| 1 |
// SPDX-FileCopyrightText: Copyright © DUNE Project contributors, see file LICENSE.md in module root |
| 2 |
// SPDX-License-Identifier: LicenseRef-GPL-2.0-only-with-DUNE-exception |
| 3 |
// -*- tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*- |
| 4 |
// vi: set et ts=4 sw=2 sts=2: |
| 5 |
#ifndef DUNE_ISTL_UMFPACK_HH |
| 6 |
#define DUNE_ISTL_UMFPACK_HH |
| 7 |
|
| 8 |
#if HAVE_SUITESPARSE_UMFPACK || defined DOXYGEN |
| 9 |
|
| 10 |
#include<complex> |
| 11 |
#include<type_traits> |
| 12 |
|
| 13 |
#include<umfpack.h> |
| 14 |
|
| 15 |
#include<dune/common/exceptions.hh> |
| 16 |
#include<dune/common/fmatrix.hh> |
| 17 |
#include<dune/common/fvector.hh> |
| 18 |
#include<dune/istl/bccsmatrixinitializer.hh> |
| 19 |
#include<dune/istl/bcrsmatrix.hh> |
| 20 |
#include<dune/istl/matrix.hh> |
| 21 |
#include<dune/istl/foreach.hh> |
| 22 |
#include<dune/istl/multitypeblockmatrix.hh> |
| 23 |
#include<dune/istl/multitypeblockvector.hh> |
| 24 |
#include<dune/istl/solvers.hh> |
| 25 |
#include<dune/istl/solvertype.hh> |
| 26 |
#include <dune/istl/solverfactory.hh> |
| 27 |
|
| 28 |
|
| 29 |
|
| 30 |
namespace Dune { |
| 31 |
/** |
| 32 |
* @addtogroup ISTL |
| 33 |
* |
| 34 |
* @{ |
| 35 |
*/ |
| 36 |
/** |
| 37 |
* @file |
| 38 |
* @author Dominic Kempf |
| 39 |
* @brief Classes for using UMFPack with ISTL matrices. |
| 40 |
*/ |
| 41 |
|
| 42 |
// FORWARD DECLARATIONS |
| 43 |
template<class M, class T, class TM, class TD, class TA> |
| 44 |
class SeqOverlappingSchwarz; |
| 45 |
|
| 46 |
template<class T, bool tag> |
| 47 |
struct SeqOverlappingSchwarzAssemblerHelper; |
| 48 |
|
| 49 |
// wrapper class for C-Function Calls in the backend. Choose the right function namespace |
| 50 |
// depending on the template parameter used. |
| 51 |
template<typename T> |
| 52 |
struct UMFPackMethodChooser |
| 53 |
{ |
| 54 |
static constexpr bool valid = false ; |
| 55 |
}; |
| 56 |
|
| 57 |
template<> |
| 58 |
struct UMFPackMethodChooser<double> |
| 59 |
{ |
| 60 |
static constexpr bool valid = true ; |
| 61 |
|
| 62 |
template<typename... A> |
| 63 |
static void defaults(A... args) |
| 64 |
{ |
| 65 |
umfpack_dl_defaults(args...); |
| 66 |
} |
| 67 |
template<typename... A> |
| 68 |
static void free_numeric(A... args) |
| 69 |
{ |
| 70 |
umfpack_dl_free_numeric(args...); |
| 71 |
} |
| 72 |
template<typename... A> |
| 73 |
static void free_symbolic(A... args) |
| 74 |
{ |
| 75 |
umfpack_dl_free_symbolic(args...); |
| 76 |
} |
| 77 |
template<typename... A> |
| 78 |
static int load_numeric(A... args) |
| 79 |
{ |
| 80 |
return umfpack_dl_load_numeric(args...); |
| 81 |
} |
| 82 |
template<typename... A> |
| 83 |
static void numeric(A... args) |
| 84 |
{ |
| 85 |
umfpack_dl_numeric(args...); |
| 86 |
} |
| 87 |
template<typename... A> |
| 88 |
static void report_info(A... args) |
| 89 |
{ |
| 90 |
umfpack_dl_report_info(args...); |
| 91 |
} |
| 92 |
template<typename... A> |
| 93 |
static void report_status(A... args) |
| 94 |
{ |
| 95 |
umfpack_dl_report_status(args...); |
| 96 |
} |
| 97 |
template<typename... A> |
| 98 |
static int save_numeric(A... args) |
| 99 |
{ |
| 100 |
return umfpack_dl_save_numeric(args...); |
| 101 |
} |
| 102 |
template<typename... A> |
| 103 |
static void solve(A... args) |
| 104 |
{ |
| 105 |
umfpack_dl_solve(args...); |
| 106 |
} |
| 107 |
template<typename... A> |
| 108 |
static void symbolic(A... args) |
| 109 |
{ |
| 110 |
umfpack_dl_symbolic(args...); |
| 111 |
} |
| 112 |
}; |
| 113 |
|
| 114 |
template<> |
| 115 |
struct UMFPackMethodChooser<std::complex<double> > |
| 116 |
{ |
| 117 |
static constexpr bool valid = true ; |
| 118 |
using size_type = SuiteSparse_long; |
| 119 |
|
| 120 |
template<typename... A> |
| 121 |
static void defaults(A... args) |
| 122 |
{ |
| 123 |
umfpack_zl_defaults(args...); |
| 124 |
} |
| 125 |
template<typename... A> |
| 126 |
static void free_numeric(A... args) |
| 127 |
{ |
| 128 |
umfpack_zl_free_numeric(args...); |
| 129 |
} |
| 130 |
template<typename... A> |
| 131 |
static void free_symbolic(A... args) |
| 132 |
{ |
| 133 |
umfpack_zl_free_symbolic(args...); |
| 134 |
} |
| 135 |
template<typename... A> |
| 136 |
static int load_numeric(A... args) |
| 137 |
{ |
| 138 |
return umfpack_zl_load_numeric(args...); |
| 139 |
} |
| 140 |
template<typename... A> |
| 141 |
static void numeric(const size_type* cs, const size_type* ri, const double* val, A... args) |
| 142 |
{ |
| 143 |
umfpack_zl_numeric(cs,ri,val,NULL,args...); |
| 144 |
} |
| 145 |
template<typename... A> |
| 146 |
static void report_info(A... args) |
| 147 |
{ |
| 148 |
umfpack_zl_report_info(args...); |
| 149 |
} |
| 150 |
template<typename... A> |
| 151 |
static void report_status(A... args) |
| 152 |
{ |
| 153 |
umfpack_zl_report_status(args...); |
| 154 |
} |
| 155 |
template<typename... A> |
| 156 |
static int save_numeric(A... args) |
| 157 |
{ |
| 158 |
return umfpack_zl_save_numeric(args...); |
| 159 |
} |
| 160 |
template<typename... A> |
| 161 |
static void solve(size_type m, const size_type* cs, const size_type* ri, std::complex<double>* val, double* x, const double* b,A... args) |
| 162 |
{ |
| 163 |
const double* cval = reinterpret_cast<const double*>(val); |
| 164 |
umfpack_zl_solve(m,cs,ri,cval,NULL,x,NULL,b,NULL,args...); |
| 165 |
} |
| 166 |
template<typename... A> |
| 167 |
static void symbolic(size_type m, size_type n, const size_type* cs, const size_type* ri, const double* val, A... args) |
| 168 |
{ |
| 169 |
umfpack_zl_symbolic(m,n,cs,ri,val,NULL,args...); |
| 170 |
} |
| 171 |
}; |
| 172 |
|
| 173 |
namespace Impl |
| 174 |
{ |
| 175 |
template<class M, class = void> |
| 176 |
struct UMFPackVectorChooser; |
| 177 |
|
| 178 |
/** @brief The type of the domain of the solver */ |
| 179 |
template<class M> using UMFPackDomainType = typename UMFPackVectorChooser<M>::domain_type; |
| 180 |
|
| 181 |
/** @brief The type of the range of the solver */ |
| 182 |
template<class M> using UMFPackRangeType = typename UMFPackVectorChooser<M>::range_type; |
| 183 |
|
| 184 |
template<class M> |
| 185 |
struct UMFPackVectorChooser<M, |
| 186 |
std::enable_if_t<(std::is_same<M,double>::value) || (std::is_same<M,std::complex<double> >::value)>> |
| 187 |
{ |
| 188 |
using domain_type = M; |
| 189 |
using range_type = M; |
| 190 |
}; |
| 191 |
|
| 192 |
template<typename T, int n, int m> |
| 193 |
struct UMFPackVectorChooser<FieldMatrix<T,n,m>, |
| 194 |
std::enable_if_t<(std::is_same<T,double>::value) || (std::is_same<T,std::complex<double> >::value)>> |
| 195 |
{ |
| 196 |
/** @brief The type of the domain of the solver */ |
| 197 |
using domain_type = FieldVector<T,m>; |
| 198 |
/** @brief The type of the range of the solver */ |
| 199 |
using range_type = FieldVector<T,n>; |
| 200 |
}; |
| 201 |
|
| 202 |
template<typename T, typename A> |
| 203 |
struct UMFPackVectorChooser<BCRSMatrix<T,A>, |
| 204 |
std::void_t<UMFPackDomainType<T>, UMFPackRangeType<T>>> |
| 205 |
{ |
| 206 |
// In case of recursive deduction (e.g., BCRSMatrix<FieldMatrix<...>, Allocator<FieldMatrix<...>>>) |
| 207 |
// the allocator needs to be converted to the sub-block allocator type too (e.g., Allocator<FieldVector<...>>). |
| 208 |
// Note that matrix allocator is assumed to be the same as the domain/range type of allocators |
| 209 |
/** @brief The type of the domain of the solver */ |
| 210 |
using domain_type = BlockVector<UMFPackDomainType<T>, typename std::allocator_traits<A>::template rebind_alloc<UMFPackDomainType<T>>>; |
| 211 |
/** @brief The type of the range of the solver */ |
| 212 |
using range_type = BlockVector<UMFPackRangeType<T>, typename std::allocator_traits<A>::template rebind_alloc<UMFPackRangeType<T>>>; |
| 213 |
}; |
| 214 |
|
| 215 |
template<typename T, typename A> |
| 216 |
struct UMFPackVectorChooser<Matrix<T,A>, |
| 217 |
std::void_t<UMFPackDomainType<T>, UMFPackRangeType<T>>> |
| 218 |
: public UMFPackVectorChooser<BCRSMatrix<T,A>, std::void_t<UMFPackDomainType<T>, UMFPackRangeType<T>>> |
| 219 |
{}; |
| 220 |
|
| 221 |
// to make the `UMFPackVectorChooser` work with `MultiTypeBlockMatrix`, we need to add an intermediate step for the rows, which are typically `MultiTypeBlockVector` |
| 222 |
template<typename FirstBlock, typename... Blocks> |
| 223 |
struct UMFPackVectorChooser<MultiTypeBlockVector<FirstBlock, Blocks...>, |
| 224 |
std::void_t<UMFPackDomainType<FirstBlock>, UMFPackRangeType<FirstBlock>, UMFPackDomainType<Blocks>...>> |
| 225 |
{ |
| 226 |
/** @brief The type of the domain of the solver */ |
| 227 |
using domain_type = MultiTypeBlockVector<UMFPackDomainType<FirstBlock>, UMFPackDomainType<Blocks>...>; |
| 228 |
/** @brief The type of the range of the solver */ |
| 229 |
using range_type = UMFPackRangeType<FirstBlock>; |
| 230 |
}; |
| 231 |
|
| 232 |
// specialization for `MultiTypeBlockMatrix` with `MultiTypeBlockVector` rows |
| 233 |
template<typename FirstRow, typename... Rows> |
| 234 |
struct UMFPackVectorChooser<MultiTypeBlockMatrix<FirstRow, Rows...>, |
| 235 |
std::void_t<UMFPackDomainType<FirstRow>, UMFPackRangeType<FirstRow>, UMFPackRangeType<Rows>...>> |
| 236 |
{ |
| 237 |
/** @brief The type of the domain of the solver */ |
| 238 |
using domain_type = UMFPackDomainType<FirstRow>; |
| 239 |
/** @brief The type of the range of the solver */ |
| 240 |
using range_type = MultiTypeBlockVector< UMFPackRangeType<FirstRow>, UMFPackRangeType<Rows>... >; |
| 241 |
}; |
| 242 |
|
| 243 |
// dummy class to represent no BitVector |
| 244 |
struct NoBitVector |
| 245 |
{}; |
| 246 |
|
| 247 |
|
| 248 |
} |
| 249 |
|
| 250 |
/** @brief The %UMFPack direct sparse solver |
| 251 |
* |
| 252 |
* Details on UMFPack can be found on |
| 253 |
* http://www.cise.ufl.edu/research/sparse/umfpack/ |
| 254 |
* |
| 255 |
* %UMFPack will always use double precision. |
| 256 |
* For complex matrices use a matrix type with std::complex<double> |
| 257 |
* as the underlying number type. |
| 258 |
* |
| 259 |
* \tparam Matrix the matrix type defining the system |
| 260 |
* |
| 261 |
* \note This will only work if dune-istl has been configured to use UMFPack |
| 262 |
*/ |
| 263 |
template<typename M> |
| 264 |
class UMFPack : public InverseOperator<Impl::UMFPackDomainType<M>,Impl::UMFPackRangeType<M>> |
| 265 |
{ |
| 266 |
using T = typename M::field_type; |
| 267 |
|
| 268 |
public: |
| 269 |
using size_type = SuiteSparse_long; |
| 270 |
|
| 271 |
/** @brief The matrix type. */ |
| 272 |
using Matrix = M; |
| 273 |
using matrix_type = M; |
| 274 |
/** @brief The corresponding (scalar) UMFPack matrix type.*/ |
| 275 |
using UMFPackMatrix = ISTL::Impl::BCCSMatrix<typename Matrix::field_type, size_type>; |
| 276 |
/** @brief Type of an associated initializer class. */ |
| 277 |
using MatrixInitializer = ISTL::Impl::BCCSMatrixInitializer<M, size_type>; |
| 278 |
/** @brief The type of the domain of the solver. */ |
| 279 |
using domain_type = Impl::UMFPackDomainType<M>; |
| 280 |
/** @brief The type of the range of the solver. */ |
| 281 |
using range_type = Impl::UMFPackRangeType<M>; |
| 282 |
|
| 283 |
//! Category of the solver (see SolverCategory::Category) |
| 284 |
SolverCategory::Category category() const override |
| 285 |
{ |
| 286 |
return SolverCategory::Category::sequential; |
| 287 |
} |
| 288 |
|
| 289 |
/** @brief Construct a solver object from a matrix |
| 290 |
* |
| 291 |
* This computes the matrix decomposition, and may take a long time |
| 292 |
* (and use a lot of memory). |
| 293 |
* |
| 294 |
* @param matrix the matrix to solve for |
| 295 |
* @param verbose [0..2] set the verbosity level, defaults to 0 |
| 296 |
*/ |
| 297 |
UMFPack(const Matrix& matrix, int verbose=0) : matrixIsLoaded_(false) |
| 298 |
{ |
| 299 |
//check whether T is a supported type |
| 300 |
static_assert((std::is_same<T,double>::value) || (std::is_same<T,std::complex<double> >::value), |
| 301 |
"Unsupported Type in UMFPack (only double and std::complex<double> supported)"); |
| 302 |
Caller::defaults(UMF_Control); |
| 303 |
setVerbosity(verbose); |
| 304 |
setMatrix(matrix); |
| 305 |
} |
| 306 |
|
| 307 |
/** @brief Constructor for compatibility with SuperLU standard constructor |
| 308 |
* |
| 309 |
* This computes the matrix decomposition, and may take a long time |
| 310 |
* (and use a lot of memory). |
| 311 |
* |
| 312 |
* @param matrix the matrix to solve for |
| 313 |
* @param verbose [0..2] set the verbosity level, defaults to 0 |
| 314 |
*/ |
| 315 |
UMFPack(const Matrix& matrix, int verbose, bool) : matrixIsLoaded_(false) |
| 316 |
{ |
| 317 |
//check whether T is a supported type |
| 318 |
static_assert((std::is_same<T,double>::value) || (std::is_same<T,std::complex<double> >::value), |
| 319 |
"Unsupported Type in UMFPack (only double and std::complex<double> supported)"); |
| 320 |
Caller::defaults(UMF_Control); |
| 321 |
setVerbosity(verbose); |
| 322 |
setMatrix(matrix); |
| 323 |
} |
| 324 |
|
| 325 |
/** @brief Construct a solver object from a matrix |
| 326 |
* |
| 327 |
* @param mat_ the matrix to solve for |
| 328 |
* @param config ParameterTree containing solver parameters. |
| 329 |
* |
| 330 |
* ParameterTree Key | Meaning |
| 331 |
* ------------------|------------ |
| 332 |
* verbose | The verbosity level. default=0 |
| 333 |
*/ |
| 334 |
UMFPack(const Matrix& mat_, const ParameterTree& config) |
| 335 |
: UMFPack(mat_, config.get<int>("verbose", 0)) |
| 336 |
{} |
| 337 |
|
| 338 |
/** @brief default constructor |
| 339 |
*/ |
| 340 |
UMFPack() : matrixIsLoaded_(false), verbosity_(0) |
| 341 |
{ |
| 342 |
//check whether T is a supported type |
| 343 |
static_assert((std::is_same<T,double>::value) || (std::is_same<T,std::complex<double> >::value), |
| 344 |
"Unsupported Type in UMFPack (only double and std::complex<double> supported)"); |
| 345 |
Caller::defaults(UMF_Control); |
| 346 |
} |
| 347 |
|
| 348 |
/** @brief Try loading a decomposition from file and do a decomposition if unsuccessful |
| 349 |
* @param mat_ the matrix to decompose when no decoposition file found |
| 350 |
* @param file the decomposition file |
| 351 |
* @param verbose the verbosity level |
| 352 |
* |
| 353 |
* Use saveDecomposition(char* file) for manually storing a decomposition. This constructor |
| 354 |
* will decompose mat_ and store the result to file if no file wasn't found in the first place. |
| 355 |
* Thus, if you always use this you will only compute the decomposition once (and when you manually |
| 356 |
* deleted the decomposition file). |
| 357 |
*/ |
| 358 |
UMFPack(const Matrix& mat_, const char* file, int verbose=0) |
| 359 |
{ |
| 360 |
//check whether T is a supported type |
| 361 |
static_assert((std::is_same<T,double>::value) || (std::is_same<T,std::complex<double> >::value), |
| 362 |
"Unsupported Type in UMFPack (only double and std::complex<double> supported)"); |
| 363 |
Caller::defaults(UMF_Control); |
| 364 |
setVerbosity(verbose); |
| 365 |
int errcode = Caller::load_numeric(&UMF_Numeric, const_cast<char*>(file)); |
| 366 |
if ((errcode == UMFPACK_ERROR_out_of_memory) || (errcode == UMFPACK_ERROR_file_IO)) |
| 367 |
{ |
| 368 |
matrixIsLoaded_ = false; |
| 369 |
setMatrix(mat_); |
| 370 |
saveDecomposition(file); |
| 371 |
} |
| 372 |
else |
| 373 |
{ |
| 374 |
matrixIsLoaded_ = true; |
| 375 |
std::cout << "UMFPack decomposition successfully loaded from " << file << std::endl; |
| 376 |
} |
| 377 |
} |
| 378 |
|
| 379 |
/** @brief try loading a decomposition from file |
| 380 |
* @param file the decomposition file |
| 381 |
* @param verbose the verbosity level |
| 382 |
* @throws Dune::Exception When not being able to load the file. Does not need knowledge of the |
| 383 |
* actual matrix! |
| 384 |
*/ |
| 385 |
UMFPack(const char* file, int verbose=0) |
| 386 |
{ |
| 387 |
//check whether T is a supported type |
| 388 |
static_assert((std::is_same<T,double>::value) || (std::is_same<T,std::complex<double> >::value), |
| 389 |
"Unsupported Type in UMFPack (only double and std::complex<double> supported)"); |
| 390 |
Caller::defaults(UMF_Control); |
| 391 |
int errcode = Caller::load_numeric(&UMF_Numeric, const_cast<char*>(file)); |
| 392 |
if (errcode == UMFPACK_ERROR_out_of_memory) |
| 393 |
DUNE_THROW(Dune::Exception, "ran out of memory while loading UMFPack decomposition"); |
| 394 |
if (errcode == UMFPACK_ERROR_file_IO) |
| 395 |
DUNE_THROW(Dune::Exception, "IO error while loading UMFPack decomposition"); |
| 396 |
matrixIsLoaded_ = true; |
| 397 |
std::cout << "UMFPack decomposition successfully loaded from " << file << std::endl; |
| 398 |
setVerbosity(verbose); |
| 399 |
} |
| 400 |
|
| 401 |
virtual ~UMFPack() |
| 402 |
{ |
| 403 |
if ((umfpackMatrix_.N() + umfpackMatrix_.M() > 0) || matrixIsLoaded_) |
| 404 |
free(); |
| 405 |
} |
| 406 |
|
| 407 |
/** |
| 408 |
* \copydoc InverseOperator::apply(X&, Y&, InverseOperatorResult&) |
| 409 |
*/ |
| 410 |
void apply(domain_type& x, range_type& b, InverseOperatorResult& res) override |
| 411 |
{ |
| 412 |
if (umfpackMatrix_.N() != b.dim()) |
| 413 |
DUNE_THROW(Dune::ISTLError, "Size of right-hand-side vector b does not match the number of matrix rows!"); |
| 414 |
if (umfpackMatrix_.M() != x.dim()) |
| 415 |
DUNE_THROW(Dune::ISTLError, "Size of solution vector x does not match the number of matrix columns!"); |
| 416 |
if (b.size() == 0) |
| 417 |
return; |
| 418 |
|
| 419 |
// we have to convert x and b into flat structures |
| 420 |
// however, this is linear in time |
| 421 |
std::vector<T> xFlat(x.dim()), bFlat(b.dim()); |
| 422 |
|
| 423 |
flatVectorForEach(x, [&](auto&& entry, auto&& offset) |
| 424 |
{ |
| 425 |
xFlat[ offset ] = entry; |
| 426 |
}); |
| 427 |
|
| 428 |
flatVectorForEach(b, [&](auto&& entry, auto&& offset) |
| 429 |
{ |
| 430 |
bFlat[ offset ] = entry; |
| 431 |
}); |
| 432 |
|
| 433 |
double UMF_Apply_Info[UMFPACK_INFO]; |
| 434 |
Caller::solve(UMFPACK_A, |
| 435 |
umfpackMatrix_.getColStart(), |
| 436 |
umfpackMatrix_.getRowIndex(), |
| 437 |
umfpackMatrix_.getValues(), |
| 438 |
reinterpret_cast<double*>(&xFlat[0]), |
| 439 |
reinterpret_cast<double*>(&bFlat[0]), |
| 440 |
UMF_Numeric, |
| 441 |
UMF_Control, |
| 442 |
UMF_Apply_Info); |
| 443 |
|
| 444 |
// copy back to blocked vector |
| 445 |
flatVectorForEach(x, [&](auto&& entry, auto offset) |
| 446 |
{ |
| 447 |
entry = xFlat[offset]; |
| 448 |
}); |
| 449 |
|
| 450 |
//this is a direct solver |
| 451 |
res.iterations = 1; |
| 452 |
res.converged = true; |
| 453 |
res.elapsed = UMF_Apply_Info[UMFPACK_SOLVE_WALLTIME]; |
| 454 |
|
| 455 |
printOnApply(UMF_Apply_Info); |
| 456 |
} |
| 457 |
|
| 458 |
/** |
| 459 |
* \copydoc InverseOperator::apply(X&,Y&,double,InverseOperatorResult&) |
| 460 |
*/ |
| 461 |
void apply (domain_type& x, range_type& b, [[maybe_unused]] double reduction, InverseOperatorResult& res) override |
| 462 |
{ |
| 463 |
apply(x,b,res); |
| 464 |
} |
| 465 |
|
| 466 |
/** |
| 467 |
* @brief additional apply method with c-arrays in analogy to superlu |
| 468 |
* @param x solution array |
| 469 |
* @param b rhs array |
| 470 |
* |
| 471 |
* NOTE If the user hands over a pure pointer, we assume that they know what they are doing, hence no copy to flat structures |
| 472 |
*/ |
| 473 |
void apply(T* x, T* b) |
| 474 |
{ |
| 475 |
double UMF_Apply_Info[UMFPACK_INFO]; |
| 476 |
Caller::solve(UMFPACK_A, |
| 477 |
umfpackMatrix_.getColStart(), |
| 478 |
umfpackMatrix_.getRowIndex(), |
| 479 |
umfpackMatrix_.getValues(), |
| 480 |
x, |
| 481 |
b, |
| 482 |
UMF_Numeric, |
| 483 |
UMF_Control, |
| 484 |
UMF_Apply_Info); |
| 485 |
printOnApply(UMF_Apply_Info); |
| 486 |
} |
| 487 |
|
| 488 |
/** @brief Set UMFPack-specific options |
| 489 |
* |
| 490 |
* This method allows to set various options that control the UMFPack solver. |
| 491 |
* More specifically, it allows to set values in the UMF_Control array. |
| 492 |
* Please see the UMFPack documentation for a list of possible options and values. |
| 493 |
* |
| 494 |
* \param option Entry in the UMF_Control array, e.g., UMFPACK_IRSTEP |
| 495 |
* \param value Corresponding value |
| 496 |
* |
| 497 |
* \throws RangeError If nonexisting option was requested |
| 498 |
*/ |
| 499 |
void setOption(unsigned int option, double value) |
| 500 |
{ |
| 501 |
if (option >= UMFPACK_CONTROL) |
| 502 |
DUNE_THROW(RangeError, "Requested non-existing UMFPack option"); |
| 503 |
|
| 504 |
UMF_Control[option] = value; |
| 505 |
} |
| 506 |
|
| 507 |
/** @brief saves a decomposition to a file |
| 508 |
* @param file the filename to save to |
| 509 |
*/ |
| 510 |
void saveDecomposition(const char* file) |
| 511 |
{ |
| 512 |
int errcode = Caller::save_numeric(UMF_Numeric, const_cast<char*>(file)); |
| 513 |
if (errcode != UMFPACK_OK) |
| 514 |
DUNE_THROW(Dune::Exception,"IO ERROR while trying to save UMFPack decomposition"); |
| 515 |
} |
| 516 |
|
| 517 |
/** @brief Initialize data from given matrix. |
| 518 |
* |
| 519 |
* \tparam BitVector a compatible bitvector to the domain_type/range_type. |
| 520 |
* Defaults to NoBitVector for backwards compatibility |
| 521 |
* |
| 522 |
* A positive bit indices that the corresponding matrix row/column is excluded from the |
| 523 |
* UMFPACK decomposition. |
| 524 |
* WARNING This is an opposite behavior of the previous implementation in `setSubMatrix`. |
| 525 |
*/ |
| 526 |
template<class BitVector = Impl::NoBitVector> |
| 527 |
void setMatrix(const Matrix& matrix, const BitVector& bitVector = {}) |
| 528 |
{ |
| 529 |
if ((umfpackMatrix_.N() + umfpackMatrix_.M() > 0) || matrixIsLoaded_) |
| 530 |
free(); |
| 531 |
if (matrix.N() == 0 or matrix.M() == 0) |
| 532 |
return; |
| 533 |
|
| 534 |
if (umfpackMatrix_.N() + umfpackMatrix_.M() + umfpackMatrix_.nonzeroes() != 0) |
| 535 |
umfpackMatrix_.free(); |
| 536 |
|
| 537 |
constexpr bool useBitVector = not std::is_same_v<BitVector,Impl::NoBitVector>; |
| 538 |
|
| 539 |
// use a dynamic flat vector for the bitset |
| 540 |
std::vector<bool> flatBitVector; |
| 541 |
// and a mapping from the compressed indices |
| 542 |
std::vector<size_type> subIndices; |
| 543 |
|
| 544 |
[[maybe_unused]] int numberOfIgnoredDofs = 0; |
| 545 |
int nonZeros = 0; |
| 546 |
|
| 547 |
if constexpr ( useBitVector ) |
| 548 |
{ |
| 549 |
auto flatSize = flatVectorForEach(bitVector, [](auto&&, auto&&){}); |
| 550 |
flatBitVector.resize(flatSize); |
| 551 |
|
| 552 |
flatVectorForEach(bitVector, [&](auto&& entry, auto&& offset) |
| 553 |
{ |
| 554 |
flatBitVector[ offset ] = entry; |
| 555 |
if ( entry ) |
| 556 |
{ |
| 557 |
numberOfIgnoredDofs++; |
| 558 |
} |
| 559 |
}); |
| 560 |
} |
| 561 |
|
| 562 |
// compute the flat dimension and the number of nonzeros of the matrix |
| 563 |
auto [flatRows,flatCols] = flatMatrixForEach( matrix, [&](auto&& /*entry*/, auto&& row, auto&& col){ |
| 564 |
// do not count ignored entries |
| 565 |
if constexpr ( useBitVector ) |
| 566 |
if ( flatBitVector[row] or flatBitVector[col] ) |
| 567 |
return; |
| 568 |
|
| 569 |
nonZeros++; |
| 570 |
}); |
| 571 |
|
| 572 |
if constexpr ( useBitVector ) |
| 573 |
{ |
| 574 |
// use the original flatRows! |
| 575 |
subIndices.resize(flatRows,std::numeric_limits<std::size_t>::max()); |
| 576 |
|
| 577 |
size_type subIndexCounter = 0; |
| 578 |
for ( size_type i=0; i<size_type(flatRows); i++ ) |
| 579 |
if ( not flatBitVector[ i ] ) |
| 580 |
subIndices[ i ] = subIndexCounter++; |
| 581 |
|
| 582 |
// update the original matrix size |
| 583 |
flatRows -= numberOfIgnoredDofs; |
| 584 |
flatCols -= numberOfIgnoredDofs; |
| 585 |
} |
| 586 |
|
| 587 |
|
| 588 |
umfpackMatrix_.setSize(flatRows,flatCols); |
| 589 |
umfpackMatrix_.Nnz_ = nonZeros; |
| 590 |
|
| 591 |
// prepare the arrays |
| 592 |
umfpackMatrix_.colstart = new size_type[flatCols+1]; |
| 593 |
umfpackMatrix_.rowindex = new size_type[nonZeros]; |
| 594 |
umfpackMatrix_.values = new T[nonZeros]; |
| 595 |
|
| 596 |
for ( size_type i=0; i<size_type(flatCols+1); i++ ) |
| 597 |
{ |
| 598 |
umfpackMatrix_.colstart[i] = 0; |
| 599 |
} |
| 600 |
|
| 601 |
// at first, we need to compute the column start indices |
| 602 |
// therefore, we count all entries in each column and in the end we accumulate everything |
| 603 |
flatMatrixForEach(matrix, [&](auto&& /*entry*/, auto&& flatRowIndex, auto&& flatColIndex) |
| 604 |
{ |
| 605 |
// do nothing if entry is excluded |
| 606 |
if constexpr ( useBitVector ) |
| 607 |
if ( flatBitVector[flatRowIndex] or flatBitVector[flatColIndex] ) |
| 608 |
return; |
| 609 |
|
| 610 |
// pick compressed or uncompressed index |
| 611 |
// compiler will hopefully do some constexpr optimization here |
| 612 |
auto colIdx = useBitVector ? subIndices[flatColIndex] : flatColIndex; |
| 613 |
|
| 614 |
umfpackMatrix_.colstart[colIdx+1]++; |
| 615 |
}); |
| 616 |
|
| 617 |
// now accumulate |
| 618 |
for ( size_type i=0; i<(size_type)flatCols; i++ ) |
| 619 |
{ |
| 620 |
umfpackMatrix_.colstart[i+1] += umfpackMatrix_.colstart[i]; |
| 621 |
} |
| 622 |
|
| 623 |
// we need a compressed position counter in each column |
| 624 |
std::vector<size_type> colPosition(flatCols,0); |
| 625 |
|
| 626 |
// now we can set the entries: the procedure below works with both row- or column major index ordering |
| 627 |
flatMatrixForEach(matrix, [&](auto&& entry, auto&& flatRowIndex, auto&& flatColIndex) |
| 628 |
{ |
| 629 |
// do nothing if entry is excluded |
| 630 |
if constexpr ( useBitVector ) |
| 631 |
if ( flatBitVector[flatRowIndex] or flatBitVector[flatColIndex] ) |
| 632 |
return; |
| 633 |
|
| 634 |
// pick compressed or uncompressed index |
| 635 |
// compiler will hopefully do some constexpr optimization here |
| 636 |
auto rowIdx = useBitVector ? subIndices[flatRowIndex] : flatRowIndex; |
| 637 |
auto colIdx = useBitVector ? subIndices[flatColIndex] : flatColIndex; |
| 638 |
|
| 639 |
// the start index of each column is already fixed |
| 640 |
auto colStart = umfpackMatrix_.colstart[colIdx]; |
| 641 |
// get the current number of picked elements in this column |
| 642 |
auto colPos = colPosition[colIdx]; |
| 643 |
// assign the corresponding row index and the value of this element |
| 644 |
umfpackMatrix_.rowindex[ colStart + colPos ] = rowIdx; |
| 645 |
umfpackMatrix_.values[ colStart + colPos ] = entry; |
| 646 |
// increase the number of picked elements in this column |
| 647 |
colPosition[colIdx]++; |
| 648 |
}); |
| 649 |
|
| 650 |
decompose(); |
| 651 |
} |
| 652 |
|
| 653 |
// Keep legacy version using a set of scalar indices |
| 654 |
// The new version using a bitVector type for marking the active matrix indices is |
| 655 |
// directly given in `setMatrix` with an additional BitVector argument. |
| 656 |
// The new version is more flexible and allows, e.g., marking single components of a matrix block. |
| 657 |
void setSubMatrix(const Matrix& _mat, const std::set<typename Matrix::size_type>& rowIndexSet) |
| 658 |
{ |
| 659 |
if ((umfpackMatrix_.N() + umfpackMatrix_.M() > 0) || matrixIsLoaded_) |
| 660 |
free(); |
| 661 |
|
| 662 |
if (umfpackMatrix_.N() + umfpackMatrix_.M() + umfpackMatrix_.nonzeroes() != 0) |
| 663 |
umfpackMatrix_.free(); |
| 664 |
|
| 665 |
umfpackMatrix_.setSize(rowIndexSet.size()*MatrixDimension<Matrix>::rowdim(_mat) / _mat.N(), |
| 666 |
rowIndexSet.size()*MatrixDimension<Matrix>::coldim(_mat) / _mat.M()); |
| 667 |
ISTL::Impl::BCCSMatrixInitializer<Matrix, SuiteSparse_long> initializer(umfpackMatrix_); |
| 668 |
|
| 669 |
copyToBCCSMatrix(initializer, ISTL::Impl::MatrixRowSubset<Matrix,std::set<typename Matrix::size_type> >(_mat,rowIndexSet)); |
| 670 |
|
| 671 |
decompose(); |
| 672 |
} |
| 673 |
|
| 674 |
/** @brief sets the verbosity level for the UMFPack solver |
| 675 |
* @param v verbosity level |
| 676 |
* The following levels are implemented: |
| 677 |
* 0 - only error messages |
| 678 |
* 1 - a bit of statistics on decomposition and solution |
| 679 |
* 2 - lots of statistics on decomposition and solution |
| 680 |
*/ |
| 681 |
void setVerbosity(int v) |
| 682 |
{ |
| 683 |
verbosity_ = v; |
| 684 |
// set the verbosity level in UMFPack |
| 685 |
if (verbosity_ == 0) |
| 686 |
UMF_Control[UMFPACK_PRL] = 1; |
| 687 |
if (verbosity_ == 1) |
| 688 |
UMF_Control[UMFPACK_PRL] = 2; |
| 689 |
if (verbosity_ == 2) |
| 690 |
UMF_Control[UMFPACK_PRL] = 4; |
| 691 |
} |
| 692 |
|
| 693 |
/** |
| 694 |
* @brief Return the matrix factorization. |
| 695 |
* @warning It is up to the user to keep consistency. |
| 696 |
*/ |
| 697 |
void* getFactorization() |
| 698 |
{ |
| 699 |
return UMF_Numeric; |
| 700 |
} |
| 701 |
|
| 702 |
/** |
| 703 |
* @brief Return the column compress matrix from UMFPack. |
| 704 |
* @warning It is up to the user to keep consistency. |
| 705 |
*/ |
| 706 |
UMFPackMatrix& getInternalMatrix() |
| 707 |
{ |
| 708 |
return umfpackMatrix_; |
| 709 |
} |
| 710 |
|
| 711 |
/** |
| 712 |
* @brief free allocated space. |
| 713 |
* @warning later calling apply will result in an error. |
| 714 |
*/ |
| 715 |
void free() |
| 716 |
{ |
| 717 |
if (!matrixIsLoaded_) |
| 718 |
{ |
| 719 |
Caller::free_symbolic(&UMF_Symbolic); |
| 720 |
umfpackMatrix_.free(); |
| 721 |
} |
| 722 |
Caller::free_numeric(&UMF_Numeric); |
| 723 |
matrixIsLoaded_ = false; |
| 724 |
} |
| 725 |
|
| 726 |
const char* name() { return "UMFPACK"; } |
| 727 |
|
| 728 |
private: |
| 729 |
typedef typename Dune::UMFPackMethodChooser<T> Caller; |
| 730 |
|
| 731 |
template<class Mat,class X, class TM, class TD, class T1> |
| 732 |
friend class SeqOverlappingSchwarz; |
| 733 |
friend struct SeqOverlappingSchwarzAssemblerHelper<UMFPack<Matrix>,true>; |
| 734 |
|
| 735 |
/** @brief computes the LU Decomposition */ |
| 736 |
void decompose() |
| 737 |
{ |
| 738 |
double UMF_Decomposition_Info[UMFPACK_INFO]; |
| 739 |
Caller::symbolic(static_cast<SuiteSparse_long>(umfpackMatrix_.N()), |
| 740 |
static_cast<SuiteSparse_long>(umfpackMatrix_.N()), |
| 741 |
umfpackMatrix_.getColStart(), |
| 742 |
umfpackMatrix_.getRowIndex(), |
| 743 |
reinterpret_cast<double*>(umfpackMatrix_.getValues()), |
| 744 |
&UMF_Symbolic, |
| 745 |
UMF_Control, |
| 746 |
UMF_Decomposition_Info); |
| 747 |
Caller::numeric(umfpackMatrix_.getColStart(), |
| 748 |
umfpackMatrix_.getRowIndex(), |
| 749 |
reinterpret_cast<double*>(umfpackMatrix_.getValues()), |
| 750 |
UMF_Symbolic, |
| 751 |
&UMF_Numeric, |
| 752 |
UMF_Control, |
| 753 |
UMF_Decomposition_Info); |
| 754 |
Caller::report_status(UMF_Control,UMF_Decomposition_Info[UMFPACK_STATUS]); |
| 755 |
if (verbosity_ == 1) |
| 756 |
{ |
| 757 |
std::cout << "[UMFPack Decomposition]" << std::endl; |
| 758 |
std::cout << "Wallclock Time taken: " << UMF_Decomposition_Info[UMFPACK_NUMERIC_WALLTIME] << " (CPU Time: " << UMF_Decomposition_Info[UMFPACK_NUMERIC_TIME] << ")" << std::endl; |
| 759 |
std::cout << "Flops taken: " << UMF_Decomposition_Info[UMFPACK_FLOPS] << std::endl; |
| 760 |
std::cout << "Peak Memory Usage: " << UMF_Decomposition_Info[UMFPACK_PEAK_MEMORY]*UMF_Decomposition_Info[UMFPACK_SIZE_OF_UNIT] << " bytes" << std::endl; |
| 761 |
std::cout << "Condition number estimate: " << 1./UMF_Decomposition_Info[UMFPACK_RCOND] << std::endl; |
| 762 |
std::cout << "Numbers of non-zeroes in decomposition: L: " << UMF_Decomposition_Info[UMFPACK_LNZ] << " U: " << UMF_Decomposition_Info[UMFPACK_UNZ] << std::endl; |
| 763 |
} |
| 764 |
if (verbosity_ == 2) |
| 765 |
{ |
| 766 |
Caller::report_info(UMF_Control,UMF_Decomposition_Info); |
| 767 |
} |
| 768 |
} |
| 769 |
|
| 770 |
void printOnApply(double* UMF_Info) |
| 771 |
{ |
| 772 |
Caller::report_status(UMF_Control,UMF_Info[UMFPACK_STATUS]); |
| 773 |
if (verbosity_ > 0) |
| 774 |
{ |
| 775 |
std::cout << "[UMFPack Solve]" << std::endl; |
| 776 |
std::cout << "Wallclock Time: " << UMF_Info[UMFPACK_SOLVE_WALLTIME] << " (CPU Time: " << UMF_Info[UMFPACK_SOLVE_TIME] << ")" << std::endl; |
| 777 |
std::cout << "Flops Taken: " << UMF_Info[UMFPACK_SOLVE_FLOPS] << std::endl; |
| 778 |
std::cout << "Iterative Refinement steps taken: " << UMF_Info[UMFPACK_IR_TAKEN] << std::endl; |
| 779 |
std::cout << "Error Estimate: " << UMF_Info[UMFPACK_OMEGA1] << " resp. " << UMF_Info[UMFPACK_OMEGA2] << std::endl; |
| 780 |
} |
| 781 |
} |
| 782 |
|
| 783 |
UMFPackMatrix umfpackMatrix_; |
| 784 |
bool matrixIsLoaded_; |
| 785 |
int verbosity_; |
| 786 |
void *UMF_Symbolic; |
| 787 |
void *UMF_Numeric; |
| 788 |
double UMF_Control[UMFPACK_CONTROL]; |
| 789 |
}; |
| 790 |
|
| 791 |
template<typename M> |
| 792 |
struct IsDirectSolver<UMFPack<M>> |
| 793 |
{ |
| 794 |
enum { value=true}; |
| 795 |
}; |
| 796 |
|
| 797 |
template<typename T, typename A> |
| 798 |
struct StoresColumnCompressed<UMFPack<BCRSMatrix<T,A> > > |
| 799 |
{ |
| 800 |
enum { value = true }; |
| 801 |
}; |
| 802 |
|
| 803 |
namespace UMFPackImpl { |
| 804 |
template<class OpTraits, class=void> struct isValidBlock : std::false_type{}; |
| 805 |
template<class OpTraits> struct isValidBlock<OpTraits, |
| 806 |
std::enable_if_t< |
| 807 |
std::is_same_v<Impl::UMFPackDomainType<typename OpTraits::matrix_type>, typename OpTraits::domain_type> |
| 808 |
&& std::is_same_v<Impl::UMFPackRangeType<typename OpTraits::matrix_type>, typename OpTraits::range_type> |
| 809 |
&& std::is_same_v<typename FieldTraits<typename OpTraits::domain_type::field_type>::real_type, double> |
| 810 |
&& std::is_same_v<typename FieldTraits<typename OpTraits::range_type::field_type>::real_type, double> |
| 811 |
>> : std::true_type {}; |
| 812 |
} |
| 813 |
DUNE_REGISTER_SOLVER("umfpack", |
| 814 |
[](auto opTraits, const auto& op, const Dune::ParameterTree& config) |
| 815 |
-> std::shared_ptr<typename decltype(opTraits)::solver_type> |
| 816 |
{ |
| 817 |
using OpTraits = decltype(opTraits); |
| 818 |
// works only for sequential operators |
| 819 |
if constexpr (OpTraits::isParallel){ |
| 820 |
if(opTraits.getCommOrThrow(op).communicator().size() > 1) |
| 821 |
DUNE_THROW(Dune::InvalidStateException, "UMFPack works only for sequential operators."); |
| 822 |
} |
| 823 |
if constexpr (OpTraits::isAssembled){ |
| 824 |
using M = typename OpTraits::matrix_type; |
| 825 |
// check if UMFPack<M>* is convertible to |
| 826 |
// InverseOperator*. This checks compatibility of the |
| 827 |
// domain and range types |
| 828 |
if constexpr (UMFPackImpl::isValidBlock<OpTraits>::value) { |
| 829 |
const auto& A = opTraits.getAssembledOpOrThrow(op); |
| 830 |
const M& mat = A->getmat(); |
| 831 |
int verbose = config.get("verbose", 0); |
| 832 |
return std::make_shared<Dune::UMFPack<M>>(mat,verbose); |
| 833 |
} |
| 834 |
} |
| 835 |
DUNE_THROW(UnsupportedType, |
| 836 |
"Unsupported Type in UMFPack (only double and std::complex<double> supported)"); |
| 837 |
return nullptr; |
| 838 |
}); |
| 839 |
} // end namespace Dune |
| 840 |
|
| 841 |
#endif // HAVE_SUITESPARSE_UMFPACK |
| 842 |
|
| 843 |
#endif //DUNE_ISTL_UMFPACK_HH |