ldl.hh

// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
// vi: set et ts=4 sw=2 sts=2:
#ifndef DUNE_ISTL_LDL_HH
#define DUNE_ISTL_LDL_HH

#if HAVE_SUITESPARSE_LDL || defined DOXYGEN

#include <iostream>
#include <type_traits>

#ifdef __cplusplus
extern "C"
{
#include "ldl.h"
#include "amd.h"
}
#endif

#include <dune/common/exceptions.hh>
#include <dune/common/unused.hh>

#include <dune/istl/colcompmatrix.hh>
#include <dune/istl/solvers.hh>
#include <dune/istl/solvertype.hh>

namespace Dune {
  /**
   * @addtogroup ISTL
   *
   * @{
   */
  /**
   * @file
   * @author Marco Agnese, Andrea Sacconi
   * @brief Class for using LDL with ISTL matrices.
   */

  // forward declarations
  template<class M, class T, class TM, class TD, class TA>
  class SeqOverlappingSchwarz;

  template<class T, bool tag>
  struct SeqOverlappingSchwarzAssemblerHelper;

  /**
   * @brief Use the %LDL package to directly solve linear systems -- empty default class
   * @tparam Matrix the matrix type defining the system
   * Details on UMFPack can be found on
   * http://www.cise.ufl.edu/research/sparse/ldl/
   */
  template<class Matrix>
  class LDL
  {};

  /**
   * @brief The %LDL direct sparse solver for matrices of type BCRSMatrix
   *
   * Specialization for the Dune::BCRSMatrix. %LDL will always go double
   * precision.
   *
   * \tparam T Number type.  Only double is supported
   * \tparam A STL-compatible allocator type
   * \tparam n Number of rows in a matrix block
   * \tparam m Number of columns in a matrix block
   *
   * \note This will only work if dune-istl has been configured to use LDL
   */
  template<typename T, typename A, int n, int m>
  class LDL<BCRSMatrix<FieldMatrix<T,n,m>,A > >
    : public InverseOperator<BlockVector<FieldVector<T,m>, typename A::template rebind<FieldVector<T,m> >::other>,
                             BlockVector<FieldVector<T,n>, typename A::template rebind<FieldVector<T,n> >::other> >
  {
    public:
    /** @brief The matrix type. */
    typedef Dune::BCRSMatrix<FieldMatrix<T,n,m>,A> Matrix;
    typedef Dune::BCRSMatrix<FieldMatrix<T,n,m>,A> matrix_type;
    /** @brief The corresponding SuperLU Matrix type. */
    typedef Dune::ColCompMatrix<Matrix> LDLMatrix;
    /** @brief Type of an associated initializer class. */
    typedef ColCompMatrixInitializer<BCRSMatrix<FieldMatrix<T,n,m>,A> > MatrixInitializer;
    /** @brief The type of the domain of the solver. */
    typedef Dune::BlockVector<FieldVector<T,m>, typename A::template rebind<FieldVector<T,m> >::other> domain_type;
    /** @brief The type of the range of the solver. */
    typedef Dune::BlockVector<FieldVector<T,n>, typename A::template rebind<FieldVector<T,n> >::other> range_type;

    /**
     * @brief Construct a solver object from a BCRSMatrix.
     *
     * This computes the matrix decomposition, and may take a long time
     * (and use a lot of memory).
     *
     * @param matrix the matrix to solve for
     * @param verbose 0 or 1 set the verbosity level, defaults to 0
     */
    LDL(const Matrix& matrix, int verbose=0) : matrixIsLoaded_(false), verbose_(verbose)
    {
      //check whether T is a supported type
      static_assert(std::is_same<T,double>::value,"Unsupported Type in LDL (only double supported)");
      setMatrix(matrix);
    }

    /**
     * @brief Constructor for compatibility with SuperLU standard constructor
     *
     * This computes the matrix decomposition, and may take a long time
     * (and use a lot of memory).
     *
     * @param matrix the matrix to solve for
     * @param verbose 0 or 1 set the verbosity level, defaults to 0
     */
    LDL(const Matrix& matrix, int verbose, bool) : matrixIsLoaded_(false), verbose_(verbose)
    {
      //check whether T is a supported type
      static_assert(std::is_same<T,double>::value,"Unsupported Type in LDL (only double supported)");
      setMatrix(matrix);
    }

    /** @brief Default constructor. */
    LDL() : matrixIsLoaded_(false), verbose_(0)
    {}

    /** @brief Default constructor. */
    virtual ~LDL()
    {
      if ((ldlMatrix_.N() + ldlMatrix_.M() > 0) || matrixIsLoaded_)
        free();
    }

    /** \copydoc InverseOperator::apply(X&, Y&, InverseOperatorResult&) */
    virtual void apply(domain_type& x, range_type& b, InverseOperatorResult& res)
    {
      const int dimMat(ldlMatrix_.N());
      ldl_perm(dimMat, Y_, reinterpret_cast<double*>(&b[0]), P_);
      ldl_lsolve(dimMat, Y_, Lp_, Li_, Lx_);
      ldl_dsolve(dimMat, Y_, D_);
      ldl_ltsolve(dimMat, Y_, Lp_, Li_, Lx_);
      ldl_permt(dimMat, reinterpret_cast<double*>(&x[0]), Y_, P_);
      // this is a direct solver
      res.iterations = 1;
      res.converged = true;
    }

    /** \copydoc InverseOperator::apply(X&,Y&,double,InverseOperatorResult&) */
    virtual void apply(domain_type& x, range_type& b, double reduction, InverseOperatorResult& res)
    {
      DUNE_UNUSED_PARAMETER(reduction);
      apply(x,b,res);
    }

    /**
     * @brief Additional apply method with c-arrays in analogy to superlu.
     * @param x solution array
     * @param b rhs array
     */
    void apply(T* x, T* b)
    {
      const int dimMat(ldlMatrix_.N());
      ldl_perm(dimMat, Y_, b, P_);
      ldl_lsolve(dimMat, Y_, Lp_, Li_, Lx_);
      ldl_dsolve(dimMat, Y_, D_);
      ldl_ltsolve(dimMat, Y_, Lp_, Li_, Lx_);
      ldl_permt(dimMat, x, Y_, P_);
    }

    void setOption(unsigned int option, double value)
    {
      DUNE_UNUSED_PARAMETER(option);
      DUNE_UNUSED_PARAMETER(value);
    }

    /** @brief Initialize data from given matrix. */
    void setMatrix(const Matrix& matrix)
    {
      if ((ldlMatrix_.N() + ldlMatrix_.M() > 0) || matrixIsLoaded_)
        free();
      ldlMatrix_ = matrix;
      decompose();
    }

    template<class S>
    void setSubMatrix(const Matrix& matrix, const S& rowIndexSet)
    {
      if ((ldlMatrix_.N() + ldlMatrix_.M() > 0) || matrixIsLoaded_)
        free();
      ldlMatrix_.setMatrix(matrix,rowIndexSet);
      decompose();
    }

    /**
     * @brief Sets the verbosity level for the solver.
     * @param v verbosity level: 0 only error messages, 1 a bit of statistics.
     */
    inline void setVerbosity(int v)
    {
      verbose_=v;
    }

    /**
     * @brief Return the column compress matrix.
     * @warning It is up to the user to keep consistency.
     */
    inline LDLMatrix& getInternalMatrix()
    {
      return ldlMatrix_;
    }

    /**
     * @brief Free allocated space.
     * @warning Later calling apply will result in an error.
     */
    void free()
    {
      delete [] D_;
      delete [] Y_;
      delete [] Lp_;
      delete [] Lx_;
      delete [] Li_;
      delete [] P_;
      delete [] Pinv_;
      ldlMatrix_.free();
      matrixIsLoaded_ = false;
    }

    /** @brief Get method name. */
    inline const char* name()
    {
      return "LDL";
    }

    /**
     * @brief Get factorization diagonal matrix D.
     * @warning It is up to the user to preserve consistency.
     */
    inline double* getD()
    {
      return D_;
    }

    /**
     * @brief Get factorization Lp.
     * @warning It is up to the user to preserve consistency.
     */
    inline int* getLp()
    {
      return Lp_;
    }

    /**
     * @brief Get factorization Li.
     * @warning It is up to the user to preserve consistency.
     */
    inline int* getLi()
    {
      return Li_;
    }

    /**
     * @brief Get factorization Lx.
     * @warning It is up to the user to preserve consistency.
     */
    inline double* getLx()
    {
      return Lx_;
    }

    private:
    template<class M,class X, class TM, class TD, class T1>
    friend class SeqOverlappingSchwarz;

    friend struct SeqOverlappingSchwarzAssemblerHelper<LDL<Matrix>,true>;

    /** @brief Computes the LDL decomposition. */
    void decompose()
    {
      // allocate vectors
      const int dimMat(ldlMatrix_.N());
      D_ = new double [dimMat];
      Y_ = new double [dimMat];
      Lp_ = new int [dimMat + 1];
      Parent_ = new int [dimMat];
      Lnz_ = new int [dimMat];
      Flag_ = new int [dimMat];
      Pattern_ = new int [dimMat];
      P_ = new int [dimMat];
      Pinv_ = new int [dimMat];

      double Info [AMD_INFO];
      if(amd_order (dimMat, ldlMatrix_.getColStart(), ldlMatrix_.getRowIndex(), P_, (double *) NULL, Info) < AMD_OK)
        DUNE_THROW(InvalidStateException,"Error: AMD failed!");
      if(verbose_ > 0)
        amd_info (Info);
      // compute the symbolic factorisation
      ldl_symbolic(dimMat, ldlMatrix_.getColStart(), ldlMatrix_.getRowIndex(), Lp_, Parent_, Lnz_, Flag_, P_, Pinv_);
      // initialise those entries of additionalVectors_ whose dimension is known only now
      Lx_ = new double [Lp_[dimMat]];
      Li_ = new int [Lp_[dimMat]];
      // compute the numeric factorisation
      const int rank(ldl_numeric(dimMat, ldlMatrix_.getColStart(), ldlMatrix_.getRowIndex(), ldlMatrix_.getValues(),
                                 Lp_, Parent_, Lnz_, Li_, Lx_, D_, Y_, Pattern_, Flag_, P_, Pinv_));
      // free temporary vectors
      delete [] Flag_;
      delete [] Pattern_;
      delete [] Parent_;
      delete [] Lnz_;

      if(rank!=dimMat)
        DUNE_THROW(InvalidStateException,"Error: LDL factorisation failed!");
    }

    LDLMatrix ldlMatrix_;
    bool matrixIsLoaded_;
    int verbose_;
    int* Lp_;
    int* Parent_;
    int* Lnz_;
    int* Flag_;
    int* Pattern_;
    int* P_;
    int* Pinv_;
    double* D_;
    double* Y_;
    double* Lx_;
    int* Li_;
  };

  template<typename T, typename A, int n, int m>
  struct IsDirectSolver<LDL<BCRSMatrix<FieldMatrix<T,n,m>,A> > >
  {
    enum {value = true};
  };

  template<typename T, typename A, int n, int m>
  struct StoresColumnCompressed<LDL<BCRSMatrix<FieldMatrix<T,n,m>,A> > >
  {
    enum {value = true};
  };

}

#endif //HAVE_SUITESPARSE_LDL
#endif //DUNE_ISTL_LDL_HH