MsqHessian.hpp

Go to the documentation of this file.

/* *****************************************************************
    MESQUITE -- The Mesh Quality Improvement Toolkit

    Copyright 2004 Sandia Corporation and Argonne National
    Laboratory.  Under the terms of Contract DE-AC04-94AL85000
    with Sandia Corporation, the U.S. Government retains certain
    rights in this software.

    This library is free software; you can redistribute it and/or
    modify it under the terms of the GNU Lesser General Public
    License as published by the Free Software Foundation; either
    version 2.1 of the License, or (at your option) any later version.

    This library is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
    Lesser General Public License for more details.

    You should have received a copy of the GNU Lesser General Public License
    (lgpl.txt) along with this library; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA

    diachin2@llnl.gov, djmelan@sandia.gov, mbrewer@sandia.gov,
    pknupp@sandia.gov, tleurent@mcs.anl.gov, tmunson@mcs.anl.gov

  ***************************************************************** */
// -*- Mode : c++; tab-width: 3; c-tab-always-indent: t; indent-tabs-mode: nil; c-basic-offset: 3
// -*-
//
//    AUTHOR: Todd Munson <tmunson@mcs.anl.gov>
//       ORG: Argonne National Laboratory
//    E-MAIL: tmunson@mcs.anl.gov
//
// ORIG-DATE:  2-Jan-03 at 11:02:19 bu Thomas Leurent
//  LAST-MOD:  5-Oct-04 by Jason Kraftcheck
//
// DESCRIPTION:
// ============
/*! \file MsqHessian.hpp

 The MsqHessian class stores a sparse hessian for a given objective
 function. The objective function must be C2 and such that its hessian
 has non-zero entries only for the duplet of derivatives corresponding
 to nodes of a same element.

 \author Thomas Leurent
*/

#ifndef MsqHessian_hpp
#define MsqHessian_hpp

#include "Mesquite.hpp"
#include "Matrix3D.hpp"
#include "PatchData.hpp"
#include "MsqTimer.hpp"

#include <iosfwd>

namespace MBMesquite
{
class ObjectiveFunction;

/*!
  \class MsqHessian
  \brief Vector3D is the object that effeciently stores the objective function
  Hessian each entry is a Matrix3D object (i.e. a vertex Hessian).
*/
class MESQUITE_EXPORT MsqHessian
{
  protected:             // data accessed directly in tests.
    Matrix3D* mEntries;  //!< CSR block entries. size: nb of nonzero blocks, i.e. mRowStart[mSize] .
    size_t* mRowStart;   //!< start of each row in mEntries. size: nb of vertices (mSize).
    size_t* mColIndex;   //!< CSR block structure: column indexes of the row entries.

    size_t mSize;  //!< number of rows (or number of columns, this is a square matrix).

    Matrix3D* mPreconditioner;
    size_t precondArraySize;

    Vector3D* mR;         //!< array used in the CG solver
    Vector3D* mZ;         //!< array used in the CG solver
    Vector3D* mP;         //!< array used in the CG solver
    Vector3D* mW;         //!< array used in the CG solver
    size_t cgArraySizes;  //!< size of arrays allocated in the CG solver.
    size_t maxCGiter;     //!< max nb of iterations of the CG solver.

  public:
    MsqHessian();
    ~MsqHessian();

    void initialize( PatchData& pd, MsqError& err );
    void initialize( const MsqHessian& other );

    inline void zero_out();

    size_t size() const
    {
        return mSize;
    }

    //! returns the diagonal blocks, memory must be allocated before call.
    void get_diagonal_blocks( std::vector< Matrix3D >& diag, MsqError& err ) const;

    Matrix3D* get_block( size_t i, size_t j );
    const Matrix3D* get_block( size_t i, size_t j ) const;

    // inline void accumulate_entries(PatchData &pd, const size_t &elem_index,
    //                               Matrix3D* const &mat3d_array, MsqError &err);

    void compute_preconditioner( MsqError& err );
    void apply_preconditioner( Vector3D zloc[], Vector3D rloc[], MsqError& err );

    void cg_solver( Vector3D x[], Vector3D b[], MsqError& err );
    //! Hessian - vector product, summed with a second vector (optional).
    friend void axpy( Vector3D res[],
                      size_t size_r,
                      const MsqHessian& H,
                      const Vector3D x[],
                      size_t size_x,
                      const Vector3D y[],
                      size_t size_y,
                      MsqError& err );

    //! r = this * x, where r and x are arrays of length size().
    void product( Vector3D* r, const Vector3D* x ) const;

    friend std::ostream& operator<<( std::ostream&, const MsqHessian& );

    inline void add( size_t row, size_t col, const Matrix3D& m, MsqError& err );

    inline void scale( double value );

    void add( const MsqHessian& other );

    // Release all storage.  Object is invalid until next call
    // to initialize(..).
    void clear();

    inline double norm() const;  //!< Frobenius norm

  private:
    MsqHessian& operator=( const MsqHessian& h );
    MsqHessian( const MsqHessian& copy );
};

inline double MsqHessian::norm() const
{
    double sum = 0.0;
    for( size_t i = 0; i < mRowStart[mSize]; ++i )
        sum += Frobenius_2( mEntries[i] );
    return sqrt( sum );
}

/*! Sets all Hessian entries to zero. This is usually used before
  starting to accumulate elements hessian in the objective function
  hessian. */
inline void MsqHessian::zero_out()
{
    if( mSize == 0 ) return;  // empty hessian.

    size_t i;
    for( i = 0; i < mRowStart[mSize]; ++i )
    {
        mEntries[i].zero();
    }
}

inline void MsqHessian::scale( double value )
{
    for( size_t i = 0; i < mRowStart[mSize]; ++i )
        mEntries[i] *= value;
}

/*! Accumulates entries of an element hessian into an objective function
  hessian. Make sure to use zero_out() before starting the accumulation
  process.

  \param pd: PatchData in that contains the element which Hessian
  we are accumulating in the Hessian matrix. This must be the same
  PatchData that was used in MsqHessian::initialize().
  \param elem_index: index of the element in the PatchData.
  \param mat3d_array: This is the upper triangular part of the element Hessian
  for all nodes, including fixed nodes, for which the entries must be null Matrix3Ds.
  \param nb_mat3d. The size of the mat3d_array: (n+1)n/2, where n is
  the number of nodes in the element.
*/
//  inline void MsqHessian::accumulate_entries(PatchData &pd, const size_t &elem_index,
//                                             Matrix3D* const &mat3d_array, MsqError &err)
//    {
//      if (&pd != origin_pd) {
//        MSQ_SETERR(err)(
//                    "Cannot accumulate elements from a different patch. "
//                    "Use MsqHessian::initialize first.",
//                    MsqError::INVALID_ARG );
//        return;
//      }
//
//      size_t nve = pd.get_element_array(err)[elem_index].vertex_count();
//      size_t* ve = pd.get_element_array(err)[elem_index].get_vertex_index_array();
//      size_t e = mAccumElemStart[elem_index];
//      size_t i, j, c = 0;
//      for (i = 0; i < nve; ++i)
//      {
//        for (j = i; j < nve; ++j)
//        {
//          if (ve[i] < mSize && ve[j] < mSize)
//          {
//            int k = mAccumulation[e++];
//            if (k >= 0)
//              mEntries[k] += mat3d_array[c];
//            else
//              mEntries[-k].plus_transpose_equal( mat3d_array[c] );
//          }
//          ++c;
//        }
//      }
//    }

inline void MsqHessian::add( size_t row, size_t col, const Matrix3D& m, MsqError& err )
{
    if( row <= col )
    {
        for( size_t i = mRowStart[row]; i != mRowStart[row + 1]; ++i )
            if( mColIndex[i] == col )
            {
                mEntries[i] += m;
                return;
            }
    }
    else
    {
        for( size_t i = mRowStart[col]; i != mRowStart[col + 1]; ++i )
            if( mColIndex[i] == row )
            {
                mEntries[i].plus_transpose_equal( m );
                return;
            }
    }

    MSQ_SETERR( err )
    ( MsqError::INVALID_ARG, "Hessian entry (%lu,%lu) does not exist.", (unsigned long)row, (unsigned long)col );
}

/*!
  \param res: array of Vector3D in which the result is stored.
  \param size_r: size of the res array.
  \param x: vector multiplied by the Hessian.
  \param size_x: size of the x array.
  \param y: vector added to the Hessian vector product. Set to 0 (NULL) if not needed.
  \param size_y: size of the y array. Set to 0 if not needed.
*/
inline void axpy( Vector3D res[],
                  size_t size_r,
                  const MsqHessian& H,
                  const Vector3D x[],
                  size_t size_x,
                  const Vector3D y[],
                  size_t size_y,
                  MsqError& /*err*/ )
{
    if( ( size_r != H.mSize ) || ( size_x != H.mSize ) || ( size_y != H.mSize && size_y != 0 ) )
    {
        // throw an error
    }

    Vector3D tmpx, tmpm;  // for cache opt.
    size_t* col     = H.mColIndex;
    const size_t nn = H.mSize;
    size_t rl;  // row length
    size_t el;  // entries index
    size_t lo;
    size_t c;  // column index
    size_t i, j;

    if( y != 0 )
    {
        for( i = 0; i < nn; ++i )
        {
            res[i] = y[i];
        }
    }
    else
    {  // y == 0
        for( i = 0; i < nn; ++i )
        {
            res[i] = 0.;
        }
    }

    el = 0;
    for( i = 0; i < nn; ++i )
    {
        rl = H.mRowStart[i + 1] - H.mRowStart[i];
        lo = *col++;

        // Diagonal entry
        tmpx = x[i];
        eqAx( tmpm, H.mEntries[el], tmpx );
        ++el;

        // Non-diagonal entries
        for( j = 1; j < rl; ++j )
        {
            c = *col++;
            //          res[i] += H.mEntries[e] * x[c];
            plusEqAx( tmpm, H.mEntries[el], x[c] );
            //          res[c] += transpose(H.mEntries[e]) * tmpxi;
            plusEqTransAx( res[c], H.mEntries[el], tmpx );
            ++el;
        }
        res[lo] += tmpm;
    }
}

/*! Computes \f$ z=M^{-1}r \f$ . */
inline void MsqHessian::apply_preconditioner( Vector3D zloc[], Vector3D rloc[], MsqError& /*err*/ )
{
    size_t m;

    for( m = 0; m < mSize; ++m )
    {
#ifdef DIAGONAL_PRECONDITIONER
        // preconditioner is identity matrix for now.
        zloc[m][0] = mPreconditioner[m][0][0] * rloc[m][0];
        zloc[m][1] = mPreconditioner[m][1][1] * rloc[m][1];
        zloc[m][2] = mPreconditioner[m][2][2] * rloc[m][2];
#else
        // z = inv(L^T) * r
        zloc[m][0] = rloc[m][0];
        zloc[m][1] = rloc[m][1] - mPreconditioner[m][0][1] * zloc[m][0];
        zloc[m][2] = rloc[m][2] - mPreconditioner[m][0][2] * zloc[m][0] - mPreconditioner[m][1][2] * zloc[m][1];

        // z = inv(D) * z
        zloc[m][0] *= mPreconditioner[m][0][0];
        zloc[m][1] *= mPreconditioner[m][1][1];
        zloc[m][2] *= mPreconditioner[m][2][2];

        // z = inv(L) * z
        zloc[m][2] = zloc[m][2];
        zloc[m][1] = zloc[m][1] - mPreconditioner[m][1][2] * zloc[m][2];
        zloc[m][0] = zloc[m][0] - mPreconditioner[m][0][1] * zloc[m][1] - mPreconditioner[m][0][2] * zloc[m][2];
#endif
    }
}

/*! Returns a pointer to the Matrix3D block at position i,j if it exist.
  Returns the NULL pointer if position i,j (0-based) is a NULL entry.
  Note that block i,j must be in the upper triangular part of the
  (symetric) hessian. */
inline Matrix3D* MsqHessian::get_block( size_t i, size_t j )
{
    size_t c;

    if( i >= mSize || j >= mSize || j < i ) return NULL;

    for( c = mRowStart[i]; c < mRowStart[i + 1]; ++c )
    {
        if( mColIndex[c] == j ) return ( mEntries + c );
    }

    // if there is no block at position i,j (zero entry).
    return NULL;
}
inline const Matrix3D* MsqHessian::get_block( size_t i, size_t j ) const
{
    return const_cast< MsqHessian* >( this )->get_block( i, j );
}

/* ------------------ I/O ----------------- */

std::ostream& operator<<( std::ostream& s, const MsqHessian& h );

}  // namespace MBMesquite

#endif  // MsqHessian_hpp