solvers.hpp

Go to the documentation of this file.

// Copyright (c) 2010-2023, Lawrence Livermore National Security, LLC. Produced
// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
// LICENSE and NOTICE for details. LLNL-CODE-806117.
//
// This file is part of the MFEM library. For more information and source code
// availability visit https://mfem.org.
//
// MFEM is free software; you can redistribute it and/or modify it under the
// terms of the BSD-3 license. We welcome feedback and contributions, see file
// CONTRIBUTING.md for details.

#ifndef MFEM_SOLVERS
#define MFEM_SOLVERS

#include "../config/config.hpp"
#include "densemat.hpp"
#include "handle.hpp"
#include <memory>

#ifdef MFEM_USE_MPI
#include <mpi.h>
#endif

#ifdef MFEM_USE_SUITESPARSE
#include "sparsemat.hpp"
#include <umfpack.h>
#include <klu.h>
#endif

namespace mfem
{

class BilinearForm;

/// Abstract base class for an iterative solver monitor
class IterativeSolverMonitor
{
protected:
   /// The last IterativeSolver to which this monitor was attached.
   const class IterativeSolver *iter_solver;

public:
   IterativeSolverMonitor() : iter_solver(nullptr) {}

   virtual ~IterativeSolverMonitor() {}

   /// Monitor the residual vector r
   virtual void MonitorResidual(int it, double norm, const Vector &r,
                                bool final)
   {
   }

   /// Monitor the solution vector x
   virtual void MonitorSolution(int it, double norm, const Vector &x,
                                bool final)
   {
   }

   /** @brief This method is invoked by IterativeSolver::SetMonitor, informing
       the monitor which IterativeSolver is using it. */
   void SetIterativeSolver(const IterativeSolver &solver)
   { iter_solver = &solver; }
};

/// Abstract base class for iterative solver
class IterativeSolver : public Solver
{
public:
   /** @brief Settings for the output behavior of the IterativeSolver.

       By default, all output is suppressed. The construction of the desired
       print level can be achieved through a builder pattern, for example

           PrintLevel().Errors().Warnings()

       constructs the print level with only errors and warnings enabled.
     */
   struct PrintLevel
   {
      /** @brief If a fatal problem has been detected the failure will be
          reported to @ref mfem::err. */
      bool errors = false;
      /** @brief If a non-fatal problem has been detected some context-specific
          information will be reported to @ref mfem::out */
      bool warnings = false;
      /** @brief Detailed information about each iteration will be reported to
          @ref mfem::out */
      bool iterations = false;
      /** @brief A summary of the solver process will be reported after the last
          iteration to @ref mfem::out */
      bool summary = false;
      /** @brief Information about the first and last iteration will be printed
          to @ref mfem::out */
      bool first_and_last = false;

      /// Initializes the print level to suppress
      PrintLevel() = default;

      /** @name Builder
         These methods are utilized to construct PrintLevel objects through a
         builder approach by chaining the function calls in this group. */
      ///@{
      PrintLevel &None() { *this = PrintLevel(); return *this; }
      PrintLevel &Warnings() { warnings=true; return *this; }
      PrintLevel &Errors() { errors=true; return *this; }
      PrintLevel &Iterations() { iterations=true; return *this; }
      PrintLevel &FirstAndLast() { first_and_last=true; return *this; }
      PrintLevel &Summary() { summary=true; return *this; }
      PrintLevel &All() { return Warnings().Errors().Iterations().FirstAndLast().Summary(); }
      ///@}
   };

#ifdef MFEM_USE_MPI
private:
   int dot_prod_type; // 0 - local, 1 - global over 'comm'
   MPI_Comm comm = MPI_COMM_NULL;
#endif

protected:
   const Operator *oper;
   Solver *prec;
   IterativeSolverMonitor *monitor = nullptr;

   /// @name Reporting (protected attributes and member functions)
   ///@{

   /** @brief (DEPRECATED) Legacy print level definition, which is left for
       compatibility with custom iterative solvers.
       @deprecated #print_options should be used instead. */
   int print_level = -1;

   /** @brief Output behavior for the iterative solver.

       This primarily controls the output behavior of the iterative solvers
       provided by this library. This member must be synchronized with
       #print_level to ensure compatibility with custom iterative solvers. */
   PrintLevel print_options;

   /// Convert a legacy print level integer to a PrintLevel object
   PrintLevel FromLegacyPrintLevel(int);

   /// @brief Use some heuristics to guess a legacy print level corresponding to
   /// the given PrintLevel.
   static int GuessLegacyPrintLevel(PrintLevel);
   ///@}

   /// @name Convergence (protected attributes)
   ///@{

   /// Limit for the number of iterations the solver is allowed to do
   int max_iter;

   /// Relative tolerance.
   double rel_tol;

   /// Absolute tolerance.
   double abs_tol;

   ///@}

   /// @name Solver statistics (protected attributes)
   /// Every IterativeSolver is expected to define these in its Mult() call.
   ///@{

   mutable int final_iter = -1;
   mutable bool converged = false;
   mutable double initial_norm = -1.0, final_norm = -1.0;

   ///@}

   double Dot(const Vector &x, const Vector &y) const;
   double Norm(const Vector &x) const { return sqrt(Dot(x, x)); }
   void Monitor(int it, double norm, const Vector& r, const Vector& x,
                bool final=false) const;

public:
   IterativeSolver();

#ifdef MFEM_USE_MPI
   IterativeSolver(MPI_Comm comm_);
#endif

   /** @name Convergence
       @brief Termination criteria for the iterative solvers.

       @details While the convergence criterion is solver specific, most of the
       provided iterative solvers use one of the following criteria

       \f$ ||r||_X \leq tol_{rel}||r_0||_X \f$,

       \f$ ||r||_X \leq tol_{abs} \f$,

       \f$ ||r||_X \leq \max\{ tol_{abs}, tol_{rel} ||r_0||_X \} \f$,

       where X denotes the space in which the norm is measured. The choice of
       X depends on the specific iterative solver.
      */
   ///@{
   void SetRelTol(double rtol) { rel_tol = rtol; }
   void SetAbsTol(double atol) { abs_tol = atol; }
   void SetMaxIter(int max_it) { max_iter = max_it; }
   ///@}

   /** @name Reporting
       These options control the internal reporting behavior into ::mfem::out
       and ::mfem::err of the iterative solvers.
    */
   ///@{

   /// @brief Legacy method to set the level of verbosity of the solver output.
   /** This is the old way to control what information will be printed to
       ::mfem::out and ::mfem::err. The behavior for the print level for all
       iterative solvers is:

       - -1: Suppress all outputs.
       -  0: Print information about all detected issues (e.g. no convergence).
       -  1: Same as level 0, but with detailed information about each
             iteration.
       -  2: Print detected issues and a summary when the solver terminates.
       -  3: Same as 2, but print also the first and last iterations.
       - >3: Custom print options which are dependent on the specific solver.

       In parallel, only rank 0 produces output.

       @note It is recommended to use @ref SetPrintLevel(PrintLevel) instead.

       @note Some derived classes, like KINSolver, redefine this method and use
       their own set of print level constants. */
   virtual void SetPrintLevel(int print_lvl);

   /// @brief Set the level of verbosity of the solver output.
   /** In parallel, only rank 0 produces outputs. Errors are output to
       ::mfem::err and all other information to ::mfem::out.

       @note Not all subclasses of IterativeSolver support all possible options.

       @note Some derived classes, like KINSolver, disable this method in favor
       of SetPrintLevel(int).

       @sa PrintLevel for possible options.
   */
   virtual void SetPrintLevel(PrintLevel);
   ///@}

   /// @name Solver statistics.
   /// These are valid after the call to Mult().
   ///@{

   /// Returns the number of iterations taken during the last call to Mult()
   int GetNumIterations() const { return final_iter; }
   /// Returns true if the last call to Mult() converged successfully.
   bool GetConverged() const { return converged; }
   /// @brief Returns the initial residual norm from the last call to Mult().
   ///
   /// This function returns the norm of the residual (or preconditioned
   /// residual, depending on the solver), computed before the start of the
   /// iteration.
   double GetInitialNorm() const { return initial_norm; }
   /// @brief Returns the final residual norm after termination of the solver
   /// during the last call to Mult().
   ///
   /// This function returns the norm of the residual (or preconditioned
   /// residual, depending on the solver), corresponding to the returned
   /// solution.
   double GetFinalNorm() const { return final_norm; }
   /// @brief Returns the final residual norm after termination of the solver
   /// during the last call to Mult(), divided by the initial residual norm.
   /// Returns -1 if one of these norms is left undefined by the solver.
   ///
   /// @sa GetFinalNorm(), GetInitialNorm()
   double GetFinalRelNorm() const
   {
      if (final_norm < 0.0 || initial_norm < 0.0) { return -1.0; }
      return final_norm / initial_norm;
   }

   ///@}

   /// This should be called before SetOperator
   virtual void SetPreconditioner(Solver &pr);

   /// Also calls SetOperator for the preconditioner if there is one
   virtual void SetOperator(const Operator &op) override;

   /// Set the iterative solver monitor
   void SetMonitor(IterativeSolverMonitor &m)
   { monitor = &m; m.SetIterativeSolver(*this); }

#ifdef MFEM_USE_MPI
   /** @brief Return the associated MPI communicator, or MPI_COMM_NULL if no
       communicator is set. */
   MPI_Comm GetComm() const
   { return dot_prod_type == 0 ? MPI_COMM_NULL : comm; }
#endif
};


/// Jacobi smoothing for a given bilinear form (no matrix necessary).
/** Useful with tensorized, partially assembled operators. Can also be defined
    by given diagonal vector. This is basic Jacobi iteration; for tolerances,
    iteration control, etc. wrap with SLISolver. */
class OperatorJacobiSmoother : public Solver
{
public:
   /** @brief Default constructor: the diagonal will be computed by subsequent
       calls to SetOperator() using the Operator method AssembleDiagonal. */
   /** In this case the array of essential tdofs will be empty. */
   OperatorJacobiSmoother(const double damping=1.0);

   /** Setup a Jacobi smoother with the diagonal of @a a obtained by calling
       a.AssembleDiagonal(). It is assumed that the underlying operator acts as
       the identity on entries in ess_tdof_list, corresponding to (assembled)
       DIAG_ONE policy or ConstrainedOperator in the matrix-free setting.

       @note For objects created with this constructor, calling SetOperator()
       will only set the internal Operator pointer to the given new Operator
       without any other changes to the object. This is done to preserve the
       original behavior of this class. */
   OperatorJacobiSmoother(const BilinearForm &a,
                          const Array<int> &ess_tdof_list,
                          const double damping=1.0);

   /** Application is by the *inverse* of the given vector. It is assumed that
       the underlying operator acts as the identity on entries in ess_tdof_list,
       corresponding to (assembled) DIAG_ONE policy or ConstrainedOperator in
       the matrix-free setting.

       @note For objects created with this constructor, calling SetOperator()
       will only set the internal Operator pointer to the given new Operator
       without any other changes to the object. This is done to preserve the
       original behavior of this class. */
   OperatorJacobiSmoother(const Vector &d,
                          const Array<int> &ess_tdof_list,
                          const double damping=1.0);

   ~OperatorJacobiSmoother() {}

   /// Replace diagonal entries with their absolute values.
   void SetPositiveDiagonal(bool pos_diag = true) { use_abs_diag = pos_diag; }

   void Mult(const Vector &x, Vector &y) const;
   void MultTranspose(const Vector &x, Vector &y) const { Mult(x, y); }

   /** @brief Recompute the diagonal using the method AssembleDiagonal of the
       given new Operator, @a op. */
   /** Note that (Par)BilinearForm operators are treated similar to the way they
       are treated in the constructor that takes a BilinearForm parameter.
       Specifically, this means that the OperatorJacobiSmoother will work with
       true-dof vectors even though the size of the BilinearForm may be
       different.

       When the new Operator, @a op, is not a (Par)BilinearForm, any previously
       set array of essential true-dofs will be thrown away because in this case
       any essential b.c. will be handled by the AssembleDiagonal method. */
   void SetOperator(const Operator &op);

private:
   Vector dinv;
   const double damping;
   const Array<int> *ess_tdof_list; // not owned; may be NULL
   mutable Vector residual;
   /// Uses absolute values of the diagonal entries.
   bool use_abs_diag = false;

   const Operator *oper; // not owned

   // To preserve the original behavior, some constructors set this flag to
   // false to disallow updating the OperatorJacobiSmoother with SetOperator.
   const bool allow_updates;

public:
   void Setup(const Vector &diag);
};

/// Chebyshev accelerated smoothing with given vector, no matrix necessary
/** Potentially useful with tensorized operators, for example. This is just a
    very basic Chebyshev iteration, if you want tolerances, iteration control,
    etc. wrap this with SLISolver. */
class OperatorChebyshevSmoother : public Solver
{
public:
   /** Application is by *inverse* of the given vector. It is assumed the
       underlying operator acts as the identity on entries in ess_tdof_list,
       corresponding to (assembled) DIAG_ONE policy or ConstrainedOperator in
       the matrix-free setting. The estimated largest eigenvalue of the
       diagonally preconditoned operator must be provided via
       max_eig_estimate. */
   OperatorChebyshevSmoother(const Operator &oper_, const Vector &d,
                             const Array<int>& ess_tdof_list,
                             int order, double max_eig_estimate);

   /// Deprecated: see pass-by-reference version above
   MFEM_DEPRECATED
   OperatorChebyshevSmoother(const Operator* oper_, const Vector &d,
                             const Array<int>& ess_tdof_list,
                             int order, double max_eig_estimate);

   /** Application is by *inverse* of the given vector. It is assumed the
       underlying operator acts as the identity on entries in ess_tdof_list,
       corresponding to (assembled) DIAG_ONE policy or ConstrainedOperator in
       the matrix-free setting. The largest eigenvalue of the diagonally
       preconditoned operator is estimated internally via a power method. The
       accuracy of the estimated eigenvalue may be controlled via
       power_iterations and power_tolerance. */
#ifdef MFEM_USE_MPI
   OperatorChebyshevSmoother(const Operator &oper_, const Vector &d,
                             const Array<int>& ess_tdof_list,
                             int order, MPI_Comm comm = MPI_COMM_NULL,
                             int power_iterations = 10,
                             double power_tolerance = 1e-8);

   /// Deprecated: see pass-by-reference version above
   MFEM_DEPRECATED
   OperatorChebyshevSmoother(const Operator* oper_, const Vector &d,
                             const Array<int>& ess_tdof_list,
                             int order, MPI_Comm comm = MPI_COMM_NULL,
                             int power_iterations = 10,
                             double power_tolerance = 1e-8);
#else
   OperatorChebyshevSmoother(const Operator &oper_, const Vector &d,
                             const Array<int>& ess_tdof_list,
                             int order, int power_iterations = 10,
                             double power_tolerance = 1e-8);

   /// Deprecated: see pass-by-reference version above
   MFEM_DEPRECATED
   OperatorChebyshevSmoother(const Operator* oper_, const Vector &d,
                             const Array<int>& ess_tdof_list,
                             int order, int power_iterations = 10,
                             double power_tolerance = 1e-8);
#endif

   ~OperatorChebyshevSmoother() {}

   void Mult(const Vector&x, Vector &y) const;

   void MultTranspose(const Vector &x, Vector &y) const { Mult(x, y); }

   void SetOperator(const Operator &op_)
   {
      oper = &op_;
   }

   void Setup();

private:
   const int order;
   double max_eig_estimate;
   const int N;
   Vector dinv;
   const Vector &diag;
   Array<double> coeffs;
   const Array<int>& ess_tdof_list;
   mutable Vector residual;
   mutable Vector helperVector;
   const Operator* oper;
};


/// Stationary linear iteration: x <- x + B (b - A x)
class SLISolver : public IterativeSolver
{
protected:
   mutable Vector r, z;

   void UpdateVectors();

public:
   SLISolver() { }

#ifdef MFEM_USE_MPI
   SLISolver(MPI_Comm comm_) : IterativeSolver(comm_) { }
#endif

   virtual void SetOperator(const Operator &op)
   { IterativeSolver::SetOperator(op); UpdateVectors(); }

   virtual void Mult(const Vector &b, Vector &x) const;
};

/// Stationary linear iteration. (tolerances are squared)
void SLI(const Operator &A, const Vector &b, Vector &x,
         int print_iter = 0, int max_num_iter = 1000,
         double RTOLERANCE = 1e-12, double ATOLERANCE = 1e-24);

/// Preconditioned stationary linear iteration. (tolerances are squared)
void SLI(const Operator &A, Solver &B, const Vector &b, Vector &x,
         int print_iter = 0, int max_num_iter = 1000,
         double RTOLERANCE = 1e-12, double ATOLERANCE = 1e-24);


/// Conjugate gradient method
class CGSolver : public IterativeSolver
{
protected:
   mutable Vector r, d, z;

   void UpdateVectors();

public:
   CGSolver() { }

#ifdef MFEM_USE_MPI
   CGSolver(MPI_Comm comm_) : IterativeSolver(comm_) { }
#endif

   virtual void SetOperator(const Operator &op)
   { IterativeSolver::SetOperator(op); UpdateVectors(); }

   virtual void Mult(const Vector &b, Vector &x) const;
};

/// Conjugate gradient method. (tolerances are squared)
void CG(const Operator &A, const Vector &b, Vector &x,
        int print_iter = 0, int max_num_iter = 1000,
        double RTOLERANCE = 1e-12, double ATOLERANCE = 1e-24);

/// Preconditioned conjugate gradient method. (tolerances are squared)
void PCG(const Operator &A, Solver &B, const Vector &b, Vector &x,
         int print_iter = 0, int max_num_iter = 1000,
         double RTOLERANCE = 1e-12, double ATOLERANCE = 1e-24);


/// GMRES method
class GMRESSolver : public IterativeSolver
{
protected:
   int m; // see SetKDim()

public:
   GMRESSolver() { m = 50; }

#ifdef MFEM_USE_MPI
   GMRESSolver(MPI_Comm comm_) : IterativeSolver(comm_) { m = 50; }
#endif

   /// Set the number of iteration to perform between restarts, default is 50.
   void SetKDim(int dim) { m = dim; }

   virtual void Mult(const Vector &b, Vector &x) const;
};

/// FGMRES method
class FGMRESSolver : public IterativeSolver
{
protected:
   int m;

public:
   FGMRESSolver() { m = 50; }

#ifdef MFEM_USE_MPI
   FGMRESSolver(MPI_Comm comm_) : IterativeSolver(comm_) { m = 50; }
#endif

   void SetKDim(int dim) { m = dim; }

   virtual void Mult(const Vector &b, Vector &x) const;
};

/// GMRES method. (tolerances are squared)
int GMRES(const Operator &A, Vector &x, const Vector &b, Solver &M,
          int &max_iter, int m, double &tol, double atol, int printit);

/// GMRES method. (tolerances are squared)
void GMRES(const Operator &A, Solver &B, const Vector &b, Vector &x,
           int print_iter = 0, int max_num_iter = 1000, int m = 50,
           double rtol = 1e-12, double atol = 1e-24);


/// BiCGSTAB method
class BiCGSTABSolver : public IterativeSolver
{
protected:
   mutable Vector p, phat, s, shat, t, v, r, rtilde;

   void UpdateVectors();

public:
   BiCGSTABSolver() { }

#ifdef MFEM_USE_MPI
   BiCGSTABSolver(MPI_Comm comm_) : IterativeSolver(comm_) { }
#endif

   virtual void SetOperator(const Operator &op)
   { IterativeSolver::SetOperator(op); UpdateVectors(); }

   virtual void Mult(const Vector &b, Vector &x) const;
};

/// BiCGSTAB method. (tolerances are squared)
int BiCGSTAB(const Operator &A, Vector &x, const Vector &b, Solver &M,
             int &max_iter, double &tol, double atol, int printit);

/// BiCGSTAB method. (tolerances are squared)
void BiCGSTAB(const Operator &A, Solver &B, const Vector &b, Vector &x,
              int print_iter = 0, int max_num_iter = 1000,
              double rtol = 1e-12, double atol = 1e-24);


/// MINRES method
class MINRESSolver : public IterativeSolver
{
protected:
   mutable Vector v0, v1, w0, w1, q;
   mutable Vector u1; // used in the preconditioned version

public:
   MINRESSolver() { }

#ifdef MFEM_USE_MPI
   MINRESSolver(MPI_Comm comm_) : IterativeSolver(comm_) { }
#endif

   virtual void SetPreconditioner(Solver &pr)
   {
      IterativeSolver::SetPreconditioner(pr);
      if (oper) { u1.SetSize(width); }
   }

   virtual void SetOperator(const Operator &op);

   virtual void Mult(const Vector &b, Vector &x) const;
};

/// MINRES method without preconditioner. (tolerances are squared)
void MINRES(const Operator &A, const Vector &b, Vector &x, int print_it = 0,
            int max_it = 1000, double rtol = 1e-12, double atol = 1e-24);

/// MINRES method with preconditioner. (tolerances are squared)
void MINRES(const Operator &A, Solver &B, const Vector &b, Vector &x,
            int print_it = 0, int max_it = 1000,
            double rtol = 1e-12, double atol = 1e-24);


/// Newton's method for solving F(x)=b for a given operator F.
/** The method GetGradient() must be implemented for the operator F.
    The preconditioner is used (in non-iterative mode) to evaluate
    the action of the inverse gradient of the operator. */
class NewtonSolver : public IterativeSolver
{
protected:
   mutable Vector r, c;
   mutable Operator *grad;

   // Adaptive linear solver rtol variables

   // Method to determine rtol, 0 means the adaptive algorithm is deactivated.
   int lin_rtol_type = 0;
   // rtol to use in first iteration
   double lin_rtol0;
   // Maximum rtol
   double lin_rtol_max;
   // Function norm ||F(x)|| of the previous iterate
   mutable double fnorm_last = 0.0;
   // Linear residual norm of the previous iterate
   mutable double lnorm_last = 0.0;
   // Forcing term (linear residual rtol) from the previous iterate
   mutable double eta_last = 0.0;
   // Eisenstat-Walker factor gamma
   double gamma;
   // Eisenstat-Walker factor alpha
   double alpha;

   /** @brief Method for the adaptive linear solver rtol invoked before the
       linear solve. */
   void AdaptiveLinRtolPreSolve(const Vector &x,
                                const int it,
                                const double fnorm) const;

   /** @brief Method for the adaptive linear solver rtol invoked after the
       linear solve. */
   void AdaptiveLinRtolPostSolve(const Vector &x,
                                 const Vector &b,
                                 const int it,
                                 const double fnorm) const;

public:
   NewtonSolver() { }

#ifdef MFEM_USE_MPI
   NewtonSolver(MPI_Comm comm_) : IterativeSolver(comm_) { }
#endif
   virtual void SetOperator(const Operator &op);

   /// Set the linear solver for inverting the Jacobian.
   /** This method is equivalent to calling SetPreconditioner(). */
   virtual void SetSolver(Solver &solver) { prec = &solver; }

   /// Solve the nonlinear system with right-hand side @a b.
   /** If `b.Size() != Height()`, then @a b is assumed to be zero. */
   virtual void Mult(const Vector &b, Vector &x) const;

   /** @brief This method can be overloaded in derived classes to implement line
       search algorithms. */
   /** The base class implementation (NewtonSolver) simply returns 1. A return
       value of 0 indicates a failure, interrupting the Newton iteration. */
   virtual double ComputeScalingFactor(const Vector &x, const Vector &b) const
   { return 1.0; }

   /** @brief This method can be overloaded in derived classes to perform
       computations that need knowledge of the newest Newton state. */
   virtual void ProcessNewState(const Vector &x) const { }

   /// Enable adaptive linear solver relative tolerance algorithm.
   /** Compute a relative tolerance for the Krylov method after each nonlinear
    iteration, based on the algorithm presented in [1].

    The maximum linear solver relative tolerance @a rtol_max should be < 1. For
    @a type 1 the parameters @a alpha and @a gamma are ignored. For @a type 2
    @a alpha has to be between 0 and 1 and @a gamma between 1 and 2.

    [1] Eisenstat, Stanley C., and Homer F. Walker. "Choosing the forcing terms
    in an inexact Newton method."
    */
   void SetAdaptiveLinRtol(const int type = 2,
                           const double rtol0 = 0.5,
                           const double rtol_max = 0.9,
                           const double alpha = 0.5 * (1.0 + sqrt(5.0)),
                           const double gamma = 1.0);
};

/** L-BFGS method for solving F(x)=b for a given operator F, by minimizing
    the norm of F(x) - b. Requires only the action of the operator F. */
class LBFGSSolver : public NewtonSolver
{
protected:
   int m = 10;
   mutable Array<Vector *> skArray, ykArray;

   void DeleteStorageVectors()
   {
      for (int i = 0; i < skArray.Size(); i++)
      {
         delete skArray[i];
         delete ykArray[i];
      }
   }

   void InitializeStorageVectors()
   {
      DeleteStorageVectors();
      skArray.SetSize(m);
      ykArray.SetSize(m);
      for (int i = 0; i < m; i++)
      {
         skArray[i] = new Vector(width);
         ykArray[i] = new Vector(width);
         skArray[i]->UseDevice(true);
         ykArray[i]->UseDevice(true);
      }
   }

public:
   LBFGSSolver() : NewtonSolver() { }

#ifdef MFEM_USE_MPI
   LBFGSSolver(MPI_Comm comm_) : NewtonSolver(comm_) { }
#endif

   virtual void SetOperator(const Operator &op)
   {
      NewtonSolver::SetOperator(op);
      InitializeStorageVectors();
   }

   void SetHistorySize(int dim)
   {
      m = dim;
      InitializeStorageVectors();
   }

   /// Solve the nonlinear system with right-hand side @a b.
   /** If `b.Size() != Height()`, then @a b is assumed to be zero. */
   virtual void Mult(const Vector &b, Vector &x) const;

   virtual void SetPreconditioner(Solver &pr)
   { MFEM_WARNING("L-BFGS won't use the given preconditioner."); }
   virtual void SetSolver(Solver &solver)
   { MFEM_WARNING("L-BFGS won't use the given solver."); }

   virtual ~LBFGSSolver() { DeleteStorageVectors(); }
};


/** Adaptive restarted GMRES.
    m_max and m_min(=1) are the maximal and minimal restart parameters.
    m_step(=1) is the step to use for going from m_max and m_min.
    cf(=0.4) is a desired convergence factor. */
int aGMRES(const Operator &A, Vector &x, const Vector &b,
           const Operator &M, int &max_iter,
           int m_max, int m_min, int m_step, double cf,
           double &tol, double &atol, int printit);


/** Defines operators and constraints for the following optimization problem:
 *
 *    Find x that minimizes the objective function F(x), subject to
 *    C(x) = c_e,
 *    d_lo <= D(x) <= d_hi,
 *    x_lo <= x <= x_hi.
 *
 *  The operators F, C, D must take input of the same size (same width).
 *  Gradients of F, C, D might be needed, depending on the OptimizationSolver.
 *  When used with Hiop, gradients of C and D must be DenseMatrices.
 *  F always returns a scalar value, see CalcObjective(), CalcObjectiveGrad().
 *  C and D can have arbitrary heights.
 *  C and D can be NULL, meaning that their constraints are not used.
 *
 *  When used in parallel, all Vectors are assumed to be true dof vectors, and
 *  the operators are expected to be defined for tdof vectors. */
class OptimizationProblem
{
protected:
   /// Not owned, some can remain unused (NULL).
   const Operator *C, *D;
   const Vector *c_e, *d_lo, *d_hi, *x_lo, *x_hi;

public:
   const int input_size;

   /// In parallel, insize is the number of the local true dofs.
   OptimizationProblem(int insize, const Operator *C_, const Operator *D_);

   /// Objective F(x). In parallel, the result should be reduced over tasks.
   virtual double CalcObjective(const Vector &x) const = 0;
   /// The result grad is expected to enter with the correct size.
   virtual void CalcObjectiveGrad(const Vector &x, Vector &grad) const
   { MFEM_ABORT("The objective gradient is not implemented."); }

   void SetEqualityConstraint(const Vector &c);
   void SetInequalityConstraint(const Vector &dl, const Vector &dh);
   void SetSolutionBounds(const Vector &xl, const Vector &xh);

   const Operator *GetC() const { return C; }
   const Operator *GetD() const { return D; }
   const Vector *GetEqualityVec() const { return c_e; }
   const Vector *GetInequalityVec_Lo() const { return d_lo; }
   const Vector *GetInequalityVec_Hi() const { return d_hi; }
   const Vector *GetBoundsVec_Lo() const { return x_lo; }
   const Vector *GetBoundsVec_Hi() const { return x_hi; }

   int GetNumConstraints() const;
};

/// Abstract solver for OptimizationProblems.
class OptimizationSolver : public IterativeSolver
{
protected:
   const OptimizationProblem *problem;

public:
   OptimizationSolver(): IterativeSolver(), problem(NULL) { }
#ifdef MFEM_USE_MPI
   OptimizationSolver(MPI_Comm comm_): IterativeSolver(comm_), problem(NULL) { }
#endif
   virtual ~OptimizationSolver() { }

   /** This function is virtual as solvers might need to perform some initial
    *  actions (e.g. validation) with the OptimizationProblem. */
   virtual void SetOptimizationProblem(const OptimizationProblem &prob)
   { problem = &prob; }

   virtual void Mult(const Vector &xt, Vector &x) const = 0;

   virtual void SetPreconditioner(Solver &pr)
   { MFEM_ABORT("Not meaningful for this solver."); }
   virtual void SetOperator(const Operator &op)
   { MFEM_ABORT("Not meaningful for this solver."); }
};

/** SLBQP optimizer:
 *  (S)ingle (L)inearly Constrained with (B)ounds (Q)uadratic (P)rogram
 *
 *    Minimize || x-x_t ||, subject to
 *    sum w_i x_i = a,
 *    x_lo <= x <= x_hi.
 */
class SLBQPOptimizer : public OptimizationSolver
{
protected:
   Vector lo, hi, w;
   double a;

   /// Solve QP at fixed lambda
   inline double solve(double l, const Vector &xt, Vector &x, int &nclip) const
   {
      add(xt, l, w, x);
      if (problem == NULL) { x.median(lo,hi); }
      else
      {
         x.median(*problem->GetBoundsVec_Lo(),
                  *problem->GetBoundsVec_Hi());
      }
      nclip++;
      if (problem == NULL) { return Dot(w, x) - a; }
      else
      {
         Vector c(1);
         // Includes parallel communication.
         problem->GetC()->Mult(x, c);

         return c(0) - (*problem->GetEqualityVec())(0);
      }
   }

   inline void print_iteration(int it, double r, double l) const;

public:
   SLBQPOptimizer() { }

#ifdef MFEM_USE_MPI
   SLBQPOptimizer(MPI_Comm comm_) : OptimizationSolver(comm_) { }
#endif

   /** Setting an OptimizationProblem will overwrite the Vectors given by
    *  SetBounds and SetLinearConstraint. The objective function remains
    *  unchanged. */
   virtual void SetOptimizationProblem(const OptimizationProblem &prob);

   void SetBounds(const Vector &lo_, const Vector &hi_);
   void SetLinearConstraint(const Vector &w_, double a_);

   /** We let the target values play the role of the initial vector xt, from
    *  which the operator generates the optimal vector x. */
   virtual void Mult(const Vector &xt, Vector &x) const;
};

/** Block ILU solver:
 *  Performs a block ILU(k) approximate factorization with specified block
 *  size. Currently only k=0 is supported. This is useful as a preconditioner
 *  for DG-type discretizations, where the system matrix has a natural
 *  (elemental) block structure.
 *
 *  In the case of DG discretizations, the block size should usually be set to
 *  either ndofs_per_element or vdim*ndofs_per_element (if the finite element
 *  space has Ordering::byVDIM). The block size must evenly divide the size of
 *  the matrix.
 *
 *  Renumbering the blocks is also supported by specifying a reordering method.
 *  Currently greedy minimum discarded fill ordering and no reordering are
 *  supported. Renumbering the blocks can lead to a much better approximate
 *  factorization.
 */
class BlockILU : public Solver
{
public:

   /// The reordering method used by the BlockILU factorization.
   enum class Reordering
   {
      MINIMUM_DISCARDED_FILL,
      NONE
   };

   /** Create an "empty" BlockILU solver. SetOperator must be called later to
    *  actually form the factorization
    */
   BlockILU(int block_size_,
            Reordering reordering_ = Reordering::MINIMUM_DISCARDED_FILL,
            int k_fill_ = 0);

   /** Create a block ILU approximate factorization for the matrix @a op.
    *  @a op should be of type either SparseMatrix or HypreParMatrix. In the
    *  case that @a op is a HypreParMatrix, the ILU factorization is performed
    *  on the diagonal blocks of the parallel decomposition.
    */
   BlockILU(const Operator &op, int block_size_ = 1,
            Reordering reordering_ = Reordering::MINIMUM_DISCARDED_FILL,
            int k_fill_ = 0);

   /** Perform the block ILU factorization for the matrix @a op.
    *  As in the constructor, @a op must either be a SparseMatrix or
    *  HypreParMatrix
    */
   void SetOperator(const Operator &op);

   /// Solve the system `LUx = b`, where `L` and `U` are the block ILU factors.
   void Mult(const Vector &b, Vector &x) const;

   /** Get the I array for the block CSR representation of the factorization.
    *  Similar to SparseMatrix::GetI(). Mostly used for testing.
    */
   int *GetBlockI() { return IB.GetData(); }

   /** Get the J array for the block CSR representation of the factorization.
    *  Similar to SparseMatrix::GetJ(). Mostly used for testing.
    */
   int *GetBlockJ() { return JB.GetData(); }

   /** Get the data array for the block CSR representation of the factorization.
    *  Similar to SparseMatrix::GetData(). Mostly used for testing.
    */
   double *GetBlockData() { return AB.Data(); }

private:
   /// Set up the block CSR structure corresponding to a sparse matrix @a A
   void CreateBlockPattern(const class SparseMatrix &A);

   /// Perform the block ILU factorization
   void Factorize();

   int block_size;

   /// Fill level for block ILU(k) factorizations. Only k=0 is supported.
   int k_fill;

   Reordering reordering;

   /// Temporary vector used in the Mult() function.
   mutable Vector y;

   /// Permutation and inverse permutation vectors for the block reordering.
   Array<int> P, Pinv;

   /** Block CSR storage of the factorization. The block upper triangular part
    *  stores the U factor. The L factor implicitly has identity on the diagonal
    *  blocks, and the rest of L is given by the strictly block lower triangular
    *  part.
    */
   Array<int> IB, ID, JB;
   DenseTensor AB;

   /// DB(i) stores the LU factorization of the i'th diagonal block
   mutable DenseTensor DB;
   /// Pivot arrays for the LU factorizations given by #DB
   mutable Array<int> ipiv;
};


/// Monitor that checks whether the residual is zero at a given set of dofs.
/** This monitor is useful for checking if the initial guess, rhs, operator, and
    preconditioner are properly setup for solving in the subspace with imposed
    essential boundary conditions. */
class ResidualBCMonitor : public IterativeSolverMonitor
{
protected:
   const Array<int> *ess_dofs_list; ///< Not owned

public:
   ResidualBCMonitor(const Array<int> &ess_dofs_list_)
      : ess_dofs_list(&ess_dofs_list_) { }

   void MonitorResidual(int it, double norm, const Vector &r,
                        bool final) override;
};


#ifdef MFEM_USE_SUITESPARSE

/// Direct sparse solver using UMFPACK
class UMFPackSolver : public Solver
{
protected:
   bool use_long_ints;
   SparseMatrix *mat;
   void *Numeric;
   SuiteSparse_long *AI, *AJ;

   void Init();

public:
   double Control[UMFPACK_CONTROL];
   mutable double Info[UMFPACK_INFO];

   /** @brief For larger matrices, if the solver fails, set the parameter @a
       use_long_ints_ = true. */
   UMFPackSolver(bool use_long_ints_ = false)
      : use_long_ints(use_long_ints_) { Init(); }
   /** @brief Factorize the given SparseMatrix using the defaults. For larger
       matrices, if the solver fails, set the parameter @a use_long_ints_ =
       true. */
   UMFPackSolver(SparseMatrix &A, bool use_long_ints_ = false)
      : use_long_ints(use_long_ints_) { Init(); SetOperator(A); }

   /** @brief Factorize the given Operator @a op which must be a SparseMatrix.

       The factorization uses the parameters set in the #Control data member.
       @note This method calls SparseMatrix::SortColumnIndices() with @a op,
       modifying the matrix if the column indices are not already sorted. */
   virtual void SetOperator(const Operator &op);

   /// Set the print level field in the #Control data member.
   void SetPrintLevel(int print_lvl) { Control[UMFPACK_PRL] = print_lvl; }

   virtual void Mult(const Vector &b, Vector &x) const;
   virtual void MultTranspose(const Vector &b, Vector &x) const;

   virtual ~UMFPackSolver();
};

/// Direct sparse solver using KLU
class KLUSolver : public Solver
{
protected:
   SparseMatrix *mat;
   klu_symbolic *Symbolic;
   klu_numeric *Numeric;

   void Init();

public:
   KLUSolver()
      : mat(0),Symbolic(0),Numeric(0)
   { Init(); }
   KLUSolver(SparseMatrix &A)
      : mat(0),Symbolic(0),Numeric(0)
   { Init(); SetOperator(A); }

   // Works on sparse matrices only; calls SparseMatrix::SortColumnIndices().
   virtual void SetOperator(const Operator &op);

   virtual void Mult(const Vector &b, Vector &x) const;
   virtual void MultTranspose(const Vector &b, Vector &x) const;

   virtual ~KLUSolver();

   mutable klu_common Common;
};

#endif // MFEM_USE_SUITESPARSE

/// Block diagonal solver for A, each block is inverted by direct solver
class DirectSubBlockSolver : public Solver
{
   SparseMatrix& block_dof;
   mutable Array<int> local_dofs;
   mutable Vector sub_rhs;
   mutable Vector sub_sol;
   std::unique_ptr<DenseMatrixInverse[]> block_solvers;
public:
   /// block_dof is a boolean matrix, block_dof(i, j) = 1 if j-th dof belongs to
   /// i-th block, block_dof(i, j) = 0 otherwise.
   DirectSubBlockSolver(const SparseMatrix& A, const SparseMatrix& block_dof);
   virtual void Mult(const Vector &x, Vector &y) const;
   virtual void SetOperator(const Operator &op) { }
};

/// Solver S such that I - A * S = (I - A * S1) * (I - A * S0).
/// That is, S = S0 + S1 - S1 * A * S0.
class ProductSolver : public Solver
{
   OperatorPtr A;
   OperatorPtr S0;
   OperatorPtr S1;
public:
   ProductSolver(Operator* A_, Solver* S0_, Solver* S1_,
                 bool ownA, bool ownS0, bool ownS1)
      : Solver(A_->NumRows()), A(A_, ownA), S0(S0_, ownS0), S1(S1_, ownS1) { }
   virtual void Mult(const Vector &x, Vector &y) const;
   virtual void MultTranspose(const Vector &x, Vector &y) const;
   virtual void SetOperator(const Operator &op) { }
};

/// Solver wrapper which orthogonalizes the input and output vector
/**
 * OrthoSolver wraps an existing Solver and orthogonalizes the input vector
 * before passing it to the Mult() method of the Solver. This is a convenience
 * implementation to handle e.g. a Poisson problem with pure Neumann boundary
 * conditions, where this procedure removes the Nullspace.
 */
class OrthoSolver : public Solver
{
private:
#ifdef MFEM_USE_MPI
   MPI_Comm mycomm;
   mutable HYPRE_BigInt global_size;
   const bool parallel;
#else
   mutable int global_size;
#endif

public:
   OrthoSolver();
#ifdef MFEM_USE_MPI
   OrthoSolver(MPI_Comm mycomm_);
#endif

   /// Set the solver used by the OrthoSolver.
   /** The action of the OrthoSolver is given by P * s * P where P is the
       projection to the subspace of vectors with zero sum. Calling this method
       is required before calling SetOperator() or Mult(). */
   void SetSolver(Solver &s);

   /// Set the Operator that is the OrthoSolver is to invert (approximately).
   /** The Operator @a op is simply forwarded to the solver object given by
       SetSolver() which needs to be called before this method. Calling this
       method is optional when the solver already has an associated Operator. */
   virtual void SetOperator(const Operator &op);

   /** @brief Perform the action of the OrthoSolver: P * solver * P where P is
       the projection to the subspace of vectors with zero sum. */
   /** @note The projection P can be written as P = I - 1 1^T / (1^T 1) where
       I is the identity matrix and 1 is the column-vector with all components
       equal to 1. */
   void Mult(const Vector &b, Vector &x) const;

private:
   Solver *solver = nullptr;

   mutable Vector b_ortho;

   void Orthogonalize(const Vector &v, Vector &v_ortho) const;
};

#ifdef MFEM_USE_MPI
/** This smoother does relaxations on an auxiliary space (determined by a map
    from the original space to the auxiliary space provided by the user).
    The smoother on the auxiliary space is a HypreSmoother. Its options can be
    modified through GetSmoother.
    For example, the space can be the nullspace of div/curl, in which case the
    smoother can be used to construct a Hiptmair smoother. */
class AuxSpaceSmoother : public Solver
{
   OperatorPtr aux_map_;
   OperatorPtr aux_system_;
   OperatorPtr aux_smoother_;
   void Mult(const Vector &x, Vector &y, bool transpose) const;
public:
   AuxSpaceSmoother(const HypreParMatrix &op, HypreParMatrix *aux_map,
                    bool op_is_symmetric = true, bool own_aux_map = false);
   virtual void Mult(const Vector &x, Vector &y) const { Mult(x, y, false); }
   virtual void MultTranspose(const Vector &x, Vector &y) const { Mult(x, y, true); }
   virtual void SetOperator(const Operator &op) { }
   HypreSmoother& GetSmoother() { return *aux_smoother_.As<HypreSmoother>(); }
   using Operator::Mult;
};
#endif // MFEM_USE_MPI

}

#endif // MFEM_SOLVERS