| 1 |
// ------------------------------------------------------------------------ |
| 2 |
// |
| 3 |
// SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 |
// Copyright (C) 1999 - 2025 by the deal.II authors |
| 5 |
// |
| 6 |
// This file is part of the deal.II library. |
| 7 |
// |
| 8 |
// Part of the source code is dual licensed under Apache-2.0 WITH |
| 9 |
// LLVM-exception OR LGPL-2.1-or-later. Detailed license information |
| 10 |
// governing the source code and code contributions can be found in |
| 11 |
// LICENSE.md and CONTRIBUTING.md at the top level directory of deal.II. |
| 12 |
// |
| 13 |
// ------------------------------------------------------------------------ |
| 14 |
|
| 15 |
#ifndef dealii_precondition_h |
| 16 |
#define dealii_precondition_h |
| 17 |
|
| 18 |
#include <deal.II/base/config.h> |
| 19 |
|
| 20 |
#include <deal.II/base/memory_space.h> |
| 21 |
#include <deal.II/base/mutex.h> |
| 22 |
#include <deal.II/base/observer_pointer.h> |
| 23 |
#include <deal.II/base/parallel.h> |
| 24 |
#include <deal.II/base/template_constraints.h> |
| 25 |
|
| 26 |
#include <deal.II/lac/affine_constraints.h> |
| 27 |
#include <deal.II/lac/block_vector_base.h> |
| 28 |
#include <deal.II/lac/diagonal_matrix.h> |
| 29 |
#include <deal.II/lac/identity_matrix.h> |
| 30 |
#include <deal.II/lac/solver_cg.h> |
| 31 |
#include <deal.II/lac/vector_memory.h> |
| 32 |
|
| 33 |
#include <Kokkos_Core.hpp> |
| 34 |
|
| 35 |
#include <limits> |
| 36 |
|
| 37 |
DEAL_II_NAMESPACE_OPEN |
| 38 |
|
| 39 |
// forward declarations |
| 40 |
#ifndef DOXYGEN |
| 41 |
template <typename number> |
| 42 |
class Vector; |
| 43 |
template <typename number> |
| 44 |
class SparseMatrix; |
| 45 |
namespace LinearAlgebra |
| 46 |
{ |
| 47 |
namespace distributed |
| 48 |
{ |
| 49 |
template <typename, typename> |
| 50 |
class Vector; |
| 51 |
template <typename, typename> |
| 52 |
class BlockVector; |
| 53 |
} // namespace distributed |
| 54 |
} // namespace LinearAlgebra |
| 55 |
#endif |
| 56 |
|
| 57 |
|
| 58 |
/** |
| 59 |
* @addtogroup Preconditioners |
| 60 |
* @{ |
| 61 |
*/ |
| 62 |
|
| 63 |
namespace internal |
| 64 |
{ |
| 65 |
/** |
| 66 |
* An enum to define the available types of eigenvalue estimation |
| 67 |
* algorithms. |
| 68 |
*/ |
| 69 |
enum class EigenvalueAlgorithm |
| 70 |
{ |
| 71 |
/** |
| 72 |
* This option runs the conjugate gradient solver and computes an |
| 73 |
* eigenvalue estimation from the underlying Lanczos space. This only |
| 74 |
* works for symmetric positive definite matrices. |
| 75 |
*/ |
| 76 |
lanczos, |
| 77 |
/** |
| 78 |
* This option runs a power iteration to estimate the largest |
| 79 |
* eigenvalue. This algorithm also works for non-symmetric matrices, |
| 80 |
* but typically gives less accurate estimates than the option 'lanczos' |
| 81 |
* for the same number of iterations because it does not take the |
| 82 |
* relation between vectors in the iterations into account |
| 83 |
* (roughly speaking the off-diagonal entries in the |
| 84 |
* tri-diagonal matrix of the Lanczos iteration). |
| 85 |
*/ |
| 86 |
power_iteration |
| 87 |
}; |
| 88 |
|
| 89 |
/** |
| 90 |
* A struct that contains information about the eigenvalue estimation |
| 91 |
* performed by the PreconditionChebyshev class. |
| 92 |
*/ |
| 93 |
struct EigenvalueInformation |
| 94 |
{ |
| 95 |
/** |
| 96 |
* Estimate for the smallest eigenvalue. |
| 97 |
*/ |
| 98 |
double min_eigenvalue_estimate; |
| 99 |
/** |
| 100 |
* Estimate for the largest eigenvalue. |
| 101 |
*/ |
| 102 |
double max_eigenvalue_estimate; |
| 103 |
/** |
| 104 |
* Number of CG iterations performed or 0. |
| 105 |
*/ |
| 106 |
unsigned int cg_iterations; |
| 107 |
/** |
| 108 |
* The degree of the Chebyshev polynomial (either as set using |
| 109 |
* AdditionalData::degree or estimated as described there). |
| 110 |
*/ |
| 111 |
unsigned int degree; |
| 112 |
/** |
| 113 |
* Constructor initializing with invalid values. |
| 114 |
*/ |
| 115 |
EigenvalueInformation() |
| 116 |
: min_eigenvalue_estimate{std::numeric_limits<double>::max()} |
| 117 |
, max_eigenvalue_estimate{std::numeric_limits<double>::lowest()} |
| 118 |
, cg_iterations{0} |
| 119 |
, degree{0} |
| 120 |
{} |
| 121 |
}; |
| 122 |
|
| 123 |
/** |
| 124 |
* Base class for PreconditionRelaxation::AdditionalData and |
| 125 |
* PreconditionChebyshev::AdditionalData with relevant information |
| 126 |
* for estimating the eigenvalues. |
| 127 |
*/ |
| 128 |
template <typename PreconditionerType> |
| 129 |
struct EigenvalueAlgorithmAdditionalData |
| 130 |
{ |
| 131 |
/** |
| 132 |
* Constructor. |
| 133 |
*/ |
| 134 |
EigenvalueAlgorithmAdditionalData( |
| 135 |
const double smoothing_range, |
| 136 |
const unsigned int eig_cg_n_iterations, |
| 137 |
const double eig_cg_residual, |
| 138 |
const double max_eigenvalue, |
| 139 |
const EigenvalueAlgorithm eigenvalue_algorithm); |
| 140 |
|
| 141 |
/** |
| 142 |
* Copy assignment operator. |
| 143 |
*/ |
| 144 |
EigenvalueAlgorithmAdditionalData<PreconditionerType> & |
| 145 |
operator=( |
| 146 |
const EigenvalueAlgorithmAdditionalData<PreconditionerType> &other_data); |
| 147 |
|
| 148 |
/** |
| 149 |
* This sets the range between the largest eigenvalue in the matrix and |
| 150 |
* the smallest eigenvalue to be treated. If the parameter is set to a |
| 151 |
* number less than 1, an estimate for the largest and for the smallest |
| 152 |
* eigenvalue will be calculated internally. For a smoothing range larger |
| 153 |
* than one, the preconditioner will act in the interval |
| 154 |
* $[\lambda_\mathrm{max}/ \tt{smoothing\_range}, \lambda_\mathrm{max}]$, |
| 155 |
* where $\lambda_\mathrm{max}$ is an estimate of the maximum eigenvalue |
| 156 |
* of the matrix. A choice of <tt>smoothing_range</tt> between 5 and 20 is |
| 157 |
* useful in case the preconditioner is used as a smoother in multigrid. |
| 158 |
*/ |
| 159 |
double smoothing_range; |
| 160 |
|
| 161 |
/** |
| 162 |
* Maximum number of CG iterations performed for finding the maximum |
| 163 |
* eigenvalue. If set to zero, no computations are performed. Instead, the |
| 164 |
* user must supply a largest eigenvalue via the variable |
| 165 |
* PreconditionRelaxation::AdditionalData::max_eigenvalue. |
| 166 |
*/ |
| 167 |
unsigned int eig_cg_n_iterations; |
| 168 |
|
| 169 |
/** |
| 170 |
* Tolerance for iterations performed for finding the maximum |
| 171 |
* eigenvalue by the eigenvalue algorithm (Lanczos or power iteration). |
| 172 |
*/ |
| 173 |
double eig_cg_residual; |
| 174 |
|
| 175 |
/** |
| 176 |
* Maximum eigenvalue to work with. Only in effect if @p |
| 177 |
* eig_cg_n_iterations is set to zero, otherwise this parameter is |
| 178 |
* ignored. |
| 179 |
*/ |
| 180 |
double max_eigenvalue; |
| 181 |
|
| 182 |
/** |
| 183 |
* Constraints to be used for the operator given. This variable is used to |
| 184 |
* zero out the correct entries when creating an initial guess. |
| 185 |
*/ |
| 186 |
dealii::AffineConstraints<double> constraints; |
| 187 |
|
| 188 |
/** |
| 189 |
* Stores the preconditioner object that the Chebyshev is wrapped around. |
| 190 |
*/ |
| 191 |
EigenvalueAlgorithm eigenvalue_algorithm; |
| 192 |
|
| 193 |
/** |
| 194 |
* Preconditioner. |
| 195 |
*/ |
| 196 |
std::shared_ptr<PreconditionerType> preconditioner; |
| 197 |
}; |
| 198 |
} // namespace internal |
| 199 |
|
| 200 |
|
| 201 |
/** |
| 202 |
* This preconditioner implements the identity operator. |
| 203 |
* This class can be used you if you want to use a linear |
| 204 |
* solver without preconditioning, because all linear solvers |
| 205 |
* require the usage of a preconditioner. |
| 206 |
* |
| 207 |
* It can be used in the following way: |
| 208 |
* |
| 209 |
* @code |
| 210 |
* SolverControl solver_control (1000, 1e-12); |
| 211 |
* SolverCG<> cg (solver_control); |
| 212 |
* cg.solve (system_matrix, solution, system_rhs, PreconditionIdentity()); |
| 213 |
* @endcode |
| 214 |
* |
| 215 |
* See the step-3 tutorial program for an example and additional explanations. |
| 216 |
* |
| 217 |
* Alternatively, the IdentityMatrix class can be used to precondition in this |
| 218 |
* way. |
| 219 |
*/ |
| 220 |
class PreconditionIdentity : public EnableObserverPointer |
| 221 |
{ |
| 222 |
public: |
| 223 |
/** |
| 224 |
* Declare type for container size. |
| 225 |
*/ |
| 226 |
using size_type = types::global_dof_index; |
| 227 |
|
| 228 |
/** |
| 229 |
* This function is only present to provide the interface of a |
| 230 |
* preconditioner to be handed to a smoother. This does nothing. |
| 231 |
*/ |
| 232 |
struct AdditionalData |
| 233 |
{ |
| 234 |
/** |
| 235 |
* Constructor. |
| 236 |
*/ |
| 237 |
AdditionalData() = default; |
| 238 |
}; |
| 239 |
|
| 240 |
/** |
| 241 |
* Constructor, sets the domain and range sizes to their defaults. |
| 242 |
*/ |
| 243 |
PreconditionIdentity(); |
| 244 |
|
| 245 |
/** |
| 246 |
* The matrix argument is ignored and here just for compatibility with more |
| 247 |
* complex preconditioners. |
| 248 |
*/ |
| 249 |
template <typename MatrixType> |
| 250 |
void |
| 251 |
initialize(const MatrixType &matrix, |
| 252 |
const AdditionalData &additional_data = AdditionalData()); |
| 253 |
|
| 254 |
/** |
| 255 |
* Apply preconditioner. |
| 256 |
*/ |
| 257 |
template <typename VectorType> |
| 258 |
void |
| 259 |
vmult(VectorType &, const VectorType &) const; |
| 260 |
|
| 261 |
/** |
| 262 |
* Apply transpose preconditioner. Since this is the identity, this function |
| 263 |
* is the same as vmult(). |
| 264 |
*/ |
| 265 |
template <typename VectorType> |
| 266 |
void |
| 267 |
Tvmult(VectorType &, const VectorType &) const; |
| 268 |
|
| 269 |
/** |
| 270 |
* Apply preconditioner, adding to the previous value. |
| 271 |
*/ |
| 272 |
template <typename VectorType> |
| 273 |
void |
| 274 |
vmult_add(VectorType &, const VectorType &) const; |
| 275 |
|
| 276 |
/** |
| 277 |
* Apply transpose preconditioner, adding. Since this is the identity, this |
| 278 |
* function is the same as vmult_add(). |
| 279 |
*/ |
| 280 |
template <typename VectorType> |
| 281 |
void |
| 282 |
Tvmult_add(VectorType &, const VectorType &) const; |
| 283 |
|
| 284 |
/** |
| 285 |
* This function is only present to provide the interface of a |
| 286 |
* preconditioner to be handed to a smoother. This does nothing. |
| 287 |
*/ |
| 288 |
void |
| 289 |
clear(); |
| 290 |
|
| 291 |
/** |
| 292 |
* Return the dimension of the codomain (or range) space. Note that the |
| 293 |
* matrix is of dimension $m \times n$. |
| 294 |
* |
| 295 |
* @note This function should only be called if the preconditioner has been |
| 296 |
* initialized. |
| 297 |
*/ |
| 298 |
size_type |
| 299 |
m() const; |
| 300 |
|
| 301 |
/** |
| 302 |
* Return the dimension of the domain space. Note that the matrix is of |
| 303 |
* dimension $m \times n$. |
| 304 |
* |
| 305 |
* @note This function should only be called if the preconditioner has been |
| 306 |
* initialized. |
| 307 |
*/ |
| 308 |
size_type |
| 309 |
n() const; |
| 310 |
|
| 311 |
private: |
| 312 |
/** |
| 313 |
* The dimension of the range space. |
| 314 |
*/ |
| 315 |
size_type n_rows; |
| 316 |
|
| 317 |
/** |
| 318 |
* The dimension of the domain space. |
| 319 |
*/ |
| 320 |
size_type n_columns; |
| 321 |
}; |
| 322 |
|
| 323 |
|
| 324 |
|
| 325 |
/** |
| 326 |
* Preconditioning with Richardson's method. This preconditioner just scales |
| 327 |
* the vector with a constant relaxation factor provided by the AdditionalData |
| 328 |
* object. |
| 329 |
* |
| 330 |
* In Krylov-space methods, this preconditioner should not have any effect. |
| 331 |
* Using SolverRichardson, the two relaxation parameters will be just |
| 332 |
* multiplied. Still, this class is useful in multigrid smoother objects |
| 333 |
* (MGSmootherRelaxation). |
| 334 |
*/ |
| 335 |
class PreconditionRichardson : public EnableObserverPointer |
| 336 |
{ |
| 337 |
public: |
| 338 |
/** |
| 339 |
* Declare type for container size. |
| 340 |
*/ |
| 341 |
using size_type = types::global_dof_index; |
| 342 |
|
| 343 |
/** |
| 344 |
* Parameters for Richardson preconditioner. |
| 345 |
*/ |
| 346 |
class AdditionalData |
| 347 |
{ |
| 348 |
public: |
| 349 |
/** |
| 350 |
* Constructor. Block size must be given since there is no reasonable |
| 351 |
* default parameter. |
| 352 |
*/ |
| 353 |
AdditionalData(const double relaxation = 1.); |
| 354 |
|
| 355 |
/** |
| 356 |
* Relaxation parameter. |
| 357 |
*/ |
| 358 |
double relaxation; |
| 359 |
}; |
| 360 |
|
| 361 |
/** |
| 362 |
* Constructor, sets the relaxation parameter, domain and range sizes to |
| 363 |
* their default. |
| 364 |
*/ |
| 365 |
PreconditionRichardson(); |
| 366 |
|
| 367 |
/** |
| 368 |
* Change the relaxation parameter. |
| 369 |
*/ |
| 370 |
void |
| 371 |
initialize(const AdditionalData ¶meters); |
| 372 |
|
| 373 |
/** |
| 374 |
* Change the relaxation parameter in a way consistent with other |
| 375 |
* preconditioners. The matrix argument is ignored and here just for |
| 376 |
* compatibility with more complex preconditioners. |
| 377 |
*/ |
| 378 |
template <typename MatrixType> |
| 379 |
void |
| 380 |
initialize(const MatrixType &matrix, const AdditionalData ¶meters); |
| 381 |
|
| 382 |
/** |
| 383 |
* Apply preconditioner. |
| 384 |
*/ |
| 385 |
template <typename VectorType> |
| 386 |
void |
| 387 |
vmult(VectorType &, const VectorType &) const; |
| 388 |
|
| 389 |
/** |
| 390 |
* Apply transpose preconditioner. Since this is the identity, this function |
| 391 |
* is the same as vmult(). |
| 392 |
*/ |
| 393 |
template <typename VectorType> |
| 394 |
void |
| 395 |
Tvmult(VectorType &, const VectorType &) const; |
| 396 |
/** |
| 397 |
* Apply preconditioner, adding to the previous value. |
| 398 |
*/ |
| 399 |
template <typename VectorType> |
| 400 |
void |
| 401 |
vmult_add(VectorType &, const VectorType &) const; |
| 402 |
|
| 403 |
/** |
| 404 |
* Apply transpose preconditioner, adding. Since this is the identity, this |
| 405 |
* function is the same as vmult_add(). |
| 406 |
*/ |
| 407 |
template <typename VectorType> |
| 408 |
void |
| 409 |
Tvmult_add(VectorType &, const VectorType &) const; |
| 410 |
|
| 411 |
/** |
| 412 |
* This function is only present to provide the interface of a |
| 413 |
* preconditioner to be handed to a smoother. This does nothing. |
| 414 |
*/ |
| 415 |
void |
| 416 |
clear() |
| 417 |
{} |
| 418 |
|
| 419 |
/** |
| 420 |
* Return the dimension of the codomain (or range) space. Note that the |
| 421 |
* matrix is of dimension $m \times n$. |
| 422 |
* |
| 423 |
* @note This function should only be called if the preconditioner has been |
| 424 |
* initialized. |
| 425 |
*/ |
| 426 |
size_type |
| 427 |
m() const; |
| 428 |
|
| 429 |
/** |
| 430 |
* Return the dimension of the domain space. Note that the matrix is of |
| 431 |
* dimension $m \times n$. |
| 432 |
* |
| 433 |
* @note This function should only be called if the preconditioner has been |
| 434 |
* initialized. |
| 435 |
*/ |
| 436 |
size_type |
| 437 |
n() const; |
| 438 |
|
| 439 |
private: |
| 440 |
/** |
| 441 |
* The relaxation parameter multiplied with the vectors. |
| 442 |
*/ |
| 443 |
double relaxation; |
| 444 |
|
| 445 |
/** |
| 446 |
* The dimension of the range space. |
| 447 |
*/ |
| 448 |
size_type n_rows; |
| 449 |
|
| 450 |
/** |
| 451 |
* The dimension of the domain space. |
| 452 |
*/ |
| 453 |
size_type n_columns; |
| 454 |
}; |
| 455 |
|
| 456 |
|
| 457 |
|
| 458 |
/** |
| 459 |
* Preconditioner using a matrix-builtin function. This class forms a |
| 460 |
* preconditioner suitable for the LAC solver classes. Since many |
| 461 |
* preconditioning methods are based on matrix entries, these have to be |
| 462 |
* implemented as member functions of the underlying matrix implementation. |
| 463 |
* This class now is intended to allow easy access to these member functions |
| 464 |
* from LAC solver classes. |
| 465 |
* |
| 466 |
* It seems that all builtin preconditioners have a relaxation parameter, so |
| 467 |
* please use PreconditionRelaxation for these. |
| 468 |
* |
| 469 |
* You will usually not want to create a named object of this type, although |
| 470 |
* possible. The most common use is like this: |
| 471 |
* @code |
| 472 |
* SolverGMRES<SparseMatrix<double> Vector<double>> gmres(control,memory,500); |
| 473 |
* |
| 474 |
* gmres.solve( |
| 475 |
* matrix, solution, right_hand_side, |
| 476 |
* PreconditionUseMatrix<SparseMatrix<double>,Vector<double> >( |
| 477 |
* matrix, &SparseMatrix<double>::template precondition_Jacobi<double>)); |
| 478 |
* @endcode |
| 479 |
* This creates an unnamed object to be passed as the fourth parameter to the |
| 480 |
* solver function of the SolverGMRES class. It assumes that the SparseMatrix |
| 481 |
* class has a function <tt>precondition_Jacobi</tt> taking two vectors |
| 482 |
* (source and destination) as parameters (Actually, there is no function like |
| 483 |
* that, the existing function takes a third parameter, denoting the |
| 484 |
* relaxation parameter; this example is therefore only meant to illustrate |
| 485 |
* the general idea). |
| 486 |
* |
| 487 |
* Note that due to the default template parameters, the above example could |
| 488 |
* be written shorter as follows: |
| 489 |
* @code |
| 490 |
* ... |
| 491 |
* gmres.solve( |
| 492 |
* matrix, solution, right_hand_side, |
| 493 |
* PreconditionUseMatrix<>( |
| 494 |
* matrix,&SparseMatrix<double>::template precondition_Jacobi<double>)); |
| 495 |
* @endcode |
| 496 |
*/ |
| 497 |
template <typename MatrixType = SparseMatrix<double>, |
| 498 |
typename VectorType = Vector<double>> |
| 499 |
class PreconditionUseMatrix : public EnableObserverPointer |
| 500 |
{ |
| 501 |
public: |
| 502 |
/** |
| 503 |
* Type of the preconditioning function of the matrix. |
| 504 |
*/ |
| 505 |
using function_ptr = void (MatrixType::*)(VectorType &, |
| 506 |
const VectorType &) const; |
| 507 |
|
| 508 |
/** |
| 509 |
* Constructor. This constructor stores a reference to the matrix object |
| 510 |
* for later use and selects a preconditioning method, which must be a |
| 511 |
* member function of that matrix. |
| 512 |
*/ |
| 513 |
PreconditionUseMatrix(const MatrixType &M, const function_ptr method); |
| 514 |
|
| 515 |
/** |
| 516 |
* Execute preconditioning. Calls the function passed to the constructor of |
| 517 |
* this object with the two arguments given here. |
| 518 |
*/ |
| 519 |
void |
| 520 |
vmult(VectorType &dst, const VectorType &src) const; |
| 521 |
|
| 522 |
private: |
| 523 |
/** |
| 524 |
* Pointer to the matrix in use. |
| 525 |
*/ |
| 526 |
const MatrixType &matrix; |
| 527 |
|
| 528 |
/** |
| 529 |
* Pointer to the preconditioning function. |
| 530 |
*/ |
| 531 |
const function_ptr precondition; |
| 532 |
}; |
| 533 |
|
| 534 |
|
| 535 |
|
| 536 |
/** |
| 537 |
* Base class for other preconditioners. Here, only some common features |
| 538 |
* Jacobi, SOR and SSOR preconditioners are implemented. For preconditioning, |
| 539 |
* refer to derived classes. |
| 540 |
* |
| 541 |
* One iteration is given as: |
| 542 |
* @f[ |
| 543 |
* x^{n+1} = x^{n} + \alpha P^{-1} (b-Ax^n). |
| 544 |
* @f] |
| 545 |
* |
| 546 |
* The relaxation parameter $\alpha$ has to be in the range: |
| 547 |
* @f[ |
| 548 |
* 0 < \alpha < \frac{2}{\lambda_{\max}(P^{-1}A)}. |
| 549 |
* @f] |
| 550 |
* Its theoretically optimal value is given by: |
| 551 |
* @f[ |
| 552 |
* \alpha := \frac{2}{\lambda_{\min}(P^{-1}A)+\lambda_{\max}(P^{-1}A)}. |
| 553 |
* @f] |
| 554 |
* |
| 555 |
* For details on the algorithm, see @cite Varga2009. |
| 556 |
* |
| 557 |
* The relaxation parameter can be set manually or can be automatically |
| 558 |
* determined, for which we use the theoretically optimal value. |
| 559 |
* For this purpose, the user needs to set the relaxation parameter |
| 560 |
* to zero. Internally, the minimum and maximum eigenvalues of the |
| 561 |
* preconditioned system are estimated by an eigenvalue algorithm, and the |
| 562 |
* resulting estimate is multiplied by the 1.2 for safety reasons. For more |
| 563 |
* details on the underlying algorithms, see PreconditionChebyshev. |
| 564 |
*/ |
| 565 |
template <typename MatrixType = SparseMatrix<double>, |
| 566 |
typename PreconditionerType = IdentityMatrix> |
| 567 |
class PreconditionRelaxation : public EnableObserverPointer |
| 568 |
{ |
| 569 |
public: |
| 570 |
/** |
| 571 |
* Declare type for container size. |
| 572 |
*/ |
| 573 |
using size_type = types::global_dof_index; |
| 574 |
|
| 575 |
/** |
| 576 |
* Class for parameters. |
| 577 |
*/ |
| 578 |
class AdditionalData |
| 579 |
: public internal::EigenvalueAlgorithmAdditionalData<PreconditionerType> |
| 580 |
{ |
| 581 |
public: |
| 582 |
using EigenvalueAlgorithm = internal::EigenvalueAlgorithm; |
| 583 |
|
| 584 |
/** |
| 585 |
* Constructor. |
| 586 |
*/ |
| 587 |
AdditionalData(const double relaxation = 1., |
| 588 |
const unsigned int n_iterations = 1, |
| 589 |
const double smoothing_range = 0., |
| 590 |
const unsigned int eig_cg_n_iterations = 8, |
| 591 |
const double eig_cg_residual = 1e-2, |
| 592 |
const double max_eigenvalue = 1, |
| 593 |
const EigenvalueAlgorithm eigenvalue_algorithm = |
| 594 |
EigenvalueAlgorithm::lanczos); |
| 595 |
|
| 596 |
/** |
| 597 |
* Relaxation parameter. |
| 598 |
*/ |
| 599 |
double relaxation; |
| 600 |
|
| 601 |
/** |
| 602 |
* Number of smoothing steps to be performed in an |
| 603 |
* invocation of vmult() or step(). |
| 604 |
*/ |
| 605 |
unsigned int n_iterations; |
| 606 |
}; |
| 607 |
|
| 608 |
/** |
| 609 |
* Initialize matrix and relaxation parameter. The matrix is just stored in |
| 610 |
* the preconditioner object. The relaxation parameter should be larger than |
| 611 |
* zero and smaller than 2 for numerical reasons. It defaults to 1. |
| 612 |
*/ |
| 613 |
void |
| 614 |
initialize(const MatrixType &A, |
| 615 |
const AdditionalData ¶meters = AdditionalData()); |
| 616 |
|
| 617 |
/** |
| 618 |
* Release the matrix and reset its pointer. |
| 619 |
*/ |
| 620 |
void |
| 621 |
clear(); |
| 622 |
|
| 623 |
/** |
| 624 |
* Return the dimension of the codomain (or range) space. Note that the |
| 625 |
* matrix is of dimension $m \times n$. |
| 626 |
*/ |
| 627 |
size_type |
| 628 |
m() const; |
| 629 |
|
| 630 |
/** |
| 631 |
* Return the dimension of the domain space. Note that the matrix is of |
| 632 |
* dimension $m \times n$. |
| 633 |
*/ |
| 634 |
size_type |
| 635 |
n() const; |
| 636 |
|
| 637 |
/** |
| 638 |
* Apply preconditioner. |
| 639 |
*/ |
| 640 |
template <typename VectorType> |
| 641 |
void |
| 642 |
vmult(VectorType &, const VectorType &) const; |
| 643 |
|
| 644 |
/** |
| 645 |
* Apply transpose preconditioner. Since this is a symmetric preconditioner, |
| 646 |
* this function is the same as vmult(). |
| 647 |
*/ |
| 648 |
template <typename VectorType> |
| 649 |
void |
| 650 |
Tvmult(VectorType &, const VectorType &) const; |
| 651 |
|
| 652 |
/** |
| 653 |
* Perform one step of the preconditioned Richardson iteration |
| 654 |
*/ |
| 655 |
template <typename VectorType> |
| 656 |
void |
| 657 |
step(VectorType &x, const VectorType &rhs) const; |
| 658 |
|
| 659 |
/** |
| 660 |
* Perform one transposed step of the preconditioned Richardson iteration. |
| 661 |
*/ |
| 662 |
template <typename VectorType> |
| 663 |
void |
| 664 |
Tstep(VectorType &x, const VectorType &rhs) const; |
| 665 |
|
| 666 |
using EigenvalueInformation = internal::EigenvalueInformation; |
| 667 |
|
| 668 |
/** |
| 669 |
* Estimate eigenvalues and set relaxation parameter. |
| 670 |
*/ |
| 671 |
template <typename VectorType> |
| 672 |
EigenvalueInformation |
| 673 |
estimate_eigenvalues(const VectorType &src) const; |
| 674 |
|
| 675 |
/** |
| 676 |
* Return the relaxation parameter. This function also allows to return |
| 677 |
* the parameter in case it was internally determined by running an eigenvalue |
| 678 |
* algorithm. |
| 679 |
*/ |
| 680 |
double |
| 681 |
get_relaxation() const; |
| 682 |
|
| 683 |
protected: |
| 684 |
/** |
| 685 |
* Pointer to the matrix object. |
| 686 |
*/ |
| 687 |
ObserverPointer<const MatrixType, PreconditionRelaxation<MatrixType>> A; |
| 688 |
|
| 689 |
/** |
| 690 |
* Stores the additional data passed to the initialize function, obtained |
| 691 |
* through a copy operation. |
| 692 |
*/ |
| 693 |
AdditionalData data; |
| 694 |
|
| 695 |
/** |
| 696 |
* Preconditioner. |
| 697 |
*/ |
| 698 |
std::shared_ptr<PreconditionerType> preconditioner; |
| 699 |
|
| 700 |
/** |
| 701 |
* Stores whether the preconditioner has been set up and eigenvalues have |
| 702 |
* been computed. |
| 703 |
*/ |
| 704 |
bool eigenvalues_are_initialized; |
| 705 |
}; |
| 706 |
|
| 707 |
|
| 708 |
|
| 709 |
#ifndef DOXYGEN |
| 710 |
|
| 711 |
namespace internal |
| 712 |
{ |
| 713 |
// a helper type-trait that leverage SFINAE to figure out if MatrixType has |
| 714 |
// ... MatrixType::vmult(VectorType &, const VectorType&, |
| 715 |
// std::function<...>, std::function<...>) const |
| 716 |
template <typename MatrixType, typename VectorType> |
| 717 |
using vmult_functions_t = decltype(std::declval<const MatrixType>().vmult( |
| 718 |
std::declval<VectorType &>(), |
| 719 |
std::declval<const VectorType &>(), |
| 720 |
std::declval< |
| 721 |
const std::function<void(const unsigned int, const unsigned int)> &>(), |
| 722 |
std::declval< |
| 723 |
const std::function<void(const unsigned int, const unsigned int)> &>())); |
| 724 |
|
| 725 |
template <typename MatrixType, |
| 726 |
typename VectorType, |
| 727 |
typename PreconditionerType> |
| 728 |
constexpr bool has_vmult_with_std_functions = |
| 729 |
is_supported_operation<vmult_functions_t, MatrixType, VectorType> && |
| 730 |
std::is_same_v<PreconditionerType, dealii::DiagonalMatrix<VectorType>> && |
| 731 |
(std::is_same_v<VectorType, |
| 732 |
dealii::Vector<typename VectorType::value_type>> || |
| 733 |
std::is_same_v< |
| 734 |
VectorType, |
| 735 |
LinearAlgebra::distributed::Vector<typename VectorType::value_type, |
| 736 |
MemorySpace::Host>>); |
| 737 |
|
| 738 |
|
| 739 |
template <typename MatrixType, typename VectorType> |
| 740 |
constexpr bool has_vmult_with_std_functions_for_precondition = |
| 741 |
is_supported_operation<vmult_functions_t, MatrixType, VectorType>; |
| 742 |
|
| 743 |
namespace PreconditionRelaxation |
| 744 |
{ |
| 745 |
template <typename T, typename VectorType> |
| 746 |
using Tvmult_t = decltype(std::declval<const T>().Tvmult( |
| 747 |
std::declval<VectorType &>(), |
| 748 |
std::declval<const VectorType &>())); |
| 749 |
|
| 750 |
template <typename T, typename VectorType> |
| 751 |
constexpr bool has_Tvmult = is_supported_operation<Tvmult_t, T, VectorType>; |
| 752 |
|
| 753 |
template <typename T, typename VectorType> |
| 754 |
using step_t = decltype(std::declval<const T>().step( |
| 755 |
std::declval<VectorType &>(), |
| 756 |
std::declval<const VectorType &>())); |
| 757 |
|
| 758 |
template <typename T, typename VectorType> |
| 759 |
constexpr bool has_step = is_supported_operation<step_t, T, VectorType>; |
| 760 |
|
| 761 |
template <typename T, typename VectorType> |
| 762 |
using step_omega_t = |
| 763 |
decltype(std::declval<const T>().step(std::declval<VectorType &>(), |
| 764 |
std::declval<const VectorType &>(), |
| 765 |
std::declval<const double>())); |
| 766 |
|
| 767 |
template <typename T, typename VectorType> |
| 768 |
constexpr bool has_step_omega = |
| 769 |
is_supported_operation<step_omega_t, T, VectorType>; |
| 770 |
|
| 771 |
template <typename T, typename VectorType> |
| 772 |
using Tstep_t = decltype(std::declval<const T>().Tstep( |
| 773 |
std::declval<VectorType &>(), |
| 774 |
std::declval<const VectorType &>())); |
| 775 |
|
| 776 |
template <typename T, typename VectorType> |
| 777 |
constexpr bool has_Tstep = is_supported_operation<Tstep_t, T, VectorType>; |
| 778 |
|
| 779 |
template <typename T, typename VectorType> |
| 780 |
using Tstep_omega_t = |
| 781 |
decltype(std::declval<const T>().Tstep(std::declval<VectorType &>(), |
| 782 |
std::declval<const VectorType &>(), |
| 783 |
std::declval<const double>())); |
| 784 |
|
| 785 |
template <typename T, typename VectorType> |
| 786 |
constexpr bool has_Tstep_omega = |
| 787 |
is_supported_operation<Tstep_omega_t, T, VectorType>; |
| 788 |
|
| 789 |
template <typename T, typename VectorType> |
| 790 |
using jacobi_step_t = decltype(std::declval<const T>().Jacobi_step( |
| 791 |
std::declval<VectorType &>(), |
| 792 |
std::declval<const VectorType &>(), |
| 793 |
std::declval<const double>())); |
| 794 |
|
| 795 |
template <typename T, typename VectorType> |
| 796 |
constexpr bool has_jacobi_step = |
| 797 |
is_supported_operation<jacobi_step_t, T, VectorType>; |
| 798 |
|
| 799 |
template <typename T, typename VectorType> |
| 800 |
using SOR_step_t = decltype(std::declval<const T>().SOR_step( |
| 801 |
std::declval<VectorType &>(), |
| 802 |
std::declval<const VectorType &>(), |
| 803 |
std::declval<const double>())); |
| 804 |
|
| 805 |
template <typename T, typename VectorType> |
| 806 |
constexpr bool has_SOR_step = |
| 807 |
is_supported_operation<SOR_step_t, T, VectorType>; |
| 808 |
|
| 809 |
template <typename T, typename VectorType> |
| 810 |
using SSOR_step_t = decltype(std::declval<const T>().SSOR_step( |
| 811 |
std::declval<VectorType &>(), |
| 812 |
std::declval<const VectorType &>(), |
| 813 |
std::declval<const double>())); |
| 814 |
|
| 815 |
template <typename T, typename VectorType> |
| 816 |
constexpr bool has_SSOR_step = |
| 817 |
is_supported_operation<SSOR_step_t, T, VectorType>; |
| 818 |
|
| 819 |
template <typename MatrixType> |
| 820 |
class PreconditionJacobiImpl |
| 821 |
{ |
| 822 |
public: |
| 823 |
PreconditionJacobiImpl(const MatrixType &A, const double relaxation) |
| 824 |
: A(&A) |
| 825 |
, relaxation(relaxation) |
| 826 |
{} |
| 827 |
|
| 828 |
template <typename VectorType> |
| 829 |
void |
| 830 |
vmult(VectorType &dst, const VectorType &src) const |
| 831 |
{ |
| 832 |
this->A->precondition_Jacobi(dst, src, this->relaxation); |
| 833 |
} |
| 834 |
|
| 835 |
template <typename VectorType> |
| 836 |
void |
| 837 |
Tvmult(VectorType &dst, const VectorType &src) const |
| 838 |
{ |
| 839 |
// call vmult, since preconditioner is symmetrical |
| 840 |
this->vmult(dst, src); |
| 841 |
} |
| 842 |
|
| 843 |
template <typename VectorType, |
| 844 |
std::enable_if_t<has_jacobi_step<MatrixType, VectorType>, |
| 845 |
MatrixType> * = nullptr> |
| 846 |
void |
| 847 |
step(VectorType &dst, const VectorType &src) const |
| 848 |
{ |
| 849 |
this->A->Jacobi_step(dst, src, this->relaxation); |
| 850 |
} |
| 851 |
|
| 852 |
template <typename VectorType, |
| 853 |
std::enable_if_t<!has_jacobi_step<MatrixType, VectorType>, |
| 854 |
MatrixType> * = nullptr> |
| 855 |
void |
| 856 |
step(VectorType &, const VectorType &) const |
| 857 |
{ |
| 858 |
AssertThrow(false, |
| 859 |
ExcMessage( |
| 860 |
"Matrix A does not provide a Jacobi_step() function!")); |
| 861 |
} |
| 862 |
|
| 863 |
template <typename VectorType> |
| 864 |
void |
| 865 |
Tstep(VectorType &dst, const VectorType &src) const |
| 866 |
{ |
| 867 |
// call step, since preconditioner is symmetrical |
| 868 |
this->step(dst, src); |
| 869 |
} |
| 870 |
|
| 871 |
private: |
| 872 |
const ObserverPointer<const MatrixType> A; |
| 873 |
const double relaxation; |
| 874 |
}; |
| 875 |
|
| 876 |
template <typename MatrixType> |
| 877 |
class PreconditionSORImpl |
| 878 |
{ |
| 879 |
public: |
| 880 |
PreconditionSORImpl(const MatrixType &A, const double relaxation) |
| 881 |
: A(&A) |
| 882 |
, relaxation(relaxation) |
| 883 |
{} |
| 884 |
|
| 885 |
template <typename VectorType> |
| 886 |
void |
| 887 |
vmult(VectorType &dst, const VectorType &src) const |
| 888 |
{ |
| 889 |
this->A->precondition_SOR(dst, src, this->relaxation); |
| 890 |
} |
| 891 |
|
| 892 |
template <typename VectorType> |
| 893 |
void |
| 894 |
Tvmult(VectorType &dst, const VectorType &src) const |
| 895 |
{ |
| 896 |
this->A->precondition_TSOR(dst, src, this->relaxation); |
| 897 |
} |
| 898 |
|
| 899 |
template <typename VectorType, |
| 900 |
std::enable_if_t<has_SOR_step<MatrixType, VectorType>, |
| 901 |
MatrixType> * = nullptr> |
| 902 |
void |
| 903 |
step(VectorType &dst, const VectorType &src) const |
| 904 |
{ |
| 905 |
this->A->SOR_step(dst, src, this->relaxation); |
| 906 |
} |
| 907 |
|
| 908 |
template <typename VectorType, |
| 909 |
std::enable_if_t<!has_SOR_step<MatrixType, VectorType>, |
| 910 |
MatrixType> * = nullptr> |
| 911 |
void |
| 912 |
step(VectorType &, const VectorType &) const |
| 913 |
{ |
| 914 |
AssertThrow(false, |
| 915 |
ExcMessage( |
| 916 |
"Matrix A does not provide a SOR_step() function!")); |
| 917 |
} |
| 918 |
|
| 919 |
template <typename VectorType, |
| 920 |
std::enable_if_t<has_SOR_step<MatrixType, VectorType>, |
| 921 |
MatrixType> * = nullptr> |
| 922 |
void |
| 923 |
Tstep(VectorType &dst, const VectorType &src) const |
| 924 |
{ |
| 925 |
this->A->TSOR_step(dst, src, this->relaxation); |
| 926 |
} |
| 927 |
|
| 928 |
template <typename VectorType, |
| 929 |
std::enable_if_t<!has_SOR_step<MatrixType, VectorType>, |
| 930 |
MatrixType> * = nullptr> |
| 931 |
void |
| 932 |
Tstep(VectorType &, const VectorType &) const |
| 933 |
{ |
| 934 |
AssertThrow(false, |
| 935 |
ExcMessage( |
| 936 |
"Matrix A does not provide a TSOR_step() function!")); |
| 937 |
} |
| 938 |
|
| 939 |
private: |
| 940 |
const ObserverPointer<const MatrixType> A; |
| 941 |
const double relaxation; |
| 942 |
}; |
| 943 |
|
| 944 |
template <typename MatrixType> |
| 945 |
class PreconditionSSORImpl |
| 946 |
{ |
| 947 |
public: |
| 948 |
using size_type = typename MatrixType::size_type; |
| 949 |
|
| 950 |
PreconditionSSORImpl(const MatrixType &A, const double relaxation) |
| 951 |
: A(&A) |
| 952 |
, relaxation(relaxation) |
| 953 |
{ |
| 954 |
// in case we have a SparseMatrix class, we can extract information |
| 955 |
// about the diagonal. |
| 956 |
const SparseMatrix<typename MatrixType::value_type> *mat = |
| 957 |
dynamic_cast<const SparseMatrix<typename MatrixType::value_type> *>( |
| 958 |
&*this->A); |
| 959 |
|
| 960 |
// calculate the positions first after the diagonal. |
| 961 |
if (mat != nullptr) |
| 962 |
{ |
| 963 |
const size_type n = this->A->n(); |
| 964 |
pos_right_of_diagonal.resize(n, static_cast<std::size_t>(-1)); |
| 965 |
for (size_type row = 0; row < n; ++row) |
| 966 |
{ |
| 967 |
// find the first element in this line which is on the right of |
| 968 |
// the diagonal. we need to precondition with the elements on |
| 969 |
// the left only. note: the first entry in each line denotes the |
| 970 |
// diagonal element, which we need not check. |
| 971 |
typename SparseMatrix< |
| 972 |
typename MatrixType::value_type>::const_iterator it = |
| 973 |
mat->begin(row) + 1; |
| 974 |
for (; it < mat->end(row); ++it) |
| 975 |
if (it->column() > row) |
| 976 |
break; |
| 977 |
pos_right_of_diagonal[row] = it - mat->begin(); |
| 978 |
} |
| 979 |
} |
| 980 |
} |
| 981 |
|
| 982 |
template <typename VectorType> |
| 983 |
void |
| 984 |
vmult(VectorType &dst, const VectorType &src) const |
| 985 |
{ |
| 986 |
this->A->precondition_SSOR(dst, |
| 987 |
src, |
| 988 |
this->relaxation, |
| 989 |
pos_right_of_diagonal); |
| 990 |
} |
| 991 |
|
| 992 |
template <typename VectorType> |
| 993 |
void |
| 994 |
Tvmult(VectorType &dst, const VectorType &src) const |
| 995 |
{ |
| 996 |
this->A->precondition_SSOR(dst, |
| 997 |
src, |
| 998 |
this->relaxation, |
| 999 |
pos_right_of_diagonal); |
| 1000 |
} |
| 1001 |
|
| 1002 |
template <typename VectorType, |
| 1003 |
std::enable_if_t<has_SSOR_step<MatrixType, VectorType>, |
| 1004 |
MatrixType> * = nullptr> |
| 1005 |
void |
| 1006 |
step(VectorType &dst, const VectorType &src) const |
| 1007 |
{ |
| 1008 |
this->A->SSOR_step(dst, src, this->relaxation); |
| 1009 |
} |
| 1010 |
|
| 1011 |
template <typename VectorType, |
| 1012 |
std::enable_if_t<!has_SSOR_step<MatrixType, VectorType>, |
| 1013 |
MatrixType> * = nullptr> |
| 1014 |
void |
| 1015 |
step(VectorType &, const VectorType &) const |
| 1016 |
{ |
| 1017 |
AssertThrow(false, |
| 1018 |
ExcMessage( |
| 1019 |
"Matrix A does not provide a SSOR_step() function!")); |
| 1020 |
} |
| 1021 |
|
| 1022 |
template <typename VectorType> |
| 1023 |
void |
| 1024 |
Tstep(VectorType &dst, const VectorType &src) const |
| 1025 |
{ |
| 1026 |
// call step, since preconditioner is symmetrical |
| 1027 |
this->step(dst, src); |
| 1028 |
} |
| 1029 |
|
| 1030 |
private: |
| 1031 |
const ObserverPointer<const MatrixType> A; |
| 1032 |
const double relaxation; |
| 1033 |
|
| 1034 |
/** |
| 1035 |
* An array that stores for each matrix row where the first position after |
| 1036 |
* the diagonal is located. |
| 1037 |
*/ |
| 1038 |
std::vector<std::size_t> pos_right_of_diagonal; |
| 1039 |
}; |
| 1040 |
|
| 1041 |
template <typename MatrixType> |
| 1042 |
class PreconditionPSORImpl |
| 1043 |
{ |
| 1044 |
public: |
| 1045 |
using size_type = typename MatrixType::size_type; |
| 1046 |
|
| 1047 |
PreconditionPSORImpl(const MatrixType &A, |
| 1048 |
const double relaxation, |
| 1049 |
const std::vector<size_type> &permutation, |
| 1050 |
const std::vector<size_type> &inverse_permutation) |
| 1051 |
: A(&A) |
| 1052 |
, relaxation(relaxation) |
| 1053 |
, permutation(permutation) |
| 1054 |
, inverse_permutation(inverse_permutation) |
| 1055 |
{} |
| 1056 |
|
| 1057 |
template <typename VectorType> |
| 1058 |
void |
| 1059 |
vmult(VectorType &dst, const VectorType &src) const |
| 1060 |
{ |
| 1061 |
dst = src; |
| 1062 |
this->A->PSOR(dst, permutation, inverse_permutation, this->relaxation); |
| 1063 |
} |
| 1064 |
|
| 1065 |
template <typename VectorType> |
| 1066 |
void |
| 1067 |
Tvmult(VectorType &dst, const VectorType &src) const |
| 1068 |
{ |
| 1069 |
dst = src; |
| 1070 |
this->A->TPSOR(dst, permutation, inverse_permutation, this->relaxation); |
| 1071 |
} |
| 1072 |
|
| 1073 |
private: |
| 1074 |
const ObserverPointer<const MatrixType> A; |
| 1075 |
const double relaxation; |
| 1076 |
|
| 1077 |
const std::vector<size_type> &permutation; |
| 1078 |
const std::vector<size_type> &inverse_permutation; |
| 1079 |
}; |
| 1080 |
|
| 1081 |
template <typename MatrixType, |
| 1082 |
typename PreconditionerType, |
| 1083 |
typename VectorType, |
| 1084 |
std::enable_if_t<has_step_omega<PreconditionerType, VectorType>, |
| 1085 |
PreconditionerType> * = nullptr> |
| 1086 |
void |
| 1087 |
step(const MatrixType &, |
| 1088 |
const PreconditionerType &preconditioner, |
| 1089 |
VectorType &dst, |
| 1090 |
const VectorType &src, |
| 1091 |
const double relaxation, |
| 1092 |
VectorType &, |
| 1093 |
VectorType &) |
| 1094 |
{ |
| 1095 |
preconditioner.step(dst, src, relaxation); |
| 1096 |
} |
| 1097 |
|
| 1098 |
template < |
| 1099 |
typename MatrixType, |
| 1100 |
typename PreconditionerType, |
| 1101 |
typename VectorType, |
| 1102 |
std::enable_if_t<!has_step_omega<PreconditionerType, VectorType> && |
| 1103 |
has_step<PreconditionerType, VectorType>, |
| 1104 |
PreconditionerType> * = nullptr> |
| 1105 |
void |
| 1106 |
step(const MatrixType &, |
| 1107 |
const PreconditionerType &preconditioner, |
| 1108 |
VectorType &dst, |
| 1109 |
const VectorType &src, |
| 1110 |
const double relaxation, |
| 1111 |
VectorType &, |
| 1112 |
VectorType &) |
| 1113 |
{ |
| 1114 |
Assert(relaxation == 1.0, ExcInternalError()); |
| 1115 |
|
| 1116 |
(void)relaxation; |
| 1117 |
|
| 1118 |
preconditioner.step(dst, src); |
| 1119 |
} |
| 1120 |
|
| 1121 |
template < |
| 1122 |
typename MatrixType, |
| 1123 |
typename PreconditionerType, |
| 1124 |
typename VectorType, |
| 1125 |
std::enable_if_t<!has_step_omega<PreconditionerType, VectorType> && |
| 1126 |
!has_step<PreconditionerType, VectorType>, |
| 1127 |
PreconditionerType> * = nullptr> |
| 1128 |
void |
| 1129 |
step(const MatrixType &A, |
| 1130 |
const PreconditionerType &preconditioner, |
| 1131 |
VectorType &dst, |
| 1132 |
const VectorType &src, |
| 1133 |
const double relaxation, |
| 1134 |
VectorType &residual, |
| 1135 |
VectorType &tmp) |
| 1136 |
{ |
| 1137 |
residual.reinit(dst, true); |
| 1138 |
tmp.reinit(dst, true); |
| 1139 |
|
| 1140 |
A.vmult(residual, dst); |
| 1141 |
residual.sadd(-1.0, 1.0, src); |
| 1142 |
|
| 1143 |
preconditioner.vmult(tmp, residual); |
| 1144 |
dst.add(relaxation, tmp); |
| 1145 |
} |
| 1146 |
|
| 1147 |
template <typename MatrixType, |
| 1148 |
typename PreconditionerType, |
| 1149 |
typename VectorType, |
| 1150 |
std::enable_if_t<has_Tstep_omega<PreconditionerType, VectorType>, |
| 1151 |
PreconditionerType> * = nullptr> |
| 1152 |
void |
| 1153 |
Tstep(const MatrixType &, |
| 1154 |
const PreconditionerType &preconditioner, |
| 1155 |
VectorType &dst, |
| 1156 |
const VectorType &src, |
| 1157 |
const double relaxation, |
| 1158 |
VectorType &, |
| 1159 |
VectorType &) |
| 1160 |
{ |
| 1161 |
preconditioner.Tstep(dst, src, relaxation); |
| 1162 |
} |
| 1163 |
|
| 1164 |
template < |
| 1165 |
typename MatrixType, |
| 1166 |
typename PreconditionerType, |
| 1167 |
typename VectorType, |
| 1168 |
std::enable_if_t<!has_Tstep_omega<PreconditionerType, VectorType> && |
| 1169 |
has_Tstep<PreconditionerType, VectorType>, |
| 1170 |
PreconditionerType> * = nullptr> |
| 1171 |
void |
| 1172 |
Tstep(const MatrixType &, |
| 1173 |
const PreconditionerType &preconditioner, |
| 1174 |
VectorType &dst, |
| 1175 |
const VectorType &src, |
| 1176 |
const double relaxation, |
| 1177 |
VectorType &, |
| 1178 |
VectorType &) |
| 1179 |
{ |
| 1180 |
Assert(relaxation == 1.0, ExcInternalError()); |
| 1181 |
|
| 1182 |
(void)relaxation; |
| 1183 |
|
| 1184 |
preconditioner.Tstep(dst, src); |
| 1185 |
} |
| 1186 |
|
| 1187 |
template <typename MatrixType, |
| 1188 |
typename VectorType, |
| 1189 |
std::enable_if_t<has_Tvmult<MatrixType, VectorType>, MatrixType> |
| 1190 |
* = nullptr> |
| 1191 |
void |
| 1192 |
Tvmult(const MatrixType &A, VectorType &dst, const VectorType &src) |
| 1193 |
{ |
| 1194 |
A.Tvmult(dst, src); |
| 1195 |
} |
| 1196 |
|
| 1197 |
template <typename MatrixType, |
| 1198 |
typename VectorType, |
| 1199 |
std::enable_if_t<!has_Tvmult<MatrixType, VectorType>, MatrixType> |
| 1200 |
* = nullptr> |
| 1201 |
void |
| 1202 |
Tvmult(const MatrixType &, VectorType &, const VectorType &) |
| 1203 |
{ |
| 1204 |
AssertThrow(false, |
| 1205 |
ExcMessage("Matrix A does not provide a Tvmult() function!")); |
| 1206 |
} |
| 1207 |
|
| 1208 |
template < |
| 1209 |
typename MatrixType, |
| 1210 |
typename PreconditionerType, |
| 1211 |
typename VectorType, |
| 1212 |
std::enable_if_t<!has_Tstep_omega<PreconditionerType, VectorType> && |
| 1213 |
!has_Tstep<PreconditionerType, VectorType>, |
| 1214 |
PreconditionerType> * = nullptr> |
| 1215 |
void |
| 1216 |
Tstep(const MatrixType &A, |
| 1217 |
const PreconditionerType &preconditioner, |
| 1218 |
VectorType &dst, |
| 1219 |
const VectorType &src, |
| 1220 |
const double relaxation, |
| 1221 |
VectorType &residual, |
| 1222 |
VectorType &tmp) |
| 1223 |
{ |
| 1224 |
residual.reinit(dst, true); |
| 1225 |
tmp.reinit(dst, true); |
| 1226 |
|
| 1227 |
Tvmult(A, residual, dst); |
| 1228 |
residual.sadd(-1.0, 1.0, src); |
| 1229 |
|
| 1230 |
Tvmult(preconditioner, tmp, residual); |
| 1231 |
dst.add(relaxation, tmp); |
| 1232 |
} |
| 1233 |
|
| 1234 |
// 0) general implementation |
| 1235 |
template <typename MatrixType, |
| 1236 |
typename PreconditionerType, |
| 1237 |
typename VectorType, |
| 1238 |
std::enable_if_t<!has_vmult_with_std_functions_for_precondition< |
| 1239 |
PreconditionerType, |
| 1240 |
VectorType>, |
| 1241 |
int> * = nullptr> |
| 1242 |
void |
| 1243 |
step_operations(const MatrixType &A, |
| 1244 |
const PreconditionerType &preconditioner, |
| 1245 |
VectorType &dst, |
| 1246 |
const VectorType &src, |
| 1247 |
const double relaxation, |
| 1248 |
VectorType &tmp1, |
| 1249 |
VectorType &tmp2, |
| 1250 |
const unsigned int i, |
| 1251 |
const bool transposed) |
| 1252 |
{ |
| 1253 |
if (i == 0) |
| 1254 |
{ |
| 1255 |
if (transposed) |
| 1256 |
Tvmult(preconditioner, dst, src); |
| 1257 |
else |
| 1258 |
preconditioner.vmult(dst, src); |
| 1259 |
|
| 1260 |
if (relaxation != 1.0) |
| 1261 |
dst *= relaxation; |
| 1262 |
} |
| 1263 |
else |
| 1264 |
{ |
| 1265 |
if (transposed) |
| 1266 |
Tstep(A, preconditioner, dst, src, relaxation, tmp1, tmp2); |
| 1267 |
else |
| 1268 |
step(A, preconditioner, dst, src, relaxation, tmp1, tmp2); |
| 1269 |
} |
| 1270 |
} |
| 1271 |
|
| 1272 |
// 1) specialized implementation with a preconditioner that accepts |
| 1273 |
// ranges |
| 1274 |
template < |
| 1275 |
typename MatrixType, |
| 1276 |
typename PreconditionerType, |
| 1277 |
typename VectorType, |
| 1278 |
std::enable_if_t< |
| 1279 |
has_vmult_with_std_functions_for_precondition<PreconditionerType, |
| 1280 |
VectorType> && |
| 1281 |
!has_vmult_with_std_functions_for_precondition<MatrixType, |
| 1282 |
VectorType>, |
| 1283 |
int> * = nullptr> |
| 1284 |
void |
| 1285 |
step_operations(const MatrixType &A, |
| 1286 |
const PreconditionerType &preconditioner, |
| 1287 |
VectorType &dst, |
| 1288 |
const VectorType &src, |
| 1289 |
const double relaxation, |
| 1290 |
VectorType &tmp, |
| 1291 |
VectorType &, |
| 1292 |
const unsigned int i, |
| 1293 |
const bool transposed) |
| 1294 |
{ |
| 1295 |
(void)transposed; |
| 1296 |
using Number = typename VectorType::value_type; |
| 1297 |
Number *dst_ptr = dst.begin(); |
| 1298 |
const Number *src_ptr = src.begin(); |
| 1299 |
|
| 1300 |
if (i == 0) |
| 1301 |
{ |
| 1302 |
preconditioner.vmult( |
| 1303 |
dst, |
| 1304 |
src, |
| 1305 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1306 |
// zero 'dst' before running the vmult operation |
| 1307 |
if (end_range > start_range) |
| 1308 |
std::memset(dst_ptr + start_range, |
| 1309 |
0, |
| 1310 |
sizeof(Number) * (end_range - start_range)); |
| 1311 |
}, |
| 1312 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1313 |
if (relaxation == 1.0) |
| 1314 |
return; // nothing to do |
| 1315 |
|
| 1316 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1317 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 1318 |
dst_ptr[i] *= relaxation; |
| 1319 |
}); |
| 1320 |
} |
| 1321 |
else |
| 1322 |
{ |
| 1323 |
tmp.reinit(src, true); |
| 1324 |
|
| 1325 |
Assert(transposed == false, ExcNotImplemented()); |
| 1326 |
|
| 1327 |
A.vmult(tmp, dst); |
| 1328 |
|
| 1329 |
preconditioner.vmult( |
| 1330 |
dst, |
| 1331 |
tmp, |
| 1332 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1333 |
const auto tmp_ptr = tmp.begin(); |
| 1334 |
|
| 1335 |
if (relaxation == 1.0) |
| 1336 |
{ |
| 1337 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1338 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 1339 |
tmp_ptr[i] = src_ptr[i] - tmp_ptr[i]; |
| 1340 |
} |
| 1341 |
else |
| 1342 |
{ |
| 1343 |
// note: we scale the residual here to be able to add into |
| 1344 |
// the dst vector, which contains the solution from the last |
| 1345 |
// iteration |
| 1346 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1347 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 1348 |
tmp_ptr[i] = relaxation * (src_ptr[i] - tmp_ptr[i]); |
| 1349 |
} |
| 1350 |
}, |
| 1351 |
[&](const unsigned int, const unsigned int) { |
| 1352 |
// nothing to do, since scaling by the relaxation factor |
| 1353 |
// has been done in the pre operation |
| 1354 |
}); |
| 1355 |
} |
| 1356 |
} |
| 1357 |
|
| 1358 |
// 2) specialized implementation with a preconditioner and a matrix that |
| 1359 |
// accepts ranges |
| 1360 |
template < |
| 1361 |
typename MatrixType, |
| 1362 |
typename PreconditionerType, |
| 1363 |
typename VectorType, |
| 1364 |
std::enable_if_t< |
| 1365 |
has_vmult_with_std_functions_for_precondition<PreconditionerType, |
| 1366 |
VectorType> && |
| 1367 |
has_vmult_with_std_functions_for_precondition<MatrixType, VectorType>, |
| 1368 |
int> * = nullptr> |
| 1369 |
void |
| 1370 |
step_operations(const MatrixType &A, |
| 1371 |
const PreconditionerType &preconditioner, |
| 1372 |
VectorType &dst, |
| 1373 |
const VectorType &src, |
| 1374 |
const double relaxation, |
| 1375 |
VectorType &tmp, |
| 1376 |
VectorType &, |
| 1377 |
const unsigned int i, |
| 1378 |
const bool transposed) |
| 1379 |
{ |
| 1380 |
(void)transposed; |
| 1381 |
using Number = typename VectorType::value_type; |
| 1382 |
|
| 1383 |
Number *dst_ptr = dst.begin(); |
| 1384 |
const Number *src_ptr = src.begin(); |
| 1385 |
|
| 1386 |
if (i == 0) |
| 1387 |
{ |
| 1388 |
preconditioner.vmult( |
| 1389 |
dst, |
| 1390 |
src, |
| 1391 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1392 |
// zero 'dst' before running the vmult operation |
| 1393 |
if (end_range > start_range) |
| 1394 |
std::memset(dst_ptr + start_range, |
| 1395 |
0, |
| 1396 |
sizeof(Number) * (end_range - start_range)); |
| 1397 |
}, |
| 1398 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1399 |
if (relaxation == 1.0) |
| 1400 |
return; // nothing to do |
| 1401 |
|
| 1402 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1403 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 1404 |
dst_ptr[i] *= relaxation; |
| 1405 |
}); |
| 1406 |
} |
| 1407 |
else |
| 1408 |
{ |
| 1409 |
tmp.reinit(src, true); |
| 1410 |
const auto tmp_ptr = tmp.begin(); |
| 1411 |
|
| 1412 |
Assert(transposed == false, ExcNotImplemented()); |
| 1413 |
|
| 1414 |
A.vmult( |
| 1415 |
tmp, |
| 1416 |
dst, |
| 1417 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1418 |
// zero 'tmp' before running the vmult |
| 1419 |
// operation |
| 1420 |
if (end_range > start_range) |
| 1421 |
std::memset(tmp_ptr + start_range, |
| 1422 |
0, |
| 1423 |
sizeof(Number) * (end_range - start_range)); |
| 1424 |
}, |
| 1425 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1426 |
if (relaxation == 1.0) |
| 1427 |
{ |
| 1428 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1429 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 1430 |
tmp_ptr[i] = src_ptr[i] - tmp_ptr[i]; |
| 1431 |
} |
| 1432 |
else |
| 1433 |
{ |
| 1434 |
// note: we scale the residual here to be able to add into |
| 1435 |
// the dst vector, which contains the solution from the last |
| 1436 |
// iteration |
| 1437 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1438 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 1439 |
tmp_ptr[i] = relaxation * (src_ptr[i] - tmp_ptr[i]); |
| 1440 |
} |
| 1441 |
}); |
| 1442 |
|
| 1443 |
preconditioner.vmult(dst, tmp, [](const auto, const auto) { |
| 1444 |
// note: `dst` vector does not have to be zeroed out |
| 1445 |
// since we add the result into it |
| 1446 |
}); |
| 1447 |
} |
| 1448 |
} |
| 1449 |
|
| 1450 |
// 3) specialized implementation for inverse-diagonal preconditioner |
| 1451 |
template < |
| 1452 |
typename MatrixType, |
| 1453 |
typename VectorType, |
| 1454 |
std::enable_if_t< |
| 1455 |
!IsBlockVector<VectorType>::value && |
| 1456 |
!std::is_same_v< |
| 1457 |
VectorType, |
| 1458 |
LinearAlgebra::distributed::Vector<typename VectorType::value_type, |
| 1459 |
MemorySpace::Default>> && |
| 1460 |
!has_vmult_with_std_functions<MatrixType, |
| 1461 |
VectorType, |
| 1462 |
dealii::DiagonalMatrix<VectorType>>, |
| 1463 |
VectorType> * = nullptr> |
| 1464 |
void |
| 1465 |
step_operations(const MatrixType &A, |
| 1466 |
const dealii::DiagonalMatrix<VectorType> &preconditioner, |
| 1467 |
VectorType &dst, |
| 1468 |
const VectorType &src, |
| 1469 |
const double relaxation, |
| 1470 |
VectorType &tmp, |
| 1471 |
VectorType &, |
| 1472 |
const unsigned int i, |
| 1473 |
const bool transposed) |
| 1474 |
{ |
| 1475 |
using Number = typename VectorType::value_type; |
| 1476 |
|
| 1477 |
if (i == 0) |
| 1478 |
{ |
| 1479 |
Number *dst_ptr = dst.begin(); |
| 1480 |
const Number *src_ptr = src.begin(); |
| 1481 |
const Number *diag_ptr = preconditioner.get_vector().begin(); |
| 1482 |
|
| 1483 |
if (relaxation == 1.0) |
| 1484 |
{ |
| 1485 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1486 |
for (unsigned int i = 0; i < dst.locally_owned_size(); ++i) |
| 1487 |
dst_ptr[i] = src_ptr[i] * diag_ptr[i]; |
| 1488 |
} |
| 1489 |
else |
| 1490 |
{ |
| 1491 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1492 |
for (unsigned int i = 0; i < dst.locally_owned_size(); ++i) |
| 1493 |
dst_ptr[i] = relaxation * src_ptr[i] * diag_ptr[i]; |
| 1494 |
} |
| 1495 |
} |
| 1496 |
else |
| 1497 |
{ |
| 1498 |
tmp.reinit(src, true); |
| 1499 |
|
| 1500 |
if (transposed) |
| 1501 |
Tvmult(A, tmp, dst); |
| 1502 |
else |
| 1503 |
A.vmult(tmp, dst); |
| 1504 |
|
| 1505 |
Number *dst_ptr = dst.begin(); |
| 1506 |
const Number *src_ptr = src.begin(); |
| 1507 |
const Number *tmp_ptr = tmp.begin(); |
| 1508 |
const Number *diag_ptr = preconditioner.get_vector().begin(); |
| 1509 |
|
| 1510 |
if (relaxation == 1.0) |
| 1511 |
{ |
| 1512 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1513 |
for (unsigned int i = 0; i < dst.locally_owned_size(); ++i) |
| 1514 |
dst_ptr[i] += (src_ptr[i] - tmp_ptr[i]) * diag_ptr[i]; |
| 1515 |
} |
| 1516 |
else |
| 1517 |
{ |
| 1518 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1519 |
for (unsigned int i = 0; i < dst.locally_owned_size(); ++i) |
| 1520 |
dst_ptr[i] += |
| 1521 |
relaxation * (src_ptr[i] - tmp_ptr[i]) * diag_ptr[i]; |
| 1522 |
} |
| 1523 |
} |
| 1524 |
} |
| 1525 |
|
| 1526 |
// 4) specialized implementation for inverse-diagonal preconditioner and |
| 1527 |
// matrix that accepts ranges |
| 1528 |
template <typename MatrixType, |
| 1529 |
typename VectorType, |
| 1530 |
std::enable_if_t<!IsBlockVector<VectorType>::value && |
| 1531 |
has_vmult_with_std_functions< |
| 1532 |
MatrixType, |
| 1533 |
VectorType, |
| 1534 |
dealii::DiagonalMatrix<VectorType>>, |
| 1535 |
VectorType> * = nullptr> |
| 1536 |
void |
| 1537 |
step_operations(const MatrixType &A, |
| 1538 |
const dealii::DiagonalMatrix<VectorType> &preconditioner, |
| 1539 |
VectorType &dst, |
| 1540 |
const VectorType &src, |
| 1541 |
const double relaxation, |
| 1542 |
VectorType &tmp, |
| 1543 |
VectorType &, |
| 1544 |
const unsigned int i, |
| 1545 |
const bool transposed) |
| 1546 |
{ |
| 1547 |
(void)transposed; |
| 1548 |
using Number = typename VectorType::value_type; |
| 1549 |
|
| 1550 |
if (i == 0) |
| 1551 |
{ |
| 1552 |
Number *dst_ptr = dst.begin(); |
| 1553 |
const Number *src_ptr = src.begin(); |
| 1554 |
const Number *diag_ptr = preconditioner.get_vector().begin(); |
| 1555 |
|
| 1556 |
if (relaxation == 1.0) |
| 1557 |
{ |
| 1558 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1559 |
for (unsigned int i = 0; i < dst.locally_owned_size(); ++i) |
| 1560 |
dst_ptr[i] = src_ptr[i] * diag_ptr[i]; |
| 1561 |
} |
| 1562 |
else |
| 1563 |
{ |
| 1564 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1565 |
for (unsigned int i = 0; i < dst.locally_owned_size(); ++i) |
| 1566 |
dst_ptr[i] = relaxation * src_ptr[i] * diag_ptr[i]; |
| 1567 |
} |
| 1568 |
} |
| 1569 |
else |
| 1570 |
{ |
| 1571 |
tmp.reinit(src, true); |
| 1572 |
|
| 1573 |
Assert(transposed == false, ExcNotImplemented()); |
| 1574 |
|
| 1575 |
A.vmult( |
| 1576 |
tmp, |
| 1577 |
dst, |
| 1578 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 1579 |
// zero 'tmp' before running the vmult operation |
| 1580 |
if (end_range > start_range) |
| 1581 |
std::memset(tmp.begin() + start_range, |
| 1582 |
0, |
| 1583 |
sizeof(Number) * (end_range - start_range)); |
| 1584 |
}, |
| 1585 |
[&](const unsigned int begin, const unsigned int end) { |
| 1586 |
const Number *dst_ptr = dst.begin(); |
| 1587 |
const Number *src_ptr = src.begin(); |
| 1588 |
Number *tmp_ptr = tmp.begin(); |
| 1589 |
const Number *diag_ptr = preconditioner.get_vector().begin(); |
| 1590 |
|
| 1591 |
// for efficiency reason, write back to temp_vector that is |
| 1592 |
// already read (avoid read-for-ownership) |
| 1593 |
if (relaxation == 1.0) |
| 1594 |
{ |
| 1595 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1596 |
for (std::size_t i = begin; i < end; ++i) |
| 1597 |
tmp_ptr[i] = |
| 1598 |
dst_ptr[i] + (src_ptr[i] - tmp_ptr[i]) * diag_ptr[i]; |
| 1599 |
} |
| 1600 |
else |
| 1601 |
{ |
| 1602 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 1603 |
for (std::size_t i = begin; i < end; ++i) |
| 1604 |
tmp_ptr[i] = dst_ptr[i] + relaxation * |
| 1605 |
(src_ptr[i] - tmp_ptr[i]) * |
| 1606 |
diag_ptr[i]; |
| 1607 |
} |
| 1608 |
}); |
| 1609 |
|
| 1610 |
tmp.swap(dst); |
| 1611 |
} |
| 1612 |
} |
| 1613 |
|
| 1614 |
} // namespace PreconditionRelaxation |
| 1615 |
} // namespace internal |
| 1616 |
|
| 1617 |
#endif |
| 1618 |
|
| 1619 |
|
| 1620 |
|
| 1621 |
/** |
| 1622 |
* Jacobi preconditioner using matrix built-in function. The |
| 1623 |
* <tt>MatrixType</tt> class used is required to have a function |
| 1624 |
* <tt>precondition_Jacobi(VectorType&, const VectorType&, double</tt>). This |
| 1625 |
* class satisfies the |
| 1626 |
* @ref ConceptRelaxationType "relaxation concept". |
| 1627 |
* |
| 1628 |
* @code |
| 1629 |
* // Declare related objects |
| 1630 |
* |
| 1631 |
* SparseMatrix<double> A; |
| 1632 |
* Vector<double> x; |
| 1633 |
* Vector<double> b; |
| 1634 |
* SolverCG<> solver(...); |
| 1635 |
* |
| 1636 |
* //...initialize and build A |
| 1637 |
* |
| 1638 |
* // Define and initialize preconditioner: |
| 1639 |
* |
| 1640 |
* PreconditionJacobi<SparseMatrix<double> > precondition; |
| 1641 |
* precondition.initialize( |
| 1642 |
* A, PreconditionJacobi<SparseMatrix<double>>::AdditionalData(.6)); |
| 1643 |
* |
| 1644 |
* solver.solve (A, x, b, precondition); |
| 1645 |
* @endcode |
| 1646 |
*/ |
| 1647 |
template <typename MatrixType = SparseMatrix<double>> |
| 1648 |
class PreconditionJacobi |
| 1649 |
: public PreconditionRelaxation< |
| 1650 |
MatrixType, |
| 1651 |
internal::PreconditionRelaxation::PreconditionJacobiImpl<MatrixType>> |
| 1652 |
{ |
| 1653 |
using PreconditionerType = |
| 1654 |
internal::PreconditionRelaxation::PreconditionJacobiImpl<MatrixType>; |
| 1655 |
using BaseClass = PreconditionRelaxation<MatrixType, PreconditionerType>; |
| 1656 |
|
| 1657 |
public: |
| 1658 |
/** |
| 1659 |
* An alias to the base class AdditionalData. |
| 1660 |
*/ |
| 1661 |
using AdditionalData = typename BaseClass::AdditionalData; |
| 1662 |
|
| 1663 |
/** |
| 1664 |
* @copydoc PreconditionRelaxation::initialize() |
| 1665 |
*/ |
| 1666 |
void |
| 1667 |
initialize(const MatrixType &A, |
| 1668 |
const AdditionalData ¶meters = AdditionalData()); |
| 1669 |
}; |
| 1670 |
|
| 1671 |
|
| 1672 |
/** |
| 1673 |
* SOR preconditioner using matrix built-in function. |
| 1674 |
* |
| 1675 |
* Assuming the matrix <i>A = D + L + U</i> is split into its diagonal |
| 1676 |
* <i>D</i> as well as the strict lower and upper triangles <i>L</i> and |
| 1677 |
* <i>U</i>, then the SOR preconditioner with relaxation parameter <i>r</i> is |
| 1678 |
* @f[ |
| 1679 |
* P^{-1} = r (D+rL)^{-1}. |
| 1680 |
* @f] |
| 1681 |
* It is this operator <i>P<sup>-1</sup></i>, which is implemented by vmult() |
| 1682 |
* through forward substitution. Analogously, Tvmult() implements the |
| 1683 |
* operation of <i>r(D+rU)<sup>-1</sup></i>. |
| 1684 |
* |
| 1685 |
* The SOR iteration itself can be directly written as |
| 1686 |
* @f[ |
| 1687 |
* x^{k+1} = x^k - r D^{-1} \bigl(L x^{k+1} + U x^k - b\bigr). |
| 1688 |
* @f] |
| 1689 |
* Using the right hand side <i>b</i> and the previous iterate <i>x</i>, this |
| 1690 |
* is the operation implemented by step(). |
| 1691 |
* |
| 1692 |
* The MatrixType class used is required to have functions |
| 1693 |
* <tt>precondition_SOR(VectorType&, const VectorType&, double)</tt> and |
| 1694 |
* <tt>precondition_TSOR(VectorType&, const VectorType&, double)</tt>. This |
| 1695 |
* class satisfies the |
| 1696 |
* @ref ConceptRelaxationType "relaxation concept". |
| 1697 |
* |
| 1698 |
* @code |
| 1699 |
* // Declare related objects |
| 1700 |
* |
| 1701 |
* SparseMatrix<double> A; |
| 1702 |
* Vector<double> x; |
| 1703 |
* Vector<double> b; |
| 1704 |
* SolverCG<> solver(...); |
| 1705 |
* |
| 1706 |
* //...initialize and build A |
| 1707 |
* |
| 1708 |
* // Define and initialize preconditioner |
| 1709 |
* |
| 1710 |
* PreconditionSOR<SparseMatrix<double> > precondition; |
| 1711 |
* precondition.initialize( |
| 1712 |
* A, PreconditionSOR<SparseMatrix<double>>::AdditionalData(.6)); |
| 1713 |
* |
| 1714 |
* solver.solve (A, x, b, precondition); |
| 1715 |
* @endcode |
| 1716 |
*/ |
| 1717 |
template <typename MatrixType = SparseMatrix<double>> |
| 1718 |
class PreconditionSOR |
| 1719 |
: public PreconditionRelaxation< |
| 1720 |
MatrixType, |
| 1721 |
internal::PreconditionRelaxation::PreconditionSORImpl<MatrixType>> |
| 1722 |
{ |
| 1723 |
using PreconditionerType = |
| 1724 |
internal::PreconditionRelaxation::PreconditionSORImpl<MatrixType>; |
| 1725 |
using BaseClass = PreconditionRelaxation<MatrixType, PreconditionerType>; |
| 1726 |
|
| 1727 |
public: |
| 1728 |
/** |
| 1729 |
* An alias to the base class AdditionalData. |
| 1730 |
*/ |
| 1731 |
using AdditionalData = typename BaseClass::AdditionalData; |
| 1732 |
|
| 1733 |
/** |
| 1734 |
* @copydoc PreconditionRelaxation::initialize() |
| 1735 |
*/ |
| 1736 |
void |
| 1737 |
initialize(const MatrixType &A, |
| 1738 |
const AdditionalData ¶meters = AdditionalData()); |
| 1739 |
}; |
| 1740 |
|
| 1741 |
|
| 1742 |
|
| 1743 |
/** |
| 1744 |
* SSOR preconditioner using matrix built-in function. The |
| 1745 |
* <tt>MatrixType</tt> class used is required to have a function |
| 1746 |
* <tt>precondition_SSOR(VectorType&, const VectorType&, double)</tt>. This |
| 1747 |
* class satisfies the |
| 1748 |
* @ref ConceptRelaxationType "relaxation concept". |
| 1749 |
* |
| 1750 |
* @code |
| 1751 |
* // Declare related objects |
| 1752 |
* |
| 1753 |
* SparseMatrix<double> A; |
| 1754 |
* Vector<double> x; |
| 1755 |
* Vector<double> b; |
| 1756 |
* SolverCG<> solver(...); |
| 1757 |
* |
| 1758 |
* //...initialize and build A |
| 1759 |
* |
| 1760 |
* // Define and initialize preconditioner |
| 1761 |
* |
| 1762 |
* PreconditionSSOR<SparseMatrix<double> > precondition; |
| 1763 |
* precondition.initialize( |
| 1764 |
* A, PreconditionSSOR<SparseMatrix<double>>::AdditionalData(.6)); |
| 1765 |
* |
| 1766 |
* solver.solve (A, x, b, precondition); |
| 1767 |
* @endcode |
| 1768 |
*/ |
| 1769 |
template <typename MatrixType = SparseMatrix<double>> |
| 1770 |
class PreconditionSSOR |
| 1771 |
: public PreconditionRelaxation< |
| 1772 |
MatrixType, |
| 1773 |
internal::PreconditionRelaxation::PreconditionSSORImpl<MatrixType>> |
| 1774 |
{ |
| 1775 |
using PreconditionerType = |
| 1776 |
internal::PreconditionRelaxation::PreconditionSSORImpl<MatrixType>; |
| 1777 |
using BaseClass = PreconditionRelaxation<MatrixType, PreconditionerType>; |
| 1778 |
|
| 1779 |
public: |
| 1780 |
/** |
| 1781 |
* An alias to the base class AdditionalData. |
| 1782 |
*/ |
| 1783 |
using AdditionalData = typename BaseClass::AdditionalData; |
| 1784 |
|
| 1785 |
/** |
| 1786 |
* Initialize matrix and relaxation parameter. The matrix is just stored in |
| 1787 |
* the preconditioner object. The relaxation parameter should be larger than |
| 1788 |
* zero and smaller than 2 for numerical reasons. It defaults to 1. |
| 1789 |
*/ |
| 1790 |
void |
| 1791 |
initialize(const MatrixType &A, |
| 1792 |
const AdditionalData ¶meters = AdditionalData()); |
| 1793 |
}; |
| 1794 |
|
| 1795 |
|
| 1796 |
/** |
| 1797 |
* Permuted SOR preconditioner using matrix built-in function. The |
| 1798 |
* <tt>MatrixType</tt> class used is required to have functions |
| 1799 |
* <tt>PSOR(VectorType&, const VectorType&, double)</tt> and |
| 1800 |
* <tt>TPSOR(VectorType&, const VectorType&, double)</tt>. |
| 1801 |
* |
| 1802 |
* @code |
| 1803 |
* // Declare related objects |
| 1804 |
* |
| 1805 |
* SparseMatrix<double> A; |
| 1806 |
* Vector<double> x; |
| 1807 |
* Vector<double> b; |
| 1808 |
* SolverCG<> solver(...); |
| 1809 |
* |
| 1810 |
* //...initialize and build A |
| 1811 |
* |
| 1812 |
* std::vector<unsigned int> permutation(x.size()); |
| 1813 |
* std::vector<unsigned int> inverse_permutation(x.size()); |
| 1814 |
* |
| 1815 |
* //...fill permutation and its inverse with reasonable values |
| 1816 |
* |
| 1817 |
* // Define and initialize preconditioner |
| 1818 |
* |
| 1819 |
* PreconditionPSOR<SparseMatrix<double> > precondition; |
| 1820 |
* precondition.initialize (A, permutation, inverse_permutation, .6); |
| 1821 |
* |
| 1822 |
* solver.solve (A, x, b, precondition); |
| 1823 |
* @endcode |
| 1824 |
*/ |
| 1825 |
template <typename MatrixType = SparseMatrix<double>> |
| 1826 |
class PreconditionPSOR |
| 1827 |
: public PreconditionRelaxation< |
| 1828 |
MatrixType, |
| 1829 |
internal::PreconditionRelaxation::PreconditionPSORImpl<MatrixType>> |
| 1830 |
{ |
| 1831 |
using PreconditionerType = |
| 1832 |
internal::PreconditionRelaxation::PreconditionPSORImpl<MatrixType>; |
| 1833 |
using BaseClass = PreconditionRelaxation<MatrixType, PreconditionerType>; |
| 1834 |
|
| 1835 |
public: |
| 1836 |
/** |
| 1837 |
* Declare type for container size. |
| 1838 |
*/ |
| 1839 |
using size_type = typename BaseClass::size_type; |
| 1840 |
|
| 1841 |
/** |
| 1842 |
* Parameters for PreconditionPSOR. |
| 1843 |
*/ |
| 1844 |
class AdditionalData |
| 1845 |
{ |
| 1846 |
public: |
| 1847 |
/** |
| 1848 |
* Constructor. For the parameters' description, see below. |
| 1849 |
* |
| 1850 |
* The permutation vectors are stored as a reference. Therefore, it has to |
| 1851 |
* be assured that the lifetime of the vector exceeds the lifetime of the |
| 1852 |
* preconditioner. |
| 1853 |
* |
| 1854 |
* The relaxation parameter should be larger than zero and smaller than 2 |
| 1855 |
* for numerical reasons. It defaults to 1. |
| 1856 |
*/ |
| 1857 |
AdditionalData(const std::vector<size_type> &permutation, |
| 1858 |
const std::vector<size_type> &inverse_permutation, |
| 1859 |
const typename BaseClass::AdditionalData ¶meters = |
| 1860 |
typename BaseClass::AdditionalData()); |
| 1861 |
|
| 1862 |
/** |
| 1863 |
* Storage for the permutation vector. |
| 1864 |
*/ |
| 1865 |
const std::vector<size_type> &permutation; |
| 1866 |
/** |
| 1867 |
* Storage for the inverse permutation vector. |
| 1868 |
*/ |
| 1869 |
const std::vector<size_type> &inverse_permutation; |
| 1870 |
/** |
| 1871 |
* Relaxation parameters |
| 1872 |
*/ |
| 1873 |
typename BaseClass::AdditionalData parameters; |
| 1874 |
}; |
| 1875 |
|
| 1876 |
/** |
| 1877 |
* Initialize matrix and relaxation parameter. The matrix is just stored in |
| 1878 |
* the preconditioner object. |
| 1879 |
* |
| 1880 |
* The permutation vector is stored as a pointer. Therefore, it has to be |
| 1881 |
* assured that the lifetime of the vector exceeds the lifetime of the |
| 1882 |
* preconditioner. |
| 1883 |
* |
| 1884 |
* The relaxation parameter should be larger than zero and smaller than 2 |
| 1885 |
* for numerical reasons. It defaults to 1. |
| 1886 |
*/ |
| 1887 |
void |
| 1888 |
initialize(const MatrixType &A, |
| 1889 |
const std::vector<size_type> &permutation, |
| 1890 |
const std::vector<size_type> &inverse_permutation, |
| 1891 |
const typename BaseClass::AdditionalData ¶meters = |
| 1892 |
typename BaseClass::AdditionalData()); |
| 1893 |
|
| 1894 |
/** |
| 1895 |
* Initialize matrix and relaxation parameter. The matrix is just stored in |
| 1896 |
* the preconditioner object. |
| 1897 |
* |
| 1898 |
* For more detail about possible parameters, see the class documentation |
| 1899 |
* and the documentation of the PreconditionPSOR::AdditionalData class. |
| 1900 |
* |
| 1901 |
* After this function is called the preconditioner is ready to be used |
| 1902 |
* (using the <code>vmult</code> function of derived classes). |
| 1903 |
*/ |
| 1904 |
void |
| 1905 |
initialize(const MatrixType &A, const AdditionalData &additional_data); |
| 1906 |
}; |
| 1907 |
|
| 1908 |
|
| 1909 |
|
| 1910 |
/** |
| 1911 |
* Preconditioning with a Chebyshev polynomial for symmetric positive definite |
| 1912 |
* matrices. This preconditioner is based on an iteration of an inner |
| 1913 |
* preconditioner of type @p PreconditionerType with coefficients that are |
| 1914 |
* adapted to optimally cover an eigenvalue range between the largest |
| 1915 |
* eigenvalue $\lambda_{\max{}}$ down to a given lower eigenvalue |
| 1916 |
* $\lambda_{\min{}}$ specified by the optional parameter |
| 1917 |
* @p smoothing_range. The algorithm is based on the following three-term |
| 1918 |
* recurrence: |
| 1919 |
* @f[ |
| 1920 |
* x^{n+1} = x^{n} + \alpha^n_0 (x^{n} - x^{n-1}) + \alpha^n_1 P^{-1} (b-Ax^n) |
| 1921 |
* \quad\text{with}\quad |
| 1922 |
* \alpha^0_0 := 0,\; |
| 1923 |
* \alpha^0_1 := \frac{2\rho_0}{\lambda_{\max}-\lambda_{\min}}\; |
| 1924 |
* \alpha^n_0 := \rho_n \rho_{n-1},\;\text{and}\; |
| 1925 |
* \alpha^n_1 := \frac{4\rho_n}{\lambda_{\max}-\lambda_{\min}}, |
| 1926 |
* @f] |
| 1927 |
* where the parameter $\rho_0$ is set to $\rho_0 = |
| 1928 |
* \frac{\lambda_{\max{}}-\lambda_{\min{}}}{\lambda_{\max{}}+\lambda_{\min{}}}$ |
| 1929 |
* for the maximal eigenvalue $\lambda_{\max{}}$ and updated via $\rho_n = |
| 1930 |
* \left(2\frac{\lambda_{\max{}}+\lambda_{\min{}}} |
| 1931 |
* {\lambda_{\max{}}-\lambda_{\min{}}} - \rho_{n-1}\right)^{-1}$. The |
| 1932 |
* Chebyshev polynomial is constructed to strongly damp the eigenvalue range |
| 1933 |
* between $\lambda_{\min{}}$ and $\lambda_{\max{}}$ and is visualized e.g. in |
| 1934 |
* Utilities::LinearAlgebra::chebyshev_filter(). |
| 1935 |
* |
| 1936 |
* The typical use case for the preconditioner is a Jacobi preconditioner |
| 1937 |
* specified through DiagonalMatrix, which is also the default value for the |
| 1938 |
* preconditioner. Note that if the degree variable is set to one, the |
| 1939 |
* Chebyshev iteration corresponds to a Jacobi preconditioner (or the |
| 1940 |
* underlying preconditioner type) with relaxation parameter according to the |
| 1941 |
* specified smoothing range. |
| 1942 |
* |
| 1943 |
* Besides the default choice of a pointwise Jacobi preconditioner, this class |
| 1944 |
* also allows for more advanced types of preconditioners, for example |
| 1945 |
* iterating block-Jacobi preconditioners in DG methods. |
| 1946 |
* |
| 1947 |
* Apart from the inner preconditioner object, this iteration does not need |
| 1948 |
* access to matrix entries, which makes it an ideal ingredient for |
| 1949 |
* matrix-free computations. In that context, this class can be used as a |
| 1950 |
* multigrid smoother that is trivially %parallel (assuming that matrix-vector |
| 1951 |
* products are %parallel and the inner preconditioner is %parallel). Its use |
| 1952 |
* is demonstrated in the step-37 and step-59 tutorial programs. |
| 1953 |
* |
| 1954 |
* <h4>Estimation of the eigenvalues</h4> |
| 1955 |
* |
| 1956 |
* The Chebyshev method relies on an estimate of the eigenvalues of the matrix |
| 1957 |
* which are computed during the first invocation of vmult(). This class |
| 1958 |
* offers several algorithms to this end, see |
| 1959 |
* PreconditionChebyshev::AdditionalData::EigenvalueAlgorithm. The default |
| 1960 |
* algorithm invokes the Lanczos method via the SolverCG class, which requires |
| 1961 |
* symmetry and positive definiteness of the (preconditioned) matrix system |
| 1962 |
* are required. Also note that deal.II needs to be configured with LAPACK |
| 1963 |
* support to use this option. The eigenvalue algorithm can be controlled by |
| 1964 |
* PreconditionChebyshev::AdditionalData::eig_cg_n_iterations specifying how |
| 1965 |
* many iterations should be performed. For all algorithms, the iterative |
| 1966 |
* process is started from an initial vector that depends on the vector |
| 1967 |
* type. For the classes dealii::Vector or |
| 1968 |
* dealii::LinearAlgebra::distributed::Vector, which have fast element access, |
| 1969 |
* it is a vector with entries `(-5.5, -4.5, -3.5, -2.5, ..., 3.5, 4.5, 5.5)` |
| 1970 |
* with appropriate epilogue and adjusted such that its mean is always zero, |
| 1971 |
* which works well for the Laplacian. This setup is stable in parallel in the |
| 1972 |
* sense that for a different number of processors but the same ordering of |
| 1973 |
* unknowns, the same initial vector and thus eigenvalue distribution will be |
| 1974 |
* computed, apart from roundoff errors. For other vector types, the initial |
| 1975 |
* vector contains all ones, scaled by the length of the vector, except for |
| 1976 |
* the very first entry that is zero, triggering high-frequency content again. |
| 1977 |
* |
| 1978 |
* The computation of eigenvalues happens the first time one of the vmult(), |
| 1979 |
* Tvmult(), step() or Tstep() functions is called or when |
| 1980 |
* estimate_eigenvalues() is called directly. In the latter case, it is |
| 1981 |
* necessary to provide a temporary vector of the same layout as the source |
| 1982 |
* and destination vectors used during application of the preconditioner. |
| 1983 |
* |
| 1984 |
* The estimates for minimum and maximum eigenvalue are taken from the |
| 1985 |
* underlying solver or eigenvalue algorithm in the given number of |
| 1986 |
* iterations, even if the solver did not converge in the requested number of |
| 1987 |
* iterations. Finally, the maximum eigenvalue is multiplied by a safety |
| 1988 |
* factor of 1.2. |
| 1989 |
* |
| 1990 |
* Due to the cost of the eigenvalue estimate, this class is most appropriate |
| 1991 |
* if it is applied repeatedly, e.g. in a smoother for a geometric multigrid |
| 1992 |
* solver, that can in turn be used to solve several linear systems. |
| 1993 |
* |
| 1994 |
* <h4>Bypassing the eigenvalue computation</h4> |
| 1995 |
* |
| 1996 |
* In some contexts, the automatic eigenvalue computation of this class may |
| 1997 |
* result in a bad quality, e.g. when the polynomial basis or numbering of |
| 1998 |
* unknowns is such that the initial vector described above is a bad |
| 1999 |
* choice. It is possible to bypass the automatic eigenvalue computation by |
| 2000 |
* setting AdditionalData::eig_cg_n_iterations to zero, and provide the |
| 2001 |
* variable AdditionalData::max_eigenvalue instead. The minimal eigenvalue is |
| 2002 |
* implicitly specified via `max_eigenvalue/smoothing_range`. |
| 2003 |
* |
| 2004 |
* <h4>Using the PreconditionChebyshev as a solver</h4> |
| 2005 |
* |
| 2006 |
* If the range <tt>[max_eigenvalue/smoothing_range, max_eigenvalue]</tt> |
| 2007 |
* contains all eigenvalues of the preconditioned matrix system and the degree |
| 2008 |
* (i.e., number of iterations) is high enough, this class can also be used as |
| 2009 |
* a direct solver. For an error estimation of the Chebyshev iteration that |
| 2010 |
* can be used to determine the number of iteration, see @cite Varga2009. |
| 2011 |
* |
| 2012 |
* In order to use Chebyshev as a solver, set the degree to |
| 2013 |
* numbers::invalid_unsigned_int to force the automatic computation of the |
| 2014 |
* number of iterations needed to reach a given target tolerance. Note that |
| 2015 |
* this currently only works for symmetric positive definite matrices with the |
| 2016 |
* eigenvalue algorithm set to the conjugate gradient algorithm. In this case, |
| 2017 |
* the target tolerance is read from the variable |
| 2018 |
* PreconditionChebyshev::AdditionalData::smoothing_range (it needs to be a |
| 2019 |
* number less than one to force any iterations obviously). |
| 2020 |
* |
| 2021 |
* For details on the algorithm, see section 5.1 of @cite Varga2009. |
| 2022 |
* |
| 2023 |
* <h4>Requirements on the templated classes</h4> |
| 2024 |
* |
| 2025 |
* The class `MatrixType` must be derived from |
| 2026 |
* EnableObserverPointer because a ObserverPointer to `MatrixType` |
| 2027 |
* is held in the class. In particular, this means that the matrix object needs |
| 2028 |
* to persist during the lifetime of PreconditionChebyshev. The preconditioner |
| 2029 |
* is held in a shared_ptr that is copied into the AdditionalData member |
| 2030 |
* variable of the class, so the variable used for initialization can safely be |
| 2031 |
* discarded after calling initialize(). Both the matrix and the preconditioner |
| 2032 |
* need to provide |
| 2033 |
* @p vmult() functions for the matrix-vector product and @p m() functions for |
| 2034 |
* accessing the number of rows in the (square) matrix. Furthermore, the |
| 2035 |
* matrix must provide <tt>el(i,i)</tt> methods for accessing the matrix |
| 2036 |
* diagonal in case the preconditioner type is DiagonalMatrix. Even though |
| 2037 |
* it is highly recommended to pass the inverse diagonal entries inside a |
| 2038 |
* separate preconditioner object for implementing the Jacobi method (which is |
| 2039 |
* the only possible way to operate this class when computing in %parallel |
| 2040 |
* with MPI because there is no knowledge about the locally stored range of |
| 2041 |
* entries that would be needed from the matrix alone), there is a backward |
| 2042 |
* compatibility function that can extract the diagonal in case of a serial |
| 2043 |
* computation. |
| 2044 |
* |
| 2045 |
* <h4>Optimized operations with specific `MatrixType` argument</h4> |
| 2046 |
* |
| 2047 |
* This class enables to embed the vector updates into the matrix-vector |
| 2048 |
* product in case the `MatrixType` supports this. To this end, the |
| 2049 |
* `VectorType` needs to be of type LinearAlgebra::distributed::Vector, the |
| 2050 |
* `PreconditionerType` needs to be DiagonalMatrix, and the class `MatrixType` |
| 2051 |
* needs to provide a function with the signature |
| 2052 |
* @code |
| 2053 |
* void MatrixType::vmult( |
| 2054 |
* VectorType &, |
| 2055 |
* const VectorType &, |
| 2056 |
* const std::function<void(const unsigned int, const unsigned int)> &, |
| 2057 |
* const std::function<void(const unsigned int, const unsigned int)> &) const |
| 2058 |
* @endcode |
| 2059 |
* where the two given functions run before and after the matrix-vector |
| 2060 |
* product, respectively. They take as arguments a sub-range among the locally |
| 2061 |
* owned elements of the vector, defined as half-open intervals. The intervals |
| 2062 |
* are designed to be scheduled close to the time the matrix-vector product |
| 2063 |
* touches upon the entries in the `src` and `dst` vectors, respectively, with |
| 2064 |
* the requirement that |
| 2065 |
* <ul> |
| 2066 |
* <li> the matrix-vector product may only access an entry in `src` or `dst` |
| 2067 |
* once the `operation_before_matrix_vector_product` has been run on that |
| 2068 |
* vector entry; </li> |
| 2069 |
* <li> `operation_after_matrix_vector_product` may run on a range of entries |
| 2070 |
* `[i,j)` once the matrix-vector product does not access the entries `[i,j)` |
| 2071 |
* in `src` and `dst` any more. </li> |
| 2072 |
* </ul> |
| 2073 |
* The motivation for this function is to increase data locality and hence |
| 2074 |
* cache usage. For the example of a class similar to the one in the step-37 |
| 2075 |
* tutorial program, the implementation is |
| 2076 |
* @code |
| 2077 |
* void |
| 2078 |
* vmult(LinearAlgebra::distributed::Vector<number> & dst, |
| 2079 |
* const LinearAlgebra::distributed::Vector<number> &src, |
| 2080 |
* const std::function<void(const unsigned int, const unsigned int)> |
| 2081 |
* &operation_before_matrix_vector_product, |
| 2082 |
* const std::function<void(const unsigned int, const unsigned int)> |
| 2083 |
* &operation_after_matrix_vector_product) const |
| 2084 |
* { |
| 2085 |
* data.cell_loop(&LaplaceOperator::local_apply, |
| 2086 |
* this, |
| 2087 |
* dst, |
| 2088 |
* src, |
| 2089 |
* operation_before_matrix_vector_product, |
| 2090 |
* operation_after_matrix_vector_product); |
| 2091 |
* } |
| 2092 |
* @endcode |
| 2093 |
* In terms of the Chebyshev iteration, the operation before the loop will |
| 2094 |
* set `dst` to zero, whereas the operation after the loop performs the |
| 2095 |
* iteration leading to $x^{n+1}$ described above, modifying the `dst` and |
| 2096 |
* `src` vectors. |
| 2097 |
*/ |
| 2098 |
template <typename MatrixType = SparseMatrix<double>, |
| 2099 |
typename VectorType = Vector<double>, |
| 2100 |
typename PreconditionerType = DiagonalMatrix<VectorType>> |
| 2101 |
class PreconditionChebyshev : public EnableObserverPointer |
| 2102 |
{ |
| 2103 |
public: |
| 2104 |
/** |
| 2105 |
* Declare type for container size. |
| 2106 |
*/ |
| 2107 |
using size_type = types::global_dof_index; |
| 2108 |
|
| 2109 |
/** |
| 2110 |
* Standardized data struct to pipe additional parameters to the |
| 2111 |
* preconditioner. |
| 2112 |
*/ |
| 2113 |
struct AdditionalData |
| 2114 |
: public internal::EigenvalueAlgorithmAdditionalData<PreconditionerType> |
| 2115 |
{ |
| 2116 |
using EigenvalueAlgorithm = internal::EigenvalueAlgorithm; |
| 2117 |
|
| 2118 |
/** |
| 2119 |
* An enum to define the available types of polynomial types. |
| 2120 |
*/ |
| 2121 |
enum class PolynomialType |
| 2122 |
{ |
| 2123 |
/** |
| 2124 |
* First-kind Chebyshev polynomials. |
| 2125 |
*/ |
| 2126 |
first_kind, |
| 2127 |
/** |
| 2128 |
* Fourth-kind Chebyshev polynomials according to @cite lottes2022optimal |
| 2129 |
* and @cite phillips2022optimal. |
| 2130 |
*/ |
| 2131 |
fourth_kind |
| 2132 |
}; |
| 2133 |
|
| 2134 |
/** |
| 2135 |
* Constructor. |
| 2136 |
*/ |
| 2137 |
AdditionalData( |
| 2138 |
const unsigned int degree = 1, |
| 2139 |
const double smoothing_range = 0., |
| 2140 |
const unsigned int eig_cg_n_iterations = 8, |
| 2141 |
const double eig_cg_residual = 1e-2, |
| 2142 |
const double max_eigenvalue = 1, |
| 2143 |
const EigenvalueAlgorithm eigenvalue_algorithm = |
| 2144 |
EigenvalueAlgorithm::lanczos, |
| 2145 |
const PolynomialType polynomial_type = PolynomialType::first_kind); |
| 2146 |
|
| 2147 |
/** |
| 2148 |
* This determines the degree of the Chebyshev polynomial. The degree of |
| 2149 |
* the polynomial gives the number of matrix-vector products to be |
| 2150 |
* performed for one application of the step() operation. During vmult(), |
| 2151 |
* the method performs `(degree-1)` matrix-vector products. Degree one |
| 2152 |
* corresponds to a damped Jacobi method. |
| 2153 |
* |
| 2154 |
* If the degree is set to numbers::invalid_unsigned_int, the algorithm |
| 2155 |
* will automatically determine the number of necessary iterations based |
| 2156 |
* on the usual Chebyshev error formula as mentioned in the discussion of |
| 2157 |
* the main class. |
| 2158 |
*/ |
| 2159 |
unsigned int degree; |
| 2160 |
|
| 2161 |
/** |
| 2162 |
* Specifies the polynomial type to be used. |
| 2163 |
*/ |
| 2164 |
PolynomialType polynomial_type; |
| 2165 |
}; |
| 2166 |
|
| 2167 |
|
| 2168 |
/** |
| 2169 |
* Constructor. |
| 2170 |
*/ |
| 2171 |
PreconditionChebyshev(); |
| 2172 |
|
| 2173 |
/** |
| 2174 |
* Initialize function. Takes the matrix which is used to form the |
| 2175 |
* preconditioner, and additional flags if there are any. This function |
| 2176 |
* works only if the input matrix has an operator <tt>el(i,i)</tt> for |
| 2177 |
* accessing all the elements in the diagonal. Alternatively, the diagonal |
| 2178 |
* can be supplied with the help of the AdditionalData field. |
| 2179 |
* |
| 2180 |
* This function calculates an estimate of the eigenvalue range of the |
| 2181 |
* matrix weighted by its diagonal using a modified CG iteration in case the |
| 2182 |
* given number of iterations is positive. |
| 2183 |
*/ |
| 2184 |
void |
| 2185 |
initialize(const MatrixType &matrix, |
| 2186 |
const AdditionalData &additional_data = AdditionalData()); |
| 2187 |
|
| 2188 |
/** |
| 2189 |
* Compute the action of the preconditioner on <tt>src</tt>, storing the |
| 2190 |
* result in <tt>dst</tt>. |
| 2191 |
*/ |
| 2192 |
void |
| 2193 |
vmult(VectorType &dst, const VectorType &src) const; |
| 2194 |
|
| 2195 |
/** |
| 2196 |
* Compute the action of the transposed preconditioner on <tt>src</tt>, |
| 2197 |
* storing the result in <tt>dst</tt>. |
| 2198 |
*/ |
| 2199 |
void |
| 2200 |
Tvmult(VectorType &dst, const VectorType &src) const; |
| 2201 |
|
| 2202 |
/** |
| 2203 |
* Perform one step of the preconditioned Richardson iteration. |
| 2204 |
*/ |
| 2205 |
void |
| 2206 |
step(VectorType &dst, const VectorType &src) const; |
| 2207 |
|
| 2208 |
/** |
| 2209 |
* Perform one transposed step of the preconditioned Richardson iteration. |
| 2210 |
*/ |
| 2211 |
void |
| 2212 |
Tstep(VectorType &dst, const VectorType &src) const; |
| 2213 |
|
| 2214 |
/** |
| 2215 |
* Resets the preconditioner. |
| 2216 |
*/ |
| 2217 |
void |
| 2218 |
clear(); |
| 2219 |
|
| 2220 |
/** |
| 2221 |
* Return the dimension of the codomain (or range) space. Note that the |
| 2222 |
* matrix is of dimension $m \times n$. |
| 2223 |
*/ |
| 2224 |
size_type |
| 2225 |
m() const; |
| 2226 |
|
| 2227 |
/** |
| 2228 |
* Return the dimension of the domain space. Note that the matrix is of |
| 2229 |
* dimension $m \times n$. |
| 2230 |
*/ |
| 2231 |
size_type |
| 2232 |
n() const; |
| 2233 |
|
| 2234 |
using EigenvalueInformation = internal::EigenvalueInformation; |
| 2235 |
|
| 2236 |
/** |
| 2237 |
* Compute eigenvalue estimates required for the preconditioner. |
| 2238 |
* |
| 2239 |
* This function is called automatically on first use of the preconditioner |
| 2240 |
* if it is not called by the user. The layout of the vector @p src is used |
| 2241 |
* to create internal temporary vectors and its content does not matter. |
| 2242 |
* |
| 2243 |
* Initializes the factors theta and delta based on an eigenvalue |
| 2244 |
* computation. If the user set provided values for the largest eigenvalue |
| 2245 |
* in AdditionalData, no computation is performed and the information given |
| 2246 |
* by the user is used. |
| 2247 |
*/ |
| 2248 |
EigenvalueInformation |
| 2249 |
estimate_eigenvalues(const VectorType &src) const; |
| 2250 |
|
| 2251 |
private: |
| 2252 |
/** |
| 2253 |
* A pointer to the underlying matrix. |
| 2254 |
*/ |
| 2255 |
ObserverPointer< |
| 2256 |
const MatrixType, |
| 2257 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>> |
| 2258 |
matrix_ptr; |
| 2259 |
|
| 2260 |
/** |
| 2261 |
* Internal vector used for the <tt>vmult</tt> operation. |
| 2262 |
*/ |
| 2263 |
mutable VectorType solution_old; |
| 2264 |
|
| 2265 |
/** |
| 2266 |
* Internal vector used for the <tt>vmult</tt> operation. |
| 2267 |
*/ |
| 2268 |
mutable VectorType temp_vector1; |
| 2269 |
|
| 2270 |
/** |
| 2271 |
* Internal vector used for the <tt>vmult</tt> operation. |
| 2272 |
*/ |
| 2273 |
mutable VectorType temp_vector2; |
| 2274 |
|
| 2275 |
/** |
| 2276 |
* Stores the additional data passed to the initialize function, obtained |
| 2277 |
* through a copy operation. |
| 2278 |
*/ |
| 2279 |
AdditionalData data; |
| 2280 |
|
| 2281 |
/** |
| 2282 |
* Average of the largest and smallest eigenvalue under consideration. |
| 2283 |
*/ |
| 2284 |
double theta; |
| 2285 |
|
| 2286 |
/** |
| 2287 |
* Half the interval length between the largest and smallest eigenvalue |
| 2288 |
* under consideration. |
| 2289 |
*/ |
| 2290 |
double delta; |
| 2291 |
|
| 2292 |
/** |
| 2293 |
* Stores whether the preconditioner has been set up and eigenvalues have |
| 2294 |
* been computed. |
| 2295 |
*/ |
| 2296 |
bool eigenvalues_are_initialized; |
| 2297 |
|
| 2298 |
/** |
| 2299 |
* A mutex to avoid that multiple vmult() invocations by different threads |
| 2300 |
* overwrite the temporary vectors. |
| 2301 |
*/ |
| 2302 |
mutable Threads::Mutex mutex; |
| 2303 |
}; |
| 2304 |
|
| 2305 |
|
| 2306 |
|
| 2307 |
/** @} */ |
| 2308 |
/* ---------------------------------- Inline functions ------------------- */ |
| 2309 |
|
| 2310 |
#ifndef DOXYGEN |
| 2311 |
|
| 2312 |
|
| 2313 |
namespace internal |
| 2314 |
{ |
| 2315 |
template <typename VectorType> |
| 2316 |
void |
| 2317 |
set_initial_guess(VectorType &vector) |
| 2318 |
{ |
| 2319 |
vector = 1. / std::sqrt(static_cast<double>(vector.size())); |
| 2320 |
if (vector.locally_owned_elements().is_element(0)) |
| 2321 |
vector(0) = 0.; |
| 2322 |
} |
| 2323 |
|
| 2324 |
template <typename Number> |
| 2325 |
void |
| 2326 |
set_initial_guess(::dealii::Vector<Number> &vector) |
| 2327 |
{ |
| 2328 |
// Choose a high-frequency mode consisting of numbers between 0 and 1 |
| 2329 |
// that is cheap to compute (cheaper than random numbers) but avoids |
| 2330 |
// obviously re-occurring numbers in multi-component systems by choosing |
| 2331 |
// a period of 11 |
| 2332 |
for (unsigned int i = 0; i < vector.size(); ++i) |
| 2333 |
vector(i) = i % 11; |
| 2334 |
|
| 2335 |
const Number mean_value = vector.mean_value(); |
| 2336 |
vector.add(-mean_value); |
| 2337 |
} |
| 2338 |
|
| 2339 |
template <typename Number, typename MemorySpace> |
| 2340 |
void |
| 2341 |
set_initial_guess( |
| 2342 |
::dealii::LinearAlgebra::distributed::BlockVector<Number, MemorySpace> |
| 2343 |
&vector) |
| 2344 |
{ |
| 2345 |
for (unsigned int block = 0; block < vector.n_blocks(); ++block) |
| 2346 |
set_initial_guess(vector.block(block)); |
| 2347 |
} |
| 2348 |
|
| 2349 |
template <typename Number, typename MemorySpace> |
| 2350 |
void |
| 2351 |
set_initial_guess( |
| 2352 |
::dealii::LinearAlgebra::distributed::Vector<Number, MemorySpace> &vector) |
| 2353 |
{ |
| 2354 |
// Choose a high-frequency mode consisting of numbers between 0 and 1 |
| 2355 |
// that is cheap to compute (cheaper than random numbers) but avoids |
| 2356 |
// obviously re-occurring numbers in multi-component systems by choosing |
| 2357 |
// a period of 11. |
| 2358 |
// Make initial guess robust with respect to number of processors |
| 2359 |
// by operating on the global index. |
| 2360 |
types::global_dof_index first_local_range = 0; |
| 2361 |
if (!vector.locally_owned_elements().is_empty()) |
| 2362 |
first_local_range = vector.locally_owned_elements().nth_index_in_set(0); |
| 2363 |
|
| 2364 |
const auto n_local_elements = vector.locally_owned_size(); |
| 2365 |
Number *values_ptr = vector.get_values(); |
| 2366 |
Kokkos::RangePolicy<typename MemorySpace::kokkos_space::execution_space, |
| 2367 |
Kokkos::IndexType<types::global_dof_index>> |
| 2368 |
policy(0, n_local_elements); |
| 2369 |
Kokkos::parallel_for( |
| 2370 |
"dealii::PreconditionChebyshev::set_initial_guess", |
| 2371 |
policy, |
| 2372 |
KOKKOS_LAMBDA(types::global_dof_index i) { |
| 2373 |
values_ptr[i] = (i + first_local_range) % 11; |
| 2374 |
}); |
| 2375 |
const Number mean_value = vector.mean_value(); |
| 2376 |
vector.add(-mean_value); |
| 2377 |
} |
| 2378 |
|
| 2379 |
struct EigenvalueTracker |
| 2380 |
{ |
| 2381 |
public: |
| 2382 |
void |
| 2383 |
slot(const std::vector<double> &eigenvalues) |
| 2384 |
{ |
| 2385 |
values = eigenvalues; |
| 2386 |
} |
| 2387 |
|
| 2388 |
std::vector<double> values; |
| 2389 |
}; |
| 2390 |
|
| 2391 |
|
| 2392 |
|
| 2393 |
template <typename MatrixType, |
| 2394 |
typename VectorType, |
| 2395 |
typename PreconditionerType> |
| 2396 |
double |
| 2397 |
power_iteration(const MatrixType &matrix, |
| 2398 |
VectorType &eigenvector, |
| 2399 |
const PreconditionerType &preconditioner, |
| 2400 |
const unsigned int n_iterations) |
| 2401 |
{ |
| 2402 |
typename VectorType::value_type eigenvalue_estimate = 0.; |
| 2403 |
eigenvector /= eigenvector.l2_norm(); |
| 2404 |
VectorType vector1, vector2; |
| 2405 |
vector1.reinit(eigenvector, true); |
| 2406 |
if (!std::is_same_v<PreconditionerType, PreconditionIdentity>) |
| 2407 |
vector2.reinit(eigenvector, true); |
| 2408 |
for (unsigned int i = 0; i < n_iterations; ++i) |
| 2409 |
{ |
| 2410 |
if (!std::is_same_v<PreconditionerType, PreconditionIdentity>) |
| 2411 |
{ |
| 2412 |
matrix.vmult(vector2, eigenvector); |
| 2413 |
preconditioner.vmult(vector1, vector2); |
| 2414 |
} |
| 2415 |
else |
| 2416 |
matrix.vmult(vector1, eigenvector); |
| 2417 |
|
| 2418 |
eigenvalue_estimate = eigenvector * vector1; |
| 2419 |
|
| 2420 |
vector1 /= vector1.l2_norm(); |
| 2421 |
eigenvector.swap(vector1); |
| 2422 |
} |
| 2423 |
return std::abs(eigenvalue_estimate); |
| 2424 |
} |
| 2425 |
|
| 2426 |
|
| 2427 |
|
| 2428 |
template <typename MatrixType, |
| 2429 |
typename VectorType, |
| 2430 |
typename PreconditionerType> |
| 2431 |
EigenvalueInformation |
| 2432 |
estimate_eigenvalues( |
| 2433 |
const EigenvalueAlgorithmAdditionalData<PreconditionerType> &data, |
| 2434 |
const MatrixType *matrix_ptr, |
| 2435 |
VectorType &solution_old, |
| 2436 |
VectorType &temp_vector1, |
| 2437 |
const unsigned int degree) |
| 2438 |
{ |
| 2439 |
Assert(data.preconditioner.get() != nullptr, ExcNotInitialized()); |
| 2440 |
|
| 2441 |
EigenvalueInformation info{}; |
| 2442 |
|
| 2443 |
if (data.eig_cg_n_iterations > 0) |
| 2444 |
{ |
| 2445 |
Assert(data.eig_cg_n_iterations > 2, |
| 2446 |
ExcMessage( |
| 2447 |
"Need to set at least two iterations to find eigenvalues.")); |
| 2448 |
|
| 2449 |
internal::EigenvalueTracker eigenvalue_tracker; |
| 2450 |
|
| 2451 |
// set an initial guess that contains some high-frequency parts (to the |
| 2452 |
// extent possible without knowing the discretization and the numbering) |
| 2453 |
// to trigger high eigenvalues according to the external function |
| 2454 |
internal::set_initial_guess(temp_vector1); |
| 2455 |
data.constraints.set_zero(temp_vector1); |
| 2456 |
|
| 2457 |
if (data.eigenvalue_algorithm == internal::EigenvalueAlgorithm::lanczos) |
| 2458 |
{ |
| 2459 |
// set a very strict tolerance to force at least two iterations |
| 2460 |
IterationNumberControl control(data.eig_cg_n_iterations, |
| 2461 |
1e-10, |
| 2462 |
false, |
| 2463 |
false); |
| 2464 |
|
| 2465 |
dealii::SolverCG<VectorType> solver(control); |
| 2466 |
solver.connect_eigenvalues_slot( |
| 2467 |
[&eigenvalue_tracker](const std::vector<double> &eigenvalues) { |
| 2468 |
eigenvalue_tracker.slot(eigenvalues); |
| 2469 |
}); |
| 2470 |
|
| 2471 |
solver.solve(*matrix_ptr, |
| 2472 |
solution_old, |
| 2473 |
temp_vector1, |
| 2474 |
*data.preconditioner); |
| 2475 |
|
| 2476 |
info.cg_iterations = control.last_step(); |
| 2477 |
} |
| 2478 |
else if (data.eigenvalue_algorithm == |
| 2479 |
internal::EigenvalueAlgorithm::power_iteration) |
| 2480 |
{ |
| 2481 |
(void)degree; |
| 2482 |
|
| 2483 |
Assert(degree != numbers::invalid_unsigned_int, |
| 2484 |
ExcMessage("Cannot estimate the minimal eigenvalue with the " |
| 2485 |
"power iteration")); |
| 2486 |
|
| 2487 |
eigenvalue_tracker.values.push_back( |
| 2488 |
internal::power_iteration(*matrix_ptr, |
| 2489 |
temp_vector1, |
| 2490 |
*data.preconditioner, |
| 2491 |
data.eig_cg_n_iterations)); |
| 2492 |
} |
| 2493 |
else |
| 2494 |
DEAL_II_NOT_IMPLEMENTED(); |
| 2495 |
|
| 2496 |
// read the eigenvalues from the attached eigenvalue tracker |
| 2497 |
if (eigenvalue_tracker.values.empty()) |
| 2498 |
info.min_eigenvalue_estimate = info.max_eigenvalue_estimate = 1.; |
| 2499 |
else |
| 2500 |
{ |
| 2501 |
info.min_eigenvalue_estimate = eigenvalue_tracker.values.front(); |
| 2502 |
|
| 2503 |
// include a safety factor since the CG method will in general not |
| 2504 |
// be converged |
| 2505 |
info.max_eigenvalue_estimate = |
| 2506 |
1.2 * eigenvalue_tracker.values.back(); |
| 2507 |
} |
| 2508 |
} |
| 2509 |
else |
| 2510 |
{ |
| 2511 |
info.max_eigenvalue_estimate = data.max_eigenvalue; |
| 2512 |
info.min_eigenvalue_estimate = |
| 2513 |
data.max_eigenvalue / data.smoothing_range; |
| 2514 |
} |
| 2515 |
|
| 2516 |
return info; |
| 2517 |
} |
| 2518 |
} // namespace internal |
| 2519 |
|
| 2520 |
|
| 2521 |
|
| 2522 |
inline PreconditionIdentity::PreconditionIdentity() |
| 2523 |
: n_rows(0) |
| 2524 |
, n_columns(0) |
| 2525 |
{} |
| 2526 |
|
| 2527 |
template <typename MatrixType> |
| 2528 |
inline void |
| 2529 |
PreconditionIdentity::initialize(const MatrixType &matrix, |
| 2530 |
const PreconditionIdentity::AdditionalData &) |
| 2531 |
{ |
| 2532 |
n_rows = matrix.m(); |
| 2533 |
n_columns = matrix.n(); |
| 2534 |
} |
| 2535 |
|
| 2536 |
|
| 2537 |
template <typename VectorType> |
| 2538 |
inline void |
| 2539 |
PreconditionIdentity::vmult(VectorType &dst, const VectorType &src) const |
| 2540 |
{ |
| 2541 |
dst = src; |
| 2542 |
} |
| 2543 |
|
| 2544 |
|
| 2545 |
|
| 2546 |
template <typename VectorType> |
| 2547 |
inline void |
| 2548 |
PreconditionIdentity::Tvmult(VectorType &dst, const VectorType &src) const |
| 2549 |
{ |
| 2550 |
dst = src; |
| 2551 |
} |
| 2552 |
|
| 2553 |
template <typename VectorType> |
| 2554 |
inline void |
| 2555 |
PreconditionIdentity::vmult_add(VectorType &dst, const VectorType &src) const |
| 2556 |
{ |
| 2557 |
dst += src; |
| 2558 |
} |
| 2559 |
|
| 2560 |
|
| 2561 |
|
| 2562 |
template <typename VectorType> |
| 2563 |
inline void |
| 2564 |
PreconditionIdentity::Tvmult_add(VectorType &dst, const VectorType &src) const |
| 2565 |
{ |
| 2566 |
dst += src; |
| 2567 |
} |
| 2568 |
|
| 2569 |
|
| 2570 |
|
| 2571 |
inline void |
| 2572 |
PreconditionIdentity::clear() |
| 2573 |
{} |
| 2574 |
|
| 2575 |
|
| 2576 |
|
| 2577 |
inline PreconditionIdentity::size_type |
| 2578 |
PreconditionIdentity::m() const |
| 2579 |
{ |
| 2580 |
Assert(n_rows != 0, ExcNotInitialized()); |
| 2581 |
return n_rows; |
| 2582 |
} |
| 2583 |
|
| 2584 |
inline PreconditionIdentity::size_type |
| 2585 |
PreconditionIdentity::n() const |
| 2586 |
{ |
| 2587 |
Assert(n_columns != 0, ExcNotInitialized()); |
| 2588 |
return n_columns; |
| 2589 |
} |
| 2590 |
|
| 2591 |
//--------------------------------------------------------------------------- |
| 2592 |
|
| 2593 |
inline PreconditionRichardson::AdditionalData::AdditionalData( |
| 2594 |
const double relaxation) |
| 2595 |
: relaxation(relaxation) |
| 2596 |
{} |
| 2597 |
|
| 2598 |
|
| 2599 |
inline PreconditionRichardson::PreconditionRichardson() |
| 2600 |
: relaxation(0) |
| 2601 |
, n_rows(0) |
| 2602 |
, n_columns(0) |
| 2603 |
{ |
| 2604 |
AdditionalData add_data; |
| 2605 |
relaxation = add_data.relaxation; |
| 2606 |
} |
| 2607 |
|
| 2608 |
|
| 2609 |
|
| 2610 |
inline void |
| 2611 |
PreconditionRichardson::initialize( |
| 2612 |
const PreconditionRichardson::AdditionalData ¶meters) |
| 2613 |
{ |
| 2614 |
relaxation = parameters.relaxation; |
| 2615 |
} |
| 2616 |
|
| 2617 |
|
| 2618 |
|
| 2619 |
template <typename MatrixType> |
| 2620 |
inline void |
| 2621 |
PreconditionRichardson::initialize( |
| 2622 |
const MatrixType &matrix, |
| 2623 |
const PreconditionRichardson::AdditionalData ¶meters) |
| 2624 |
{ |
| 2625 |
relaxation = parameters.relaxation; |
| 2626 |
n_rows = matrix.m(); |
| 2627 |
n_columns = matrix.n(); |
| 2628 |
} |
| 2629 |
|
| 2630 |
|
| 2631 |
|
| 2632 |
template <typename VectorType> |
| 2633 |
inline void |
| 2634 |
PreconditionRichardson::vmult(VectorType &dst, const VectorType &src) const |
| 2635 |
{ |
| 2636 |
static_assert( |
| 2637 |
std::is_same_v<size_type, typename VectorType::size_type>, |
| 2638 |
"PreconditionRichardson and VectorType must have the same size_type."); |
| 2639 |
|
| 2640 |
dst.equ(relaxation, src); |
| 2641 |
} |
| 2642 |
|
| 2643 |
|
| 2644 |
|
| 2645 |
template <typename VectorType> |
| 2646 |
inline void |
| 2647 |
PreconditionRichardson::Tvmult(VectorType &dst, const VectorType &src) const |
| 2648 |
{ |
| 2649 |
static_assert( |
| 2650 |
std::is_same_v<size_type, typename VectorType::size_type>, |
| 2651 |
"PreconditionRichardson and VectorType must have the same size_type."); |
| 2652 |
|
| 2653 |
dst.equ(relaxation, src); |
| 2654 |
} |
| 2655 |
|
| 2656 |
template <typename VectorType> |
| 2657 |
inline void |
| 2658 |
PreconditionRichardson::vmult_add(VectorType &dst, const VectorType &src) const |
| 2659 |
{ |
| 2660 |
static_assert( |
| 2661 |
std::is_same_v<size_type, typename VectorType::size_type>, |
| 2662 |
"PreconditionRichardson and VectorType must have the same size_type."); |
| 2663 |
|
| 2664 |
dst.add(relaxation, src); |
| 2665 |
} |
| 2666 |
|
| 2667 |
|
| 2668 |
|
| 2669 |
template <typename VectorType> |
| 2670 |
inline void |
| 2671 |
PreconditionRichardson::Tvmult_add(VectorType &dst, const VectorType &src) const |
| 2672 |
{ |
| 2673 |
static_assert( |
| 2674 |
std::is_same_v<size_type, typename VectorType::size_type>, |
| 2675 |
"PreconditionRichardson and VectorType must have the same size_type."); |
| 2676 |
|
| 2677 |
dst.add(relaxation, src); |
| 2678 |
} |
| 2679 |
|
| 2680 |
inline PreconditionRichardson::size_type |
| 2681 |
PreconditionRichardson::m() const |
| 2682 |
{ |
| 2683 |
Assert(n_rows != 0, ExcNotInitialized()); |
| 2684 |
return n_rows; |
| 2685 |
} |
| 2686 |
|
| 2687 |
inline PreconditionRichardson::size_type |
| 2688 |
PreconditionRichardson::n() const |
| 2689 |
{ |
| 2690 |
Assert(n_columns != 0, ExcNotInitialized()); |
| 2691 |
return n_columns; |
| 2692 |
} |
| 2693 |
|
| 2694 |
//--------------------------------------------------------------------------- |
| 2695 |
|
| 2696 |
template <typename MatrixType, typename PreconditionerType> |
| 2697 |
inline void |
| 2698 |
PreconditionRelaxation<MatrixType, PreconditionerType>::initialize( |
| 2699 |
const MatrixType &rA, |
| 2700 |
const AdditionalData ¶meters) |
| 2701 |
{ |
| 2702 |
A = &rA; |
| 2703 |
eigenvalues_are_initialized = false; |
| 2704 |
|
| 2705 |
Assert(parameters.preconditioner, ExcNotInitialized()); |
| 2706 |
|
| 2707 |
this->data = parameters; |
| 2708 |
} |
| 2709 |
|
| 2710 |
|
| 2711 |
template <typename MatrixType, typename PreconditionerType> |
| 2712 |
inline void |
| 2713 |
PreconditionRelaxation<MatrixType, PreconditionerType>::clear() |
| 2714 |
{ |
| 2715 |
eigenvalues_are_initialized = false; |
| 2716 |
A = nullptr; |
| 2717 |
data.relaxation = 1.0; |
| 2718 |
data.preconditioner = nullptr; |
| 2719 |
} |
| 2720 |
|
| 2721 |
template <typename MatrixType, typename PreconditionerType> |
| 2722 |
inline |
| 2723 |
typename PreconditionRelaxation<MatrixType, PreconditionerType>::size_type |
| 2724 |
PreconditionRelaxation<MatrixType, PreconditionerType>::m() const |
| 2725 |
{ |
| 2726 |
Assert(A != nullptr, ExcNotInitialized()); |
| 2727 |
return A->m(); |
| 2728 |
} |
| 2729 |
|
| 2730 |
template <typename MatrixType, typename PreconditionerType> |
| 2731 |
inline |
| 2732 |
typename PreconditionRelaxation<MatrixType, PreconditionerType>::size_type |
| 2733 |
PreconditionRelaxation<MatrixType, PreconditionerType>::n() const |
| 2734 |
{ |
| 2735 |
Assert(A != nullptr, ExcNotInitialized()); |
| 2736 |
return A->n(); |
| 2737 |
} |
| 2738 |
|
| 2739 |
template <typename MatrixType, typename PreconditionerType> |
| 2740 |
template <typename VectorType> |
| 2741 |
inline void |
| 2742 |
PreconditionRelaxation<MatrixType, PreconditionerType>::vmult( |
| 2743 |
VectorType &dst, |
| 2744 |
const VectorType &src) const |
| 2745 |
{ |
| 2746 |
Assert(this->A != nullptr, ExcNotInitialized()); |
| 2747 |
Assert(data.preconditioner != nullptr, ExcNotInitialized()); |
| 2748 |
|
| 2749 |
if (eigenvalues_are_initialized == false) |
| 2750 |
estimate_eigenvalues(src); |
| 2751 |
|
| 2752 |
VectorType tmp1, tmp2; |
| 2753 |
|
| 2754 |
for (unsigned int i = 0; i < data.n_iterations; ++i) |
| 2755 |
internal::PreconditionRelaxation::step_operations(*A, |
| 2756 |
*data.preconditioner, |
| 2757 |
dst, |
| 2758 |
src, |
| 2759 |
data.relaxation, |
| 2760 |
tmp1, |
| 2761 |
tmp2, |
| 2762 |
i, |
| 2763 |
false); |
| 2764 |
} |
| 2765 |
|
| 2766 |
template <typename MatrixType, typename PreconditionerType> |
| 2767 |
template <typename VectorType> |
| 2768 |
inline void |
| 2769 |
PreconditionRelaxation<MatrixType, PreconditionerType>::Tvmult( |
| 2770 |
VectorType &dst, |
| 2771 |
const VectorType &src) const |
| 2772 |
{ |
| 2773 |
Assert(this->A != nullptr, ExcNotInitialized()); |
| 2774 |
Assert(data.preconditioner != nullptr, ExcNotInitialized()); |
| 2775 |
|
| 2776 |
if (eigenvalues_are_initialized == false) |
| 2777 |
estimate_eigenvalues(src); |
| 2778 |
|
| 2779 |
VectorType tmp1, tmp2; |
| 2780 |
|
| 2781 |
for (unsigned int i = 0; i < data.n_iterations; ++i) |
| 2782 |
internal::PreconditionRelaxation::step_operations( |
| 2783 |
*A, *data.preconditioner, dst, src, data.relaxation, tmp1, tmp2, i, true); |
| 2784 |
} |
| 2785 |
|
| 2786 |
template <typename MatrixType, typename PreconditionerType> |
| 2787 |
template <typename VectorType> |
| 2788 |
inline void |
| 2789 |
PreconditionRelaxation<MatrixType, PreconditionerType>::step( |
| 2790 |
VectorType &dst, |
| 2791 |
const VectorType &src) const |
| 2792 |
{ |
| 2793 |
Assert(this->A != nullptr, ExcNotInitialized()); |
| 2794 |
Assert(data.preconditioner != nullptr, ExcNotInitialized()); |
| 2795 |
|
| 2796 |
if (eigenvalues_are_initialized == false) |
| 2797 |
estimate_eigenvalues(src); |
| 2798 |
|
| 2799 |
VectorType tmp1, tmp2; |
| 2800 |
|
| 2801 |
for (unsigned int i = 1; i <= data.n_iterations; ++i) |
| 2802 |
internal::PreconditionRelaxation::step_operations(*A, |
| 2803 |
*data.preconditioner, |
| 2804 |
dst, |
| 2805 |
src, |
| 2806 |
data.relaxation, |
| 2807 |
tmp1, |
| 2808 |
tmp2, |
| 2809 |
i, |
| 2810 |
false); |
| 2811 |
} |
| 2812 |
|
| 2813 |
template <typename MatrixType, typename PreconditionerType> |
| 2814 |
template <typename VectorType> |
| 2815 |
inline void |
| 2816 |
PreconditionRelaxation<MatrixType, PreconditionerType>::Tstep( |
| 2817 |
VectorType &dst, |
| 2818 |
const VectorType &src) const |
| 2819 |
{ |
| 2820 |
Assert(this->A != nullptr, ExcNotInitialized()); |
| 2821 |
Assert(data.preconditioner != nullptr, ExcNotInitialized()); |
| 2822 |
|
| 2823 |
if (eigenvalues_are_initialized == false) |
| 2824 |
estimate_eigenvalues(src); |
| 2825 |
|
| 2826 |
VectorType tmp1, tmp2; |
| 2827 |
|
| 2828 |
for (unsigned int i = 1; i <= data.n_iterations; ++i) |
| 2829 |
internal::PreconditionRelaxation::step_operations( |
| 2830 |
*A, *data.preconditioner, dst, src, data.relaxation, tmp1, tmp2, i, true); |
| 2831 |
} |
| 2832 |
|
| 2833 |
template <typename MatrixType, typename PreconditionerType> |
| 2834 |
template <typename VectorType> |
| 2835 |
inline internal::EigenvalueInformation |
| 2836 |
PreconditionRelaxation<MatrixType, PreconditionerType>::estimate_eigenvalues( |
| 2837 |
const VectorType &src) const |
| 2838 |
{ |
| 2839 |
Assert(eigenvalues_are_initialized == false, ExcInternalError()); |
| 2840 |
|
| 2841 |
EigenvalueInformation info; |
| 2842 |
|
| 2843 |
if (data.relaxation == 0.0) |
| 2844 |
{ |
| 2845 |
VectorType solution_old, temp_vector1; |
| 2846 |
|
| 2847 |
solution_old.reinit(src); |
| 2848 |
temp_vector1.reinit(src, true); |
| 2849 |
|
| 2850 |
info = internal::estimate_eigenvalues<MatrixType>( |
| 2851 |
data, A, solution_old, temp_vector1, data.n_iterations); |
| 2852 |
|
| 2853 |
const double alpha = |
| 2854 |
(data.smoothing_range > 1. ? |
| 2855 |
info.max_eigenvalue_estimate / data.smoothing_range : |
| 2856 |
std::min(0.9 * info.max_eigenvalue_estimate, |
| 2857 |
info.min_eigenvalue_estimate)); |
| 2858 |
|
| 2859 |
const_cast<PreconditionRelaxation<MatrixType, PreconditionerType> *>(this) |
| 2860 |
->data.relaxation = 2.0 / (alpha + info.max_eigenvalue_estimate); |
| 2861 |
} |
| 2862 |
|
| 2863 |
const_cast<PreconditionRelaxation<MatrixType, PreconditionerType> *>(this) |
| 2864 |
->eigenvalues_are_initialized = true; |
| 2865 |
|
| 2866 |
return info; |
| 2867 |
} |
| 2868 |
|
| 2869 |
template <typename MatrixType, typename PreconditionerType> |
| 2870 |
double |
| 2871 |
PreconditionRelaxation<MatrixType, PreconditionerType>::get_relaxation() const |
| 2872 |
{ |
| 2873 |
return data.relaxation; |
| 2874 |
} |
| 2875 |
|
| 2876 |
|
| 2877 |
//--------------------------------------------------------------------------- |
| 2878 |
|
| 2879 |
template <typename MatrixType> |
| 2880 |
inline void |
| 2881 |
PreconditionJacobi<MatrixType>::initialize(const MatrixType &A, |
| 2882 |
const AdditionalData ¶meters_in) |
| 2883 |
{ |
| 2884 |
Assert(parameters_in.preconditioner == nullptr, ExcInternalError()); |
| 2885 |
Assert( |
| 2886 |
parameters_in.relaxation != 0.0, |
| 2887 |
ExcMessage( |
| 2888 |
"Relaxation cannot automatically be determined by PreconditionJacobi.")); |
| 2889 |
|
| 2890 |
AdditionalData parameters; |
| 2891 |
parameters.relaxation = 1.0; |
| 2892 |
parameters.n_iterations = parameters_in.n_iterations; |
| 2893 |
parameters.preconditioner = |
| 2894 |
std::make_shared<PreconditionerType>(A, parameters_in.relaxation); |
| 2895 |
|
| 2896 |
this->BaseClass::initialize(A, parameters); |
| 2897 |
} |
| 2898 |
|
| 2899 |
//--------------------------------------------------------------------------- |
| 2900 |
|
| 2901 |
template <typename MatrixType> |
| 2902 |
inline void |
| 2903 |
PreconditionSOR<MatrixType>::initialize(const MatrixType &A, |
| 2904 |
const AdditionalData ¶meters_in) |
| 2905 |
{ |
| 2906 |
Assert(parameters_in.preconditioner == nullptr, ExcInternalError()); |
| 2907 |
Assert( |
| 2908 |
parameters_in.relaxation != 0.0, |
| 2909 |
ExcMessage( |
| 2910 |
"Relaxation cannot automatically be determined by PreconditionSOR.")); |
| 2911 |
|
| 2912 |
AdditionalData parameters; |
| 2913 |
parameters.relaxation = 1.0; |
| 2914 |
parameters.n_iterations = parameters_in.n_iterations; |
| 2915 |
parameters.preconditioner = |
| 2916 |
std::make_shared<PreconditionerType>(A, parameters_in.relaxation); |
| 2917 |
|
| 2918 |
this->BaseClass::initialize(A, parameters); |
| 2919 |
} |
| 2920 |
|
| 2921 |
//--------------------------------------------------------------------------- |
| 2922 |
|
| 2923 |
template <typename MatrixType> |
| 2924 |
inline void |
| 2925 |
PreconditionSSOR<MatrixType>::initialize(const MatrixType &A, |
| 2926 |
const AdditionalData ¶meters_in) |
| 2927 |
{ |
| 2928 |
Assert(parameters_in.preconditioner == nullptr, ExcInternalError()); |
| 2929 |
Assert( |
| 2930 |
parameters_in.relaxation != 0.0, |
| 2931 |
ExcMessage( |
| 2932 |
"Relaxation cannot automatically be determined by PreconditionSSOR.")); |
| 2933 |
|
| 2934 |
AdditionalData parameters; |
| 2935 |
parameters.relaxation = 1.0; |
| 2936 |
parameters.n_iterations = parameters_in.n_iterations; |
| 2937 |
parameters.preconditioner = |
| 2938 |
std::make_shared<PreconditionerType>(A, parameters_in.relaxation); |
| 2939 |
|
| 2940 |
this->BaseClass::initialize(A, parameters); |
| 2941 |
} |
| 2942 |
|
| 2943 |
|
| 2944 |
|
| 2945 |
//--------------------------------------------------------------------------- |
| 2946 |
|
| 2947 |
template <typename MatrixType> |
| 2948 |
inline void |
| 2949 |
PreconditionPSOR<MatrixType>::initialize( |
| 2950 |
const MatrixType &A, |
| 2951 |
const std::vector<size_type> &p, |
| 2952 |
const std::vector<size_type> &ip, |
| 2953 |
const typename BaseClass::AdditionalData ¶meters_in) |
| 2954 |
{ |
| 2955 |
Assert(parameters_in.preconditioner == nullptr, ExcInternalError()); |
| 2956 |
Assert( |
| 2957 |
parameters_in.relaxation != 0.0, |
| 2958 |
ExcMessage( |
| 2959 |
"Relaxation cannot automatically be determined by PreconditionPSOR.")); |
| 2960 |
|
| 2961 |
typename BaseClass::AdditionalData parameters; |
| 2962 |
parameters.relaxation = 1.0; |
| 2963 |
parameters.n_iterations = parameters_in.n_iterations; |
| 2964 |
parameters.preconditioner = |
| 2965 |
std::make_shared<PreconditionerType>(A, parameters_in.relaxation, p, ip); |
| 2966 |
|
| 2967 |
this->BaseClass::initialize(A, parameters); |
| 2968 |
} |
| 2969 |
|
| 2970 |
|
| 2971 |
template <typename MatrixType> |
| 2972 |
inline void |
| 2973 |
PreconditionPSOR<MatrixType>::initialize(const MatrixType &A, |
| 2974 |
const AdditionalData &additional_data) |
| 2975 |
{ |
| 2976 |
initialize(A, |
| 2977 |
additional_data.permutation, |
| 2978 |
additional_data.inverse_permutation, |
| 2979 |
additional_data.parameters); |
| 2980 |
} |
| 2981 |
|
| 2982 |
template <typename MatrixType> |
| 2983 |
PreconditionPSOR<MatrixType>::AdditionalData::AdditionalData( |
| 2984 |
const std::vector<size_type> &permutation, |
| 2985 |
const std::vector<size_type> &inverse_permutation, |
| 2986 |
const typename PreconditionPSOR<MatrixType>::BaseClass::AdditionalData |
| 2987 |
¶meters) |
| 2988 |
: permutation(permutation) |
| 2989 |
, inverse_permutation(inverse_permutation) |
| 2990 |
, parameters(parameters) |
| 2991 |
{} |
| 2992 |
|
| 2993 |
|
| 2994 |
//--------------------------------------------------------------------------- |
| 2995 |
|
| 2996 |
|
| 2997 |
template <typename MatrixType, typename VectorType> |
| 2998 |
PreconditionUseMatrix<MatrixType, VectorType>::PreconditionUseMatrix( |
| 2999 |
const MatrixType &M, |
| 3000 |
const function_ptr method) |
| 3001 |
: matrix(M) |
| 3002 |
, precondition(method) |
| 3003 |
{} |
| 3004 |
|
| 3005 |
|
| 3006 |
|
| 3007 |
template <typename MatrixType, typename VectorType> |
| 3008 |
void |
| 3009 |
PreconditionUseMatrix<MatrixType, VectorType>::vmult( |
| 3010 |
VectorType &dst, |
| 3011 |
const VectorType &src) const |
| 3012 |
{ |
| 3013 |
(matrix.*precondition)(dst, src); |
| 3014 |
} |
| 3015 |
|
| 3016 |
//--------------------------------------------------------------------------- |
| 3017 |
|
| 3018 |
namespace internal |
| 3019 |
{ |
| 3020 |
|
| 3021 |
template <typename PreconditionerType> |
| 3022 |
inline EigenvalueAlgorithmAdditionalData<PreconditionerType>:: |
| 3023 |
EigenvalueAlgorithmAdditionalData( |
| 3024 |
const double smoothing_range, |
| 3025 |
const unsigned int eig_cg_n_iterations, |
| 3026 |
const double eig_cg_residual, |
| 3027 |
const double max_eigenvalue, |
| 3028 |
const EigenvalueAlgorithm eigenvalue_algorithm) |
| 3029 |
: smoothing_range(smoothing_range) |
| 3030 |
, eig_cg_n_iterations(eig_cg_n_iterations) |
| 3031 |
, eig_cg_residual(eig_cg_residual) |
| 3032 |
, max_eigenvalue(max_eigenvalue) |
| 3033 |
, eigenvalue_algorithm(eigenvalue_algorithm) |
| 3034 |
{} |
| 3035 |
|
| 3036 |
|
| 3037 |
|
| 3038 |
template <typename PreconditionerType> |
| 3039 |
inline EigenvalueAlgorithmAdditionalData<PreconditionerType> & |
| 3040 |
EigenvalueAlgorithmAdditionalData<PreconditionerType>::operator=( |
| 3041 |
const EigenvalueAlgorithmAdditionalData &other_data) |
| 3042 |
{ |
| 3043 |
smoothing_range = other_data.smoothing_range; |
| 3044 |
eig_cg_n_iterations = other_data.eig_cg_n_iterations; |
| 3045 |
eig_cg_residual = other_data.eig_cg_residual; |
| 3046 |
max_eigenvalue = other_data.max_eigenvalue; |
| 3047 |
preconditioner = other_data.preconditioner; |
| 3048 |
eigenvalue_algorithm = other_data.eigenvalue_algorithm; |
| 3049 |
constraints.copy_from(other_data.constraints); |
| 3050 |
|
| 3051 |
return *this; |
| 3052 |
} |
| 3053 |
} // namespace internal |
| 3054 |
|
| 3055 |
template <typename MatrixType, typename PreconditionerType> |
| 3056 |
inline PreconditionRelaxation<MatrixType, PreconditionerType>::AdditionalData:: |
| 3057 |
AdditionalData(const double relaxation, |
| 3058 |
const unsigned int n_iterations, |
| 3059 |
const double smoothing_range, |
| 3060 |
const unsigned int eig_cg_n_iterations, |
| 3061 |
const double eig_cg_residual, |
| 3062 |
const double max_eigenvalue, |
| 3063 |
const EigenvalueAlgorithm eigenvalue_algorithm) |
| 3064 |
: internal::EigenvalueAlgorithmAdditionalData<PreconditionerType>( |
| 3065 |
smoothing_range, |
| 3066 |
eig_cg_n_iterations, |
| 3067 |
eig_cg_residual, |
| 3068 |
max_eigenvalue, |
| 3069 |
eigenvalue_algorithm) |
| 3070 |
, relaxation(relaxation) |
| 3071 |
, n_iterations(n_iterations) |
| 3072 |
{} |
| 3073 |
|
| 3074 |
|
| 3075 |
|
| 3076 |
//--------------------------------------------------------------------------- |
| 3077 |
|
| 3078 |
namespace internal |
| 3079 |
{ |
| 3080 |
namespace PreconditionChebyshevImplementation |
| 3081 |
{ |
| 3082 |
// for deal.II vectors, perform updates for Chebyshev preconditioner all |
| 3083 |
// at once to reduce memory transfer. Here, we select between general |
| 3084 |
// vectors and deal.II vectors where we expand the loop over the (local) |
| 3085 |
// size of the vector |
| 3086 |
|
| 3087 |
// generic part for non-deal.II vectors |
| 3088 |
template <typename VectorType, typename PreconditionerType> |
| 3089 |
inline void |
| 3090 |
vector_updates(const VectorType &rhs, |
| 3091 |
const PreconditionerType &preconditioner, |
| 3092 |
const unsigned int iteration_index, |
| 3093 |
const double factor1, |
| 3094 |
const double factor2, |
| 3095 |
VectorType &solution_old, |
| 3096 |
VectorType &temp_vector1, |
| 3097 |
VectorType &temp_vector2, |
| 3098 |
VectorType &solution) |
| 3099 |
{ |
| 3100 |
if (iteration_index == 0) |
| 3101 |
{ |
| 3102 |
solution.equ(factor2, rhs); |
| 3103 |
preconditioner.vmult(solution_old, solution); |
| 3104 |
} |
| 3105 |
else if (iteration_index == 1) |
| 3106 |
{ |
| 3107 |
// compute t = P^{-1} * (b-A*x^{n}) |
| 3108 |
temp_vector1.sadd(-1.0, 1.0, rhs); |
| 3109 |
preconditioner.vmult(solution_old, temp_vector1); |
| 3110 |
|
| 3111 |
// compute x^{n+1} = x^{n} + f_1 * x^{n} + f_2 * t |
| 3112 |
solution_old.sadd(factor2, 1 + factor1, solution); |
| 3113 |
} |
| 3114 |
else |
| 3115 |
{ |
| 3116 |
// compute t = P^{-1} * (b-A*x^{n}) |
| 3117 |
temp_vector1.sadd(-1.0, 1.0, rhs); |
| 3118 |
preconditioner.vmult(temp_vector2, temp_vector1); |
| 3119 |
|
| 3120 |
// compute x^{n+1} = x^{n} + f_1 * (x^{n}-x^{n-1}) + f_2 * t |
| 3121 |
solution_old.sadd(-factor1, factor2, temp_vector2); |
| 3122 |
solution_old.add(1 + factor1, solution); |
| 3123 |
} |
| 3124 |
|
| 3125 |
solution.swap(solution_old); |
| 3126 |
} |
| 3127 |
|
| 3128 |
// generic part for deal.II vectors |
| 3129 |
template < |
| 3130 |
typename Number, |
| 3131 |
typename PreconditionerType, |
| 3132 |
std::enable_if_t< |
| 3133 |
!has_vmult_with_std_functions_for_precondition< |
| 3134 |
PreconditionerType, |
| 3135 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host>>, |
| 3136 |
int> * = nullptr> |
| 3137 |
inline void |
| 3138 |
vector_updates( |
| 3139 |
const LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &rhs, |
| 3140 |
const PreconditionerType &preconditioner, |
| 3141 |
const unsigned int iteration_index, |
| 3142 |
const double factor1_, |
| 3143 |
const double factor2_, |
| 3144 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3145 |
&solution_old, |
| 3146 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3147 |
&temp_vector1, |
| 3148 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3149 |
&temp_vector2, |
| 3150 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &solution) |
| 3151 |
{ |
| 3152 |
const Number factor1 = factor1_; |
| 3153 |
const Number factor1_plus_1 = 1. + factor1_; |
| 3154 |
const Number factor2 = factor2_; |
| 3155 |
|
| 3156 |
if (iteration_index == 0) |
| 3157 |
{ |
| 3158 |
// compute t = P^{-1} * (b) |
| 3159 |
preconditioner.vmult(solution_old, rhs); |
| 3160 |
|
| 3161 |
// compute x^{n+1} = f_2 * t |
| 3162 |
const auto solution_old_ptr = solution_old.begin(); |
| 3163 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3164 |
for (unsigned int i = 0; i < solution_old.locally_owned_size(); ++i) |
| 3165 |
solution_old_ptr[i] = solution_old_ptr[i] * factor2; |
| 3166 |
} |
| 3167 |
else if (iteration_index == 1) |
| 3168 |
{ |
| 3169 |
// compute t = P^{-1} * (b-A*x^{n}) |
| 3170 |
temp_vector1.sadd(-1.0, 1.0, rhs); |
| 3171 |
|
| 3172 |
preconditioner.vmult(solution_old, temp_vector1); |
| 3173 |
|
| 3174 |
// compute x^{n+1} = x^{n} + f_1 * x^{n} + f_2 * t |
| 3175 |
const auto solution_ptr = solution.begin(); |
| 3176 |
const auto solution_old_ptr = solution_old.begin(); |
| 3177 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3178 |
for (unsigned int i = 0; i < solution_old.locally_owned_size(); ++i) |
| 3179 |
solution_old_ptr[i] = |
| 3180 |
factor1_plus_1 * solution_ptr[i] + solution_old_ptr[i] * factor2; |
| 3181 |
} |
| 3182 |
else |
| 3183 |
{ |
| 3184 |
// compute t = P^{-1} * (b-A*x^{n}) |
| 3185 |
temp_vector1.sadd(-1.0, 1.0, rhs); |
| 3186 |
|
| 3187 |
preconditioner.vmult(temp_vector2, temp_vector1); |
| 3188 |
|
| 3189 |
// compute x^{n+1} = x^{n} + f_1 * (x^{n}-x^{n-1}) + f_2 * t |
| 3190 |
const auto solution_ptr = solution.begin(); |
| 3191 |
const auto solution_old_ptr = solution_old.begin(); |
| 3192 |
const auto temp_vector2_ptr = temp_vector2.begin(); |
| 3193 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3194 |
for (unsigned int i = 0; i < solution_old.locally_owned_size(); ++i) |
| 3195 |
solution_old_ptr[i] = factor1_plus_1 * solution_ptr[i] - |
| 3196 |
factor1 * solution_old_ptr[i] + |
| 3197 |
temp_vector2_ptr[i] * factor2; |
| 3198 |
} |
| 3199 |
|
| 3200 |
solution.swap(solution_old); |
| 3201 |
} |
| 3202 |
|
| 3203 |
template < |
| 3204 |
typename Number, |
| 3205 |
typename PreconditionerType, |
| 3206 |
std::enable_if_t< |
| 3207 |
has_vmult_with_std_functions_for_precondition< |
| 3208 |
PreconditionerType, |
| 3209 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host>>, |
| 3210 |
int> * = nullptr> |
| 3211 |
inline void |
| 3212 |
vector_updates( |
| 3213 |
const LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &rhs, |
| 3214 |
const PreconditionerType &preconditioner, |
| 3215 |
const unsigned int iteration_index, |
| 3216 |
const double factor1_, |
| 3217 |
const double factor2_, |
| 3218 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3219 |
&solution_old, |
| 3220 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3221 |
&temp_vector1, |
| 3222 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3223 |
&temp_vector2, |
| 3224 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &solution) |
| 3225 |
{ |
| 3226 |
const Number factor1 = factor1_; |
| 3227 |
const Number factor1_plus_1 = 1. + factor1_; |
| 3228 |
const Number factor2 = factor2_; |
| 3229 |
|
| 3230 |
const auto rhs_ptr = rhs.begin(); |
| 3231 |
const auto temp_vector1_ptr = temp_vector1.begin(); |
| 3232 |
const auto temp_vector2_ptr = temp_vector2.begin(); |
| 3233 |
const auto solution_ptr = solution.begin(); |
| 3234 |
const auto solution_old_ptr = solution_old.begin(); |
| 3235 |
|
| 3236 |
if (iteration_index == 0) |
| 3237 |
{ |
| 3238 |
preconditioner.vmult( |
| 3239 |
solution, |
| 3240 |
rhs, |
| 3241 |
[&](const auto start_range, const auto end_range) { |
| 3242 |
if (end_range > start_range) |
| 3243 |
std::memset(solution_ptr + start_range, |
| 3244 |
0, |
| 3245 |
sizeof(Number) * (end_range - start_range)); |
| 3246 |
}, |
| 3247 |
[&](const auto begin, const auto end) { |
| 3248 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3249 |
for (std::size_t i = begin; i < end; ++i) |
| 3250 |
solution_ptr[i] *= factor2; |
| 3251 |
}); |
| 3252 |
} |
| 3253 |
else |
| 3254 |
{ |
| 3255 |
preconditioner.vmult( |
| 3256 |
temp_vector2, |
| 3257 |
temp_vector1, |
| 3258 |
[&](const auto begin, const auto end) { |
| 3259 |
if (end > begin) |
| 3260 |
std::memset(temp_vector2_ptr + begin, |
| 3261 |
0, |
| 3262 |
sizeof(Number) * (end - begin)); |
| 3263 |
|
| 3264 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3265 |
for (std::size_t i = begin; i < end; ++i) |
| 3266 |
temp_vector1_ptr[i] = rhs_ptr[i] - temp_vector1_ptr[i]; |
| 3267 |
}, |
| 3268 |
[&](const auto begin, const auto end) { |
| 3269 |
if (iteration_index == 1) |
| 3270 |
{ |
| 3271 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3272 |
for (std::size_t i = begin; i < end; ++i) |
| 3273 |
temp_vector2_ptr[i] = factor1_plus_1 * solution_ptr[i] + |
| 3274 |
factor2 * temp_vector2_ptr[i]; |
| 3275 |
} |
| 3276 |
else |
| 3277 |
{ |
| 3278 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3279 |
for (std::size_t i = begin; i < end; ++i) |
| 3280 |
temp_vector2_ptr[i] = factor1_plus_1 * solution_ptr[i] - |
| 3281 |
factor1 * solution_old_ptr[i] + |
| 3282 |
factor2 * temp_vector2_ptr[i]; |
| 3283 |
} |
| 3284 |
}); |
| 3285 |
} |
| 3286 |
|
| 3287 |
if (iteration_index > 0) |
| 3288 |
{ |
| 3289 |
solution_old.swap(temp_vector2); |
| 3290 |
solution_old.swap(solution); |
| 3291 |
} |
| 3292 |
} |
| 3293 |
|
| 3294 |
// worker routine for deal.II vectors. Because of vectorization, we need |
| 3295 |
// to put the loop into an extra structure because the virtual function of |
| 3296 |
// VectorUpdatesRange prevents the compiler from applying vectorization. |
| 3297 |
template <typename Number> |
| 3298 |
struct VectorUpdater |
| 3299 |
{ |
| 3300 |
VectorUpdater(const Number *rhs, |
| 3301 |
const Number *matrix_diagonal_inverse, |
| 3302 |
const unsigned int iteration_index, |
| 3303 |
const Number factor1, |
| 3304 |
const Number factor2, |
| 3305 |
Number *solution_old, |
| 3306 |
Number *tmp_vector, |
| 3307 |
Number *solution) |
| 3308 |
: rhs(rhs) |
| 3309 |
, matrix_diagonal_inverse(matrix_diagonal_inverse) |
| 3310 |
, iteration_index(iteration_index) |
| 3311 |
, factor1(factor1) |
| 3312 |
, factor2(factor2) |
| 3313 |
, solution_old(solution_old) |
| 3314 |
, tmp_vector(tmp_vector) |
| 3315 |
, solution(solution) |
| 3316 |
{} |
| 3317 |
|
| 3318 |
void |
| 3319 |
apply_to_subrange(const std::size_t begin, const std::size_t end) const |
| 3320 |
{ |
| 3321 |
// To circumvent a bug in gcc |
| 3322 |
// (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63945), we create |
| 3323 |
// copies of the variables factor1 and factor2 and do not check based on |
| 3324 |
// factor1. |
| 3325 |
const Number factor1 = this->factor1; |
| 3326 |
const Number factor1_plus_1 = 1. + this->factor1; |
| 3327 |
const Number factor2 = this->factor2; |
| 3328 |
if (iteration_index == 0) |
| 3329 |
{ |
| 3330 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3331 |
for (std::size_t i = begin; i < end; ++i) |
| 3332 |
solution[i] = factor2 * matrix_diagonal_inverse[i] * rhs[i]; |
| 3333 |
} |
| 3334 |
else if (iteration_index == 1) |
| 3335 |
{ |
| 3336 |
// x^{n+1} = x^{n} + f_1 * x^{n} + f_2 * P^{-1} * (b-A*x^{n}) |
| 3337 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3338 |
for (std::size_t i = begin; i < end; ++i) |
| 3339 |
// for efficiency reason, write back to temp_vector that is |
| 3340 |
// already read (avoid read-for-ownership) |
| 3341 |
tmp_vector[i] = |
| 3342 |
factor1_plus_1 * solution[i] + |
| 3343 |
factor2 * matrix_diagonal_inverse[i] * (rhs[i] - tmp_vector[i]); |
| 3344 |
} |
| 3345 |
else |
| 3346 |
{ |
| 3347 |
// x^{n+1} = x^{n} + f_1 * (x^{n}-x^{n-1}) |
| 3348 |
// + f_2 * P^{-1} * (b-A*x^{n}) |
| 3349 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3350 |
for (std::size_t i = begin; i < end; ++i) |
| 3351 |
// for efficiency reason, write back to temp_vector, which is |
| 3352 |
// already modified during vmult (in best case, the modified |
| 3353 |
// values are not written back to main memory yet so that |
| 3354 |
// we do not have to pay additional costs for writing and |
| 3355 |
// read-for-ownership) |
| 3356 |
tmp_vector[i] = |
| 3357 |
factor1_plus_1 * solution[i] - factor1 * solution_old[i] + |
| 3358 |
factor2 * matrix_diagonal_inverse[i] * (rhs[i] - tmp_vector[i]); |
| 3359 |
} |
| 3360 |
} |
| 3361 |
|
| 3362 |
const Number *rhs; |
| 3363 |
const Number *matrix_diagonal_inverse; |
| 3364 |
const unsigned int iteration_index; |
| 3365 |
const Number factor1; |
| 3366 |
const Number factor2; |
| 3367 |
mutable Number *solution_old; |
| 3368 |
mutable Number *tmp_vector; |
| 3369 |
mutable Number *solution; |
| 3370 |
}; |
| 3371 |
|
| 3372 |
template <typename Number> |
| 3373 |
struct VectorUpdatesRange : public ::dealii::parallel::ParallelForInteger |
| 3374 |
{ |
| 3375 |
VectorUpdatesRange(const VectorUpdater<Number> &updater, |
| 3376 |
const std::size_t size) |
| 3377 |
: updater(updater) |
| 3378 |
{ |
| 3379 |
if (size < internal::VectorImplementation::minimum_parallel_grain_size) |
| 3380 |
VectorUpdatesRange::apply_to_subrange(0, size); |
| 3381 |
else |
| 3382 |
apply_parallel( |
| 3383 |
0, |
| 3384 |
size, |
| 3385 |
internal::VectorImplementation::minimum_parallel_grain_size); |
| 3386 |
} |
| 3387 |
|
| 3388 |
~VectorUpdatesRange() override = default; |
| 3389 |
|
| 3390 |
virtual void |
| 3391 |
apply_to_subrange(const std::size_t begin, |
| 3392 |
const std::size_t end) const override |
| 3393 |
{ |
| 3394 |
updater.apply_to_subrange(begin, end); |
| 3395 |
} |
| 3396 |
|
| 3397 |
const VectorUpdater<Number> &updater; |
| 3398 |
}; |
| 3399 |
|
| 3400 |
// selection for diagonal matrix around deal.II vector |
| 3401 |
template <typename Number> |
| 3402 |
inline void |
| 3403 |
vector_updates( |
| 3404 |
const ::dealii::Vector<Number> &rhs, |
| 3405 |
const dealii::DiagonalMatrix<::dealii::Vector<Number>> &jacobi, |
| 3406 |
const unsigned int iteration_index, |
| 3407 |
const double factor1, |
| 3408 |
const double factor2, |
| 3409 |
::dealii::Vector<Number> &solution_old, |
| 3410 |
::dealii::Vector<Number> &temp_vector1, |
| 3411 |
::dealii::Vector<Number> &, |
| 3412 |
::dealii::Vector<Number> &solution) |
| 3413 |
{ |
| 3414 |
VectorUpdater<Number> upd(rhs.begin(), |
| 3415 |
jacobi.get_vector().begin(), |
| 3416 |
iteration_index, |
| 3417 |
factor1, |
| 3418 |
factor2, |
| 3419 |
solution_old.begin(), |
| 3420 |
temp_vector1.begin(), |
| 3421 |
solution.begin()); |
| 3422 |
VectorUpdatesRange<Number>(upd, rhs.size()); |
| 3423 |
|
| 3424 |
// swap vectors x^{n+1}->x^{n}, given the updates in the function above |
| 3425 |
if (iteration_index == 0) |
| 3426 |
{ |
| 3427 |
// nothing to do here because we can immediately write into the |
| 3428 |
// solution vector without remembering any of the other vectors |
| 3429 |
} |
| 3430 |
else |
| 3431 |
{ |
| 3432 |
solution.swap(temp_vector1); |
| 3433 |
solution_old.swap(temp_vector1); |
| 3434 |
} |
| 3435 |
} |
| 3436 |
|
| 3437 |
// selection for diagonal matrix around parallel deal.II vector |
| 3438 |
template <typename Number> |
| 3439 |
inline void |
| 3440 |
vector_updates( |
| 3441 |
const LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &rhs, |
| 3442 |
const dealii::DiagonalMatrix< |
| 3443 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host>> &jacobi, |
| 3444 |
const unsigned int iteration_index, |
| 3445 |
const double factor1, |
| 3446 |
const double factor2, |
| 3447 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3448 |
&solution_old, |
| 3449 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> |
| 3450 |
&temp_vector1, |
| 3451 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &, |
| 3452 |
LinearAlgebra::distributed::Vector<Number, MemorySpace::Host> &solution) |
| 3453 |
{ |
| 3454 |
VectorUpdater<Number> upd(rhs.begin(), |
| 3455 |
jacobi.get_vector().begin(), |
| 3456 |
iteration_index, |
| 3457 |
factor1, |
| 3458 |
factor2, |
| 3459 |
solution_old.begin(), |
| 3460 |
temp_vector1.begin(), |
| 3461 |
solution.begin()); |
| 3462 |
VectorUpdatesRange<Number>(upd, rhs.locally_owned_size()); |
| 3463 |
|
| 3464 |
// swap vectors x^{n+1}->x^{n}, given the updates in the function above |
| 3465 |
if (iteration_index == 0) |
| 3466 |
{ |
| 3467 |
// nothing to do here because we can immediately write into the |
| 3468 |
// solution vector without remembering any of the other vectors |
| 3469 |
} |
| 3470 |
else |
| 3471 |
{ |
| 3472 |
solution.swap(temp_vector1); |
| 3473 |
solution_old.swap(temp_vector1); |
| 3474 |
} |
| 3475 |
} |
| 3476 |
|
| 3477 |
// We need to have a separate declaration for static const members |
| 3478 |
|
| 3479 |
// general case and the case that the preconditioner can work on |
| 3480 |
// ranges (covered by vector_updates()) |
| 3481 |
template < |
| 3482 |
typename MatrixType, |
| 3483 |
typename VectorType, |
| 3484 |
typename PreconditionerType, |
| 3485 |
std::enable_if_t< |
| 3486 |
!has_vmult_with_std_functions<MatrixType, |
| 3487 |
VectorType, |
| 3488 |
PreconditionerType> && |
| 3489 |
!(has_vmult_with_std_functions_for_precondition<PreconditionerType, |
| 3490 |
VectorType> && |
| 3491 |
has_vmult_with_std_functions_for_precondition<MatrixType, |
| 3492 |
VectorType>), |
| 3493 |
int> * = nullptr> |
| 3494 |
inline void |
| 3495 |
vmult_and_update(const MatrixType &matrix, |
| 3496 |
const PreconditionerType &preconditioner, |
| 3497 |
const VectorType &rhs, |
| 3498 |
const unsigned int iteration_index, |
| 3499 |
const double factor1, |
| 3500 |
const double factor2, |
| 3501 |
VectorType &solution, |
| 3502 |
VectorType &solution_old, |
| 3503 |
VectorType &temp_vector1, |
| 3504 |
VectorType &temp_vector2) |
| 3505 |
{ |
| 3506 |
if (iteration_index > 0) |
| 3507 |
matrix.vmult(temp_vector1, solution); |
| 3508 |
vector_updates(rhs, |
| 3509 |
preconditioner, |
| 3510 |
iteration_index, |
| 3511 |
factor1, |
| 3512 |
factor2, |
| 3513 |
solution_old, |
| 3514 |
temp_vector1, |
| 3515 |
temp_vector2, |
| 3516 |
solution); |
| 3517 |
} |
| 3518 |
|
| 3519 |
// case that both the operator and the preconditioner can work on |
| 3520 |
// subranges |
| 3521 |
template < |
| 3522 |
typename MatrixType, |
| 3523 |
typename VectorType, |
| 3524 |
typename PreconditionerType, |
| 3525 |
std::enable_if_t< |
| 3526 |
!has_vmult_with_std_functions<MatrixType, |
| 3527 |
VectorType, |
| 3528 |
PreconditionerType> && |
| 3529 |
(has_vmult_with_std_functions_for_precondition<PreconditionerType, |
| 3530 |
VectorType> && |
| 3531 |
has_vmult_with_std_functions_for_precondition<MatrixType, |
| 3532 |
VectorType>), |
| 3533 |
int> * = nullptr> |
| 3534 |
inline void |
| 3535 |
vmult_and_update(const MatrixType &matrix, |
| 3536 |
const PreconditionerType &preconditioner, |
| 3537 |
const VectorType &rhs, |
| 3538 |
const unsigned int iteration_index, |
| 3539 |
const double factor1_, |
| 3540 |
const double factor2_, |
| 3541 |
VectorType &solution, |
| 3542 |
VectorType &solution_old, |
| 3543 |
VectorType &temp_vector1, |
| 3544 |
VectorType &temp_vector2) |
| 3545 |
{ |
| 3546 |
using Number = typename VectorType::value_type; |
| 3547 |
|
| 3548 |
const Number factor1 = factor1_; |
| 3549 |
const Number factor1_plus_1 = 1. + factor1_; |
| 3550 |
const Number factor2 = factor2_; |
| 3551 |
|
| 3552 |
if (iteration_index == 0) |
| 3553 |
{ |
| 3554 |
preconditioner.vmult( |
| 3555 |
solution, |
| 3556 |
rhs, |
| 3557 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3558 |
// zero 'solution' before running the vmult operation |
| 3559 |
if (end_range > start_range) |
| 3560 |
std::memset(solution.begin() + start_range, |
| 3561 |
0, |
| 3562 |
sizeof(Number) * (end_range - start_range)); |
| 3563 |
}, |
| 3564 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3565 |
const auto solution_ptr = solution.begin(); |
| 3566 |
|
| 3567 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3568 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 3569 |
solution_ptr[i] *= factor2; |
| 3570 |
}); |
| 3571 |
} |
| 3572 |
else |
| 3573 |
{ |
| 3574 |
temp_vector1.reinit(rhs, true); |
| 3575 |
temp_vector2.reinit(rhs, true); |
| 3576 |
|
| 3577 |
// 1) compute residual (including operator application) |
| 3578 |
matrix.vmult( |
| 3579 |
temp_vector1, |
| 3580 |
solution, |
| 3581 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3582 |
// zero 'temp_vector1' before running the vmult |
| 3583 |
// operation |
| 3584 |
if (end_range > start_range) |
| 3585 |
std::memset(temp_vector1.begin() + start_range, |
| 3586 |
0, |
| 3587 |
sizeof(Number) * (end_range - start_range)); |
| 3588 |
}, |
| 3589 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3590 |
const auto rhs_ptr = rhs.begin(); |
| 3591 |
const auto tmp_ptr = temp_vector1.begin(); |
| 3592 |
|
| 3593 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3594 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 3595 |
tmp_ptr[i] = rhs_ptr[i] - tmp_ptr[i]; |
| 3596 |
}); |
| 3597 |
|
| 3598 |
// 2) perform vector updates (including preconditioner application) |
| 3599 |
preconditioner.vmult( |
| 3600 |
temp_vector2, |
| 3601 |
temp_vector1, |
| 3602 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3603 |
// zero 'temp_vector2' before running the vmult |
| 3604 |
// operation |
| 3605 |
if (end_range > start_range) |
| 3606 |
std::memset(temp_vector2.begin() + start_range, |
| 3607 |
0, |
| 3608 |
sizeof(Number) * (end_range - start_range)); |
| 3609 |
}, |
| 3610 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3611 |
const auto solution_ptr = solution.begin(); |
| 3612 |
const auto solution_old_ptr = solution_old.begin(); |
| 3613 |
const auto tmp_ptr = temp_vector2.begin(); |
| 3614 |
|
| 3615 |
if (iteration_index == 1) |
| 3616 |
{ |
| 3617 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3618 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 3619 |
tmp_ptr[i] = |
| 3620 |
factor1_plus_1 * solution_ptr[i] + factor2 * tmp_ptr[i]; |
| 3621 |
} |
| 3622 |
else |
| 3623 |
{ |
| 3624 |
DEAL_II_OPENMP_SIMD_PRAGMA |
| 3625 |
for (std::size_t i = start_range; i < end_range; ++i) |
| 3626 |
tmp_ptr[i] = factor1_plus_1 * solution_ptr[i] - |
| 3627 |
factor1 * solution_old_ptr[i] + |
| 3628 |
factor2 * tmp_ptr[i]; |
| 3629 |
} |
| 3630 |
}); |
| 3631 |
|
| 3632 |
solution.swap(temp_vector2); |
| 3633 |
solution_old.swap(temp_vector2); |
| 3634 |
} |
| 3635 |
} |
| 3636 |
|
| 3637 |
// case that the operator can work on subranges and the preconditioner |
| 3638 |
// is a diagonal |
| 3639 |
template <typename MatrixType, |
| 3640 |
typename VectorType, |
| 3641 |
typename PreconditionerType, |
| 3642 |
std::enable_if_t<has_vmult_with_std_functions<MatrixType, |
| 3643 |
VectorType, |
| 3644 |
PreconditionerType>, |
| 3645 |
int> * = nullptr> |
| 3646 |
inline void |
| 3647 |
vmult_and_update(const MatrixType &matrix, |
| 3648 |
const PreconditionerType &preconditioner, |
| 3649 |
const VectorType &rhs, |
| 3650 |
const unsigned int iteration_index, |
| 3651 |
const double factor1, |
| 3652 |
const double factor2, |
| 3653 |
VectorType &solution, |
| 3654 |
VectorType &solution_old, |
| 3655 |
VectorType &temp_vector1, |
| 3656 |
VectorType &) |
| 3657 |
{ |
| 3658 |
using Number = typename VectorType::value_type; |
| 3659 |
VectorUpdater<Number> updater(rhs.begin(), |
| 3660 |
preconditioner.get_vector().begin(), |
| 3661 |
iteration_index, |
| 3662 |
factor1, |
| 3663 |
factor2, |
| 3664 |
solution_old.begin(), |
| 3665 |
temp_vector1.begin(), |
| 3666 |
solution.begin()); |
| 3667 |
if (iteration_index > 0) |
| 3668 |
matrix.vmult( |
| 3669 |
temp_vector1, |
| 3670 |
solution, |
| 3671 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3672 |
// zero 'temp_vector1' before running the vmult |
| 3673 |
// operation |
| 3674 |
if (end_range > start_range) |
| 3675 |
std::memset(temp_vector1.begin() + start_range, |
| 3676 |
0, |
| 3677 |
sizeof(Number) * (end_range - start_range)); |
| 3678 |
}, |
| 3679 |
[&](const unsigned int start_range, const unsigned int end_range) { |
| 3680 |
if (end_range > start_range) |
| 3681 |
updater.apply_to_subrange(start_range, end_range); |
| 3682 |
}); |
| 3683 |
else |
| 3684 |
updater.apply_to_subrange(0U, solution.locally_owned_size()); |
| 3685 |
|
| 3686 |
// swap vectors x^{n+1}->x^{n}, given the updates in the function above |
| 3687 |
if (iteration_index == 0) |
| 3688 |
{ |
| 3689 |
// nothing to do here because we can immediately write into the |
| 3690 |
// solution vector without remembering any of the other vectors |
| 3691 |
} |
| 3692 |
else |
| 3693 |
{ |
| 3694 |
solution.swap(temp_vector1); |
| 3695 |
solution_old.swap(temp_vector1); |
| 3696 |
} |
| 3697 |
} |
| 3698 |
|
| 3699 |
template <typename MatrixType, typename PreconditionerType> |
| 3700 |
inline void |
| 3701 |
initialize_preconditioner( |
| 3702 |
const MatrixType & /*matrix*/, |
| 3703 |
std::shared_ptr<PreconditionerType> &preconditioner) |
| 3704 |
{ |
| 3705 |
(void)preconditioner; |
| 3706 |
AssertThrow(preconditioner.get() != nullptr, ExcNotInitialized()); |
| 3707 |
} |
| 3708 |
|
| 3709 |
template <typename MatrixType, typename VectorType> |
| 3710 |
inline void |
| 3711 |
initialize_preconditioner( |
| 3712 |
const MatrixType &matrix, |
| 3713 |
std::shared_ptr<dealii::DiagonalMatrix<VectorType>> &preconditioner) |
| 3714 |
{ |
| 3715 |
if (preconditioner.get() == nullptr || preconditioner->m() != matrix.m()) |
| 3716 |
{ |
| 3717 |
if (preconditioner.get() == nullptr) |
| 3718 |
preconditioner = |
| 3719 |
std::make_shared<dealii::DiagonalMatrix<VectorType>>(); |
| 3720 |
|
| 3721 |
Assert( |
| 3722 |
preconditioner->m() == 0, |
| 3723 |
ExcMessage( |
| 3724 |
"Preconditioner appears to be initialized but not sized correctly")); |
| 3725 |
|
| 3726 |
// This part only works in serial |
| 3727 |
if (preconditioner->m() != matrix.m()) |
| 3728 |
{ |
| 3729 |
preconditioner->get_vector().reinit(matrix.m()); |
| 3730 |
for (typename VectorType::size_type i = 0; i < matrix.m(); ++i) |
| 3731 |
preconditioner->get_vector()(i) = 1. / matrix.el(i, i); |
| 3732 |
} |
| 3733 |
} |
| 3734 |
} |
| 3735 |
} // namespace PreconditionChebyshevImplementation |
| 3736 |
} // namespace internal |
| 3737 |
|
| 3738 |
|
| 3739 |
|
| 3740 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3741 |
inline PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>:: |
| 3742 |
AdditionalData::AdditionalData(const unsigned int degree, |
| 3743 |
const double smoothing_range, |
| 3744 |
const unsigned int eig_cg_n_iterations, |
| 3745 |
const double eig_cg_residual, |
| 3746 |
const double max_eigenvalue, |
| 3747 |
const EigenvalueAlgorithm eigenvalue_algorithm, |
| 3748 |
const PolynomialType polynomial_type) |
| 3749 |
: internal::EigenvalueAlgorithmAdditionalData<PreconditionerType>( |
| 3750 |
smoothing_range, |
| 3751 |
eig_cg_n_iterations, |
| 3752 |
eig_cg_residual, |
| 3753 |
max_eigenvalue, |
| 3754 |
eigenvalue_algorithm) |
| 3755 |
, degree(degree) |
| 3756 |
, polynomial_type(polynomial_type) |
| 3757 |
{} |
| 3758 |
|
| 3759 |
|
| 3760 |
|
| 3761 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3762 |
inline PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>:: |
| 3763 |
PreconditionChebyshev() |
| 3764 |
: theta(1.) |
| 3765 |
, delta(1.) |
| 3766 |
, eigenvalues_are_initialized(false) |
| 3767 |
{ |
| 3768 |
static_assert( |
| 3769 |
std::is_same_v<size_type, typename VectorType::size_type>, |
| 3770 |
"PreconditionChebyshev and VectorType must have the same size_type."); |
| 3771 |
} |
| 3772 |
|
| 3773 |
|
| 3774 |
|
| 3775 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3776 |
inline void |
| 3777 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::initialize( |
| 3778 |
const MatrixType &matrix, |
| 3779 |
const AdditionalData &additional_data) |
| 3780 |
{ |
| 3781 |
matrix_ptr = &matrix; |
| 3782 |
data = additional_data; |
| 3783 |
Assert(data.degree > 0, |
| 3784 |
ExcMessage("The degree of the Chebyshev method must be positive.")); |
| 3785 |
internal::PreconditionChebyshevImplementation::initialize_preconditioner( |
| 3786 |
matrix, data.preconditioner); |
| 3787 |
eigenvalues_are_initialized = false; |
| 3788 |
} |
| 3789 |
|
| 3790 |
|
| 3791 |
|
| 3792 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3793 |
inline void |
| 3794 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::clear() |
| 3795 |
{ |
| 3796 |
eigenvalues_are_initialized = false; |
| 3797 |
theta = delta = 1.0; |
| 3798 |
matrix_ptr = nullptr; |
| 3799 |
{ |
| 3800 |
VectorType empty_vector; |
| 3801 |
solution_old.reinit(empty_vector); |
| 3802 |
temp_vector1.reinit(empty_vector); |
| 3803 |
temp_vector2.reinit(empty_vector); |
| 3804 |
} |
| 3805 |
data.preconditioner.reset(); |
| 3806 |
} |
| 3807 |
|
| 3808 |
|
| 3809 |
|
| 3810 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3811 |
inline typename internal::EigenvalueInformation |
| 3812 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>:: |
| 3813 |
estimate_eigenvalues(const VectorType &src) const |
| 3814 |
{ |
| 3815 |
Assert(eigenvalues_are_initialized == false, ExcInternalError()); |
| 3816 |
|
| 3817 |
solution_old.reinit(src); |
| 3818 |
temp_vector1.reinit(src, true); |
| 3819 |
|
| 3820 |
auto info = internal::estimate_eigenvalues<MatrixType>( |
| 3821 |
data, matrix_ptr, solution_old, temp_vector1, data.degree); |
| 3822 |
|
| 3823 |
const double alpha = (data.smoothing_range > 1. ? |
| 3824 |
info.max_eigenvalue_estimate / data.smoothing_range : |
| 3825 |
std::min(0.9 * info.max_eigenvalue_estimate, |
| 3826 |
info.min_eigenvalue_estimate)); |
| 3827 |
|
| 3828 |
// in case the user set the degree to invalid unsigned int, we have to |
| 3829 |
// determine the number of necessary iterations from the Chebyshev error |
| 3830 |
// estimate, given the target tolerance specified by smoothing_range. This |
| 3831 |
// estimate is based on the error formula given in section 5.1 of |
| 3832 |
// R. S. Varga, Matrix iterative analysis, 2nd ed., Springer, 2009 |
| 3833 |
if (data.degree == numbers::invalid_unsigned_int) |
| 3834 |
{ |
| 3835 |
const double actual_range = info.max_eigenvalue_estimate / alpha; |
| 3836 |
const double sigma = (1. - std::sqrt(1. / actual_range)) / |
| 3837 |
(1. + std::sqrt(1. / actual_range)); |
| 3838 |
const double eps = data.smoothing_range; |
| 3839 |
const_cast< |
| 3840 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType> *>( |
| 3841 |
this) |
| 3842 |
->data.degree = |
| 3843 |
1 + static_cast<unsigned int>( |
| 3844 |
std::log(1. / eps + std::sqrt(1. / eps / eps - 1.)) / |
| 3845 |
std::log(1. / sigma)); |
| 3846 |
} |
| 3847 |
|
| 3848 |
info.degree = data.degree; |
| 3849 |
|
| 3850 |
const_cast< |
| 3851 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType> *>(this) |
| 3852 |
->delta = |
| 3853 |
(data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) ? |
| 3854 |
(info.max_eigenvalue_estimate) : |
| 3855 |
((info.max_eigenvalue_estimate - alpha) * 0.5); |
| 3856 |
const_cast< |
| 3857 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType> *>(this) |
| 3858 |
->theta = (info.max_eigenvalue_estimate + alpha) * 0.5; |
| 3859 |
|
| 3860 |
// We do not need the second temporary vector in case we have a |
| 3861 |
// DiagonalMatrix as preconditioner and use deal.II's own vectors |
| 3862 |
using NumberType = typename VectorType::value_type; |
| 3863 |
if (std::is_same_v<PreconditionerType, dealii::DiagonalMatrix<VectorType>> == |
| 3864 |
false || |
| 3865 |
(std::is_same_v<VectorType, dealii::Vector<NumberType>> == false && |
| 3866 |
((std::is_same_v< |
| 3867 |
VectorType, |
| 3868 |
LinearAlgebra::distributed::Vector<NumberType, MemorySpace::Host>> == |
| 3869 |
false) || |
| 3870 |
(std::is_same_v<VectorType, |
| 3871 |
LinearAlgebra::distributed:: |
| 3872 |
Vector<NumberType, MemorySpace::Default>> == false)))) |
| 3873 |
temp_vector2.reinit(src, true); |
| 3874 |
else |
| 3875 |
{ |
| 3876 |
VectorType empty_vector; |
| 3877 |
temp_vector2.reinit(empty_vector); |
| 3878 |
} |
| 3879 |
|
| 3880 |
const_cast< |
| 3881 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType> *>(this) |
| 3882 |
->eigenvalues_are_initialized = true; |
| 3883 |
|
| 3884 |
return info; |
| 3885 |
} |
| 3886 |
|
| 3887 |
|
| 3888 |
|
| 3889 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3890 |
inline void |
| 3891 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::vmult( |
| 3892 |
VectorType &solution, |
| 3893 |
const VectorType &rhs) const |
| 3894 |
{ |
| 3895 |
std::lock_guard<std::mutex> lock(mutex); |
| 3896 |
if (eigenvalues_are_initialized == false) |
| 3897 |
estimate_eigenvalues(rhs); |
| 3898 |
|
| 3899 |
internal::PreconditionChebyshevImplementation::vmult_and_update( |
| 3900 |
*matrix_ptr, |
| 3901 |
*data.preconditioner, |
| 3902 |
rhs, |
| 3903 |
0, |
| 3904 |
0., |
| 3905 |
(data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) ? |
| 3906 |
(4. / (3. * delta)) : |
| 3907 |
(1. / theta), |
| 3908 |
solution, |
| 3909 |
solution_old, |
| 3910 |
temp_vector1, |
| 3911 |
temp_vector2); |
| 3912 |
|
| 3913 |
// if delta is zero, we do not need to iterate because the updates will be |
| 3914 |
// zero |
| 3915 |
if (data.degree < 2 || std::abs(delta) < 1e-40) |
| 3916 |
return; |
| 3917 |
|
| 3918 |
double rhok = delta / theta, sigma = theta / delta; |
| 3919 |
for (unsigned int k = 0; k < data.degree - 1; ++k) |
| 3920 |
{ |
| 3921 |
double factor1 = 0.0; |
| 3922 |
double factor2 = 0.0; |
| 3923 |
|
| 3924 |
if (data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) |
| 3925 |
{ |
| 3926 |
factor1 = (2 * k + 1.) / (2 * k + 5.); |
| 3927 |
factor2 = (8 * k + 12.) / (delta * (2 * k + 5.)); |
| 3928 |
} |
| 3929 |
else |
| 3930 |
{ |
| 3931 |
const double rhokp = 1. / (2. * sigma - rhok); |
| 3932 |
factor1 = rhokp * rhok; |
| 3933 |
factor2 = 2. * rhokp / delta; |
| 3934 |
rhok = rhokp; |
| 3935 |
} |
| 3936 |
|
| 3937 |
internal::PreconditionChebyshevImplementation::vmult_and_update( |
| 3938 |
*matrix_ptr, |
| 3939 |
*data.preconditioner, |
| 3940 |
rhs, |
| 3941 |
k + 1, |
| 3942 |
factor1, |
| 3943 |
factor2, |
| 3944 |
solution, |
| 3945 |
solution_old, |
| 3946 |
temp_vector1, |
| 3947 |
temp_vector2); |
| 3948 |
} |
| 3949 |
} |
| 3950 |
|
| 3951 |
|
| 3952 |
|
| 3953 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 3954 |
inline void |
| 3955 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::Tvmult( |
| 3956 |
VectorType &solution, |
| 3957 |
const VectorType &rhs) const |
| 3958 |
{ |
| 3959 |
std::lock_guard<std::mutex> lock(mutex); |
| 3960 |
if (eigenvalues_are_initialized == false) |
| 3961 |
estimate_eigenvalues(rhs); |
| 3962 |
|
| 3963 |
internal::PreconditionChebyshevImplementation::vector_updates( |
| 3964 |
rhs, |
| 3965 |
*data.preconditioner, |
| 3966 |
0, |
| 3967 |
0., |
| 3968 |
(data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) ? |
| 3969 |
(4. / (3. * delta)) : |
| 3970 |
(1. / theta), |
| 3971 |
solution_old, |
| 3972 |
temp_vector1, |
| 3973 |
temp_vector2, |
| 3974 |
solution); |
| 3975 |
|
| 3976 |
if (data.degree < 2 || std::abs(delta) < 1e-40) |
| 3977 |
return; |
| 3978 |
|
| 3979 |
double rhok = delta / theta, sigma = theta / delta; |
| 3980 |
for (unsigned int k = 0; k < data.degree - 1; ++k) |
| 3981 |
{ |
| 3982 |
double factor1 = 0.0; |
| 3983 |
double factor2 = 0.0; |
| 3984 |
|
| 3985 |
if (data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) |
| 3986 |
{ |
| 3987 |
factor1 = (2 * k + 1.) / (2 * k + 5.); |
| 3988 |
factor2 = (8 * k + 12.) / (delta * (2 * k + 5.)); |
| 3989 |
} |
| 3990 |
else |
| 3991 |
{ |
| 3992 |
const double rhokp = 1. / (2. * sigma - rhok); |
| 3993 |
factor1 = rhokp * rhok; |
| 3994 |
factor2 = 2. * rhokp / delta; |
| 3995 |
rhok = rhokp; |
| 3996 |
} |
| 3997 |
|
| 3998 |
matrix_ptr->Tvmult(temp_vector1, solution); |
| 3999 |
internal::PreconditionChebyshevImplementation::vector_updates( |
| 4000 |
rhs, |
| 4001 |
*data.preconditioner, |
| 4002 |
k + 1, |
| 4003 |
factor1, |
| 4004 |
factor2, |
| 4005 |
solution_old, |
| 4006 |
temp_vector1, |
| 4007 |
temp_vector2, |
| 4008 |
solution); |
| 4009 |
} |
| 4010 |
} |
| 4011 |
|
| 4012 |
|
| 4013 |
|
| 4014 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 4015 |
inline void |
| 4016 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::step( |
| 4017 |
VectorType &solution, |
| 4018 |
const VectorType &rhs) const |
| 4019 |
{ |
| 4020 |
std::lock_guard<std::mutex> lock(mutex); |
| 4021 |
if (eigenvalues_are_initialized == false) |
| 4022 |
estimate_eigenvalues(rhs); |
| 4023 |
|
| 4024 |
internal::PreconditionChebyshevImplementation::vmult_and_update( |
| 4025 |
*matrix_ptr, |
| 4026 |
*data.preconditioner, |
| 4027 |
rhs, |
| 4028 |
1, |
| 4029 |
0., |
| 4030 |
(data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) ? |
| 4031 |
(4. / (3. * delta)) : |
| 4032 |
(1. / theta), |
| 4033 |
solution, |
| 4034 |
solution_old, |
| 4035 |
temp_vector1, |
| 4036 |
temp_vector2); |
| 4037 |
|
| 4038 |
if (data.degree < 2 || std::abs(delta) < 1e-40) |
| 4039 |
return; |
| 4040 |
|
| 4041 |
double rhok = delta / theta, sigma = theta / delta; |
| 4042 |
for (unsigned int k = 0; k < data.degree - 1; ++k) |
| 4043 |
{ |
| 4044 |
double factor1 = 0.0; |
| 4045 |
double factor2 = 0.0; |
| 4046 |
|
| 4047 |
if (data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) |
| 4048 |
{ |
| 4049 |
factor1 = (2 * k + 1.) / (2 * k + 5.); |
| 4050 |
factor2 = (8 * k + 12.) / (delta * (2 * k + 5.)); |
| 4051 |
} |
| 4052 |
else |
| 4053 |
{ |
| 4054 |
const double rhokp = 1. / (2. * sigma - rhok); |
| 4055 |
factor1 = rhokp * rhok; |
| 4056 |
factor2 = 2. * rhokp / delta; |
| 4057 |
rhok = rhokp; |
| 4058 |
} |
| 4059 |
|
| 4060 |
internal::PreconditionChebyshevImplementation::vmult_and_update( |
| 4061 |
*matrix_ptr, |
| 4062 |
*data.preconditioner, |
| 4063 |
rhs, |
| 4064 |
k + 2, |
| 4065 |
factor1, |
| 4066 |
factor2, |
| 4067 |
solution, |
| 4068 |
solution_old, |
| 4069 |
temp_vector1, |
| 4070 |
temp_vector2); |
| 4071 |
} |
| 4072 |
} |
| 4073 |
|
| 4074 |
|
| 4075 |
|
| 4076 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 4077 |
inline void |
| 4078 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::Tstep( |
| 4079 |
VectorType &solution, |
| 4080 |
const VectorType &rhs) const |
| 4081 |
{ |
| 4082 |
std::lock_guard<std::mutex> lock(mutex); |
| 4083 |
if (eigenvalues_are_initialized == false) |
| 4084 |
estimate_eigenvalues(rhs); |
| 4085 |
|
| 4086 |
matrix_ptr->Tvmult(temp_vector1, solution); |
| 4087 |
internal::PreconditionChebyshevImplementation::vector_updates( |
| 4088 |
rhs, |
| 4089 |
*data.preconditioner, |
| 4090 |
1, |
| 4091 |
0., |
| 4092 |
(data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) ? |
| 4093 |
(4. / (3. * delta)) : |
| 4094 |
(1. / theta), |
| 4095 |
solution_old, |
| 4096 |
temp_vector1, |
| 4097 |
temp_vector2, |
| 4098 |
solution); |
| 4099 |
|
| 4100 |
if (data.degree < 2 || std::abs(delta) < 1e-40) |
| 4101 |
return; |
| 4102 |
|
| 4103 |
double rhok = delta / theta, sigma = theta / delta; |
| 4104 |
for (unsigned int k = 0; k < data.degree - 1; ++k) |
| 4105 |
{ |
| 4106 |
double factor1 = 0.0; |
| 4107 |
double factor2 = 0.0; |
| 4108 |
|
| 4109 |
if (data.polynomial_type == AdditionalData::PolynomialType::fourth_kind) |
| 4110 |
{ |
| 4111 |
factor1 = (2 * k + 1.) / (2 * k + 5.); |
| 4112 |
factor2 = (8 * k + 12.) / (delta * (2 * k + 5.)); |
| 4113 |
} |
| 4114 |
else |
| 4115 |
{ |
| 4116 |
const double rhokp = 1. / (2. * sigma - rhok); |
| 4117 |
factor1 = rhokp * rhok; |
| 4118 |
factor2 = 2. * rhokp / delta; |
| 4119 |
rhok = rhokp; |
| 4120 |
} |
| 4121 |
|
| 4122 |
matrix_ptr->Tvmult(temp_vector1, solution); |
| 4123 |
internal::PreconditionChebyshevImplementation::vector_updates( |
| 4124 |
rhs, |
| 4125 |
*data.preconditioner, |
| 4126 |
k + 2, |
| 4127 |
factor1, |
| 4128 |
factor2, |
| 4129 |
solution_old, |
| 4130 |
temp_vector1, |
| 4131 |
temp_vector2, |
| 4132 |
solution); |
| 4133 |
} |
| 4134 |
} |
| 4135 |
|
| 4136 |
|
| 4137 |
|
| 4138 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 4139 |
inline typename PreconditionChebyshev<MatrixType, |
| 4140 |
VectorType, |
| 4141 |
PreconditionerType>::size_type |
| 4142 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::m() const |
| 4143 |
{ |
| 4144 |
Assert(matrix_ptr != nullptr, ExcNotInitialized()); |
| 4145 |
return matrix_ptr->m(); |
| 4146 |
} |
| 4147 |
|
| 4148 |
|
| 4149 |
|
| 4150 |
template <typename MatrixType, typename VectorType, typename PreconditionerType> |
| 4151 |
inline typename PreconditionChebyshev<MatrixType, |
| 4152 |
VectorType, |
| 4153 |
PreconditionerType>::size_type |
| 4154 |
PreconditionChebyshev<MatrixType, VectorType, PreconditionerType>::n() const |
| 4155 |
{ |
| 4156 |
Assert(matrix_ptr != nullptr, ExcNotInitialized()); |
| 4157 |
return matrix_ptr->n(); |
| 4158 |
} |
| 4159 |
|
| 4160 |
#endif // DOXYGEN |
| 4161 |
|
| 4162 |
DEAL_II_NAMESPACE_CLOSE |
| 4163 |
|
| 4164 |
#endif |