| 1 |
// SPDX-FileCopyrightText: Copyright © DUNE Project contributors, see file LICENSE.md in module root |
| 2 |
// SPDX-License-Identifier: LicenseRef-GPL-2.0-only-with-DUNE-exception |
| 3 |
// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*- |
| 4 |
// vi: set et ts=4 sw=2 sts=2: |
| 5 |
#ifndef DUNE_ISTL_FASTAMG_HH |
| 6 |
#define DUNE_ISTL_FASTAMG_HH |
| 7 |
|
| 8 |
#include <memory> |
| 9 |
#include <dune/common/exceptions.hh> |
| 10 |
#include <dune/common/typetraits.hh> |
| 11 |
#include <dune/istl/paamg/smoother.hh> |
| 12 |
#include <dune/istl/paamg/transfer.hh> |
| 13 |
#include <dune/istl/paamg/matrixhierarchy.hh> |
| 14 |
#include <dune/istl/solvers.hh> |
| 15 |
#include <dune/istl/scalarproducts.hh> |
| 16 |
#include <dune/istl/superlu.hh> |
| 17 |
#include <dune/istl/umfpack.hh> |
| 18 |
#include <dune/istl/solvertype.hh> |
| 19 |
#include <dune/istl/io.hh> |
| 20 |
#include <dune/istl/preconditioners.hh> |
| 21 |
|
| 22 |
#include "fastamgsmoother.hh" |
| 23 |
|
| 24 |
/** @file |
| 25 |
* @author Markus Blatt |
| 26 |
* @brief A fast AMG method, that currently only allows only Gauss-Seidel |
| 27 |
* smoothing and is currently purely sequential. It |
| 28 |
* combines one Gauss-Seidel presmoothing sweep with |
| 29 |
* the defect calculation to reduce memory transfers. |
| 30 |
*/ |
| 31 |
|
| 32 |
namespace Dune |
| 33 |
{ |
| 34 |
namespace Amg |
| 35 |
{ |
| 36 |
/** |
| 37 |
* @defgroup ISTL_FSAMG Fast (sequential) Algebraic Multigrid |
| 38 |
* @ingroup ISTL_Prec |
| 39 |
* @brief An Algebraic Multigrid based on Agglomeration that saves memory bandwidth. |
| 40 |
*/ |
| 41 |
|
| 42 |
/** |
| 43 |
* @addtogroup ISTL_FSAMG |
| 44 |
* |
| 45 |
* @{ |
| 46 |
*/ |
| 47 |
|
| 48 |
/** |
| 49 |
* @brief A fast (sequential) algebraic multigrid based on agglomeration that saves memory bandwidth. |
| 50 |
* |
| 51 |
* It combines one Gauss-Seidel smoothing sweep with |
| 52 |
* the defect calculation to reduce memory transfers. |
| 53 |
* \tparam M The matrix-operator type, e.g. MatrixAdapter or AssembledLinearOperator |
| 54 |
* \tparam X The vector type |
| 55 |
* \tparam PI Currently ignored. |
| 56 |
* \tparam A An allocator for X |
| 57 |
*/ |
| 58 |
template<class M, class X, class PI=SequentialInformation, class A=std::allocator<X> > |
| 59 |
class FastAMG : public Preconditioner<X,X> |
| 60 |
{ |
| 61 |
public: |
| 62 |
/** @brief The matrix operator type. */ |
| 63 |
typedef M Operator; |
| 64 |
/** |
| 65 |
* @brief The type of the parallel information. |
| 66 |
* Either OwnerOverlapCommunication or another type |
| 67 |
* describing the parallel data distribution and |
| 68 |
* providing communication methods. |
| 69 |
*/ |
| 70 |
typedef PI ParallelInformation; |
| 71 |
/** @brief The operator hierarchy type. */ |
| 72 |
typedef MatrixHierarchy<M, ParallelInformation, A> OperatorHierarchy; |
| 73 |
/** @brief The parallal data distribution hierarchy type. */ |
| 74 |
typedef typename OperatorHierarchy::ParallelInformationHierarchy ParallelInformationHierarchy; |
| 75 |
|
| 76 |
/** @brief The domain type. */ |
| 77 |
typedef X Domain; |
| 78 |
/** @brief The range type. */ |
| 79 |
typedef X Range; |
| 80 |
/** @brief the type of the coarse solver. */ |
| 81 |
typedef InverseOperator<X,X> CoarseSolver; |
| 82 |
|
| 83 |
/** |
| 84 |
* @brief Construct a new amg with a specific coarse solver. |
| 85 |
* @param matrices The already set-up matrix hierarchy. |
| 86 |
* @param coarseSolver The set up solver to use on the coarse |
| 87 |
* grid, must match the coarse matrix in the matrix hierarchy. |
| 88 |
* @param parms The parameters for the AMG. |
| 89 |
*/ |
| 90 |
FastAMG(OperatorHierarchy& matrices, CoarseSolver& coarseSolver, |
| 91 |
const Parameters& parms, |
| 92 |
bool symmetric=true); |
| 93 |
|
| 94 |
/** |
| 95 |
* @brief Construct an AMG with an inexact coarse solver based on the smoother. |
| 96 |
* |
| 97 |
* As coarse solver a preconditioned CG method with the smoother as preconditioner |
| 98 |
* will be used. The matrix hierarchy is built automatically. |
| 99 |
* @param fineOperator The operator on the fine level. |
| 100 |
* @param criterion The criterion describing the coarsening strategy. E. g. SymmetricCriterion |
| 101 |
* or UnsymmetricCriterion, and providing the parameters. |
| 102 |
* @param parms The parameters for the AMG. |
| 103 |
* @param pinfo The information about the parallel distribution of the data. |
| 104 |
*/ |
| 105 |
template<class C> |
| 106 |
FastAMG(std::shared_ptr<const Operator> fineOperator, |
| 107 |
const C& criterion, |
| 108 |
const Parameters& parms=Parameters(), |
| 109 |
bool symmetric=true, |
| 110 |
const ParallelInformation& pinfo=ParallelInformation()); |
| 111 |
|
| 112 |
/** |
| 113 |
* @brief Construct an AMG with an inexact coarse solver based on the smoother. |
| 114 |
* |
| 115 |
* As coarse solver a preconditioned CG method with the smoother as preconditioner |
| 116 |
* will be used. The matrix hierarchy is built automatically. |
| 117 |
* @param fineOperator The operator on the fine level. |
| 118 |
* @param criterion The criterion describing the coarsening strategy. E. g. SymmetricCriterion |
| 119 |
* or UnsymmetricCriterion, and providing the parameters. |
| 120 |
* @param parms The parameters for the AMG. |
| 121 |
* |
| 122 |
* NOTE: This constructor uses the default constructed `ParallelInformation`. |
| 123 |
*/ |
| 124 |
template<class C> |
| 125 |
FastAMG(const Operator& fineOperator, |
| 126 |
const C& criterion, |
| 127 |
const Parameters& parms=Parameters(), |
| 128 |
bool symmetric=true, |
| 129 |
const ParallelInformation& pinfo=ParallelInformation()) |
| 130 |
: FastAMG(stackobject_to_shared_ptr(fineOperator), |
| 131 |
criterion, parms, symmetric, pinfo) |
| 132 |
{} |
| 133 |
|
| 134 |
/** |
| 135 |
* @brief Copy constructor. |
| 136 |
*/ |
| 137 |
FastAMG(const FastAMG& amg); |
| 138 |
|
| 139 |
/** \copydoc Preconditioner::pre */ |
| 140 |
void pre(Domain& x, Range& b); |
| 141 |
|
| 142 |
/** \copydoc Preconditioner::apply */ |
| 143 |
void apply(Domain& v, const Range& d); |
| 144 |
|
| 145 |
//! Category of the preconditioner (see SolverCategory::Category) |
| 146 |
virtual SolverCategory::Category category() const |
| 147 |
{ |
| 148 |
return SolverCategory::sequential; |
| 149 |
} |
| 150 |
|
| 151 |
/** \copydoc Preconditioner::post */ |
| 152 |
void post(Domain& x); |
| 153 |
|
| 154 |
/** |
| 155 |
* @brief Get the aggregate number of each unknown on the coarsest level. |
| 156 |
* @param cont The random access container to store the numbers in. |
| 157 |
*/ |
| 158 |
template<class A1> |
| 159 |
void getCoarsestAggregateNumbers(std::vector<std::size_t,A1>& cont); |
| 160 |
|
| 161 |
std::size_t levels(); |
| 162 |
|
| 163 |
std::size_t maxlevels(); |
| 164 |
|
| 165 |
/** |
| 166 |
* @brief Recalculate the matrix hierarchy. |
| 167 |
* |
| 168 |
* It is assumed that the coarsening for the changed fine level |
| 169 |
* matrix would yield the same aggregates. In this case it suffices |
| 170 |
* to recalculate all the Galerkin products for the matrices of the |
| 171 |
* coarser levels. |
| 172 |
*/ |
| 173 |
void recalculateHierarchy() |
| 174 |
{ |
| 175 |
matrices_->recalculateGalerkin(NegateSet<typename PI::OwnerSet>()); |
| 176 |
} |
| 177 |
|
| 178 |
/** |
| 179 |
* @brief Check whether the coarse solver used is a direct solver. |
| 180 |
* @return True if the coarse level solver is a direct solver. |
| 181 |
*/ |
| 182 |
bool usesDirectCoarseLevelSolver() const; |
| 183 |
|
| 184 |
private: |
| 185 |
/** |
| 186 |
* @brief Create matrix and smoother hierarchies. |
| 187 |
* @param criterion The coarsening criterion. |
| 188 |
* @param matrix The fine level matrix operator. |
| 189 |
* @param pinfo The fine level parallel information. |
| 190 |
*/ |
| 191 |
template<class C> |
| 192 |
void createHierarchies(C& criterion, |
| 193 |
std::shared_ptr<const Operator> fineOperator, |
| 194 |
const PI& pinfo); |
| 195 |
|
| 196 |
/** |
| 197 |
* @brief A struct that holds the context of the current level. |
| 198 |
* |
| 199 |
* These are the iterators to the smoother, matrix, parallel information, |
| 200 |
* and so on needed for the computations on the current level. |
| 201 |
*/ |
| 202 |
struct LevelContext |
| 203 |
{ |
| 204 |
/** |
| 205 |
* @brief The iterator over the matrices. |
| 206 |
*/ |
| 207 |
typename OperatorHierarchy::ParallelMatrixHierarchy::ConstIterator matrix; |
| 208 |
/** |
| 209 |
* @brief The iterator over the parallel information. |
| 210 |
*/ |
| 211 |
typename ParallelInformationHierarchy::Iterator pinfo; |
| 212 |
/** |
| 213 |
* @brief The iterator over the redistribution information. |
| 214 |
*/ |
| 215 |
typename OperatorHierarchy::RedistributeInfoList::const_iterator redist; |
| 216 |
/** |
| 217 |
* @brief The iterator over the aggregates maps. |
| 218 |
*/ |
| 219 |
typename OperatorHierarchy::AggregatesMapList::const_iterator aggregates; |
| 220 |
/** |
| 221 |
* @brief The iterator over the left hand side. |
| 222 |
*/ |
| 223 |
typename Hierarchy<Domain,A>::Iterator lhs; |
| 224 |
/** |
| 225 |
* @brief The iterator over the residuals. |
| 226 |
*/ |
| 227 |
typename Hierarchy<Domain,A>::Iterator residual; |
| 228 |
/** |
| 229 |
* @brief The iterator over the right hand sided. |
| 230 |
*/ |
| 231 |
typename Hierarchy<Range,A>::Iterator rhs; |
| 232 |
/** |
| 233 |
* @brief The level index. |
| 234 |
*/ |
| 235 |
std::size_t level; |
| 236 |
}; |
| 237 |
|
| 238 |
/** @brief Multigrid cycle on a level. */ |
| 239 |
void mgc(LevelContext& levelContext, Domain& x, const Range& b); |
| 240 |
|
| 241 |
/** |
| 242 |
* @brief Apply pre smoothing on the current level. |
| 243 |
* @param levelContext The context with the iterators for the level. |
| 244 |
* @param x The left hand side at the current level. |
| 245 |
* @param b The right hand side at the current level. |
| 246 |
*/ |
| 247 |
void presmooth(LevelContext& levelContext, Domain& x, const Range& b); |
| 248 |
|
| 249 |
/** |
| 250 |
* @brief Apply post smoothing on the current level. |
| 251 |
* @param levelContext The context with the iterators for the level. |
| 252 |
* @param x The left hand side at the current level. |
| 253 |
* @param b The right hand side at the current level. |
| 254 |
*/ |
| 255 |
void postsmooth(LevelContext& levelContext, Domain& x, const Range& b); |
| 256 |
|
| 257 |
/** |
| 258 |
* @brief Move the iterators to the finer level |
| 259 |
* @param levelContext The context with the iterators for the level. |
| 260 |
* @param processedFineLevel whether the process did compute on the finer level |
| 261 |
* @param fineX The vector to add the coarse level correction to. |
| 262 |
*/ |
| 263 |
void moveToFineLevel(LevelContext& levelContext, bool processedFineLevel, |
| 264 |
Domain& fineX); |
| 265 |
|
| 266 |
/** |
| 267 |
* @brief Move the iterators to the coarser level. |
| 268 |
* @param levelContext The context with the iterators for the level. |
| 269 |
*/ |
| 270 |
bool moveToCoarseLevel(LevelContext& levelContext); |
| 271 |
|
| 272 |
/** |
| 273 |
* @brief Initialize iterators over levels with fine level. |
| 274 |
* @param levelContext The context with the iterators for the level. |
| 275 |
*/ |
| 276 |
void initIteratorsWithFineLevel(LevelContext& levelContext); |
| 277 |
|
| 278 |
/** @brief The matrix we solve. */ |
| 279 |
std::shared_ptr<OperatorHierarchy> matrices_; |
| 280 |
/** @brief The solver of the coarsest level. */ |
| 281 |
std::shared_ptr<CoarseSolver> solver_; |
| 282 |
/** @brief The right hand side of our problem. */ |
| 283 |
std::shared_ptr<Hierarchy<Range,A>> rhs_; |
| 284 |
/** @brief The left approximate solution of our problem. */ |
| 285 |
std::shared_ptr<Hierarchy<Domain,A>> lhs_; |
| 286 |
/** @brief The current residual. */ |
| 287 |
std::shared_ptr<Hierarchy<Domain,A>> residual_; |
| 288 |
|
| 289 |
/** @brief The type of the scalar product for the coarse solver. */ |
| 290 |
using ScalarProduct = Dune::ScalarProduct<X>; |
| 291 |
/** @brief Scalar product on the coarse level. */ |
| 292 |
std::shared_ptr<ScalarProduct> scalarProduct_; |
| 293 |
/** @brief Gamma, 1 for V-cycle and 2 for W-cycle. */ |
| 294 |
std::size_t gamma_; |
| 295 |
/** @brief The number of pre and postsmoothing steps. */ |
| 296 |
std::size_t preSteps_; |
| 297 |
/** @brief The number of postsmoothing steps. */ |
| 298 |
std::size_t postSteps_; |
| 299 |
std::size_t level; |
| 300 |
bool buildHierarchy_; |
| 301 |
bool symmetric; |
| 302 |
bool coarsesolverconverged; |
| 303 |
typedef SeqSSOR<typename M::matrix_type,X,X> Smoother; |
| 304 |
typedef std::shared_ptr<Smoother> SmootherPointer; |
| 305 |
SmootherPointer coarseSmoother_; |
| 306 |
/** @brief The verbosity level. */ |
| 307 |
std::size_t verbosity_; |
| 308 |
}; |
| 309 |
|
| 310 |
template<class M, class X, class PI, class A> |
| 311 |
FastAMG<M,X,PI,A>::FastAMG(const FastAMG& amg) |
| 312 |
: matrices_(amg.matrices_), solver_(amg.solver_), |
| 313 |
rhs_(), lhs_(), residual_(), scalarProduct_(amg.scalarProduct_), |
| 314 |
gamma_(amg.gamma_), preSteps_(amg.preSteps_), postSteps_(amg.postSteps_), |
| 315 |
symmetric(amg.symmetric), coarsesolverconverged(amg.coarsesolverconverged), |
| 316 |
coarseSmoother_(amg.coarseSmoother_), verbosity_(amg.verbosity_) |
| 317 |
{} |
| 318 |
|
| 319 |
template<class M, class X, class PI, class A> |
| 320 |
FastAMG<M,X,PI,A>::FastAMG(OperatorHierarchy& matrices, CoarseSolver& coarseSolver, |
| 321 |
const Parameters& parms, bool symmetric_) |
| 322 |
: matrices_(stackobject_to_shared_ptr(matrices)), solver_(&coarseSolver), |
| 323 |
rhs_(), lhs_(), residual_(), scalarProduct_(), |
| 324 |
gamma_(parms.getGamma()), preSteps_(parms.getNoPreSmoothSteps()), |
| 325 |
postSteps_(parms.getNoPostSmoothSteps()), buildHierarchy_(false), |
| 326 |
symmetric(symmetric_), coarsesolverconverged(true), |
| 327 |
coarseSmoother_(), verbosity_(parms.debugLevel()) |
| 328 |
{ |
| 329 |
if(preSteps_>1||postSteps_>1) |
| 330 |
{ |
| 331 |
std::cerr<<"WARNING only one step of smoothing is supported!"<<std::endl; |
| 332 |
preSteps_=postSteps_=0; |
| 333 |
} |
| 334 |
assert(matrices_->isBuilt()); |
| 335 |
static_assert(std::is_same<PI,SequentialInformation>::value, |
| 336 |
"Currently only sequential runs are supported"); |
| 337 |
} |
| 338 |
template<class M, class X, class PI, class A> |
| 339 |
template<class C> |
| 340 |
FastAMG<M,X,PI,A>::FastAMG(std::shared_ptr<const Operator> fineOperator, |
| 341 |
const C& criterion, |
| 342 |
const Parameters& parms, |
| 343 |
bool symmetric_, |
| 344 |
const PI& pinfo) |
| 345 |
: solver_(), rhs_(), lhs_(), residual_(), scalarProduct_(), gamma_(parms.getGamma()), |
| 346 |
preSteps_(parms.getNoPreSmoothSteps()), postSteps_(parms.getNoPostSmoothSteps()), |
| 347 |
buildHierarchy_(true), |
| 348 |
symmetric(symmetric_), coarsesolverconverged(true), |
| 349 |
coarseSmoother_(), verbosity_(criterion.debugLevel()) |
| 350 |
{ |
| 351 |
if(preSteps_>1||postSteps_>1) |
| 352 |
{ |
| 353 |
std::cerr<<"WARNING only one step of smoothing is supported!"<<std::endl; |
| 354 |
preSteps_=postSteps_=1; |
| 355 |
} |
| 356 |
static_assert(std::is_same<PI,SequentialInformation>::value, |
| 357 |
"Currently only sequential runs are supported"); |
| 358 |
// TODO: reestablish compile time checks. |
| 359 |
//static_assert(static_cast<int>(PI::category)==static_cast<int>(S::category), |
| 360 |
// "Matrix and Solver must match in terms of category!"); |
| 361 |
createHierarchies(criterion, std::move(fineOperator), pinfo); |
| 362 |
} |
| 363 |
|
| 364 |
template<class M, class X, class PI, class A> |
| 365 |
template<class C> |
| 366 |
void FastAMG<M,X,PI,A>::createHierarchies(C& criterion, |
| 367 |
std::shared_ptr<const Operator> fineOperator, |
| 368 |
const PI& pinfo) |
| 369 |
{ |
| 370 |
Timer watch; |
| 371 |
matrices_ = std::make_shared<OperatorHierarchy>( |
| 372 |
std::const_pointer_cast<Operator>(std::move(fineOperator)), |
| 373 |
stackobject_to_shared_ptr(const_cast<PI&>(pinfo))); |
| 374 |
|
| 375 |
matrices_->template build<NegateSet<typename PI::OwnerSet> >(criterion); |
| 376 |
|
| 377 |
if(verbosity_>0 && matrices_->parallelInformation().finest()->communicator().rank()==0) |
| 378 |
std::cout<<"Building Hierarchy of "<<matrices_->maxlevels()<<" levels took "<<watch.elapsed()<<" seconds."<<std::endl; |
| 379 |
|
| 380 |
if(buildHierarchy_ && matrices_->levels()==matrices_->maxlevels()) { |
| 381 |
// We have the carsest level. Create the coarse Solver |
| 382 |
typedef typename SmootherTraits<Smoother>::Arguments SmootherArgs; |
| 383 |
SmootherArgs sargs; |
| 384 |
sargs.iterations = 1; |
| 385 |
|
| 386 |
typename ConstructionTraits<Smoother>::Arguments cargs; |
| 387 |
cargs.setArgs(sargs); |
| 388 |
if(matrices_->redistributeInformation().back().isSetup()) { |
| 389 |
// Solve on the redistributed partitioning |
| 390 |
cargs.setMatrix(matrices_->matrices().coarsest().getRedistributed().getmat()); |
| 391 |
cargs.setComm(matrices_->parallelInformation().coarsest().getRedistributed()); |
| 392 |
}else{ |
| 393 |
cargs.setMatrix(matrices_->matrices().coarsest()->getmat()); |
| 394 |
cargs.setComm(*matrices_->parallelInformation().coarsest()); |
| 395 |
} |
| 396 |
|
| 397 |
coarseSmoother_ = ConstructionTraits<Smoother>::construct(cargs); |
| 398 |
scalarProduct_ = createScalarProduct<X>(cargs.getComm(),category()); |
| 399 |
|
| 400 |
#if HAVE_SUPERLU|| HAVE_SUITESPARSE_UMFPACK |
| 401 |
#if HAVE_SUITESPARSE_UMFPACK |
| 402 |
#define DIRECTSOLVER UMFPack |
| 403 |
#else |
| 404 |
#define DIRECTSOLVER SuperLU |
| 405 |
#endif |
| 406 |
// Use superlu if we are purely sequential or with only one processor on the coarsest level. |
| 407 |
if(std::is_same<ParallelInformation,SequentialInformation>::value // sequential mode |
| 408 |
|| matrices_->parallelInformation().coarsest()->communicator().size()==1 //parallel mode and only one processor |
| 409 |
|| (matrices_->parallelInformation().coarsest().isRedistributed() |
| 410 |
&& matrices_->parallelInformation().coarsest().getRedistributed().communicator().size()==1 |
| 411 |
&& matrices_->parallelInformation().coarsest().getRedistributed().communicator().size()>0)) { // redistribute and 1 proc |
| 412 |
if(verbosity_>0 && matrices_->parallelInformation().coarsest()->communicator().rank()==0) |
| 413 |
std::cout<<"Using superlu"<<std::endl; |
| 414 |
if(matrices_->parallelInformation().coarsest().isRedistributed()) |
| 415 |
{ |
| 416 |
if(matrices_->matrices().coarsest().getRedistributed().getmat().N()>0) |
| 417 |
// We are still participating on this level |
| 418 |
solver_.reset(new DIRECTSOLVER<typename M::matrix_type>(matrices_->matrices().coarsest().getRedistributed().getmat(), false, false)); |
| 419 |
else |
| 420 |
solver_.reset(); |
| 421 |
}else |
| 422 |
solver_.reset(new DIRECTSOLVER<typename M::matrix_type>(matrices_->matrices().coarsest()->getmat(), false, false)); |
| 423 |
}else |
| 424 |
#undef DIRECTSOLVER |
| 425 |
#endif // HAVE_SUPERLU|| HAVE_SUITESPARSE_UMFPACK |
| 426 |
{ |
| 427 |
if(matrices_->parallelInformation().coarsest().isRedistributed()) |
| 428 |
{ |
| 429 |
if(matrices_->matrices().coarsest().getRedistributed().getmat().N()>0) |
| 430 |
// We are still participating on this level |
| 431 |
solver_.reset(new BiCGSTABSolver<X>(const_cast<M&>(matrices_->matrices().coarsest().getRedistributed()), |
| 432 |
*scalarProduct_, |
| 433 |
*coarseSmoother_, 1E-2, 1000, 0)); |
| 434 |
else |
| 435 |
solver_.reset(); |
| 436 |
}else |
| 437 |
solver_.reset(new BiCGSTABSolver<X>(const_cast<M&>(*matrices_->matrices().coarsest()), |
| 438 |
*scalarProduct_, |
| 439 |
*coarseSmoother_, 1E-2, 1000, 0)); |
| 440 |
} |
| 441 |
} |
| 442 |
|
| 443 |
if(verbosity_>0 && matrices_->parallelInformation().finest()->communicator().rank()==0) |
| 444 |
std::cout<<"Building Hierarchy of "<<matrices_->maxlevels()<<" levels took "<<watch.elapsed()<<" seconds."<<std::endl; |
| 445 |
} |
| 446 |
|
| 447 |
|
| 448 |
template<class M, class X, class PI, class A> |
| 449 |
void FastAMG<M,X,PI,A>::pre(Domain& x, Range& b) |
| 450 |
{ |
| 451 |
Timer watch, watch1; |
| 452 |
// Detect Matrix rows where all offdiagonal entries are |
| 453 |
// zero and set x such that A_dd*x_d=b_d |
| 454 |
// Thus users can be more careless when setting up their linear |
| 455 |
// systems. |
| 456 |
typedef typename M::matrix_type Matrix; |
| 457 |
typedef typename Matrix::ConstRowIterator RowIter; |
| 458 |
typedef typename Matrix::ConstColIterator ColIter; |
| 459 |
typedef typename Matrix::block_type Block; |
| 460 |
Block zero; |
| 461 |
zero=typename Matrix::field_type(); |
| 462 |
|
| 463 |
const Matrix& mat=matrices_->matrices().finest()->getmat(); |
| 464 |
for(RowIter row=mat.begin(); row!=mat.end(); ++row) { |
| 465 |
bool isDirichlet = true; |
| 466 |
bool hasDiagonal = false; |
| 467 |
ColIter diag; |
| 468 |
for(ColIter col=row->begin(); col!=row->end(); ++col) { |
| 469 |
if(row.index()==col.index()) { |
| 470 |
diag = col; |
| 471 |
hasDiagonal = (*col != zero); |
| 472 |
}else{ |
| 473 |
if(*col!=zero) |
| 474 |
isDirichlet = false; |
| 475 |
} |
| 476 |
} |
| 477 |
if(isDirichlet && hasDiagonal) |
| 478 |
{ |
| 479 |
if constexpr (Dune::IsNumber<Block>::value) |
| 480 |
x[row.index()] = b[row.index()]/(*diag); |
| 481 |
else |
| 482 |
diag->solve(x[row.index()], b[row.index()]); |
| 483 |
} |
| 484 |
} |
| 485 |
if (verbosity_>0) |
| 486 |
std::cout<<" Preprocessing Dirichlet took "<<watch1.elapsed()<<std::endl; |
| 487 |
watch1.reset(); |
| 488 |
// No smoother to make x consistent! Do it by hand |
| 489 |
matrices_->parallelInformation().coarsest()->copyOwnerToAll(x,x); |
| 490 |
rhs_ = std::make_shared<Hierarchy<Range,A>>(std::make_shared<Range>(b)); |
| 491 |
lhs_ = std::make_shared<Hierarchy<Domain,A>>(std::make_shared<Domain>(x)); |
| 492 |
residual_ = std::make_shared<Hierarchy<Domain,A>>(std::make_shared<Domain>(x)); |
| 493 |
matrices_->coarsenVector(*rhs_); |
| 494 |
matrices_->coarsenVector(*lhs_); |
| 495 |
matrices_->coarsenVector(*residual_); |
| 496 |
|
| 497 |
// The preconditioner might change x and b. So we have to |
| 498 |
// copy the changes to the original vectors. |
| 499 |
x = *lhs_->finest(); |
| 500 |
b = *rhs_->finest(); |
| 501 |
} |
| 502 |
template<class M, class X, class PI, class A> |
| 503 |
std::size_t FastAMG<M,X,PI,A>::levels() |
| 504 |
{ |
| 505 |
return matrices_->levels(); |
| 506 |
} |
| 507 |
template<class M, class X, class PI, class A> |
| 508 |
std::size_t FastAMG<M,X,PI,A>::maxlevels() |
| 509 |
{ |
| 510 |
return matrices_->maxlevels(); |
| 511 |
} |
| 512 |
|
| 513 |
/** \copydoc Preconditioner::apply */ |
| 514 |
template<class M, class X, class PI, class A> |
| 515 |
void FastAMG<M,X,PI,A>::apply(Domain& v, const Range& d) |
| 516 |
{ |
| 517 |
LevelContext levelContext; |
| 518 |
// Init all iterators for the current level |
| 519 |
initIteratorsWithFineLevel(levelContext); |
| 520 |
|
| 521 |
assert(v.two_norm()==0); |
| 522 |
|
| 523 |
level=0; |
| 524 |
if(matrices_->maxlevels()==1){ |
| 525 |
// The coarse solver might modify the d! |
| 526 |
Range b(d); |
| 527 |
mgc(levelContext, v, b); |
| 528 |
}else |
| 529 |
mgc(levelContext, v, d); |
| 530 |
if(postSteps_==0||matrices_->maxlevels()==1) |
| 531 |
levelContext.pinfo->copyOwnerToAll(v, v); |
| 532 |
} |
| 533 |
|
| 534 |
template<class M, class X, class PI, class A> |
| 535 |
void FastAMG<M,X,PI,A>::initIteratorsWithFineLevel(LevelContext& levelContext) |
| 536 |
{ |
| 537 |
levelContext.matrix = matrices_->matrices().finest(); |
| 538 |
levelContext.pinfo = matrices_->parallelInformation().finest(); |
| 539 |
levelContext.redist = |
| 540 |
matrices_->redistributeInformation().begin(); |
| 541 |
levelContext.aggregates = matrices_->aggregatesMaps().begin(); |
| 542 |
levelContext.lhs = lhs_->finest(); |
| 543 |
levelContext.residual = residual_->finest(); |
| 544 |
levelContext.rhs = rhs_->finest(); |
| 545 |
levelContext.level=0; |
| 546 |
} |
| 547 |
|
| 548 |
template<class M, class X, class PI, class A> |
| 549 |
bool FastAMG<M,X,PI,A> |
| 550 |
::moveToCoarseLevel(LevelContext& levelContext) |
| 551 |
{ |
| 552 |
bool processNextLevel=true; |
| 553 |
|
| 554 |
if(levelContext.redist->isSetup()) { |
| 555 |
throw "bla"; |
| 556 |
levelContext.redist->redistribute(static_cast<const Range&>(*levelContext.residual), |
| 557 |
levelContext.residual.getRedistributed()); |
| 558 |
processNextLevel = levelContext.residual.getRedistributed().size()>0; |
| 559 |
if(processNextLevel) { |
| 560 |
//restrict defect to coarse level right hand side. |
| 561 |
++levelContext.pinfo; |
| 562 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 563 |
::restrictVector(*(*levelContext.aggregates), *levelContext.rhs, |
| 564 |
static_cast<const Range&>(levelContext.residual.getRedistributed()), |
| 565 |
*levelContext.pinfo); |
| 566 |
} |
| 567 |
}else{ |
| 568 |
//restrict defect to coarse level right hand side. |
| 569 |
++levelContext.rhs; |
| 570 |
++levelContext.pinfo; |
| 571 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 572 |
::restrictVector(*(*levelContext.aggregates), *levelContext.rhs, |
| 573 |
static_cast<const Range&>(*levelContext.residual), *levelContext.pinfo); |
| 574 |
} |
| 575 |
|
| 576 |
if(processNextLevel) { |
| 577 |
// prepare coarse system |
| 578 |
++levelContext.residual; |
| 579 |
++levelContext.lhs; |
| 580 |
++levelContext.matrix; |
| 581 |
++levelContext.level; |
| 582 |
++levelContext.redist; |
| 583 |
|
| 584 |
if(levelContext.matrix != matrices_->matrices().coarsest() || matrices_->levels()<matrices_->maxlevels()) { |
| 585 |
// next level is not the globally coarsest one |
| 586 |
++levelContext.aggregates; |
| 587 |
} |
| 588 |
// prepare the lhs on the next level |
| 589 |
*levelContext.lhs=0; |
| 590 |
*levelContext.residual=0; |
| 591 |
} |
| 592 |
return processNextLevel; |
| 593 |
} |
| 594 |
|
| 595 |
template<class M, class X, class PI, class A> |
| 596 |
void FastAMG<M,X,PI,A> |
| 597 |
::moveToFineLevel(LevelContext& levelContext, bool processNextLevel, Domain& x) |
| 598 |
{ |
| 599 |
if(processNextLevel) { |
| 600 |
if(levelContext.matrix != matrices_->matrices().coarsest() || matrices_->levels()<matrices_->maxlevels()) { |
| 601 |
// previous level is not the globally coarsest one |
| 602 |
--levelContext.aggregates; |
| 603 |
} |
| 604 |
--levelContext.redist; |
| 605 |
--levelContext.level; |
| 606 |
//prolongate and add the correction (update is in coarse left hand side) |
| 607 |
--levelContext.matrix; |
| 608 |
--levelContext.residual; |
| 609 |
|
| 610 |
} |
| 611 |
|
| 612 |
typename Hierarchy<Domain,A>::Iterator coarseLhs = levelContext.lhs--; |
| 613 |
if(levelContext.redist->isSetup()) { |
| 614 |
|
| 615 |
// Need to redistribute during prolongate |
| 616 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 617 |
::prolongateVector(*(*levelContext.aggregates), *coarseLhs, x, |
| 618 |
levelContext.lhs.getRedistributed(), |
| 619 |
matrices_->getProlongationDampingFactor(), |
| 620 |
*levelContext.pinfo, *levelContext.redist); |
| 621 |
}else{ |
| 622 |
Transfer<typename OperatorHierarchy::AggregatesMap::AggregateDescriptor,Range,ParallelInformation> |
| 623 |
::prolongateVector(*(*levelContext.aggregates), *coarseLhs, x, |
| 624 |
matrices_->getProlongationDampingFactor(), *levelContext.pinfo); |
| 625 |
|
| 626 |
// printvector(std::cout, *lhs, "prolongated coarse grid correction", "lhs", 10, 10, 10); |
| 627 |
} |
| 628 |
|
| 629 |
|
| 630 |
if(processNextLevel) { |
| 631 |
--levelContext.rhs; |
| 632 |
} |
| 633 |
|
| 634 |
} |
| 635 |
|
| 636 |
|
| 637 |
template<class M, class X, class PI, class A> |
| 638 |
void FastAMG<M,X,PI,A> |
| 639 |
::presmooth(LevelContext& levelContext, Domain& x, const Range& b) |
| 640 |
{ |
| 641 |
constexpr auto bl = blockLevel<typename M::matrix_type>(); |
| 642 |
GaussSeidelPresmoothDefect<bl>::apply(levelContext.matrix->getmat(), |
| 643 |
x, |
| 644 |
*levelContext.residual, |
| 645 |
b); |
| 646 |
} |
| 647 |
|
| 648 |
template<class M, class X, class PI, class A> |
| 649 |
void FastAMG<M,X,PI,A> |
| 650 |
::postsmooth(LevelContext& levelContext, Domain& x, const Range& b) |
| 651 |
{ |
| 652 |
constexpr auto bl = blockLevel<typename M::matrix_type>(); |
| 653 |
GaussSeidelPostsmoothDefect<bl> |
| 654 |
::apply(levelContext.matrix->getmat(), x, *levelContext.residual, b); |
| 655 |
} |
| 656 |
|
| 657 |
|
| 658 |
template<class M, class X, class PI, class A> |
| 659 |
bool FastAMG<M,X,PI,A>::usesDirectCoarseLevelSolver() const |
| 660 |
{ |
| 661 |
return IsDirectSolver< CoarseSolver>::value; |
| 662 |
} |
| 663 |
|
| 664 |
template<class M, class X, class PI, class A> |
| 665 |
void FastAMG<M,X,PI,A>::mgc(LevelContext& levelContext, Domain& v, const Range& b){ |
| 666 |
|
| 667 |
if(levelContext.matrix == matrices_->matrices().coarsest() && levels()==maxlevels()) { |
| 668 |
// Solve directly |
| 669 |
InverseOperatorResult res; |
| 670 |
res.converged=true; // If we do not compute this flag will not get updated |
| 671 |
if(levelContext.redist->isSetup()) { |
| 672 |
levelContext.redist->redistribute(b, levelContext.rhs.getRedistributed()); |
| 673 |
if(levelContext.rhs.getRedistributed().size()>0) { |
| 674 |
// We are still participating in the computation |
| 675 |
levelContext.pinfo.getRedistributed().copyOwnerToAll(levelContext.rhs.getRedistributed(), |
| 676 |
levelContext.rhs.getRedistributed()); |
| 677 |
solver_->apply(levelContext.lhs.getRedistributed(), levelContext.rhs.getRedistributed(), res); |
| 678 |
} |
| 679 |
levelContext.redist->redistributeBackward(v, levelContext.lhs.getRedistributed()); |
| 680 |
levelContext.pinfo->copyOwnerToAll(v, v); |
| 681 |
}else{ |
| 682 |
levelContext.pinfo->copyOwnerToAll(b, b); |
| 683 |
solver_->apply(v, const_cast<Range&>(b), res); |
| 684 |
} |
| 685 |
|
| 686 |
// printvector(std::cout, *lhs, "coarse level update", "u", 10, 10, 10); |
| 687 |
// printvector(std::cout, *rhs, "coarse level rhs", "rhs", 10, 10, 10); |
| 688 |
if (!res.converged) |
| 689 |
coarsesolverconverged = false; |
| 690 |
}else{ |
| 691 |
// presmoothing |
| 692 |
presmooth(levelContext, v, b); |
| 693 |
// printvector(std::cout, *lhs, "update", "u", 10, 10, 10); |
| 694 |
// printvector(std::cout, *residual, "post presmooth residual", "r", 10); |
| 695 |
#ifndef DUNE_AMG_NO_COARSEGRIDCORRECTION |
| 696 |
bool processNextLevel = moveToCoarseLevel(levelContext); |
| 697 |
|
| 698 |
if(processNextLevel) { |
| 699 |
// next level |
| 700 |
for(std::size_t i=0; i<gamma_; i++) |
| 701 |
mgc(levelContext, *levelContext.lhs, *levelContext.rhs); |
| 702 |
} |
| 703 |
|
| 704 |
moveToFineLevel(levelContext, processNextLevel, v); |
| 705 |
#else |
| 706 |
*lhs=0; |
| 707 |
#endif |
| 708 |
|
| 709 |
if(levelContext.matrix == matrices_->matrices().finest()) { |
| 710 |
coarsesolverconverged = matrices_->parallelInformation().finest()->communicator().prod(coarsesolverconverged); |
| 711 |
if(!coarsesolverconverged) |
| 712 |
DUNE_THROW(MathError, "Coarse solver did not converge"); |
| 713 |
} |
| 714 |
|
| 715 |
postsmooth(levelContext, v, b); |
| 716 |
} |
| 717 |
} |
| 718 |
|
| 719 |
|
| 720 |
/** \copydoc Preconditioner::post */ |
| 721 |
template<class M, class X, class PI, class A> |
| 722 |
void FastAMG<M,X,PI,A>::post([[maybe_unused]] Domain& x) |
| 723 |
{ |
| 724 |
lhs_=nullptr; |
| 725 |
rhs_=nullptr; |
| 726 |
residual_=nullptr; |
| 727 |
} |
| 728 |
|
| 729 |
template<class M, class X, class PI, class A> |
| 730 |
template<class A1> |
| 731 |
void FastAMG<M,X,PI,A>::getCoarsestAggregateNumbers(std::vector<std::size_t,A1>& cont) |
| 732 |
{ |
| 733 |
matrices_->getCoarsestAggregatesOnFinest(cont); |
| 734 |
} |
| 735 |
|
| 736 |
} // end namespace Amg |
| 737 |
} // end namespace Dune |
| 738 |
|
| 739 |
#endif |