• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_DGMRES_H
11 #define EIGEN_DGMRES_H
12 
13 #include <Eigen/Eigenvalues>
14 
15 namespace Eigen {
16 
17 template< typename _MatrixType,
18           typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
19 class DGMRES;
20 
21 namespace internal {
22 
23 template< typename _MatrixType, typename _Preconditioner>
24 struct traits<DGMRES<_MatrixType,_Preconditioner> >
25 {
26   typedef _MatrixType MatrixType;
27   typedef _Preconditioner Preconditioner;
28 };
29 
30 /** \brief Computes a permutation vector to have a sorted sequence
31   * \param vec The vector to reorder.
32   * \param perm gives the sorted sequence on output. Must be initialized with 0..n-1
33   * \param ncut Put  the ncut smallest elements at the end of the vector
34   * WARNING This is an expensive sort, so should be used only
35   * for small size vectors
36   * TODO Use modified QuickSplit or std::nth_element to get the smallest values
37   */
38 template <typename VectorType, typename IndexType>
39 void sortWithPermutation (VectorType& vec, IndexType& perm, typename IndexType::Scalar& ncut)
40 {
41   eigen_assert(vec.size() == perm.size());
42   typedef typename IndexType::Scalar Index;
43   bool flag;
44   for (Index k  = 0; k < ncut; k++)
45   {
46     flag = false;
47     for (Index j = 0; j < vec.size()-1; j++)
48     {
49       if ( vec(perm(j)) < vec(perm(j+1)) )
50       {
51         std::swap(perm(j),perm(j+1));
52         flag = true;
53       }
54       if (!flag) break; // The vector is in sorted order
55     }
56   }
57 }
58 
59 }
60 /**
61  * \ingroup IterativeLInearSolvers_Module
62  * \brief A Restarted GMRES with deflation.
63  * This class implements a modification of the GMRES solver for
64  * sparse linear systems. The basis is built with modified
65  * Gram-Schmidt. At each restart, a few approximated eigenvectors
66  * corresponding to the smallest eigenvalues are used to build a
67  * preconditioner for the next cycle. This preconditioner
68  * for deflation can be combined with any other preconditioner,
69  * the IncompleteLUT for instance. The preconditioner is applied
70  * at right of the matrix and the combination is multiplicative.
71  *
72  * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
73  * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
74  * Typical usage :
75  * \code
76  * SparseMatrix<double> A;
77  * VectorXd x, b;
78  * //Fill A and b ...
79  * DGMRES<SparseMatrix<double> > solver;
80  * solver.set_restart(30); // Set restarting value
81  * solver.setEigenv(1); // Set the number of eigenvalues to deflate
82  * solver.compute(A);
83  * x = solver.solve(b);
84  * \endcode
85  *
86  * DGMRES can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
87  *
88  * References :
89  * [1] D. NUENTSA WAKAM and F. PACULL, Memory Efficient Hybrid
90  *  Algebraic Solvers for Linear Systems Arising from Compressible
91  *  Flows, Computers and Fluids, In Press,
92  *  http://dx.doi.org/10.1016/j.compfluid.2012.03.023
93  * [2] K. Burrage and J. Erhel, On the performance of various
94  * adaptive preconditioned GMRES strategies, 5(1998), 101-121.
95  * [3] J. Erhel, K. Burrage and B. Pohl, Restarted GMRES
96  *  preconditioned by deflation,J. Computational and Applied
97  *  Mathematics, 69(1996), 303-318.
98 
99  *
100  */
101 template< typename _MatrixType, typename _Preconditioner>
102 class DGMRES : public IterativeSolverBase<DGMRES<_MatrixType,_Preconditioner> >
103 {
104     typedef IterativeSolverBase<DGMRES> Base;
105     using Base::matrix;
106     using Base::m_error;
107     using Base::m_iterations;
108     using Base::m_info;
109     using Base::m_isInitialized;
110     using Base::m_tolerance;
111   public:
112     using Base::_solve_impl;
113     typedef _MatrixType MatrixType;
114     typedef typename MatrixType::Scalar Scalar;
115     typedef typename MatrixType::Index Index;
116     typedef typename MatrixType::StorageIndex StorageIndex;
117     typedef typename MatrixType::RealScalar RealScalar;
118     typedef _Preconditioner Preconditioner;
119     typedef Matrix<Scalar,Dynamic,Dynamic> DenseMatrix;
120     typedef Matrix<RealScalar,Dynamic,Dynamic> DenseRealMatrix;
121     typedef Matrix<Scalar,Dynamic,1> DenseVector;
122     typedef Matrix<RealScalar,Dynamic,1> DenseRealVector;
123     typedef Matrix<std::complex<RealScalar>, Dynamic, 1> ComplexVector;
124 
125 
126   /** Default constructor. */
127   DGMRES() : Base(),m_restart(30),m_neig(0),m_r(0),m_maxNeig(5),m_isDeflAllocated(false),m_isDeflInitialized(false) {}
128 
129   /** Initialize the solver with matrix \a A for further \c Ax=b solving.
130     *
131     * This constructor is a shortcut for the default constructor followed
132     * by a call to compute().
133     *
134     * \warning this class stores a reference to the matrix A as well as some
135     * precomputed values that depend on it. Therefore, if \a A is changed
136     * this class becomes invalid. Call compute() to update it with the new
137     * matrix A, or modify a copy of A.
138     */
139   template<typename MatrixDerived>
140   explicit DGMRES(const EigenBase<MatrixDerived>& A) : Base(A.derived()), m_restart(30),m_neig(0),m_r(0),m_maxNeig(5),m_isDeflAllocated(false),m_isDeflInitialized(false) {}
141 
142   ~DGMRES() {}
143 
144   /** \internal */
145   template<typename Rhs,typename Dest>
146   void _solve_with_guess_impl(const Rhs& b, Dest& x) const
147   {
148     bool failed = false;
149     for(int j=0; j<b.cols(); ++j)
150     {
151       m_iterations = Base::maxIterations();
152       m_error = Base::m_tolerance;
153 
154       typename Dest::ColXpr xj(x,j);
155       dgmres(matrix(), b.col(j), xj, Base::m_preconditioner);
156     }
157     m_info = failed ? NumericalIssue
158            : m_error <= Base::m_tolerance ? Success
159            : NoConvergence;
160     m_isInitialized = true;
161   }
162 
163   /** \internal */
164   template<typename Rhs,typename Dest>
165   void _solve_impl(const Rhs& b, MatrixBase<Dest>& x) const
166   {
167     x = b;
168     _solve_with_guess_impl(b,x.derived());
169   }
170   /**
171    * Get the restart value
172     */
173   int restart() { return m_restart; }
174 
175   /**
176    * Set the restart value (default is 30)
177    */
178   void set_restart(const int restart) { m_restart=restart; }
179 
180   /**
181    * Set the number of eigenvalues to deflate at each restart
182    */
183   void setEigenv(const int neig)
184   {
185     m_neig = neig;
186     if (neig+1 > m_maxNeig) m_maxNeig = neig+1; // To allow for complex conjugates
187   }
188 
189   /**
190    * Get the size of the deflation subspace size
191    */
192   int deflSize() {return m_r; }
193 
194   /**
195    * Set the maximum size of the deflation subspace
196    */
197   void setMaxEigenv(const int maxNeig) { m_maxNeig = maxNeig; }
198 
199   protected:
200     // DGMRES algorithm
201     template<typename Rhs, typename Dest>
202     void dgmres(const MatrixType& mat,const Rhs& rhs, Dest& x, const Preconditioner& precond) const;
203     // Perform one cycle of GMRES
204     template<typename Dest>
205     int dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, int& nbIts) const;
206     // Compute data to use for deflation
207     int dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const;
208     // Apply deflation to a vector
209     template<typename RhsType, typename DestType>
210     int dgmresApplyDeflation(const RhsType& In, DestType& Out) const;
211     ComplexVector schurValues(const ComplexSchur<DenseMatrix>& schurofH) const;
212     ComplexVector schurValues(const RealSchur<DenseMatrix>& schurofH) const;
213     // Init data for deflation
214     void dgmresInitDeflation(Index& rows) const;
215     mutable DenseMatrix m_V; // Krylov basis vectors
216     mutable DenseMatrix m_H; // Hessenberg matrix
217     mutable DenseMatrix m_Hes; // Initial hessenberg matrix wihout Givens rotations applied
218     mutable Index m_restart; // Maximum size of the Krylov subspace
219     mutable DenseMatrix m_U; // Vectors that form the basis of the invariant subspace
220     mutable DenseMatrix m_MU; // matrix operator applied to m_U (for next cycles)
221     mutable DenseMatrix m_T; /* T=U^T*M^{-1}*A*U */
222     mutable PartialPivLU<DenseMatrix> m_luT; // LU factorization of m_T
223     mutable StorageIndex m_neig; //Number of eigenvalues to extract at each restart
224     mutable int m_r; // Current number of deflated eigenvalues, size of m_U
225     mutable int m_maxNeig; // Maximum number of eigenvalues to deflate
226     mutable RealScalar m_lambdaN; //Modulus of the largest eigenvalue of A
227     mutable bool m_isDeflAllocated;
228     mutable bool m_isDeflInitialized;
229 
230     //Adaptive strategy
231     mutable RealScalar m_smv; // Smaller multiple of the remaining number of steps allowed
232     mutable bool m_force; // Force the use of deflation at each restart
233 
234 };
235 /**
236  * \brief Perform several cycles of restarted GMRES with modified Gram Schmidt,
237  *
238  * A right preconditioner is used combined with deflation.
239  *
240  */
241 template< typename _MatrixType, typename _Preconditioner>
242 template<typename Rhs, typename Dest>
243 void DGMRES<_MatrixType, _Preconditioner>::dgmres(const MatrixType& mat,const Rhs& rhs, Dest& x,
244               const Preconditioner& precond) const
245 {
246   //Initialization
247   int n = mat.rows();
248   DenseVector r0(n);
249   int nbIts = 0;
250   m_H.resize(m_restart+1, m_restart);
251   m_Hes.resize(m_restart, m_restart);
252   m_V.resize(n,m_restart+1);
253   //Initial residual vector and intial norm
254   x = precond.solve(x);
255   r0 = rhs - mat * x;
256   RealScalar beta = r0.norm();
257   RealScalar normRhs = rhs.norm();
258   m_error = beta/normRhs;
259   if(m_error < m_tolerance)
260     m_info = Success;
261   else
262     m_info = NoConvergence;
263 
264   // Iterative process
265   while (nbIts < m_iterations && m_info == NoConvergence)
266   {
267     dgmresCycle(mat, precond, x, r0, beta, normRhs, nbIts);
268 
269     // Compute the new residual vector for the restart
270     if (nbIts < m_iterations && m_info == NoConvergence)
271       r0 = rhs - mat * x;
272   }
273 }
274 
275 /**
276  * \brief Perform one restart cycle of DGMRES
277  * \param mat The coefficient matrix
278  * \param precond The preconditioner
279  * \param x the new approximated solution
280  * \param r0 The initial residual vector
281  * \param beta The norm of the residual computed so far
282  * \param normRhs The norm of the right hand side vector
283  * \param nbIts The number of iterations
284  */
285 template< typename _MatrixType, typename _Preconditioner>
286 template<typename Dest>
287 int DGMRES<_MatrixType, _Preconditioner>::dgmresCycle(const MatrixType& mat, const Preconditioner& precond, Dest& x, DenseVector& r0, RealScalar& beta, const RealScalar& normRhs, int& nbIts) const
288 {
289   //Initialization
290   DenseVector g(m_restart+1); // Right hand side of the least square problem
291   g.setZero();
292   g(0) = Scalar(beta);
293   m_V.col(0) = r0/beta;
294   m_info = NoConvergence;
295   std::vector<JacobiRotation<Scalar> >gr(m_restart); // Givens rotations
296   int it = 0; // Number of inner iterations
297   int n = mat.rows();
298   DenseVector tv1(n), tv2(n);  //Temporary vectors
299   while (m_info == NoConvergence && it < m_restart && nbIts < m_iterations)
300   {
301     // Apply preconditioner(s) at right
302     if (m_isDeflInitialized )
303     {
304       dgmresApplyDeflation(m_V.col(it), tv1); // Deflation
305       tv2 = precond.solve(tv1);
306     }
307     else
308     {
309       tv2 = precond.solve(m_V.col(it)); // User's selected preconditioner
310     }
311     tv1 = mat * tv2;
312 
313     // Orthogonalize it with the previous basis in the basis using modified Gram-Schmidt
314     Scalar coef;
315     for (int i = 0; i <= it; ++i)
316     {
317       coef = tv1.dot(m_V.col(i));
318       tv1 = tv1 - coef * m_V.col(i);
319       m_H(i,it) = coef;
320       m_Hes(i,it) = coef;
321     }
322     // Normalize the vector
323     coef = tv1.norm();
324     m_V.col(it+1) = tv1/coef;
325     m_H(it+1, it) = coef;
326 //     m_Hes(it+1,it) = coef;
327 
328     // FIXME Check for happy breakdown
329 
330     // Update Hessenberg matrix with Givens rotations
331     for (int i = 1; i <= it; ++i)
332     {
333       m_H.col(it).applyOnTheLeft(i-1,i,gr[i-1].adjoint());
334     }
335     // Compute the new plane rotation
336     gr[it].makeGivens(m_H(it, it), m_H(it+1,it));
337     // Apply the new rotation
338     m_H.col(it).applyOnTheLeft(it,it+1,gr[it].adjoint());
339     g.applyOnTheLeft(it,it+1, gr[it].adjoint());
340 
341     beta = std::abs(g(it+1));
342     m_error = beta/normRhs;
343     // std::cerr << nbIts << " Relative Residual Norm " << m_error << std::endl;
344     it++; nbIts++;
345 
346     if (m_error < m_tolerance)
347     {
348       // The method has converged
349       m_info = Success;
350       break;
351     }
352   }
353 
354   // Compute the new coefficients by solving the least square problem
355 //   it++;
356   //FIXME  Check first if the matrix is singular ... zero diagonal
357   DenseVector nrs(m_restart);
358   nrs = m_H.topLeftCorner(it,it).template triangularView<Upper>().solve(g.head(it));
359 
360   // Form the new solution
361   if (m_isDeflInitialized)
362   {
363     tv1 = m_V.leftCols(it) * nrs;
364     dgmresApplyDeflation(tv1, tv2);
365     x = x + precond.solve(tv2);
366   }
367   else
368     x = x + precond.solve(m_V.leftCols(it) * nrs);
369 
370   // Go for a new cycle and compute data for deflation
371   if(nbIts < m_iterations && m_info == NoConvergence && m_neig > 0 && (m_r+m_neig) < m_maxNeig)
372     dgmresComputeDeflationData(mat, precond, it, m_neig);
373   return 0;
374 
375 }
376 
377 
378 template< typename _MatrixType, typename _Preconditioner>
379 void DGMRES<_MatrixType, _Preconditioner>::dgmresInitDeflation(Index& rows) const
380 {
381   m_U.resize(rows, m_maxNeig);
382   m_MU.resize(rows, m_maxNeig);
383   m_T.resize(m_maxNeig, m_maxNeig);
384   m_lambdaN = 0.0;
385   m_isDeflAllocated = true;
386 }
387 
388 template< typename _MatrixType, typename _Preconditioner>
389 inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_MatrixType, _Preconditioner>::schurValues(const ComplexSchur<DenseMatrix>& schurofH) const
390 {
391   return schurofH.matrixT().diagonal();
392 }
393 
394 template< typename _MatrixType, typename _Preconditioner>
395 inline typename DGMRES<_MatrixType, _Preconditioner>::ComplexVector DGMRES<_MatrixType, _Preconditioner>::schurValues(const RealSchur<DenseMatrix>& schurofH) const
396 {
397   typedef typename MatrixType::Index Index;
398   const DenseMatrix& T = schurofH.matrixT();
399   Index it = T.rows();
400   ComplexVector eig(it);
401   Index j = 0;
402   while (j < it-1)
403   {
404     if (T(j+1,j) ==Scalar(0))
405     {
406       eig(j) = std::complex<RealScalar>(T(j,j),RealScalar(0));
407       j++;
408     }
409     else
410     {
411       eig(j) = std::complex<RealScalar>(T(j,j),T(j+1,j));
412       eig(j+1) = std::complex<RealScalar>(T(j,j+1),T(j+1,j+1));
413       j++;
414     }
415   }
416   if (j < it-1) eig(j) = std::complex<RealScalar>(T(j,j),RealScalar(0));
417   return eig;
418 }
419 
420 template< typename _MatrixType, typename _Preconditioner>
421 int DGMRES<_MatrixType, _Preconditioner>::dgmresComputeDeflationData(const MatrixType& mat, const Preconditioner& precond, const Index& it, StorageIndex& neig) const
422 {
423   // First, find the Schur form of the Hessenberg matrix H
424   typename internal::conditional<NumTraits<Scalar>::IsComplex, ComplexSchur<DenseMatrix>, RealSchur<DenseMatrix> >::type schurofH;
425   bool computeU = true;
426   DenseMatrix matrixQ(it,it);
427   matrixQ.setIdentity();
428   schurofH.computeFromHessenberg(m_Hes.topLeftCorner(it,it), matrixQ, computeU);
429 
430   ComplexVector eig(it);
431   Matrix<StorageIndex,Dynamic,1>perm(it);
432   eig = this->schurValues(schurofH);
433 
434   // Reorder the absolute values of Schur values
435   DenseRealVector modulEig(it);
436   for (int j=0; j<it; ++j) modulEig(j) = std::abs(eig(j));
437   perm.setLinSpaced(it,0,it-1);
438   internal::sortWithPermutation(modulEig, perm, neig);
439 
440   if (!m_lambdaN)
441   {
442     m_lambdaN = (std::max)(modulEig.maxCoeff(), m_lambdaN);
443   }
444   //Count the real number of extracted eigenvalues (with complex conjugates)
445   int nbrEig = 0;
446   while (nbrEig < neig)
447   {
448     if(eig(perm(it-nbrEig-1)).imag() == RealScalar(0)) nbrEig++;
449     else nbrEig += 2;
450   }
451   // Extract the  Schur vectors corresponding to the smallest Ritz values
452   DenseMatrix Sr(it, nbrEig);
453   Sr.setZero();
454   for (int j = 0; j < nbrEig; j++)
455   {
456     Sr.col(j) = schurofH.matrixU().col(perm(it-j-1));
457   }
458 
459   // Form the Schur vectors of the initial matrix using the Krylov basis
460   DenseMatrix X;
461   X = m_V.leftCols(it) * Sr;
462   if (m_r)
463   {
464    // Orthogonalize X against m_U using modified Gram-Schmidt
465    for (int j = 0; j < nbrEig; j++)
466      for (int k =0; k < m_r; k++)
467       X.col(j) = X.col(j) - (m_U.col(k).dot(X.col(j)))*m_U.col(k);
468   }
469 
470   // Compute m_MX = A * M^-1 * X
471   Index m = m_V.rows();
472   if (!m_isDeflAllocated)
473     dgmresInitDeflation(m);
474   DenseMatrix MX(m, nbrEig);
475   DenseVector tv1(m);
476   for (int j = 0; j < nbrEig; j++)
477   {
478     tv1 = mat * X.col(j);
479     MX.col(j) = precond.solve(tv1);
480   }
481 
482   //Update m_T = [U'MU U'MX; X'MU X'MX]
483   m_T.block(m_r, m_r, nbrEig, nbrEig) = X.transpose() * MX;
484   if(m_r)
485   {
486     m_T.block(0, m_r, m_r, nbrEig) = m_U.leftCols(m_r).transpose() * MX;
487     m_T.block(m_r, 0, nbrEig, m_r) = X.transpose() * m_MU.leftCols(m_r);
488   }
489 
490   // Save X into m_U and m_MX in m_MU
491   for (int j = 0; j < nbrEig; j++) m_U.col(m_r+j) = X.col(j);
492   for (int j = 0; j < nbrEig; j++) m_MU.col(m_r+j) = MX.col(j);
493   // Increase the size of the invariant subspace
494   m_r += nbrEig;
495 
496   // Factorize m_T into m_luT
497   m_luT.compute(m_T.topLeftCorner(m_r, m_r));
498 
499   //FIXME CHeck if the factorization was correctly done (nonsingular matrix)
500   m_isDeflInitialized = true;
501   return 0;
502 }
503 template<typename _MatrixType, typename _Preconditioner>
504 template<typename RhsType, typename DestType>
505 int DGMRES<_MatrixType, _Preconditioner>::dgmresApplyDeflation(const RhsType &x, DestType &y) const
506 {
507   DenseVector x1 = m_U.leftCols(m_r).transpose() * x;
508   y = x + m_U.leftCols(m_r) * ( m_lambdaN * m_luT.solve(x1) - x1);
509   return 0;
510 }
511 
512 } // end namespace Eigen
513 #endif
514