solvers.hh

Go to the documentation of this file.

// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
// vi: set et ts=4 sw=2 sts=2:
 
#ifndef DUNE_ISTL_SOLVERS_HH
#define DUNE_ISTL_SOLVERS_HH
 
#include <array>
#include <cmath>
#include <complex>
#include <iostream>
#include <memory>
#include <type_traits>
#include <vector>
 
#include <dune/common/exceptions.hh>
#include <dune/common/math.hh>
#include <dune/common/simd/io.hh>
#include <dune/common/simd/simd.hh>
#include <dune/common/std/type_traits.hh>
#include <dune/common/timer.hh>
 
#include <dune/istl/allocator.hh>
#include <dune/istl/bcrsmatrix.hh>
#include <dune/istl/eigenvalue/arpackpp.hh>
#include <dune/istl/istlexception.hh>
#include <dune/istl/operators.hh>
#include <dune/istl/preconditioner.hh>
#include <dune/istl/scalarproducts.hh>
#include <dune/istl/solver.hh>
#include <dune/istl/solverregistry.hh>
 
namespace Dune {
   //=====================================================================
  // Implementation of this interface
  //=====================================================================
 
  template<class X>
  class LoopSolver : public IterativeSolver<X,X> {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
    // copy base class constructors
    using IterativeSolver<X,X>::IterativeSolver;
 
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      Iteration iteration(*this, res);
      _prec->pre(x,b);
 
      // overwrite b with defect
      _op->applyscaleadd(-1,x,b);
 
      // compute norm, \todo parallelization
      real_type def = _sp->norm(b);
      if(iteration.step(0, def)){
        _prec->post(x);
        return;
      }
      // prepare preconditioner
 
      // allocate correction vector
      X v(x);
 
      // iteration loop
      int i=1;
      for ( ; i<=_maxit; i++ )
      {
        v = 0;                      // clear correction
        _prec->apply(v,b);           // apply preconditioner
        x += v;                     // update solution
        _op->applyscaleadd(-1,v,b);  // update defect
        def=_sp->norm(b);  // comp defect norm
        if(iteration.step(i, def))
          break;
      }
 
      // postprocess preconditioner
      _prec->post(x);
    }
 
  protected:
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("loopsolver", defaultIterativeSolverCreator<Dune::LoopSolver>());
 
 
  // all these solvers are taken from the SUMO library
  template<class X>
  class GradientSolver : public IterativeSolver<X,X> {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
    // copy base class constructors
    using IterativeSolver<X,X>::IterativeSolver;
 
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      Iteration iteration(*this, res);
      _prec->pre(x,b);             // prepare preconditioner
 
      _op->applyscaleadd(-1,x,b);  // overwrite b with defec
 
      real_type def = _sp->norm(b); // compute norm
      if(iteration.step(0, def)){
        _prec->post(x);
        return;
      }
 
      X p(x);                     // create local vectors
      X q(b);
 
      int i=1;   // loop variables
      field_type lambda;
      for ( ; i<=_maxit; i++ )
      {
        p = 0;                      // clear correction
        _prec->apply(p,b);           // apply preconditioner
        _op->apply(p,q);             // q=Ap
        lambda = _sp->dot(p,b)/_sp->dot(q,p); // minimization
        x.axpy(lambda,p);           // update solution
        b.axpy(-lambda,q);          // update defect
 
        def =_sp->norm(b); // comp defect norm
        if(iteration.step(i, def))
          break;
      }
      // postprocess preconditioner
      _prec->post(x);
    }
 
  protected:
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("gradientsolver", defaultIterativeSolverCreator<Dune::GradientSolver>());
 
  template<class X>
  class CGSolver : public IterativeSolver<X,X> {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
    // copy base class constructors
    using IterativeSolver<X,X>::IterativeSolver;
 
  private:
    using typename IterativeSolver<X,X>::scalar_real_type;
 
  protected:
 
    static constexpr bool enableConditionEstimate = (std::is_same_v<field_type,float> || std::is_same_v<field_type,double>);
    using enableConditionEstimate_t [[deprecated("Use bool enableConditionEstimate instead. Will be removed after Dune 2.8")]]
      = std::bool_constant<enableConditionEstimate>;
 
  public:
 
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
 
    CGSolver (LinearOperator<X,X>& op, Preconditioner<X,X>& prec,
      scalar_real_type reduction, int maxit, int verbose, bool condition_estimate) : IterativeSolver<X,X>(op, prec, reduction, maxit, verbose),
      condition_estimate_(condition_estimate)
    {
      if (condition_estimate && !enableConditionEstimate) {
        condition_estimate_ = false;
        std::cerr << "WARNING: Condition estimate was disabled. It is only available for double and float field types!" << std::endl;
      }
    }
 
    CGSolver (LinearOperator<X,X>& op, ScalarProduct<X>& sp, Preconditioner<X,X>& prec,
      scalar_real_type reduction, int maxit, int verbose, bool condition_estimate) : IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose),
      condition_estimate_(condition_estimate)
    {
      if (condition_estimate && !(std::is_same<field_type,float>::value || std::is_same<field_type,double>::value)) {
        condition_estimate_ = false;
        std::cerr << "WARNING: Condition estimate was disabled. It is only available for double and float field types!" << std::endl;
      }
    }
 
    CGSolver (std::shared_ptr<LinearOperator<X,X>> op, std::shared_ptr<ScalarProduct<X>> sp,
              std::shared_ptr<Preconditioner<X,X>> prec,
              scalar_real_type reduction, int maxit, int verbose, bool condition_estimate)
      : IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose),
      condition_estimate_(condition_estimate)
    {
      if (condition_estimate && !(std::is_same<field_type,float>::value || std::is_same<field_type,double>::value)) {
        condition_estimate_ = false;
        std::cerr << "WARNING: Condition estimate was disabled. It is only available for double and float field types!" << std::endl;
      }
    }
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      Iteration iteration(*this,res);
      _prec->pre(x,b);             // prepare preconditioner
 
      _op->applyscaleadd(-1,x,b);  // overwrite b with defect
 
      real_type def = _sp->norm(b); // compute norm
      if(iteration.step(0, def)){
        _prec->post(x);
        return;
      }
 
      X p(x);              // the search direction
      X q(x);              // a temporary vector
 
      // Remember lambda and beta values for condition estimate
      std::vector<real_type> lambdas(0);
      std::vector<real_type> betas(0);
 
      // some local variables
      field_type rho,rholast,lambda,alpha,beta;
 
      // determine initial search direction
      p = 0;                          // clear correction
      _prec->apply(p,b);               // apply preconditioner
      rholast = _sp->dot(p,b);         // orthogonalization
 
      // the loop
      int i=1;
      for ( ; i<=_maxit; i++ )
      {
        // minimize in given search direction p
        _op->apply(p,q);             // q=Ap
        alpha = _sp->dot(p,q);       // scalar product
        lambda = rholast/alpha;     // minimization
        if constexpr (enableConditionEstimate)
          if (condition_estimate_)
            lambdas.push_back(std::real(lambda));
        x.axpy(lambda,p);           // update solution
        b.axpy(-lambda,q);          // update defect
 
        // convergence test
        def=_sp->norm(b); // comp defect norm
        if(iteration.step(i, def))
          break;
 
        // determine new search direction
        q = 0;                      // clear correction
        _prec->apply(q,b);           // apply preconditioner
        rho = _sp->dot(q,b);         // orthogonalization
        beta = rho/rholast;         // scaling factor
        if constexpr (enableConditionEstimate)
          if (condition_estimate_)
            betas.push_back(std::real(beta));
        p *= beta;                  // scale old search direction
        p += q;                     // orthogonalization with correction
        rholast = rho;              // remember rho for recurrence
      }
 
      _prec->post(x);                  // postprocess preconditioner
 
      if (condition_estimate_) {
#if HAVE_ARPACKPP
        if constexpr (enableConditionEstimate) {
          using std::sqrt;
 
          // Build T matrix which has extreme eigenvalues approximating
          // those of the original system
          // (see Y. Saad, Iterative methods for sparse linear systems)
 
          COND_MAT T(i, i, COND_MAT::row_wise);
 
          for (auto row = T.createbegin(); row != T.createend(); ++row) {
            if (row.index() > 0)
              row.insert(row.index()-1);
            row.insert(row.index());
            if (row.index() < T.N() - 1)
              row.insert(row.index()+1);
          }
          for (int row = 0; row < i; ++row) {
            if (row > 0) {
              T[row][row-1] = sqrt(betas[row-1]) / lambdas[row-1];
            }
 
            T[row][row] = 1.0 / lambdas[row];
            if (row > 0) {
              T[row][row] += betas[row-1] / lambdas[row-1];
            }
 
            if (row < i - 1) {
              T[row][row+1] = sqrt(betas[row]) / lambdas[row];
            }
          }
 
          // Compute largest and smallest eigenvalue of T matrix and return as estimate
          Dune::ArPackPlusPlus_Algorithms<COND_MAT, COND_VEC> arpack(T);
 
          real_type eps = 0.0;
          COND_VEC eigv;
          real_type min_eigv, max_eigv;
          arpack.computeSymMinMagnitude (eps, eigv, min_eigv);
          arpack.computeSymMaxMagnitude (eps, eigv, max_eigv);
 
          res.condition_estimate = max_eigv / min_eigv;
 
          if (this->_verbose > 0) {
            std::cout << "Min eigv estimate: " << Simd::io(min_eigv) << '\n';
            std::cout << "Max eigv estimate: " << Simd::io(max_eigv) << '\n';
            std::cout << "Condition estimate: "
                      << Simd::io(max_eigv / min_eigv) << std::endl;
          }
        }
#else
      std::cerr << "WARNING: Condition estimate was requested. This requires ARPACK, but ARPACK was not found!" << std::endl;
#endif
      }
    }
 
  private:
    bool condition_estimate_ = false;
 
    // Matrix and vector types used for condition estimate
    typedef Dune::BCRSMatrix<Dune::FieldMatrix<real_type,1,1> > COND_MAT;
    typedef Dune::BlockVector<Dune::FieldVector<real_type,1> > COND_VEC;
 
  protected:
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("cgsolver", defaultIterativeSolverCreator<Dune::CGSolver>());
 
  // Ronald Kriemanns BiCG-STAB implementation from Sumo
  template<class X>
  class BiCGSTABSolver : public IterativeSolver<X,X> {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
    // copy base class constructors
    using IterativeSolver<X,X>::IterativeSolver;
 
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      using std::abs;
      const Simd::Scalar<real_type> EPSILON=1e-80;
      using std::abs;
      double it;
      field_type rho, rho_new, alpha, beta, h, omega;
      real_type norm;
 
      //
      // get vectors and matrix
      //
      X& r=b;
      X p(x);
      X v(x);
      X t(x);
      X y(x);
      X rt(x);
 
      //
      // begin iteration
      //
 
      // r = r - Ax; rt = r
      Iteration<double> iteration(*this,res);
      _prec->pre(x,r);             // prepare preconditioner
 
      _op->applyscaleadd(-1,x,r);  // overwrite b with defect
 
      rt=r;
 
      norm = _sp->norm(r);
      if(iteration.step(0, norm)){
        _prec->post(x);
        return;
      }
      p=0;
      v=0;
 
      rho   = 1;
      alpha = 1;
      omega = 1;
 
      //
      // iteration
      //
 
      for (it = 0.5; it < _maxit; it+=.5)
      {
        //
        // preprocess, set vecsizes etc.
        //
 
        // rho_new = < rt , r >
        rho_new = _sp->dot(rt,r);
 
        // look if breakdown occurred
        if (Simd::allTrue(abs(rho) <= EPSILON))
          DUNE_THROW(SolverAbort,"breakdown in BiCGSTAB - rho "
                     << Simd::io(rho) << " <= EPSILON " << EPSILON
                     << " after " << it << " iterations");
        if (Simd::allTrue(abs(omega) <= EPSILON))
          DUNE_THROW(SolverAbort,"breakdown in BiCGSTAB - omega "
                     << Simd::io(omega) << " <= EPSILON " << EPSILON
                     << " after " << it << " iterations");
 
 
        if (it<1)
          p = r;
        else
        {
          beta = ( rho_new / rho ) * ( alpha / omega );
          p.axpy(-omega,v); // p = r + beta (p - omega*v)
          p *= beta;
          p += r;
        }
 
        // y = W^-1 * p
        y = 0;
        _prec->apply(y,p);           // apply preconditioner
 
        // v = A * y
        _op->apply(y,v);
 
        // alpha = rho_new / < rt, v >
        h = _sp->dot(rt,v);
 
        if ( Simd::allTrue(abs(h) < EPSILON) )
          DUNE_THROW(SolverAbort,"abs(h) < EPSILON in BiCGSTAB - abs(h) "
                     << Simd::io(abs(h)) << " < EPSILON " << EPSILON
                     << " after " << it << " iterations");
 
        alpha = rho_new / h;
 
        // apply first correction to x
        // x <- x + alpha y
        x.axpy(alpha,y);
 
        // r = r - alpha*v
        r.axpy(-alpha,v);
 
        //
        // test stop criteria
        //
 
        norm = _sp->norm(r);
        if(iteration.step(it, norm)){
          break;
        }
 
        it+=.5;
 
        // y = W^-1 * r
        y = 0;
        _prec->apply(y,r);
 
        // t = A * y
        _op->apply(y,t);
 
        // omega = < t, r > / < t, t >
        omega = _sp->dot(t,r)/_sp->dot(t,t);
 
        // apply second correction to x
        // x <- x + omega y
        x.axpy(omega,y);
 
        // r = s - omega*t (remember : r = s)
        r.axpy(-omega,t);
 
        rho = rho_new;
 
        //
        // test stop criteria
        //
 
        norm = _sp->norm(r);
        if(iteration.step(it, norm)){
          break;
        }
      } // end for
 
      _prec->post(x);                  // postprocess preconditioner
    }
 
  protected:
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    template<class CountType>
    using Iteration = typename IterativeSolver<X,X>::template Iteration<CountType>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("bicgstabsolver", defaultIterativeSolverCreator<Dune::BiCGSTABSolver>());
 
  template<class X>
  class MINRESSolver : public IterativeSolver<X,X> {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
    // copy base class constructors
    using IterativeSolver<X,X>::IterativeSolver;
 
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      using std::sqrt;
      using std::abs;
      Iteration iteration(*this, res);
      // prepare preconditioner
      _prec->pre(x,b);
 
      // overwrite rhs with defect
      _op->applyscaleadd(-1,x,b);
 
      // compute residual norm
      real_type def = _sp->norm(b);
      if(iteration.step(0, def)){
        _prec->post(x);
        return;
      }
 
      // recurrence coefficients as computed in Lanczos algorithm
      field_type alpha, beta;
        // diagonal entries of givens rotation
      std::array<real_type,2> c{{0.0,0.0}};
        // off-diagonal entries of givens rotation
      std::array<field_type,2> s{{0.0,0.0}};
 
      // recurrence coefficients (column k of tridiag matrix T_k)
      std::array<field_type,3> T{{0.0,0.0,0.0}};
 
      // the rhs vector of the min problem
      std::array<field_type,2> xi{{1.0,0.0}};
 
      // some temporary vectors
      X z(b), dummy(b);
 
      // initialize and clear correction
      z = 0.0;
      _prec->apply(z,b);
 
      // beta is real and positive in exact arithmetic
      // since it is the norm of the basis vectors (in unpreconditioned case)
      beta = sqrt(_sp->dot(b,z));
      field_type beta0 = beta;
 
      // the search directions
      std::array<X,3> p{{b,b,b}};
      p[0] = 0.0;
      p[1] = 0.0;
      p[2] = 0.0;
 
      // orthonormal basis vectors (in unpreconditioned case)
      std::array<X,3> q{{b,b,b}};
      q[0] = 0.0;
      q[1] *= real_type(1.0)/beta;
      q[2] = 0.0;
 
      z *= real_type(1.0)/beta;
 
      // the loop
      int i = 1;
      for( ; i<=_maxit; i++) {
 
        dummy = z;
        int i1 = i%3,
          i0 = (i1+2)%3,
          i2 = (i1+1)%3;
 
        // symmetrically preconditioned Lanczos algorithm (see Greenbaum p.121)
        _op->apply(z,q[i2]); // q[i2] = Az
        q[i2].axpy(-beta,q[i0]);
        // alpha is real since it is the diagonal entry of the hermitian tridiagonal matrix
        // from the Lanczos Algorithm
        // so the order in the scalar product doesn't matter even for the complex case
        alpha = _sp->dot(z,q[i2]);
        q[i2].axpy(-alpha,q[i1]);
 
        z = 0.0;
        _prec->apply(z,q[i2]);
 
        // beta is real and positive in exact arithmetic
        // since it is the norm of the basis vectors (in unpreconditioned case)
        beta = sqrt(_sp->dot(q[i2],z));
 
        q[i2] *= real_type(1.0)/beta;
        z *= real_type(1.0)/beta;
 
        // QR Factorization of recurrence coefficient matrix
        // apply previous givens rotations to last column of T
        T[1] = T[2];
        if(i>2) {
          T[0] = s[i%2]*T[1];
          T[1] = c[i%2]*T[1];
        }
        if(i>1) {
          T[2] = c[(i+1)%2]*alpha - s[(i+1)%2]*T[1];
          T[1] = c[(i+1)%2]*T[1] + s[(i+1)%2]*alpha;
        }
        else
          T[2] = alpha;
 
        // update QR factorization
        generateGivensRotation(T[2],beta,c[i%2],s[i%2]);
        // to last column of T_k
        T[2] = c[i%2]*T[2] + s[i%2]*beta;
        // and to the rhs xi of the min problem
        xi[i%2] = -s[i%2]*xi[(i+1)%2];
        xi[(i+1)%2] *= c[i%2];
 
        // compute correction direction
        p[i2] = dummy;
        p[i2].axpy(-T[1],p[i1]);
        p[i2].axpy(-T[0],p[i0]);
        p[i2] *= real_type(1.0)/T[2];
 
        // apply correction/update solution
        x.axpy(beta0*xi[(i+1)%2],p[i2]);
 
        // remember beta_old
        T[2] = beta;
 
        // check for convergence
        // the last entry in the rhs of the min-problem is the residual
        def = abs(beta0*xi[i%2]);
        if(iteration.step(i, def)){
          break;
        }
      } // end for
 
        // postprocess preconditioner
        _prec->post(x);
    }
 
  private:
 
    void generateGivensRotation(field_type &dx, field_type &dy, real_type &cs, field_type &sn)
    {
      using std::sqrt;
      using std::abs;
      using std::max;
      using std::min;
      const real_type eps = 1e-15;
      real_type norm_dx = abs(dx);
      real_type norm_dy = abs(dy);
      real_type norm_max = max(norm_dx, norm_dy);
      real_type norm_min = min(norm_dx, norm_dy);
      real_type temp = norm_min/norm_max;
      // we rewrite the code in a vectorizable fashion
      cs = Simd::cond(norm_dy < eps,
        real_type(1.0),
        Simd::cond(norm_dx < eps,
          real_type(0.0),
          Simd::cond(norm_dy > norm_dx,
            real_type(1.0)/sqrt(real_type(1.0) + temp*temp)*temp,
            real_type(1.0)/sqrt(real_type(1.0) + temp*temp)
          )));
      sn = Simd::cond(norm_dy < eps,
        field_type(0.0),
        Simd::cond(norm_dx < eps,
          field_type(1.0),
          Simd::cond(norm_dy > norm_dx,
            // dy and dx are real in exact arithmetic
            // thus dx*dy is real so we can explicitly enforce it
            field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*dx*dy/norm_dx/norm_dy,
            // dy and dx is real in exact arithmetic
            // so we don't have to conjugate both of them
            field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*dy/dx
          )));
    }
 
  protected:
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("minressolver", defaultIterativeSolverCreator<Dune::MINRESSolver>());
 
  template<class X, class Y=X, class F = Y>
  class RestartedGMResSolver : public IterativeSolver<X,Y>
  {
  public:
    using typename IterativeSolver<X,Y>::domain_type;
    using typename IterativeSolver<X,Y>::range_type;
    using typename IterativeSolver<X,Y>::field_type;
    using typename IterativeSolver<X,Y>::real_type;
 
  protected:
    using typename IterativeSolver<X,X>::scalar_real_type;
 
    using fAlloc = ReboundAllocatorType<X,field_type>;
    using rAlloc = ReboundAllocatorType<X,real_type>;
 
  public:
 
    RestartedGMResSolver (LinearOperator<X,Y>& op, Preconditioner<X,Y>& prec, scalar_real_type reduction, int restart, int maxit, int verbose) :
      IterativeSolver<X,Y>::IterativeSolver(op,prec,reduction,maxit,verbose),
      _restart(restart)
    {}
 
    RestartedGMResSolver (LinearOperator<X,Y>& op, ScalarProduct<X>& sp, Preconditioner<X,Y>& prec, scalar_real_type reduction, int restart, int maxit, int verbose) :
      IterativeSolver<X,Y>::IterativeSolver(op,sp,prec,reduction,maxit,verbose),
      _restart(restart)
    {}
 
    RestartedGMResSolver (std::shared_ptr<LinearOperator<X,Y> > op, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) :
      IterativeSolver<X,Y>::IterativeSolver(op,prec,configuration),
      _restart(configuration.get<int>("restart"))
    {}
 
    RestartedGMResSolver (std::shared_ptr<LinearOperator<X,Y> > op, std::shared_ptr<ScalarProduct<X> > sp, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) :
      IterativeSolver<X,Y>::IterativeSolver(op,sp,prec,configuration),
      _restart(configuration.get<int>("restart"))
    {}
 
    RestartedGMResSolver (std::shared_ptr<LinearOperator<X,Y>> op,
                          std::shared_ptr<ScalarProduct<X>> sp,
                          std::shared_ptr<Preconditioner<X,Y>> prec,
                          scalar_real_type reduction, int restart, int maxit, int verbose) :
      IterativeSolver<X,Y>::IterativeSolver(op,sp,prec,reduction,maxit,verbose),
      _restart(restart)
    {}
 
    virtual void apply (X& x, Y& b, InverseOperatorResult& res)
    {
      apply(x,b,Simd::max(_reduction),res);
    }
 
    virtual void apply (X& x, Y& b, [[maybe_unused]] double reduction, InverseOperatorResult& res)
    {
      using std::abs;
      const Simd::Scalar<real_type> EPSILON = 1e-80;
      const int m = _restart;
      real_type norm = 0.0;
      int j = 1;
      std::vector<field_type,fAlloc> s(m+1), sn(m);
      std::vector<real_type,rAlloc> cs(m);
      // need copy of rhs if GMRes has to be restarted
      Y b2(b);
      // helper vector
      Y w(b);
      std::vector< std::vector<field_type,fAlloc> > H(m+1,s);
      std::vector<F> v(m+1,b);
 
      Iteration iteration(*this,res);
 
      // clear solver statistics and set res.converged to false
      _prec->pre(x,b);
 
      // calculate defect and overwrite rhs with it
      _op->applyscaleadd(-1.0,x,b); // b -= Ax
      // calculate preconditioned defect
      v[0] = 0.0; _prec->apply(v[0],b); // r = W^-1 b
      norm = _sp->norm(v[0]);
      if(iteration.step(0, norm)){
        _prec->post(x);
        return;
      }
 
      while(j <= _maxit && res.converged != true) {
 
        int i = 0;
        v[0] *= real_type(1.0)/norm;
        s[0] = norm;
        for(i=1; i<m+1; i++)
          s[i] = 0.0;
 
        for(i=0; i < m && j <= _maxit && res.converged != true; i++, j++) {
          w = 0.0;
          // use v[i+1] as temporary vector
          v[i+1] = 0.0;
          // do Arnoldi algorithm
          _op->apply(v[i],v[i+1]);
          _prec->apply(w,v[i+1]);
          for(int k=0; k<i+1; k++) {
            // notice that _sp->dot(v[k],w) = v[k]\adjoint w
            // so one has to pay attention to the order
            // in the scalar product for the complex case
            // doing the modified Gram-Schmidt algorithm
            H[k][i] = _sp->dot(v[k],w);
            // w -= H[k][i] * v[k]
            w.axpy(-H[k][i],v[k]);
          }
          H[i+1][i] = _sp->norm(w);
          if(Simd::allTrue(abs(H[i+1][i]) < EPSILON))
            DUNE_THROW(SolverAbort,
                       "breakdown in GMRes - |w| == 0.0 after " << j << " iterations");
 
          // normalize new vector
          v[i+1] = w; v[i+1] *= real_type(1.0)/H[i+1][i];
 
          // update QR factorization
          for(int k=0; k<i; k++)
            applyPlaneRotation(H[k][i],H[k+1][i],cs[k],sn[k]);
 
          // compute new givens rotation
          generatePlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]);
          // finish updating QR factorization
          applyPlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]);
          applyPlaneRotation(s[i],s[i+1],cs[i],sn[i]);
 
          // norm of the defect is the last component the vector s
          norm = abs(s[i+1]);
 
          iteration.step(j, norm);
 
        } // end for
 
        // calculate update vector
        w = 0.0;
        update(w,i,H,s,v);
        // and current iterate
        x += w;
 
        // restart GMRes if convergence was not achieved,
        // i.e. linear defect has not reached desired reduction
        // and if j < _maxit (do not restart on last iteration)
        if( res.converged != true && j < _maxit ) {
 
          if(_verbose > 0)
            std::cout << "=== GMRes::restart" << std::endl;
          // get saved rhs
          b = b2;
          // calculate new defect
          _op->applyscaleadd(-1.0,x,b); // b -= Ax;
          // calculate preconditioned defect
          v[0] = 0.0;
          _prec->apply(v[0],b);
          norm = _sp->norm(v[0]);
        }
 
      } //end while
 
      // postprocess preconditioner
      _prec->post(x);
    }
 
  protected :
 
    void update(X& w, int i,
                const std::vector<std::vector<field_type,fAlloc> >& H,
                const std::vector<field_type,fAlloc>& s,
                const std::vector<X>& v) {
      // solution vector of the upper triangular system
      std::vector<field_type,fAlloc> y(s);
 
      // backsolve
      for(int a=i-1; a>=0; a--) {
        field_type rhs(s[a]);
        for(int b=a+1; b<i; b++)
          rhs -= H[a][b]*y[b];
        y[a] = rhs/H[a][a];
 
        // compute update on the fly
        // w += y[a]*v[a]
        w.axpy(y[a],v[a]);
      }
    }
 
    template<typename T>
    typename std::enable_if<std::is_same<field_type,real_type>::value,T>::type conjugate(const T& t) {
      return t;
    }
 
    template<typename T>
    typename std::enable_if<!std::is_same<field_type,real_type>::value,T>::type conjugate(const T& t) {
      using std::conj;
      return conj(t);
    }
 
    void
    generatePlaneRotation(field_type &dx, field_type &dy, real_type &cs, field_type &sn)
    {
      using std::sqrt;
      using std::abs;
      using std::max;
      using std::min;
      const real_type eps = 1e-15;
      real_type norm_dx = abs(dx);
      real_type norm_dy = abs(dy);
      real_type norm_max = max(norm_dx, norm_dy);
      real_type norm_min = min(norm_dx, norm_dy);
      real_type temp = norm_min/norm_max;
      // we rewrite the code in a vectorizable fashion
      cs = Simd::cond(norm_dy < eps,
        real_type(1.0),
        Simd::cond(norm_dx < eps,
          real_type(0.0),
          Simd::cond(norm_dy > norm_dx,
            real_type(1.0)/sqrt(real_type(1.0) + temp*temp)*temp,
            real_type(1.0)/sqrt(real_type(1.0) + temp*temp)
          )));
      sn = Simd::cond(norm_dy < eps,
        field_type(0.0),
        Simd::cond(norm_dx < eps,
          field_type(1.0),
          Simd::cond(norm_dy > norm_dx,
            field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*dx*conjugate(dy)/norm_dx/norm_dy,
            field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*conjugate(dy/dx)
          )));
    }
 
 
    void
    applyPlaneRotation(field_type &dx, field_type &dy, real_type &cs, field_type &sn)
    {
      field_type temp  =  cs * dx + sn * dy;
      dy = -conjugate(sn) * dx + cs * dy;
      dx = temp;
    }
 
    using IterativeSolver<X,Y>::_op;
    using IterativeSolver<X,Y>::_prec;
    using IterativeSolver<X,Y>::_sp;
    using IterativeSolver<X,Y>::_reduction;
    using IterativeSolver<X,Y>::_maxit;
    using IterativeSolver<X,Y>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
    int _restart;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("restartedgmressolver", defaultIterativeSolverCreator<Dune::RestartedGMResSolver>());
 
  template<class X, class Y=X, class F = Y>
  class RestartedFlexibleGMResSolver : public RestartedGMResSolver<X,Y>
  {
  public:
    using typename RestartedGMResSolver<X,Y>::domain_type;
    using typename RestartedGMResSolver<X,Y>::range_type;
    using typename RestartedGMResSolver<X,Y>::field_type;
    using typename RestartedGMResSolver<X,Y>::real_type;
 
  private:
    using typename RestartedGMResSolver<X,Y>::scalar_real_type;
 
    using fAlloc = typename RestartedGMResSolver<X,Y>::fAlloc;
    using rAlloc = typename RestartedGMResSolver<X,Y>::rAlloc;
 
  public:
    // copy base class constructors
    using RestartedGMResSolver<X,Y>::RestartedGMResSolver;
 
    // don't shadow four-argument version of apply defined in the base class
    using RestartedGMResSolver<X,Y>::apply;
 
    void apply (X& x, Y& b, double reduction, InverseOperatorResult& res) override
    {
      using std::abs;
      const Simd::Scalar<real_type> EPSILON = 1e-80;
      const int m = _restart;
      real_type norm = 0.0;
      int i, j = 1, k;
      std::vector<field_type,fAlloc> s(m+1), sn(m);
      std::vector<real_type,rAlloc> cs(m);
      // helper vector
      Y tmp(b);
      std::vector< std::vector<field_type,fAlloc> > H(m+1,s);
      std::vector<F> v(m+1,b);
      std::vector<X> w(m+1,b);
 
      Iteration iteration(*this,res);
      // setup preconditioner if it does something in pre
 
      // calculate residual and overwrite a copy of the rhs with it
      _prec->pre(x, b);
      v[0] = b;
      _op->applyscaleadd(-1.0, x, v[0]); // b -= Ax
 
      norm = _sp->norm(v[0]); // the residual norm
      if(iteration.step(0, norm)){
        _prec->post(x);
        return;
      }
 
      // start iterations
      res.converged = false;;
      while(j <= _maxit && res.converged != true)
      {
        v[0] *= (1.0 / norm);
        s[0] = norm;
        for(i=1; i<m+1; ++i)
          s[i] = 0.0;
 
        // inner loop
        for(i=0; i < m && j <= _maxit && res.converged != true; i++, j++)
        {
          w[i] = 0.0;
          // compute wi = M^-1*vi (also called zi)
          _prec->apply(w[i], v[i]);
          // compute vi = A*wi
          // use v[i+1] as temporary vector for w
          _op->apply(w[i], v[i+1]);
          // do Arnoldi algorithm
          for(int kk=0; kk<i+1; kk++)
          {
            // notice that _sp->dot(v[k],v[i+1]) = v[k]\adjoint v[i+1]
            // so one has to pay attention to the order
            // in the scalar product for the complex case
            // doing the modified Gram-Schmidt algorithm
            H[kk][i] = _sp->dot(v[kk],v[i+1]);
            // w -= H[k][i] * v[kk]
            v[i+1].axpy(-H[kk][i], v[kk]);
          }
          H[i+1][i] = _sp->norm(v[i+1]);
          if(Simd::allTrue(abs(H[i+1][i]) < EPSILON))
            DUNE_THROW(SolverAbort, "breakdown in fGMRes - |w| (-> "
                                     << w[i] << ") == 0.0 after "
                                     << j << " iterations");
 
          // v[i+1] = w*1/H[i+1][i]
          v[i+1] *= real_type(1.0)/H[i+1][i];
 
          // update QR factorization
          for(k=0; k<i; k++)
            this->applyPlaneRotation(H[k][i],H[k+1][i],cs[k],sn[k]);
 
          // compute new givens rotation
          this->generatePlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]);
 
          // finish updating QR factorization
          this->applyPlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]);
          this->applyPlaneRotation(s[i],s[i+1],cs[i],sn[i]);
 
          // norm of the residual is the last component of vector s
          using std::abs;
          norm = abs(s[i+1]);
          iteration.step(j, norm);
        } // end inner for loop
 
        // calculate update vector
        tmp = 0.0;
        this->update(tmp, i, H, s, w);
        // and update current iterate
        x += tmp;
 
        // restart fGMRes if convergence was not achieved,
        // i.e. linear residual has not reached desired reduction
        // and if still j < _maxit (do not restart on last iteration)
        if( res.converged != true && j < _maxit)
        {
          if (_verbose > 0)
            std::cout << "=== fGMRes::restart" << std::endl;
          // get rhs
          v[0] = b;
          // calculate new defect
          _op->applyscaleadd(-1.0, x,v[0]); // b -= Ax;
          // calculate preconditioned defect
          norm = _sp->norm(v[0]); // update the residual norm
        }
 
      } // end outer while loop
 
      // post-process preconditioner
      _prec->post(x);
    }
 
private:
    using RestartedGMResSolver<X,Y>::_op;
    using RestartedGMResSolver<X,Y>::_prec;
    using RestartedGMResSolver<X,Y>::_sp;
    using RestartedGMResSolver<X,Y>::_reduction;
    using RestartedGMResSolver<X,Y>::_maxit;
    using RestartedGMResSolver<X,Y>::_verbose;
    using RestartedGMResSolver<X,Y>::_restart;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("restartedflexiblegmressolver", defaultIterativeSolverCreator<Dune::RestartedFlexibleGMResSolver>());
 
  template<class X>
  class GeneralizedPCGSolver : public IterativeSolver<X,X>
  {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
  private:
    using typename IterativeSolver<X,X>::scalar_real_type;
 
    using fAlloc = ReboundAllocatorType<X,field_type>;
 
  public:
 
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
 
    GeneralizedPCGSolver (LinearOperator<X,X>& op, Preconditioner<X,X>& prec, scalar_real_type reduction, int maxit, int verbose, int restart = 10) :
      IterativeSolver<X,X>::IterativeSolver(op,prec,reduction,maxit,verbose),
      _restart(restart)
    {}
 
    GeneralizedPCGSolver (LinearOperator<X,X>& op, ScalarProduct<X>& sp, Preconditioner<X,X>& prec, scalar_real_type reduction, int maxit, int verbose, int restart = 10) :
      IterativeSolver<X,X>::IterativeSolver(op,sp,prec,reduction,maxit,verbose),
      _restart(restart)
    {}
 
 
    GeneralizedPCGSolver (std::shared_ptr<LinearOperator<X,X> > op, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) :
      IterativeSolver<X,X>::IterativeSolver(op,prec,configuration),
      _restart(configuration.get<int>("restart"))
    {}
 
    GeneralizedPCGSolver (std::shared_ptr<LinearOperator<X,X> > op, std::shared_ptr<ScalarProduct<X> > sp, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) :
      IterativeSolver<X,X>::IterativeSolver(op,sp,prec,configuration),
      _restart(configuration.get<int>("restart"))
    {}
    GeneralizedPCGSolver (std::shared_ptr<LinearOperator<X,X>> op,
                          std::shared_ptr<ScalarProduct<X>> sp,
                          std::shared_ptr<Preconditioner<X,X>> prec,
                          scalar_real_type reduction, int maxit, int verbose,
                          int restart = 10) :
      IterativeSolver<X,X>::IterativeSolver(op,sp,prec,reduction,maxit,verbose),
      _restart(restart)
    {}
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      Iteration iteration(*this, res);
      _prec->pre(x,b);                 // prepare preconditioner
      _op->applyscaleadd(-1,x,b);      // overwrite b with defect
 
      std::vector<std::shared_ptr<X> > p(_restart);
      std::vector<field_type,fAlloc> pp(_restart);
      X q(x);                  // a temporary vector
      X prec_res(x);           // a temporary vector for preconditioner output
 
      p[0].reset(new X(x));
 
      real_type def = _sp->norm(b);    // compute norm
      if(iteration.step(0, def)){
        _prec->post(x);
        return;
      }
      // some local variables
      field_type rho, lambda;
 
      int i=0;
      int ii=0;
      // determine initial search direction
      *(p[0]) = 0;                              // clear correction
      _prec->apply(*(p[0]),b);                   // apply preconditioner
      rho = _sp->dot(*(p[0]),b);             // orthogonalization
      _op->apply(*(p[0]),q);                 // q=Ap
      pp[0] = _sp->dot(*(p[0]),q);           // scalar product
      lambda = rho/pp[0];         // minimization
      x.axpy(lambda,*(p[0]));               // update solution
      b.axpy(-lambda,q);              // update defect
 
      // convergence test
      def=_sp->norm(b);    // comp defect norm
      ++i;
      if(iteration.step(i, def)){
        _prec->post(x);
        return;
      }
 
      while(i<_maxit) {
        // the loop
        int end=std::min(_restart, _maxit-i+1);
        for (ii=1; ii<end; ++ii )
        {
          //std::cout<<" ii="<<ii<<" i="<<i<<std::endl;
          // compute next conjugate direction
          prec_res = 0;                                  // clear correction
          _prec->apply(prec_res,b);                       // apply preconditioner
 
          p[ii].reset(new X(prec_res));
          _op->apply(prec_res, q);
 
          for(int j=0; j<ii; ++j) {
            rho =_sp->dot(q,*(p[j]))/pp[j];
            p[ii]->axpy(-rho, *(p[j]));
          }
 
          // minimize in given search direction
          _op->apply(*(p[ii]),q);                     // q=Ap
          pp[ii] = _sp->dot(*(p[ii]),q);               // scalar product
          rho = _sp->dot(*(p[ii]),b);                 // orthogonalization
          lambda = rho/pp[ii];             // minimization
          x.axpy(lambda,*(p[ii]));                   // update solution
          b.axpy(-lambda,q);                  // update defect
 
          // convergence test
          def = _sp->norm(b);        // comp defect norm
 
          ++i;
          iteration.step(i, def);
        }
        if(res.converged)
          break;
        if(end==_restart) {
          *(p[0])=*(p[_restart-1]);
          pp[0]=pp[_restart-1];
        }
      }
 
      // postprocess preconditioner
      _prec->post(x);
 
    }
 
  private:
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
    int _restart;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("generalizedpcgsolver", defaultIterativeSolverCreator<Dune::GeneralizedPCGSolver>());
 
  template<class X>
  class RestartedFCGSolver : public IterativeSolver<X,X> {
  public:
    using typename IterativeSolver<X,X>::domain_type;
    using typename IterativeSolver<X,X>::range_type;
    using typename IterativeSolver<X,X>::field_type;
    using typename IterativeSolver<X,X>::real_type;
 
  private:
    using typename IterativeSolver<X,X>::scalar_real_type;
 
  public:
    // don't shadow four-argument version of apply defined in the base class
    using IterativeSolver<X,X>::apply;
    RestartedFCGSolver (LinearOperator<X,X>& op, Preconditioner<X,X>& prec,
                        scalar_real_type reduction, int maxit, int verbose, int mmax = 10) : IterativeSolver<X,X>(op, prec, reduction, maxit, verbose), _mmax(mmax)
    {
    }
 
    RestartedFCGSolver (LinearOperator<X,X>& op, ScalarProduct<X>& sp, Preconditioner<X,X>& prec,
                        scalar_real_type reduction, int maxit, int verbose, int mmax = 10) : IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose), _mmax(mmax)
    {
    }
 
    RestartedFCGSolver (std::shared_ptr<LinearOperator<X,X>> op,
                        std::shared_ptr<ScalarProduct<X>> sp,
                        std::shared_ptr<Preconditioner<X,X>> prec,
                        scalar_real_type reduction, int maxit, int verbose,
                        int mmax = 10)
      : IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose), _mmax(mmax)
    {}
 
    RestartedFCGSolver (std::shared_ptr<LinearOperator<X,X>> op,
                        std::shared_ptr<Preconditioner<X,X>> prec,
                        const ParameterTree& config)
      : IterativeSolver<X,X>(op, prec, config), _mmax(config.get("mmax", 10))
    {}
 
    RestartedFCGSolver (std::shared_ptr<LinearOperator<X,X>> op,
                        std::shared_ptr<ScalarProduct<X>> sp,
                        std::shared_ptr<Preconditioner<X,X>> prec,
                        const ParameterTree& config)
      : IterativeSolver<X,X>(op, sp, prec, config), _mmax(config.get("mmax", 10))
    {}
 
    virtual void apply (X& x, X& b, InverseOperatorResult& res)
    {
      using rAlloc = ReboundAllocatorType<X,field_type>;
      res.clear();
      Iteration iteration(*this,res);
      _prec->pre(x,b);             // prepare preconditioner
      _op->applyscaleadd(-1,x,b); // overwrite b with defect
 
      //arrays for interim values:
      std::vector<X> d(_mmax+1, x);                      // array for directions
      std::vector<X> Ad(_mmax+1, x);                    // array for Ad[i]
      std::vector<field_type,rAlloc> ddotAd(_mmax+1,0); // array for <d[i],Ad[i]>
      X w(x);
 
      real_type def = _sp->norm(b); // compute norm
      if(iteration.step(0, def)){
        _prec->post(x);
        return;
      }
 
      // some local variables
      field_type alpha;
 
      // the loop
      int i=1;
      int i_bounded=0;
      while(i<=_maxit && !res.converged) {
        for (; i_bounded <= _mmax && i<= _maxit; i_bounded++) {
          d[i_bounded] = 0;                   // reset search direction
          _prec->apply(d[i_bounded], b);     // apply preconditioner
          w = d[i_bounded];                 // copy of current d[i]
          // orthogonalization with previous directions
          orthogonalizations(i_bounded,Ad,w,ddotAd,d);
 
          //saving interim values for future calculating
          _op->apply(d[i_bounded], Ad[i_bounded]);                    // save Ad[i]
          ddotAd[i_bounded]=_sp->dot(d[i_bounded],Ad[i_bounded]);    // save <d[i],Ad[i]>
          alpha = _sp->dot(d[i_bounded], b)/ddotAd[i_bounded];      // <d[i],b>/<d[i],Ad[i]>
 
          //update solution and defect
          x.axpy(alpha, d[i_bounded]);
          b.axpy(-alpha, Ad[i_bounded]);
 
          // convergence test
          def = _sp->norm(b); // comp defect norm
 
          iteration.step(i, def);
          i++;
        }
        //restart: exchange first and last stored values
        cycle(Ad,d,ddotAd,i_bounded);
      }
 
      //correct i which is wrong if convergence was not achieved.
      i=std::min(_maxit,i);
 
      _prec->post(x);                  // postprocess preconditioner
    }
 
  private:
    //This function is called every iteration to orthogonalize against the last search directions
    virtual void orthogonalizations(const int& i_bounded,const std::vector<X>& Ad, const X& w, const std::vector<field_type,ReboundAllocatorType<X,field_type>>& ddotAd,std::vector<X>& d) {
      // The RestartedFCGSolver uses only values with lower array index;
      for (int k = 0; k < i_bounded; k++) {
        d[i_bounded].axpy(-_sp->dot(Ad[k], w) / ddotAd[k], d[k]); // d[i] -= <<Ad[k],w>/<d[k],Ad[k]>>d[k]
      }
    }
 
    // This function is called every mmax iterations to handle limited array sizes.
    virtual void cycle(std::vector<X>& Ad,std::vector<X>& d,std::vector<field_type,ReboundAllocatorType<X,field_type> >& ddotAd,int& i_bounded) {
      // Reset loop index and exchange the first and last arrays
      i_bounded = 1;
      std::swap(Ad[0], Ad[_mmax]);
      std::swap(d[0], d[_mmax]);
      std::swap(ddotAd[0], ddotAd[_mmax]);
    }
 
  protected:
    int _mmax;
    using IterativeSolver<X,X>::_op;
    using IterativeSolver<X,X>::_prec;
    using IterativeSolver<X,X>::_sp;
    using IterativeSolver<X,X>::_reduction;
    using IterativeSolver<X,X>::_maxit;
    using IterativeSolver<X,X>::_verbose;
    using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("restartedfcgsolver", defaultIterativeSolverCreator<Dune::RestartedFCGSolver>());
 
  template<class X>
  class CompleteFCGSolver : public RestartedFCGSolver<X> {
  public:
    using typename RestartedFCGSolver<X>::domain_type;
    using typename RestartedFCGSolver<X>::range_type;
    using typename RestartedFCGSolver<X>::field_type;
    using typename RestartedFCGSolver<X>::real_type;
 
    // copy base class constructors
    using RestartedFCGSolver<X>::RestartedFCGSolver;
 
    // don't shadow four-argument version of apply defined in the base class
    using RestartedFCGSolver<X>::apply;
 
    // just a minor part of the RestartedFCGSolver apply method will be modified
    virtual void apply (X& x, X& b, InverseOperatorResult& res) override {
      // reset limiter of orthogonalization loop
      _k_limit = 0;
      this->RestartedFCGSolver<X>::apply(x,b,res);
    };
 
  private:
    // This function is called every iteration to orthogonalize against the last search directions.
    virtual void orthogonalizations(const int& i_bounded,const std::vector<X>& Ad, const X& w, const std::vector<field_type,ReboundAllocatorType<X,field_type>>& ddotAd,std::vector<X>& d) override {
      // This FCGSolver uses values with higher array indexes too, if existent.
      for (int k = 0; k < _k_limit; k++) {
        if(i_bounded!=k)
          d[i_bounded].axpy(-_sp->dot(Ad[k], w) / ddotAd[k], d[k]); // d[i] -= <<Ad[k],w>/<d[k],Ad[k]>>d[k]
      }
      // The loop limit increase, if array is not completely filled.
      if(_k_limit<=i_bounded)
        _k_limit++;
 
    };
 
    // This function is called every mmax iterations to handle limited array sizes.
    virtual void cycle(std::vector<X>& Ad,std::vector<X>& d,std::vector<field_type,ReboundAllocatorType<X,field_type> >& ddotAd,int& i_bounded) override {
      // Only the loop index i_bounded return to 0, if it reached mmax.
      i_bounded = 0;
      // Now all arrays are filled and the loop in void orthogonalizations can use the whole arrays.
      _k_limit = Ad.size();
    };
 
    int _k_limit = 0;
 
  protected:
    using RestartedFCGSolver<X>::_mmax;
    using RestartedFCGSolver<X>::_op;
    using RestartedFCGSolver<X>::_prec;
    using RestartedFCGSolver<X>::_sp;
    using RestartedFCGSolver<X>::_reduction;
    using RestartedFCGSolver<X>::_maxit;
    using RestartedFCGSolver<X>::_verbose;
  };
  DUNE_REGISTER_ITERATIVE_SOLVER("completefcgsolver", defaultIterativeSolverCreator<Dune::CompleteFCGSolver>());
} // end namespace
 
#endif