-=(const DenseBase<Other>& other) { return *this = m_matrix - other.derived(); }
+    /** \sa MatrixBase::operator*=() */
+    TriangularView&  operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = m_matrix * other; }
+    /** \sa MatrixBase::operator/=() */
+    TriangularView&  operator/=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = m_matrix / other; }
+
+    /** \sa MatrixBase::fill() */
+    void fill(const Scalar& value) { setConstant(value); }
+    /** \sa MatrixBase::setConstant() */
+    TriangularView& setConstant(const Scalar& value)
+    { return *this = MatrixType::Constant(rows(), cols(), value); }
+    /** \sa MatrixBase::setZero() */
+    TriangularView& setZero() { return setConstant(Scalar(0)); }
+    /** \sa MatrixBase::setOnes() */
+    TriangularView& setOnes() { return setConstant(Scalar(1)); }
+
+    /** \sa MatrixBase::coeff()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    inline Scalar coeff(Index row, Index col) const
+    {
+      Base::check_coordinates_internal(row, col);
+      return m_matrix.coeff(row, col);
+    }
+
+    /** \sa MatrixBase::coeffRef()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      Base::check_coordinates_internal(row, col);
+      return m_matrix.const_cast_derived().coeffRef(row, col);
+    }
+
+    const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; }
+    MatrixTypeNestedCleaned& nestedExpression() { return *const_cast<MatrixTypeNestedCleaned*>(&m_matrix); }
+
+    /** Assigns a triangular matrix to a triangular part of a dense matrix */
+    template<typename OtherDerived>
+    TriangularView& operator=(const TriangularBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    TriangularView& operator=(const MatrixBase<OtherDerived>& other);
+
+    TriangularView& operator=(const TriangularView& other)
+    { return *this = other.nestedExpression(); }
+
+    template<typename OtherDerived>
+    void lazyAssign(const TriangularBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    void lazyAssign(const MatrixBase<OtherDerived>& other);
+
+    /** \sa MatrixBase::conjugate() */
+    inline TriangularView<MatrixConjugateReturnType,Mode> conjugate()
+    { return m_matrix.conjugate(); }
+    /** \sa MatrixBase::conjugate() const */
+    inline const TriangularView<MatrixConjugateReturnType,Mode> conjugate() const
+    { return m_matrix.conjugate(); }
+
+    /** \sa MatrixBase::adjoint() const */
+    inline const TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> adjoint() const
+    { return m_matrix.adjoint(); }
+
+    /** \sa MatrixBase::transpose() */
+    inline TriangularView<Transpose<MatrixType>,TransposeMode> transpose()
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      return m_matrix.const_cast_derived().transpose();
+    }
+    /** \sa MatrixBase::transpose() const */
+    inline const TriangularView<Transpose<MatrixType>,TransposeMode> transpose() const
+    {
+      return m_matrix.transpose();
+    }
+
+    /** Efficient triangular matrix times vector/matrix product */
+    template<typename OtherDerived>
+    TriangularProduct<Mode,true,MatrixType,false,OtherDerived, OtherDerived::IsVectorAtCompileTime>
+    operator*(const MatrixBase<OtherDerived>& rhs) const
+    {
+      return TriangularProduct
+              <Mode,true,MatrixType,false,OtherDerived,OtherDerived::IsVectorAtCompileTime>
+              (m_matrix, rhs.derived());
+    }
+
+    /** Efficient vector/matrix times triangular matrix product */
+    template<typename OtherDerived> friend
+    TriangularProduct<Mode,false,OtherDerived,OtherDerived::IsVectorAtCompileTime,MatrixType,false>
+    operator*(const MatrixBase<OtherDerived>& lhs, const TriangularView& rhs)
+    {
+      return TriangularProduct
+              <Mode,false,OtherDerived,OtherDerived::IsVectorAtCompileTime,MatrixType,false>
+              (lhs.derived(),rhs.m_matrix);
+    }
+
+    #ifdef EIGEN2_SUPPORT
+    template<typename OtherDerived>
+    struct eigen2_product_return_type
+    {
+      typedef typename TriangularView<MatrixType,Mode>::DenseMatrixType DenseMatrixType;
+      typedef typename OtherDerived::PlainObject::DenseType OtherPlainObject;
+      typedef typename ProductReturnType<DenseMatrixType, OtherPlainObject>::Type ProdRetType;
+      typedef typename ProdRetType::PlainObject type;
+    };
+    template<typename OtherDerived>
+    const typename eigen2_product_return_type<OtherDerived>::type
+    operator*(const EigenBase<OtherDerived>& rhs) const
+    {
+      typename OtherDerived::PlainObject::DenseType rhsPlainObject;
+      rhs.evalTo(rhsPlainObject);
+      return this->toDenseMatrix() * rhsPlainObject;
+    }
+    template<typename OtherMatrixType>
+    bool isApprox(const TriangularView<OtherMatrixType, Mode>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
+    {
+      return this->toDenseMatrix().isApprox(other.toDenseMatrix(), precision);
+    }
+    template<typename OtherDerived>
+    bool isApprox(const MatrixBase<OtherDerived>& other, typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision()) const
+    {
+      return this->toDenseMatrix().isApprox(other, precision);
+    }
+    #endif // EIGEN2_SUPPORT
+
+    template<int Side, typename Other>
+    inline const internal::triangular_solve_retval<Side,TriangularView, Other>
+    solve(const MatrixBase<Other>& other) const;
+
+    template<int Side, typename OtherDerived>
+    void solveInPlace(const MatrixBase<OtherDerived>& other) const;
+
+    template<typename Other>
+    inline const internal::triangular_solve_retval<OnTheLeft,TriangularView, Other> 
+    solve(const MatrixBase<Other>& other) const
+    { return solve<OnTheLeft>(other); }
+
+    template<typename OtherDerived>
+    void solveInPlace(const MatrixBase<OtherDerived>& other) const
+    { return solveInPlace<OnTheLeft>(other); }
+
+    const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
+    {
+      EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+    SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView()
+    {
+      EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+
+    template<typename OtherDerived>
+    void swap(TriangularBase<OtherDerived> const & other)
+    {
+      TriangularView<SwapWrapper<MatrixType>,Mode>(const_cast<MatrixType&>(m_matrix)).lazyAssign(other.derived());
+    }
+
+    template<typename OtherDerived>
+    void swap(MatrixBase<OtherDerived> const & other)
+    {
+      SwapWrapper<MatrixType> swaper(const_cast<MatrixType&>(m_matrix));
+      TriangularView<SwapWrapper<MatrixType>,Mode>(swaper).lazyAssign(other.derived());
+    }
+
+    Scalar determinant() const
+    {
+      if (Mode & UnitDiag)
+        return 1;
+      else if (Mode & ZeroDiag)
+        return 0;
+      else
+        return m_matrix.diagonal().prod();
+    }
+    
+    // TODO simplify the following:
+    template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_STRONG_INLINE TriangularView& operator=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
+    {
+      setZero();
+      return assignProduct(other,1);
+    }
+    
+    template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_STRONG_INLINE TriangularView& operator+=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
+    {
+      return assignProduct(other,1);
+    }
+    
+    template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_STRONG_INLINE TriangularView& operator-=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
+    {
+      return assignProduct(other,-1);
+    }
+    
+    
+    template<typename ProductDerived>
+    EIGEN_STRONG_INLINE TriangularView& operator=(const ScaledProduct<ProductDerived>& other)
+    {
+      setZero();
+      return assignProduct(other,other.alpha());
+    }
+    
+    template<typename ProductDerived>
+    EIGEN_STRONG_INLINE TriangularView& operator+=(const ScaledProduct<ProductDerived>& other)
+    {
+      return assignProduct(other,other.alpha());
+    }
+    
+    template<typename ProductDerived>
+    EIGEN_STRONG_INLINE TriangularView& operator-=(const ScaledProduct<ProductDerived>& other)
+    {
+      return assignProduct(other,-other.alpha());
+    }
+    
+  protected:
+    
+    template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase<ProductDerived, Lhs,Rhs>& prod, const Scalar& alpha);
+
+    MatrixTypeNested m_matrix;
+};
+
+/***************************************************************************
+* Implementation of triangular evaluation/assignment
+***************************************************************************/
+
+namespace internal {
+
+template<typename Derived1, typename Derived2, unsigned int Mode, int UnrollCount, bool ClearOpposite>
+struct triangular_assignment_selector
+{
+  enum {
+    col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
+    row = (UnrollCount-1) % Derived1::RowsAtCompileTime
+  };
+  
+  typedef typename Derived1::Scalar Scalar;
+
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    triangular_assignment_selector<Derived1, Derived2, Mode, UnrollCount-1, ClearOpposite>::run(dst, src);
+
+    eigen_assert( Mode == Upper || Mode == Lower
+            || Mode == StrictlyUpper || Mode == StrictlyLower
+            || Mode == UnitUpper || Mode == UnitLower);
+    if((Mode == Upper && row <= col)
+    || (Mode == Lower && row >= col)
+    || (Mode == StrictlyUpper && row < col)
+    || (Mode == StrictlyLower && row > col)
+    || (Mode == UnitUpper && row < col)
+    || (Mode == UnitLower && row > col))
+      dst.copyCoeff(row, col, src);
+    else if(ClearOpposite)
+    {
+      if (Mode&UnitDiag && row==col)
+        dst.coeffRef(row, col) = Scalar(1);
+      else
+        dst.coeffRef(row, col) = Scalar(0);
+    }
+  }
+};
+
+// prevent buggy user code from causing an infinite recursion
+template<typename Derived1, typename Derived2, unsigned int Mode, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, Mode, 0, ClearOpposite>
+{
+  static inline void run(Derived1 &, const Derived2 &) {}
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, Upper, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  typedef typename Derived1::Scalar Scalar;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      Index maxi = (std::min)(j, dst.rows()-1);
+      for(Index i = 0; i <= maxi; ++i)
+        dst.copyCoeff(i, j, src);
+      if (ClearOpposite)
+        for(Index i = maxi+1; i < dst.rows(); ++i)
+          dst.coeffRef(i, j) = Scalar(0);
+    }
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, Lower, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      for(Index i = j; i < dst.rows(); ++i)
+        dst.copyCoeff(i, j, src);
+      Index maxi = (std::min)(j, dst.rows());
+      if (ClearOpposite)
+        for(Index i = 0; i < maxi; ++i)
+          dst.coeffRef(i, j) = static_cast<typename Derived1::Scalar>(0);
+    }
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, StrictlyUpper, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  typedef typename Derived1::Scalar Scalar;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      Index maxi = (std::min)(j, dst.rows());
+      for(Index i = 0; i < maxi; ++i)
+        dst.copyCoeff(i, j, src);
+      if (ClearOpposite)
+        for(Index i = maxi; i < dst.rows(); ++i)
+          dst.coeffRef(i, j) = Scalar(0);
+    }
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, StrictlyLower, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      for(Index i = j+1; i < dst.rows(); ++i)
+        dst.copyCoeff(i, j, src);
+      Index maxi = (std::min)(j, dst.rows()-1);
+      if (ClearOpposite)
+        for(Index i = 0; i <= maxi; ++i)
+          dst.coeffRef(i, j) = static_cast<typename Derived1::Scalar>(0);
+    }
+  }
+};
+
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, UnitUpper, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      Index maxi = (std::min)(j, dst.rows());
+      for(Index i = 0; i < maxi; ++i)
+        dst.copyCoeff(i, j, src);
+      if (ClearOpposite)
+      {
+        for(Index i = maxi+1; i < dst.rows(); ++i)
+          dst.coeffRef(i, j) = 0;
+      }
+    }
+    dst.diagonal().setOnes();
+  }
+};
+template<typename Derived1, typename Derived2, bool ClearOpposite>
+struct triangular_assignment_selector<Derived1, Derived2, UnitLower, Dynamic, ClearOpposite>
+{
+  typedef typename Derived1::Index Index;
+  static inline void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(Index j = 0; j < dst.cols(); ++j)
+    {
+      Index maxi = (std::min)(j, dst.rows());
+      for(Index i = maxi+1; i < dst.rows(); ++i)
+        dst.copyCoeff(i, j, src);
+      if (ClearOpposite)
+      {
+        for(Index i = 0; i < maxi; ++i)
+          dst.coeffRef(i, j) = 0;
+      }
+    }
+    dst.diagonal().setOnes();
+  }
+};
+
+} // end namespace internal
+
+// FIXME should we keep that possibility
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+inline TriangularView<MatrixType, Mode>&
+TriangularView<MatrixType, Mode>::operator=(const MatrixBase<OtherDerived>& other)
+{
+  if(OtherDerived::Flags & EvalBeforeAssigningBit)
+  {
+    typename internal::plain_matrix_type<OtherDerived>::type other_evaluated(other.rows(), other.cols());
+    other_evaluated.template triangularView<Mode>().lazyAssign(other.derived());
+    lazyAssign(other_evaluated);
+  }
+  else
+    lazyAssign(other.derived());
+  return *this;
+}
+
+// FIXME should we keep that possibility
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+void TriangularView<MatrixType, Mode>::lazyAssign(const MatrixBase<OtherDerived>& other)
+{
+  enum {
+    unroll = MatrixType::SizeAtCompileTime != Dynamic
+          && internal::traits<OtherDerived>::CoeffReadCost != Dynamic
+          && MatrixType::SizeAtCompileTime*internal::traits<OtherDerived>::CoeffReadCost/2 <= EIGEN_UNROLLING_LIMIT
+  };
+  eigen_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols());
+
+  internal::triangular_assignment_selector
+    <MatrixType, OtherDerived, int(Mode),
+    unroll ? int(MatrixType::SizeAtCompileTime) : Dynamic,
+    false // do not change the opposite triangular part
+    >::run(m_matrix.const_cast_derived(), other.derived());
+}
+
+
+
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+inline TriangularView<MatrixType, Mode>&
+TriangularView<MatrixType, Mode>::operator=(const TriangularBase<OtherDerived>& other)
+{
+  eigen_assert(Mode == int(OtherDerived::Mode));
+  if(internal::traits<OtherDerived>::Flags & EvalBeforeAssigningBit)
+  {
+    typename OtherDerived::DenseMatrixType other_evaluated(other.rows(), other.cols());
+    other_evaluated.template triangularView<Mode>().lazyAssign(other.derived().nestedExpression());
+    lazyAssign(other_evaluated);
+  }
+  else
+    lazyAssign(other.derived().nestedExpression());
+  return *this;
+}
+
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+void TriangularView<MatrixType, Mode>::lazyAssign(const TriangularBase<OtherDerived>& other)
+{
+  enum {
+    unroll = MatrixType::SizeAtCompileTime != Dynamic
+                   && internal::traits<OtherDerived>::CoeffReadCost != Dynamic
+                   && MatrixType::SizeAtCompileTime * internal::traits<OtherDerived>::CoeffReadCost / 2
+                        <= EIGEN_UNROLLING_LIMIT
+  };
+  eigen_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols());
+
+  internal::triangular_assignment_selector
+    <MatrixType, OtherDerived, int(Mode),
+    unroll ? int(MatrixType::SizeAtCompileTime) : Dynamic,
+    false // preserve the opposite triangular part
+    >::run(m_matrix.const_cast_derived(), other.derived().nestedExpression());
+}
+
+/***************************************************************************
+* Implementation of TriangularBase methods
+***************************************************************************/
+
+/** Assigns a triangular or selfadjoint matrix to a dense matrix.
+  * If the matrix is triangular, the opposite part is set to zero. */
+template<typename Derived>
+template<typename DenseDerived>
+void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
+{
+  if(internal::traits<Derived>::Flags & EvalBeforeAssigningBit)
+  {
+    typename internal::plain_matrix_type<Derived>::type other_evaluated(rows(), cols());
+    evalToLazy(other_evaluated);
+    other.derived().swap(other_evaluated);
+  }
+  else
+    evalToLazy(other.derived());
+}
+
+/** Assigns a triangular or selfadjoint matrix to a dense matrix.
+  * If the matrix is triangular, the opposite part is set to zero. */
+template<typename Derived>
+template<typename DenseDerived>
+void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
+{
+  enum {
+    unroll = DenseDerived::SizeAtCompileTime != Dynamic
+                   && internal::traits<Derived>::CoeffReadCost != Dynamic
+                   && DenseDerived::SizeAtCompileTime * internal::traits<Derived>::CoeffReadCost / 2
+                        <= EIGEN_UNROLLING_LIMIT
+  };
+  other.derived().resize(this->rows(), this->cols());
+
+  internal::triangular_assignment_selector
+    <DenseDerived, typename internal::traits<Derived>::MatrixTypeNestedCleaned, Derived::Mode,
+    unroll ? int(DenseDerived::SizeAtCompileTime) : Dynamic,
+    true // clear the opposite triangular part
+    >::run(other.derived(), derived().nestedExpression());
+}
+
+/***************************************************************************
+* Implementation of TriangularView methods
+***************************************************************************/
+
+/***************************************************************************
+* Implementation of MatrixBase methods
+***************************************************************************/
+
+#ifdef EIGEN2_SUPPORT
+
+// implementation of part<>(), including the SelfAdjoint case.
+
+namespace internal {
+template<typename MatrixType, unsigned int Mode>
+struct eigen2_part_return_type
+{
+  typedef TriangularView<MatrixType, Mode> type;
+};
+
+template<typename MatrixType>
+struct eigen2_part_return_type<MatrixType, SelfAdjoint>
+{
+  typedef SelfAdjointView<MatrixType, Upper> type;
+};
+}
+
+/** \deprecated use MatrixBase::triangularView() */
+template<typename Derived>
+template<unsigned int Mode>
+const typename internal::eigen2_part_return_type<Derived, Mode>::type MatrixBase<Derived>::part() const
+{
+  return derived();
+}
+
+/** \deprecated use MatrixBase::triangularView() */
+template<typename Derived>
+template<unsigned int Mode>
+typename internal::eigen2_part_return_type<Derived, Mode>::type MatrixBase<Derived>::part()
+{
+  return derived();
+}
+#endif
+
+/**
+  * \returns an expression of a triangular view extracted from the current matrix
+  *
+  * The parameter \a Mode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper,
+  * \c #Lower, \c #StrictlyLower, \c #UnitLower.
+  *
+  * Example: \include MatrixBase_extract.cpp
+  * Output: \verbinclude MatrixBase_extract.out
+  *
+  * \sa class TriangularView
+  */
+template<typename Derived>
+template<unsigned int Mode>
+typename MatrixBase<Derived>::template TriangularViewReturnType<Mode>::Type
+MatrixBase<Derived>::triangularView()
+{
+  return derived();
+}
+
+/** This is the const version of MatrixBase::triangularView() */
+template<typename Derived>
+template<unsigned int Mode>
+typename MatrixBase<Derived>::template ConstTriangularViewReturnType<Mode>::Type
+MatrixBase<Derived>::triangularView() const
+{
+  return derived();
+}
+
+/** \returns true if *this is approximately equal to an upper triangular matrix,
+  *          within the precision given by \a prec.
+  *
+  * \sa isLowerTriangular()
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isUpperTriangular(RealScalar prec) const
+{
+  RealScalar maxAbsOnUpperPart = static_cast<RealScalar>(-1);
+  for(Index j = 0; j < cols(); ++j)
+  {
+    Index maxi = (std::min)(j, rows()-1);
+    for(Index i = 0; i <= maxi; ++i)
+    {
+      RealScalar absValue = internal::abs(coeff(i,j));
+      if(absValue > maxAbsOnUpperPart) maxAbsOnUpperPart = absValue;
+    }
+  }
+  RealScalar threshold = maxAbsOnUpperPart * prec;
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = j+1; i < rows(); ++i)
+      if(internal::abs(coeff(i, j)) > threshold) return false;
+  return true;
+}
+
+/** \returns true if *this is approximately equal to a lower triangular matrix,
+  *          within the precision given by \a prec.
+  *
+  * \sa isUpperTriangular()
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isLowerTriangular(RealScalar prec) const
+{
+  RealScalar maxAbsOnLowerPart = static_cast<RealScalar>(-1);
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = j; i < rows(); ++i)
+    {
+      RealScalar absValue = internal::abs(coeff(i,j));
+      if(absValue > maxAbsOnLowerPart) maxAbsOnLowerPart = absValue;
+    }
+  RealScalar threshold = maxAbsOnLowerPart * prec;
+  for(Index j = 1; j < cols(); ++j)
+  {
+    Index maxi = (std::min)(j, rows()-1);
+    for(Index i = 0; i < maxi; ++i)
+      if(internal::abs(coeff(i, j)) > threshold) return false;
+  }
+  return true;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULARMATRIX_H

densecrf/include/Eigen/src/Core/VectorBlock.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// 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_VECTORBLOCK_H
+#define EIGEN_VECTORBLOCK_H
+
+namespace Eigen { 
+
+/** \class VectorBlock
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a fixed-size or dynamic-size sub-vector
+  *
+  * \param VectorType the type of the object in which we are taking a sub-vector
+  * \param Size size of the sub-vector we are taking at compile time (optional)
+  *
+  * This class represents an expression of either a fixed-size or dynamic-size sub-vector.
+  * It is the return type of DenseBase::segment(Index,Index) and DenseBase::segment<int>(Index) and
+  * most of the time this is the only way it is used.
+  *
+  * However, if you want to directly maniputate sub-vector expressions,
+  * for instance if you want to write a function returning such an expression, you
+  * will need to use this class.
+  *
+  * Here is an example illustrating the dynamic case:
+  * \include class_VectorBlock.cpp
+  * Output: \verbinclude class_VectorBlock.out
+  *
+  * \note Even though this expression has dynamic size, in the case where \a VectorType
+  * has fixed size, this expression inherits a fixed maximal size which means that evaluating
+  * it does not cause a dynamic memory allocation.
+  *
+  * Here is an example illustrating the fixed-size case:
+  * \include class_FixedVectorBlock.cpp
+  * Output: \verbinclude class_FixedVectorBlock.out
+  *
+  * \sa class Block, DenseBase::segment(Index,Index,Index,Index), DenseBase::segment(Index,Index)
+  */
+
+namespace internal {
+template<typename VectorType, int Size>
+struct traits<VectorBlock<VectorType, Size> >
+  : public traits<Block<VectorType,
+                     traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     traits<VectorType>::Flags & RowMajorBit ? Size : 1> >
+{
+};
+}
+
+template<typename VectorType, int Size> class VectorBlock
+  : public Block<VectorType,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? Size : 1>
+{
+    typedef Block<VectorType,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? Size : 1> Base;
+    enum {
+      IsColVector = !(internal::traits<VectorType>::Flags & RowMajorBit)
+    };
+  public:
+    EIGEN_DENSE_PUBLIC_INTERFACE(VectorBlock)
+
+    using Base::operator=;
+
+    /** Dynamic-size constructor
+      */
+    inline VectorBlock(VectorType& vector, Index start, Index size)
+      : Base(vector,
+             IsColVector ? start : 0, IsColVector ? 0 : start,
+             IsColVector ? size  : 1, IsColVector ? 1 : size)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock);
+    }
+
+    /** Fixed-size constructor
+      */
+    inline VectorBlock(VectorType& vector, Index start)
+      : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock);
+    }
+};
+
+
+/** \returns a dynamic-size expression of a segment (i.e. a vector block) in *this.
+  *
+  * \only_for_vectors
+  *
+  * \param start the first coefficient in the segment
+  * \param size the number of coefficients in the segment
+  *
+  * Example: \include MatrixBase_segment_int_int.cpp
+  * Output: \verbinclude MatrixBase_segment_int_int.out
+  *
+  * \note Even though the returned expression has dynamic size, in the case
+  * when it is applied to a fixed-size vector, it inherits a fixed maximal size,
+  * which means that evaluating it does not cause a dynamic memory allocation.
+  *
+  * \sa class Block, segment(Index)
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::SegmentReturnType
+DenseBase<Derived>::segment(Index start, Index size)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return SegmentReturnType(derived(), start, size);
+}
+
+/** This is the const version of segment(Index,Index).*/
+template<typename Derived>
+inline typename DenseBase<Derived>::ConstSegmentReturnType
+DenseBase<Derived>::segment(Index start, Index size) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return ConstSegmentReturnType(derived(), start, size);
+}
+
+/** \returns a dynamic-size expression of the first coefficients of *this.
+  *
+  * \only_for_vectors
+  *
+  * \param size the number of coefficients in the block
+  *
+  * Example: \include MatrixBase_start_int.cpp
+  * Output: \verbinclude MatrixBase_start_int.out
+  *
+  * \note Even though the returned expression has dynamic size, in the case
+  * when it is applied to a fixed-size vector, it inherits a fixed maximal size,
+  * which means that evaluating it does not cause a dynamic memory allocation.
+  *
+  * \sa class Block, block(Index,Index)
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::SegmentReturnType
+DenseBase<Derived>::head(Index size)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return SegmentReturnType(derived(), 0, size);
+}
+
+/** This is the const version of head(Index).*/
+template<typename Derived>
+inline typename DenseBase<Derived>::ConstSegmentReturnType
+DenseBase<Derived>::head(Index size) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return ConstSegmentReturnType(derived(), 0, size);
+}
+
+/** \returns a dynamic-size expression of the last coefficients of *this.
+  *
+  * \only_for_vectors
+  *
+  * \param size the number of coefficients in the block
+  *
+  * Example: \include MatrixBase_end_int.cpp
+  * Output: \verbinclude MatrixBase_end_int.out
+  *
+  * \note Even though the returned expression has dynamic size, in the case
+  * when it is applied to a fixed-size vector, it inherits a fixed maximal size,
+  * which means that evaluating it does not cause a dynamic memory allocation.
+  *
+  * \sa class Block, block(Index,Index)
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::SegmentReturnType
+DenseBase<Derived>::tail(Index size)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return SegmentReturnType(derived(), this->size() - size, size);
+}
+
+/** This is the const version of tail(Index).*/
+template<typename Derived>
+inline typename DenseBase<Derived>::ConstSegmentReturnType
+DenseBase<Derived>::tail(Index size) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return ConstSegmentReturnType(derived(), this->size() - size, size);
+}
+
+/** \returns a fixed-size expression of a segment (i.e. a vector block) in \c *this
+  *
+  * \only_for_vectors
+  *
+  * The template parameter \a Size is the number of coefficients in the block
+  *
+  * \param start the index of the first element of the sub-vector
+  *
+  * Example: \include MatrixBase_template_int_segment.cpp
+  * Output: \verbinclude MatrixBase_template_int_segment.out
+  *
+  * \sa class Block
+  */
+template<typename Derived>
+template<int Size>
+inline typename DenseBase<Derived>::template FixedSegmentReturnType<Size>::Type
+DenseBase<Derived>::segment(Index start)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename FixedSegmentReturnType<Size>::Type(derived(), start);
+}
+
+/** This is the const version of segment<int>(Index).*/
+template<typename Derived>
+template<int Size>
+inline typename DenseBase<Derived>::template ConstFixedSegmentReturnType<Size>::Type
+DenseBase<Derived>::segment(Index start) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename ConstFixedSegmentReturnType<Size>::Type(derived(), start);
+}
+
+/** \returns a fixed-size expression of the first coefficients of *this.
+  *
+  * \only_for_vectors
+  *
+  * The template parameter \a Size is the number of coefficients in the block
+  *
+  * Example: \include MatrixBase_template_int_start.cpp
+  * Output: \verbinclude MatrixBase_template_int_start.out
+  *
+  * \sa class Block
+  */
+template<typename Derived>
+template<int Size>
+inline typename DenseBase<Derived>::template FixedSegmentReturnType<Size>::Type
+DenseBase<Derived>::head()
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename FixedSegmentReturnType<Size>::Type(derived(), 0);
+}
+
+/** This is the const version of head<int>().*/
+template<typename Derived>
+template<int Size>
+inline typename DenseBase<Derived>::template ConstFixedSegmentReturnType<Size>::Type
+DenseBase<Derived>::head() const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename ConstFixedSegmentReturnType<Size>::Type(derived(), 0);
+}
+
+/** \returns a fixed-size expression of the last coefficients of *this.
+  *
+  * \only_for_vectors
+  *
+  * The template parameter \a Size is the number of coefficients in the block
+  *
+  * Example: \include MatrixBase_template_int_end.cpp
+  * Output: \verbinclude MatrixBase_template_int_end.out
+  *
+  * \sa class Block
+  */
+template<typename Derived>
+template<int Size>
+inline typename DenseBase<Derived>::template FixedSegmentReturnType<Size>::Type
+DenseBase<Derived>::tail()
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename FixedSegmentReturnType<Size>::Type(derived(), size() - Size);
+}
+
+/** This is the const version of tail<int>.*/
+template<typename Derived>
+template<int Size>
+inline typename DenseBase<Derived>::template ConstFixedSegmentReturnType<Size>::Type
+DenseBase<Derived>::tail() const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename ConstFixedSegmentReturnType<Size>::Type(derived(), size() - Size);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_VECTORBLOCK_H

densecrf/include/Eigen/src/Core/VectorwiseOp.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// 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_PARTIAL_REDUX_H
+#define EIGEN_PARTIAL_REDUX_H
+
+namespace Eigen { 
+
+/** \class PartialReduxExpr
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression of a partially reduxed matrix
+  *
+  * \tparam MatrixType the type of the matrix we are applying the redux operation
+  * \tparam MemberOp type of the member functor
+  * \tparam Direction indicates the direction of the redux (#Vertical or #Horizontal)
+  *
+  * This class represents an expression of a partial redux operator of a matrix.
+  * It is the return type of some VectorwiseOp functions,
+  * and most of the time this is the only way it is used.
+  *
+  * \sa class VectorwiseOp
+  */
+
+template< typename MatrixType, typename MemberOp, int Direction>
+class PartialReduxExpr;
+
+namespace internal {
+template<typename MatrixType, typename MemberOp, int Direction>
+struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
+ : traits<MatrixType>
+{
+  typedef typename MemberOp::result_type Scalar;
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename traits<MatrixType>::XprKind XprKind;
+  typedef typename MatrixType::Scalar InputScalar;
+  typedef typename nested<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
+  enum {
+    RowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
+    Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits,
+    Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0),
+    TraversalSize = Direction==Vertical ? RowsAtCompileTime : ColsAtCompileTime
+  };
+  #if EIGEN_GNUC_AT_LEAST(3,4)
+  typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
+  #else
+  typedef typename MemberOp::template Cost<InputScalar,TraversalSize> CostOpType;
+  #endif
+  enum {
+    CoeffReadCost = TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value)
+  };
+};
+}
+
+template< typename MatrixType, typename MemberOp, int Direction>
+class PartialReduxExpr : internal::no_assignment_operator,
+  public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<PartialReduxExpr>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(PartialReduxExpr)
+    typedef typename internal::traits<PartialReduxExpr>::MatrixTypeNested MatrixTypeNested;
+    typedef typename internal::traits<PartialReduxExpr>::_MatrixTypeNested _MatrixTypeNested;
+
+    PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
+      : m_matrix(mat), m_functor(func) {}
+
+    Index rows() const { return (Direction==Vertical   ? 1 : m_matrix.rows()); }
+    Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); }
+
+    EIGEN_STRONG_INLINE const Scalar coeff(Index i, Index j) const
+    {
+      if (Direction==Vertical)
+        return m_functor(m_matrix.col(j));
+      else
+        return m_functor(m_matrix.row(i));
+    }
+
+    const Scalar coeff(Index index) const
+    {
+      if (Direction==Vertical)
+        return m_functor(m_matrix.col(index));
+      else
+        return m_functor(m_matrix.row(index));
+    }
+
+  protected:
+    MatrixTypeNested m_matrix;
+    const MemberOp m_functor;
+};
+
+#define EIGEN_MEMBER_FUNCTOR(MEMBER,COST)                               \
+  template <typename ResultType>                                        \
+  struct member_##MEMBER {                                           \
+    EIGEN_EMPTY_STRUCT_CTOR(member_##MEMBER)                         \
+    typedef ResultType result_type;                                     \
+    template<typename Scalar, int Size> struct Cost                     \
+    { enum { value = COST }; };                                         \
+    template<typename XprType>                                          \
+    EIGEN_STRONG_INLINE ResultType operator()(const XprType& mat) const \
+    { return mat.MEMBER(); } \
+  }
+
+namespace internal {
+
+EIGEN_MEMBER_FUNCTOR(squaredNorm, Size * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(norm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(stableNorm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(blueNorm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(hypotNorm, (Size-1) * functor_traits<scalar_hypot_op<Scalar> >::Cost );
+EIGEN_MEMBER_FUNCTOR(sum, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(mean, (Size-1)*NumTraits<Scalar>::AddCost + NumTraits<Scalar>::MulCost);
+EIGEN_MEMBER_FUNCTOR(minCoeff, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(maxCoeff, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(all, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(any, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(count, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost);
+
+
+template <typename BinaryOp, typename Scalar>
+struct member_redux {
+  typedef typename result_of<
+                     BinaryOp(Scalar)
+                   >::type  result_type;
+  template<typename _Scalar, int Size> struct Cost
+  { enum { value = (Size-1) * functor_traits<BinaryOp>::Cost }; };
+  member_redux(const BinaryOp func) : m_functor(func) {}
+  template<typename Derived>
+  inline result_type operator()(const DenseBase<Derived>& mat) const
+  { return mat.redux(m_functor); }
+  const BinaryOp m_functor;
+};
+}
+
+/** \class VectorwiseOp
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression providing partial reduction operations
+  *
+  * \param ExpressionType the type of the object on which to do partial reductions
+  * \param Direction indicates the direction of the redux (#Vertical or #Horizontal)
+  *
+  * This class represents a pseudo expression with partial reduction features.
+  * It is the return type of DenseBase::colwise() and DenseBase::rowwise()
+  * and most of the time this is the only way it is used.
+  *
+  * Example: \include MatrixBase_colwise.cpp
+  * Output: \verbinclude MatrixBase_colwise.out
+  *
+  * \sa DenseBase::colwise(), DenseBase::rowwise(), class PartialReduxExpr
+  */
+template<typename ExpressionType, int Direction> class VectorwiseOp
+{
+  public:
+
+    typedef typename ExpressionType::Scalar Scalar;
+    typedef typename ExpressionType::RealScalar RealScalar;
+    typedef typename ExpressionType::Index Index;
+    typedef typename internal::conditional<internal::must_nest_by_value<ExpressionType>::ret,
+        ExpressionType, ExpressionType&>::type ExpressionTypeNested;
+    typedef typename internal::remove_all<ExpressionTypeNested>::type ExpressionTypeNestedCleaned;
+
+    template<template<typename _Scalar> class Functor,
+                      typename Scalar=typename internal::traits<ExpressionType>::Scalar> struct ReturnType
+    {
+      typedef PartialReduxExpr<ExpressionType,
+                               Functor<Scalar>,
+                               Direction
+                              > Type;
+    };
+
+    template<typename BinaryOp> struct ReduxReturnType
+    {
+      typedef PartialReduxExpr<ExpressionType,
+                               internal::member_redux<BinaryOp,typename internal::traits<ExpressionType>::Scalar>,
+                               Direction
+                              > Type;
+    };
+
+    enum {
+      IsVertical   = (Direction==Vertical) ? 1 : 0,
+      IsHorizontal = (Direction==Horizontal) ? 1 : 0
+    };
+
+  protected:
+
+    /** \internal
+      * \returns the i-th subvector according to the \c Direction */
+    typedef typename internal::conditional<Direction==Vertical,
+                               typename ExpressionType::ColXpr,
+                               typename ExpressionType::RowXpr>::type SubVector;
+    SubVector subVector(Index i)
+    {
+      return SubVector(m_matrix.derived(),i);
+    }
+
+    /** \internal
+      * \returns the number of subvectors in the direction \c Direction */
+    Index subVectors() const
+    { return Direction==Vertical?m_matrix.cols():m_matrix.rows(); }
+
+    template<typename OtherDerived> struct ExtendedType {
+      typedef Replicate<OtherDerived,
+                        Direction==Vertical   ? 1 : ExpressionType::RowsAtCompileTime,
+                        Direction==Horizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
+    };
+
+    /** \internal
+      * Replicates a vector to match the size of \c *this */
+    template<typename OtherDerived>
+    typename ExtendedType<OtherDerived>::Type
+    extendedTo(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Vertical, OtherDerived::MaxColsAtCompileTime==1),
+                          YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(Direction==Horizontal, OtherDerived::MaxRowsAtCompileTime==1),
+                          YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
+      return typename ExtendedType<OtherDerived>::Type
+                      (other.derived(),
+                       Direction==Vertical   ? 1 : m_matrix.rows(),
+                       Direction==Horizontal ? 1 : m_matrix.cols());
+    }
+
+  public:
+
+    inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
+
+    /** \internal */
+    inline const ExpressionType& _expression() const { return m_matrix; }
+
+    /** \returns a row or column vector expression of \c *this reduxed by \a func
+      *
+      * The template parameter \a BinaryOp is the type of the functor
+      * of the custom redux operator. Note that func must be an associative operator.
+      *
+      * \sa class VectorwiseOp, DenseBase::colwise(), DenseBase::rowwise()
+      */
+    template<typename BinaryOp>
+    const typename ReduxReturnType<BinaryOp>::Type
+    redux(const BinaryOp& func = BinaryOp()) const
+    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), func); }
+
+    /** \returns a row (or column) vector expression of the smallest coefficient
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_minCoeff.cpp
+      * Output: \verbinclude PartialRedux_minCoeff.out
+      *
+      * \sa DenseBase::minCoeff() */
+    const typename ReturnType<internal::member_minCoeff>::Type minCoeff() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression of the largest coefficient
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_maxCoeff.cpp
+      * Output: \verbinclude PartialRedux_maxCoeff.out
+      *
+      * \sa DenseBase::maxCoeff() */
+    const typename ReturnType<internal::member_maxCoeff>::Type maxCoeff() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression of the squared norm
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_squaredNorm.cpp
+      * Output: \verbinclude PartialRedux_squaredNorm.out
+      *
+      * \sa DenseBase::squaredNorm() */
+    const typename ReturnType<internal::member_squaredNorm,RealScalar>::Type squaredNorm() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_norm.cpp
+      * Output: \verbinclude PartialRedux_norm.out
+      *
+      * \sa DenseBase::norm() */
+    const typename ReturnType<internal::member_norm,RealScalar>::Type norm() const
+    { return _expression(); }
+
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression, using
+      * blue's algorithm.
+      *
+      * \sa DenseBase::blueNorm() */
+    const typename ReturnType<internal::member_blueNorm,RealScalar>::Type blueNorm() const
+    { return _expression(); }
+
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression, avoiding
+      * underflow and overflow.
+      *
+      * \sa DenseBase::stableNorm() */
+    const typename ReturnType<internal::member_stableNorm,RealScalar>::Type stableNorm() const
+    { return _expression(); }
+
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression, avoiding
+      * underflow and overflow using a concatenation of hypot() calls.
+      *
+      * \sa DenseBase::hypotNorm() */
+    const typename ReturnType<internal::member_hypotNorm,RealScalar>::Type hypotNorm() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression of the sum
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_sum.cpp
+      * Output: \verbinclude PartialRedux_sum.out
+      *
+      * \sa DenseBase::sum() */
+    const typename ReturnType<internal::member_sum>::Type sum() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression of the mean
+    * of each column (or row) of the referenced expression.
+    *
+    * \sa DenseBase::mean() */
+    const typename ReturnType<internal::member_mean>::Type mean() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression representing
+      * whether \b all coefficients of each respective column (or row) are \c true.
+      *
+      * \sa DenseBase::all() */
+    const typename ReturnType<internal::member_all>::Type all() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression representing
+      * whether \b at \b least one coefficient of each respective column (or row) is \c true.
+      *
+      * \sa DenseBase::any() */
+    const typename ReturnType<internal::member_any>::Type any() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression representing
+      * the number of \c true coefficients of each respective column (or row).
+      *
+      * Example: \include PartialRedux_count.cpp
+      * Output: \verbinclude PartialRedux_count.out
+      *
+      * \sa DenseBase::count() */
+    const PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> count() const
+    { return _expression(); }
+
+    /** \returns a row (or column) vector expression of the product
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_prod.cpp
+      * Output: \verbinclude PartialRedux_prod.out
+      *
+      * \sa DenseBase::prod() */
+    const typename ReturnType<internal::member_prod>::Type prod() const
+    { return _expression(); }
+
+
+    /** \returns a matrix expression
+      * where each column (or row) are reversed.
+      *
+      * Example: \include Vectorwise_reverse.cpp
+      * Output: \verbinclude Vectorwise_reverse.out
+      *
+      * \sa DenseBase::reverse() */
+    const Reverse<ExpressionType, Direction> reverse() const
+    { return Reverse<ExpressionType, Direction>( _expression() ); }
+
+    typedef Replicate<ExpressionType,Direction==Vertical?Dynamic:1,Direction==Horizontal?Dynamic:1> ReplicateReturnType;
+    const ReplicateReturnType replicate(Index factor) const;
+
+    /**
+      * \return an expression of the replication of each column (or row) of \c *this
+      *
+      * Example: \include DirectionWise_replicate.cpp
+      * Output: \verbinclude DirectionWise_replicate.out
+      *
+      * \sa VectorwiseOp::replicate(Index), DenseBase::replicate(), class Replicate
+      */
+    // NOTE implemented here because of sunstudio's compilation errors
+    template<int Factor> const Replicate<ExpressionType,(IsVertical?Factor:1),(IsHorizontal?Factor:1)>
+    replicate(Index factor = Factor) const
+    {
+      return Replicate<ExpressionType,Direction==Vertical?Factor:1,Direction==Horizontal?Factor:1>
+          (_expression(),Direction==Vertical?factor:1,Direction==Horizontal?factor:1);
+    }
+
+/////////// Artithmetic operators ///////////
+
+    /** Copies the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived>
+    ExpressionType& operator=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      //eigen_assert((m_matrix.isNull()) == (other.isNull())); FIXME
+      return const_cast<ExpressionType&>(m_matrix = extendedTo(other.derived()));
+    }
+
+    /** Adds the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived>
+    ExpressionType& operator+=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return const_cast<ExpressionType&>(m_matrix += extendedTo(other.derived()));
+    }
+
+    /** Substracts the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived>
+    ExpressionType& operator-=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return const_cast<ExpressionType&>(m_matrix -= extendedTo(other.derived()));
+    }
+
+    /** Multiples each subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    ExpressionType& operator*=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      m_matrix *= extendedTo(other.derived());
+      return const_cast<ExpressionType&>(m_matrix);
+    }
+
+    /** Divides each subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    ExpressionType& operator/=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      m_matrix /= extendedTo(other.derived());
+      return const_cast<ExpressionType&>(m_matrix);
+    }
+
+    /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    CwiseBinaryOp<internal::scalar_sum_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator+(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix + extendedTo(other.derived());
+    }
+
+    /** Returns the expression of the difference between each subvector of \c *this and the vector \a other */
+    template<typename OtherDerived>
+    CwiseBinaryOp<internal::scalar_difference_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator-(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix - extendedTo(other.derived());
+    }
+
+    /** Returns the expression where each subvector is the product of the vector \a other
+      * by the corresponding subvector of \c *this */
+    template<typename OtherDerived> EIGEN_STRONG_INLINE
+    CwiseBinaryOp<internal::scalar_product_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator*(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix * extendedTo(other.derived());
+    }
+
+    /** Returns the expression where each subvector is the quotient of the corresponding
+      * subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator/(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix / extendedTo(other.derived());
+    }
+
+/////////// Geometry module ///////////
+
+    #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS
+    Homogeneous<ExpressionType,Direction> homogeneous() const;
+    #endif
+
+    typedef typename ExpressionType::PlainObject CrossReturnType;
+    template<typename OtherDerived>
+    const CrossReturnType cross(const MatrixBase<OtherDerived>& other) const;
+
+    enum {
+      HNormalized_Size = Direction==Vertical ? internal::traits<ExpressionType>::RowsAtCompileTime
+                                             : internal::traits<ExpressionType>::ColsAtCompileTime,
+      HNormalized_SizeMinusOne = HNormalized_Size==Dynamic ? Dynamic : HNormalized_Size-1
+    };
+    typedef Block<const ExpressionType,
+                  Direction==Vertical   ? int(HNormalized_SizeMinusOne)
+                                        : int(internal::traits<ExpressionType>::RowsAtCompileTime),
+                  Direction==Horizontal ? int(HNormalized_SizeMinusOne)
+                                        : int(internal::traits<ExpressionType>::ColsAtCompileTime)>
+            HNormalized_Block;
+    typedef Block<const ExpressionType,
+                  Direction==Vertical   ? 1 : int(internal::traits<ExpressionType>::RowsAtCompileTime),
+                  Direction==Horizontal ? 1 : int(internal::traits<ExpressionType>::ColsAtCompileTime)>
+            HNormalized_Factors;
+    typedef CwiseBinaryOp<internal::scalar_quotient_op<typename internal::traits<ExpressionType>::Scalar>,
+                const HNormalized_Block,
+                const Replicate<HNormalized_Factors,
+                  Direction==Vertical   ? HNormalized_SizeMinusOne : 1,
+                  Direction==Horizontal ? HNormalized_SizeMinusOne : 1> >
+            HNormalizedReturnType;
+
+    const HNormalizedReturnType hnormalized() const;
+
+  protected:
+    ExpressionTypeNested m_matrix;
+};
+
+/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+  *
+  * Example: \include MatrixBase_colwise.cpp
+  * Output: \verbinclude MatrixBase_colwise.out
+  *
+  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+  */
+template<typename Derived>
+inline const typename DenseBase<Derived>::ConstColwiseReturnType
+DenseBase<Derived>::colwise() const
+{
+  return derived();
+}
+
+/** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
+  *
+  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::ColwiseReturnType
+DenseBase<Derived>::colwise()
+{
+  return derived();
+}
+
+/** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+  *
+  * Example: \include MatrixBase_rowwise.cpp
+  * Output: \verbinclude MatrixBase_rowwise.out
+  *
+  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+  */
+template<typename Derived>
+inline const typename DenseBase<Derived>::ConstRowwiseReturnType
+DenseBase<Derived>::rowwise() const
+{
+  return derived();
+}
+
+/** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
+  *
+  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::RowwiseReturnType
+DenseBase<Derived>::rowwise()
+{
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARTIAL_REDUX_H

densecrf/include/Eigen/src/Core/Visitor.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 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_VISITOR_H
+#define EIGEN_VISITOR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Visitor, typename Derived, int UnrollCount>
+struct visitor_impl
+{
+  enum {
+    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
+    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+  };
+
+  static inline void run(const Derived &mat, Visitor& visitor)
+  {
+    visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor);
+    visitor(mat.coeff(row, col), row, col);
+  }
+};
+
+template<typename Visitor, typename Derived>
+struct visitor_impl<Visitor, Derived, 1>
+{
+  static inline void run(const Derived &mat, Visitor& visitor)
+  {
+    return visitor.init(mat.coeff(0, 0), 0, 0);
+  }
+};
+
+template<typename Visitor, typename Derived>
+struct visitor_impl<Visitor, Derived, Dynamic>
+{
+  typedef typename Derived::Index Index;
+  static inline void run(const Derived& mat, Visitor& visitor)
+  {
+    visitor.init(mat.coeff(0,0), 0, 0);
+    for(Index i = 1; i < mat.rows(); ++i)
+      visitor(mat.coeff(i, 0), i, 0);
+    for(Index j = 1; j < mat.cols(); ++j)
+      for(Index i = 0; i < mat.rows(); ++i)
+        visitor(mat.coeff(i, j), i, j);
+  }
+};
+
+} // end namespace internal
+
+/** Applies the visitor \a visitor to the whole coefficients of the matrix or vector.
+  *
+  * The template parameter \a Visitor is the type of the visitor and provides the following interface:
+  * \code
+  * struct MyVisitor {
+  *   // called for the first coefficient
+  *   void init(const Scalar& value, Index i, Index j);
+  *   // called for all other coefficients
+  *   void operator() (const Scalar& value, Index i, Index j);
+  * };
+  * \endcode
+  *
+  * \note compared to one or two \em for \em loops, visitors offer automatic
+  * unrolling for small fixed size matrix.
+  *
+  * \sa minCoeff(Index*,Index*), maxCoeff(Index*,Index*), DenseBase::redux()
+  */
+template<typename Derived>
+template<typename Visitor>
+void DenseBase<Derived>::visit(Visitor& visitor) const
+{
+  enum { unroll = SizeAtCompileTime != Dynamic
+                   && CoeffReadCost != Dynamic
+                   && (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
+                   && SizeAtCompileTime * CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
+                      <= EIGEN_UNROLLING_LIMIT };
+  return internal::visitor_impl<Visitor, Derived,
+      unroll ? int(SizeAtCompileTime) : Dynamic
+    >::run(derived(), visitor);
+}
+
+namespace internal {
+
+/** \internal
+  * \brief Base class to implement min and max visitors
+  */
+template <typename Derived>
+struct coeff_visitor
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  Index row, col;
+  Scalar res;
+  inline void init(const Scalar& value, Index i, Index j)
+  {
+    res = value;
+    row = i;
+    col = j;
+  }
+};
+
+/** \internal
+  * \brief Visitor computing the min coefficient with its value and coordinates
+  *
+  * \sa DenseBase::minCoeff(Index*, Index*)
+  */
+template <typename Derived>
+struct min_coeff_visitor : coeff_visitor<Derived>
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  void operator() (const Scalar& value, Index i, Index j)
+  {
+    if(value < this->res)
+    {
+      this->res = value;
+      this->row = i;
+      this->col = j;
+    }
+  }
+};
+
+template<typename Scalar>
+struct functor_traits<min_coeff_visitor<Scalar> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost
+  };
+};
+
+/** \internal
+  * \brief Visitor computing the max coefficient with its value and coordinates
+  *
+  * \sa DenseBase::maxCoeff(Index*, Index*)
+  */
+template <typename Derived>
+struct max_coeff_visitor : coeff_visitor<Derived>
+{
+  typedef typename Derived::Index Index;
+  typedef typename Derived::Scalar Scalar;
+  void operator() (const Scalar& value, Index i, Index j)
+  {
+    if(value > this->res)
+    {
+      this->res = value;
+      this->row = i;
+      this->col = j;
+    }
+  }
+};
+
+template<typename Scalar>
+struct functor_traits<max_coeff_visitor<Scalar> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost
+  };
+};
+
+} // end namespace internal
+
+/** \returns the minimum of all coefficients of *this
+  * and puts in *row and *col its location.
+  *
+  * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visitor(), DenseBase::minCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::minCoeff(IndexType* row, IndexType* col) const
+{
+  internal::min_coeff_visitor<Derived> minVisitor;
+  this->visit(minVisitor);
+  *row = minVisitor.row;
+  if (col) *col = minVisitor.col;
+  return minVisitor.res;
+}
+
+/** \returns the minimum of all coefficients of *this
+  * and puts in *index its location.
+  *
+  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::minCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::minCoeff(IndexType* index) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  internal::min_coeff_visitor<Derived> minVisitor;
+  this->visit(minVisitor);
+  *index = (RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row;
+  return minVisitor.res;
+}
+
+/** \returns the maximum of all coefficients of *this
+  * and puts in *row and *col its location.
+  *
+  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::maxCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::maxCoeff(IndexType* row, IndexType* col) const
+{
+  internal::max_coeff_visitor<Derived> maxVisitor;
+  this->visit(maxVisitor);
+  *row = maxVisitor.row;
+  if (col) *col = maxVisitor.col;
+  return maxVisitor.res;
+}
+
+/** \returns the maximum of all coefficients of *this
+  * and puts in *index its location.
+  *
+  * \sa DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::maxCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::maxCoeff(IndexType* index) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  internal::max_coeff_visitor<Derived> maxVisitor;
+  this->visit(maxVisitor);
+  *index = (RowsAtCompileTime==1) ? maxVisitor.col : maxVisitor.row;
+  return maxVisitor.res;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_VISITOR_H

densecrf/include/Eigen/src/Core/arch/AltiVec/CMakeLists.txt

+FILE(GLOB Eigen_Core_arch_AltiVec_SRCS "*.h")
+
+INSTALL(FILES
+  ${Eigen_Core_arch_AltiVec_SRCS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AltiVec COMPONENT Devel
+)

densecrf/include/Eigen/src/Core/arch/AltiVec/Complex.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 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_COMPLEX_ALTIVEC_H
+#define EIGEN_COMPLEX_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((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (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) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 1));//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 2), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 3));//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+
+//---------- float ----------
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  Packet4f  v;
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; };
+
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  /* On AltiVec we cannot load 64-bit registers, so wa have to take care of alignment */
+  if((ptrdiff_t(&from) % 16) == 0)
+    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);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_add(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_sub(a.v,b.v)); }
+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((Packet4f)vec_xor((Packet4ui)a.v, 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_COMPLEX_RE);
+  // Get the imaginary parts of a
+  v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM);
+  // 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);
+  
+  return Packet2cf(vec_add(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_and(a.v, vec_nor(b.v,b.v))); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*     from)
+{
+  return pset1<Packet2cf>(*from);
+}
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { vec_dstt((float *)addr, DST_CTRL(2,2,32), DST_CHAN); }
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore((float *)&res, a.v);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet4f rev_a;
+  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2);
+  return Packet2cf(rev_a);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  Packet4f b;
+  b = (Packet4f) vec_sld(a.v, a.v, 8);
+  b = padd(a.v, b);
+  return pfirst(Packet2cf(b));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  Packet4f b1, b2;
+  
+  b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8);
+  b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8);
+  b2 = (Packet4f) vec_sld(b2, b2, 8);
+  b2 = padd(b1, b2);
+
+  return Packet2cf(b2);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  Packet4f b;
+  Packet2cf prod;
+  b = (Packet4f) vec_sld(a.v, a.v, 8);
+  prod = pmul(a, Packet2cf(b));
+
+  return pfirst(prod);
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset==1)
+    {
+      first.v = vec_sld(first.v, second.v, 8);
+    }
+  }
+};
+
+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<> 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);
+  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))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x)
+{
+  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV));
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_ALTIVEC_H

densecrf/include/Eigen/src/Core/arch/AltiVec/PacketMath.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Konstantinos Margaritis <markos@codex.gr>
+//
+// 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_PACKET_MATH_ALTIVEC_H
+#define EIGEN_PACKET_MATH_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
+#endif
+
+#ifndef EIGEN_HAS_FUSE_CJMADD
+#define EIGEN_HAS_FUSE_CJMADD 1
+#endif
+
+// 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
+#endif
+
+typedef __vector float          Packet4f;
+typedef __vector int            Packet4i;
+typedef __vector unsigned int   Packet4ui;
+typedef __vector __bool int     Packet4bi;
+typedef __vector short int      Packet8i;
+typedef __vector unsigned char  Packet16uc;
+
+// We don't want to write the same code all the time, but we need to reuse the constants
+// and it doesn't really work to declare them global, so we define macros instead
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
+  Packet4f p4f_##NAME = (Packet4f) vec_splat_s32(X)
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = vec_splat_s32(X)
+
+#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 DST_CHAN 1
+#define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
+
+// Define global static constants:
+static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 };
+static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 };
+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);
+static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7};
+
+static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0);
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0);
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1);
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16);
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1);
+static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0);
+static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1);
+
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet4f type;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+
+    // FIXME check the Has*
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 0,
+    HasSqrt = 0
+  };
+};
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet4i type;
+  enum {
+    // FIXME check the Has*
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4
+  };
+};
+
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; };
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
+/*
+inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
+{
+  union {
+    Packet4f   v;
+    float n[4];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
+{
+  union {
+    Packet4i   v;
+    int n[4];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
+{
+  union {
+    Packet4ui   v;
+    unsigned int n[4];
+  } vt;
+  vt.v = 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 {
+    Packet4bi v;
+    unsigned int n[4];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
+  return s;
+}
+*/
+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);
+  vc = vec_splat(vc, 0);
+  return vc;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
+  int EIGEN_ALIGN16 ai[4];
+  ai[0] = from;
+  Packet4i vc = vec_ld(0, ai);
+  vc = vec_splat(vc, 0);
+  return vc;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)     { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
+
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); }
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); }
+/* Commented out: it's actually slower than processing it scalar
+ *
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
+{
+  // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec
+  //Set up constants, variables
+  Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel;
+
+  // Get the absolute values
+  a1  = vec_abs(a);
+  b1  = vec_abs(b);
+
+  // Get the signs using xor
+  Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO);
+
+  // Do the multiplication for the asbolute values.
+  bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 );
+  low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1);
+  high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO);
+  high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16);
+  prod = vec_add( low_prod, high_prod );
+
+  // NOR the product and select only the negative elements according to the sign mask
+  prod_ = vec_nor(prod, prod);
+  prod_ = vec_sel(p4i_ZERO, prod_, sgn);
+
+  // Add 1 to the result to get the negative numbers
+  v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn);
+  prod_ = vec_add(prod_, v1sel);
+
+  // Merge the results back to the final vector.
+  prod = vec_sel(prod, prod_, sgn);
+
+  return prod;
+}
+*/
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f t, y_0, y_1, res;
+
+  // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
+  y_0 = vec_re(b);
+
+  // Do one Newton-Raphson iteration to get the needed accuracy
+  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;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
+{ eigen_assert(false && "packet integer division are not supported by AltiVec");
+  return pset1<Packet4i>(0);
+}
+
+// for some weird raisons, it has to be overloaded for packet of integers
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
+
+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); }
+
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
+
+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); }
+
+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
+  LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
+  mask = vec_lvsl(0, from);                        // create the permute mask
+  return (Packet4f) vec_perm(MSQ, LSQ, mask);           // align the data
+
+}
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* 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
+  LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
+  mask = vec_lvsl(0, from);                        // create the permute mask
+  return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
+}
+
+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);
+}
+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);
+}
+
+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 void pstoreu<float>(float*  to, const Packet4f& from)
+{
+  EIGEN_DEBUG_UNALIGNED_STORE
+  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
+  // Warning: not thread safe!
+  Packet16uc MSQ, LSQ, edges;
+  Packet16uc edgeAlign, align;
+
+  MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
+  LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
+  edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
+  edges=vec_perm(LSQ,MSQ,edgeAlign);                        // extract the edges
+  align = vec_lvsr( 0, to );                                // permute map to misalign data
+  MSQ = vec_perm(edges,(Packet16uc)from,align);             // misalign the data (MSQ)
+  LSQ = vec_perm((Packet16uc)from,edges,align);             // misalign the data (LSQ)
+  vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
+  vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
+}
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from)
+{
+  EIGEN_DEBUG_UNALIGNED_STORE
+  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
+  // Warning: not thread safe!
+  Packet16uc MSQ, LSQ, edges;
+  Packet16uc edgeAlign, align;
+
+  MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
+  LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
+  edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
+  edges=vec_perm(LSQ, MSQ, edgeAlign);                      // extract the edges
+  align = vec_lvsr( 0, to );                                // permute map to misalign data
+  MSQ = vec_perm(edges, (Packet16uc) from, align);          // misalign the data (MSQ)
+  LSQ = vec_perm((Packet16uc) from, edges, align);          // misalign the data (LSQ)
+  vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
+  vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
+}
+
+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); }
+
+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 pabs(const Packet4f& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
+
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, sum;
+  b   = (Packet4f) vec_sld(a, a, 8);
+  sum = vec_add(a, b);
+  b   = (Packet4f) vec_sld(sum, sum, 4);
+  sum = vec_add(sum, b);
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  Packet4f v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = vec_add(sum[0], sum[1]);
+  // Lines 2+3
+  sum[1] = vec_add(sum[2], sum[3]);
+  // Add the results
+  sum[0] = vec_add(sum[0], sum[1]);
+
+  return sum[0];
+}
+
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i sum;
+  sum = vec_sums(a, p4i_ZERO);
+  sum = vec_sld(sum, p4i_ZERO, 12);
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  Packet4i v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = vec_add(sum[0], sum[1]);
+  // Lines 2+3
+  sum[1] = vec_add(sum[2], sum[3]);
+  // Add the results
+  sum[0] = vec_add(sum[0], sum[1]);
+
+  return sum[0];
+}
+
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  Packet4f prod;
+  prod = pmul(a, (Packet4f)vec_sld(a, a, 8));
+  return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4)));
+}
+
+template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+{
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  return aux[0] * aux[1] * aux[2] * aux[3];
+}
+
+// min
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, res;
+  b = vec_min(a, vec_sld(a, a, 8));
+  res = vec_min(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b = vec_min(a, vec_sld(a, a, 8));
+  res = vec_min(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+// max
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, res;
+  b = vec_max(a, vec_sld(a, a, 8));
+  res = vec_max(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b = vec_max(a, vec_sld(a, a, 8));
+  res = vec_max(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<int Offset>
+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);
+  }
+};
+
+template<int Offset>
+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);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_ALTIVEC_H

densecrf/include/Eigen/src/Core/arch/CMakeLists.txt

+ADD_SUBDIRECTORY(SSE)
+ADD_SUBDIRECTORY(AltiVec)
+ADD_SUBDIRECTORY(NEON)
+ADD_SUBDIRECTORY(Default)

densecrf/include/Eigen/src/Core/arch/Default/CMakeLists.txt

+FILE(GLOB Eigen_Core_arch_Default_SRCS "*.h")
+
+INSTALL(FILES
+  ${Eigen_Core_arch_Default_SRCS}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/Default COMPONENT Devel
+)

densecrf/include/Eigen/src/Core/arch/Default/Settings.h

+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// 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/.
+
+
+/* All the parameters defined in this file can be specialized in the
+ * architecture specific files, and/or by the user.
+ * More to come... */
+
+#ifndef EIGEN_DEFAULT_SETTINGS_H
+#define EIGEN_DEFAULT_SETTINGS_H
+
+/** Defines the maximal loop size to enable meta unrolling of loops.
+  * Note that the value here is expressed in Eigen's own notion of "number of FLOPS",
+  * it does not correspond to the number of iterations or the number of instructions
+  */
+#ifndef EIGEN_UNROLLING_LIMIT
+#define EIGEN_UNROLLING_LIMIT 100
+#endif
+
+/** Defines the threshold between a "small" and a "large" matrix.
+  * This threshold is mainly used to select the proper product implementation.
+  */
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+/** Defines the maximal width of the blocks used in the triangular product and solver
+  * for vectors (level 2 blas xTRMV and xTRSV). The default is 8.
+  */
+#ifndef EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH
+#define EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH 8
+#endif
+
+
+/** Defines the default number of registers available for that architecture.
+  * Currently it must be 8 or 16. Other values will fail.
+  */
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8
+#endif
+
+#endif // EIGEN_DEFAULT_SETTINGS_H