fastamg.hh

Go to the documentation of this file.

// SPDX-FileCopyrightText: Copyright © DUNE Project contributors, see file LICENSE.md in module root
// SPDX-License-Identifier: LicenseRef-GPL-2.0-only-with-DUNE-exception
// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*-
// vi: set et ts=4 sw=2 sts=2:
#ifndef DUNE_ISTL_FASTAMG_HH
#define DUNE_ISTL_FASTAMG_HH

#include <memory>
#include <dune/common/exceptions.hh>
#include <dune/common/typetraits.hh>
#include <dune/istl/paamg/smoother.hh>
#include <dune/istl/paamg/transfer.hh>
#include <dune/istl/paamg/matrixhierarchy.hh>
#include <dune/istl/solvers.hh>
#include <dune/istl/scalarproducts.hh>
#include <dune/istl/superlu.hh>
#include <dune/istl/umfpack.hh>
#include <dune/istl/solvertype.hh>
#include <dune/istl/io.hh>
#include <dune/istl/preconditioners.hh>

#include "fastamgsmoother.hh"

namespace Dune
{
  namespace Amg
  {
    template<class M, class X, class PI=SequentialInformation, class A=std::allocator<X> >
    class FastAMG : public Preconditioner<X,X>
    {
    public:
      typedef M Operator;
      typedef PI ParallelInformation;
      typedef MatrixHierarchy<M, ParallelInformation, A> OperatorHierarchy;
      typedef typename OperatorHierarchy::ParallelInformationHierarchy ParallelInformationHierarchy;

      typedef X Domain;
      typedef X Range;
      typedef InverseOperator<X,X> CoarseSolver;

      FastAMG(OperatorHierarchy& matrices, CoarseSolver& coarseSolver,
              const Parameters& parms,
              bool symmetric=true);

      template<class C>
      FastAMG(std::shared_ptr<const Operator> fineOperator,
              const C& criterion,
              const Parameters& parms=Parameters(),
              bool symmetric=true,
              const ParallelInformation& pinfo=ParallelInformation());

      template<class C>
      FastAMG(const Operator& fineOperator,
              const C& criterion,
              const Parameters& parms=Parameters(),
              bool symmetric=true,
              const ParallelInformation& pinfo=ParallelInformation())
        : FastAMG(stackobject_to_shared_ptr(fineOperator),
                  criterion, parms, symmetric, pinfo)
      {}

      FastAMG(const FastAMG& amg);

      void pre(Domain& x, Range& b);

      void apply(Domain& v, const Range& d);

      virtual SolverCategory::Category category() const
      {
        return SolverCategory::sequential;
      }

      void post(Domain& x);

      template<class A1>
      void getCoarsestAggregateNumbers(std::vector<std::size_t,A1>& cont);

      std::size_t levels();

      std::size_t maxlevels();

      void recalculateHierarchy()
      {
        matrices_->recalculateGalerkin(NegateSet<typename PI::OwnerSet>());
      }

      bool usesDirectCoarseLevelSolver() const;

    private:
      template<class C>
      void createHierarchies(C& criterion,
                             std::shared_ptr<const Operator> fineOperator,
                             const PI& pinfo);

      struct LevelContext
      {
        typename OperatorHierarchy::ParallelMatrixHierarchy::ConstIterator matrix;
        typename ParallelInformationHierarchy::Iterator pinfo;
        typename OperatorHierarchy::RedistributeInfoList::const_iterator redist;
        typename OperatorHierarchy::AggregatesMapList::const_iterator aggregates;
        typename Hierarchy<Domain,A>::Iterator lhs;
        typename Hierarchy<Domain,A>::Iterator residual;
        typename Hierarchy<Range,A>::Iterator rhs;
        std::size_t level;
      };

      void mgc(LevelContext& levelContext, Domain& x, const Range& b);

      void presmooth(LevelContext& levelContext, Domain& x, const Range& b);

      void postsmooth(LevelContext& levelContext, Domain& x, const Range& b);

      void moveToFineLevel(LevelContext& levelContext, bool processedFineLevel,
                           Domain& fineX);

      bool moveToCoarseLevel(LevelContext& levelContext);

      void initIteratorsWithFineLevel(LevelContext& levelContext);

      std::shared_ptr<OperatorHierarchy> matrices_;
      std::shared_ptr<CoarseSolver> solver_;
      std::shared_ptr<Hierarchy<Range,A>> rhs_;
      std::shared_ptr<Hierarchy<Domain,A>> lhs_;
      std::shared_ptr<Hierarchy<Domain,A>> residual_;

      using ScalarProduct = Dune::ScalarProduct<X>;
      std::shared_ptr<ScalarProduct> scalarProduct_;
      std::size_t gamma_;
      std::size_t preSteps_;
      std::size_t postSteps_;
      std::size_t level;
      bool buildHierarchy_;
      bool symmetric;
      bool coarsesolverconverged;
      typedef SeqSSOR<typename M::matrix_type,X,X> Smoother;
      typedef std::shared_ptr<Smoother> SmootherPointer;
      SmootherPointer coarseSmoother_;
      std::size_t verbosity_;
    };

    template<class M, class X, class PI, class A>
    FastAMG<M,X,PI,A>::FastAMG(const FastAMG& amg)
    : matrices_(amg.matrices_), solver_(amg.solver_),
      rhs_(), lhs_(), residual_(), scalarProduct_(amg.scalarProduct_),
      gamma_(amg.gamma_), preSteps_(amg.preSteps_), postSteps_(amg.postSteps_),
      symmetric(amg.symmetric), coarsesolverconverged(amg.coarsesolverconverged),
      coarseSmoother_(amg.coarseSmoother_), verbosity_(amg.verbosity_)
    {}

    template<class M, class X, class PI, class A>
    FastAMG<M,X,PI,A>::FastAMG(OperatorHierarchy& matrices, CoarseSolver& coarseSolver,
                               const Parameters& parms, bool symmetric_)
      : matrices_(stackobject_to_shared_ptr(matrices)), solver_(&coarseSolver),
        rhs_(), lhs_(), residual_(), scalarProduct_(),
        gamma_(parms.getGamma()), preSteps_(parms.getNoPreSmoothSteps()),
        postSteps_(parms.getNoPostSmoothSteps()), buildHierarchy_(false),
        symmetric(symmetric_), coarsesolverconverged(true),
        coarseSmoother_(), verbosity_(parms.debugLevel())
    {
      if(preSteps_>1||postSteps_>1)
      {
        std::cerr<<"WARNING only one step of smoothing is supported!"<<std::endl;
        preSteps_=postSteps_=0;
      }
      assert(matrices_->isBuilt());
      static_assert(std::is_same<PI,SequentialInformation>::value,
                    "Currently only sequential runs are supported");
    }
    template<class M, class X, class PI, class A>
    template<class C>
    FastAMG<M,X,PI,A>::FastAMG(std::shared_ptr<const Operator> fineOperator,
                               const C& criterion,
                               const Parameters& parms,
                               bool symmetric_,
                               const PI& pinfo)
      : solver_(), rhs_(), lhs_(), residual_(), scalarProduct_(), gamma_(parms.getGamma()),
        preSteps_(parms.getNoPreSmoothSteps()), postSteps_(parms.getNoPostSmoothSteps()),
        buildHierarchy_(true),
        symmetric(symmetric_), coarsesolverconverged(true),
        coarseSmoother_(), verbosity_(criterion.debugLevel())
    {
      if(preSteps_>1||postSteps_>1)
      {
        std::cerr<<"WARNING only one step of smoothing is supported!"<<std::endl;
        preSteps_=postSteps_=1;
      }
      static_assert(std::is_same<PI,SequentialInformation>::value,
                    "Currently only sequential runs are supported");
      // TODO: reestablish compile time checks.
      //static_assert(static_cast<int>(PI::category)==static_cast<int>(S::category),
      //             "Matrix and Solver must match in terms of category!");
      createHierarchies(criterion, std::move(fineOperator), pinfo);
    }

    template<class M, class X, class PI, class A>
    template<class C>
    void FastAMG<M,X,PI,A>::createHierarchies(C& criterion,
      std::shared_ptr<const Operator> fineOperator,
      const PI& pinfo)
    {
      Timer watch;
      matrices_ = std::make_shared<OperatorHierarchy>(
        std::const_pointer_cast<Operator>(std::move(fineOperator)),
        stackobject_to_shared_ptr(const_cast<PI&>(pinfo)));

      matrices_->template build<NegateSet<typename PI::OwnerSet> >(criterion);

      if(verbosity_>0 && matrices_->parallelInformation().finest()->communicator().rank()==0)
        std::cout<<"Building Hierarchy of "<<matrices_->maxlevels()<<" levels took "<<watch.elapsed()<<" seconds."<<std::endl;

      if(buildHierarchy_ && matrices_->levels()==matrices_->maxlevels()) {
        // We have the carsest level. Create the coarse Solver
        typedef typename SmootherTraits<Smoother>::Arguments SmootherArgs;
        SmootherArgs sargs;
        sargs.iterations = 1;

        typename ConstructionTraits<Smoother>::Arguments cargs;
        cargs.setArgs(sargs);
        if(matrices_->redistributeInformation().back().isSetup()) {
          // Solve on the redistributed partitioning
          cargs.setMatrix(matrices_->matrices().coarsest().getRedistributed().getmat());
          cargs.setComm(matrices_->parallelInformation().coarsest().getRedistributed());
        }else{
          cargs.setMatrix(matrices_->matrices().coarsest()->getmat());
          cargs.setComm(*matrices_->parallelInformation().coarsest());
        }

        coarseSmoother_ = ConstructionTraits<Smoother>::construct(cargs);
        scalarProduct_ = createScalarProduct<X>(cargs.getComm(),category());

#if HAVE_SUPERLU|| HAVE_SUITESPARSE_UMFPACK
#if HAVE_SUITESPARSE_UMFPACK
#define DIRECTSOLVER UMFPack
#else
#define DIRECTSOLVER SuperLU
#endif
        // Use superlu if we are purely sequential or with only one processor on the coarsest level.
        if(std::is_same<ParallelInformation,SequentialInformation>::value // sequential mode
           || matrices_->parallelInformation().coarsest()->communicator().size()==1 //parallel mode and only one processor
           || (matrices_->parallelInformation().coarsest().isRedistributed()
               && matrices_->parallelInformation().coarsest().getRedistributed().communicator().size()==1
               && matrices_->parallelInformation().coarsest().getRedistributed().communicator().size()>0)) { // redistribute and 1 proc
          if(verbosity_>0 && matrices_->parallelInformation().coarsest()->communicator().rank()==0)
            std::cout<<"Using superlu"<<std::endl;
          if(matrices_->parallelInformation().coarsest().isRedistributed())
          {
            if(matrices_->matrices().coarsest().getRedistributed().getmat().N()>0)
              // We are still participating on this level
              solver_.reset(new DIRECTSOLVER<typename M::matrix_type>(matrices_->matrices().coarsest().getRedistributed().getmat(), false, false));
            else
              solver_.reset();
          }else
            solver_.reset(new DIRECTSOLVER<typename M::matrix_type>(matrices_->matrices().coarsest()->getmat(), false, false));
        }else
#undef DIRECTSOLVER
#endif // HAVE_SUPERLU|| HAVE_SUITESPARSE_UMFPACK
        {
          if(matrices_->parallelInformation().coarsest().isRedistributed())
          {
            if(matrices_->matrices().coarsest().getRedistributed().getmat().N()>0)
              // We are still participating on this level
              solver_.reset(new BiCGSTABSolver<X>(const_cast<M&>(matrices_->matrices().coarsest().getRedistributed()),
                                                  *scalarProduct_,
                                                  *coarseSmoother_, 1E-2, 1000, 0));
            else
              solver_.reset();
          }else
            solver_.reset(new BiCGSTABSolver<X>(const_cast<M&>(*matrices_->matrices().coarsest()),
                                                *scalarProduct_,
                                                *coarseSmoother_, 1E-2, 1000, 0));
        }
      }

      if(verbosity_>0 && matrices_->parallelInformation().finest()->communicator().rank()==0)
        std::cout<<"Building Hierarchy of "<<matrices_->maxlevels()<<" levels took "<<watch.elapsed()<<" seconds."<<std::endl;
    }


    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>::pre(Domain& x, Range& b)
    {
      Timer watch, watch1;
      // Detect Matrix rows where all offdiagonal entries are
      // zero and set x such that  A_dd*x_d=b_d
      // Thus users can be more careless when setting up their linear
      // systems.
      typedef typename M::matrix_type Matrix;
      typedef typename Matrix::ConstRowIterator RowIter;
      typedef typename Matrix::ConstColIterator ColIter;
      typedef typename Matrix::block_type Block;
      Block zero;
      zero=typename Matrix::field_type();

      const Matrix& mat=matrices_->matrices().finest()->getmat();
      for(RowIter row=mat.begin(); row!=mat.end(); ++row) {
        bool isDirichlet = true;
        bool hasDiagonal = false;
        ColIter diag;
        for(ColIter col=row->begin(); col!=row->end(); ++col) {
          if(row.index()==col.index()) {
            diag = col;
            hasDiagonal = (*col != zero);
          }else{
            if(*col!=zero)
              isDirichlet = false;
          }
        }
        if(isDirichlet && hasDiagonal)
        {
          if constexpr (Dune::IsNumber<Block>::value)
            x[row.index()] = b[row.index()]/(*diag);
          else
            diag->solve(x[row.index()], b[row.index()]);
        }
      }
      if (verbosity_>0)
        std::cout<<" Preprocessing Dirichlet took "<<watch1.elapsed()<<std::endl;
      watch1.reset();
      // No smoother to make x consistent! Do it by hand
      matrices_->parallelInformation().coarsest()->copyOwnerToAll(x,x);
      rhs_ = std::make_shared<Hierarchy<Range,A>>(std::make_shared<Range>(b));
      lhs_ = std::make_shared<Hierarchy<Domain,A>>(std::make_shared<Domain>(x));
      residual_ = std::make_shared<Hierarchy<Domain,A>>(std::make_shared<Domain>(x));
      matrices_->coarsenVector(*rhs_);
      matrices_->coarsenVector(*lhs_);
      matrices_->coarsenVector(*residual_);

      // The preconditioner might change x and b. So we have to
      // copy the changes to the original vectors.
      x = *lhs_->finest();
      b = *rhs_->finest();
    }
    template<class M, class X, class PI, class A>
    std::size_t FastAMG<M,X,PI,A>::levels()
    {
      return matrices_->levels();
    }
    template<class M, class X, class PI, class A>
    std::size_t FastAMG<M,X,PI,A>::maxlevels()
    {
      return matrices_->maxlevels();
    }

    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>::apply(Domain& v, const Range& d)
    {
      LevelContext levelContext;
      // Init all iterators for the current level
      initIteratorsWithFineLevel(levelContext);

      assert(v.two_norm()==0);

      level=0;
      if(matrices_->maxlevels()==1){
        // The coarse solver might modify the d!
        Range b(d);
        mgc(levelContext, v, b);
      }else
        mgc(levelContext, v, d);
      if(postSteps_==0||matrices_->maxlevels()==1)
        levelContext.pinfo->copyOwnerToAll(v, v);
    }

    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>::initIteratorsWithFineLevel(LevelContext& levelContext)
    {
      levelContext.matrix = matrices_->matrices().finest();
      levelContext.pinfo = matrices_->parallelInformation().finest();
      levelContext.redist =
        matrices_->redistributeInformation().begin();
      levelContext.aggregates = matrices_->aggregatesMaps().begin();
      levelContext.lhs = lhs_->finest();
      levelContext.residual = residual_->finest();
      levelContext.rhs = rhs_->finest();
      levelContext.level=0;
    }

    template<class M, class X, class PI, class A>
    bool FastAMG<M,X,PI,A>
    ::moveToCoarseLevel(LevelContext& levelContext)
    {
      bool processNextLevel=true;

      if(levelContext.redist->isSetup()) {
        throw "bla";
        levelContext.redist->redistribute(static_cast<const Range&>(*levelContext.residual),
                                          levelContext.residual.getRedistributed());
        processNextLevel = levelContext.residual.getRedistributed().size()>0;
        if(processNextLevel) {
          //restrict defect to coarse level right hand side.
          ++levelContext.pinfo;
          Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation>
          ::restrictVector(*(*levelContext.aggregates), *levelContext.rhs,
                           static_cast<const Range&>(levelContext.residual.getRedistributed()),
                           *levelContext.pinfo);
        }
      }else{
        //restrict defect to coarse level right hand side.
        ++levelContext.rhs;
        ++levelContext.pinfo;
        Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation>
        ::restrictVector(*(*levelContext.aggregates), *levelContext.rhs,
                         static_cast<const Range&>(*levelContext.residual), *levelContext.pinfo);
      }

      if(processNextLevel) {
        // prepare coarse system
        ++levelContext.residual;
        ++levelContext.lhs;
        ++levelContext.matrix;
        ++levelContext.level;
        ++levelContext.redist;

        if(levelContext.matrix != matrices_->matrices().coarsest() || matrices_->levels()<matrices_->maxlevels()) {
          // next level is not the globally coarsest one
          ++levelContext.aggregates;
        }
        // prepare the lhs on the next level
        *levelContext.lhs=0;
        *levelContext.residual=0;
      }
      return processNextLevel;
    }

    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>
    ::moveToFineLevel(LevelContext& levelContext, bool processNextLevel, Domain& x)
    {
      if(processNextLevel) {
        if(levelContext.matrix != matrices_->matrices().coarsest() || matrices_->levels()<matrices_->maxlevels()) {
          // previous level is not the globally coarsest one
          --levelContext.aggregates;
        }
        --levelContext.redist;
        --levelContext.level;
        //prolongate and add the correction (update is in coarse left hand side)
        --levelContext.matrix;
        --levelContext.residual;

      }

      typename Hierarchy<Domain,A>::Iterator coarseLhs = levelContext.lhs--;
      if(levelContext.redist->isSetup()) {

        // Need to redistribute during prolongate
        Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation>
        ::prolongateVector(*(*levelContext.aggregates), *coarseLhs, x,
                           levelContext.lhs.getRedistributed(),
                           matrices_->getProlongationDampingFactor(),
                           *levelContext.pinfo, *levelContext.redist);
      }else{
        Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation>
        ::prolongateVector(*(*levelContext.aggregates), *coarseLhs, x,
                           matrices_->getProlongationDampingFactor(), *levelContext.pinfo);

        // printvector(std::cout, *lhs, "prolongated coarse grid correction", "lhs", 10, 10, 10);
      }


      if(processNextLevel) {
        --levelContext.rhs;
      }

    }


    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>
    ::presmooth(LevelContext& levelContext, Domain& x, const Range& b)
    {
      constexpr auto bl = blockLevel<typename M::matrix_type>();
      GaussSeidelPresmoothDefect<bl>::apply(levelContext.matrix->getmat(),
                                            x,
                                            *levelContext.residual,
                                            b);
    }

    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>
    ::postsmooth(LevelContext& levelContext, Domain& x, const Range& b)
    {
      constexpr auto bl = blockLevel<typename M::matrix_type>();
      GaussSeidelPostsmoothDefect<bl>
      ::apply(levelContext.matrix->getmat(), x, *levelContext.residual, b);
    }


    template<class M, class X, class PI, class A>
    bool FastAMG<M,X,PI,A>::usesDirectCoarseLevelSolver() const
    {
      return IsDirectSolver< CoarseSolver>::value;
    }

    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>::mgc(LevelContext& levelContext, Domain& v, const Range& b){

      if(levelContext.matrix == matrices_->matrices().coarsest() && levels()==maxlevels()) {
        // Solve directly
        InverseOperatorResult res;
        res.converged=true; // If we do not compute this flag will not get updated
        if(levelContext.redist->isSetup()) {
          levelContext.redist->redistribute(b, levelContext.rhs.getRedistributed());
          if(levelContext.rhs.getRedistributed().size()>0) {
            // We are still participating in the computation
            levelContext.pinfo.getRedistributed().copyOwnerToAll(levelContext.rhs.getRedistributed(),
                                                                 levelContext.rhs.getRedistributed());
            solver_->apply(levelContext.lhs.getRedistributed(), levelContext.rhs.getRedistributed(), res);
          }
          levelContext.redist->redistributeBackward(v, levelContext.lhs.getRedistributed());
          levelContext.pinfo->copyOwnerToAll(v, v);
        }else{
          levelContext.pinfo->copyOwnerToAll(b, b);
          solver_->apply(v, const_cast<Range&>(b), res);
        }

        // printvector(std::cout, *lhs, "coarse level update", "u", 10, 10, 10);
        // printvector(std::cout, *rhs, "coarse level rhs", "rhs", 10, 10, 10);
        if (!res.converged)
          coarsesolverconverged = false;
      }else{
        // presmoothing
        presmooth(levelContext, v, b);
        // printvector(std::cout, *lhs, "update", "u", 10, 10, 10);
        // printvector(std::cout, *residual, "post presmooth residual", "r", 10);
#ifndef DUNE_AMG_NO_COARSEGRIDCORRECTION
        bool processNextLevel = moveToCoarseLevel(levelContext);

        if(processNextLevel) {
          // next level
          for(std::size_t i=0; i<gamma_; i++)
            mgc(levelContext, *levelContext.lhs, *levelContext.rhs);
        }

        moveToFineLevel(levelContext, processNextLevel, v);
#else
        *lhs=0;
#endif

        if(levelContext.matrix == matrices_->matrices().finest()) {
          coarsesolverconverged = matrices_->parallelInformation().finest()->communicator().prod(coarsesolverconverged);
          if(!coarsesolverconverged)
            DUNE_THROW(MathError, "Coarse solver did not converge");
        }

        postsmooth(levelContext, v, b);
      }
    }


    template<class M, class X, class PI, class A>
    void FastAMG<M,X,PI,A>::post([[maybe_unused]] Domain& x)
    {
      lhs_=nullptr;
      rhs_=nullptr;
      residual_=nullptr;
    }

    template<class M, class X, class PI, class A>
    template<class A1>
    void FastAMG<M,X,PI,A>::getCoarsestAggregateNumbers(std::vector<std::size_t,A1>& cont)
    {
      matrices_->getCoarsestAggregatesOnFinest(cont);
    }

  } // end namespace Amg
} // end namespace Dune

#endif