| 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 |
|
| 6 |
#ifndef DUNE_ISTL_SOLVERS_HH |
| 7 |
#define DUNE_ISTL_SOLVERS_HH |
| 8 |
|
| 9 |
#include <array> |
| 10 |
#include <cmath> |
| 11 |
#include <complex> |
| 12 |
#include <iostream> |
| 13 |
#include <memory> |
| 14 |
#include <type_traits> |
| 15 |
#include <vector> |
| 16 |
|
| 17 |
#include <dune/common/exceptions.hh> |
| 18 |
#include <dune/common/math.hh> |
| 19 |
#include <dune/common/simd/io.hh> |
| 20 |
#include <dune/common/simd/simd.hh> |
| 21 |
#include <dune/common/std/type_traits.hh> |
| 22 |
#include <dune/common/timer.hh> |
| 23 |
|
| 24 |
#include <dune/istl/allocator.hh> |
| 25 |
#include <dune/istl/bcrsmatrix.hh> |
| 26 |
#include <dune/istl/eigenvalue/arpackpp.hh> |
| 27 |
#include <dune/istl/istlexception.hh> |
| 28 |
#include <dune/istl/operators.hh> |
| 29 |
#include <dune/istl/preconditioner.hh> |
| 30 |
#include <dune/istl/scalarproducts.hh> |
| 31 |
#include <dune/istl/solver.hh> |
| 32 |
#include <dune/istl/solverregistry.hh> |
| 33 |
|
| 34 |
namespace Dune { |
| 35 |
/** @addtogroup ISTL_Solvers |
| 36 |
@{ |
| 37 |
*/ |
| 38 |
|
| 39 |
/** \file |
| 40 |
|
| 41 |
\brief Implementations of the inverse operator interface. |
| 42 |
|
| 43 |
This file provides various preconditioned Krylov methods. |
| 44 |
*/ |
| 45 |
|
| 46 |
//===================================================================== |
| 47 |
// Implementation of this interface |
| 48 |
//===================================================================== |
| 49 |
|
| 50 |
/*! |
| 51 |
\brief Preconditioned loop solver. |
| 52 |
|
| 53 |
Implements a preconditioned loop. |
| 54 |
Using this class every Preconditioner can be turned |
| 55 |
into a solver. The solver will apply one preconditioner |
| 56 |
step in each iteration loop. |
| 57 |
*/ |
| 58 |
template<class X> |
| 59 |
class LoopSolver : public IterativeSolver<X,X> { |
| 60 |
public: |
| 61 |
using typename IterativeSolver<X,X>::domain_type; |
| 62 |
using typename IterativeSolver<X,X>::range_type; |
| 63 |
using typename IterativeSolver<X,X>::field_type; |
| 64 |
using typename IterativeSolver<X,X>::real_type; |
| 65 |
|
| 66 |
// copy base class constructors |
| 67 |
using IterativeSolver<X,X>::IterativeSolver; |
| 68 |
|
| 69 |
// don't shadow four-argument version of apply defined in the base class |
| 70 |
using IterativeSolver<X,X>::apply; |
| 71 |
|
| 72 |
//! \copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 73 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 74 |
{ |
| 75 |
Iteration iteration(*this, res); |
| 76 |
_prec->pre(x,b); |
| 77 |
|
| 78 |
// overwrite b with defect |
| 79 |
_op->applyscaleadd(-1,x,b); |
| 80 |
|
| 81 |
// compute norm, \todo parallelization |
| 82 |
real_type def = _sp->norm(b); |
| 83 |
if(iteration.step(0, def)){ |
| 84 |
_prec->post(x); |
| 85 |
return; |
| 86 |
} |
| 87 |
// prepare preconditioner |
| 88 |
|
| 89 |
// allocate correction vector |
| 90 |
X v(x); |
| 91 |
|
| 92 |
// iteration loop |
| 93 |
int i=1; |
| 94 |
for ( ; i<=_maxit; i++ ) |
| 95 |
{ |
| 96 |
v = 0; // clear correction |
| 97 |
_prec->apply(v,b); // apply preconditioner |
| 98 |
x += v; // update solution |
| 99 |
_op->applyscaleadd(-1,v,b); // update defect |
| 100 |
def=_sp->norm(b); // comp defect norm |
| 101 |
if(iteration.step(i, def)) |
| 102 |
break; |
| 103 |
} |
| 104 |
|
| 105 |
// postprocess preconditioner |
| 106 |
_prec->post(x); |
| 107 |
} |
| 108 |
|
| 109 |
protected: |
| 110 |
using IterativeSolver<X,X>::_op; |
| 111 |
using IterativeSolver<X,X>::_prec; |
| 112 |
using IterativeSolver<X,X>::_sp; |
| 113 |
using IterativeSolver<X,X>::_reduction; |
| 114 |
using IterativeSolver<X,X>::_maxit; |
| 115 |
using IterativeSolver<X,X>::_verbose; |
| 116 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 117 |
}; |
| 118 |
DUNE_REGISTER_SOLVER("loopsolver", defaultIterativeSolverCreator<Dune::LoopSolver>()); |
| 119 |
|
| 120 |
|
| 121 |
// all these solvers are taken from the SUMO library |
| 122 |
//! gradient method |
| 123 |
template<class X> |
| 124 |
class GradientSolver : public IterativeSolver<X,X> { |
| 125 |
public: |
| 126 |
using typename IterativeSolver<X,X>::domain_type; |
| 127 |
using typename IterativeSolver<X,X>::range_type; |
| 128 |
using typename IterativeSolver<X,X>::field_type; |
| 129 |
using typename IterativeSolver<X,X>::real_type; |
| 130 |
|
| 131 |
// copy base class constructors |
| 132 |
using IterativeSolver<X,X>::IterativeSolver; |
| 133 |
|
| 134 |
// don't shadow four-argument version of apply defined in the base class |
| 135 |
using IterativeSolver<X,X>::apply; |
| 136 |
|
| 137 |
/*! |
| 138 |
\brief Apply inverse operator. |
| 139 |
|
| 140 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 141 |
*/ |
| 142 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 143 |
{ |
| 144 |
Iteration iteration(*this, res); |
| 145 |
_prec->pre(x,b); // prepare preconditioner |
| 146 |
|
| 147 |
_op->applyscaleadd(-1,x,b); // overwrite b with defec |
| 148 |
|
| 149 |
real_type def = _sp->norm(b); // compute norm |
| 150 |
if(iteration.step(0, def)){ |
| 151 |
_prec->post(x); |
| 152 |
return; |
| 153 |
} |
| 154 |
|
| 155 |
X p(x); // create local vectors |
| 156 |
X q(b); |
| 157 |
|
| 158 |
int i=1; // loop variables |
| 159 |
field_type lambda; |
| 160 |
for ( ; i<=_maxit; i++ ) |
| 161 |
{ |
| 162 |
p = 0; // clear correction |
| 163 |
_prec->apply(p,b); // apply preconditioner |
| 164 |
_op->apply(p,q); // q=Ap |
| 165 |
auto alpha = _sp->dot(q,p); |
| 166 |
lambda = Simd::cond(def==field_type(0.), |
| 167 |
field_type(0.), // no need for minimization if def is already 0 |
| 168 |
field_type(_sp->dot(p,b)/alpha)); // minimization |
| 169 |
x.axpy(lambda,p); // update solution |
| 170 |
b.axpy(-lambda,q); // update defect |
| 171 |
|
| 172 |
def =_sp->norm(b); // comp defect norm |
| 173 |
if(iteration.step(i, def)) |
| 174 |
break; |
| 175 |
} |
| 176 |
// postprocess preconditioner |
| 177 |
_prec->post(x); |
| 178 |
} |
| 179 |
|
| 180 |
protected: |
| 181 |
using IterativeSolver<X,X>::_op; |
| 182 |
using IterativeSolver<X,X>::_prec; |
| 183 |
using IterativeSolver<X,X>::_sp; |
| 184 |
using IterativeSolver<X,X>::_reduction; |
| 185 |
using IterativeSolver<X,X>::_maxit; |
| 186 |
using IterativeSolver<X,X>::_verbose; |
| 187 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 188 |
}; |
| 189 |
DUNE_REGISTER_SOLVER("gradientsolver", defaultIterativeSolverCreator<Dune::GradientSolver>()); |
| 190 |
|
| 191 |
//! \brief conjugate gradient method |
| 192 |
template<class X> |
| 193 |
class CGSolver : public IterativeSolver<X,X> { |
| 194 |
public: |
| 195 |
using typename IterativeSolver<X,X>::domain_type; |
| 196 |
using typename IterativeSolver<X,X>::range_type; |
| 197 |
using typename IterativeSolver<X,X>::field_type; |
| 198 |
using typename IterativeSolver<X,X>::real_type; |
| 199 |
|
| 200 |
// copy base class constructors |
| 201 |
using IterativeSolver<X,X>::IterativeSolver; |
| 202 |
|
| 203 |
private: |
| 204 |
using typename IterativeSolver<X,X>::scalar_real_type; |
| 205 |
|
| 206 |
protected: |
| 207 |
|
| 208 |
static constexpr bool enableConditionEstimate = (std::is_same_v<field_type,float> || std::is_same_v<field_type,double>); |
| 209 |
|
| 210 |
public: |
| 211 |
|
| 212 |
// don't shadow four-argument version of apply defined in the base class |
| 213 |
using IterativeSolver<X,X>::apply; |
| 214 |
|
| 215 |
/*! |
| 216 |
\brief Constructor to initialize a CG solver. |
| 217 |
\copydetails IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&, Preconditioner<X,Y>&, scalar_real_type, int, int) |
| 218 |
\param condition_estimate Whether to calculate an estimate of the condition number. |
| 219 |
The estimate is given in the InverseOperatorResult returned by apply(). |
| 220 |
This is only supported for float and double field types. |
| 221 |
*/ |
| 222 |
CGSolver (const LinearOperator<X,X>& op, Preconditioner<X,X>& prec, |
| 223 |
scalar_real_type reduction, int maxit, int verbose, bool condition_estimate) : IterativeSolver<X,X>(op, prec, reduction, maxit, verbose), |
| 224 |
condition_estimate_(condition_estimate) |
| 225 |
{ |
| 226 |
if (condition_estimate && !enableConditionEstimate) { |
| 227 |
condition_estimate_ = false; |
| 228 |
std::cerr << "WARNING: Condition estimate was disabled. It is only available for double and float field types!" << std::endl; |
| 229 |
} |
| 230 |
} |
| 231 |
|
| 232 |
/*! |
| 233 |
\brief Constructor to initialize a CG solver. |
| 234 |
\copydetails IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&, const ScalarProduct<X>&, Preconditioner<X,Y>&, scalar_real_type, int, int) |
| 235 |
\param condition_estimate Whether to calculate an estimate of the condition number. |
| 236 |
The estimate is given in the InverseOperatorResult returned by apply(). |
| 237 |
This is only supported for float and double field types. |
| 238 |
*/ |
| 239 |
CGSolver (const LinearOperator<X,X>& op, const ScalarProduct<X>& sp, Preconditioner<X,X>& prec, |
| 240 |
scalar_real_type reduction, int maxit, int verbose, bool condition_estimate) : IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose), |
| 241 |
condition_estimate_(condition_estimate) |
| 242 |
{ |
| 243 |
if (condition_estimate && !(std::is_same<field_type,float>::value || std::is_same<field_type,double>::value)) { |
| 244 |
condition_estimate_ = false; |
| 245 |
std::cerr << "WARNING: Condition estimate was disabled. It is only available for double and float field types!" << std::endl; |
| 246 |
} |
| 247 |
} |
| 248 |
|
| 249 |
/*! |
| 250 |
\brief Constructor to initialize a CG solver. |
| 251 |
|
| 252 |
\copydetails IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>,std::shared_ptr<const ScalarProduct<X>>,std::shared_ptr<Preconditioner<X,Y>>,scalar_real_type,int,int) |
| 253 |
\param condition_estimate Whether to calculate an estimate of the condition number. |
| 254 |
The estimate is given in the InverseOperatorResult returned by apply(). |
| 255 |
This is only supported for float and double field types. |
| 256 |
*/ |
| 257 |
CGSolver (std::shared_ptr<const LinearOperator<X,X>> op, std::shared_ptr<ScalarProduct<X>> sp, |
| 258 |
std::shared_ptr<Preconditioner<X,X>> prec, |
| 259 |
scalar_real_type reduction, int maxit, int verbose, bool condition_estimate) |
| 260 |
: IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose), |
| 261 |
condition_estimate_(condition_estimate) |
| 262 |
{ |
| 263 |
if (condition_estimate && !(std::is_same<field_type,float>::value || std::is_same<field_type,double>::value)) { |
| 264 |
condition_estimate_ = false; |
| 265 |
std::cerr << "WARNING: Condition estimate was disabled. It is only available for double and float field types!" << std::endl; |
| 266 |
} |
| 267 |
} |
| 268 |
|
| 269 |
/*! |
| 270 |
\brief Apply inverse operator. |
| 271 |
|
| 272 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 273 |
|
| 274 |
\note Currently, the CGSolver aborts when a NaN or infinite defect is |
| 275 |
detected. However, -ffinite-math-only (implied by -ffast-math) |
| 276 |
can inhibit a result from becoming NaN that really should be NaN. |
| 277 |
E.g. numeric_limits<double>::quiet_NaN()*0.0==0.0 with gcc-5.3 |
| 278 |
-ffast-math. |
| 279 |
*/ |
| 280 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 281 |
{ |
| 282 |
Iteration iteration(*this,res); |
| 283 |
_prec->pre(x,b); // prepare preconditioner |
| 284 |
|
| 285 |
_op->applyscaleadd(-1,x,b); // overwrite b with defect |
| 286 |
|
| 287 |
real_type def = _sp->norm(b); // compute norm |
| 288 |
if(iteration.step(0, def)){ |
| 289 |
_prec->post(x); |
| 290 |
return; |
| 291 |
} |
| 292 |
|
| 293 |
X p(x); // the search direction |
| 294 |
X q(x); // a temporary vector |
| 295 |
|
| 296 |
// Remember lambda and beta values for condition estimate |
| 297 |
std::vector<real_type> lambdas(0); |
| 298 |
std::vector<real_type> betas(0); |
| 299 |
|
| 300 |
// some local variables |
| 301 |
field_type rho,rholast,lambda,alpha,beta; |
| 302 |
|
| 303 |
// determine initial search direction |
| 304 |
p = 0; // clear correction |
| 305 |
_prec->apply(p,b); // apply preconditioner |
| 306 |
rholast = _sp->dot(p,b); // orthogonalization |
| 307 |
|
| 308 |
// the loop |
| 309 |
int i=1; |
| 310 |
for ( ; i<=_maxit; i++ ) |
| 311 |
{ |
| 312 |
// minimize in given search direction p |
| 313 |
_op->apply(p,q); // q=Ap |
| 314 |
alpha = _sp->dot(p,q); // scalar product |
| 315 |
lambda = Simd::cond(def==field_type(0.), field_type(0.), field_type(rholast/alpha)); // minimization |
| 316 |
if constexpr (enableConditionEstimate) |
| 317 |
if (condition_estimate_) |
| 318 |
lambdas.push_back(std::real(lambda)); |
| 319 |
x.axpy(lambda,p); // update solution |
| 320 |
b.axpy(-lambda,q); // update defect |
| 321 |
|
| 322 |
// convergence test |
| 323 |
def=_sp->norm(b); // comp defect norm |
| 324 |
if(iteration.step(i, def)) |
| 325 |
break; |
| 326 |
|
| 327 |
// determine new search direction |
| 328 |
q = 0; // clear correction |
| 329 |
_prec->apply(q,b); // apply preconditioner |
| 330 |
rho = _sp->dot(q,b); // orthogonalization |
| 331 |
beta = Simd::cond(def==field_type(0.), field_type(0.), field_type(rho/rholast)); // scaling factor |
| 332 |
if constexpr (enableConditionEstimate) |
| 333 |
if (condition_estimate_) |
| 334 |
betas.push_back(std::real(beta)); |
| 335 |
p *= beta; // scale old search direction |
| 336 |
p += q; // orthogonalization with correction |
| 337 |
rholast = rho; // remember rho for recurrence |
| 338 |
} |
| 339 |
|
| 340 |
_prec->post(x); // postprocess preconditioner |
| 341 |
|
| 342 |
if (condition_estimate_) { |
| 343 |
#if HAVE_ARPACKPP |
| 344 |
if constexpr (enableConditionEstimate) { |
| 345 |
using std::sqrt; |
| 346 |
|
| 347 |
// Build T matrix which has extreme eigenvalues approximating |
| 348 |
// those of the original system |
| 349 |
// (see Y. Saad, Iterative methods for sparse linear systems) |
| 350 |
|
| 351 |
COND_MAT T(i, i, COND_MAT::row_wise); |
| 352 |
|
| 353 |
for (auto row = T.createbegin(); row != T.createend(); ++row) { |
| 354 |
if (row.index() > 0) |
| 355 |
row.insert(row.index()-1); |
| 356 |
row.insert(row.index()); |
| 357 |
if (row.index() < T.N() - 1) |
| 358 |
row.insert(row.index()+1); |
| 359 |
} |
| 360 |
for (int row = 0; row < i; ++row) { |
| 361 |
if (row > 0) { |
| 362 |
T[row][row-1] = sqrt(betas[row-1]) / lambdas[row-1]; |
| 363 |
} |
| 364 |
|
| 365 |
T[row][row] = 1.0 / lambdas[row]; |
| 366 |
if (row > 0) { |
| 367 |
T[row][row] += betas[row-1] / lambdas[row-1]; |
| 368 |
} |
| 369 |
|
| 370 |
if (row < i - 1) { |
| 371 |
T[row][row+1] = sqrt(betas[row]) / lambdas[row]; |
| 372 |
} |
| 373 |
} |
| 374 |
|
| 375 |
// Compute largest and smallest eigenvalue of T matrix and return as estimate |
| 376 |
Dune::ArPackPlusPlus_Algorithms<COND_MAT, COND_VEC> arpack(T); |
| 377 |
|
| 378 |
real_type eps = 0.0; |
| 379 |
COND_VEC eigv; |
| 380 |
real_type min_eigv, max_eigv; |
| 381 |
arpack.computeSymMinMagnitude (eps, eigv, min_eigv); |
| 382 |
arpack.computeSymMaxMagnitude (eps, eigv, max_eigv); |
| 383 |
|
| 384 |
res.condition_estimate = max_eigv / min_eigv; |
| 385 |
|
| 386 |
if (this->_verbose > 0) { |
| 387 |
std::cout << "Min eigv estimate: " << Simd::io(min_eigv) << '\n'; |
| 388 |
std::cout << "Max eigv estimate: " << Simd::io(max_eigv) << '\n'; |
| 389 |
std::cout << "Condition estimate: " |
| 390 |
<< Simd::io(max_eigv / min_eigv) << std::endl; |
| 391 |
} |
| 392 |
} |
| 393 |
#else |
| 394 |
std::cerr << "WARNING: Condition estimate was requested. This requires ARPACK, but ARPACK was not found!" << std::endl; |
| 395 |
#endif |
| 396 |
} |
| 397 |
} |
| 398 |
|
| 399 |
private: |
| 400 |
bool condition_estimate_ = false; |
| 401 |
|
| 402 |
// Matrix and vector types used for condition estimate |
| 403 |
typedef Dune::BCRSMatrix<Dune::FieldMatrix<real_type,1,1> > COND_MAT; |
| 404 |
typedef Dune::BlockVector<Dune::FieldVector<real_type,1> > COND_VEC; |
| 405 |
|
| 406 |
protected: |
| 407 |
using IterativeSolver<X,X>::_op; |
| 408 |
using IterativeSolver<X,X>::_prec; |
| 409 |
using IterativeSolver<X,X>::_sp; |
| 410 |
using IterativeSolver<X,X>::_reduction; |
| 411 |
using IterativeSolver<X,X>::_maxit; |
| 412 |
using IterativeSolver<X,X>::_verbose; |
| 413 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 414 |
}; |
| 415 |
DUNE_REGISTER_SOLVER("cgsolver", defaultIterativeSolverCreator<Dune::CGSolver>()); |
| 416 |
|
| 417 |
// Ronald Kriemanns BiCG-STAB implementation from Sumo |
| 418 |
//! \brief Bi-conjugate Gradient Stabilized (BiCG-STAB) |
| 419 |
template<class X> |
| 420 |
class BiCGSTABSolver : public IterativeSolver<X,X> { |
| 421 |
public: |
| 422 |
using typename IterativeSolver<X,X>::domain_type; |
| 423 |
using typename IterativeSolver<X,X>::range_type; |
| 424 |
using typename IterativeSolver<X,X>::field_type; |
| 425 |
using typename IterativeSolver<X,X>::real_type; |
| 426 |
|
| 427 |
// copy base class constructors |
| 428 |
using IterativeSolver<X,X>::IterativeSolver; |
| 429 |
|
| 430 |
// don't shadow four-argument version of apply defined in the base class |
| 431 |
using IterativeSolver<X,X>::apply; |
| 432 |
|
| 433 |
/*! |
| 434 |
\brief Apply inverse operator. |
| 435 |
|
| 436 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 437 |
|
| 438 |
\note Currently, the BiCGSTABSolver aborts when it detects a breakdown. |
| 439 |
*/ |
| 440 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 441 |
{ |
| 442 |
using std::abs; |
| 443 |
const Simd::Scalar<real_type> EPSILON=1e-80; |
| 444 |
using std::abs; |
| 445 |
double it; |
| 446 |
field_type rho, rho_new, alpha, beta, h, omega; |
| 447 |
real_type norm; |
| 448 |
|
| 449 |
// |
| 450 |
// get vectors and matrix |
| 451 |
// |
| 452 |
X& r=b; |
| 453 |
X p(x); |
| 454 |
X v(x); |
| 455 |
X t(x); |
| 456 |
X y(x); |
| 457 |
X rt(x); |
| 458 |
|
| 459 |
// |
| 460 |
// begin iteration |
| 461 |
// |
| 462 |
|
| 463 |
// r = r - Ax; rt = r |
| 464 |
Iteration<double> iteration(*this,res); |
| 465 |
_prec->pre(x,r); // prepare preconditioner |
| 466 |
|
| 467 |
_op->applyscaleadd(-1,x,r); // overwrite b with defect |
| 468 |
|
| 469 |
rt=r; |
| 470 |
|
| 471 |
norm = _sp->norm(r); |
| 472 |
if(iteration.step(0, norm)){ |
| 473 |
_prec->post(x); |
| 474 |
return; |
| 475 |
} |
| 476 |
p=0; |
| 477 |
v=0; |
| 478 |
|
| 479 |
rho = 1; |
| 480 |
alpha = 1; |
| 481 |
omega = 1; |
| 482 |
|
| 483 |
// |
| 484 |
// iteration |
| 485 |
// |
| 486 |
|
| 487 |
for (it = 0.5; it < _maxit; it+=.5) |
| 488 |
{ |
| 489 |
// |
| 490 |
// preprocess, set vecsizes etc. |
| 491 |
// |
| 492 |
|
| 493 |
// rho_new = < rt , r > |
| 494 |
rho_new = _sp->dot(rt,r); |
| 495 |
|
| 496 |
// look if breakdown occurred |
| 497 |
if (Simd::allTrue(abs(rho) <= EPSILON)) |
| 498 |
DUNE_THROW(SolverAbort,"breakdown in BiCGSTAB - rho " |
| 499 |
<< Simd::io(rho) << " <= EPSILON " << EPSILON |
| 500 |
<< " after " << it << " iterations"); |
| 501 |
if (Simd::allTrue(abs(omega) <= EPSILON)) |
| 502 |
DUNE_THROW(SolverAbort,"breakdown in BiCGSTAB - omega " |
| 503 |
<< Simd::io(omega) << " <= EPSILON " << EPSILON |
| 504 |
<< " after " << it << " iterations"); |
| 505 |
|
| 506 |
|
| 507 |
if (it<1) |
| 508 |
p = r; |
| 509 |
else |
| 510 |
{ |
| 511 |
beta = Simd::cond(norm==field_type(0.), |
| 512 |
field_type(0.), // no need for orthogonalization if norm is already 0 |
| 513 |
field_type(( rho_new / rho ) * ( alpha / omega ))); |
| 514 |
p.axpy(-omega,v); // p = r + beta (p - omega*v) |
| 515 |
p *= beta; |
| 516 |
p += r; |
| 517 |
} |
| 518 |
|
| 519 |
// y = W^-1 * p |
| 520 |
y = 0; |
| 521 |
_prec->apply(y,p); // apply preconditioner |
| 522 |
|
| 523 |
// v = A * y |
| 524 |
_op->apply(y,v); |
| 525 |
|
| 526 |
// alpha = rho_new / < rt, v > |
| 527 |
h = _sp->dot(rt,v); |
| 528 |
|
| 529 |
if ( Simd::allTrue(abs(h) < EPSILON) ) |
| 530 |
DUNE_THROW(SolverAbort,"abs(h) < EPSILON in BiCGSTAB - abs(h) " |
| 531 |
<< Simd::io(abs(h)) << " < EPSILON " << EPSILON |
| 532 |
<< " after " << it << " iterations"); |
| 533 |
|
| 534 |
alpha = Simd::cond(norm==field_type(0.), |
| 535 |
field_type(0.), |
| 536 |
field_type(rho_new / h)); |
| 537 |
|
| 538 |
// apply first correction to x |
| 539 |
// x <- x + alpha y |
| 540 |
x.axpy(alpha,y); |
| 541 |
|
| 542 |
// r = r - alpha*v |
| 543 |
r.axpy(-alpha,v); |
| 544 |
|
| 545 |
// |
| 546 |
// test stop criteria |
| 547 |
// |
| 548 |
|
| 549 |
norm = _sp->norm(r); |
| 550 |
if(iteration.step(it, norm)){ |
| 551 |
break; |
| 552 |
} |
| 553 |
|
| 554 |
it+=.5; |
| 555 |
|
| 556 |
// y = W^-1 * r |
| 557 |
y = 0; |
| 558 |
_prec->apply(y,r); |
| 559 |
|
| 560 |
// t = A * y |
| 561 |
_op->apply(y,t); |
| 562 |
|
| 563 |
// omega = < t, r > / < t, t > |
| 564 |
h = _sp->dot(t,t); |
| 565 |
omega = Simd::cond(norm==field_type(0.), |
| 566 |
field_type(0.), |
| 567 |
field_type(_sp->dot(t,r)/h)); |
| 568 |
|
| 569 |
// apply second correction to x |
| 570 |
// x <- x + omega y |
| 571 |
x.axpy(omega,y); |
| 572 |
|
| 573 |
// r = s - omega*t (remember : r = s) |
| 574 |
r.axpy(-omega,t); |
| 575 |
|
| 576 |
rho = rho_new; |
| 577 |
|
| 578 |
// |
| 579 |
// test stop criteria |
| 580 |
// |
| 581 |
|
| 582 |
norm = _sp->norm(r); |
| 583 |
if(iteration.step(it, norm)){ |
| 584 |
break; |
| 585 |
} |
| 586 |
} // end for |
| 587 |
|
| 588 |
_prec->post(x); // postprocess preconditioner |
| 589 |
} |
| 590 |
|
| 591 |
protected: |
| 592 |
using IterativeSolver<X,X>::_op; |
| 593 |
using IterativeSolver<X,X>::_prec; |
| 594 |
using IterativeSolver<X,X>::_sp; |
| 595 |
using IterativeSolver<X,X>::_reduction; |
| 596 |
using IterativeSolver<X,X>::_maxit; |
| 597 |
using IterativeSolver<X,X>::_verbose; |
| 598 |
template<class CountType> |
| 599 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<CountType>; |
| 600 |
}; |
| 601 |
DUNE_REGISTER_SOLVER("bicgstabsolver", defaultIterativeSolverCreator<Dune::BiCGSTABSolver>()); |
| 602 |
|
| 603 |
/*! \brief Minimal Residual Method (MINRES) |
| 604 |
|
| 605 |
Symmetrically Preconditioned MINRES as in A. Greenbaum, 'Iterative Methods for Solving Linear Systems', pp. 121 |
| 606 |
Iterative solver for symmetric indefinite operators. |
| 607 |
Note that in order to ensure the (symmetrically) preconditioned system to remain symmetric, the preconditioner has to be spd. |
| 608 |
*/ |
| 609 |
template<class X> |
| 610 |
class MINRESSolver : public IterativeSolver<X,X> { |
| 611 |
public: |
| 612 |
using typename IterativeSolver<X,X>::domain_type; |
| 613 |
using typename IterativeSolver<X,X>::range_type; |
| 614 |
using typename IterativeSolver<X,X>::field_type; |
| 615 |
using typename IterativeSolver<X,X>::real_type; |
| 616 |
|
| 617 |
// copy base class constructors |
| 618 |
using IterativeSolver<X,X>::IterativeSolver; |
| 619 |
|
| 620 |
// don't shadow four-argument version of apply defined in the base class |
| 621 |
using IterativeSolver<X,X>::apply; |
| 622 |
|
| 623 |
/*! |
| 624 |
\brief Apply inverse operator. |
| 625 |
|
| 626 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 627 |
*/ |
| 628 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 629 |
{ |
| 630 |
using std::sqrt; |
| 631 |
using std::abs; |
| 632 |
Iteration iteration(*this, res); |
| 633 |
// prepare preconditioner |
| 634 |
_prec->pre(x,b); |
| 635 |
|
| 636 |
// overwrite rhs with defect |
| 637 |
_op->applyscaleadd(-1.0,x,b); // b -= Ax |
| 638 |
|
| 639 |
// some temporary vectors |
| 640 |
X z(b), dummy(b); |
| 641 |
z = 0.0; |
| 642 |
|
| 643 |
// calculate preconditioned defect |
| 644 |
_prec->apply(z,b); // r = W^-1 (b - Ax) |
| 645 |
real_type def = _sp->norm(z); |
| 646 |
if (iteration.step(0, def)){ |
| 647 |
_prec->post(x); |
| 648 |
return; |
| 649 |
} |
| 650 |
|
| 651 |
// recurrence coefficients as computed in Lanczos algorithm |
| 652 |
field_type alpha, beta; |
| 653 |
// diagonal entries of givens rotation |
| 654 |
std::array<real_type,2> c{{0.0,0.0}}; |
| 655 |
// off-diagonal entries of givens rotation |
| 656 |
std::array<field_type,2> s{{0.0,0.0}}; |
| 657 |
|
| 658 |
// recurrence coefficients (column k of tridiag matrix T_k) |
| 659 |
std::array<field_type,3> T{{0.0,0.0,0.0}}; |
| 660 |
|
| 661 |
// the rhs vector of the min problem |
| 662 |
std::array<field_type,2> xi{{1.0,0.0}}; |
| 663 |
|
| 664 |
// beta is real and positive in exact arithmetic |
| 665 |
// since it is the norm of the basis vectors (in unpreconditioned case) |
| 666 |
beta = sqrt(_sp->dot(b,z)); |
| 667 |
field_type beta0 = beta; |
| 668 |
|
| 669 |
// the search directions |
| 670 |
std::array<X,3> p{{b,b,b}}; |
| 671 |
p[0] = 0.0; |
| 672 |
p[1] = 0.0; |
| 673 |
p[2] = 0.0; |
| 674 |
|
| 675 |
// orthonormal basis vectors (in unpreconditioned case) |
| 676 |
std::array<X,3> q{{b,b,b}}; |
| 677 |
q[0] = 0.0; |
| 678 |
q[1] *= Simd::cond(def==field_type(0.), |
| 679 |
field_type(0.), |
| 680 |
field_type(real_type(1.0)/beta)); |
| 681 |
q[2] = 0.0; |
| 682 |
|
| 683 |
z *= Simd::cond(def==field_type(0.), |
| 684 |
field_type(0.), |
| 685 |
field_type(real_type(1.0)/beta)); |
| 686 |
|
| 687 |
// the loop |
| 688 |
int i = 1; |
| 689 |
for( ; i<=_maxit; i++) { |
| 690 |
|
| 691 |
dummy = z; |
| 692 |
int i1 = i%3, |
| 693 |
i0 = (i1+2)%3, |
| 694 |
i2 = (i1+1)%3; |
| 695 |
|
| 696 |
// symmetrically preconditioned Lanczos algorithm (see Greenbaum p.121) |
| 697 |
_op->apply(z,q[i2]); // q[i2] = Az |
| 698 |
q[i2].axpy(-beta,q[i0]); |
| 699 |
// alpha is real since it is the diagonal entry of the hermitian tridiagonal matrix |
| 700 |
// from the Lanczos Algorithm |
| 701 |
// so the order in the scalar product doesn't matter even for the complex case |
| 702 |
alpha = _sp->dot(z,q[i2]); |
| 703 |
q[i2].axpy(-alpha,q[i1]); |
| 704 |
|
| 705 |
z = 0.0; |
| 706 |
_prec->apply(z,q[i2]); |
| 707 |
|
| 708 |
// beta is real and positive in exact arithmetic |
| 709 |
// since it is the norm of the basis vectors (in unpreconditioned case) |
| 710 |
beta = sqrt(_sp->dot(q[i2],z)); |
| 711 |
|
| 712 |
q[i2] *= Simd::cond(def==field_type(0.), |
| 713 |
field_type(0.), |
| 714 |
field_type(real_type(1.0)/beta)); |
| 715 |
z *= Simd::cond(def==field_type(0.), |
| 716 |
field_type(0.), |
| 717 |
field_type(real_type(1.0)/beta)); |
| 718 |
|
| 719 |
// QR Factorization of recurrence coefficient matrix |
| 720 |
// apply previous givens rotations to last column of T |
| 721 |
T[1] = T[2]; |
| 722 |
if(i>2) { |
| 723 |
T[0] = s[i%2]*T[1]; |
| 724 |
T[1] = c[i%2]*T[1]; |
| 725 |
} |
| 726 |
if(i>1) { |
| 727 |
T[2] = c[(i+1)%2]*alpha - s[(i+1)%2]*T[1]; |
| 728 |
T[1] = c[(i+1)%2]*T[1] + s[(i+1)%2]*alpha; |
| 729 |
} |
| 730 |
else |
| 731 |
T[2] = alpha; |
| 732 |
|
| 733 |
// update QR factorization |
| 734 |
generateGivensRotation(T[2],beta,c[i%2],s[i%2]); |
| 735 |
// to last column of T_k |
| 736 |
T[2] = c[i%2]*T[2] + s[i%2]*beta; |
| 737 |
// and to the rhs xi of the min problem |
| 738 |
xi[i%2] = -s[i%2]*xi[(i+1)%2]; |
| 739 |
xi[(i+1)%2] *= c[i%2]; |
| 740 |
|
| 741 |
// compute correction direction |
| 742 |
p[i2] = dummy; |
| 743 |
p[i2].axpy(-T[1],p[i1]); |
| 744 |
p[i2].axpy(-T[0],p[i0]); |
| 745 |
p[i2] *= real_type(1.0)/T[2]; |
| 746 |
|
| 747 |
// apply correction/update solution |
| 748 |
x.axpy(beta0*xi[(i+1)%2],p[i2]); |
| 749 |
|
| 750 |
// remember beta_old |
| 751 |
T[2] = beta; |
| 752 |
|
| 753 |
// check for convergence |
| 754 |
// the last entry in the rhs of the min-problem is the residual |
| 755 |
def = abs(beta0*xi[i%2]); |
| 756 |
if(iteration.step(i, def)){ |
| 757 |
break; |
| 758 |
} |
| 759 |
} // end for |
| 760 |
|
| 761 |
// postprocess preconditioner |
| 762 |
_prec->post(x); |
| 763 |
} |
| 764 |
|
| 765 |
private: |
| 766 |
|
| 767 |
void generateGivensRotation(field_type &dx, field_type &dy, real_type &cs, field_type &sn) |
| 768 |
{ |
| 769 |
using std::sqrt; |
| 770 |
using std::abs; |
| 771 |
using std::max; |
| 772 |
using std::min; |
| 773 |
const real_type eps = 1e-15; |
| 774 |
real_type norm_dx = abs(dx); |
| 775 |
real_type norm_dy = abs(dy); |
| 776 |
real_type norm_max = max(norm_dx, norm_dy); |
| 777 |
real_type norm_min = min(norm_dx, norm_dy); |
| 778 |
real_type temp = norm_min/norm_max; |
| 779 |
// we rewrite the code in a vectorizable fashion |
| 780 |
cs = Simd::cond(norm_dy < eps, |
| 781 |
real_type(1.0), |
| 782 |
real_type(Simd::cond(norm_dx < eps, |
| 783 |
real_type(0.0), |
| 784 |
real_type(Simd::cond(norm_dy > norm_dx, |
| 785 |
real_type(1.0)/sqrt(real_type(1.0) + temp*temp)*temp, |
| 786 |
real_type(1.0)/sqrt(real_type(1.0) + temp*temp) |
| 787 |
))))); |
| 788 |
sn = Simd::cond(norm_dy < eps, |
| 789 |
field_type(0.0), |
| 790 |
field_type(Simd::cond(norm_dx < eps, |
| 791 |
field_type(1.0), |
| 792 |
field_type(Simd::cond(norm_dy > norm_dx, |
| 793 |
// dy and dx are real in exact arithmetic |
| 794 |
// thus dx*dy is real so we can explicitly enforce it |
| 795 |
field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*dx*dy/norm_dx/norm_dy, |
| 796 |
// dy and dx is real in exact arithmetic |
| 797 |
// so we don't have to conjugate both of them |
| 798 |
field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*dy/dx |
| 799 |
))))); |
| 800 |
} |
| 801 |
|
| 802 |
protected: |
| 803 |
using IterativeSolver<X,X>::_op; |
| 804 |
using IterativeSolver<X,X>::_prec; |
| 805 |
using IterativeSolver<X,X>::_sp; |
| 806 |
using IterativeSolver<X,X>::_reduction; |
| 807 |
using IterativeSolver<X,X>::_maxit; |
| 808 |
using IterativeSolver<X,X>::_verbose; |
| 809 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 810 |
}; |
| 811 |
DUNE_REGISTER_SOLVER("minressolver", defaultIterativeSolverCreator<Dune::MINRESSolver>()); |
| 812 |
|
| 813 |
/** |
| 814 |
\brief implements the Generalized Minimal Residual (GMRes) method |
| 815 |
|
| 816 |
GMRes solves the unsymmetric linear system Ax = b using the |
| 817 |
Generalized Minimal Residual method as described the SIAM Templates |
| 818 |
book (http://www.netlib.org/templates/templates.pdf). |
| 819 |
|
| 820 |
\tparam X trial vector, vector type of the solution |
| 821 |
\tparam Y test vector, vector type of the RHS |
| 822 |
\tparam F vector type for orthonormal basis of Krylov space |
| 823 |
|
| 824 |
*/ |
| 825 |
|
| 826 |
template<class X, class Y=X, class F = Y> |
| 827 |
class RestartedGMResSolver : public IterativeSolver<X,Y> |
| 828 |
{ |
| 829 |
public: |
| 830 |
using typename IterativeSolver<X,Y>::domain_type; |
| 831 |
using typename IterativeSolver<X,Y>::range_type; |
| 832 |
using typename IterativeSolver<X,Y>::field_type; |
| 833 |
using typename IterativeSolver<X,Y>::real_type; |
| 834 |
|
| 835 |
protected: |
| 836 |
using typename IterativeSolver<X,X>::scalar_real_type; |
| 837 |
|
| 838 |
//! \brief field_type Allocator retrieved from domain type |
| 839 |
using fAlloc = ReboundAllocatorType<X,field_type>; |
| 840 |
//! \brief real_type Allocator retrieved from domain type |
| 841 |
using rAlloc = ReboundAllocatorType<X,real_type>; |
| 842 |
|
| 843 |
public: |
| 844 |
|
| 845 |
/*! |
| 846 |
\brief Set up RestartedGMResSolver solver. |
| 847 |
|
| 848 |
\copydoc IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&,Preconditioner<X,Y>&,scalar_real_type,int,int) |
| 849 |
\param restart number of GMRes cycles before restart |
| 850 |
*/ |
| 851 |
RestartedGMResSolver (const LinearOperator<X,Y>& op, Preconditioner<X,Y>& prec, scalar_real_type reduction, int restart, int maxit, int verbose) : |
| 852 |
IterativeSolver<X,Y>::IterativeSolver(op,prec,reduction,maxit,verbose), |
| 853 |
_restart(restart) |
| 854 |
{} |
| 855 |
|
| 856 |
/*! |
| 857 |
\brief Set up RestartedGMResSolver solver. |
| 858 |
|
| 859 |
\copydoc IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&,const ScalarProduct<X>&,Preconditioner<X,Y>&,scalar_real_type,int,int) |
| 860 |
\param restart number of GMRes cycles before restart |
| 861 |
*/ |
| 862 |
RestartedGMResSolver (const LinearOperator<X,Y>& op, const ScalarProduct<X>& sp, Preconditioner<X,Y>& prec, scalar_real_type reduction, int restart, int maxit, int verbose) : |
| 863 |
IterativeSolver<X,Y>::IterativeSolver(op,sp,prec,reduction,maxit,verbose), |
| 864 |
_restart(restart) |
| 865 |
{} |
| 866 |
|
| 867 |
/*! |
| 868 |
\brief Constructor. |
| 869 |
|
| 870 |
\copydoc IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>,std::shared_ptr<Preconditioner<X,X>>,const ParameterTree&) |
| 871 |
|
| 872 |
Additional parameter: |
| 873 |
ParameterTree Key | Meaning |
| 874 |
------------------|------------ |
| 875 |
restart | number of GMRes cycles before restart |
| 876 |
|
| 877 |
See \ref ISTL_Factory for the ParameterTree layout and examples. |
| 878 |
*/ |
| 879 |
RestartedGMResSolver (std::shared_ptr<const LinearOperator<X,Y> > op, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) : |
| 880 |
IterativeSolver<X,Y>::IterativeSolver(op,prec,configuration), |
| 881 |
_restart(configuration.get<int>("restart")) |
| 882 |
{} |
| 883 |
|
| 884 |
RestartedGMResSolver (std::shared_ptr<const LinearOperator<X,Y> > op, std::shared_ptr<const ScalarProduct<X> > sp, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) : |
| 885 |
IterativeSolver<X,Y>::IterativeSolver(op,sp,prec,configuration), |
| 886 |
_restart(configuration.get<int>("restart")) |
| 887 |
{} |
| 888 |
|
| 889 |
/*! |
| 890 |
\brief Set up RestartedGMResSolver solver. |
| 891 |
|
| 892 |
\copydetails IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>,std::shared_ptr<const ScalarProduct<X>>,std::shared_ptr<Preconditioner<X,Y>>,scalar_real_type,int,int) |
| 893 |
\param restart number of GMRes cycles before restart |
| 894 |
*/ |
| 895 |
RestartedGMResSolver (std::shared_ptr<const LinearOperator<X,Y>> op, |
| 896 |
std::shared_ptr<const ScalarProduct<X>> sp, |
| 897 |
std::shared_ptr<Preconditioner<X,Y>> prec, |
| 898 |
scalar_real_type reduction, int restart, int maxit, int verbose) : |
| 899 |
IterativeSolver<X,Y>::IterativeSolver(op,sp,prec,reduction,maxit,verbose), |
| 900 |
_restart(restart) |
| 901 |
{} |
| 902 |
|
| 903 |
/*! |
| 904 |
\brief Apply inverse operator. |
| 905 |
|
| 906 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 907 |
|
| 908 |
\note Currently, the RestartedGMResSolver aborts when it detects a |
| 909 |
breakdown. |
| 910 |
*/ |
| 911 |
void apply (X& x, Y& b, InverseOperatorResult& res) override |
| 912 |
{ |
| 913 |
apply(x,b,Simd::max(_reduction),res); |
| 914 |
} |
| 915 |
|
| 916 |
/*! |
| 917 |
\brief Apply inverse operator. |
| 918 |
|
| 919 |
\copydoc InverseOperator::apply(X&,Y&,double,InverseOperatorResult&) |
| 920 |
|
| 921 |
\note Currently, the RestartedGMResSolver aborts when it detects a |
| 922 |
breakdown. |
| 923 |
*/ |
| 924 |
void apply (X& x, Y& b, [[maybe_unused]] double reduction, InverseOperatorResult& res) override |
| 925 |
{ |
| 926 |
using std::abs; |
| 927 |
const Simd::Scalar<real_type> EPSILON = 1e-80; |
| 928 |
const int m = _restart; |
| 929 |
real_type norm = 0.0; |
| 930 |
int j = 1; |
| 931 |
std::vector<field_type,fAlloc> s(m+1), sn(m); |
| 932 |
std::vector<real_type,rAlloc> cs(m); |
| 933 |
// need copy of rhs if GMRes has to be restarted |
| 934 |
Y b2(b); |
| 935 |
// helper vector |
| 936 |
Y w(b); |
| 937 |
std::vector< std::vector<field_type,fAlloc> > H(m+1,s); |
| 938 |
std::vector<F> v(m+1,b); |
| 939 |
|
| 940 |
Iteration iteration(*this,res); |
| 941 |
|
| 942 |
// clear solver statistics and set res.converged to false |
| 943 |
_prec->pre(x,b); |
| 944 |
|
| 945 |
// calculate defect and overwrite rhs with it |
| 946 |
_op->applyscaleadd(-1.0,x,b); // b -= Ax |
| 947 |
// calculate preconditioned defect |
| 948 |
v[0] = 0.0; _prec->apply(v[0],b); // r = W^-1 b |
| 949 |
norm = _sp->norm(v[0]); |
| 950 |
if(iteration.step(0, norm)){ |
| 951 |
_prec->post(x); |
| 952 |
return; |
| 953 |
} |
| 954 |
|
| 955 |
while(j <= _maxit && res.converged != true) { |
| 956 |
|
| 957 |
int i = 0; |
| 958 |
v[0] *= Simd::cond(norm==real_type(0.), |
| 959 |
real_type(0.), |
| 960 |
real_type(real_type(1.0)/norm)); |
| 961 |
s[0] = norm; |
| 962 |
for(i=1; i<m+1; i++) |
| 963 |
s[i] = 0.0; |
| 964 |
|
| 965 |
for(i=0; i < m && j <= _maxit && res.converged != true; i++, j++) { |
| 966 |
w = 0.0; |
| 967 |
// use v[i+1] as temporary vector |
| 968 |
v[i+1] = 0.0; |
| 969 |
// do Arnoldi algorithm |
| 970 |
_op->apply(v[i],v[i+1]); |
| 971 |
_prec->apply(w,v[i+1]); |
| 972 |
for(int k=0; k<i+1; k++) { |
| 973 |
// notice that _sp->dot(v[k],w) = v[k]\adjoint w |
| 974 |
// so one has to pay attention to the order |
| 975 |
// in the scalar product for the complex case |
| 976 |
// doing the modified Gram-Schmidt algorithm |
| 977 |
H[k][i] = _sp->dot(v[k],w); |
| 978 |
// w -= H[k][i] * v[k] |
| 979 |
w.axpy(-H[k][i],v[k]); |
| 980 |
} |
| 981 |
H[i+1][i] = _sp->norm(w); |
| 982 |
if(Simd::allTrue(abs(H[i+1][i]) < EPSILON)) |
| 983 |
DUNE_THROW(SolverAbort, |
| 984 |
"breakdown in GMRes - |w| == 0.0 after " << j << " iterations"); |
| 985 |
|
| 986 |
// normalize new vector |
| 987 |
v[i+1] = w; |
| 988 |
v[i+1] *= Simd::cond(norm==real_type(0.), |
| 989 |
field_type(0.), |
| 990 |
field_type(real_type(1.0)/H[i+1][i])); |
| 991 |
|
| 992 |
// update QR factorization |
| 993 |
for(int k=0; k<i; k++) |
| 994 |
applyPlaneRotation(H[k][i],H[k+1][i],cs[k],sn[k]); |
| 995 |
|
| 996 |
// compute new givens rotation |
| 997 |
generatePlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]); |
| 998 |
// finish updating QR factorization |
| 999 |
applyPlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]); |
| 1000 |
applyPlaneRotation(s[i],s[i+1],cs[i],sn[i]); |
| 1001 |
|
| 1002 |
// norm of the defect is the last component the vector s |
| 1003 |
norm = abs(s[i+1]); |
| 1004 |
|
| 1005 |
iteration.step(j, norm); |
| 1006 |
|
| 1007 |
} // end for |
| 1008 |
|
| 1009 |
// calculate update vector |
| 1010 |
w = 0.0; |
| 1011 |
update(w,i,H,s,v); |
| 1012 |
// and current iterate |
| 1013 |
x += w; |
| 1014 |
|
| 1015 |
// restart GMRes if convergence was not achieved, |
| 1016 |
// i.e. linear defect has not reached desired reduction |
| 1017 |
// and if j < _maxit (do not restart on last iteration) |
| 1018 |
if( res.converged != true && j < _maxit ) { |
| 1019 |
|
| 1020 |
if(_verbose > 0) |
| 1021 |
std::cout << "=== GMRes::restart" << std::endl; |
| 1022 |
// get saved rhs |
| 1023 |
b = b2; |
| 1024 |
// calculate new defect |
| 1025 |
_op->applyscaleadd(-1.0,x,b); // b -= Ax; |
| 1026 |
// calculate preconditioned defect |
| 1027 |
v[0] = 0.0; |
| 1028 |
_prec->apply(v[0],b); |
| 1029 |
norm = _sp->norm(v[0]); |
| 1030 |
} |
| 1031 |
|
| 1032 |
} //end while |
| 1033 |
|
| 1034 |
// postprocess preconditioner |
| 1035 |
_prec->post(x); |
| 1036 |
} |
| 1037 |
|
| 1038 |
protected : |
| 1039 |
|
| 1040 |
void update(X& w, int i, |
| 1041 |
const std::vector<std::vector<field_type,fAlloc> >& H, |
| 1042 |
const std::vector<field_type,fAlloc>& s, |
| 1043 |
const std::vector<X>& v) { |
| 1044 |
// solution vector of the upper triangular system |
| 1045 |
std::vector<field_type,fAlloc> y(s); |
| 1046 |
|
| 1047 |
// backsolve |
| 1048 |
for(int a=i-1; a>=0; a--) { |
| 1049 |
field_type rhs(s[a]); |
| 1050 |
for(int b=a+1; b<i; b++) |
| 1051 |
rhs -= H[a][b]*y[b]; |
| 1052 |
y[a] = Simd::cond(rhs==field_type(0.), |
| 1053 |
field_type(0.), |
| 1054 |
field_type(rhs/H[a][a])); |
| 1055 |
|
| 1056 |
// compute update on the fly |
| 1057 |
// w += y[a]*v[a] |
| 1058 |
w.axpy(y[a],v[a]); |
| 1059 |
} |
| 1060 |
} |
| 1061 |
|
| 1062 |
template<typename T> |
| 1063 |
typename std::enable_if<std::is_same<field_type,real_type>::value,T>::type conjugate(const T& t) { |
| 1064 |
return t; |
| 1065 |
} |
| 1066 |
|
| 1067 |
template<typename T> |
| 1068 |
typename std::enable_if<!std::is_same<field_type,real_type>::value,T>::type conjugate(const T& t) { |
| 1069 |
using std::conj; |
| 1070 |
return conj(t); |
| 1071 |
} |
| 1072 |
|
| 1073 |
void |
| 1074 |
generatePlaneRotation(field_type &dx, field_type &dy, real_type &cs, field_type &sn) |
| 1075 |
{ |
| 1076 |
using std::sqrt; |
| 1077 |
using std::abs; |
| 1078 |
using std::max; |
| 1079 |
using std::min; |
| 1080 |
const real_type eps = 1e-15; |
| 1081 |
real_type norm_dx = abs(dx); |
| 1082 |
real_type norm_dy = abs(dy); |
| 1083 |
real_type norm_max = max(norm_dx, norm_dy); |
| 1084 |
real_type norm_min = min(norm_dx, norm_dy); |
| 1085 |
real_type temp = norm_min/norm_max; |
| 1086 |
// we rewrite the code in a vectorizable fashion |
| 1087 |
cs = Simd::cond(norm_dy < eps, |
| 1088 |
real_type(1.0), |
| 1089 |
Simd::cond(norm_dx < eps, |
| 1090 |
real_type(0.0), |
| 1091 |
real_type(Simd::cond(norm_dy > norm_dx, |
| 1092 |
real_type(1.0)/sqrt(real_type(1.0) + temp*temp)*temp, |
| 1093 |
real_type(1.0)/sqrt(real_type(1.0) + temp*temp) |
| 1094 |
)))); |
| 1095 |
sn = Simd::cond(norm_dy < eps, |
| 1096 |
field_type(0.0), |
| 1097 |
Simd::cond(norm_dx < eps, |
| 1098 |
field_type(1.0), |
| 1099 |
field_type(Simd::cond(norm_dy > norm_dx, |
| 1100 |
field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*dx*conjugate(dy)/norm_dx/norm_dy, |
| 1101 |
field_type(1.0)/sqrt(real_type(1.0) + temp*temp)*conjugate(dy/dx) |
| 1102 |
)))); |
| 1103 |
} |
| 1104 |
|
| 1105 |
|
| 1106 |
void |
| 1107 |
applyPlaneRotation(field_type &dx, field_type &dy, real_type &cs, field_type &sn) |
| 1108 |
{ |
| 1109 |
field_type temp = cs * dx + sn * dy; |
| 1110 |
dy = -conjugate(sn) * dx + cs * dy; |
| 1111 |
dx = temp; |
| 1112 |
} |
| 1113 |
|
| 1114 |
using IterativeSolver<X,Y>::_op; |
| 1115 |
using IterativeSolver<X,Y>::_prec; |
| 1116 |
using IterativeSolver<X,Y>::_sp; |
| 1117 |
using IterativeSolver<X,Y>::_reduction; |
| 1118 |
using IterativeSolver<X,Y>::_maxit; |
| 1119 |
using IterativeSolver<X,Y>::_verbose; |
| 1120 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 1121 |
int _restart; |
| 1122 |
}; |
| 1123 |
DUNE_REGISTER_SOLVER("restartedgmressolver", defaultIterativeSolverCreator<Dune::RestartedGMResSolver>()); |
| 1124 |
|
| 1125 |
/** |
| 1126 |
\brief implements the Flexible Generalized Minimal Residual (FGMRes) method (right preconditioned) |
| 1127 |
|
| 1128 |
FGMRes solves the right-preconditioned unsymmetric linear system Ax = b using the |
| 1129 |
Flexible Generalized Minimal Residual method. It is flexible because the preconditioner can change in every iteration, |
| 1130 |
which allows to use Krylov solvers without fixed number of iterations as preconditioners. Needs more memory than GMRes. |
| 1131 |
|
| 1132 |
\tparam X trial vector, vector type of the solution |
| 1133 |
\tparam Y test vector, vector type of the RHS |
| 1134 |
\tparam F vector type for orthonormal basis of Krylov space |
| 1135 |
|
| 1136 |
*/ |
| 1137 |
|
| 1138 |
template<class X, class Y=X, class F = Y> |
| 1139 |
class RestartedFlexibleGMResSolver : public RestartedGMResSolver<X,Y> |
| 1140 |
{ |
| 1141 |
public: |
| 1142 |
using typename RestartedGMResSolver<X,Y>::domain_type; |
| 1143 |
using typename RestartedGMResSolver<X,Y>::range_type; |
| 1144 |
using typename RestartedGMResSolver<X,Y>::field_type; |
| 1145 |
using typename RestartedGMResSolver<X,Y>::real_type; |
| 1146 |
|
| 1147 |
private: |
| 1148 |
using typename RestartedGMResSolver<X,Y>::scalar_real_type; |
| 1149 |
|
| 1150 |
//! \brief field_type Allocator retrieved from domain type |
| 1151 |
using fAlloc = typename RestartedGMResSolver<X,Y>::fAlloc; |
| 1152 |
//! \brief real_type Allocator retrieved from domain type |
| 1153 |
using rAlloc = typename RestartedGMResSolver<X,Y>::rAlloc; |
| 1154 |
|
| 1155 |
public: |
| 1156 |
// copy base class constructors |
| 1157 |
using RestartedGMResSolver<X,Y>::RestartedGMResSolver; |
| 1158 |
|
| 1159 |
// don't shadow four-argument version of apply defined in the base class |
| 1160 |
using RestartedGMResSolver<X,Y>::apply; |
| 1161 |
|
| 1162 |
/*! |
| 1163 |
\brief Apply inverse operator. |
| 1164 |
|
| 1165 |
\copydoc InverseOperator::apply(X&,Y&,double,InverseOperatorResult&) |
| 1166 |
|
| 1167 |
\note Currently, the RestartedFlexibleGMResSolver aborts when it detects a |
| 1168 |
breakdown. |
| 1169 |
*/ |
| 1170 |
void apply (X& x, Y& b, [[maybe_unused]] double reduction, InverseOperatorResult& res) override |
| 1171 |
{ |
| 1172 |
using std::abs; |
| 1173 |
const Simd::Scalar<real_type> EPSILON = 1e-80; |
| 1174 |
const int m = _restart; |
| 1175 |
real_type norm = 0.0; |
| 1176 |
int i, j = 1, k; |
| 1177 |
std::vector<field_type,fAlloc> s(m+1), sn(m); |
| 1178 |
std::vector<real_type,rAlloc> cs(m); |
| 1179 |
// helper vector |
| 1180 |
Y tmp(b); |
| 1181 |
std::vector< std::vector<field_type,fAlloc> > H(m+1,s); |
| 1182 |
std::vector<F> v(m+1,b); |
| 1183 |
std::vector<X> w(m+1,b); |
| 1184 |
|
| 1185 |
Iteration iteration(*this,res); |
| 1186 |
// setup preconditioner if it does something in pre |
| 1187 |
|
| 1188 |
// calculate residual and overwrite a copy of the rhs with it |
| 1189 |
_prec->pre(x, b); |
| 1190 |
v[0] = b; |
| 1191 |
_op->applyscaleadd(-1.0, x, v[0]); // b -= Ax |
| 1192 |
|
| 1193 |
norm = _sp->norm(v[0]); // the residual norm |
| 1194 |
if(iteration.step(0, norm)){ |
| 1195 |
_prec->post(x); |
| 1196 |
return; |
| 1197 |
} |
| 1198 |
|
| 1199 |
// start iterations |
| 1200 |
res.converged = false;; |
| 1201 |
while(j <= _maxit && res.converged != true) |
| 1202 |
{ |
| 1203 |
v[0] *= (1.0 / norm); |
| 1204 |
s[0] = norm; |
| 1205 |
for(i=1; i<m+1; ++i) |
| 1206 |
s[i] = 0.0; |
| 1207 |
|
| 1208 |
// inner loop |
| 1209 |
for(i=0; i < m && j <= _maxit && res.converged != true; i++, j++) |
| 1210 |
{ |
| 1211 |
w[i] = 0.0; |
| 1212 |
// compute wi = M^-1*vi (also called zi) |
| 1213 |
_prec->apply(w[i], v[i]); |
| 1214 |
// compute vi = A*wi |
| 1215 |
// use v[i+1] as temporary vector for w |
| 1216 |
_op->apply(w[i], v[i+1]); |
| 1217 |
// do Arnoldi algorithm |
| 1218 |
for(int kk=0; kk<i+1; kk++) |
| 1219 |
{ |
| 1220 |
// notice that _sp->dot(v[k],v[i+1]) = v[k]\adjoint v[i+1] |
| 1221 |
// so one has to pay attention to the order |
| 1222 |
// in the scalar product for the complex case |
| 1223 |
// doing the modified Gram-Schmidt algorithm |
| 1224 |
H[kk][i] = _sp->dot(v[kk],v[i+1]); |
| 1225 |
// w -= H[k][i] * v[kk] |
| 1226 |
v[i+1].axpy(-H[kk][i], v[kk]); |
| 1227 |
} |
| 1228 |
H[i+1][i] = _sp->norm(v[i+1]); |
| 1229 |
if(Simd::allTrue(abs(H[i+1][i]) < EPSILON)) |
| 1230 |
DUNE_THROW(SolverAbort, "breakdown in fGMRes - |w| (-> " |
| 1231 |
<< w[i] << ") == 0.0 after " |
| 1232 |
<< j << " iterations"); |
| 1233 |
|
| 1234 |
// v[i+1] = w*1/H[i+1][i] |
| 1235 |
v[i+1] *= real_type(1.0)/H[i+1][i]; |
| 1236 |
|
| 1237 |
// update QR factorization |
| 1238 |
for(k=0; k<i; k++) |
| 1239 |
this->applyPlaneRotation(H[k][i],H[k+1][i],cs[k],sn[k]); |
| 1240 |
|
| 1241 |
// compute new givens rotation |
| 1242 |
this->generatePlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]); |
| 1243 |
|
| 1244 |
// finish updating QR factorization |
| 1245 |
this->applyPlaneRotation(H[i][i],H[i+1][i],cs[i],sn[i]); |
| 1246 |
this->applyPlaneRotation(s[i],s[i+1],cs[i],sn[i]); |
| 1247 |
|
| 1248 |
// norm of the residual is the last component of vector s |
| 1249 |
using std::abs; |
| 1250 |
norm = abs(s[i+1]); |
| 1251 |
iteration.step(j, norm); |
| 1252 |
} // end inner for loop |
| 1253 |
|
| 1254 |
// calculate update vector |
| 1255 |
tmp = 0.0; |
| 1256 |
this->update(tmp, i, H, s, w); |
| 1257 |
// and update current iterate |
| 1258 |
x += tmp; |
| 1259 |
|
| 1260 |
// restart fGMRes if convergence was not achieved, |
| 1261 |
// i.e. linear residual has not reached desired reduction |
| 1262 |
// and if still j < _maxit (do not restart on last iteration) |
| 1263 |
if( res.converged != true && j < _maxit) |
| 1264 |
{ |
| 1265 |
if (_verbose > 0) |
| 1266 |
std::cout << "=== fGMRes::restart" << std::endl; |
| 1267 |
// get rhs |
| 1268 |
v[0] = b; |
| 1269 |
// calculate new defect |
| 1270 |
_op->applyscaleadd(-1.0, x,v[0]); // b -= Ax; |
| 1271 |
// calculate preconditioned defect |
| 1272 |
norm = _sp->norm(v[0]); // update the residual norm |
| 1273 |
} |
| 1274 |
|
| 1275 |
} // end outer while loop |
| 1276 |
|
| 1277 |
// post-process preconditioner |
| 1278 |
_prec->post(x); |
| 1279 |
} |
| 1280 |
|
| 1281 |
private: |
| 1282 |
using RestartedGMResSolver<X,Y>::_op; |
| 1283 |
using RestartedGMResSolver<X,Y>::_prec; |
| 1284 |
using RestartedGMResSolver<X,Y>::_sp; |
| 1285 |
using RestartedGMResSolver<X,Y>::_reduction; |
| 1286 |
using RestartedGMResSolver<X,Y>::_maxit; |
| 1287 |
using RestartedGMResSolver<X,Y>::_verbose; |
| 1288 |
using RestartedGMResSolver<X,Y>::_restart; |
| 1289 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 1290 |
}; |
| 1291 |
DUNE_REGISTER_SOLVER("restartedflexiblegmressolver", defaultIterativeSolverCreator<Dune::RestartedFlexibleGMResSolver>()); |
| 1292 |
|
| 1293 |
/** |
| 1294 |
* @brief Generalized preconditioned conjugate gradient solver. |
| 1295 |
* |
| 1296 |
* A preconditioned conjugate gradient that allows |
| 1297 |
* the preconditioner to change between iterations. |
| 1298 |
* |
| 1299 |
* One example for such preconditioner is AMG when used without |
| 1300 |
* a direct coarse solver. In this case the number of iterations |
| 1301 |
* performed on the coarsest level might change between applications. |
| 1302 |
* |
| 1303 |
* In contrast to CGSolver the search directions are stored and |
| 1304 |
* the orthogonalization is done explicitly. |
| 1305 |
*/ |
| 1306 |
template<class X> |
| 1307 |
class GeneralizedPCGSolver : public IterativeSolver<X,X> |
| 1308 |
{ |
| 1309 |
public: |
| 1310 |
using typename IterativeSolver<X,X>::domain_type; |
| 1311 |
using typename IterativeSolver<X,X>::range_type; |
| 1312 |
using typename IterativeSolver<X,X>::field_type; |
| 1313 |
using typename IterativeSolver<X,X>::real_type; |
| 1314 |
|
| 1315 |
private: |
| 1316 |
using typename IterativeSolver<X,X>::scalar_real_type; |
| 1317 |
|
| 1318 |
//! \brief field_type Allocator retrieved from domain type |
| 1319 |
using fAlloc = ReboundAllocatorType<X,field_type>; |
| 1320 |
|
| 1321 |
public: |
| 1322 |
|
| 1323 |
// don't shadow four-argument version of apply defined in the base class |
| 1324 |
using IterativeSolver<X,X>::apply; |
| 1325 |
|
| 1326 |
/*! |
| 1327 |
\brief Set up nonlinear preconditioned conjugate gradient solver. |
| 1328 |
|
| 1329 |
\copydetails IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&,Preconditioner<X,Y>&,scalar_real_type,int,int) |
| 1330 |
\param restart number of GMRes cycles before restart |
| 1331 |
*/ |
| 1332 |
GeneralizedPCGSolver (const LinearOperator<X,X>& op, Preconditioner<X,X>& prec, scalar_real_type reduction, int maxit, int verbose, int restart = 10) : |
| 1333 |
IterativeSolver<X,X>::IterativeSolver(op,prec,reduction,maxit,verbose), |
| 1334 |
_restart(restart) |
| 1335 |
{} |
| 1336 |
|
| 1337 |
/*! |
| 1338 |
\brief Set up nonlinear preconditioned conjugate gradient solver. |
| 1339 |
|
| 1340 |
\copydetails IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&,const ScalarProduct<X>&,Preconditioner<X,Y>&,scalar_real_type,int,int) |
| 1341 |
\param restart When to restart the construction of |
| 1342 |
the Krylov search space. |
| 1343 |
*/ |
| 1344 |
GeneralizedPCGSolver (const LinearOperator<X,X>& op, const ScalarProduct<X>& sp, Preconditioner<X,X>& prec, scalar_real_type reduction, int maxit, int verbose, int restart = 10) : |
| 1345 |
IterativeSolver<X,X>::IterativeSolver(op,sp,prec,reduction,maxit,verbose), |
| 1346 |
_restart(restart) |
| 1347 |
{} |
| 1348 |
|
| 1349 |
|
| 1350 |
/*! |
| 1351 |
\brief Set up nonlinear preconditioned conjugate gradient solver. |
| 1352 |
|
| 1353 |
\copydetails IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>,std::shared_ptr<Preconditioner<X,X>>,const ParameterTree&) |
| 1354 |
|
| 1355 |
Additional parameter: |
| 1356 |
ParameterTree Key | Meaning |
| 1357 |
------------------|------------ |
| 1358 |
restart | number of PCG cycles before restart |
| 1359 |
|
| 1360 |
See \ref ISTL_Factory for the ParameterTree layout and examples. |
| 1361 |
*/ |
| 1362 |
GeneralizedPCGSolver (std::shared_ptr<const LinearOperator<X,X> > op, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) : |
| 1363 |
IterativeSolver<X,X>::IterativeSolver(op,prec,configuration), |
| 1364 |
_restart(configuration.get<int>("restart")) |
| 1365 |
{} |
| 1366 |
|
| 1367 |
GeneralizedPCGSolver (std::shared_ptr<const LinearOperator<X,X> > op, std::shared_ptr<const ScalarProduct<X> > sp, std::shared_ptr<Preconditioner<X,X> > prec, const ParameterTree& configuration) : |
| 1368 |
IterativeSolver<X,X>::IterativeSolver(op,sp,prec,configuration), |
| 1369 |
_restart(configuration.get<int>("restart")) |
| 1370 |
{} |
| 1371 |
/*! |
| 1372 |
\brief Set up nonlinear preconditioned conjugate gradient solver. |
| 1373 |
|
| 1374 |
\copydetails IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>,std::shared_ptr<const ScalarProduct<X>>,std::shared_ptr<Preconditioner<X,Y>>,scalar_real_type,int,int) |
| 1375 |
\param restart When to restart the construction of the Krylov search space. |
| 1376 |
*/ |
| 1377 |
GeneralizedPCGSolver (std::shared_ptr<const LinearOperator<X,X>> op, |
| 1378 |
std::shared_ptr<const ScalarProduct<X>> sp, |
| 1379 |
std::shared_ptr<Preconditioner<X,X>> prec, |
| 1380 |
scalar_real_type reduction, int maxit, int verbose, |
| 1381 |
int restart = 10) : |
| 1382 |
IterativeSolver<X,X>::IterativeSolver(op,sp,prec,reduction,maxit,verbose), |
| 1383 |
_restart(restart) |
| 1384 |
{} |
| 1385 |
|
| 1386 |
/*! |
| 1387 |
\brief Apply inverse operator. |
| 1388 |
|
| 1389 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 1390 |
*/ |
| 1391 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 1392 |
{ |
| 1393 |
Iteration iteration(*this, res); |
| 1394 |
_prec->pre(x,b); // prepare preconditioner |
| 1395 |
_op->applyscaleadd(-1,x,b); // overwrite b with defect |
| 1396 |
|
| 1397 |
std::vector<std::shared_ptr<X> > p(_restart); |
| 1398 |
std::vector<field_type,fAlloc> pp(_restart); |
| 1399 |
X q(x); // a temporary vector |
| 1400 |
X prec_res(x); // a temporary vector for preconditioner output |
| 1401 |
|
| 1402 |
p[0].reset(new X(x)); |
| 1403 |
|
| 1404 |
real_type def = _sp->norm(b); // compute norm |
| 1405 |
if(iteration.step(0, def)){ |
| 1406 |
_prec->post(x); |
| 1407 |
return; |
| 1408 |
} |
| 1409 |
// some local variables |
| 1410 |
field_type rho, lambda; |
| 1411 |
|
| 1412 |
int i=0; |
| 1413 |
// determine initial search direction |
| 1414 |
*(p[0]) = 0; // clear correction |
| 1415 |
_prec->apply(*(p[0]),b); // apply preconditioner |
| 1416 |
rho = _sp->dot(*(p[0]),b); // orthogonalization |
| 1417 |
_op->apply(*(p[0]),q); // q=Ap |
| 1418 |
pp[0] = _sp->dot(*(p[0]),q); // scalar product |
| 1419 |
lambda = rho/pp[0]; // minimization |
| 1420 |
x.axpy(lambda,*(p[0])); // update solution |
| 1421 |
b.axpy(-lambda,q); // update defect |
| 1422 |
|
| 1423 |
// convergence test |
| 1424 |
def=_sp->norm(b); // comp defect norm |
| 1425 |
++i; |
| 1426 |
if(iteration.step(i, def)){ |
| 1427 |
_prec->post(x); |
| 1428 |
return; |
| 1429 |
} |
| 1430 |
|
| 1431 |
while(i<_maxit) { |
| 1432 |
// the loop |
| 1433 |
int end=std::min(_restart, _maxit-i+1); |
| 1434 |
for (int ii = 1; ii < end; ++ii) |
| 1435 |
{ |
| 1436 |
//std::cout<<" ii="<<ii<<" i="<<i<<std::endl; |
| 1437 |
// compute next conjugate direction |
| 1438 |
prec_res = 0; // clear correction |
| 1439 |
_prec->apply(prec_res,b); // apply preconditioner |
| 1440 |
|
| 1441 |
p[ii].reset(new X(prec_res)); |
| 1442 |
_op->apply(prec_res, q); |
| 1443 |
|
| 1444 |
for(int j=0; j<ii; ++j) { |
| 1445 |
rho =_sp->dot(q,*(p[j]))/pp[j]; |
| 1446 |
p[ii]->axpy(-rho, *(p[j])); |
| 1447 |
} |
| 1448 |
|
| 1449 |
// minimize in given search direction |
| 1450 |
_op->apply(*(p[ii]),q); // q=Ap |
| 1451 |
pp[ii] = _sp->dot(*(p[ii]),q); // scalar product |
| 1452 |
rho = _sp->dot(*(p[ii]),b); // orthogonalization |
| 1453 |
lambda = rho/pp[ii]; // minimization |
| 1454 |
x.axpy(lambda,*(p[ii])); // update solution |
| 1455 |
b.axpy(-lambda,q); // update defect |
| 1456 |
|
| 1457 |
// convergence test |
| 1458 |
def = _sp->norm(b); // comp defect norm |
| 1459 |
|
| 1460 |
++i; |
| 1461 |
iteration.step(i, def); |
| 1462 |
} |
| 1463 |
if(res.converged) |
| 1464 |
break; |
| 1465 |
if(end==_restart) { |
| 1466 |
*(p[0])=*(p[_restart-1]); |
| 1467 |
pp[0]=pp[_restart-1]; |
| 1468 |
} |
| 1469 |
} |
| 1470 |
|
| 1471 |
// postprocess preconditioner |
| 1472 |
_prec->post(x); |
| 1473 |
|
| 1474 |
} |
| 1475 |
|
| 1476 |
private: |
| 1477 |
using IterativeSolver<X,X>::_op; |
| 1478 |
using IterativeSolver<X,X>::_prec; |
| 1479 |
using IterativeSolver<X,X>::_sp; |
| 1480 |
using IterativeSolver<X,X>::_reduction; |
| 1481 |
using IterativeSolver<X,X>::_maxit; |
| 1482 |
using IterativeSolver<X,X>::_verbose; |
| 1483 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 1484 |
int _restart; |
| 1485 |
}; |
| 1486 |
DUNE_REGISTER_SOLVER("generalizedpcgsolver", defaultIterativeSolverCreator<Dune::GeneralizedPCGSolver>()); |
| 1487 |
|
| 1488 |
/*! \brief Accelerated flexible conjugate gradient method |
| 1489 |
|
| 1490 |
Flexible conjugate gradient method as in Y. Notay 'Flexible conjugate Gradients', |
| 1491 |
SIAM J. Sci. Comput Vol. 22, No.4, pp. 1444-1460 |
| 1492 |
|
| 1493 |
This solver discard cyclic all old directions to speed up computing. |
| 1494 |
In exact arithmetic it is exactly the same as the GeneralizedPCGSolver, |
| 1495 |
but it is much faster, depending on the operator and dimension. |
| 1496 |
On the other hand for large mmax it uses noticeably more memory. |
| 1497 |
|
| 1498 |
*/ |
| 1499 |
template<class X> |
| 1500 |
class RestartedFCGSolver : public IterativeSolver<X,X> { |
| 1501 |
public: |
| 1502 |
using typename IterativeSolver<X,X>::domain_type; |
| 1503 |
using typename IterativeSolver<X,X>::range_type; |
| 1504 |
using typename IterativeSolver<X,X>::field_type; |
| 1505 |
using typename IterativeSolver<X,X>::real_type; |
| 1506 |
|
| 1507 |
private: |
| 1508 |
using typename IterativeSolver<X,X>::scalar_real_type; |
| 1509 |
|
| 1510 |
public: |
| 1511 |
// don't shadow four-argument version of apply defined in the base class |
| 1512 |
using IterativeSolver<X,X>::apply; |
| 1513 |
/*! |
| 1514 |
\brief Constructor to initialize a RestartedFCG solver. |
| 1515 |
\copydetails IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&, Preconditioner<X,Y>&, scalar_real_type, int, int) |
| 1516 |
\param mmax is the maximal number of previous vectors which are orthogonalized against the new search direction. |
| 1517 |
*/ |
| 1518 |
RestartedFCGSolver (const LinearOperator<X,X>& op, Preconditioner<X,X>& prec, |
| 1519 |
scalar_real_type reduction, int maxit, int verbose, int mmax = 10) : IterativeSolver<X,X>(op, prec, reduction, maxit, verbose), _mmax(mmax) |
| 1520 |
{ |
| 1521 |
} |
| 1522 |
|
| 1523 |
/*! |
| 1524 |
\brief Constructor to initialize a RestartedFCG solver. |
| 1525 |
\copydetails IterativeSolver::IterativeSolver(const LinearOperator<X,Y>&, const ScalarProduct<X>&, Preconditioner<X,Y>&, scalar_real_type, int, int) |
| 1526 |
\param mmax is the maximal number of previous vectors which are orthogonalized against the new search direction. |
| 1527 |
*/ |
| 1528 |
RestartedFCGSolver (const LinearOperator<X,X>& op, const ScalarProduct<X>& sp, Preconditioner<X,X>& prec, |
| 1529 |
scalar_real_type reduction, int maxit, int verbose, int mmax = 10) : IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose), _mmax(mmax) |
| 1530 |
{ |
| 1531 |
} |
| 1532 |
|
| 1533 |
/*! |
| 1534 |
\brief Constructor to initialize a RestartedFCG solver. |
| 1535 |
\copydetails IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>, std::shared_ptr<const ScalarProduct<X>>, std::shared_ptr<Preconditioner<X,Y>>, scalar_real_type, int, int) |
| 1536 |
\param mmax is the maximal number of previous vectors which are orthogonalized against the new search direction. |
| 1537 |
*/ |
| 1538 |
RestartedFCGSolver (std::shared_ptr<const LinearOperator<X,X>> op, |
| 1539 |
std::shared_ptr<const ScalarProduct<X>> sp, |
| 1540 |
std::shared_ptr<Preconditioner<X,X>> prec, |
| 1541 |
scalar_real_type reduction, int maxit, int verbose, |
| 1542 |
int mmax = 10) |
| 1543 |
: IterativeSolver<X,X>(op, sp, prec, reduction, maxit, verbose), _mmax(mmax) |
| 1544 |
{} |
| 1545 |
|
| 1546 |
/*! |
| 1547 |
\brief Constructor to initialize a RestartedFCG solver. |
| 1548 |
|
| 1549 |
\copydoc IterativeSolver::IterativeSolver(std::shared_ptr<const LinearOperator<X,Y>>,std::shared_ptr<Preconditioner<X,X>>,const ParameterTree&) |
| 1550 |
|
| 1551 |
Additional parameter: |
| 1552 |
ParameterTree Key | Meaning |
| 1553 |
------------------|------------ |
| 1554 |
mmax | number of FCG cycles before restart. default=10 |
| 1555 |
|
| 1556 |
See \ref ISTL_Factory for the ParameterTree layout and examples. |
| 1557 |
*/ |
| 1558 |
RestartedFCGSolver (std::shared_ptr<const LinearOperator<X,X>> op, |
| 1559 |
std::shared_ptr<Preconditioner<X,X>> prec, |
| 1560 |
const ParameterTree& configuration) |
| 1561 |
: IterativeSolver<X,X>(op, prec, configuration), _mmax(configuration.get("mmax", 10)) |
| 1562 |
{} |
| 1563 |
|
| 1564 |
RestartedFCGSolver (std::shared_ptr<const LinearOperator<X,X>> op, |
| 1565 |
std::shared_ptr<const ScalarProduct<X>> sp, |
| 1566 |
std::shared_ptr<Preconditioner<X,X>> prec, |
| 1567 |
const ParameterTree& config) |
| 1568 |
: IterativeSolver<X,X>(op, sp, prec, config), _mmax(config.get("mmax", 10)) |
| 1569 |
{} |
| 1570 |
|
| 1571 |
/*! |
| 1572 |
\brief Apply inverse operator. |
| 1573 |
|
| 1574 |
\copydoc InverseOperator::apply(X&,Y&,InverseOperatorResult&) |
| 1575 |
|
| 1576 |
\note Currently, the RestartedFCGSolver aborts when a NaN or infinite defect is |
| 1577 |
detected. However, -ffinite-math-only (implied by -ffast-math) |
| 1578 |
can inhibit a result from becoming NaN that really should be NaN. |
| 1579 |
E.g. numeric_limits<double>::quiet_NaN()*0.0==0.0 with gcc-5.3 |
| 1580 |
-ffast-math. |
| 1581 |
*/ |
| 1582 |
|
| 1583 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 1584 |
{ |
| 1585 |
using rAlloc = ReboundAllocatorType<X,field_type>; |
| 1586 |
res.clear(); |
| 1587 |
Iteration iteration(*this,res); |
| 1588 |
_prec->pre(x,b); // prepare preconditioner |
| 1589 |
_op->applyscaleadd(-1,x,b); // overwrite b with defect |
| 1590 |
|
| 1591 |
//arrays for interim values: |
| 1592 |
std::vector<X> d(_mmax+1, x); // array for directions |
| 1593 |
std::vector<X> Ad(_mmax+1, x); // array for Ad[i] |
| 1594 |
std::vector<field_type,rAlloc> ddotAd(_mmax+1,0); // array for <d[i],Ad[i]> |
| 1595 |
X w(x); |
| 1596 |
|
| 1597 |
real_type def = _sp->norm(b); // compute norm |
| 1598 |
if(iteration.step(0, def)){ |
| 1599 |
_prec->post(x); |
| 1600 |
return; |
| 1601 |
} |
| 1602 |
|
| 1603 |
// some local variables |
| 1604 |
field_type alpha; |
| 1605 |
|
| 1606 |
// the loop |
| 1607 |
int i=1; |
| 1608 |
int i_bounded=0; |
| 1609 |
while(i<=_maxit && !res.converged) { |
| 1610 |
for (; i_bounded <= _mmax && i<= _maxit; i_bounded++) { |
| 1611 |
d[i_bounded] = 0; // reset search direction |
| 1612 |
_prec->apply(d[i_bounded], b); // apply preconditioner |
| 1613 |
w = d[i_bounded]; // copy of current d[i] |
| 1614 |
// orthogonalization with previous directions |
| 1615 |
orthogonalizations(i_bounded,Ad,w,ddotAd,d); |
| 1616 |
|
| 1617 |
//saving interim values for future calculating |
| 1618 |
_op->apply(d[i_bounded], Ad[i_bounded]); // save Ad[i] |
| 1619 |
ddotAd[i_bounded]=_sp->dot(d[i_bounded],Ad[i_bounded]); // save <d[i],Ad[i]> |
| 1620 |
alpha = _sp->dot(d[i_bounded], b)/ddotAd[i_bounded]; // <d[i],b>/<d[i],Ad[i]> |
| 1621 |
|
| 1622 |
//update solution and defect |
| 1623 |
x.axpy(alpha, d[i_bounded]); |
| 1624 |
b.axpy(-alpha, Ad[i_bounded]); |
| 1625 |
|
| 1626 |
// convergence test |
| 1627 |
def = _sp->norm(b); // comp defect norm |
| 1628 |
|
| 1629 |
iteration.step(i, def); |
| 1630 |
i++; |
| 1631 |
} |
| 1632 |
//restart: exchange first and last stored values |
| 1633 |
cycle(Ad,d,ddotAd,i_bounded); |
| 1634 |
} |
| 1635 |
|
| 1636 |
//correct i which is wrong if convergence was not achieved. |
| 1637 |
i=std::min(_maxit,i); |
| 1638 |
|
| 1639 |
_prec->post(x); // postprocess preconditioner |
| 1640 |
} |
| 1641 |
|
| 1642 |
private: |
| 1643 |
//This function is called every iteration to orthogonalize against the last search directions |
| 1644 |
virtual void orthogonalizations(const int& i_bounded,const std::vector<X>& Ad, const X& w, const std::vector<field_type,ReboundAllocatorType<X,field_type>>& ddotAd,std::vector<X>& d) { |
| 1645 |
// The RestartedFCGSolver uses only values with lower array index; |
| 1646 |
for (int k = 0; k < i_bounded; k++) { |
| 1647 |
d[i_bounded].axpy(-_sp->dot(Ad[k], w) / ddotAd[k], d[k]); // d[i] -= <<Ad[k],w>/<d[k],Ad[k]>>d[k] |
| 1648 |
} |
| 1649 |
} |
| 1650 |
|
| 1651 |
// This function is called every mmax iterations to handle limited array sizes. |
| 1652 |
virtual void cycle(std::vector<X>& Ad,std::vector<X>& d,std::vector<field_type,ReboundAllocatorType<X,field_type> >& ddotAd,int& i_bounded) { |
| 1653 |
// Reset loop index and exchange the first and last arrays |
| 1654 |
i_bounded = 1; |
| 1655 |
std::swap(Ad[0], Ad[_mmax]); |
| 1656 |
std::swap(d[0], d[_mmax]); |
| 1657 |
std::swap(ddotAd[0], ddotAd[_mmax]); |
| 1658 |
} |
| 1659 |
|
| 1660 |
protected: |
| 1661 |
int _mmax; |
| 1662 |
using IterativeSolver<X,X>::_op; |
| 1663 |
using IterativeSolver<X,X>::_prec; |
| 1664 |
using IterativeSolver<X,X>::_sp; |
| 1665 |
using IterativeSolver<X,X>::_reduction; |
| 1666 |
using IterativeSolver<X,X>::_maxit; |
| 1667 |
using IterativeSolver<X,X>::_verbose; |
| 1668 |
using Iteration = typename IterativeSolver<X,X>::template Iteration<unsigned int>; |
| 1669 |
}; |
| 1670 |
DUNE_REGISTER_SOLVER("restartedfcgsolver", defaultIterativeSolverCreator<Dune::RestartedFCGSolver>()); |
| 1671 |
|
| 1672 |
/*! \brief Complete flexible conjugate gradient method |
| 1673 |
|
| 1674 |
This solver is a simple modification of the RestartedFCGSolver and, if possible, uses mmax old directions. |
| 1675 |
It uses noticeably more memory, but provides more stability for preconditioner changes. |
| 1676 |
|
| 1677 |
*/ |
| 1678 |
template<class X> |
| 1679 |
class CompleteFCGSolver : public RestartedFCGSolver<X> { |
| 1680 |
public: |
| 1681 |
using typename RestartedFCGSolver<X>::domain_type; |
| 1682 |
using typename RestartedFCGSolver<X>::range_type; |
| 1683 |
using typename RestartedFCGSolver<X>::field_type; |
| 1684 |
using typename RestartedFCGSolver<X>::real_type; |
| 1685 |
|
| 1686 |
// copy base class constructors |
| 1687 |
using RestartedFCGSolver<X>::RestartedFCGSolver; |
| 1688 |
|
| 1689 |
// don't shadow four-argument version of apply defined in the base class |
| 1690 |
using RestartedFCGSolver<X>::apply; |
| 1691 |
|
| 1692 |
// just a minor part of the RestartedFCGSolver apply method will be modified |
| 1693 |
void apply (X& x, X& b, InverseOperatorResult& res) override |
| 1694 |
{ |
| 1695 |
// reset limiter of orthogonalization loop |
| 1696 |
_k_limit = 0; |
| 1697 |
this->RestartedFCGSolver<X>::apply(x,b,res); |
| 1698 |
}; |
| 1699 |
|
| 1700 |
private: |
| 1701 |
// This function is called every iteration to orthogonalize against the last search directions. |
| 1702 |
void orthogonalizations(const int& i_bounded,const std::vector<X>& Ad, const X& w, const std::vector<field_type,ReboundAllocatorType<X,field_type>>& ddotAd,std::vector<X>& d) override { |
| 1703 |
// This FCGSolver uses values with higher array indexes too, if existent. |
| 1704 |
for (int k = 0; k < _k_limit; k++) { |
| 1705 |
if(i_bounded!=k) |
| 1706 |
d[i_bounded].axpy(-_sp->dot(Ad[k], w) / ddotAd[k], d[k]); // d[i] -= <<Ad[k],w>/<d[k],Ad[k]>>d[k] |
| 1707 |
} |
| 1708 |
// The loop limit increase, if array is not completely filled. |
| 1709 |
if(_k_limit<=i_bounded) |
| 1710 |
_k_limit++; |
| 1711 |
|
| 1712 |
}; |
| 1713 |
|
| 1714 |
// This function is called every mmax iterations to handle limited array sizes. |
| 1715 |
void cycle(std::vector<X>& Ad, [[maybe_unused]] std::vector<X>& d, [[maybe_unused]] std::vector<field_type,ReboundAllocatorType<X,field_type> >& ddotAd,int& i_bounded) override { |
| 1716 |
// Only the loop index i_bounded return to 0, if it reached mmax. |
| 1717 |
i_bounded = 0; |
| 1718 |
// Now all arrays are filled and the loop in void orthogonalizations can use the whole arrays. |
| 1719 |
_k_limit = Ad.size(); |
| 1720 |
}; |
| 1721 |
|
| 1722 |
int _k_limit = 0; |
| 1723 |
|
| 1724 |
protected: |
| 1725 |
using RestartedFCGSolver<X>::_mmax; |
| 1726 |
using RestartedFCGSolver<X>::_op; |
| 1727 |
using RestartedFCGSolver<X>::_prec; |
| 1728 |
using RestartedFCGSolver<X>::_sp; |
| 1729 |
using RestartedFCGSolver<X>::_reduction; |
| 1730 |
using RestartedFCGSolver<X>::_maxit; |
| 1731 |
using RestartedFCGSolver<X>::_verbose; |
| 1732 |
}; |
| 1733 |
DUNE_REGISTER_SOLVER("completefcgsolver", defaultIterativeSolverCreator<Dune::CompleteFCGSolver>()); |
| 1734 |
/** @} end documentation */ |
| 1735 |
} // end namespace |
| 1736 |
|
| 1737 |
#endif |