| 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_affine_constraints_templates_h |
| 16 |
#define dealii_affine_constraints_templates_h |
| 17 |
|
| 18 |
#include <deal.II/base/config.h> |
| 19 |
|
| 20 |
#include <deal.II/base/memory_consumption.h> |
| 21 |
#include <deal.II/base/parallel.h> |
| 22 |
#include <deal.II/base/table.h> |
| 23 |
#include <deal.II/base/thread_local_storage.h> |
| 24 |
#include <deal.II/base/thread_management.h> |
| 25 |
#include <deal.II/base/trilinos_utilities.h> |
| 26 |
|
| 27 |
#include <deal.II/lac/affine_constraints.h> |
| 28 |
#include <deal.II/lac/block_sparse_matrix.h> |
| 29 |
#include <deal.II/lac/block_sparse_matrix_ez.h> |
| 30 |
#include <deal.II/lac/block_sparsity_pattern.h> |
| 31 |
#include <deal.II/lac/block_vector.h> |
| 32 |
#include <deal.II/lac/chunk_sparse_matrix.h> |
| 33 |
#include <deal.II/lac/diagonal_matrix.h> |
| 34 |
#include <deal.II/lac/dynamic_sparsity_pattern.h> |
| 35 |
#include <deal.II/lac/full_matrix.h> |
| 36 |
#include <deal.II/lac/la_parallel_block_vector.h> |
| 37 |
#include <deal.II/lac/la_parallel_vector.h> |
| 38 |
#include <deal.II/lac/matrix_block.h> |
| 39 |
#include <deal.II/lac/petsc_block_sparse_matrix.h> |
| 40 |
#include <deal.II/lac/petsc_block_vector.h> |
| 41 |
#include <deal.II/lac/petsc_sparse_matrix.h> |
| 42 |
#include <deal.II/lac/petsc_vector.h> |
| 43 |
#include <deal.II/lac/sparse_matrix.h> |
| 44 |
#include <deal.II/lac/sparse_matrix_ez.h> |
| 45 |
#include <deal.II/lac/sparsity_pattern.h> |
| 46 |
#include <deal.II/lac/trilinos_block_sparse_matrix.h> |
| 47 |
#include <deal.II/lac/trilinos_parallel_block_vector.h> |
| 48 |
#include <deal.II/lac/trilinos_sparse_matrix.h> |
| 49 |
#include <deal.II/lac/trilinos_tpetra_block_sparse_matrix.h> |
| 50 |
#include <deal.II/lac/trilinos_tpetra_block_vector.h> |
| 51 |
#include <deal.II/lac/trilinos_tpetra_sparse_matrix.h> |
| 52 |
#include <deal.II/lac/trilinos_tpetra_vector.h> |
| 53 |
#include <deal.II/lac/trilinos_vector.h> |
| 54 |
|
| 55 |
#include <boost/container/small_vector.hpp> |
| 56 |
#include <boost/range/iterator_range_core.hpp> |
| 57 |
#include <boost/serialization/complex.hpp> |
| 58 |
#include <boost/serialization/utility.hpp> |
| 59 |
|
| 60 |
#include <Kokkos_Core.hpp> |
| 61 |
|
| 62 |
#include <algorithm> |
| 63 |
#include <complex> |
| 64 |
#include <iomanip> |
| 65 |
#include <numeric> |
| 66 |
#include <ostream> |
| 67 |
#include <set> |
| 68 |
|
| 69 |
DEAL_II_NAMESPACE_OPEN |
| 70 |
|
| 71 |
|
| 72 |
|
| 73 |
template <typename number> |
| 74 |
std::size_t |
| 75 |
AffineConstraints<number>::ConstraintLine::memory_consumption() const |
| 76 |
{ |
| 77 |
return (MemoryConsumption::memory_consumption(index) + |
| 78 |
MemoryConsumption::memory_consumption(entries) + |
| 79 |
MemoryConsumption::memory_consumption(inhomogeneity)); |
| 80 |
} |
| 81 |
|
| 82 |
|
| 83 |
|
| 84 |
template <typename number> |
| 85 |
void |
| 86 |
AffineConstraints<number>::add_constraint( |
| 87 |
const size_type constrained_dof, |
| 88 |
const ArrayView<const std::pair<size_type, number>> &dependencies, |
| 89 |
const number inhomogeneity) |
| 90 |
{ |
| 91 |
Assert(sorted == false, ExcMatrixIsClosed()); |
| 92 |
Assert(is_constrained(constrained_dof) == false, |
| 93 |
ExcMessage("You cannot add a constraint for a degree of freedom " |
| 94 |
"that is already constrained.")); |
| 95 |
|
| 96 |
// In debug mode, make sure that columns don't show up more than |
| 97 |
// once. Do this by sorting an array of the column indices. Try to |
| 98 |
// avoid memory allocation by using a vector type that allocates up |
| 99 |
// to 25 entries on the stack -- this is enough for the constraints |
| 100 |
// of Q4 elements in 3d, and so should cover the vast majority of |
| 101 |
// cases. If we have a constraint with more dependencies, then |
| 102 |
// that's just going to require a heap allocation. |
| 103 |
if constexpr (running_in_debug_mode()) |
| 104 |
{ |
| 105 |
{ |
| 106 |
boost::container::small_vector<size_type, 25> column_indices; |
| 107 |
column_indices.reserve(dependencies.size()); |
| 108 |
for (const auto &d : dependencies) |
| 109 |
column_indices.emplace_back(d.first); |
| 110 |
std::sort(column_indices.begin(), column_indices.end()); |
| 111 |
Assert( |
| 112 |
std::adjacent_find(column_indices.begin(), column_indices.end()) == |
| 113 |
column_indices.end(), |
| 114 |
ExcMessage( |
| 115 |
"You are trying to insert a constraint that lists the same " |
| 116 |
"degree of freedom more than once on the right hand side. This is " |
| 117 |
"not allowed.")); |
| 118 |
} |
| 119 |
} |
| 120 |
|
| 121 |
|
| 122 |
// The following can happen when we compute with distributed meshes and dof |
| 123 |
// handlers and we constrain a degree of freedom whose number we don't have |
| 124 |
// locally. if we don't abort here the program will try to allocate several |
| 125 |
// terabytes of memory to resize the various arrays below :-) |
| 126 |
Assert(constrained_dof != numbers::invalid_size_type, ExcInternalError()); |
| 127 |
|
| 128 |
// if necessary enlarge vector of existing entries for cache |
| 129 |
const size_type line_index = calculate_line_index(constrained_dof); |
| 130 |
if (line_index >= lines_cache.size()) |
| 131 |
lines_cache.resize(std::max(2 * static_cast<size_type>(lines_cache.size()), |
| 132 |
line_index + 1), |
| 133 |
numbers::invalid_size_type); |
| 134 |
|
| 135 |
// Push a new line to the end of the list and fill it with the |
| 136 |
// provided information: |
| 137 |
ConstraintLine &constraint = lines.emplace_back(); |
| 138 |
constraint.index = constrained_dof; |
| 139 |
constraint.entries.reserve(dependencies.size()); |
| 140 |
for (const auto &[column, weight] : dependencies) |
| 141 |
{ |
| 142 |
Assert(column != constrained_dof, |
| 143 |
ExcMessage("You cannot constrain a degree of freedom against " |
| 144 |
"itself.")); |
| 145 |
constraint.entries.emplace_back(column, weight); |
| 146 |
} |
| 147 |
constraint.inhomogeneity = inhomogeneity; |
| 148 |
|
| 149 |
// Record the new constraint in the cache: |
| 150 |
Assert(lines_cache[line_index] == numbers::invalid_size_type, |
| 151 |
ExcInternalError()); |
| 152 |
lines_cache[line_index] = lines.size() - 1; |
| 153 |
} |
| 154 |
|
| 155 |
|
| 156 |
|
| 157 |
template <typename number> |
| 158 |
void |
| 159 |
AffineConstraints<number>::constrain_dof_to_zero( |
| 160 |
const size_type constrained_dof) |
| 161 |
{ |
| 162 |
Assert(sorted == false, ExcMatrixIsClosed()); |
| 163 |
|
| 164 |
// The following can happen when we compute with distributed meshes and dof |
| 165 |
// handlers and we constrain a degree of freedom whose number we don't have |
| 166 |
// locally. if we don't abort here the program will try to allocate several |
| 167 |
// terabytes of memory to resize the various arrays below :-) |
| 168 |
Assert(constrained_dof != numbers::invalid_size_type, ExcInternalError()); |
| 169 |
|
| 170 |
// if necessary enlarge vector of existing entries for cache |
| 171 |
const size_type line_index = calculate_line_index(constrained_dof); |
| 172 |
if (line_index >= lines_cache.size()) |
| 173 |
lines_cache.resize(std::max(2 * static_cast<size_type>(lines_cache.size()), |
| 174 |
line_index + 1), |
| 175 |
numbers::invalid_size_type); |
| 176 |
|
| 177 |
// Let's check whether the DoF is already constrained. This is only allowed |
| 178 |
// if it had previously been constrained to zero, and only then. |
| 179 |
if (lines_cache[line_index] != numbers::invalid_size_type) |
| 180 |
{ |
| 181 |
Assert(lines[lines_cache[line_index]].entries.empty() && |
| 182 |
(lines[lines_cache[line_index]].inhomogeneity == number(0.)), |
| 183 |
ExcMessage("You cannot constrain a degree of freedom " |
| 184 |
"to zero that is is already constrained to " |
| 185 |
"something else.")); |
| 186 |
} |
| 187 |
else |
| 188 |
{ |
| 189 |
// Push a new line to the end of the list and fill it with the |
| 190 |
// provided information: |
| 191 |
ConstraintLine &constraint = lines.emplace_back(); |
| 192 |
constraint.index = constrained_dof; |
| 193 |
constraint.inhomogeneity = 0.; |
| 194 |
|
| 195 |
// Record the new constraint in the cache: |
| 196 |
Assert(lines_cache[line_index] == numbers::invalid_size_type, |
| 197 |
ExcInternalError()); |
| 198 |
lines_cache[line_index] = lines.size() - 1; |
| 199 |
} |
| 200 |
} |
| 201 |
|
| 202 |
|
| 203 |
|
| 204 |
template <typename number> |
| 205 |
typename AffineConstraints<number>::LineRange |
| 206 |
AffineConstraints<number>::get_lines() const |
| 207 |
{ |
| 208 |
return boost::make_iterator_range(lines.begin(), lines.end()); |
| 209 |
} |
| 210 |
|
| 211 |
|
| 212 |
|
| 213 |
template <typename number> |
| 214 |
bool |
| 215 |
AffineConstraints<number>::is_consistent_in_parallel( |
| 216 |
const std::vector<IndexSet> &locally_owned_dofs, |
| 217 |
const IndexSet &locally_active_dofs, |
| 218 |
const MPI_Comm mpi_communicator, |
| 219 |
const bool verbose) const |
| 220 |
{ |
| 221 |
// Helper to return a ConstraintLine object that belongs to row @p row. |
| 222 |
// If @p row is not constrained or not stored locally, return an empty |
| 223 |
// constraint object that would correspond to a zero constraints |
| 224 |
auto get_line = [&](const size_type line_n) -> ConstraintLine { |
| 225 |
const size_type line_index = calculate_line_index(line_n); |
| 226 |
if (line_index >= lines_cache.size() || |
| 227 |
lines_cache[line_index] == numbers::invalid_size_type) |
| 228 |
{ |
| 229 |
const ConstraintLine empty = {line_n, {}, 0.0}; |
| 230 |
return empty; |
| 231 |
} |
| 232 |
else |
| 233 |
return lines[lines_cache[line_index]]; |
| 234 |
}; |
| 235 |
|
| 236 |
// identify non-owned rows and send to owner: |
| 237 |
std::map<unsigned int, std::vector<ConstraintLine>> to_send; |
| 238 |
|
| 239 |
const unsigned int myid = |
| 240 |
dealii::Utilities::MPI::this_mpi_process(mpi_communicator); |
| 241 |
const unsigned int nproc = |
| 242 |
dealii::Utilities::MPI::n_mpi_processes(mpi_communicator); |
| 243 |
|
| 244 |
// We will send all locally active dofs that are not locally owned for |
| 245 |
// checking. Note that we allow constraints to differ on locally_relevant (and |
| 246 |
// not active) DoFs. |
| 247 |
// NOLINTNEXTLINE(performance-unnecessary-copy-initialization) |
| 248 |
IndexSet non_owned = locally_active_dofs; |
| 249 |
non_owned.subtract_set(locally_owned_dofs[myid]); |
| 250 |
for (unsigned int owner = 0; owner < nproc; ++owner) |
| 251 |
{ |
| 252 |
// find all lines to send to @p owner |
| 253 |
const IndexSet indices_to_send = non_owned & locally_owned_dofs[owner]; |
| 254 |
for (const auto row_idx : indices_to_send) |
| 255 |
{ |
| 256 |
to_send[owner].emplace_back(get_line(row_idx)); |
| 257 |
} |
| 258 |
} |
| 259 |
|
| 260 |
const std::map<unsigned int, std::vector<ConstraintLine>> received = |
| 261 |
Utilities::MPI::some_to_some(mpi_communicator, to_send); |
| 262 |
|
| 263 |
unsigned int inconsistent = 0; |
| 264 |
|
| 265 |
// from each processor: |
| 266 |
for (const auto &[sender, constraints] : received) |
| 267 |
{ |
| 268 |
// for each incoming line: |
| 269 |
for (const ConstraintLine &constraint : constraints) |
| 270 |
{ |
| 271 |
const ConstraintLine reference = get_line(constraint.index); |
| 272 |
|
| 273 |
if (constraint.inhomogeneity != reference.inhomogeneity) |
| 274 |
{ |
| 275 |
++inconsistent; |
| 276 |
|
| 277 |
if (verbose) |
| 278 |
std::cout << "Proc " << myid << " got line " << constraint.index |
| 279 |
<< " from " << sender << " inhomogeneity " |
| 280 |
<< constraint.inhomogeneity |
| 281 |
<< " != " << reference.inhomogeneity << std::endl; |
| 282 |
} |
| 283 |
else if (constraint.entries != reference.entries) |
| 284 |
{ |
| 285 |
++inconsistent; |
| 286 |
if (verbose) |
| 287 |
std::cout << "Proc " << myid << " got line " << constraint.index |
| 288 |
<< " from " << sender << " with wrong values!" |
| 289 |
<< std::endl; |
| 290 |
} |
| 291 |
} |
| 292 |
} |
| 293 |
|
| 294 |
const unsigned int total = |
| 295 |
Utilities::MPI::sum(inconsistent, mpi_communicator); |
| 296 |
if (verbose && total > 0 && myid == 0) |
| 297 |
std::cout << total << " inconsistent lines discovered!" << std::endl; |
| 298 |
return total == 0; |
| 299 |
} |
| 300 |
|
| 301 |
|
| 302 |
namespace internal |
| 303 |
{ |
| 304 |
// A helper function that sorts and normalizes the constraints |
| 305 |
// provided through the function's argument: |
| 306 |
template <typename number> |
| 307 |
void |
| 308 |
sort_and_make_unique( |
| 309 |
std::vector<typename dealii::AffineConstraints<number>::ConstraintLine> |
| 310 |
&constraints) |
| 311 |
{ |
| 312 |
using ConstraintLine = |
| 313 |
typename dealii::AffineConstraints<number>::ConstraintLine; |
| 314 |
|
| 315 |
if (constraints.empty()) |
| 316 |
return; |
| 317 |
|
| 318 |
// First sort the array of constraints by their index: |
| 319 |
std::sort(constraints.begin(), |
| 320 |
constraints.end(), |
| 321 |
[](const ConstraintLine &l1, const ConstraintLine &l2) { |
| 322 |
return l1.index < l2.index; |
| 323 |
}); |
| 324 |
|
| 325 |
// It is possible that two processes have computed constraints |
| 326 |
// for the same DoF differently (for example, in parallel because |
| 327 |
// they have visited different faces of the same cell because |
| 328 |
// on different processes, different neighbor cells are artificial. |
| 329 |
// |
| 330 |
// This is ok, as long as after resolution of all chains of |
| 331 |
// constraints we end up with the same constraint. But at this |
| 332 |
// point, we just don't know yet. We deal with this by simply |
| 333 |
// dropping constraints for DoFs for which there is a previous |
| 334 |
// constraint in the list: |
| 335 |
constraints.erase(std::unique(constraints.begin(), |
| 336 |
constraints.end(), |
| 337 |
[](const ConstraintLine &a, |
| 338 |
const ConstraintLine &b) { |
| 339 |
return (a.index == b.index); |
| 340 |
}), |
| 341 |
constraints.end()); |
| 342 |
|
| 343 |
// For those constraints that survive, sort the right hand side arrays: |
| 344 |
for (ConstraintLine &entry : constraints) |
| 345 |
std::sort(entry.entries.begin(), |
| 346 |
entry.entries.end(), |
| 347 |
[](const auto &l1, const auto &l2) { |
| 348 |
return l1.first < l2.first; |
| 349 |
}); |
| 350 |
} |
| 351 |
|
| 352 |
template <typename number> |
| 353 |
void |
| 354 |
compute_locally_relevant_constraints( |
| 355 |
const dealii::AffineConstraints<number> &constraints_in, |
| 356 |
const IndexSet &locally_owned_dofs, |
| 357 |
const IndexSet &locally_relevant_dofs, |
| 358 |
const MPI_Comm mpi_communicator, |
| 359 |
const bool first_run, |
| 360 |
std::vector<typename dealii::AffineConstraints<number>::ConstraintLine> |
| 361 |
&locally_relevant_constraints) |
| 362 |
{ |
| 363 |
// The result vector filled step by step. |
| 364 |
using ConstraintLine = |
| 365 |
typename dealii::AffineConstraints<number>::ConstraintLine; |
| 366 |
|
| 367 |
#ifndef DEAL_II_WITH_MPI |
| 368 |
AssertThrow(false, ExcNotImplemented()); // one should not come here |
| 369 |
(void)constraints_in; |
| 370 |
(void)locally_owned_dofs; |
| 371 |
(void)locally_relevant_dofs; |
| 372 |
(void)mpi_communicator; |
| 373 |
(void)first_run; |
| 374 |
(void)locally_relevant_constraints; |
| 375 |
#else |
| 376 |
|
| 377 |
[[maybe_unused]] const unsigned int my_rank = |
| 378 |
Utilities::MPI::this_mpi_process(mpi_communicator); |
| 379 |
|
| 380 |
// step 0: Collect the indices of constrained DoFs we know of: |
| 381 |
IndexSet my_constraint_indices(locally_owned_dofs.size()); |
| 382 |
{ |
| 383 |
std::vector<types::global_dof_index> indices; |
| 384 |
indices.reserve(constraints_in.n_constraints()); |
| 385 |
for (const ConstraintLine &line : constraints_in.get_lines()) |
| 386 |
if (locally_owned_dofs.is_element(line.index) == false) |
| 387 |
indices.push_back(line.index); |
| 388 |
std::sort(indices.begin(), indices.end()); |
| 389 |
my_constraint_indices.add_indices(indices.begin(), indices.end()); |
| 390 |
} |
| 391 |
|
| 392 |
// step 1: Identify the owners of DoFs we know to be constrained but do |
| 393 |
// not own. |
| 394 |
const auto [owners_of_my_constraints, constrained_indices_by_ranks] = |
| 395 |
Utilities::MPI::compute_index_owner_and_requesters(locally_owned_dofs, |
| 396 |
my_constraint_indices, |
| 397 |
mpi_communicator); |
| 398 |
|
| 399 |
// step 2: Collect all locally owned constraints into a data structure |
| 400 |
// that we can send to other processes that want to know |
| 401 |
// about them. |
| 402 |
if (first_run) |
| 403 |
{ |
| 404 |
std::map<unsigned int, std::vector<ConstraintLine>> send_data; |
| 405 |
|
| 406 |
// For each constraint we know of but owned by another process, create a |
| 407 |
// copy of the constraint and add it to the list of things to send to |
| 408 |
// other processes: |
| 409 |
for (unsigned int constraint_index = 0; |
| 410 |
constraint_index < owners_of_my_constraints.size(); |
| 411 |
++constraint_index) |
| 412 |
{ |
| 413 |
ConstraintLine entry; |
| 414 |
|
| 415 |
const types::global_dof_index index = |
| 416 |
my_constraint_indices.nth_index_in_set(constraint_index); |
| 417 |
|
| 418 |
entry.index = index; |
| 419 |
if (const std::vector<std::pair<types::global_dof_index, number>> |
| 420 |
*constraints = constraints_in.get_constraint_entries(index)) |
| 421 |
entry.entries = *constraints; |
| 422 |
entry.inhomogeneity = constraints_in.get_inhomogeneity(index); |
| 423 |
|
| 424 |
|
| 425 |
Assert(owners_of_my_constraints[constraint_index] != my_rank, |
| 426 |
ExcInternalError()); |
| 427 |
send_data[owners_of_my_constraints[constraint_index]].push_back( |
| 428 |
entry); |
| 429 |
} |
| 430 |
|
| 431 |
// Now exchange this data between processes: |
| 432 |
const std::map<unsigned int, std::vector<ConstraintLine>> |
| 433 |
received_data = |
| 434 |
Utilities::MPI::some_to_some(mpi_communicator, send_data); |
| 435 |
|
| 436 |
// Finally join things with the constraints we know about and own |
| 437 |
// ourselves, collate everything we received into one array, sort, |
| 438 |
// and make it unique: |
| 439 |
for (const ConstraintLine &line : constraints_in.get_lines()) |
| 440 |
if (locally_owned_dofs.is_element(line.index)) |
| 441 |
locally_relevant_constraints.push_back(line); |
| 442 |
for (const auto &[rank, constraints] : received_data) |
| 443 |
locally_relevant_constraints.insert( |
| 444 |
locally_relevant_constraints.end(), |
| 445 |
constraints.begin(), |
| 446 |
constraints.end()); |
| 447 |
|
| 448 |
sort_and_make_unique<number>(locally_relevant_constraints); |
| 449 |
} |
| 450 |
|
| 451 |
// step 3: communicate constraints so that each process knows how the |
| 452 |
// locally relevant dofs are constrained |
| 453 |
{ |
| 454 |
// First determine the owners of those locally relevant dofs we |
| 455 |
// don't own ourselves. |
| 456 |
IndexSet locally_relevant_dofs_non_local = locally_relevant_dofs; |
| 457 |
locally_relevant_dofs_non_local.subtract_set(locally_owned_dofs); |
| 458 |
|
| 459 |
// We are, however, not actually interested in who owns a specific |
| 460 |
// DoF we have among our locally-stored-but-not-locally-owned constraints. |
| 461 |
// Rather, we want to know who requested information about our own |
| 462 |
// locally-owned DoFs. That's because we will want to send our own |
| 463 |
// locally-owned constraints to those processes for which these are |
| 464 |
// in their own locally-relevant index sets. |
| 465 |
const auto [_, requesters_and_requested_constraints] = |
| 466 |
Utilities::MPI::compute_index_owner_and_requesters( |
| 467 |
locally_owned_dofs, |
| 468 |
locally_relevant_dofs_non_local, |
| 469 |
mpi_communicator); |
| 470 |
|
| 471 |
std::map<unsigned int, std::vector<ConstraintLine>> send_data; |
| 472 |
for (const auto &[destination, requested_indices] : |
| 473 |
requesters_and_requested_constraints) |
| 474 |
{ |
| 475 |
Assert(destination != my_rank, ExcInternalError()); |
| 476 |
Assert(requested_indices.is_subset_of(locally_owned_dofs), |
| 477 |
ExcInternalError()); |
| 478 |
|
| 479 |
std::vector<ConstraintLine> data; |
| 480 |
|
| 481 |
IndexSet::ElementIterator i = requested_indices.begin(); |
| 482 |
|
| 483 |
typename std::vector<ConstraintLine>::const_iterator j = |
| 484 |
std::lower_bound(locally_relevant_constraints.begin(), |
| 485 |
locally_relevant_constraints.end(), |
| 486 |
*i, |
| 487 |
[&](const auto &a, const auto &b) { |
| 488 |
return a.index < b; |
| 489 |
}); |
| 490 |
|
| 491 |
while ((i != requested_indices.end()) && |
| 492 |
(j != locally_relevant_constraints.end())) |
| 493 |
{ |
| 494 |
if (*i == j->index) |
| 495 |
{ |
| 496 |
data.push_back(*j); |
| 497 |
++i; |
| 498 |
++j; |
| 499 |
} |
| 500 |
else if (*i < j->index) |
| 501 |
++i; |
| 502 |
else |
| 503 |
++j; |
| 504 |
} |
| 505 |
|
| 506 |
send_data[destination] = std::move(data); |
| 507 |
} |
| 508 |
|
| 509 |
const std::map<unsigned int, std::vector<ConstraintLine>> received_data = |
| 510 |
Utilities::MPI::some_to_some(mpi_communicator, send_data); |
| 511 |
|
| 512 |
// Finally collate everything into one array, sort, and make it unique: |
| 513 |
for (const auto &[sender, constraints] : received_data) |
| 514 |
locally_relevant_constraints.insert(locally_relevant_constraints.end(), |
| 515 |
constraints.begin(), |
| 516 |
constraints.end()); |
| 517 |
|
| 518 |
sort_and_make_unique<number>(locally_relevant_constraints); |
| 519 |
} |
| 520 |
|
| 521 |
#endif |
| 522 |
} |
| 523 |
} // namespace internal |
| 524 |
|
| 525 |
|
| 526 |
|
| 527 |
template <typename number> |
| 528 |
void |
| 529 |
AffineConstraints<number>::make_consistent_in_parallel( |
| 530 |
const IndexSet &locally_owned_dofs, |
| 531 |
const IndexSet &constraints_to_make_consistent_, |
| 532 |
const MPI_Comm mpi_communicator) |
| 533 |
{ |
| 534 |
// We need to resolve chains of constraints in this function, and that |
| 535 |
// is exactly what close() does. So call this function at the beginning |
| 536 |
// to make sure we start working with constraints that are already |
| 537 |
// resolved to the degree possible before exchanging information. |
| 538 |
// We do this here so that it is also done if we choose the early |
| 539 |
// exit in the following statement. |
| 540 |
if (sorted == false) |
| 541 |
close(); |
| 542 |
|
| 543 |
if (Utilities::MPI::n_mpi_processes(mpi_communicator) == 1) |
| 544 |
return; // Nothing to do, since serial. |
| 545 |
|
| 546 |
Assert(this->local_lines.size() > 0, |
| 547 |
ExcMessage( |
| 548 |
"This functionality requires that the AffineConstraints object " |
| 549 |
"knows for which degrees of freedom it can store constraints. " |
| 550 |
"Please initialize this object with the corresponding index sets.")); |
| 551 |
|
| 552 |
// Container for indices that are constrained or that other indices are |
| 553 |
// constrained against. Generously reserve memory for it, assuming the worst |
| 554 |
// case that each constraint line involves two distinct DoF indices. |
| 555 |
std::vector<types::global_dof_index> constrained_indices; |
| 556 |
constrained_indices.reserve(2 * n_constraints()); |
| 557 |
|
| 558 |
// This IndexSet keeps track of the locally stored constraints on this |
| 559 |
// AffineConstraints object: If we have received constrained indices that we |
| 560 |
// don't know of, we need to expand our playing field. |
| 561 |
IndexSet locally_stored_constraints; |
| 562 |
|
| 563 |
// This IndexSet contains those DoFs about which we want to know all |
| 564 |
// constraints. We will receive constraints for these DoFs against other DoFs, |
| 565 |
// which might be constrained themselves. To successfully resolve all chains |
| 566 |
// of constraints, we need to know the constraints of all these DoFs, and we |
| 567 |
// keep track of the relevant DoFs here. |
| 568 |
IndexSet constraints_to_make_consistent = constraints_to_make_consistent_; |
| 569 |
|
| 570 |
// This IndexSet stores DoFs from the previous iteration. If the old and new |
| 571 |
// index sets match, we have converged. |
| 572 |
IndexSet constraints_made_consistent; |
| 573 |
|
| 574 |
std::vector<typename dealii::AffineConstraints<number>::ConstraintLine> |
| 575 |
imported_constraints; |
| 576 |
|
| 577 |
const unsigned int max_iterations = 10; |
| 578 |
unsigned int iteration_count = 0; |
| 579 |
for (; iteration_count < max_iterations; ++iteration_count) |
| 580 |
{ |
| 581 |
// 1) Get all locally relevant constraints we need to know about: |
| 582 |
internal::compute_locally_relevant_constraints( |
| 583 |
*this, |
| 584 |
locally_owned_dofs, |
| 585 |
constraints_to_make_consistent, |
| 586 |
mpi_communicator, |
| 587 |
iteration_count == 0, |
| 588 |
imported_constraints); |
| 589 |
|
| 590 |
// 2) Add untracked DoFs to the index sets. |
| 591 |
constrained_indices.clear(); |
| 592 |
for (const auto &line : imported_constraints) |
| 593 |
{ |
| 594 |
constrained_indices.push_back(line.index); |
| 595 |
for (const auto &entry : line.entries) |
| 596 |
constrained_indices.push_back(entry.first); |
| 597 |
} |
| 598 |
std::sort(constrained_indices.begin(), constrained_indices.end()); |
| 599 |
|
| 600 |
locally_stored_constraints = this->local_lines; |
| 601 |
locally_stored_constraints.add_indices(constrained_indices.begin(), |
| 602 |
constrained_indices.end()); |
| 603 |
|
| 604 |
constraints_made_consistent = constraints_to_make_consistent; |
| 605 |
constraints_to_make_consistent.add_indices(constrained_indices.begin(), |
| 606 |
constrained_indices.end()); |
| 607 |
|
| 608 |
// 3) Clear and refill this constraint matrix. Also resolve chains: |
| 609 |
this->reinit(locally_owned_dofs, locally_stored_constraints); |
| 610 |
for (const auto &line : imported_constraints) |
| 611 |
this->add_constraint(line.index, line.entries, line.inhomogeneity); |
| 612 |
this->close(); |
| 613 |
|
| 614 |
// 4) Stop loop if converged. |
| 615 |
const bool constraints_converged = |
| 616 |
(Utilities::MPI::min( |
| 617 |
(constraints_to_make_consistent == constraints_made_consistent ? 1 : |
| 618 |
0), |
| 619 |
mpi_communicator) == 1); |
| 620 |
if (constraints_converged) |
| 621 |
break; |
| 622 |
} |
| 623 |
|
| 624 |
AssertThrow(iteration_count < max_iterations, |
| 625 |
ExcMessage( |
| 626 |
"make_consistent_in_parallel() did not converge after " + |
| 627 |
Utilities::to_string(max_iterations) + " iterations.")); |
| 628 |
|
| 629 |
Assert(this->is_consistent_in_parallel( |
| 630 |
Utilities::MPI::all_gather(mpi_communicator, locally_owned_dofs), |
| 631 |
constraints_to_make_consistent, |
| 632 |
mpi_communicator), |
| 633 |
ExcInternalError()); |
| 634 |
} |
| 635 |
|
| 636 |
|
| 637 |
|
| 638 |
template <typename number> |
| 639 |
void |
| 640 |
AffineConstraints<number>::add_lines(const std::set<size_type> &lines) |
| 641 |
{ |
| 642 |
for (const size_type &line : lines) |
| 643 |
add_line(line); |
| 644 |
} |
| 645 |
|
| 646 |
|
| 647 |
|
| 648 |
template <typename number> |
| 649 |
void |
| 650 |
AffineConstraints<number>::add_lines(const std::vector<bool> &lines) |
| 651 |
{ |
| 652 |
for (size_type i = 0; i < lines.size(); ++i) |
| 653 |
if (lines[i] == true) |
| 654 |
add_line(i); |
| 655 |
} |
| 656 |
|
| 657 |
|
| 658 |
|
| 659 |
template <typename number> |
| 660 |
void |
| 661 |
AffineConstraints<number>::add_lines(const IndexSet &lines) |
| 662 |
{ |
| 663 |
for (size_type i = 0; i < lines.n_elements(); ++i) |
| 664 |
add_line(lines.nth_index_in_set(i)); |
| 665 |
} |
| 666 |
|
| 667 |
|
| 668 |
|
| 669 |
template <typename number> |
| 670 |
void |
| 671 |
AffineConstraints<number>::add_entries( |
| 672 |
const size_type constrained_dof_index, |
| 673 |
const std::vector<std::pair<size_type, number>> &col_weight_pairs) |
| 674 |
{ |
| 675 |
Assert(sorted == false, ExcMatrixIsClosed()); |
| 676 |
Assert(is_constrained(constrained_dof_index), |
| 677 |
ExcLineInexistent(constrained_dof_index)); |
| 678 |
|
| 679 |
ConstraintLine &line = |
| 680 |
lines[lines_cache[calculate_line_index(constrained_dof_index)]]; |
| 681 |
Assert(line.index == constrained_dof_index, ExcInternalError()); |
| 682 |
|
| 683 |
// if in debug mode, check whether an entry for this column already |
| 684 |
// exists and if its the same as the one entered at present |
| 685 |
// |
| 686 |
// in any case: skip this entry if an entry for this column already |
| 687 |
// exists, since we don't want to enter it twice |
| 688 |
line.entries.reserve(line.entries.size() + col_weight_pairs.size()); |
| 689 |
for (const std::pair<size_type, number> &col_weight_pair : col_weight_pairs) |
| 690 |
{ |
| 691 |
Assert(constrained_dof_index != col_weight_pair.first, |
| 692 |
ExcMessage("Can't constrain a degree of freedom to itself")); |
| 693 |
bool entry_exists = false; |
| 694 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 695 |
if (entry.first == col_weight_pair.first) |
| 696 |
{ |
| 697 |
// entry exists, break innermost loop |
| 698 |
Assert(entry.second == col_weight_pair.second, |
| 699 |
ExcEntryAlreadyExists(constrained_dof_index, |
| 700 |
col_weight_pair.first, |
| 701 |
entry.second, |
| 702 |
col_weight_pair.second)); |
| 703 |
entry_exists = true; |
| 704 |
break; |
| 705 |
} |
| 706 |
|
| 707 |
if (entry_exists == false) |
| 708 |
line.entries.push_back(col_weight_pair); |
| 709 |
} |
| 710 |
} |
| 711 |
|
| 712 |
|
| 713 |
|
| 714 |
template <typename number> |
| 715 |
void |
| 716 |
AffineConstraints<number>::add_selected_constraints( |
| 717 |
const AffineConstraints &constraints, |
| 718 |
const IndexSet &filter) |
| 719 |
{ |
| 720 |
if (constraints.n_constraints() == 0) |
| 721 |
return; |
| 722 |
|
| 723 |
Assert(filter.size() > constraints.lines.back().index, |
| 724 |
ExcMessage( |
| 725 |
"The filter must be larger than the given constraints object.")); |
| 726 |
for (const ConstraintLine &line : constraints.lines) |
| 727 |
if (filter.is_element(line.index)) |
| 728 |
{ |
| 729 |
const size_type row = filter.index_within_set(line.index); |
| 730 |
add_line(row); |
| 731 |
set_inhomogeneity(row, line.inhomogeneity); |
| 732 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 733 |
if (filter.is_element(entry.first)) |
| 734 |
add_entry(row, filter.index_within_set(entry.first), entry.second); |
| 735 |
} |
| 736 |
} |
| 737 |
|
| 738 |
|
| 739 |
|
| 740 |
template <typename number> |
| 741 |
void |
| 742 |
AffineConstraints<number>::close() |
| 743 |
{ |
| 744 |
if (sorted == true) |
| 745 |
return; |
| 746 |
|
| 747 |
// sort the lines |
| 748 |
std::sort(lines.begin(), |
| 749 |
lines.end(), |
| 750 |
[](const ConstraintLine &l1, const ConstraintLine &l2) { |
| 751 |
return l1.index < l2.index; |
| 752 |
}); |
| 753 |
|
| 754 |
// update list of pointers and give the vector a sharp size since we |
| 755 |
// won't modify the size any more after this point. |
| 756 |
{ |
| 757 |
std::vector<size_type> new_lines(lines_cache.size(), |
| 758 |
numbers::invalid_size_type); |
| 759 |
size_type counter = 0; |
| 760 |
for (const ConstraintLine &line : lines) |
| 761 |
{ |
| 762 |
new_lines[calculate_line_index(line.index)] = counter; |
| 763 |
++counter; |
| 764 |
} |
| 765 |
lines_cache = std::move(new_lines); |
| 766 |
} |
| 767 |
|
| 768 |
// in debug mode: check whether we really set the pointers correctly. |
| 769 |
for (size_type i = 0; i < lines_cache.size(); ++i) |
| 770 |
if (lines_cache[i] != numbers::invalid_size_type) |
| 771 |
Assert(i == calculate_line_index(lines[lines_cache[i]].index), |
| 772 |
ExcInternalError()); |
| 773 |
|
| 774 |
// The second part is that we need to work on the individual lines. |
| 775 |
// Let us start by stripping zero entries. That would mean that in the linear |
| 776 |
// constraint for a node, x_i = ax_1 + bx_2 + ..., another node times 0 |
| 777 |
// appears. obviously, 0*something can be omitted. |
| 778 |
// |
| 779 |
// This can be done in parallel: |
| 780 |
parallel::apply_to_subranges( |
| 781 |
lines.begin(), |
| 782 |
lines.end(), |
| 783 |
[](const typename std::vector<ConstraintLine>::iterator begin, |
| 784 |
const typename std::vector<ConstraintLine>::iterator end) { |
| 785 |
for (ConstraintLine &line : boost::iterator_range< |
| 786 |
typename std::vector<ConstraintLine>::iterator>(begin, end)) |
| 787 |
line.entries.erase( |
| 788 |
std::remove_if(line.entries.begin(), |
| 789 |
line.entries.end(), |
| 790 |
[](const std::pair<size_type, number> &p) { |
| 791 |
return p.second == number(0.); |
| 792 |
}), |
| 793 |
line.entries.end()); |
| 794 |
}, |
| 795 |
/* grainsize = */ 100); |
| 796 |
|
| 797 |
|
| 798 |
|
| 799 |
// In debug mode we are computing an estimate for the maximum number |
| 800 |
// of constraints so that we can bail out if there is a cycle in the |
| 801 |
// constraints (which is easier than searching for cycles in the graph). |
| 802 |
// |
| 803 |
// Let us figure out the largest dof index. This is an upper bound for the |
| 804 |
// number of constraints because it is an approximation for the number of dofs |
| 805 |
// in our system. |
| 806 |
[[maybe_unused]] size_type largest_idx = 0; |
| 807 |
if constexpr (running_in_debug_mode()) |
| 808 |
{ |
| 809 |
for (const ConstraintLine &line : lines) |
| 810 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 811 |
largest_idx = std::max(largest_idx, entry.first); |
| 812 |
} |
| 813 |
|
| 814 |
// replace references to dofs that are themselves constrained. note that |
| 815 |
// because we may replace references to other dofs that may themselves be |
| 816 |
// constrained to third ones, we have to iterate over all this until we |
| 817 |
// replace no chains of constraints any more |
| 818 |
// |
| 819 |
// the iteration replaces references to constrained degrees of freedom by |
| 820 |
// second-order references. for example if x3=x0/2+x2/2 and x2=x0/2+x1/2, |
| 821 |
// then the new list will be x3=x0/2+x0/4+x1/4. note that x0 appear |
| 822 |
// twice. we will throw this duplicate out in the following step, where |
| 823 |
// we sort the list so that throwing out duplicates becomes much more |
| 824 |
// efficient. also, we have to do it only once, rather than in each |
| 825 |
// iteration |
| 826 |
[[maybe_unused]] size_type iteration = 0; |
| 827 |
bool chained_constraint_replaced = false; |
| 828 |
std::vector<bool> line_finalized(lines.size(), false); |
| 829 |
do |
| 830 |
{ |
| 831 |
chained_constraint_replaced = false; |
| 832 |
for (unsigned int line_index = 0; line_index < lines.size(); ++line_index) |
| 833 |
if (line_finalized[line_index] == false) |
| 834 |
{ |
| 835 |
ConstraintLine &line = lines[line_index]; |
| 836 |
|
| 837 |
// we need to keep track of how many replacements we do in this |
| 838 |
// line, because we can end up in a cycle A->B->C->A without the |
| 839 |
// number of entries growing. |
| 840 |
[[maybe_unused]] size_type n_replacements = 0; |
| 841 |
|
| 842 |
// loop over all entries of this line (including ones that we |
| 843 |
// have appended in this go around) and see whether they are |
| 844 |
// further constrained. ignore elements that we don't store on |
| 845 |
// the current processor. |
| 846 |
const unsigned int n_original_entries = line.entries.size(); |
| 847 |
bool has_sub_constraints = false; |
| 848 |
const size_type lines_cache_size = lines_cache.size(); |
| 849 |
for (unsigned int entry = 0; entry < n_original_entries; ++entry) |
| 850 |
{ |
| 851 |
const size_type dof_index = |
| 852 |
calculate_line_index(line.entries[entry].first); |
| 853 |
|
| 854 |
if (dof_index < lines_cache_size && |
| 855 |
lines_cache[dof_index] != numbers::invalid_size_type) |
| 856 |
{ |
| 857 |
// Now we found an index that is further constrained, so |
| 858 |
// we need to resolve the entries. We may replace some of |
| 859 |
// them, and we may add some, but we only walk through the |
| 860 |
// original entries (and also don't touch the |
| 861 |
// replacements) even though in principle we could also do |
| 862 |
// the ones we have added or replaced. There is no harm in |
| 863 |
// doing so: we will simply treat those in the next round |
| 864 |
// around of the outer iteration. |
| 865 |
has_sub_constraints = true; |
| 866 |
|
| 867 |
const number weight = line.entries[entry].second; |
| 868 |
Assert(line.entries[entry].first != line.index, |
| 869 |
ExcMessage("Cycle in constraints detected!")); |
| 870 |
|
| 871 |
const ConstraintLine &constrained_line = |
| 872 |
lines[lines_cache[dof_index]]; |
| 873 |
Assert(constrained_line.index == line.entries[entry].first, |
| 874 |
ExcInternalError()); |
| 875 |
|
| 876 |
// now we have to replace an entry by its expansion. we do |
| 877 |
// that by overwriting the entry by the first entry of the |
| 878 |
// expansion and adding the remaining ones to the end, |
| 879 |
// where we will later process them once more |
| 880 |
// |
| 881 |
// we can of course only do that if the DoF that we are |
| 882 |
// currently handling is constrained by a linear combination |
| 883 |
// of other dofs: |
| 884 |
if (constrained_line.entries.size() > 0) |
| 885 |
{ |
| 886 |
for (size_type i = 0; |
| 887 |
i < constrained_line.entries.size(); |
| 888 |
++i) |
| 889 |
Assert(line.entries[entry].first != |
| 890 |
constrained_line.entries[i].first, |
| 891 |
ExcMessage("Cycle in constraints detected!")); |
| 892 |
|
| 893 |
// replace first entry, then tack the rest to the end |
| 894 |
// of the list |
| 895 |
line.entries[entry] = std::pair<size_type, number>( |
| 896 |
constrained_line.entries[0].first, |
| 897 |
constrained_line.entries[0].second * weight); |
| 898 |
|
| 899 |
for (size_type i = 1; |
| 900 |
i < constrained_line.entries.size(); |
| 901 |
++i) |
| 902 |
line.entries.emplace_back( |
| 903 |
constrained_line.entries[i].first, |
| 904 |
constrained_line.entries[i].second * weight); |
| 905 |
|
| 906 |
if constexpr (library_build_mode == |
| 907 |
LibraryBuildMode::debug) |
| 908 |
{ |
| 909 |
// keep track of how many entries we replace in this |
| 910 |
// line. If we do more than there are constraints or |
| 911 |
// dofs in our system, we must have a cycle. |
| 912 |
++n_replacements; |
| 913 |
Assert(n_replacements / 2 < largest_idx, |
| 914 |
ExcMessage( |
| 915 |
"Cycle in constraints detected!")); |
| 916 |
} |
| 917 |
} |
| 918 |
else |
| 919 |
// the DoF that we encountered is not constrained by a |
| 920 |
// linear combination of other dofs but is equal to just |
| 921 |
// the inhomogeneity (i.e. its chain of entries is |
| 922 |
// empty). in that case, we can't just overwrite the |
| 923 |
// current entry, but we have to actually eliminate |
| 924 |
// it. we do not want to change the loop length above we |
| 925 |
// do so by setting the 'first' entry to |
| 926 |
// invalid_size_type here to finally remove entries in a |
| 927 |
// second loop |
| 928 |
{ |
| 929 |
line.entries[entry].first = numbers::invalid_size_type; |
| 930 |
} |
| 931 |
|
| 932 |
line.inhomogeneity += |
| 933 |
constrained_line.inhomogeneity * weight; |
| 934 |
} |
| 935 |
} |
| 936 |
|
| 937 |
// If none of the entries in the current line refer to DoFs that are |
| 938 |
// themselves constrained, then we can move on: |
| 939 |
if (has_sub_constraints == false) |
| 940 |
line_finalized[line_index] = true; |
| 941 |
else |
| 942 |
{ |
| 943 |
chained_constraint_replaced = true; |
| 944 |
|
| 945 |
// Now delete the elements we have marked for deletion. |
| 946 |
auto remaining_entries = line.entries.begin(); |
| 947 |
for (const auto &entry : line.entries) |
| 948 |
if (entry.first != numbers::invalid_size_type) |
| 949 |
{ |
| 950 |
*remaining_entries = entry; |
| 951 |
++remaining_entries; |
| 952 |
} |
| 953 |
line.entries.erase(remaining_entries, line.entries.end()); |
| 954 |
} |
| 955 |
} |
| 956 |
|
| 957 |
if constexpr (running_in_debug_mode()) |
| 958 |
{ |
| 959 |
// increase iteration count. note that we should not iterate more |
| 960 |
// times than there are constraints, since this puts a natural upper |
| 961 |
// bound on the length of constraint chains |
| 962 |
++iteration; |
| 963 |
Assert(iteration <= lines.size() + 1, ExcInternalError()); |
| 964 |
} |
| 965 |
} |
| 966 |
while (chained_constraint_replaced == true); |
| 967 |
|
| 968 |
// Finally sort the entries and re-scale them if necessary. in this step, |
| 969 |
// we also throw out duplicates as mentioned above. moreover, as some |
| 970 |
// entries might have had zero weights, we replace them by a vector with |
| 971 |
// sharp sizes. |
| 972 |
// |
| 973 |
// This is again an operation that works on each line separately. It can be |
| 974 |
// run in parallel: |
| 975 |
parallel::apply_to_subranges( |
| 976 |
lines.begin(), |
| 977 |
lines.end(), |
| 978 |
[](const typename std::vector<ConstraintLine>::iterator &begin, |
| 979 |
const typename std::vector<ConstraintLine>::iterator &end) { |
| 980 |
for (ConstraintLine &line : boost::iterator_range< |
| 981 |
typename std::vector<ConstraintLine>::iterator>(begin, end)) |
| 982 |
{ |
| 983 |
unsigned int duplicates = 0; |
| 984 |
bool is_sorted_without_duplicates = true; |
| 985 |
for (unsigned int i = 1; i < line.entries.size(); ++i) |
| 986 |
if (!(line.entries[i - 1].first < line.entries[i].first)) |
| 987 |
{ |
| 988 |
is_sorted_without_duplicates = false; |
| 989 |
break; |
| 990 |
} |
| 991 |
if (is_sorted_without_duplicates == false) |
| 992 |
{ |
| 993 |
std::sort(line.entries.begin(), |
| 994 |
line.entries.end(), |
| 995 |
[](const std::pair<unsigned int, number> &a, |
| 996 |
const std::pair<unsigned int, number> &b) -> bool { |
| 997 |
// Just look at the index, ignore the value. |
| 998 |
return a.first < b.first; |
| 999 |
}); |
| 1000 |
|
| 1001 |
// loop over the now sorted list and see whether any of the |
| 1002 |
// entries references the same dofs more than once in order to |
| 1003 |
// find how many non-duplicate entries we have. This lets us |
| 1004 |
// allocate the correct amount of memory for the constraint |
| 1005 |
// entries. |
| 1006 |
for (size_type i = 1; i < line.entries.size(); ++i) |
| 1007 |
if (line.entries[i].first == line.entries[i - 1].first) |
| 1008 |
++duplicates; |
| 1009 |
} |
| 1010 |
|
| 1011 |
if (duplicates > 0 || (line.entries.size() < line.entries.capacity())) |
| 1012 |
{ |
| 1013 |
typename ConstraintLine::Entries new_entries; |
| 1014 |
|
| 1015 |
// if we have no duplicates, copy verbatim the entries. this way, |
| 1016 |
// the final size is of the vector is correct. |
| 1017 |
if (duplicates == 0) |
| 1018 |
new_entries = line.entries; |
| 1019 |
else |
| 1020 |
{ |
| 1021 |
// otherwise, we need to go through the list and resolve the |
| 1022 |
// duplicates |
| 1023 |
new_entries.reserve(line.entries.size() - duplicates); |
| 1024 |
new_entries.push_back(line.entries[0]); |
| 1025 |
for (size_type j = 1; j < line.entries.size(); ++j) |
| 1026 |
if (line.entries[j].first == line.entries[j - 1].first) |
| 1027 |
{ |
| 1028 |
Assert(new_entries.back().first == |
| 1029 |
line.entries[j].first, |
| 1030 |
ExcInternalError()); |
| 1031 |
new_entries.back().second += line.entries[j].second; |
| 1032 |
} |
| 1033 |
else |
| 1034 |
new_entries.push_back(line.entries[j]); |
| 1035 |
|
| 1036 |
Assert(new_entries.size() == line.entries.size() - duplicates, |
| 1037 |
ExcInternalError()); |
| 1038 |
|
| 1039 |
// make sure there are really no duplicates left and that the |
| 1040 |
// list is still sorted |
| 1041 |
for (size_type j = 1; j < new_entries.size(); ++j) |
| 1042 |
{ |
| 1043 |
Assert(new_entries[j].first != new_entries[j - 1].first, |
| 1044 |
ExcInternalError()); |
| 1045 |
Assert(new_entries[j].first > new_entries[j - 1].first, |
| 1046 |
ExcInternalError()); |
| 1047 |
} |
| 1048 |
} |
| 1049 |
|
| 1050 |
// replace old list of constraints for this dof by the new one |
| 1051 |
line.entries.swap(new_entries); |
| 1052 |
} |
| 1053 |
|
| 1054 |
// Finally do the following check: if the sum of weights for the |
| 1055 |
// constraints is close to one, but not exactly one, then rescale all |
| 1056 |
// the weights so that they sum up to 1. this adds a little numerical |
| 1057 |
// stability and avoids all sorts of problems where the actual value |
| 1058 |
// is close to, but not quite what we expected |
| 1059 |
// |
| 1060 |
// the case where the weights don't quite sum up happens when we |
| 1061 |
// compute the interpolation weights "on the fly", i.e. not from |
| 1062 |
// precomputed tables. in this case, the interpolation weights are |
| 1063 |
// also subject to round-off |
| 1064 |
number sum = 0.; |
| 1065 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 1066 |
sum += entry.second; |
| 1067 |
if (std::abs(sum - number(1.)) < 1.e-13 && |
| 1068 |
std::abs(sum - number(1.)) > 0.) |
| 1069 |
{ |
| 1070 |
const number inverse_sum = number(1.) / sum; |
| 1071 |
for (std::pair<size_type, number> &entry : line.entries) |
| 1072 |
entry.second *= inverse_sum; |
| 1073 |
line.inhomogeneity *= inverse_sum; |
| 1074 |
} |
| 1075 |
} |
| 1076 |
}, |
| 1077 |
/* grainsize = */ 100); |
| 1078 |
|
| 1079 |
// if in debug mode: check that no dof is constrained to another dof that |
| 1080 |
// is also constrained. exclude dofs from this check whose constraint |
| 1081 |
// lines are not stored on the local processor |
| 1082 |
Assert(std::none_of(lines.begin(), |
| 1083 |
lines.end(), |
| 1084 |
[this](const ConstraintLine &line) { |
| 1085 |
for (const std::pair<size_type, number> &entry : |
| 1086 |
line.entries) |
| 1087 |
if ((local_lines.size() == 0) || |
| 1088 |
(local_lines.is_element(entry.first))) |
| 1089 |
{ |
| 1090 |
// make sure that entry->first is not the index of |
| 1091 |
// a line itself |
| 1092 |
const bool is_cycle = is_constrained(entry.first); |
| 1093 |
if (is_cycle) |
| 1094 |
return true; |
| 1095 |
} |
| 1096 |
return false; |
| 1097 |
}), |
| 1098 |
ExcMessage("The constraints represented by this object have a cycle. " |
| 1099 |
"This is not allowed.")); |
| 1100 |
|
| 1101 |
// Now also compute which indices we need in distribute(). |
| 1102 |
// |
| 1103 |
// This processor owns only part of the vectors we will work on. One may think |
| 1104 |
// that every processor should be able to simply communicate those elements it |
| 1105 |
// owns and for which it knows that they act as sources to constrained DoFs to |
| 1106 |
// the owner of these DoFs. This would lead to a scheme where all we need to |
| 1107 |
// do is to add some local elements to (possibly non-local) ones and then call |
| 1108 |
// compress(). |
| 1109 |
// |
| 1110 |
// Alas, this scheme does not work as evidenced by the disaster of |
| 1111 |
// bug #51 (originally stored in the Google Code repository, but now |
| 1112 |
// unfortunately no longer available because that platform has gone |
| 1113 |
// away) and reversion of one attempt that implements this in svn |
| 1114 |
// r29662. Rather, we need to get a vector that has all the |
| 1115 |
// *sources* or constraints we own locally, possibly as ghost vector |
| 1116 |
// elements, then read from them, and finally throw away the ghosted |
| 1117 |
// vector. Implement this in the following. |
| 1118 |
needed_elements_for_distribute = locally_owned_dofs; |
| 1119 |
|
| 1120 |
if (needed_elements_for_distribute != |
| 1121 |
complete_index_set(locally_owned_dofs.size())) |
| 1122 |
{ |
| 1123 |
std::vector<types::global_dof_index> additional_elements; |
| 1124 |
for (const ConstraintLine &line : lines) |
| 1125 |
if (locally_owned_dofs.is_element(line.index)) |
| 1126 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 1127 |
if (!locally_owned_dofs.is_element(entry.first)) |
| 1128 |
additional_elements.emplace_back(entry.first); |
| 1129 |
std::sort(additional_elements.begin(), additional_elements.end()); |
| 1130 |
needed_elements_for_distribute.add_indices(additional_elements.begin(), |
| 1131 |
additional_elements.end()); |
| 1132 |
} |
| 1133 |
|
| 1134 |
sorted = true; |
| 1135 |
} |
| 1136 |
|
| 1137 |
|
| 1138 |
|
| 1139 |
template <typename number> |
| 1140 |
bool |
| 1141 |
AffineConstraints<number>::is_closed() const |
| 1142 |
{ |
| 1143 |
return sorted || (n_constraints() == 0); |
| 1144 |
} |
| 1145 |
|
| 1146 |
|
| 1147 |
|
| 1148 |
template <typename number> |
| 1149 |
bool |
| 1150 |
AffineConstraints<number>::is_closed(const MPI_Comm comm) const |
| 1151 |
{ |
| 1152 |
return Utilities::MPI::min(static_cast<unsigned int>(is_closed()), comm) == 1; |
| 1153 |
} |
| 1154 |
|
| 1155 |
|
| 1156 |
|
| 1157 |
template <typename number> |
| 1158 |
void |
| 1159 |
AffineConstraints<number>::shift(const size_type offset) |
| 1160 |
{ |
| 1161 |
if (local_lines.size() == 0) |
| 1162 |
lines_cache.insert(lines_cache.begin(), offset, numbers::invalid_size_type); |
| 1163 |
else |
| 1164 |
{ |
| 1165 |
// shift local_lines |
| 1166 |
IndexSet new_local_lines(local_lines.size() + offset); |
| 1167 |
new_local_lines.add_indices(local_lines, offset); |
| 1168 |
std::swap(local_lines, new_local_lines); |
| 1169 |
} |
| 1170 |
|
| 1171 |
for (ConstraintLine &line : lines) |
| 1172 |
{ |
| 1173 |
line.index += offset; |
| 1174 |
for (std::pair<size_type, number> &entry : line.entries) |
| 1175 |
entry.first += offset; |
| 1176 |
} |
| 1177 |
|
| 1178 |
if constexpr (running_in_debug_mode()) |
| 1179 |
{ |
| 1180 |
// make sure that lines, lines_cache and local_lines |
| 1181 |
// are still linked correctly |
| 1182 |
for (size_type index = 0; index < lines_cache.size(); ++index) |
| 1183 |
Assert(lines_cache[index] == numbers::invalid_size_type || |
| 1184 |
calculate_line_index(lines[lines_cache[index]].index) == index, |
| 1185 |
ExcInternalError()); |
| 1186 |
} |
| 1187 |
} |
| 1188 |
|
| 1189 |
|
| 1190 |
|
| 1191 |
template <typename number> |
| 1192 |
AffineConstraints<number> |
| 1193 |
AffineConstraints<number>::get_view(const IndexSet &mask) const |
| 1194 |
{ |
| 1195 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 1196 |
|
| 1197 |
AffineConstraints<number> view; |
| 1198 |
|
| 1199 |
// If index sets were associated with the current object, take views |
| 1200 |
// of those as well: |
| 1201 |
if (locally_owned_dofs != IndexSet()) |
| 1202 |
{ |
| 1203 |
Assert(locally_owned_dofs.size() == mask.size(), |
| 1204 |
ExcMessage("This operation only makes sense if the index " |
| 1205 |
"space described by the mask is the same as " |
| 1206 |
"the index space associated with the " |
| 1207 |
"AffineConstraints object into which you take " |
| 1208 |
"a view.")); |
| 1209 |
Assert(local_lines.size() == mask.size(), |
| 1210 |
ExcMessage("This operation only makes sense if the index " |
| 1211 |
"space described by the mask is the same as " |
| 1212 |
"the index space associated with the " |
| 1213 |
"AffineConstraints object into which you take " |
| 1214 |
"a view.")); |
| 1215 |
|
| 1216 |
view.reinit(locally_owned_dofs.get_view(mask), |
| 1217 |
local_lines.get_view(mask)); |
| 1218 |
} |
| 1219 |
|
| 1220 |
for (const ConstraintLine &line : lines) |
| 1221 |
if (mask.is_element(line.index)) |
| 1222 |
{ |
| 1223 |
if constexpr (running_in_debug_mode()) |
| 1224 |
{ |
| 1225 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 1226 |
{ |
| 1227 |
Assert( |
| 1228 |
mask.is_element(entry.first), |
| 1229 |
ExcMessage( |
| 1230 |
"In creating a view of an AffineConstraints " |
| 1231 |
"object, the constraint on degree of freedom " + |
| 1232 |
std::to_string(line.index) + " (which corresponds to the " + |
| 1233 |
std::to_string(mask.index_within_set(line.index)) + |
| 1234 |
"th degree of freedom selected in the mask) " |
| 1235 |
"is constrained against degree of freedom " + |
| 1236 |
std::to_string(entry.first) + |
| 1237 |
", but this degree of freedom is not listed in the mask and " |
| 1238 |
"consequently cannot be transcribed into the index space " |
| 1239 |
"of the output object.")); |
| 1240 |
} |
| 1241 |
} |
| 1242 |
|
| 1243 |
std::vector<std::pair<size_type, number>> translated_entries = |
| 1244 |
line.entries; |
| 1245 |
for (auto &entry : translated_entries) |
| 1246 |
entry.first = mask.index_within_set(entry.first); |
| 1247 |
|
| 1248 |
view.add_constraint(mask.index_within_set(line.index), |
| 1249 |
translated_entries, |
| 1250 |
line.inhomogeneity); |
| 1251 |
} |
| 1252 |
|
| 1253 |
view.close(); |
| 1254 |
return view; |
| 1255 |
} |
| 1256 |
|
| 1257 |
|
| 1258 |
|
| 1259 |
template <typename number> |
| 1260 |
void |
| 1261 |
AffineConstraints<number>::clear() |
| 1262 |
{ |
| 1263 |
{ |
| 1264 |
std::vector<ConstraintLine> tmp; |
| 1265 |
lines.swap(tmp); |
| 1266 |
} |
| 1267 |
|
| 1268 |
{ |
| 1269 |
std::vector<size_type> tmp; |
| 1270 |
lines_cache.swap(tmp); |
| 1271 |
} |
| 1272 |
|
| 1273 |
locally_owned_dofs = {}; |
| 1274 |
local_lines = {}; |
| 1275 |
needed_elements_for_distribute = {}; |
| 1276 |
|
| 1277 |
sorted = false; |
| 1278 |
} |
| 1279 |
|
| 1280 |
|
| 1281 |
|
| 1282 |
template <typename number> |
| 1283 |
void |
| 1284 |
AffineConstraints<number>::reinit() |
| 1285 |
{ |
| 1286 |
reinit(IndexSet(), IndexSet()); |
| 1287 |
} |
| 1288 |
|
| 1289 |
|
| 1290 |
|
| 1291 |
template <typename number> |
| 1292 |
void |
| 1293 |
AffineConstraints<number>::reinit(const IndexSet &locally_stored_constraints) |
| 1294 |
{ |
| 1295 |
reinit(locally_stored_constraints, locally_stored_constraints); |
| 1296 |
} |
| 1297 |
|
| 1298 |
|
| 1299 |
|
| 1300 |
template <typename number> |
| 1301 |
void |
| 1302 |
AffineConstraints<number>::reinit(const IndexSet &locally_owned_dofs, |
| 1303 |
const IndexSet &locally_stored_constraints) |
| 1304 |
{ |
| 1305 |
// First clear previous content |
| 1306 |
clear(); |
| 1307 |
|
| 1308 |
// Then set the objects that describe the index sets of DoFs we care about: |
| 1309 |
Assert(locally_owned_dofs.is_subset_of(locally_stored_constraints), |
| 1310 |
ExcMessage("The set of locally stored constraints needs to be a " |
| 1311 |
"superset of the locally owned DoFs.")); |
| 1312 |
|
| 1313 |
this->locally_owned_dofs = locally_owned_dofs; |
| 1314 |
this->local_lines = locally_stored_constraints; |
| 1315 |
|
| 1316 |
// make sure the IndexSet is compressed. Otherwise this can lead to crashes |
| 1317 |
// that are hard to find (only happen in release mode). |
| 1318 |
// see tests/mpi/affine_constraints_crash_01 |
| 1319 |
this->locally_owned_dofs.compress(); |
| 1320 |
this->local_lines.compress(); |
| 1321 |
} |
| 1322 |
|
| 1323 |
|
| 1324 |
|
| 1325 |
template <typename number> |
| 1326 |
bool |
| 1327 |
AffineConstraints<number>::is_identity_constrained(const size_type line_n) const |
| 1328 |
{ |
| 1329 |
if (is_constrained(line_n) == false) |
| 1330 |
return false; |
| 1331 |
|
| 1332 |
const ConstraintLine &line = lines[lines_cache[calculate_line_index(line_n)]]; |
| 1333 |
Assert(line.index == line_n, ExcInternalError()); |
| 1334 |
|
| 1335 |
// return if an entry for this line was found and if it has only one |
| 1336 |
// entry equal to 1.0 |
| 1337 |
return (line.entries.size() == 1) && (line.entries[0].second == number(1.0)); |
| 1338 |
} |
| 1339 |
|
| 1340 |
|
| 1341 |
template <typename number> |
| 1342 |
bool |
| 1343 |
AffineConstraints<number>::are_identity_constrained( |
| 1344 |
const size_type line_n_1, |
| 1345 |
const size_type line_n_2) const |
| 1346 |
{ |
| 1347 |
if (is_constrained(line_n_1) == true) |
| 1348 |
{ |
| 1349 |
const ConstraintLine &line = |
| 1350 |
lines[lines_cache[calculate_line_index(line_n_1)]]; |
| 1351 |
Assert(line.index == line_n_1, ExcInternalError()); |
| 1352 |
|
| 1353 |
// return if an entry for this line was found and if it has only one |
| 1354 |
// entry equal to 1.0 and that one is index2 |
| 1355 |
return ((line.entries.size() == 1) && |
| 1356 |
(line.entries[0].first == line_n_2) && |
| 1357 |
(line.entries[0].second == number(1.0))); |
| 1358 |
} |
| 1359 |
else if (is_constrained(line_n_2) == true) |
| 1360 |
{ |
| 1361 |
const ConstraintLine &line = |
| 1362 |
lines[lines_cache[calculate_line_index(line_n_2)]]; |
| 1363 |
Assert(line.index == line_n_2, ExcInternalError()); |
| 1364 |
|
| 1365 |
// return if an entry for this line was found and if it has only one |
| 1366 |
// entry equal to 1.0 and that one is line_n_1 |
| 1367 |
return ((line.entries.size() == 1) && |
| 1368 |
(line.entries[0].first == line_n_1) && |
| 1369 |
(line.entries[0].second == number(1.0))); |
| 1370 |
} |
| 1371 |
else |
| 1372 |
return false; |
| 1373 |
} |
| 1374 |
|
| 1375 |
|
| 1376 |
|
| 1377 |
template <typename number> |
| 1378 |
typename AffineConstraints<number>::size_type |
| 1379 |
AffineConstraints<number>::max_constraint_indirections() const |
| 1380 |
{ |
| 1381 |
size_type return_value = 0; |
| 1382 |
for (const ConstraintLine &line : lines) |
| 1383 |
// use static cast, since typeof(size)==std::size_t, which is != |
| 1384 |
// size_type on AIX |
| 1385 |
return_value = |
| 1386 |
std::max(return_value, static_cast<size_type>(line.entries.size())); |
| 1387 |
|
| 1388 |
return return_value; |
| 1389 |
} |
| 1390 |
|
| 1391 |
|
| 1392 |
|
| 1393 |
template <typename number> |
| 1394 |
bool |
| 1395 |
AffineConstraints<number>::has_inhomogeneities() const |
| 1396 |
{ |
| 1397 |
for (const ConstraintLine &line : lines) |
| 1398 |
if (line.inhomogeneity != number(0.)) |
| 1399 |
return true; |
| 1400 |
|
| 1401 |
return false; |
| 1402 |
} |
| 1403 |
|
| 1404 |
|
| 1405 |
|
| 1406 |
template <typename number> |
| 1407 |
void |
| 1408 |
AffineConstraints<number>::print(std::ostream &out) const |
| 1409 |
{ |
| 1410 |
for (const ConstraintLine &line : lines) |
| 1411 |
{ |
| 1412 |
// output the list of constraints as pairs of dofs and their weights |
| 1413 |
if (line.entries.size() > 0) |
| 1414 |
{ |
| 1415 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 1416 |
out << " " << line.index << ' ' << entry.first << ": " |
| 1417 |
<< entry.second << '\n'; |
| 1418 |
|
| 1419 |
// print out inhomogeneity. |
| 1420 |
if (line.inhomogeneity != number(0.)) |
| 1421 |
out << " " << line.index << ": " << line.inhomogeneity << '\n'; |
| 1422 |
} |
| 1423 |
else |
| 1424 |
// but also output something if the constraint simply reads |
| 1425 |
// x[13]=0, i.e. where the right hand side is not a linear |
| 1426 |
// combination of other dofs |
| 1427 |
{ |
| 1428 |
if (line.inhomogeneity != number(0.)) |
| 1429 |
out << " " << line.index << " = " << line.inhomogeneity << '\n'; |
| 1430 |
else |
| 1431 |
out << " " << line.index << " = 0\n"; |
| 1432 |
} |
| 1433 |
} |
| 1434 |
|
| 1435 |
AssertThrow(out.fail() == false, ExcIO()); |
| 1436 |
} |
| 1437 |
|
| 1438 |
|
| 1439 |
|
| 1440 |
template <typename number> |
| 1441 |
void |
| 1442 |
AffineConstraints<number>::write_dot(std::ostream &out) const |
| 1443 |
{ |
| 1444 |
out << "digraph constraints {" << std::endl; |
| 1445 |
for (size_type i = 0; i != lines.size(); ++i) |
| 1446 |
{ |
| 1447 |
// same concept as in the previous function |
| 1448 |
if (lines[i].entries.size() > 0) |
| 1449 |
for (size_type j = 0; j < lines[i].entries.size(); ++j) |
| 1450 |
out << " " << lines[i].index << "->" << lines[i].entries[j].first |
| 1451 |
<< "; // weight: " << lines[i].entries[j].second << '\n'; |
| 1452 |
else |
| 1453 |
out << " " << lines[i].index << '\n'; |
| 1454 |
} |
| 1455 |
out << '}' << std::endl; |
| 1456 |
} |
| 1457 |
|
| 1458 |
|
| 1459 |
|
| 1460 |
template <typename number> |
| 1461 |
std::size_t |
| 1462 |
AffineConstraints<number>::memory_consumption() const |
| 1463 |
{ |
| 1464 |
return (MemoryConsumption::memory_consumption(lines) + |
| 1465 |
MemoryConsumption::memory_consumption(lines_cache) + |
| 1466 |
MemoryConsumption::memory_consumption(sorted) + |
| 1467 |
MemoryConsumption::memory_consumption(local_lines)); |
| 1468 |
} |
| 1469 |
|
| 1470 |
|
| 1471 |
|
| 1472 |
template <typename number> |
| 1473 |
void |
| 1474 |
AffineConstraints<number>::resolve_indices( |
| 1475 |
std::vector<types::global_dof_index> &indices) const |
| 1476 |
{ |
| 1477 |
const unsigned int indices_size = indices.size(); |
| 1478 |
const std::vector<std::pair<types::global_dof_index, number>> *line_ptr; |
| 1479 |
for (unsigned int i = 0; i < indices_size; ++i) |
| 1480 |
{ |
| 1481 |
line_ptr = get_constraint_entries(indices[i]); |
| 1482 |
// if the index is constraint, the constraints indices are added to the |
| 1483 |
// indices vector |
| 1484 |
if (line_ptr != nullptr) |
| 1485 |
{ |
| 1486 |
const unsigned int line_size = line_ptr->size(); |
| 1487 |
for (unsigned int j = 0; j < line_size; ++j) |
| 1488 |
indices.push_back((*line_ptr)[j].first); |
| 1489 |
} |
| 1490 |
} |
| 1491 |
|
| 1492 |
// keep only the unique elements |
| 1493 |
std::sort(indices.begin(), indices.end()); |
| 1494 |
indices.erase(std::unique(indices.begin(), indices.end()), indices.end()); |
| 1495 |
} |
| 1496 |
|
| 1497 |
|
| 1498 |
|
| 1499 |
template <typename number> |
| 1500 |
void |
| 1501 |
AffineConstraints<number>::condense(SparsityPattern &sparsity) const |
| 1502 |
{ |
| 1503 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 1504 |
Assert(sparsity.is_compressed() == false, ExcMatrixIsClosed()); |
| 1505 |
Assert(sparsity.n_rows() == sparsity.n_cols(), ExcNotQuadratic()); |
| 1506 |
|
| 1507 |
// store for each index whether it must be distributed or not. If entry |
| 1508 |
// is numbers::invalid_unsigned_int, no distribution is necessary. |
| 1509 |
// otherwise, the number states which line in the AffineConstraints object |
| 1510 |
// handles this index |
| 1511 |
std::vector<size_type> distribute(sparsity.n_rows(), |
| 1512 |
numbers::invalid_size_type); |
| 1513 |
|
| 1514 |
for (size_type c = 0; c < lines.size(); ++c) |
| 1515 |
distribute[lines[c].index] = c; |
| 1516 |
|
| 1517 |
const size_type n_rows = sparsity.n_rows(); |
| 1518 |
for (size_type row = 0; row < n_rows; ++row) |
| 1519 |
{ |
| 1520 |
if (distribute[row] == numbers::invalid_size_type) |
| 1521 |
{ |
| 1522 |
// regular line. loop over cols all valid cols. note that this |
| 1523 |
// changes the line we are presently working on: we add additional |
| 1524 |
// entries. these are put to the end of the row. however, as |
| 1525 |
// constrained nodes cannot be constrained to other constrained |
| 1526 |
// nodes, nothing will happen if we run into these added nodes, as |
| 1527 |
// they can't be distributed further. we might store the position of |
| 1528 |
// the last old entry and stop work there, but since operating on |
| 1529 |
// the newly added ones only takes two comparisons (column index |
| 1530 |
// valid, distribute[column] necessarily |
| 1531 |
// ==numbers::invalid_size_type), it is cheaper to not do so and |
| 1532 |
// run right until the end of the line |
| 1533 |
for (SparsityPattern::iterator entry = sparsity.begin(row); |
| 1534 |
((entry != sparsity.end(row)) && entry->is_valid_entry()); |
| 1535 |
++entry) |
| 1536 |
{ |
| 1537 |
const size_type column = entry->column(); |
| 1538 |
|
| 1539 |
if (distribute[column] != numbers::invalid_size_type) |
| 1540 |
{ |
| 1541 |
// distribute entry at regular row @p{row} and irregular |
| 1542 |
// column sparsity.colnums[j] |
| 1543 |
for (size_type q = 0; |
| 1544 |
q != lines[distribute[column]].entries.size(); |
| 1545 |
++q) |
| 1546 |
sparsity.add(row, |
| 1547 |
lines[distribute[column]].entries[q].first); |
| 1548 |
} |
| 1549 |
} |
| 1550 |
} |
| 1551 |
else |
| 1552 |
// row must be distributed. note that here the present row is not |
| 1553 |
// touched (unlike above) |
| 1554 |
{ |
| 1555 |
for (SparsityPattern::iterator entry = sparsity.begin(row); |
| 1556 |
(entry != sparsity.end(row)) && entry->is_valid_entry(); |
| 1557 |
++entry) |
| 1558 |
{ |
| 1559 |
const size_type column = entry->column(); |
| 1560 |
if (distribute[column] == numbers::invalid_size_type) |
| 1561 |
// distribute entry at irregular row @p{row} and regular |
| 1562 |
// column sparsity.colnums[j] |
| 1563 |
for (size_type q = 0; |
| 1564 |
q != lines[distribute[row]].entries.size(); |
| 1565 |
++q) |
| 1566 |
sparsity.add(lines[distribute[row]].entries[q].first, column); |
| 1567 |
else |
| 1568 |
// distribute entry at irregular row @p{row} and irregular |
| 1569 |
// column sparsity.get_column_numbers()[j] |
| 1570 |
for (size_type p = 0; |
| 1571 |
p != lines[distribute[row]].entries.size(); |
| 1572 |
++p) |
| 1573 |
for (size_type q = 0; |
| 1574 |
q != lines[distribute[column]].entries.size(); |
| 1575 |
++q) |
| 1576 |
sparsity.add(lines[distribute[row]].entries[p].first, |
| 1577 |
lines[distribute[column]].entries[q].first); |
| 1578 |
} |
| 1579 |
} |
| 1580 |
} |
| 1581 |
|
| 1582 |
sparsity.compress(); |
| 1583 |
} |
| 1584 |
|
| 1585 |
|
| 1586 |
|
| 1587 |
template <typename number> |
| 1588 |
void |
| 1589 |
AffineConstraints<number>::condense(BlockSparsityPattern &sparsity) const |
| 1590 |
{ |
| 1591 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 1592 |
Assert(sparsity.is_compressed() == false, ExcMatrixIsClosed()); |
| 1593 |
Assert(sparsity.n_rows() == sparsity.n_cols(), ExcNotQuadratic()); |
| 1594 |
Assert(sparsity.n_block_rows() == sparsity.n_block_cols(), ExcNotQuadratic()); |
| 1595 |
Assert(sparsity.get_column_indices() == sparsity.get_row_indices(), |
| 1596 |
ExcNotQuadratic()); |
| 1597 |
|
| 1598 |
const BlockIndices &index_mapping = sparsity.get_column_indices(); |
| 1599 |
|
| 1600 |
const size_type n_blocks = sparsity.n_block_rows(); |
| 1601 |
|
| 1602 |
// store for each index whether it must be distributed or not. If entry |
| 1603 |
// is numbers::invalid_unsigned_int, no distribution is necessary. |
| 1604 |
// otherwise, the number states which line in the AffineConstraints object |
| 1605 |
// handles this index |
| 1606 |
std::vector<size_type> distribute(sparsity.n_rows(), |
| 1607 |
numbers::invalid_size_type); |
| 1608 |
|
| 1609 |
for (size_type c = 0; c < lines.size(); ++c) |
| 1610 |
distribute[lines[c].index] = c; |
| 1611 |
|
| 1612 |
const size_type n_rows = sparsity.n_rows(); |
| 1613 |
for (size_type row = 0; row < n_rows; ++row) |
| 1614 |
{ |
| 1615 |
// get index of this row within the blocks |
| 1616 |
const std::pair<size_type, size_type> block_index = |
| 1617 |
index_mapping.global_to_local(row); |
| 1618 |
const size_type block_row = block_index.first; |
| 1619 |
|
| 1620 |
if (distribute[row] == numbers::invalid_size_type) |
| 1621 |
// regular line. loop over all columns and see whether this column |
| 1622 |
// must be distributed |
| 1623 |
{ |
| 1624 |
// to loop over all entries in this row, we have to loop over all |
| 1625 |
// blocks in this blockrow and the corresponding row therein |
| 1626 |
for (size_type block_col = 0; block_col < n_blocks; ++block_col) |
| 1627 |
{ |
| 1628 |
const SparsityPattern &block_sparsity = |
| 1629 |
sparsity.block(block_row, block_col); |
| 1630 |
|
| 1631 |
for (SparsityPattern::const_iterator entry = |
| 1632 |
block_sparsity.begin(block_index.second); |
| 1633 |
(entry != block_sparsity.end(block_index.second)) && |
| 1634 |
entry->is_valid_entry(); |
| 1635 |
++entry) |
| 1636 |
{ |
| 1637 |
const size_type global_col = |
| 1638 |
index_mapping.local_to_global(block_col, entry->column()); |
| 1639 |
|
| 1640 |
if (distribute[global_col] != numbers::invalid_size_type) |
| 1641 |
// distribute entry at regular row @p{row} and |
| 1642 |
// irregular column global_col |
| 1643 |
{ |
| 1644 |
for (size_type q = 0; |
| 1645 |
q != lines[distribute[global_col]].entries.size(); |
| 1646 |
++q) |
| 1647 |
sparsity.add( |
| 1648 |
row, lines[distribute[global_col]].entries[q].first); |
| 1649 |
} |
| 1650 |
} |
| 1651 |
} |
| 1652 |
} |
| 1653 |
else |
| 1654 |
{ |
| 1655 |
// row must be distributed. split the whole row into the chunks |
| 1656 |
// defined by the blocks |
| 1657 |
for (size_type block_col = 0; block_col < n_blocks; ++block_col) |
| 1658 |
{ |
| 1659 |
const SparsityPattern &block_sparsity = |
| 1660 |
sparsity.block(block_row, block_col); |
| 1661 |
|
| 1662 |
for (SparsityPattern::const_iterator entry = |
| 1663 |
block_sparsity.begin(block_index.second); |
| 1664 |
(entry != block_sparsity.end(block_index.second)) && |
| 1665 |
entry->is_valid_entry(); |
| 1666 |
++entry) |
| 1667 |
{ |
| 1668 |
const size_type global_col = |
| 1669 |
index_mapping.local_to_global(block_col, entry->column()); |
| 1670 |
|
| 1671 |
if (distribute[global_col] == numbers::invalid_size_type) |
| 1672 |
// distribute entry at irregular row @p{row} and |
| 1673 |
// regular column global_col. |
| 1674 |
{ |
| 1675 |
for (size_type q = 0; |
| 1676 |
q != lines[distribute[row]].entries.size(); |
| 1677 |
++q) |
| 1678 |
sparsity.add(lines[distribute[row]].entries[q].first, |
| 1679 |
global_col); |
| 1680 |
} |
| 1681 |
else |
| 1682 |
// distribute entry at irregular row @p{row} and |
| 1683 |
// irregular column @p{global_col} |
| 1684 |
{ |
| 1685 |
for (size_type p = 0; |
| 1686 |
p != lines[distribute[row]].entries.size(); |
| 1687 |
++p) |
| 1688 |
for (size_type q = 0; |
| 1689 |
q != lines[distribute[global_col]].entries.size(); |
| 1690 |
++q) |
| 1691 |
sparsity.add( |
| 1692 |
lines[distribute[row]].entries[p].first, |
| 1693 |
lines[distribute[global_col]].entries[q].first); |
| 1694 |
} |
| 1695 |
} |
| 1696 |
} |
| 1697 |
} |
| 1698 |
} |
| 1699 |
|
| 1700 |
sparsity.compress(); |
| 1701 |
} |
| 1702 |
|
| 1703 |
|
| 1704 |
|
| 1705 |
template <typename number> |
| 1706 |
void |
| 1707 |
AffineConstraints<number>::condense(DynamicSparsityPattern &sparsity) const |
| 1708 |
{ |
| 1709 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 1710 |
Assert(sparsity.n_rows() == sparsity.n_cols(), ExcNotQuadratic()); |
| 1711 |
|
| 1712 |
// store for each index whether it must be distributed or not. If entry |
| 1713 |
// is numbers::invalid_unsigned_int, no distribution is necessary. |
| 1714 |
// otherwise, the number states which line in the constraints object |
| 1715 |
// handles this index |
| 1716 |
std::vector<size_type> distribute(sparsity.n_rows(), |
| 1717 |
numbers::invalid_size_type); |
| 1718 |
|
| 1719 |
for (size_type c = 0; c < lines.size(); ++c) |
| 1720 |
distribute[lines[c].index] = c; |
| 1721 |
|
| 1722 |
const size_type n_rows = sparsity.n_rows(); |
| 1723 |
for (size_type row = 0; row < n_rows; ++row) |
| 1724 |
{ |
| 1725 |
if (distribute[row] == numbers::invalid_size_type) |
| 1726 |
// regular line. loop over cols. note that as we proceed to |
| 1727 |
// distribute cols, the loop may get longer |
| 1728 |
for (size_type j = 0; j < sparsity.row_length(row); ++j) |
| 1729 |
{ |
| 1730 |
const size_type column = sparsity.column_number(row, j); |
| 1731 |
|
| 1732 |
if (distribute[column] != numbers::invalid_size_type) |
| 1733 |
{ |
| 1734 |
// distribute entry at regular row @p{row} and irregular |
| 1735 |
// column column. note that this changes the line we are |
| 1736 |
// presently working on: we add additional entries. if we |
| 1737 |
// add another entry at a column behind the present one, we |
| 1738 |
// will encounter it later on (but since it can't be |
| 1739 |
// further constrained, won't have to do anything about |
| 1740 |
// it). if we add it up front of the present column, we |
| 1741 |
// will find the present column later on again as it was |
| 1742 |
// shifted back (again nothing happens, in particular no |
| 1743 |
// endless loop, as when we encounter it the second time we |
| 1744 |
// won't be able to add more entries as they all already |
| 1745 |
// exist, but we do the same work more often than |
| 1746 |
// necessary, and the loop gets longer), so move the cursor |
| 1747 |
// one to the right in the case that we add an entry up |
| 1748 |
// front that did not exist before. check whether it |
| 1749 |
// existed before by tracking the length of this row |
| 1750 |
size_type old_rowlength = sparsity.row_length(row); |
| 1751 |
for (size_type q = 0; |
| 1752 |
q != lines[distribute[column]].entries.size(); |
| 1753 |
++q) |
| 1754 |
{ |
| 1755 |
const size_type new_col = |
| 1756 |
lines[distribute[column]].entries[q].first; |
| 1757 |
|
| 1758 |
sparsity.add(row, new_col); |
| 1759 |
|
| 1760 |
const size_type new_rowlength = sparsity.row_length(row); |
| 1761 |
if ((new_col < column) && (old_rowlength != new_rowlength)) |
| 1762 |
++j; |
| 1763 |
old_rowlength = new_rowlength; |
| 1764 |
} |
| 1765 |
} |
| 1766 |
} |
| 1767 |
else |
| 1768 |
// row must be distributed |
| 1769 |
for (size_type j = 0; j < sparsity.row_length(row); ++j) |
| 1770 |
{ |
| 1771 |
const size_type column = sparsity.column_number(row, j); |
| 1772 |
|
| 1773 |
if (distribute[column] == numbers::invalid_size_type) |
| 1774 |
// distribute entry at irregular row @p{row} and regular |
| 1775 |
// column sparsity.colnums[j] |
| 1776 |
for (size_type q = 0; q != lines[distribute[row]].entries.size(); |
| 1777 |
++q) |
| 1778 |
sparsity.add(lines[distribute[row]].entries[q].first, column); |
| 1779 |
else |
| 1780 |
// distribute entry at irregular row @p{row} and irregular |
| 1781 |
// column sparsity.get_column_numbers()[j] |
| 1782 |
for (size_type p = 0; p != lines[distribute[row]].entries.size(); |
| 1783 |
++p) |
| 1784 |
for (size_type q = 0; |
| 1785 |
q != lines[distribute[sparsity.column_number(row, j)]] |
| 1786 |
.entries.size(); |
| 1787 |
++q) |
| 1788 |
sparsity.add(lines[distribute[row]].entries[p].first, |
| 1789 |
lines[distribute[sparsity.column_number(row, j)]] |
| 1790 |
.entries[q] |
| 1791 |
.first); |
| 1792 |
} |
| 1793 |
} |
| 1794 |
} |
| 1795 |
|
| 1796 |
|
| 1797 |
|
| 1798 |
template <typename number> |
| 1799 |
void |
| 1800 |
AffineConstraints<number>::condense(BlockDynamicSparsityPattern &sparsity) const |
| 1801 |
{ |
| 1802 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 1803 |
Assert(sparsity.n_rows() == sparsity.n_cols(), ExcNotQuadratic()); |
| 1804 |
Assert(sparsity.n_block_rows() == sparsity.n_block_cols(), ExcNotQuadratic()); |
| 1805 |
Assert(sparsity.get_column_indices() == sparsity.get_row_indices(), |
| 1806 |
ExcNotQuadratic()); |
| 1807 |
|
| 1808 |
const BlockIndices &index_mapping = sparsity.get_column_indices(); |
| 1809 |
|
| 1810 |
const size_type n_blocks = sparsity.n_block_rows(); |
| 1811 |
|
| 1812 |
// store for each index whether it must be distributed or not. If entry |
| 1813 |
// is numbers::invalid_unsigned_int, no distribution is necessary. |
| 1814 |
// otherwise, the number states which line in the constraints object |
| 1815 |
// handles this index |
| 1816 |
std::vector<size_type> distribute(sparsity.n_rows(), |
| 1817 |
numbers::invalid_size_type); |
| 1818 |
|
| 1819 |
for (size_type c = 0; c < lines.size(); ++c) |
| 1820 |
distribute[lines[c].index] = static_cast<signed int>(c); |
| 1821 |
|
| 1822 |
const size_type n_rows = sparsity.n_rows(); |
| 1823 |
for (size_type row = 0; row < n_rows; ++row) |
| 1824 |
{ |
| 1825 |
// get index of this row within the blocks |
| 1826 |
const std::pair<size_type, size_type> block_index = |
| 1827 |
index_mapping.global_to_local(row); |
| 1828 |
const size_type block_row = block_index.first; |
| 1829 |
const size_type local_row = block_index.second; |
| 1830 |
|
| 1831 |
if (distribute[row] == numbers::invalid_size_type) |
| 1832 |
// regular line. loop over all columns and see whether this column |
| 1833 |
// must be distributed. note that as we proceed to distribute cols, |
| 1834 |
// the loop over cols may get longer. |
| 1835 |
// |
| 1836 |
// don't try to be clever here as in the algorithm for the |
| 1837 |
// DynamicSparsityPattern, as that would be much more |
| 1838 |
// complicated here. after all, we know that compressed patterns |
| 1839 |
// are inefficient... |
| 1840 |
{ |
| 1841 |
// to loop over all entries in this row, we have to loop over all |
| 1842 |
// blocks in this blockrow and the corresponding row therein |
| 1843 |
for (size_type block_col = 0; block_col < n_blocks; ++block_col) |
| 1844 |
{ |
| 1845 |
const DynamicSparsityPattern &block_sparsity = |
| 1846 |
sparsity.block(block_row, block_col); |
| 1847 |
|
| 1848 |
for (size_type j = 0; j < block_sparsity.row_length(local_row); |
| 1849 |
++j) |
| 1850 |
{ |
| 1851 |
const size_type global_col = index_mapping.local_to_global( |
| 1852 |
block_col, block_sparsity.column_number(local_row, j)); |
| 1853 |
|
| 1854 |
if (distribute[global_col] != numbers::invalid_size_type) |
| 1855 |
// distribute entry at regular row @p{row} and |
| 1856 |
// irregular column global_col |
| 1857 |
{ |
| 1858 |
for (size_type q = 0; |
| 1859 |
q != lines[distribute[global_col]].entries.size(); |
| 1860 |
++q) |
| 1861 |
sparsity.add( |
| 1862 |
row, lines[distribute[global_col]].entries[q].first); |
| 1863 |
} |
| 1864 |
} |
| 1865 |
} |
| 1866 |
} |
| 1867 |
else |
| 1868 |
{ |
| 1869 |
// row must be distributed. split the whole row into the chunks |
| 1870 |
// defined by the blocks |
| 1871 |
for (size_type block_col = 0; block_col < n_blocks; ++block_col) |
| 1872 |
{ |
| 1873 |
const DynamicSparsityPattern &block_sparsity = |
| 1874 |
sparsity.block(block_row, block_col); |
| 1875 |
|
| 1876 |
for (size_type j = 0; j < block_sparsity.row_length(local_row); |
| 1877 |
++j) |
| 1878 |
{ |
| 1879 |
const size_type global_col = index_mapping.local_to_global( |
| 1880 |
block_col, block_sparsity.column_number(local_row, j)); |
| 1881 |
|
| 1882 |
if (distribute[global_col] == numbers::invalid_size_type) |
| 1883 |
// distribute entry at irregular row @p{row} and |
| 1884 |
// regular column global_col. |
| 1885 |
{ |
| 1886 |
for (size_type q = 0; |
| 1887 |
q != lines[distribute[row]].entries.size(); |
| 1888 |
++q) |
| 1889 |
sparsity.add(lines[distribute[row]].entries[q].first, |
| 1890 |
global_col); |
| 1891 |
} |
| 1892 |
else |
| 1893 |
// distribute entry at irregular row @p{row} and |
| 1894 |
// irregular column @p{global_col} |
| 1895 |
{ |
| 1896 |
for (size_type p = 0; |
| 1897 |
p != lines[distribute[row]].entries.size(); |
| 1898 |
++p) |
| 1899 |
for (size_type q = 0; |
| 1900 |
q != lines[distribute[global_col]].entries.size(); |
| 1901 |
++q) |
| 1902 |
sparsity.add( |
| 1903 |
lines[distribute[row]].entries[p].first, |
| 1904 |
lines[distribute[global_col]].entries[q].first); |
| 1905 |
} |
| 1906 |
} |
| 1907 |
} |
| 1908 |
} |
| 1909 |
} |
| 1910 |
} |
| 1911 |
|
| 1912 |
|
| 1913 |
|
| 1914 |
template <typename number> |
| 1915 |
void |
| 1916 |
AffineConstraints<number>::condense(SparseMatrix<number> &uncondensed) const |
| 1917 |
{ |
| 1918 |
Vector<number> dummy(0); |
| 1919 |
condense(uncondensed, dummy); |
| 1920 |
} |
| 1921 |
|
| 1922 |
|
| 1923 |
|
| 1924 |
template <typename number> |
| 1925 |
void |
| 1926 |
AffineConstraints<number>::condense( |
| 1927 |
BlockSparseMatrix<number> &uncondensed) const |
| 1928 |
{ |
| 1929 |
BlockVector<number> dummy(0); |
| 1930 |
condense(uncondensed, dummy); |
| 1931 |
} |
| 1932 |
|
| 1933 |
|
| 1934 |
|
| 1935 |
template <typename number> |
| 1936 |
template <typename VectorType> |
| 1937 |
void |
| 1938 |
AffineConstraints<number>::condense(const VectorType &vec_ghosted, |
| 1939 |
VectorType &vec) const |
| 1940 |
{ |
| 1941 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 1942 |
|
| 1943 |
// if this is called with different arguments, we need to copy the data |
| 1944 |
// over: |
| 1945 |
if (&vec != &vec_ghosted) |
| 1946 |
vec = vec_ghosted; |
| 1947 |
|
| 1948 |
// distribute all entries, and set them to zero. do so in two loops |
| 1949 |
// because in the first one we need to add to elements and in the second |
| 1950 |
// one we need to set elements to zero. for parallel vectors, this can |
| 1951 |
// only work if we can put a compress() in between, but we don't want to |
| 1952 |
// call compress() twice per entry |
| 1953 |
for (const ConstraintLine &line : lines) |
| 1954 |
{ |
| 1955 |
// in case the constraint is inhomogeneous, this function is not |
| 1956 |
// appropriate. Throw an exception. |
| 1957 |
Assert(line.inhomogeneity == number(0.), |
| 1958 |
ExcMessage("Inhomogeneous constraint cannot be condensed " |
| 1959 |
"without any matrix specified.")); |
| 1960 |
|
| 1961 |
const typename VectorType::value_type old_value = vec_ghosted(line.index); |
| 1962 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 1963 |
if (vec.in_local_range(entry.first) == true) |
| 1964 |
vec(entry.first) += |
| 1965 |
(static_cast<typename VectorType::value_type>(old_value) * |
| 1966 |
entry.second); |
| 1967 |
} |
| 1968 |
|
| 1969 |
vec.compress(VectorOperation::add); |
| 1970 |
|
| 1971 |
for (const ConstraintLine &line : lines) |
| 1972 |
if (vec.in_local_range(line.index) == true) |
| 1973 |
vec(line.index) = 0.; |
| 1974 |
|
| 1975 |
vec.compress(VectorOperation::insert); |
| 1976 |
} |
| 1977 |
|
| 1978 |
|
| 1979 |
|
| 1980 |
template <typename number> |
| 1981 |
template <typename VectorType> |
| 1982 |
void |
| 1983 |
AffineConstraints<number>::condense(VectorType &vec) const |
| 1984 |
{ |
| 1985 |
condense(vec, vec); |
| 1986 |
} |
| 1987 |
|
| 1988 |
|
| 1989 |
|
| 1990 |
template <typename number> |
| 1991 |
template <typename VectorType> |
| 1992 |
void |
| 1993 |
AffineConstraints<number>::condense(SparseMatrix<number> &uncondensed, |
| 1994 |
VectorType &vec) const |
| 1995 |
{ |
| 1996 |
// check whether we work on real vectors or we just used a dummy when |
| 1997 |
// calling the other function above. |
| 1998 |
const bool use_vectors = vec.size() == 0 ? false : true; |
| 1999 |
|
| 2000 |
const SparsityPattern &sparsity = uncondensed.get_sparsity_pattern(); |
| 2001 |
|
| 2002 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 2003 |
Assert(sparsity.is_compressed() == true, ExcMatrixNotClosed()); |
| 2004 |
Assert(sparsity.n_rows() == sparsity.n_cols(), ExcNotQuadratic()); |
| 2005 |
if (use_vectors == true) |
| 2006 |
AssertDimension(vec.size(), sparsity.n_rows()); |
| 2007 |
|
| 2008 |
number average_diagonal = 0.; |
| 2009 |
for (size_type i = 0; i < uncondensed.m(); ++i) |
| 2010 |
average_diagonal += std::abs(uncondensed.diag_element(i)); |
| 2011 |
average_diagonal /= uncondensed.m(); |
| 2012 |
|
| 2013 |
// store for each index whether it must be distributed or not. If entry |
| 2014 |
// is invalid_size_type, no distribution is necessary. otherwise, the |
| 2015 |
// number states which line in the constraints object handles this index |
| 2016 |
std::vector<size_type> distribute(sparsity.n_rows(), |
| 2017 |
numbers::invalid_size_type); |
| 2018 |
|
| 2019 |
for (size_type c = 0; c < lines.size(); ++c) |
| 2020 |
distribute[lines[c].index] = c; |
| 2021 |
|
| 2022 |
const size_type n_rows = sparsity.n_rows(); |
| 2023 |
for (size_type row = 0; row < n_rows; ++row) |
| 2024 |
{ |
| 2025 |
if (distribute[row] == numbers::invalid_size_type) |
| 2026 |
// regular line. loop over cols |
| 2027 |
{ |
| 2028 |
for (typename SparseMatrix<number>::iterator entry = |
| 2029 |
uncondensed.begin(row); |
| 2030 |
entry != uncondensed.end(row); |
| 2031 |
++entry) |
| 2032 |
{ |
| 2033 |
const size_type column = entry->column(); |
| 2034 |
|
| 2035 |
// end of row reached? this should not happen, since we only |
| 2036 |
// operate on compressed matrices! |
| 2037 |
Assert(column != SparsityPattern::invalid_entry, |
| 2038 |
ExcMatrixNotClosed()); |
| 2039 |
|
| 2040 |
if (distribute[column] != numbers::invalid_size_type) |
| 2041 |
// distribute entry at regular row @p row and irregular |
| 2042 |
// column sparsity.get_column_numbers()[j]; set old entry |
| 2043 |
// to zero |
| 2044 |
{ |
| 2045 |
for (size_type q = 0; |
| 2046 |
q != lines[distribute[column]].entries.size(); |
| 2047 |
++q) |
| 2048 |
{ |
| 2049 |
// need a temporary variable to avoid errors like no |
| 2050 |
// known conversion from 'complex<typename |
| 2051 |
// ProductType<float, double>::type>' to 'const |
| 2052 |
// complex<float>' for 3rd argument |
| 2053 |
number v = static_cast<number>(entry->value()); |
| 2054 |
v *= lines[distribute[column]].entries[q].second; |
| 2055 |
uncondensed.add( |
| 2056 |
row, lines[distribute[column]].entries[q].first, v); |
| 2057 |
} |
| 2058 |
|
| 2059 |
// need to subtract this element from the vector. this |
| 2060 |
// corresponds to an explicit elimination in the |
| 2061 |
// respective row of the inhomogeneous constraint in the |
| 2062 |
// matrix with Gauss elimination |
| 2063 |
if (use_vectors == true) |
| 2064 |
vec(row) -= static_cast<number>(entry->value()) * |
| 2065 |
lines[distribute[column]].inhomogeneity; |
| 2066 |
|
| 2067 |
// set old value to zero |
| 2068 |
entry->value() = 0.; |
| 2069 |
} |
| 2070 |
} |
| 2071 |
} |
| 2072 |
else |
| 2073 |
// row must be distributed |
| 2074 |
{ |
| 2075 |
for (typename SparseMatrix<number>::iterator entry = |
| 2076 |
uncondensed.begin(row); |
| 2077 |
entry != uncondensed.end(row); |
| 2078 |
++entry) |
| 2079 |
{ |
| 2080 |
const size_type column = entry->column(); |
| 2081 |
|
| 2082 |
// end of row reached? this should not happen, since we only |
| 2083 |
// operate on compressed matrices! |
| 2084 |
Assert(column != SparsityPattern::invalid_entry, |
| 2085 |
ExcMatrixNotClosed()); |
| 2086 |
|
| 2087 |
if (distribute[column] == numbers::invalid_size_type) |
| 2088 |
// distribute entry at irregular row @p row and regular |
| 2089 |
// column column. set old entry to zero |
| 2090 |
{ |
| 2091 |
for (size_type q = 0; |
| 2092 |
q != lines[distribute[row]].entries.size(); |
| 2093 |
++q) |
| 2094 |
{ |
| 2095 |
// need a temporary variable to avoid errors like |
| 2096 |
// no known conversion from 'complex<typename |
| 2097 |
// ProductType<float, double>::type>' to 'const |
| 2098 |
// complex<float>' for 3rd argument |
| 2099 |
number v = static_cast<number>(entry->value()); |
| 2100 |
v *= lines[distribute[row]].entries[q].second; |
| 2101 |
uncondensed.add(lines[distribute[row]].entries[q].first, |
| 2102 |
column, |
| 2103 |
v); |
| 2104 |
} |
| 2105 |
|
| 2106 |
// set old entry to zero |
| 2107 |
entry->value() = 0.; |
| 2108 |
} |
| 2109 |
else |
| 2110 |
// distribute entry at irregular row @p row and irregular |
| 2111 |
// column @p column set old entry to one on main diagonal, |
| 2112 |
// zero otherwise |
| 2113 |
{ |
| 2114 |
for (size_type p = 0; |
| 2115 |
p != lines[distribute[row]].entries.size(); |
| 2116 |
++p) |
| 2117 |
{ |
| 2118 |
for (size_type q = 0; |
| 2119 |
q != lines[distribute[column]].entries.size(); |
| 2120 |
++q) |
| 2121 |
{ |
| 2122 |
// need a temporary variable to avoid errors like |
| 2123 |
// no known conversion from 'complex<typename |
| 2124 |
// ProductType<float, double>::type>' to 'const |
| 2125 |
// complex<float>' for 3rd argument |
| 2126 |
number v = static_cast<number>(entry->value()); |
| 2127 |
v *= lines[distribute[row]].entries[p].second * |
| 2128 |
lines[distribute[column]].entries[q].second; |
| 2129 |
uncondensed.add( |
| 2130 |
lines[distribute[row]].entries[p].first, |
| 2131 |
lines[distribute[column]].entries[q].first, |
| 2132 |
v); |
| 2133 |
} |
| 2134 |
|
| 2135 |
if (use_vectors == true) |
| 2136 |
vec(lines[distribute[row]].entries[p].first) -= |
| 2137 |
static_cast<number>(entry->value()) * |
| 2138 |
lines[distribute[row]].entries[p].second * |
| 2139 |
lines[distribute[column]].inhomogeneity; |
| 2140 |
} |
| 2141 |
|
| 2142 |
// set old entry to correct value |
| 2143 |
entry->value() = (row == column ? average_diagonal : 0.); |
| 2144 |
} |
| 2145 |
} |
| 2146 |
|
| 2147 |
// take care of vector |
| 2148 |
if (use_vectors == true) |
| 2149 |
{ |
| 2150 |
for (size_type q = 0; q != lines[distribute[row]].entries.size(); |
| 2151 |
++q) |
| 2152 |
vec(lines[distribute[row]].entries[q].first) += |
| 2153 |
(vec(row) * lines[distribute[row]].entries[q].second); |
| 2154 |
|
| 2155 |
vec(lines[distribute[row]].index) = 0.; |
| 2156 |
} |
| 2157 |
} |
| 2158 |
} |
| 2159 |
} |
| 2160 |
|
| 2161 |
|
| 2162 |
|
| 2163 |
template <typename number> |
| 2164 |
template <typename BlockVectorType> |
| 2165 |
void |
| 2166 |
AffineConstraints<number>::condense(BlockSparseMatrix<number> &uncondensed, |
| 2167 |
BlockVectorType &vec) const |
| 2168 |
{ |
| 2169 |
// check whether we work on real vectors or we just used a dummy when |
| 2170 |
// calling the other function above. |
| 2171 |
const bool use_vectors = vec.n_blocks() == 0 ? false : true; |
| 2172 |
|
| 2173 |
const size_type blocks = uncondensed.n_block_rows(); |
| 2174 |
|
| 2175 |
const BlockSparsityPattern &sparsity = uncondensed.get_sparsity_pattern(); |
| 2176 |
|
| 2177 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 2178 |
Assert(sparsity.is_compressed() == true, ExcMatrixNotClosed()); |
| 2179 |
Assert(sparsity.n_rows() == sparsity.n_cols(), ExcNotQuadratic()); |
| 2180 |
Assert(sparsity.n_block_rows() == sparsity.n_block_cols(), ExcNotQuadratic()); |
| 2181 |
Assert(sparsity.n_block_rows() == sparsity.n_block_cols(), ExcNotQuadratic()); |
| 2182 |
Assert(sparsity.get_column_indices() == sparsity.get_row_indices(), |
| 2183 |
ExcNotQuadratic()); |
| 2184 |
|
| 2185 |
if (use_vectors == true) |
| 2186 |
{ |
| 2187 |
AssertDimension(vec.size(), sparsity.n_rows()); |
| 2188 |
AssertDimension(vec.n_blocks(), sparsity.n_block_rows()); |
| 2189 |
} |
| 2190 |
|
| 2191 |
number average_diagonal = 0.; |
| 2192 |
for (size_type b = 0; b < uncondensed.n_block_rows(); ++b) |
| 2193 |
for (size_type i = 0; i < uncondensed.block(b, b).m(); ++i) |
| 2194 |
average_diagonal += std::abs(uncondensed.block(b, b).diag_element(i)); |
| 2195 |
average_diagonal /= uncondensed.m(); |
| 2196 |
|
| 2197 |
const BlockIndices &index_mapping = sparsity.get_column_indices(); |
| 2198 |
|
| 2199 |
// store for each index whether it must be distributed or not. If entry |
| 2200 |
// is numbers::invalid_size_type, no distribution is necessary. |
| 2201 |
// otherwise, the number states which line in the constraints object |
| 2202 |
// handles this index |
| 2203 |
std::vector<size_type> distribute(sparsity.n_rows(), |
| 2204 |
numbers::invalid_size_type); |
| 2205 |
|
| 2206 |
for (size_type c = 0; c < lines.size(); ++c) |
| 2207 |
distribute[lines[c].index] = c; |
| 2208 |
|
| 2209 |
const size_type n_rows = sparsity.n_rows(); |
| 2210 |
for (size_type row = 0; row < n_rows; ++row) |
| 2211 |
{ |
| 2212 |
// get index of this row within the blocks |
| 2213 |
const std::pair<size_type, size_type> block_index = |
| 2214 |
index_mapping.global_to_local(row); |
| 2215 |
const size_type block_row = block_index.first; |
| 2216 |
|
| 2217 |
if (distribute[row] == numbers::invalid_size_type) |
| 2218 |
// regular line. loop over all columns and see whether this column |
| 2219 |
// must be distributed |
| 2220 |
{ |
| 2221 |
// to loop over all entries in this row, we have to loop over all |
| 2222 |
// blocks in this blockrow and the corresponding row therein |
| 2223 |
for (size_type block_col = 0; block_col < blocks; ++block_col) |
| 2224 |
{ |
| 2225 |
for (typename SparseMatrix<number>::iterator entry = |
| 2226 |
uncondensed.block(block_row, block_col) |
| 2227 |
.begin(block_index.second); |
| 2228 |
entry != uncondensed.block(block_row, block_col) |
| 2229 |
.end(block_index.second); |
| 2230 |
++entry) |
| 2231 |
{ |
| 2232 |
const size_type global_col = |
| 2233 |
index_mapping.local_to_global(block_col, entry->column()); |
| 2234 |
|
| 2235 |
if (distribute[global_col] != numbers::invalid_size_type) |
| 2236 |
// distribute entry at regular row @p row and irregular |
| 2237 |
// column global_col; set old entry to zero |
| 2238 |
{ |
| 2239 |
const number old_value = entry->value(); |
| 2240 |
|
| 2241 |
for (size_type q = 0; |
| 2242 |
q != lines[distribute[global_col]].entries.size(); |
| 2243 |
++q) |
| 2244 |
uncondensed.add( |
| 2245 |
row, |
| 2246 |
lines[distribute[global_col]].entries[q].first, |
| 2247 |
old_value * |
| 2248 |
lines[distribute[global_col]].entries[q].second); |
| 2249 |
|
| 2250 |
// need to subtract this element from the vector. |
| 2251 |
// this corresponds to an explicit elimination in the |
| 2252 |
// respective row of the inhomogeneous constraint in |
| 2253 |
// the matrix with Gauss elimination |
| 2254 |
if (use_vectors == true) |
| 2255 |
vec(row) -= static_cast<number>(entry->value()) * |
| 2256 |
lines[distribute[global_col]].inhomogeneity; |
| 2257 |
|
| 2258 |
entry->value() = 0.; |
| 2259 |
} |
| 2260 |
} |
| 2261 |
} |
| 2262 |
} |
| 2263 |
else |
| 2264 |
{ |
| 2265 |
// row must be distributed. split the whole row into the chunks |
| 2266 |
// defined by the blocks |
| 2267 |
for (size_type block_col = 0; block_col < blocks; ++block_col) |
| 2268 |
{ |
| 2269 |
for (typename SparseMatrix<number>::iterator entry = |
| 2270 |
uncondensed.block(block_row, block_col) |
| 2271 |
.begin(block_index.second); |
| 2272 |
entry != uncondensed.block(block_row, block_col) |
| 2273 |
.end(block_index.second); |
| 2274 |
++entry) |
| 2275 |
{ |
| 2276 |
const size_type global_col = |
| 2277 |
index_mapping.local_to_global(block_col, entry->column()); |
| 2278 |
|
| 2279 |
if (distribute[global_col] == numbers::invalid_size_type) |
| 2280 |
// distribute entry at irregular row @p row and regular |
| 2281 |
// column global_col. set old entry to zero |
| 2282 |
{ |
| 2283 |
const number old_value = entry->value(); |
| 2284 |
|
| 2285 |
for (size_type q = 0; |
| 2286 |
q != lines[distribute[row]].entries.size(); |
| 2287 |
++q) |
| 2288 |
uncondensed.add( |
| 2289 |
lines[distribute[row]].entries[q].first, |
| 2290 |
global_col, |
| 2291 |
old_value * lines[distribute[row]].entries[q].second); |
| 2292 |
|
| 2293 |
entry->value() = 0.; |
| 2294 |
} |
| 2295 |
else |
| 2296 |
// distribute entry at irregular row @p row and |
| 2297 |
// irregular column @p global_col set old entry to one |
| 2298 |
// if on main diagonal, zero otherwise |
| 2299 |
{ |
| 2300 |
const number old_value = entry->value(); |
| 2301 |
|
| 2302 |
for (size_type p = 0; |
| 2303 |
p != lines[distribute[row]].entries.size(); |
| 2304 |
++p) |
| 2305 |
{ |
| 2306 |
for (size_type q = 0; |
| 2307 |
q != |
| 2308 |
lines[distribute[global_col]].entries.size(); |
| 2309 |
++q) |
| 2310 |
uncondensed.add( |
| 2311 |
lines[distribute[row]].entries[p].first, |
| 2312 |
lines[distribute[global_col]].entries[q].first, |
| 2313 |
old_value * |
| 2314 |
lines[distribute[row]].entries[p].second * |
| 2315 |
lines[distribute[global_col]] |
| 2316 |
.entries[q] |
| 2317 |
.second); |
| 2318 |
|
| 2319 |
if (use_vectors == true) |
| 2320 |
vec(lines[distribute[row]].entries[p].first) -= |
| 2321 |
old_value * |
| 2322 |
lines[distribute[row]].entries[p].second * |
| 2323 |
lines[distribute[global_col]].inhomogeneity; |
| 2324 |
} |
| 2325 |
|
| 2326 |
entry->value() = |
| 2327 |
(row == global_col ? average_diagonal : 0.); |
| 2328 |
} |
| 2329 |
} |
| 2330 |
} |
| 2331 |
|
| 2332 |
// take care of vector |
| 2333 |
if (use_vectors == true) |
| 2334 |
{ |
| 2335 |
for (size_type q = 0; q != lines[distribute[row]].entries.size(); |
| 2336 |
++q) |
| 2337 |
vec(lines[distribute[row]].entries[q].first) += |
| 2338 |
(vec(row) * lines[distribute[row]].entries[q].second); |
| 2339 |
|
| 2340 |
vec(lines[distribute[row]].index) = 0.; |
| 2341 |
} |
| 2342 |
} |
| 2343 |
} |
| 2344 |
} |
| 2345 |
|
| 2346 |
|
| 2347 |
|
| 2348 |
// TODO: I'm sure the following could be made more elegant by using a bit of |
| 2349 |
// introspection using static member variables of the various vector |
| 2350 |
// classes to dispatch between the different functions, rather than using |
| 2351 |
// knowledge of the individual types |
| 2352 |
|
| 2353 |
// number of functions to select the right implementation for set_zero(). |
| 2354 |
namespace internal |
| 2355 |
{ |
| 2356 |
namespace AffineConstraintsImplementation |
| 2357 |
{ |
| 2358 |
using size_type = types::global_dof_index; |
| 2359 |
|
| 2360 |
template <typename VectorType> |
| 2361 |
void |
| 2362 |
set_zero_parallel(const std::vector<size_type> &cm, |
| 2363 |
VectorType &vec, |
| 2364 |
size_type shift = 0) |
| 2365 |
{ |
| 2366 |
Assert(!vec.has_ghost_elements(), ExcInternalError()); |
| 2367 |
IndexSet locally_owned = vec.locally_owned_elements(); |
| 2368 |
for (const size_type index : cm) |
| 2369 |
{ |
| 2370 |
// If shift>0 then we are working on a part of a BlockVector |
| 2371 |
// so vec(i) is actually the global entry i+shift. |
| 2372 |
// We first make sure the line falls into the range of vec, |
| 2373 |
// then check if is part of the local part of the vector, before |
| 2374 |
// finally setting value to 0. |
| 2375 |
if (index < shift) |
| 2376 |
continue; |
| 2377 |
const size_type idx = index - shift; |
| 2378 |
if (idx < vec.size() && locally_owned.is_element(idx)) |
| 2379 |
internal::ElementAccess<VectorType>::set(0., idx, vec); |
| 2380 |
} |
| 2381 |
} |
| 2382 |
|
| 2383 |
template <typename number> |
| 2384 |
void |
| 2385 |
set_zero_parallel(const std::vector<size_type> &cm, |
| 2386 |
LinearAlgebra::distributed::Vector<number> &vec, |
| 2387 |
size_type shift = 0) |
| 2388 |
{ |
| 2389 |
for (const size_type index : cm) |
| 2390 |
{ |
| 2391 |
// If shift>0 then we are working on a part of a BlockVector |
| 2392 |
// so vec(i) is actually the global entry i+shift. |
| 2393 |
// We first make sure the line falls into the range of vec, |
| 2394 |
// then check if is part of the local part of the vector, before |
| 2395 |
// finally setting the value to 0. |
| 2396 |
if (index < shift) |
| 2397 |
continue; |
| 2398 |
const size_type idx = index - shift; |
| 2399 |
if (vec.in_local_range(idx)) |
| 2400 |
vec(idx) = 0.; |
| 2401 |
} |
| 2402 |
vec.zero_out_ghost_values(); |
| 2403 |
} |
| 2404 |
|
| 2405 |
template <typename number> |
| 2406 |
void |
| 2407 |
set_zero_parallel( |
| 2408 |
const std::vector<size_type> &cm, |
| 2409 |
LinearAlgebra::distributed::Vector<number, MemorySpace::Default> &vec, |
| 2410 |
size_type shift = 0) |
| 2411 |
{ |
| 2412 |
Assert(shift == 0, ExcNotImplemented()); |
| 2413 |
std::vector<size_type> constrained_local_dofs_host; |
| 2414 |
constrained_local_dofs_host.reserve(cm.size()); |
| 2415 |
|
| 2416 |
for (const auto global_index : cm) |
| 2417 |
if (vec.in_local_range(global_index)) |
| 2418 |
constrained_local_dofs_host.push_back( |
| 2419 |
vec.get_partitioner()->global_to_local(global_index)); |
| 2420 |
|
| 2421 |
const int n_constraints = constrained_local_dofs_host.size(); |
| 2422 |
Kokkos::View<size_type *, MemorySpace::Default::kokkos_space> |
| 2423 |
constrained_local_dofs_device( |
| 2424 |
Kokkos::view_alloc(Kokkos::WithoutInitializing, |
| 2425 |
"constrained_local_dofs_device"), |
| 2426 |
n_constraints); |
| 2427 |
Kokkos::deep_copy(constrained_local_dofs_device, |
| 2428 |
Kokkos::View<size_type *, Kokkos::HostSpace>( |
| 2429 |
constrained_local_dofs_host.data(), |
| 2430 |
constrained_local_dofs_host.size())); |
| 2431 |
|
| 2432 |
using ExecutionSpace = |
| 2433 |
MemorySpace::Default::kokkos_space::execution_space; |
| 2434 |
ExecutionSpace exec; |
| 2435 |
auto *local_values = vec.get_values(); |
| 2436 |
Kokkos::parallel_for( |
| 2437 |
"dealii::set_zero_parallel", |
| 2438 |
Kokkos::RangePolicy<ExecutionSpace>(exec, 0, n_constraints), |
| 2439 |
KOKKOS_LAMBDA(int i) { |
| 2440 |
local_values[constrained_local_dofs_device[i]] = 0; |
| 2441 |
}); |
| 2442 |
|
| 2443 |
vec.zero_out_ghost_values(); |
| 2444 |
} |
| 2445 |
|
| 2446 |
template <typename VectorType> |
| 2447 |
void |
| 2448 |
set_zero_in_parallel(const std::vector<size_type> &cm, |
| 2449 |
VectorType &vec, |
| 2450 |
std::bool_constant<false>) |
| 2451 |
{ |
| 2452 |
set_zero_parallel(cm, vec, 0); |
| 2453 |
} |
| 2454 |
|
| 2455 |
// in parallel for BlockVectors |
| 2456 |
template <typename VectorType> |
| 2457 |
void |
| 2458 |
set_zero_in_parallel(const std::vector<size_type> &cm, |
| 2459 |
VectorType &vec, |
| 2460 |
std::bool_constant<true>) |
| 2461 |
{ |
| 2462 |
size_type start_shift = 0; |
| 2463 |
for (size_type j = 0; j < vec.n_blocks(); ++j) |
| 2464 |
{ |
| 2465 |
set_zero_parallel(cm, vec.block(j), start_shift); |
| 2466 |
start_shift += vec.block(j).size(); |
| 2467 |
} |
| 2468 |
} |
| 2469 |
|
| 2470 |
template <typename VectorType> |
| 2471 |
void |
| 2472 |
set_zero_serial(const std::vector<size_type> &cm, VectorType &vec) |
| 2473 |
{ |
| 2474 |
for (const size_type index : cm) |
| 2475 |
vec(index) = 0.; |
| 2476 |
} |
| 2477 |
|
| 2478 |
template <typename VectorType> |
| 2479 |
void |
| 2480 |
set_zero_all(const std::vector<size_type> &cm, VectorType &vec) |
| 2481 |
{ |
| 2482 |
set_zero_in_parallel<VectorType>( |
| 2483 |
cm, vec, std::bool_constant<IsBlockVector<VectorType>::value>()); |
| 2484 |
vec.compress(VectorOperation::insert); |
| 2485 |
} |
| 2486 |
|
| 2487 |
template <class T> |
| 2488 |
void |
| 2489 |
set_zero_all(const std::vector<size_type> &cm, dealii::Vector<T> &vec) |
| 2490 |
{ |
| 2491 |
set_zero_serial(cm, vec); |
| 2492 |
} |
| 2493 |
|
| 2494 |
template <class T> |
| 2495 |
void |
| 2496 |
set_zero_all(const std::vector<size_type> &cm, dealii::BlockVector<T> &vec) |
| 2497 |
{ |
| 2498 |
set_zero_serial(cm, vec); |
| 2499 |
} |
| 2500 |
} // namespace AffineConstraintsImplementation |
| 2501 |
} // namespace internal |
| 2502 |
|
| 2503 |
|
| 2504 |
|
| 2505 |
template <typename number> |
| 2506 |
template <typename VectorType> |
| 2507 |
void |
| 2508 |
AffineConstraints<number>::distribute_local_to_global( |
| 2509 |
const Vector<number> &local_vector, |
| 2510 |
const std::vector<size_type> &local_dof_indices, |
| 2511 |
VectorType &global_vector, |
| 2512 |
const FullMatrix<number> &local_matrix) const |
| 2513 |
{ |
| 2514 |
distribute_local_to_global(local_vector, |
| 2515 |
local_dof_indices, |
| 2516 |
local_dof_indices, |
| 2517 |
global_vector, |
| 2518 |
local_matrix, |
| 2519 |
true); |
| 2520 |
} |
| 2521 |
|
| 2522 |
|
| 2523 |
|
| 2524 |
template <typename number> |
| 2525 |
template <typename VectorType> |
| 2526 |
void |
| 2527 |
AffineConstraints<number>::distribute_local_to_global( |
| 2528 |
const Vector<number> &local_vector, |
| 2529 |
const std::vector<size_type> &local_dof_indices_row, |
| 2530 |
const std::vector<size_type> &local_dof_indices_col, |
| 2531 |
VectorType &global_vector, |
| 2532 |
const FullMatrix<number> &local_matrix, |
| 2533 |
bool diagonal) const |
| 2534 |
{ |
| 2535 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 2536 |
AssertDimension(local_vector.size(), local_dof_indices_row.size()); |
| 2537 |
AssertDimension(local_matrix.m(), local_dof_indices_row.size()); |
| 2538 |
AssertDimension(local_matrix.n(), local_dof_indices_col.size()); |
| 2539 |
Assert(global_vector.has_ghost_elements() == false, ExcGhostsPresent()); |
| 2540 |
|
| 2541 |
// diagonal checks if we have only one index set (if both are equal |
| 2542 |
// diagonal should be set to true). |
| 2543 |
// If true we do both, assembly of the right hand side (next lines) |
| 2544 |
// and (see further below) modifications of the right hand side |
| 2545 |
// according to the inhomogeneous constraints. |
| 2546 |
// Otherwise we only modify the right hand side according to |
| 2547 |
// local_matrix and the inhomogeneous constraints, and omit the vector add. |
| 2548 |
|
| 2549 |
const size_type m_local_dofs = local_dof_indices_row.size(); |
| 2550 |
const size_type n_local_dofs = local_dof_indices_col.size(); |
| 2551 |
if (lines.empty()) |
| 2552 |
{ |
| 2553 |
if (diagonal) |
| 2554 |
global_vector.add(local_dof_indices_row, local_vector); |
| 2555 |
} |
| 2556 |
else |
| 2557 |
for (size_type i = 0; i < n_local_dofs; ++i) |
| 2558 |
{ |
| 2559 |
// check whether the current index is |
| 2560 |
// constrained. if not, just write the entry |
| 2561 |
// into the vector. otherwise, need to resolve |
| 2562 |
// the constraint |
| 2563 |
if (is_constrained(local_dof_indices_col[i]) == false) |
| 2564 |
{ |
| 2565 |
if (diagonal) |
| 2566 |
global_vector(local_dof_indices_row[i]) += local_vector(i); |
| 2567 |
continue; |
| 2568 |
} |
| 2569 |
|
| 2570 |
// find the constraint line to the given |
| 2571 |
// global dof index |
| 2572 |
const size_type line_index = |
| 2573 |
calculate_line_index(local_dof_indices_col[i]); |
| 2574 |
AssertIndexRange(line_index, lines_cache.size()); |
| 2575 |
AssertIndexRange(lines_cache[line_index], lines.size()); |
| 2576 |
const ConstraintLine &position = lines[lines_cache[line_index]]; |
| 2577 |
|
| 2578 |
// Gauss elimination of the matrix columns with the inhomogeneity. |
| 2579 |
// Go through them one by one and again check whether they are |
| 2580 |
// constrained. If so, distribute the constraint |
| 2581 |
const auto val = position.inhomogeneity; |
| 2582 |
if (val != number(0.)) |
| 2583 |
for (size_type j = 0; j < m_local_dofs; ++j) |
| 2584 |
{ |
| 2585 |
if (is_constrained(local_dof_indices_row[j]) == false) |
| 2586 |
{ |
| 2587 |
global_vector(local_dof_indices_row[j]) -= |
| 2588 |
val * local_matrix(j, i); |
| 2589 |
continue; |
| 2590 |
} |
| 2591 |
|
| 2592 |
const number matrix_entry = local_matrix(j, i); |
| 2593 |
|
| 2594 |
if (matrix_entry == number()) |
| 2595 |
continue; |
| 2596 |
|
| 2597 |
const ConstraintLine &position_j = |
| 2598 |
lines[lines_cache[calculate_line_index( |
| 2599 |
local_dof_indices_row[j])]]; |
| 2600 |
|
| 2601 |
for (size_type q = 0; q < position_j.entries.size(); ++q) |
| 2602 |
{ |
| 2603 |
Assert(!(!local_lines.size() || |
| 2604 |
local_lines.is_element( |
| 2605 |
position_j.entries[q].first)) || |
| 2606 |
is_constrained(position_j.entries[q].first) == false, |
| 2607 |
ExcMessage("Tried to distribute to a fixed dof.")); |
| 2608 |
global_vector(position_j.entries[q].first) -= |
| 2609 |
val * position_j.entries[q].second * matrix_entry; |
| 2610 |
} |
| 2611 |
} |
| 2612 |
|
| 2613 |
// now distribute the constraint, |
| 2614 |
// but make sure we don't touch |
| 2615 |
// the entries of fixed dofs |
| 2616 |
if (diagonal) |
| 2617 |
{ |
| 2618 |
for (size_type j = 0; j < position.entries.size(); ++j) |
| 2619 |
{ |
| 2620 |
Assert(!(!local_lines.size() || |
| 2621 |
local_lines.is_element(position.entries[j].first)) || |
| 2622 |
is_constrained(position.entries[j].first) == false, |
| 2623 |
ExcMessage("Tried to distribute to a fixed dof.")); |
| 2624 |
global_vector(position.entries[j].first) += |
| 2625 |
local_vector(i) * position.entries[j].second; |
| 2626 |
} |
| 2627 |
} |
| 2628 |
} |
| 2629 |
} |
| 2630 |
|
| 2631 |
namespace internal |
| 2632 |
{ |
| 2633 |
// create an output vector that consists of the input vector's locally owned |
| 2634 |
// elements plus some ghost elements that need to be imported from elsewhere |
| 2635 |
// |
| 2636 |
// this is an operation that is different for all vector types and so we |
| 2637 |
// need a few overloads |
| 2638 |
#ifdef DEAL_II_WITH_TRILINOS |
| 2639 |
inline void |
| 2640 |
import_vector_with_ghost_elements( |
| 2641 |
const TrilinosWrappers::MPI::Vector &vec, |
| 2642 |
const IndexSet & /*locally_owned_elements*/, |
| 2643 |
const IndexSet &needed_elements, |
| 2644 |
TrilinosWrappers::MPI::Vector &output, |
| 2645 |
const std::bool_constant<false> /*is_block_vector*/) |
| 2646 |
{ |
| 2647 |
Assert(!vec.has_ghost_elements(), ExcGhostsPresent()); |
| 2648 |
# ifdef DEAL_II_WITH_MPI |
| 2649 |
const Epetra_MpiComm *mpi_comm = |
| 2650 |
dynamic_cast<const Epetra_MpiComm *>(&vec.trilinos_vector().Comm()); |
| 2651 |
|
| 2652 |
Assert(mpi_comm != nullptr, ExcInternalError()); |
| 2653 |
output.reinit(needed_elements, mpi_comm->GetMpiComm()); |
| 2654 |
# else |
| 2655 |
output.reinit(needed_elements, MPI_COMM_SELF); |
| 2656 |
# endif |
| 2657 |
output = vec; |
| 2658 |
} |
| 2659 |
#endif |
| 2660 |
|
| 2661 |
|
| 2662 |
|
| 2663 |
#ifdef DEAL_II_TRILINOS_WITH_TPETRA |
| 2664 |
template <typename Number, typename MemorySpace> |
| 2665 |
inline void |
| 2666 |
import_vector_with_ghost_elements( |
| 2667 |
const LinearAlgebra::TpetraWrappers::Vector<Number, MemorySpace> &vec, |
| 2668 |
const IndexSet &locally_owned_elements, |
| 2669 |
const IndexSet &needed_elements, |
| 2670 |
LinearAlgebra::TpetraWrappers::Vector<Number, MemorySpace> &output, |
| 2671 |
const std::bool_constant<false> /*is_block_vector*/) |
| 2672 |
{ |
| 2673 |
Assert(!vec.has_ghost_elements(), ExcGhostsPresent()); |
| 2674 |
IndexSet parallel_partitioner = locally_owned_elements; |
| 2675 |
parallel_partitioner.add_indices(needed_elements); |
| 2676 |
|
| 2677 |
const MPI_Comm mpi_comm = Utilities::Trilinos::teuchos_comm_to_mpi_comm( |
| 2678 |
vec.trilinos_vector().getMap()->getComm()); |
| 2679 |
|
| 2680 |
output.reinit(locally_owned_elements, needed_elements, mpi_comm); |
| 2681 |
|
| 2682 |
output = vec; |
| 2683 |
} |
| 2684 |
#endif // DEAL_II_TRILINOS_WITH_TPETRA |
| 2685 |
|
| 2686 |
|
| 2687 |
|
| 2688 |
#ifdef DEAL_II_WITH_PETSC |
| 2689 |
inline void |
| 2690 |
import_vector_with_ghost_elements( |
| 2691 |
const PETScWrappers::MPI::Vector &vec, |
| 2692 |
const IndexSet &locally_owned_elements, |
| 2693 |
const IndexSet &needed_elements, |
| 2694 |
PETScWrappers::MPI::Vector &output, |
| 2695 |
const std::bool_constant<false> /*is_block_vector*/) |
| 2696 |
{ |
| 2697 |
output.reinit(locally_owned_elements, |
| 2698 |
needed_elements, |
| 2699 |
vec.get_mpi_communicator()); |
| 2700 |
output = vec; |
| 2701 |
} |
| 2702 |
#endif |
| 2703 |
|
| 2704 |
|
| 2705 |
|
| 2706 |
template <typename number> |
| 2707 |
void |
| 2708 |
import_vector_with_ghost_elements( |
| 2709 |
const LinearAlgebra::distributed::Vector<number> &vec, |
| 2710 |
const IndexSet &locally_owned_elements, |
| 2711 |
const IndexSet &needed_elements, |
| 2712 |
LinearAlgebra::distributed::Vector<number> &output, |
| 2713 |
const std::bool_constant<false> /*is_block_vector*/) |
| 2714 |
{ |
| 2715 |
// TODO: the in vector might already have all elements. need to find a |
| 2716 |
// way to efficiently avoid the copy then |
| 2717 |
const_cast<LinearAlgebra::distributed::Vector<number> &>(vec) |
| 2718 |
.zero_out_ghost_values(); |
| 2719 |
output.reinit(locally_owned_elements, |
| 2720 |
needed_elements, |
| 2721 |
vec.get_mpi_communicator()); |
| 2722 |
output = vec; |
| 2723 |
output.update_ghost_values(); |
| 2724 |
} |
| 2725 |
|
| 2726 |
|
| 2727 |
|
| 2728 |
// all other vector non-block vector types are sequential and we should |
| 2729 |
// not have this function called at all -- so throw an exception |
| 2730 |
template <typename Vector> |
| 2731 |
void |
| 2732 |
import_vector_with_ghost_elements( |
| 2733 |
const Vector & /*vec*/, |
| 2734 |
const IndexSet & /*locally_owned_elements*/, |
| 2735 |
const IndexSet & /*needed_elements*/, |
| 2736 |
Vector & /*output*/, |
| 2737 |
const std::bool_constant<false> /*is_block_vector*/) |
| 2738 |
{ |
| 2739 |
Assert(false, ExcMessage("We shouldn't even get here!")); |
| 2740 |
} |
| 2741 |
|
| 2742 |
// for block vectors, simply dispatch to the individual blocks |
| 2743 |
template <typename VectorType> |
| 2744 |
void |
| 2745 |
import_vector_with_ghost_elements( |
| 2746 |
const VectorType &vec, |
| 2747 |
const IndexSet &locally_owned_elements, |
| 2748 |
const IndexSet &needed_elements, |
| 2749 |
VectorType &output, |
| 2750 |
const std::bool_constant<true> /*is_block_vector*/) |
| 2751 |
{ |
| 2752 |
output.reinit(vec.n_blocks()); |
| 2753 |
|
| 2754 |
types::global_dof_index block_start = 0; |
| 2755 |
for (unsigned int b = 0; b < vec.n_blocks(); ++b) |
| 2756 |
{ |
| 2757 |
import_vector_with_ghost_elements( |
| 2758 |
vec.block(b), |
| 2759 |
locally_owned_elements.get_view(block_start, |
| 2760 |
block_start + vec.block(b).size()), |
| 2761 |
needed_elements.get_view(block_start, |
| 2762 |
block_start + vec.block(b).size()), |
| 2763 |
output.block(b), |
| 2764 |
std::bool_constant<false>()); |
| 2765 |
block_start += vec.block(b).size(); |
| 2766 |
} |
| 2767 |
|
| 2768 |
output.collect_sizes(); |
| 2769 |
} |
| 2770 |
} // namespace internal |
| 2771 |
|
| 2772 |
|
| 2773 |
namespace internal |
| 2774 |
{ |
| 2775 |
namespace AffineConstraints |
| 2776 |
{ |
| 2777 |
template <typename VectorType, |
| 2778 |
std::enable_if_t<internal::is_block_vector<VectorType> == false> |
| 2779 |
* = nullptr> |
| 2780 |
MPI_Comm |
| 2781 |
get_mpi_communicator(const VectorType &vec) |
| 2782 |
{ |
| 2783 |
return vec.get_mpi_communicator(); |
| 2784 |
} |
| 2785 |
|
| 2786 |
|
| 2787 |
|
| 2788 |
template <typename VectorType, |
| 2789 |
std::enable_if_t<internal::is_block_vector<VectorType> == true> |
| 2790 |
* = nullptr> |
| 2791 |
MPI_Comm |
| 2792 |
get_mpi_communicator(const VectorType &vec) |
| 2793 |
{ |
| 2794 |
Assert(vec.n_blocks() > 0, ExcInternalError()); |
| 2795 |
return vec.block(0).get_mpi_communicator(); |
| 2796 |
} |
| 2797 |
|
| 2798 |
|
| 2799 |
|
| 2800 |
template <typename V> |
| 2801 |
using is_compressed_op = decltype(std::declval<V>().is_compressed()); |
| 2802 |
} // namespace AffineConstraints |
| 2803 |
} // namespace internal |
| 2804 |
|
| 2805 |
|
| 2806 |
|
| 2807 |
template <typename number> |
| 2808 |
template <typename VectorType> |
| 2809 |
void |
| 2810 |
AffineConstraints<number>::distribute(VectorType &vec) const |
| 2811 |
{ |
| 2812 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 2813 |
|
| 2814 |
if constexpr (internal::is_supported_operation< |
| 2815 |
internal::AffineConstraints::is_compressed_op, |
| 2816 |
VectorType>) |
| 2817 |
Assert( |
| 2818 |
vec.is_compressed(), |
| 2819 |
ExcMessage( |
| 2820 |
"To distribute a parallel vector, the vector must " |
| 2821 |
"not contain any previously written-to or added-to entries that " |
| 2822 |
"have not already been sent to these entries' owners. But " |
| 2823 |
"the vector you passed in has such entries. You need to call " |
| 2824 |
"compress() on the vector first, before passing it to this function.")); |
| 2825 |
|
| 2826 |
// if the vector type supports parallel storage and if the vector actually |
| 2827 |
// does store only part of the vector, distributing is slightly more |
| 2828 |
// complicated. we might be able to skip the complicated part if one |
| 2829 |
// processor owns everything and pretend that this is a sequential vector, |
| 2830 |
// but it is difficult for the other processors to know whether they should |
| 2831 |
// not do anything or if other processors will create a temporary vector, |
| 2832 |
// exchange data (requiring communication, maybe even with the processors |
| 2833 |
// that do not own anything because of that particular parallel model), and |
| 2834 |
// call compress() finally. the first case here is for the complicated case, |
| 2835 |
// the last else is for the simple case (sequential vector) |
| 2836 |
const IndexSet vec_owned_elements = vec.locally_owned_elements(); |
| 2837 |
|
| 2838 |
if constexpr (dealii::is_serial_vector<VectorType>::value == false) |
| 2839 |
{ |
| 2840 |
// First check whether there are any constraints at all. If |
| 2841 |
// there aren't, then there is nothing we need to do and we can |
| 2842 |
// save the considerable effort below to communicate information |
| 2843 |
// -- this is for sure the case on the first (not yet adaptively |
| 2844 |
// refined) level of computations, for example: |
| 2845 |
if (Utilities::MPI::sum(lines.size(), |
| 2846 |
internal::AffineConstraints::get_mpi_communicator( |
| 2847 |
vec)) > 0) |
| 2848 |
{ |
| 2849 |
// We are looking at parallel vectors here. Of course, they could be |
| 2850 |
// used in sequential contexts, but at least if we know that we are |
| 2851 |
// in a parallel context with more than one process, we should check |
| 2852 |
// that the user of this class provided index sets to the constructor |
| 2853 |
// or the reinit() functions: |
| 2854 |
if (Utilities::MPI::n_mpi_processes( |
| 2855 |
internal::AffineConstraints::get_mpi_communicator(vec)) > 1) |
| 2856 |
{ |
| 2857 |
Assert(local_lines != IndexSet(), |
| 2858 |
ExcMessage( |
| 2859 |
"You are using the AffineConstraints class in a " |
| 2860 |
"program that is running in parallel. In this case, " |
| 2861 |
"it is inefficient to store all constraints: This class " |
| 2862 |
"should really only store constraints for those degrees " |
| 2863 |
"of freedom that are locally relevant. It is considered " |
| 2864 |
"a bug not to do that. Please call either the " |
| 2865 |
"constructor of this class, or one of its reinit() " |
| 2866 |
"functions that provide this class with index sets " |
| 2867 |
"of the locally owned and the locally relevant " |
| 2868 |
"degrees of freedom.")); |
| 2869 |
|
| 2870 |
Assert(vec_owned_elements.size() == locally_owned_dofs.size(), |
| 2871 |
ExcMessage("You have previously initialized this " |
| 2872 |
"AffineConstraints object with an index set " |
| 2873 |
"that stated that vectors have size " + |
| 2874 |
std::to_string(locally_owned_dofs.size()) + |
| 2875 |
" entries, but you are now calling " |
| 2876 |
"AffineConstraints::distribute() with a vector " |
| 2877 |
"of size " + |
| 2878 |
std::to_string(vec_owned_elements.size()) + |
| 2879 |
".")); |
| 2880 |
} |
| 2881 |
|
| 2882 |
|
| 2883 |
// Check that the set of indices we will import is a superset of |
| 2884 |
// the locally-owned ones. This *should* be the case if, as one |
| 2885 |
// would expect, the AffineConstraint object was initialized |
| 2886 |
// with a locally-relevant index set that is indeed a superset |
| 2887 |
// of the locally-owned indices. But you never know what people |
| 2888 |
// pass as arguments... |
| 2889 |
if constexpr (running_in_debug_mode()) |
| 2890 |
{ |
| 2891 |
if (needed_elements_for_distribute != IndexSet()) |
| 2892 |
{ |
| 2893 |
Assert(vec_owned_elements.size() == |
| 2894 |
needed_elements_for_distribute.size(), |
| 2895 |
ExcMessage( |
| 2896 |
"You have previously initialized this " |
| 2897 |
"AffineConstraints object with an index set " |
| 2898 |
"that stated that vectors have size " + |
| 2899 |
std::to_string(locally_owned_dofs.size()) + |
| 2900 |
" entries, but you are now calling " |
| 2901 |
"AffineConstraints::distribute() with a vector " |
| 2902 |
"of size " + |
| 2903 |
std::to_string(vec_owned_elements.size()) + ".")); |
| 2904 |
|
| 2905 |
for (const auto i : vec_owned_elements) |
| 2906 |
Assert(needed_elements_for_distribute.is_element(i), |
| 2907 |
ExcInternalError()); |
| 2908 |
} |
| 2909 |
} |
| 2910 |
|
| 2911 |
VectorType ghosted_vector; |
| 2912 |
|
| 2913 |
// It is possible that the user is using a parallel vector type, |
| 2914 |
// but is running a non-parallel program (for example, step-31 |
| 2915 |
// does this). In this case, it is (perhaps?) not a bug to not |
| 2916 |
// set an IndexSet for the local_lines and the |
| 2917 |
// locally_owned_lines -- they are simply both empty sets in |
| 2918 |
// that case. If that is so, we could just assign |
| 2919 |
// 'ghosted_vector=vec;'. But this is dangerous. Not having set |
| 2920 |
// index sets could also have been a bug, and we would get |
| 2921 |
// downstream errors about accessing elements that are not |
| 2922 |
// locally available. Rather, if no index sets were provided, |
| 2923 |
// simply import the *entire* vector. |
| 2924 |
if (local_lines != IndexSet()) |
| 2925 |
{ |
| 2926 |
Assert(vec_owned_elements.size() == |
| 2927 |
needed_elements_for_distribute.size(), |
| 2928 |
ExcInternalError()); |
| 2929 |
|
| 2930 |
Assert(needed_elements_for_distribute != IndexSet(), |
| 2931 |
ExcInternalError()); |
| 2932 |
internal::import_vector_with_ghost_elements( |
| 2933 |
vec, |
| 2934 |
vec_owned_elements, |
| 2935 |
needed_elements_for_distribute, |
| 2936 |
ghosted_vector, |
| 2937 |
std::bool_constant<IsBlockVector<VectorType>::value>()); |
| 2938 |
} |
| 2939 |
else |
| 2940 |
{ |
| 2941 |
Assert(needed_elements_for_distribute == IndexSet(), |
| 2942 |
ExcInternalError()); |
| 2943 |
|
| 2944 |
// TODO: We should really consider it a bug if this parallel |
| 2945 |
// truly is distributed (and not just a parallel vector type |
| 2946 |
// used for a sequential program). Assert that we are really |
| 2947 |
// working in a sequential context. |
| 2948 |
|
| 2949 |
internal::import_vector_with_ghost_elements( |
| 2950 |
vec, |
| 2951 |
vec_owned_elements, |
| 2952 |
complete_index_set(vec_owned_elements.size()), |
| 2953 |
ghosted_vector, |
| 2954 |
std::bool_constant<IsBlockVector<VectorType>::value>()); |
| 2955 |
} |
| 2956 |
|
| 2957 |
for (const ConstraintLine &line : lines) |
| 2958 |
if (vec_owned_elements.is_element(line.index)) |
| 2959 |
{ |
| 2960 |
typename VectorType::value_type new_value = line.inhomogeneity; |
| 2961 |
for (const std::pair<size_type, number> &entry : line.entries) |
| 2962 |
new_value += |
| 2963 |
(static_cast<typename VectorType::value_type>( |
| 2964 |
internal::ElementAccess<VectorType>::get(ghosted_vector, |
| 2965 |
entry.first)) * |
| 2966 |
entry.second); |
| 2967 |
AssertIsFinite(new_value); |
| 2968 |
internal::ElementAccess<VectorType>::set(new_value, |
| 2969 |
line.index, |
| 2970 |
vec); |
| 2971 |
} |
| 2972 |
|
| 2973 |
// now compress to communicate the entries that we added to |
| 2974 |
// and that weren't to local processors to the owner |
| 2975 |
// |
| 2976 |
// this shouldn't be strictly necessary but it probably doesn't |
| 2977 |
// hurt either |
| 2978 |
vec.compress(VectorOperation::insert); |
| 2979 |
} |
| 2980 |
} |
| 2981 |
else |
| 2982 |
// purely sequential vector (either because the type doesn't |
| 2983 |
// support anything else or because it's completely stored |
| 2984 |
// locally) |
| 2985 |
{ |
| 2986 |
for (const ConstraintLine &next_constraint : lines) |
| 2987 |
{ |
| 2988 |
// fill entry in line |
| 2989 |
// next_constraint.index by adding the |
| 2990 |
// different contributions |
| 2991 |
typename VectorType::value_type new_value = |
| 2992 |
next_constraint.inhomogeneity; |
| 2993 |
for (const std::pair<size_type, number> &entry : |
| 2994 |
next_constraint.entries) |
| 2995 |
new_value += |
| 2996 |
(static_cast<typename VectorType::value_type>( |
| 2997 |
internal::ElementAccess<VectorType>::get(vec, entry.first)) * |
| 2998 |
entry.second); |
| 2999 |
AssertIsFinite(new_value); |
| 3000 |
internal::ElementAccess<VectorType>::set(new_value, |
| 3001 |
next_constraint.index, |
| 3002 |
vec); |
| 3003 |
} |
| 3004 |
} |
| 3005 |
} |
| 3006 |
|
| 3007 |
// Some helper definitions for the local_to_global functions. |
| 3008 |
namespace internal |
| 3009 |
{ |
| 3010 |
namespace AffineConstraints |
| 3011 |
{ |
| 3012 |
inline Distributing::Distributing(const size_type global_row, |
| 3013 |
const size_type local_row) |
| 3014 |
: global_row(global_row) |
| 3015 |
, local_row(local_row) |
| 3016 |
, constraint_position(numbers::invalid_size_type) |
| 3017 |
{} |
| 3018 |
|
| 3019 |
|
| 3020 |
|
| 3021 |
inline Distributing::Distributing(const Distributing &in) |
| 3022 |
: constraint_position(numbers::invalid_size_type) |
| 3023 |
{ |
| 3024 |
*this = (in); |
| 3025 |
} |
| 3026 |
|
| 3027 |
|
| 3028 |
|
| 3029 |
inline Distributing & |
| 3030 |
Distributing::operator=(const Distributing &in) |
| 3031 |
{ |
| 3032 |
global_row = in.global_row; |
| 3033 |
local_row = in.local_row; |
| 3034 |
// the constraints pointer should not contain any data here. |
| 3035 |
Assert(constraint_position == numbers::invalid_size_type, |
| 3036 |
ExcInternalError()); |
| 3037 |
|
| 3038 |
if (in.constraint_position != numbers::invalid_size_type) |
| 3039 |
{ |
| 3040 |
constraint_position = in.constraint_position; |
| 3041 |
in.constraint_position = numbers::invalid_size_type; |
| 3042 |
} |
| 3043 |
return *this; |
| 3044 |
} |
| 3045 |
|
| 3046 |
|
| 3047 |
|
| 3048 |
template <typename number> |
| 3049 |
DataCache<number>::DataCache() |
| 3050 |
: row_length(8) |
| 3051 |
{} |
| 3052 |
|
| 3053 |
|
| 3054 |
|
| 3055 |
template <typename number> |
| 3056 |
void |
| 3057 |
DataCache<number>::reinit() |
| 3058 |
{ |
| 3059 |
individual_size.resize(0); |
| 3060 |
data.resize(0); |
| 3061 |
} |
| 3062 |
|
| 3063 |
|
| 3064 |
|
| 3065 |
template <typename number> |
| 3066 |
size_type |
| 3067 |
DataCache<number>::insert_new_index( |
| 3068 |
const std::pair<size_type, number> &pair) |
| 3069 |
{ |
| 3070 |
Assert(row_length > 0, ExcInternalError()); |
| 3071 |
const unsigned int index = individual_size.size(); |
| 3072 |
individual_size.push_back(1); |
| 3073 |
data.resize(individual_size.size() * row_length); |
| 3074 |
data[index * row_length] = pair; |
| 3075 |
individual_size[index] = 1; |
| 3076 |
return index; |
| 3077 |
} |
| 3078 |
|
| 3079 |
|
| 3080 |
|
| 3081 |
template <typename number> |
| 3082 |
void |
| 3083 |
DataCache<number>::append_index(const size_type index, |
| 3084 |
const std::pair<size_type, number> &pair) |
| 3085 |
{ |
| 3086 |
AssertIndexRange(index, individual_size.size()); |
| 3087 |
const size_type my_length = individual_size[index]; |
| 3088 |
if (my_length == row_length) |
| 3089 |
{ |
| 3090 |
AssertDimension(data.size(), individual_size.size() * row_length); |
| 3091 |
// no space left in this row, need to double row_length and |
| 3092 |
// rearrange the data items. Move all items to the right except the |
| 3093 |
// first one, starting at the back. Since individual_size contains |
| 3094 |
// at least one element when we get here, subtracting 1 works fine. |
| 3095 |
data.resize(2 * data.size()); |
| 3096 |
for (size_type i = individual_size.size() - 1; i > 0; --i) |
| 3097 |
{ |
| 3098 |
const auto ptr = data.data(); |
| 3099 |
std::move_backward(ptr + i * row_length, |
| 3100 |
ptr + i * row_length + individual_size[i], |
| 3101 |
ptr + i * 2 * row_length + individual_size[i]); |
| 3102 |
} |
| 3103 |
row_length *= 2; |
| 3104 |
} |
| 3105 |
data[index * row_length + my_length] = pair; |
| 3106 |
individual_size[index] = my_length + 1; |
| 3107 |
} |
| 3108 |
|
| 3109 |
|
| 3110 |
|
| 3111 |
template <typename number> |
| 3112 |
size_type |
| 3113 |
DataCache<number>::get_size(const size_type index) const |
| 3114 |
{ |
| 3115 |
return individual_size[index]; |
| 3116 |
} |
| 3117 |
|
| 3118 |
|
| 3119 |
|
| 3120 |
template <typename number> |
| 3121 |
const std::pair<size_type, number> * |
| 3122 |
DataCache<number>::get_entry(const size_type index) const |
| 3123 |
{ |
| 3124 |
return &data[index * row_length]; |
| 3125 |
} |
| 3126 |
|
| 3127 |
|
| 3128 |
|
| 3129 |
template <typename number> |
| 3130 |
GlobalRowsFromLocal<number>::GlobalRowsFromLocal() |
| 3131 |
: n_active_rows(0) |
| 3132 |
, n_inhomogeneous_rows(0) |
| 3133 |
{} |
| 3134 |
|
| 3135 |
|
| 3136 |
|
| 3137 |
template <typename number> |
| 3138 |
void |
| 3139 |
GlobalRowsFromLocal<number>::reinit(const size_type n_local_rows) |
| 3140 |
{ |
| 3141 |
total_row_indices.resize(n_local_rows); |
| 3142 |
for (unsigned int i = 0; i < n_local_rows; ++i) |
| 3143 |
total_row_indices[i].constraint_position = numbers::invalid_size_type; |
| 3144 |
n_active_rows = n_local_rows; |
| 3145 |
n_inhomogeneous_rows = 0; |
| 3146 |
data_cache.reinit(); |
| 3147 |
} |
| 3148 |
|
| 3149 |
|
| 3150 |
|
| 3151 |
template <typename number> |
| 3152 |
void |
| 3153 |
GlobalRowsFromLocal<number>::print(std::ostream &os) |
| 3154 |
{ |
| 3155 |
os << "Active rows " << n_active_rows << std::endl |
| 3156 |
<< "Constr rows " << n_constraints() << std::endl |
| 3157 |
<< "Inhom rows " << n_inhomogeneous_rows << std::endl |
| 3158 |
<< "Local: "; |
| 3159 |
for (const auto &total_row_index : total_row_indices) |
| 3160 |
os << ' ' << std::setw(4) << total_row_index.local_row; |
| 3161 |
os << std::endl << "Global:"; |
| 3162 |
for (const auto &total_row_index : total_row_indices) |
| 3163 |
os << ' ' << std::setw(4) << total_row_index.global_row; |
| 3164 |
os << std::endl << "ConPos:"; |
| 3165 |
for (const auto &total_row_index : total_row_indices) |
| 3166 |
os << ' ' << std::setw(4) << total_row_index.constraint_position; |
| 3167 |
os << std::endl; |
| 3168 |
} |
| 3169 |
|
| 3170 |
|
| 3171 |
|
| 3172 |
template <typename number> |
| 3173 |
size_type |
| 3174 |
GlobalRowsFromLocal<number>::size() const |
| 3175 |
{ |
| 3176 |
return n_active_rows; |
| 3177 |
} |
| 3178 |
|
| 3179 |
|
| 3180 |
|
| 3181 |
template <typename number> |
| 3182 |
size_type |
| 3183 |
GlobalRowsFromLocal<number>::size(const size_type counter_index) const |
| 3184 |
{ |
| 3185 |
return (total_row_indices[counter_index].constraint_position == |
| 3186 |
numbers::invalid_size_type ? |
| 3187 |
0 : |
| 3188 |
data_cache.get_size( |
| 3189 |
total_row_indices[counter_index].constraint_position)); |
| 3190 |
} |
| 3191 |
|
| 3192 |
|
| 3193 |
|
| 3194 |
template <typename number> |
| 3195 |
size_type |
| 3196 |
GlobalRowsFromLocal<number>::global_row(const size_type counter_index) const |
| 3197 |
{ |
| 3198 |
return total_row_indices[counter_index].global_row; |
| 3199 |
} |
| 3200 |
|
| 3201 |
|
| 3202 |
|
| 3203 |
template <typename number> |
| 3204 |
size_type & |
| 3205 |
GlobalRowsFromLocal<number>::global_row(const size_type counter_index) |
| 3206 |
{ |
| 3207 |
return total_row_indices[counter_index].global_row; |
| 3208 |
} |
| 3209 |
|
| 3210 |
|
| 3211 |
|
| 3212 |
template <typename number> |
| 3213 |
size_type |
| 3214 |
GlobalRowsFromLocal<number>::local_row(const size_type counter_index) const |
| 3215 |
{ |
| 3216 |
return total_row_indices[counter_index].local_row; |
| 3217 |
} |
| 3218 |
|
| 3219 |
|
| 3220 |
|
| 3221 |
template <typename number> |
| 3222 |
size_type & |
| 3223 |
GlobalRowsFromLocal<number>::local_row(const size_type counter_index) |
| 3224 |
{ |
| 3225 |
return total_row_indices[counter_index].local_row; |
| 3226 |
} |
| 3227 |
|
| 3228 |
|
| 3229 |
|
| 3230 |
template <typename number> |
| 3231 |
size_type |
| 3232 |
GlobalRowsFromLocal<number>::local_row( |
| 3233 |
const size_type counter_index, |
| 3234 |
const size_type index_in_constraint) const |
| 3235 |
{ |
| 3236 |
return (data_cache.get_entry(total_row_indices[counter_index] |
| 3237 |
.constraint_position)[index_in_constraint]) |
| 3238 |
.first; |
| 3239 |
} |
| 3240 |
|
| 3241 |
|
| 3242 |
|
| 3243 |
template <typename number> |
| 3244 |
number |
| 3245 |
GlobalRowsFromLocal<number>::constraint_value( |
| 3246 |
const size_type counter_index, |
| 3247 |
const size_type index_in_constraint) const |
| 3248 |
{ |
| 3249 |
return (data_cache.get_entry(total_row_indices[counter_index] |
| 3250 |
.constraint_position)[index_in_constraint]) |
| 3251 |
.second; |
| 3252 |
} |
| 3253 |
|
| 3254 |
|
| 3255 |
|
| 3256 |
template <typename number> |
| 3257 |
bool |
| 3258 |
GlobalRowsFromLocal<number>::have_indirect_rows() const |
| 3259 |
{ |
| 3260 |
return data_cache.individual_size.empty() == false; |
| 3261 |
} |
| 3262 |
|
| 3263 |
|
| 3264 |
|
| 3265 |
template <typename number> |
| 3266 |
void |
| 3267 |
GlobalRowsFromLocal<number>::insert_constraint( |
| 3268 |
const size_type constrained_local_dof) |
| 3269 |
{ |
| 3270 |
--n_active_rows; |
| 3271 |
total_row_indices[n_active_rows].local_row = constrained_local_dof; |
| 3272 |
total_row_indices[n_active_rows].global_row = numbers::invalid_size_type; |
| 3273 |
} |
| 3274 |
|
| 3275 |
|
| 3276 |
|
| 3277 |
template <typename number> |
| 3278 |
size_type |
| 3279 |
GlobalRowsFromLocal<number>::n_constraints() const |
| 3280 |
{ |
| 3281 |
return total_row_indices.size() - n_active_rows; |
| 3282 |
} |
| 3283 |
|
| 3284 |
|
| 3285 |
|
| 3286 |
template <typename number> |
| 3287 |
size_type |
| 3288 |
GlobalRowsFromLocal<number>::n_inhomogeneities() const |
| 3289 |
{ |
| 3290 |
return n_inhomogeneous_rows; |
| 3291 |
} |
| 3292 |
|
| 3293 |
|
| 3294 |
|
| 3295 |
template <typename number> |
| 3296 |
void |
| 3297 |
GlobalRowsFromLocal<number>::set_ith_constraint_inhomogeneous( |
| 3298 |
const size_type i) |
| 3299 |
{ |
| 3300 |
Assert(i >= n_inhomogeneous_rows, ExcInternalError()); |
| 3301 |
std::swap(total_row_indices[n_active_rows + i], |
| 3302 |
total_row_indices[n_active_rows + n_inhomogeneous_rows]); |
| 3303 |
++n_inhomogeneous_rows; |
| 3304 |
} |
| 3305 |
|
| 3306 |
|
| 3307 |
|
| 3308 |
template <typename number> |
| 3309 |
size_type |
| 3310 |
GlobalRowsFromLocal<number>::constraint_origin(size_type i) const |
| 3311 |
{ |
| 3312 |
return total_row_indices[n_active_rows + i].local_row; |
| 3313 |
} |
| 3314 |
|
| 3315 |
|
| 3316 |
// a function that appends an additional row to the list of values, or |
| 3317 |
// appends a value to an already existing row. Similar functionality as for |
| 3318 |
// std::map<size_type,Distributing>, but here done for a |
| 3319 |
// std::vector<Distributing>, much faster for short lists as we have them |
| 3320 |
// here |
| 3321 |
template <typename number> |
| 3322 |
inline void |
| 3323 |
GlobalRowsFromLocal<number>::insert_index(const size_type global_row, |
| 3324 |
const size_type local_row, |
| 3325 |
const number constraint_value) |
| 3326 |
{ |
| 3327 |
using index_iterator = std::vector<Distributing>::iterator; |
| 3328 |
index_iterator pos, pos1; |
| 3329 |
Distributing row_value(global_row); |
| 3330 |
std::pair<size_type, number> constraint(local_row, constraint_value); |
| 3331 |
|
| 3332 |
// check whether the list was really sorted before entering here |
| 3333 |
for (size_type i = 1; i < n_active_rows; ++i) |
| 3334 |
Assert(total_row_indices[i - 1] < total_row_indices[i], |
| 3335 |
ExcInternalError()); |
| 3336 |
|
| 3337 |
pos = Utilities::lower_bound(total_row_indices.begin(), |
| 3338 |
total_row_indices.begin() + n_active_rows, |
| 3339 |
row_value); |
| 3340 |
if (pos->global_row == global_row) |
| 3341 |
pos1 = pos; |
| 3342 |
else |
| 3343 |
{ |
| 3344 |
pos1 = total_row_indices.insert(pos, row_value); |
| 3345 |
++n_active_rows; |
| 3346 |
} |
| 3347 |
|
| 3348 |
if (pos1->constraint_position == numbers::invalid_size_type) |
| 3349 |
pos1->constraint_position = data_cache.insert_new_index(constraint); |
| 3350 |
else |
| 3351 |
data_cache.append_index(pos1->constraint_position, constraint); |
| 3352 |
} |
| 3353 |
|
| 3354 |
// this sort algorithm sorts std::vector<Distributing>, but does not take |
| 3355 |
// the constraints into account. this means that in case that constraints |
| 3356 |
// are already inserted, this function does not work as expected. Use |
| 3357 |
// shellsort, which is very fast in case the indices are already sorted |
| 3358 |
// (which is the usual case with DG elements), and not too slow in other |
| 3359 |
// cases |
| 3360 |
template <typename number> |
| 3361 |
inline void |
| 3362 |
GlobalRowsFromLocal<number>::sort() |
| 3363 |
{ |
| 3364 |
size_type i, j, j2, temp, templ, istep; |
| 3365 |
size_type step; |
| 3366 |
|
| 3367 |
// check whether the constraints are really empty. |
| 3368 |
const size_type length = size(); |
| 3369 |
|
| 3370 |
// make sure that we are in the range of the vector |
| 3371 |
AssertIndexRange(length, total_row_indices.size() + 1); |
| 3372 |
for (size_type i = 0; i < length; ++i) |
| 3373 |
Assert(total_row_indices[i].constraint_position == |
| 3374 |
numbers::invalid_size_type, |
| 3375 |
ExcInternalError()); |
| 3376 |
|
| 3377 |
step = length / 2; |
| 3378 |
while (step > 0) |
| 3379 |
{ |
| 3380 |
for (i = step; i < length; ++i) |
| 3381 |
{ |
| 3382 |
istep = step; |
| 3383 |
j = i; |
| 3384 |
j2 = j - istep; |
| 3385 |
temp = total_row_indices[i].global_row; |
| 3386 |
templ = total_row_indices[i].local_row; |
| 3387 |
if (total_row_indices[j2].global_row > temp) |
| 3388 |
{ |
| 3389 |
while ((j >= istep) && |
| 3390 |
(total_row_indices[j2].global_row > temp)) |
| 3391 |
{ |
| 3392 |
total_row_indices[j].global_row = |
| 3393 |
total_row_indices[j2].global_row; |
| 3394 |
total_row_indices[j].local_row = |
| 3395 |
total_row_indices[j2].local_row; |
| 3396 |
j = j2; |
| 3397 |
j2 -= istep; |
| 3398 |
} |
| 3399 |
total_row_indices[j].global_row = temp; |
| 3400 |
total_row_indices[j].local_row = templ; |
| 3401 |
} |
| 3402 |
} |
| 3403 |
step = step >> 1; |
| 3404 |
} |
| 3405 |
} |
| 3406 |
|
| 3407 |
|
| 3408 |
|
| 3409 |
/** |
| 3410 |
* This class is an accessor class to scratch data that is used |
| 3411 |
* during calls to distribute_local_to_global and |
| 3412 |
* add_entries_local_to_global. In order to avoid frequent memory |
| 3413 |
* allocation, we keep the data alive from one call to the next in |
| 3414 |
* a static variable. Since we want to allow for different number |
| 3415 |
* types in matrices, this is a template. |
| 3416 |
* |
| 3417 |
* Since each thread gets its private version of scratch data out of the |
| 3418 |
* ThreadLocalStorage, no conflicting access can occur. For this to be |
| 3419 |
* valid, we need to make sure that no call within |
| 3420 |
* distribute_local_to_global is made that by itself can spawn tasks. |
| 3421 |
* Otherwise, we might end up in a situation where several threads fight for |
| 3422 |
* the data. |
| 3423 |
*/ |
| 3424 |
template <typename number> |
| 3425 |
class ScratchDataAccessor |
| 3426 |
{ |
| 3427 |
public: |
| 3428 |
/** |
| 3429 |
* Constructor. Takes the scratch data object for the current |
| 3430 |
* thread out of the provided object and marks it as used. |
| 3431 |
*/ |
| 3432 |
ScratchDataAccessor( |
| 3433 |
Threads::ThreadLocalStorage<ScratchData<number>> &tls_scratch_data) |
| 3434 |
: my_scratch_data(&tls_scratch_data.get()) |
| 3435 |
{ |
| 3436 |
Assert(my_scratch_data->in_use == false, |
| 3437 |
ExcMessage( |
| 3438 |
"Access to thread-local scratch data tried, but it is already " |
| 3439 |
"in use")); |
| 3440 |
my_scratch_data->in_use = true; |
| 3441 |
} |
| 3442 |
|
| 3443 |
/** |
| 3444 |
* Destructor. Mark scratch data as available again. |
| 3445 |
*/ |
| 3446 |
~ScratchDataAccessor() |
| 3447 |
{ |
| 3448 |
my_scratch_data->in_use = false; |
| 3449 |
} |
| 3450 |
|
| 3451 |
/** |
| 3452 |
* Dereferencing operator. |
| 3453 |
*/ |
| 3454 |
ScratchData<number> & |
| 3455 |
operator*() |
| 3456 |
{ |
| 3457 |
return *my_scratch_data; |
| 3458 |
} |
| 3459 |
|
| 3460 |
/** |
| 3461 |
* Dereferencing operator. |
| 3462 |
*/ |
| 3463 |
ScratchData<number> * |
| 3464 |
operator->() |
| 3465 |
{ |
| 3466 |
return my_scratch_data; |
| 3467 |
} |
| 3468 |
|
| 3469 |
private: |
| 3470 |
ScratchData<number> *my_scratch_data; |
| 3471 |
}; |
| 3472 |
|
| 3473 |
|
| 3474 |
// function for block matrices: Find out where in the list of local dofs |
| 3475 |
// (sorted according to global ids) the individual blocks start. |
| 3476 |
// Transform the global indices to block-local indices in order to be |
| 3477 |
// able to use functions like vector.block(1)(block_local_id), instead of |
| 3478 |
// vector(global_id). This avoids transforming indices one-by-one later |
| 3479 |
// on. |
| 3480 |
template <typename number, typename BlockType> |
| 3481 |
inline void |
| 3482 |
make_block_starts(const BlockType &block_object, |
| 3483 |
GlobalRowsFromLocal<number> &global_rows, |
| 3484 |
std::vector<size_type> &block_starts) |
| 3485 |
{ |
| 3486 |
AssertDimension(block_starts.size(), block_object.n_block_rows() + 1); |
| 3487 |
|
| 3488 |
using row_iterator = std::vector<Distributing>::iterator; |
| 3489 |
row_iterator block_indices = global_rows.total_row_indices.begin(); |
| 3490 |
|
| 3491 |
const size_type num_blocks = block_object.n_block_rows(); |
| 3492 |
const size_type n_active_rows = global_rows.size(); |
| 3493 |
|
| 3494 |
// find end of rows. |
| 3495 |
block_starts[0] = 0; |
| 3496 |
for (size_type i = 1; i < num_blocks; ++i) |
| 3497 |
{ |
| 3498 |
row_iterator first_block = Utilities::lower_bound( |
| 3499 |
block_indices, |
| 3500 |
global_rows.total_row_indices.begin() + n_active_rows, |
| 3501 |
Distributing(block_object.get_row_indices().block_start(i))); |
| 3502 |
block_starts[i] = first_block - global_rows.total_row_indices.begin(); |
| 3503 |
block_indices = first_block; |
| 3504 |
} |
| 3505 |
block_starts[num_blocks] = n_active_rows; |
| 3506 |
|
| 3507 |
// transform row indices to block-local index space |
| 3508 |
for (size_type i = block_starts[1]; i < n_active_rows; ++i) |
| 3509 |
global_rows.global_row(i) = |
| 3510 |
block_object.get_row_indices() |
| 3511 |
.global_to_local(global_rows.global_row(i)) |
| 3512 |
.second; |
| 3513 |
} |
| 3514 |
|
| 3515 |
// same as before, but for std::vector<uint> instead of |
| 3516 |
// GlobalRowsFromLocal. Used in functions for sparsity patterns. |
| 3517 |
template <typename BlockType> |
| 3518 |
inline void |
| 3519 |
make_block_starts(const BlockType &block_object, |
| 3520 |
std::vector<size_type> &row_indices, |
| 3521 |
std::vector<size_type> &block_starts) |
| 3522 |
{ |
| 3523 |
AssertDimension(block_starts.size(), block_object.n_block_rows() + 1); |
| 3524 |
|
| 3525 |
using row_iterator = std::vector<size_type>::iterator; |
| 3526 |
row_iterator col_indices = row_indices.begin(); |
| 3527 |
|
| 3528 |
const size_type num_blocks = block_object.n_block_rows(); |
| 3529 |
|
| 3530 |
// find end of rows. |
| 3531 |
block_starts[0] = 0; |
| 3532 |
for (size_type i = 1; i < num_blocks; ++i) |
| 3533 |
{ |
| 3534 |
row_iterator first_block = |
| 3535 |
Utilities::lower_bound(col_indices, |
| 3536 |
row_indices.end(), |
| 3537 |
block_object.get_row_indices().block_start( |
| 3538 |
i)); |
| 3539 |
block_starts[i] = first_block - row_indices.begin(); |
| 3540 |
col_indices = first_block; |
| 3541 |
} |
| 3542 |
block_starts[num_blocks] = row_indices.size(); |
| 3543 |
|
| 3544 |
// transform row indices to local index space |
| 3545 |
for (size_type i = block_starts[1]; i < row_indices.size(); ++i) |
| 3546 |
row_indices[i] = |
| 3547 |
block_object.get_row_indices().global_to_local(row_indices[i]).second; |
| 3548 |
} |
| 3549 |
|
| 3550 |
// resolves constraints of one column at the innermost loop. goes through |
| 3551 |
// the origin of each global entry and finds out which data we need to |
| 3552 |
// collect. |
| 3553 |
template <typename number> |
| 3554 |
inline number |
| 3555 |
resolve_matrix_entry(const GlobalRowsFromLocal<number> &global_rows, |
| 3556 |
const GlobalRowsFromLocal<number> &global_cols, |
| 3557 |
const size_type i, |
| 3558 |
const size_type j, |
| 3559 |
const size_type loc_row, |
| 3560 |
const FullMatrix<number> &local_matrix) |
| 3561 |
{ |
| 3562 |
const size_type loc_col = global_cols.local_row(j); |
| 3563 |
number col_val; |
| 3564 |
|
| 3565 |
// case 1: row has direct contribution in local matrix. decide whether col |
| 3566 |
// has a direct contribution. if not, set the value to zero. |
| 3567 |
if (loc_row != numbers::invalid_size_type) |
| 3568 |
{ |
| 3569 |
col_val = ((loc_col != numbers::invalid_size_type) ? |
| 3570 |
local_matrix(loc_row, loc_col) : |
| 3571 |
0); |
| 3572 |
|
| 3573 |
// account for indirect contributions by constraints in column |
| 3574 |
for (size_type p = 0; p < global_cols.size(j); ++p) |
| 3575 |
col_val += (local_matrix(loc_row, global_cols.local_row(j, p)) * |
| 3576 |
global_cols.constraint_value(j, p)); |
| 3577 |
} |
| 3578 |
|
| 3579 |
// case 2: row has no direct contribution in local matrix |
| 3580 |
else |
| 3581 |
col_val = 0; |
| 3582 |
|
| 3583 |
// account for indirect contributions by constraints in row, going through |
| 3584 |
// the direct and indirect references in the given column. |
| 3585 |
for (size_type q = 0; q < global_rows.size(i); ++q) |
| 3586 |
{ |
| 3587 |
number add_this = |
| 3588 |
(loc_col != numbers::invalid_size_type) ? |
| 3589 |
local_matrix(global_rows.local_row(i, q), loc_col) : |
| 3590 |
0; |
| 3591 |
|
| 3592 |
for (size_type p = 0; p < global_cols.size(j); ++p) |
| 3593 |
add_this += (local_matrix(global_rows.local_row(i, q), |
| 3594 |
global_cols.local_row(j, p)) * |
| 3595 |
global_cols.constraint_value(j, p)); |
| 3596 |
col_val += add_this * global_rows.constraint_value(i, q); |
| 3597 |
} |
| 3598 |
return col_val; |
| 3599 |
} |
| 3600 |
|
| 3601 |
// computes all entries that need to be written into global_rows[i]. Lists |
| 3602 |
// the resulting values in val_ptr, and the corresponding column indices in |
| 3603 |
// col_ptr. |
| 3604 |
template <typename number> |
| 3605 |
inline void |
| 3606 |
resolve_matrix_row(const GlobalRowsFromLocal<number> &global_rows, |
| 3607 |
const GlobalRowsFromLocal<number> &global_cols, |
| 3608 |
const size_type i, |
| 3609 |
const size_type column_start, |
| 3610 |
const size_type column_end, |
| 3611 |
const FullMatrix<number> &local_matrix, |
| 3612 |
size_type *&col_ptr, |
| 3613 |
number *&val_ptr) |
| 3614 |
{ |
| 3615 |
if (column_end == column_start) |
| 3616 |
return; |
| 3617 |
|
| 3618 |
AssertIndexRange(column_end - 1, global_cols.size()); |
| 3619 |
const size_type loc_row = global_rows.local_row(i); |
| 3620 |
|
| 3621 |
// fast function if there are no indirect references to any of the local |
| 3622 |
// rows at all on this set of dofs (saves a lot of checks). the only check |
| 3623 |
// we actually need to perform is whether the matrix element is zero. |
| 3624 |
if (global_rows.have_indirect_rows() == false && |
| 3625 |
global_cols.have_indirect_rows() == false) |
| 3626 |
{ |
| 3627 |
AssertIndexRange(loc_row, local_matrix.m()); |
| 3628 |
const number *matrix_ptr = &local_matrix(loc_row, 0); |
| 3629 |
|
| 3630 |
for (size_type j = column_start; j < column_end; ++j) |
| 3631 |
{ |
| 3632 |
const size_type loc_col = global_cols.local_row(j); |
| 3633 |
AssertIndexRange(loc_col, local_matrix.n()); |
| 3634 |
const number col_val = matrix_ptr[loc_col]; |
| 3635 |
if (col_val != number()) |
| 3636 |
{ |
| 3637 |
*val_ptr++ = static_cast<number>(col_val); |
| 3638 |
*col_ptr++ = global_cols.global_row(j); |
| 3639 |
} |
| 3640 |
} |
| 3641 |
} |
| 3642 |
|
| 3643 |
// more difficult part when there are indirect references and when we need |
| 3644 |
// to do some more checks. |
| 3645 |
else |
| 3646 |
{ |
| 3647 |
for (size_type j = column_start; j < column_end; ++j) |
| 3648 |
{ |
| 3649 |
number col_val = resolve_matrix_entry( |
| 3650 |
global_rows, global_cols, i, j, loc_row, local_matrix); |
| 3651 |
|
| 3652 |
// if we got some nontrivial value, append it to the array of |
| 3653 |
// values. |
| 3654 |
if (col_val != number()) |
| 3655 |
{ |
| 3656 |
*val_ptr++ = static_cast<number>(col_val); |
| 3657 |
*col_ptr++ = global_cols.global_row(j); |
| 3658 |
} |
| 3659 |
} |
| 3660 |
} |
| 3661 |
} |
| 3662 |
|
| 3663 |
// specialized function that can write into the row of a |
| 3664 |
// SparseMatrix<number>. |
| 3665 |
namespace dealiiSparseMatrix |
| 3666 |
{ |
| 3667 |
template <typename SparseMatrixIterator, typename LocalType> |
| 3668 |
inline void |
| 3669 |
add_value(const LocalType value, |
| 3670 |
const size_type row, |
| 3671 |
const size_type column, |
| 3672 |
SparseMatrixIterator &matrix_values) |
| 3673 |
{ |
| 3674 |
if (value != LocalType()) |
| 3675 |
{ |
| 3676 |
while (matrix_values->column() < column) |
| 3677 |
++matrix_values; |
| 3678 |
|
| 3679 |
// Ensure that we haven't walked off the end of the row and |
| 3680 |
// accidentally found the column in a later row. This should |
| 3681 |
// not have happened since we believed we know that we enter |
| 3682 |
// entries sorted by column, and that all columns are represented |
| 3683 |
// in the current matrix row. But it's worth checking. |
| 3684 |
Assert(matrix_values->row() == row, ExcInternalError()); |
| 3685 |
|
| 3686 |
// Then also assert that the column index actually exists. |
| 3687 |
Assert(matrix_values->column() == column, |
| 3688 |
typename SparseMatrix< |
| 3689 |
typename SparseMatrixIterator::MatrixType::value_type>:: |
| 3690 |
ExcInvalidIndex(row, column)); |
| 3691 |
|
| 3692 |
// Now so convinced, let us add the relevant value: |
| 3693 |
matrix_values->value() += value; |
| 3694 |
} |
| 3695 |
} |
| 3696 |
} // namespace dealiiSparseMatrix |
| 3697 |
|
| 3698 |
// similar as before, now with shortcut for deal.II sparse matrices. this |
| 3699 |
// lets us avoid using extra arrays, and does all the operations just in |
| 3700 |
// place, i.e., in the respective matrix row |
| 3701 |
template <typename number> |
| 3702 |
inline void |
| 3703 |
resolve_matrix_row(const GlobalRowsFromLocal<number> &global_rows, |
| 3704 |
const size_type i, |
| 3705 |
const size_type column_start, |
| 3706 |
const size_type column_end, |
| 3707 |
const FullMatrix<number> &local_matrix, |
| 3708 |
SparseMatrix<number> *sparse_matrix) |
| 3709 |
{ |
| 3710 |
if (column_end == column_start) |
| 3711 |
return; |
| 3712 |
|
| 3713 |
AssertIndexRange(column_end - 1, global_rows.size()); |
| 3714 |
const SparsityPattern &sparsity = sparse_matrix->get_sparsity_pattern(); |
| 3715 |
|
| 3716 |
if (sparsity.n_nonzero_elements() == 0) |
| 3717 |
return; |
| 3718 |
|
| 3719 |
const size_type row = global_rows.global_row(i); |
| 3720 |
const size_type loc_row = global_rows.local_row(i); |
| 3721 |
|
| 3722 |
typename SparseMatrix<number>::iterator matrix_values = |
| 3723 |
sparse_matrix->begin(row); |
| 3724 |
const bool optimize_diagonal = sparsity.n_rows() == sparsity.n_cols(); |
| 3725 |
|
| 3726 |
// distinguish three cases about what can happen for checking whether the |
| 3727 |
// diagonal is the first element of the row. this avoids if statements at |
| 3728 |
// the innermost loop positions |
| 3729 |
|
| 3730 |
if (!optimize_diagonal) // case 1: no diagonal optimization in matrix |
| 3731 |
{ |
| 3732 |
if (global_rows.have_indirect_rows() == false) |
| 3733 |
{ |
| 3734 |
AssertIndexRange(loc_row, local_matrix.m()); |
| 3735 |
const number *matrix_ptr = &local_matrix(loc_row, 0); |
| 3736 |
|
| 3737 |
for (size_type j = column_start; j < column_end; ++j) |
| 3738 |
{ |
| 3739 |
const size_type loc_col = global_rows.local_row(j); |
| 3740 |
const number col_val = matrix_ptr[loc_col]; |
| 3741 |
dealiiSparseMatrix::add_value(col_val, |
| 3742 |
row, |
| 3743 |
global_rows.global_row(j), |
| 3744 |
matrix_values); |
| 3745 |
} |
| 3746 |
} |
| 3747 |
else |
| 3748 |
{ |
| 3749 |
for (size_type j = column_start; j < column_end; ++j) |
| 3750 |
{ |
| 3751 |
number col_val = resolve_matrix_entry( |
| 3752 |
global_rows, global_rows, i, j, loc_row, local_matrix); |
| 3753 |
dealiiSparseMatrix::add_value(col_val, |
| 3754 |
row, |
| 3755 |
global_rows.global_row(j), |
| 3756 |
matrix_values); |
| 3757 |
} |
| 3758 |
} |
| 3759 |
} |
| 3760 |
else if (i >= column_start && i < column_end) // case 2: can split loop |
| 3761 |
{ |
| 3762 |
++matrix_values; // jump over diagonal element |
| 3763 |
if (global_rows.have_indirect_rows() == false) |
| 3764 |
{ |
| 3765 |
AssertIndexRange(loc_row, local_matrix.m()); |
| 3766 |
const number *matrix_ptr = &local_matrix(loc_row, 0); |
| 3767 |
|
| 3768 |
sparse_matrix->begin(row)->value() += matrix_ptr[loc_row]; |
| 3769 |
for (size_type j = column_start; j < i; ++j) |
| 3770 |
{ |
| 3771 |
const size_type loc_col = global_rows.local_row(j); |
| 3772 |
const number col_val = matrix_ptr[loc_col]; |
| 3773 |
dealiiSparseMatrix::add_value(col_val, |
| 3774 |
row, |
| 3775 |
global_rows.global_row(j), |
| 3776 |
matrix_values); |
| 3777 |
} |
| 3778 |
for (size_type j = i + 1; j < column_end; ++j) |
| 3779 |
{ |
| 3780 |
const size_type loc_col = global_rows.local_row(j); |
| 3781 |
const number col_val = matrix_ptr[loc_col]; |
| 3782 |
dealiiSparseMatrix::add_value(col_val, |
| 3783 |
row, |
| 3784 |
global_rows.global_row(j), |
| 3785 |
matrix_values); |
| 3786 |
} |
| 3787 |
} |
| 3788 |
else |
| 3789 |
{ |
| 3790 |
sparse_matrix->begin(row)->value() += resolve_matrix_entry( |
| 3791 |
global_rows, global_rows, i, i, loc_row, local_matrix); |
| 3792 |
for (size_type j = column_start; j < i; ++j) |
| 3793 |
{ |
| 3794 |
number col_val = resolve_matrix_entry( |
| 3795 |
global_rows, global_rows, i, j, loc_row, local_matrix); |
| 3796 |
dealiiSparseMatrix::add_value(col_val, |
| 3797 |
row, |
| 3798 |
global_rows.global_row(j), |
| 3799 |
matrix_values); |
| 3800 |
} |
| 3801 |
for (size_type j = i + 1; j < column_end; ++j) |
| 3802 |
{ |
| 3803 |
number col_val = resolve_matrix_entry( |
| 3804 |
global_rows, global_rows, i, j, loc_row, local_matrix); |
| 3805 |
dealiiSparseMatrix::add_value(col_val, |
| 3806 |
row, |
| 3807 |
global_rows.global_row(j), |
| 3808 |
matrix_values); |
| 3809 |
} |
| 3810 |
} |
| 3811 |
} |
| 3812 |
// case 3: can't say - need to check inside the loop |
| 3813 |
else if (global_rows.have_indirect_rows() == false) |
| 3814 |
{ |
| 3815 |
++matrix_values; // jump over diagonal element |
| 3816 |
AssertIndexRange(loc_row, local_matrix.m()); |
| 3817 |
const number *matrix_ptr = &local_matrix(loc_row, 0); |
| 3818 |
|
| 3819 |
for (size_type j = column_start; j < column_end; ++j) |
| 3820 |
{ |
| 3821 |
const size_type loc_col = global_rows.local_row(j); |
| 3822 |
const number col_val = matrix_ptr[loc_col]; |
| 3823 |
if (row == global_rows.global_row(j)) |
| 3824 |
sparse_matrix->begin(row)->value() += col_val; |
| 3825 |
else |
| 3826 |
dealiiSparseMatrix::add_value(col_val, |
| 3827 |
row, |
| 3828 |
global_rows.global_row(j), |
| 3829 |
matrix_values); |
| 3830 |
} |
| 3831 |
} |
| 3832 |
else |
| 3833 |
{ |
| 3834 |
++matrix_values; // jump over diagonal element |
| 3835 |
for (size_type j = column_start; j < column_end; ++j) |
| 3836 |
{ |
| 3837 |
number col_val = resolve_matrix_entry( |
| 3838 |
global_rows, global_rows, i, j, loc_row, local_matrix); |
| 3839 |
if (row == global_rows.global_row(j)) |
| 3840 |
sparse_matrix->begin(row)->value() += col_val; |
| 3841 |
else |
| 3842 |
dealiiSparseMatrix::add_value(col_val, |
| 3843 |
row, |
| 3844 |
global_rows.global_row(j), |
| 3845 |
matrix_values); |
| 3846 |
} |
| 3847 |
} |
| 3848 |
} |
| 3849 |
|
| 3850 |
// Same function to resolve all entries that will be added to the given |
| 3851 |
// global row global_rows[i] as before, now for sparsity pattern |
| 3852 |
template <typename number> |
| 3853 |
inline void |
| 3854 |
resolve_matrix_row(const GlobalRowsFromLocal<number> &global_rows, |
| 3855 |
const size_type i, |
| 3856 |
const size_type column_start, |
| 3857 |
const size_type column_end, |
| 3858 |
const Table<2, bool> &dof_mask, |
| 3859 |
std::vector<size_type>::iterator &col_ptr) |
| 3860 |
{ |
| 3861 |
if (column_end == column_start) |
| 3862 |
return; |
| 3863 |
|
| 3864 |
const size_type loc_row = global_rows.local_row(i); |
| 3865 |
|
| 3866 |
// fast function if there are no indirect references to any of the local |
| 3867 |
// rows at all on this set of dofs |
| 3868 |
if (global_rows.have_indirect_rows() == false) |
| 3869 |
{ |
| 3870 |
Assert(loc_row < dof_mask.n_rows(), ExcInternalError()); |
| 3871 |
|
| 3872 |
for (size_type j = column_start; j < column_end; ++j) |
| 3873 |
{ |
| 3874 |
const size_type loc_col = global_rows.local_row(j); |
| 3875 |
Assert(loc_col < dof_mask.n_cols(), ExcInternalError()); |
| 3876 |
|
| 3877 |
if (dof_mask(loc_row, loc_col) == true) |
| 3878 |
*col_ptr++ = global_rows.global_row(j); |
| 3879 |
} |
| 3880 |
} |
| 3881 |
|
| 3882 |
// slower functions when there are indirect references and when we need to |
| 3883 |
// do some more checks. |
| 3884 |
else |
| 3885 |
{ |
| 3886 |
for (size_type j = column_start; j < column_end; ++j) |
| 3887 |
{ |
| 3888 |
const size_type loc_col = global_rows.local_row(j); |
| 3889 |
if (loc_row != numbers::invalid_size_type) |
| 3890 |
{ |
| 3891 |
Assert(loc_row < dof_mask.n_rows(), ExcInternalError()); |
| 3892 |
if (loc_col != numbers::invalid_size_type) |
| 3893 |
{ |
| 3894 |
Assert(loc_col < dof_mask.n_cols(), ExcInternalError()); |
| 3895 |
if (dof_mask(loc_row, loc_col) == true) |
| 3896 |
goto add_this_index; |
| 3897 |
} |
| 3898 |
|
| 3899 |
for (size_type p = 0; p < global_rows.size(j); ++p) |
| 3900 |
if (dof_mask(loc_row, global_rows.local_row(j, p)) == true) |
| 3901 |
goto add_this_index; |
| 3902 |
} |
| 3903 |
|
| 3904 |
for (size_type q = 0; q < global_rows.size(i); ++q) |
| 3905 |
{ |
| 3906 |
if (loc_col != numbers::invalid_size_type) |
| 3907 |
{ |
| 3908 |
Assert(loc_col < dof_mask.n_cols(), ExcInternalError()); |
| 3909 |
if (dof_mask(global_rows.local_row(i, q), loc_col) == |
| 3910 |
true) |
| 3911 |
goto add_this_index; |
| 3912 |
} |
| 3913 |
|
| 3914 |
for (size_type p = 0; p < global_rows.size(j); ++p) |
| 3915 |
if (dof_mask(global_rows.local_row(i, q), |
| 3916 |
global_rows.local_row(j, p)) == true) |
| 3917 |
goto add_this_index; |
| 3918 |
} |
| 3919 |
|
| 3920 |
continue; |
| 3921 |
// if we got some nontrivial value, append it to the array of |
| 3922 |
// values. |
| 3923 |
add_this_index: |
| 3924 |
*col_ptr++ = global_rows.global_row(j); |
| 3925 |
} |
| 3926 |
} |
| 3927 |
} |
| 3928 |
|
| 3929 |
// to make sure that the global matrix remains invertible, we need to do |
| 3930 |
// something with the diagonal elements. Add the average of the |
| 3931 |
// absolute values of the local matrix diagonals, so the resulting entry |
| 3932 |
// will always be positive and furthermore be in the same order of magnitude |
| 3933 |
// as the other elements of the matrix. If all local matrix diagonals are |
| 3934 |
// zero, add the l1 norm of the local matrix divided by the matrix size |
| 3935 |
// to the diagonal of the global matrix. If the entire local matrix is zero, |
| 3936 |
// add 1 to the diagonal of the global matrix. |
| 3937 |
// |
| 3938 |
// note that this also captures the special case that a dof is both |
| 3939 |
// constrained and fixed (this can happen for hanging nodes in 3d that also |
| 3940 |
// happen to be on the boundary). in that case, following the program flow |
| 3941 |
// in distribute_local_to_global, it is realized that when distributing the |
| 3942 |
// row and column no elements of the matrix are actually touched if all the |
| 3943 |
// degrees of freedom to which this dof is constrained are also constrained |
| 3944 |
// (the usual case with hanging nodes in 3d). however, in the line below, we |
| 3945 |
// do actually do something with this dof |
| 3946 |
template <typename number, typename MatrixType, typename VectorType> |
| 3947 |
inline void |
| 3948 |
set_matrix_diagonals( |
| 3949 |
const internal::AffineConstraints::GlobalRowsFromLocal<number> |
| 3950 |
&global_rows, |
| 3951 |
const std::vector<size_type> &local_dof_indices, |
| 3952 |
const FullMatrix<number> &local_matrix, |
| 3953 |
const dealii::AffineConstraints<number> &constraints, |
| 3954 |
MatrixType &global_matrix, |
| 3955 |
VectorType &global_vector, |
| 3956 |
bool use_inhomogeneities_for_rhs) |
| 3957 |
{ |
| 3958 |
if (global_rows.n_constraints() > 0) |
| 3959 |
{ |
| 3960 |
number average_diagonal = number(); |
| 3961 |
for (size_type i = 0; i < local_matrix.m(); ++i) |
| 3962 |
average_diagonal += std::abs(local_matrix(i, i)); |
| 3963 |
average_diagonal /= static_cast<number>(local_matrix.m()); |
| 3964 |
|
| 3965 |
// handle the case that all diagonal elements are zero |
| 3966 |
if (average_diagonal == static_cast<number>(0.)) |
| 3967 |
{ |
| 3968 |
average_diagonal = static_cast<number>(local_matrix.l1_norm()) / |
| 3969 |
static_cast<number>(local_matrix.m()); |
| 3970 |
// if the entire matrix is zero, use 1. for the diagonal |
| 3971 |
if (average_diagonal == static_cast<number>(0.)) |
| 3972 |
average_diagonal = static_cast<number>(1.); |
| 3973 |
} |
| 3974 |
|
| 3975 |
for (size_type i = 0; i < global_rows.n_constraints(); ++i) |
| 3976 |
{ |
| 3977 |
const size_type local_row = global_rows.constraint_origin(i); |
| 3978 |
const size_type global_row = local_dof_indices[local_row]; |
| 3979 |
|
| 3980 |
const number current_diagonal = |
| 3981 |
local_matrix(local_row, local_row); |
| 3982 |
if (std::abs(current_diagonal) != 0.) |
| 3983 |
{ |
| 3984 |
global_matrix.add(global_row, |
| 3985 |
global_row, |
| 3986 |
std::abs(current_diagonal)); |
| 3987 |
// if the use_inhomogeneities_for_rhs flag is set to true, the |
| 3988 |
// inhomogeneities are used to create the global vector. |
| 3989 |
// instead of fill in a zero in the ith components with an |
| 3990 |
// inhomogeneity, we set those to: |
| 3991 |
// inhomogeneity(i)*global_matrix (i,i). |
| 3992 |
if (use_inhomogeneities_for_rhs == true) |
| 3993 |
global_vector(global_row) += |
| 3994 |
current_diagonal * |
| 3995 |
constraints.get_inhomogeneity(global_row); |
| 3996 |
} |
| 3997 |
else |
| 3998 |
{ |
| 3999 |
global_matrix.add(global_row, global_row, average_diagonal); |
| 4000 |
|
| 4001 |
if (use_inhomogeneities_for_rhs == true) |
| 4002 |
global_vector(global_row) += |
| 4003 |
average_diagonal * |
| 4004 |
constraints.get_inhomogeneity(global_row); |
| 4005 |
} |
| 4006 |
} |
| 4007 |
} |
| 4008 |
} |
| 4009 |
|
| 4010 |
// similar function as the one above for setting matrix diagonals, but now |
| 4011 |
// doing that for sparsity patterns when setting them up using |
| 4012 |
// add_entries_local_to_global(). In case we keep constrained entries, add |
| 4013 |
// all the rows and columns related to the constrained dof, otherwise just |
| 4014 |
// add the diagonal |
| 4015 |
template <typename number> |
| 4016 |
inline void |
| 4017 |
set_sparsity_diagonals( |
| 4018 |
const internal::AffineConstraints::GlobalRowsFromLocal<number> |
| 4019 |
&global_rows, |
| 4020 |
const std::vector<size_type> &local_dof_indices, |
| 4021 |
const Table<2, bool> &dof_mask, |
| 4022 |
const bool keep_constrained_entries, |
| 4023 |
ScratchData<number> &scratch_data, |
| 4024 |
SparsityPatternBase &sparsity_pattern) |
| 4025 |
{ |
| 4026 |
// if we got constraints, need to add the diagonal element and, if the |
| 4027 |
// user requested so, also the rest of the entries in rows and columns |
| 4028 |
// that have been left out above |
| 4029 |
if (global_rows.n_constraints() > 0) |
| 4030 |
{ |
| 4031 |
std::vector<std::pair<size_type, size_type>> &cell_entries = |
| 4032 |
scratch_data.new_entries; |
| 4033 |
cell_entries.resize(0); |
| 4034 |
cell_entries.reserve(local_dof_indices.size()); |
| 4035 |
for (size_type i = 0; i < global_rows.n_constraints(); ++i) |
| 4036 |
{ |
| 4037 |
const size_type local_row = global_rows.constraint_origin(i); |
| 4038 |
const size_type global_row = local_dof_indices[local_row]; |
| 4039 |
if (keep_constrained_entries == true) |
| 4040 |
{ |
| 4041 |
for (size_type j = 0; j < local_dof_indices.size(); ++j) |
| 4042 |
{ |
| 4043 |
if (dof_mask(local_row, j) == true) |
| 4044 |
cell_entries.emplace_back(global_row, |
| 4045 |
local_dof_indices[j]); |
| 4046 |
if (dof_mask(j, local_row) == true) |
| 4047 |
cell_entries.emplace_back(local_dof_indices[j], |
| 4048 |
global_row); |
| 4049 |
} |
| 4050 |
} |
| 4051 |
else |
| 4052 |
// don't keep constrained entries - just add the diagonal. |
| 4053 |
cell_entries.emplace_back(global_row, global_row); |
| 4054 |
} |
| 4055 |
sparsity_pattern.add_entries(make_array_view(cell_entries)); |
| 4056 |
} |
| 4057 |
} |
| 4058 |
|
| 4059 |
} // end of namespace AffineConstraints |
| 4060 |
} // end of namespace internal |
| 4061 |
|
| 4062 |
|
| 4063 |
|
| 4064 |
// Basic idea of setting up a list of |
| 4065 |
// all global dofs: first find all rows and columns |
| 4066 |
// that we are going to write touch, |
| 4067 |
// and then go through the |
| 4068 |
// lines and collect all the local rows that |
| 4069 |
// are related to it. |
| 4070 |
template <typename number> |
| 4071 |
void |
| 4072 |
AffineConstraints<number>::make_sorted_row_list( |
| 4073 |
const std::vector<size_type> &local_dof_indices, |
| 4074 |
internal::AffineConstraints::GlobalRowsFromLocal<number> &global_rows) const |
| 4075 |
{ |
| 4076 |
const size_type n_local_dofs = local_dof_indices.size(); |
| 4077 |
AssertDimension(n_local_dofs, global_rows.size()); |
| 4078 |
|
| 4079 |
// when distributing the local data to the global matrix, we can quite |
| 4080 |
// cheaply sort the indices (obviously, this introduces the need for |
| 4081 |
// allocating some memory on the way, but we need to do this only for rows, |
| 4082 |
// whereas the distribution process itself goes over rows and columns). This |
| 4083 |
// has the advantage that when writing into the global matrix, we can make |
| 4084 |
// use of the sortedness. |
| 4085 |
|
| 4086 |
// so the first step is to create a sorted list of all row values that are |
| 4087 |
// possible. these values are either the rows from unconstrained dofs, or |
| 4088 |
// some indices introduced by dofs constrained to a combination of some |
| 4089 |
// other dofs. regarding the data type, choose a <tt>std::vector</tt> of a |
| 4090 |
// pair of unsigned ints (for global columns) and internal data (containing |
| 4091 |
// local columns + possible jumps from constraints). Choosing |
| 4092 |
// <tt>std::map</tt> or anything else M.K. knows of would be much more |
| 4093 |
// expensive here! |
| 4094 |
|
| 4095 |
// cache whether we have to resolve any indirect rows generated from |
| 4096 |
// resolving constrained dofs. |
| 4097 |
size_type added_rows = 0; |
| 4098 |
|
| 4099 |
// first add the indices in an unsorted way and only keep track of the |
| 4100 |
// constraints that appear. They are resolved in a second step. |
| 4101 |
for (size_type i = 0; i < n_local_dofs; ++i) |
| 4102 |
{ |
| 4103 |
if (is_constrained(local_dof_indices[i]) == false) |
| 4104 |
{ |
| 4105 |
global_rows.global_row(added_rows) = local_dof_indices[i]; |
| 4106 |
global_rows.local_row(added_rows++) = i; |
| 4107 |
} |
| 4108 |
else |
| 4109 |
global_rows.insert_constraint(i); |
| 4110 |
} |
| 4111 |
global_rows.sort(); |
| 4112 |
|
| 4113 |
const size_type n_constrained_rows = n_local_dofs - added_rows; |
| 4114 |
for (size_type i = 0; i < n_constrained_rows; ++i) |
| 4115 |
{ |
| 4116 |
const size_type local_row = global_rows.constraint_origin(i); |
| 4117 |
AssertIndexRange(local_row, n_local_dofs); |
| 4118 |
const size_type global_row = local_dof_indices[local_row]; |
| 4119 |
Assert(is_constrained(global_row), ExcInternalError()); |
| 4120 |
const ConstraintLine &position = |
| 4121 |
lines[lines_cache[calculate_line_index(global_row)]]; |
| 4122 |
if (position.inhomogeneity != number(0.)) |
| 4123 |
global_rows.set_ith_constraint_inhomogeneous(i); |
| 4124 |
for (size_type q = 0; q < position.entries.size(); ++q) |
| 4125 |
global_rows.insert_index(position.entries[q].first, |
| 4126 |
local_row, |
| 4127 |
position.entries[q].second); |
| 4128 |
} |
| 4129 |
} |
| 4130 |
|
| 4131 |
|
| 4132 |
|
| 4133 |
// Same function as before, but now do only extract the global indices that |
| 4134 |
// come from the local ones without storing their origin. Used for sparsity |
| 4135 |
// pattern generation. |
| 4136 |
template <typename number> |
| 4137 |
inline void |
| 4138 |
AffineConstraints<number>::make_sorted_row_list( |
| 4139 |
const std::vector<size_type> &local_dof_indices, |
| 4140 |
std::vector<size_type> &active_dofs) const |
| 4141 |
{ |
| 4142 |
const size_type n_local_dofs = local_dof_indices.size(); |
| 4143 |
size_type added_rows = 0; |
| 4144 |
for (size_type i = 0; i < n_local_dofs; ++i) |
| 4145 |
{ |
| 4146 |
if (is_constrained(local_dof_indices[i]) == false) |
| 4147 |
{ |
| 4148 |
active_dofs[added_rows++] = local_dof_indices[i]; |
| 4149 |
continue; |
| 4150 |
} |
| 4151 |
|
| 4152 |
active_dofs[n_local_dofs - i + added_rows - 1] = i; |
| 4153 |
} |
| 4154 |
std::sort(active_dofs.begin(), active_dofs.begin() + added_rows); |
| 4155 |
|
| 4156 |
const size_type n_constrained_dofs = n_local_dofs - added_rows; |
| 4157 |
for (size_type i = n_constrained_dofs; i > 0; --i) |
| 4158 |
{ |
| 4159 |
const size_type local_row = active_dofs.back(); |
| 4160 |
|
| 4161 |
// remove constrained entry since we are going to resolve it in place |
| 4162 |
active_dofs.pop_back(); |
| 4163 |
const size_type global_row = local_dof_indices[local_row]; |
| 4164 |
const ConstraintLine &position = |
| 4165 |
lines[lines_cache[calculate_line_index(global_row)]]; |
| 4166 |
for (size_type q = 0; q < position.entries.size(); ++q) |
| 4167 |
{ |
| 4168 |
const size_type new_index = position.entries[q].first; |
| 4169 |
// in case all dofs are constrained, we might insert at |
| 4170 |
// active_dofs.begin(), but we should never insert before that |
| 4171 |
AssertIndexRange(i - 1, active_dofs.size() + 1); |
| 4172 |
if (i > active_dofs.size() || |
| 4173 |
active_dofs[active_dofs.size() - i] < new_index) |
| 4174 |
active_dofs.insert(active_dofs.end() - i + 1, new_index); |
| 4175 |
|
| 4176 |
// make binary search to find where to put the new index in order to |
| 4177 |
// keep the list sorted |
| 4178 |
else |
| 4179 |
{ |
| 4180 |
std::vector<size_type>::iterator it = |
| 4181 |
Utilities::lower_bound(active_dofs.begin(), |
| 4182 |
active_dofs.end() - i + 1, |
| 4183 |
new_index); |
| 4184 |
if (*it != new_index) |
| 4185 |
active_dofs.insert(it, new_index); |
| 4186 |
} |
| 4187 |
} |
| 4188 |
} |
| 4189 |
} |
| 4190 |
|
| 4191 |
|
| 4192 |
|
| 4193 |
// Resolve the constraints from the vector and apply inhomogeneities. |
| 4194 |
template <typename number> |
| 4195 |
template <typename MatrixScalar, typename VectorScalar> |
| 4196 |
inline typename ProductType<VectorScalar, MatrixScalar>::type |
| 4197 |
AffineConstraints<number>::resolve_vector_entry( |
| 4198 |
const size_type i, |
| 4199 |
const internal::AffineConstraints::GlobalRowsFromLocal<number> &global_rows, |
| 4200 |
const Vector<VectorScalar> &local_vector, |
| 4201 |
const std::vector<size_type> &local_dof_indices, |
| 4202 |
const FullMatrix<MatrixScalar> &local_matrix) const |
| 4203 |
{ |
| 4204 |
const size_type loc_row = global_rows.local_row(i); |
| 4205 |
const size_type n_inhomogeneous_rows = global_rows.n_inhomogeneities(); |
| 4206 |
typename ProductType<VectorScalar, MatrixScalar>::type val = 0; |
| 4207 |
// has a direct contribution from some local entry. If we have inhomogeneous |
| 4208 |
// constraints, compute the contribution of the inhomogeneity in the current |
| 4209 |
// row. |
| 4210 |
if (loc_row != numbers::invalid_size_type) |
| 4211 |
{ |
| 4212 |
val = local_vector(loc_row); |
| 4213 |
for (size_type i = 0; i < n_inhomogeneous_rows; ++i) |
| 4214 |
val -= (local_matrix(loc_row, global_rows.constraint_origin(i)) * |
| 4215 |
lines[lines_cache[calculate_line_index( |
| 4216 |
local_dof_indices[global_rows.constraint_origin(i)])]] |
| 4217 |
.inhomogeneity); |
| 4218 |
} |
| 4219 |
|
| 4220 |
// go through the indirect contributions |
| 4221 |
for (size_type q = 0; q < global_rows.size(i); ++q) |
| 4222 |
{ |
| 4223 |
const size_type loc_row_q = global_rows.local_row(i, q); |
| 4224 |
typename ProductType<VectorScalar, MatrixScalar>::type add_this = |
| 4225 |
local_vector(loc_row_q); |
| 4226 |
for (size_type k = 0; k < n_inhomogeneous_rows; ++k) |
| 4227 |
add_this -= |
| 4228 |
(local_matrix(loc_row_q, global_rows.constraint_origin(k)) * |
| 4229 |
lines[lines_cache[calculate_line_index( |
| 4230 |
local_dof_indices[global_rows.constraint_origin(k)])]] |
| 4231 |
.inhomogeneity); |
| 4232 |
val += add_this * global_rows.constraint_value(i, q); |
| 4233 |
} |
| 4234 |
return val; |
| 4235 |
} |
| 4236 |
|
| 4237 |
|
| 4238 |
|
| 4239 |
// internal implementation for distribute_local_to_global for standard |
| 4240 |
// (non-block) matrices |
| 4241 |
template <typename number> |
| 4242 |
template <typename MatrixType, typename VectorType> |
| 4243 |
void |
| 4244 |
AffineConstraints<number>::distribute_local_to_global( |
| 4245 |
const FullMatrix<number> &local_matrix, |
| 4246 |
const Vector<number> &local_vector, |
| 4247 |
const std::vector<size_type> &local_dof_indices, |
| 4248 |
MatrixType &global_matrix, |
| 4249 |
VectorType &global_vector, |
| 4250 |
const bool use_inhomogeneities_for_rhs, |
| 4251 |
const std::bool_constant<false>) const |
| 4252 |
{ |
| 4253 |
// FIXME: static_assert MatrixType::value_type == number |
| 4254 |
|
| 4255 |
// check whether we work on real vectors or we just used a dummy when |
| 4256 |
// calling the other function above. |
| 4257 |
const bool use_vectors = |
| 4258 |
(local_vector.size() == 0 && global_vector.size() == 0) ? false : true; |
| 4259 |
const bool use_dealii_matrix = |
| 4260 |
std::is_same_v<MatrixType, SparseMatrix<number>>; |
| 4261 |
|
| 4262 |
AssertDimension(local_matrix.n(), local_dof_indices.size()); |
| 4263 |
AssertDimension(local_matrix.m(), local_dof_indices.size()); |
| 4264 |
Assert(global_vector.has_ghost_elements() == false, ExcGhostsPresent()); |
| 4265 |
Assert(global_matrix.m() == global_matrix.n(), ExcNotQuadratic()); |
| 4266 |
if (use_vectors == true) |
| 4267 |
{ |
| 4268 |
AssertDimension(local_matrix.m(), local_vector.size()); |
| 4269 |
AssertDimension(global_matrix.m(), global_vector.size()); |
| 4270 |
} |
| 4271 |
Assert(lines.empty() || sorted == true, ExcMatrixNotClosed()); |
| 4272 |
|
| 4273 |
const size_type n_local_dofs = local_dof_indices.size(); |
| 4274 |
|
| 4275 |
typename internal::AffineConstraints::ScratchDataAccessor<number> |
| 4276 |
scratch_data(this->scratch_data); |
| 4277 |
|
| 4278 |
internal::AffineConstraints::GlobalRowsFromLocal<number> &global_rows = |
| 4279 |
scratch_data->global_rows; |
| 4280 |
global_rows.reinit(n_local_dofs); |
| 4281 |
make_sorted_row_list(local_dof_indices, global_rows); |
| 4282 |
|
| 4283 |
const size_type n_actual_dofs = global_rows.size(); |
| 4284 |
|
| 4285 |
// create arrays for the column data (indices and values) that will then be |
| 4286 |
// written into the matrix. Shortcut for deal.II sparse matrix. We can use |
| 4287 |
// the scratch data if we have a double matrix. Otherwise, we need to create |
| 4288 |
// an array in any case since we cannot know about the actual data type in |
| 4289 |
// the AffineConstraints class (unless we do cast). This involves a little |
| 4290 |
// bit of logic to determine the type of the matrix value. |
| 4291 |
std::vector<size_type> &cols = scratch_data->columns; |
| 4292 |
std::vector<number> &vals = scratch_data->values; |
| 4293 |
// create arrays for writing into the vector as well |
| 4294 |
std::vector<size_type> &vector_indices = scratch_data->vector_indices; |
| 4295 |
std::vector<typename VectorType::value_type> &vector_values = |
| 4296 |
scratch_data->vector_values; |
| 4297 |
vector_indices.resize(n_actual_dofs); |
| 4298 |
vector_values.resize(n_actual_dofs); |
| 4299 |
SparseMatrix<number> *sparse_matrix = |
| 4300 |
dynamic_cast<SparseMatrix<number> *>(&global_matrix); |
| 4301 |
if (use_dealii_matrix == false) |
| 4302 |
{ |
| 4303 |
cols.resize(n_actual_dofs); |
| 4304 |
vals.resize(n_actual_dofs); |
| 4305 |
} |
| 4306 |
else |
| 4307 |
Assert(sparse_matrix != nullptr, ExcInternalError()); |
| 4308 |
|
| 4309 |
// now do the actual job. go through all the global rows that we will touch |
| 4310 |
// and call resolve_matrix_row for each of those. |
| 4311 |
size_type local_row_n = 0; |
| 4312 |
for (size_type i = 0; i < n_actual_dofs; ++i) |
| 4313 |
{ |
| 4314 |
const size_type row = global_rows.global_row(i); |
| 4315 |
|
| 4316 |
// calculate all the data that will be written into the matrix row. |
| 4317 |
if (use_dealii_matrix == false) |
| 4318 |
{ |
| 4319 |
size_type *col_ptr = cols.data(); |
| 4320 |
// cast is uncritical here and only used to avoid compiler |
| 4321 |
// warnings. We never access a non-double array |
| 4322 |
number *val_ptr = vals.data(); |
| 4323 |
internal::AffineConstraints::resolve_matrix_row(global_rows, |
| 4324 |
global_rows, |
| 4325 |
i, |
| 4326 |
0, |
| 4327 |
n_actual_dofs, |
| 4328 |
local_matrix, |
| 4329 |
col_ptr, |
| 4330 |
val_ptr); |
| 4331 |
const size_type n_values = col_ptr - cols.data(); |
| 4332 |
if (n_values > 0) |
| 4333 |
global_matrix.add(row, |
| 4334 |
n_values, |
| 4335 |
cols.data(), |
| 4336 |
vals.data(), |
| 4337 |
/* elide zero additions */ false, |
| 4338 |
/* sorted by column index */ true); |
| 4339 |
} |
| 4340 |
else |
| 4341 |
internal::AffineConstraints::resolve_matrix_row( |
| 4342 |
global_rows, i, 0, n_actual_dofs, local_matrix, sparse_matrix); |
| 4343 |
|
| 4344 |
// now to the vectors. besides doing the same job as we did above (i.e., |
| 4345 |
// distribute the content of the local vector into the global one), need |
| 4346 |
// to account for inhomogeneities here: this corresponds to eliminating |
| 4347 |
// the respective column in the local matrix with value on the right |
| 4348 |
// hand side. |
| 4349 |
if (use_vectors == true) |
| 4350 |
{ |
| 4351 |
const typename VectorType::value_type val = resolve_vector_entry( |
| 4352 |
i, global_rows, local_vector, local_dof_indices, local_matrix); |
| 4353 |
AssertIsFinite(val); |
| 4354 |
|
| 4355 |
if (val != typename VectorType::value_type()) |
| 4356 |
{ |
| 4357 |
vector_indices[local_row_n] = row; |
| 4358 |
vector_values[local_row_n] = val; |
| 4359 |
++local_row_n; |
| 4360 |
} |
| 4361 |
} |
| 4362 |
} |
| 4363 |
// Drop the elements of vector_indices and vector_values that we do not use |
| 4364 |
// (we may always elide writing zero values to vectors) |
| 4365 |
const size_type n_local_rows = local_row_n; |
| 4366 |
vector_indices.resize(n_local_rows); |
| 4367 |
vector_values.resize(n_local_rows); |
| 4368 |
|
| 4369 |
// While the standard case is that these types are equal, they need not be, so |
| 4370 |
// only do a bulk update if they are. Note that the types in the arguments to |
| 4371 |
// add must be equal if we have a Trilinos or PETSc vector but do not have to |
| 4372 |
// be if we have a deal.II native vector: one could further optimize this for |
| 4373 |
// Vector, LinearAlgebra::distributed::vector, etc. |
| 4374 |
if constexpr (std::is_same_v<typename VectorType::value_type, number>) |
| 4375 |
{ |
| 4376 |
global_vector.add(vector_indices, vector_values); |
| 4377 |
} |
| 4378 |
else |
| 4379 |
{ |
| 4380 |
for (size_type row_n = 0; row_n < n_local_rows; ++row_n) |
| 4381 |
{ |
| 4382 |
global_vector(vector_indices[row_n]) += |
| 4383 |
static_cast<typename VectorType::value_type>(vector_values[row_n]); |
| 4384 |
} |
| 4385 |
} |
| 4386 |
|
| 4387 |
internal::AffineConstraints::set_matrix_diagonals( |
| 4388 |
global_rows, |
| 4389 |
local_dof_indices, |
| 4390 |
local_matrix, |
| 4391 |
*this, |
| 4392 |
global_matrix, |
| 4393 |
global_vector, |
| 4394 |
use_inhomogeneities_for_rhs); |
| 4395 |
} |
| 4396 |
|
| 4397 |
|
| 4398 |
|
| 4399 |
// similar function as above, but now specialized for block matrices. See the |
| 4400 |
// other function for additional comments. |
| 4401 |
template <typename number> |
| 4402 |
template <typename MatrixType, typename VectorType> |
| 4403 |
void |
| 4404 |
AffineConstraints<number>::distribute_local_to_global( |
| 4405 |
const FullMatrix<number> &local_matrix, |
| 4406 |
const Vector<number> &local_vector, |
| 4407 |
const std::vector<size_type> &local_dof_indices, |
| 4408 |
MatrixType &global_matrix, |
| 4409 |
VectorType &global_vector, |
| 4410 |
const bool use_inhomogeneities_for_rhs, |
| 4411 |
const std::bool_constant<true>) const |
| 4412 |
{ |
| 4413 |
const bool use_vectors = |
| 4414 |
(local_vector.size() == 0 && global_vector.size() == 0) ? false : true; |
| 4415 |
const bool use_dealii_matrix = |
| 4416 |
std::is_same_v<MatrixType, BlockSparseMatrix<number>>; |
| 4417 |
|
| 4418 |
AssertDimension(local_matrix.n(), local_dof_indices.size()); |
| 4419 |
AssertDimension(local_matrix.m(), local_dof_indices.size()); |
| 4420 |
Assert(global_vector.has_ghost_elements() == false, ExcGhostsPresent()); |
| 4421 |
Assert(global_matrix.m() == global_matrix.n(), ExcNotQuadratic()); |
| 4422 |
Assert(global_matrix.n_block_rows() == global_matrix.n_block_cols(), |
| 4423 |
ExcNotQuadratic()); |
| 4424 |
if (use_vectors == true) |
| 4425 |
{ |
| 4426 |
AssertDimension(local_matrix.m(), local_vector.size()); |
| 4427 |
AssertDimension(global_matrix.m(), global_vector.size()); |
| 4428 |
} |
| 4429 |
Assert(sorted == true, ExcMatrixNotClosed()); |
| 4430 |
|
| 4431 |
typename internal::AffineConstraints::ScratchDataAccessor<number> |
| 4432 |
scratch_data(this->scratch_data); |
| 4433 |
|
| 4434 |
const size_type n_local_dofs = local_dof_indices.size(); |
| 4435 |
internal::AffineConstraints::GlobalRowsFromLocal<number> &global_rows = |
| 4436 |
scratch_data->global_rows; |
| 4437 |
global_rows.reinit(n_local_dofs); |
| 4438 |
|
| 4439 |
make_sorted_row_list(local_dof_indices, global_rows); |
| 4440 |
const size_type n_actual_dofs = global_rows.size(); |
| 4441 |
|
| 4442 |
std::vector<size_type> &global_indices = scratch_data->vector_indices; |
| 4443 |
if (use_vectors == true) |
| 4444 |
{ |
| 4445 |
global_indices.resize(n_actual_dofs); |
| 4446 |
for (size_type i = 0; i < n_actual_dofs; ++i) |
| 4447 |
global_indices[i] = global_rows.global_row(i); |
| 4448 |
} |
| 4449 |
|
| 4450 |
// additional construct that also takes care of block indices. |
| 4451 |
const size_type num_blocks = global_matrix.n_block_rows(); |
| 4452 |
std::vector<size_type> &block_starts = scratch_data->block_starts; |
| 4453 |
block_starts.resize(num_blocks + 1); |
| 4454 |
internal::AffineConstraints::make_block_starts(global_matrix, |
| 4455 |
global_rows, |
| 4456 |
block_starts); |
| 4457 |
|
| 4458 |
std::vector<size_type> &cols = scratch_data->columns; |
| 4459 |
std::vector<number> &vals = scratch_data->values; |
| 4460 |
if (use_dealii_matrix == false) |
| 4461 |
{ |
| 4462 |
cols.resize(n_actual_dofs); |
| 4463 |
vals.resize(n_actual_dofs); |
| 4464 |
} |
| 4465 |
|
| 4466 |
// the basic difference to the non-block variant from now onwards is that we |
| 4467 |
// go through the blocks of the matrix separately, which allows us to set |
| 4468 |
// the block entries individually |
| 4469 |
for (size_type block = 0; block < num_blocks; ++block) |
| 4470 |
{ |
| 4471 |
const size_type next_block = block_starts[block + 1]; |
| 4472 |
for (size_type i = block_starts[block]; i < next_block; ++i) |
| 4473 |
{ |
| 4474 |
const size_type row = global_rows.global_row(i); |
| 4475 |
|
| 4476 |
for (size_type block_col = 0; block_col < num_blocks; ++block_col) |
| 4477 |
{ |
| 4478 |
const size_type start_block = block_starts[block_col], |
| 4479 |
end_block = block_starts[block_col + 1]; |
| 4480 |
if (use_dealii_matrix == false) |
| 4481 |
{ |
| 4482 |
size_type *col_ptr = cols.data(); |
| 4483 |
number *val_ptr = vals.data(); |
| 4484 |
internal::AffineConstraints::resolve_matrix_row(global_rows, |
| 4485 |
global_rows, |
| 4486 |
i, |
| 4487 |
start_block, |
| 4488 |
end_block, |
| 4489 |
local_matrix, |
| 4490 |
col_ptr, |
| 4491 |
val_ptr); |
| 4492 |
const size_type n_values = col_ptr - cols.data(); |
| 4493 |
if (n_values > 0) |
| 4494 |
global_matrix.block(block, block_col) |
| 4495 |
.add( |
| 4496 |
row, n_values, cols.data(), vals.data(), false, true); |
| 4497 |
} |
| 4498 |
else |
| 4499 |
{ |
| 4500 |
SparseMatrix<number> *sparse_matrix = |
| 4501 |
dynamic_cast<SparseMatrix<number> *>( |
| 4502 |
&global_matrix.block(block, block_col)); |
| 4503 |
Assert(sparse_matrix != nullptr, ExcInternalError()); |
| 4504 |
internal::AffineConstraints::resolve_matrix_row( |
| 4505 |
global_rows, |
| 4506 |
i, |
| 4507 |
start_block, |
| 4508 |
end_block, |
| 4509 |
local_matrix, |
| 4510 |
sparse_matrix); |
| 4511 |
} |
| 4512 |
} |
| 4513 |
|
| 4514 |
if (use_vectors == true) |
| 4515 |
{ |
| 4516 |
const number val = resolve_vector_entry( |
| 4517 |
i, global_rows, local_vector, local_dof_indices, local_matrix); |
| 4518 |
|
| 4519 |
if (val != number()) |
| 4520 |
global_vector(global_indices[i]) += |
| 4521 |
static_cast<typename VectorType::value_type>(val); |
| 4522 |
} |
| 4523 |
} |
| 4524 |
} |
| 4525 |
|
| 4526 |
internal::AffineConstraints::set_matrix_diagonals( |
| 4527 |
global_rows, |
| 4528 |
local_dof_indices, |
| 4529 |
local_matrix, |
| 4530 |
*this, |
| 4531 |
global_matrix, |
| 4532 |
global_vector, |
| 4533 |
use_inhomogeneities_for_rhs); |
| 4534 |
} |
| 4535 |
|
| 4536 |
|
| 4537 |
|
| 4538 |
template <typename number> |
| 4539 |
template <typename MatrixType> |
| 4540 |
void |
| 4541 |
AffineConstraints<number>::distribute_local_to_global( |
| 4542 |
const FullMatrix<number> &local_matrix, |
| 4543 |
const std::vector<size_type> &row_indices, |
| 4544 |
const std::vector<size_type> &col_indices, |
| 4545 |
MatrixType &global_matrix) const |
| 4546 |
{ |
| 4547 |
distribute_local_to_global( |
| 4548 |
local_matrix, row_indices, *this, col_indices, global_matrix); |
| 4549 |
} |
| 4550 |
|
| 4551 |
|
| 4552 |
|
| 4553 |
template <typename number> |
| 4554 |
template <typename MatrixType> |
| 4555 |
void |
| 4556 |
AffineConstraints<number>::distribute_local_to_global( |
| 4557 |
const FullMatrix<number> &local_matrix, |
| 4558 |
const std::vector<size_type> &row_indices, |
| 4559 |
const AffineConstraints<number> &col_constraint_matrix, |
| 4560 |
const std::vector<size_type> &col_indices, |
| 4561 |
MatrixType &global_matrix) const |
| 4562 |
{ |
| 4563 |
AssertDimension(local_matrix.m(), row_indices.size()); |
| 4564 |
AssertDimension(local_matrix.n(), col_indices.size()); |
| 4565 |
|
| 4566 |
const size_type n_local_row_dofs = row_indices.size(); |
| 4567 |
const size_type n_local_col_dofs = col_indices.size(); |
| 4568 |
|
| 4569 |
typename internal::AffineConstraints::ScratchDataAccessor< |
| 4570 |
typename MatrixType::value_type> |
| 4571 |
scratch_data(this->scratch_data); |
| 4572 |
|
| 4573 |
internal::AffineConstraints::GlobalRowsFromLocal<number> &global_rows = |
| 4574 |
scratch_data->global_rows; |
| 4575 |
global_rows.reinit(n_local_row_dofs); |
| 4576 |
|
| 4577 |
internal::AffineConstraints::GlobalRowsFromLocal<number> &global_cols = |
| 4578 |
scratch_data->global_columns; |
| 4579 |
global_cols.reinit(n_local_col_dofs); |
| 4580 |
|
| 4581 |
make_sorted_row_list(row_indices, global_rows); |
| 4582 |
col_constraint_matrix.make_sorted_row_list(col_indices, global_cols); |
| 4583 |
|
| 4584 |
const size_type n_actual_row_dofs = global_rows.size(); |
| 4585 |
const size_type n_actual_col_dofs = global_cols.size(); |
| 4586 |
|
| 4587 |
// create arrays for the column data (indices and values) that will then be |
| 4588 |
// written into the matrix. Shortcut for deal.II sparse matrix |
| 4589 |
std::vector<size_type> &cols = scratch_data->columns; |
| 4590 |
std::vector<number> &vals = scratch_data->values; |
| 4591 |
cols.resize(n_actual_col_dofs); |
| 4592 |
vals.resize(n_actual_col_dofs); |
| 4593 |
|
| 4594 |
// now do the actual job. |
| 4595 |
for (size_type i = 0; i < n_actual_row_dofs; ++i) |
| 4596 |
{ |
| 4597 |
const size_type row = global_rows.global_row(i); |
| 4598 |
|
| 4599 |
// calculate all the data that will be written into the matrix row. |
| 4600 |
size_type *col_ptr = cols.data(); |
| 4601 |
number *val_ptr = vals.data(); |
| 4602 |
internal::AffineConstraints::resolve_matrix_row(global_rows, |
| 4603 |
global_cols, |
| 4604 |
i, |
| 4605 |
0, |
| 4606 |
n_actual_col_dofs, |
| 4607 |
local_matrix, |
| 4608 |
col_ptr, |
| 4609 |
val_ptr); |
| 4610 |
const size_type n_values = col_ptr - cols.data(); |
| 4611 |
if (n_values > 0) |
| 4612 |
global_matrix.add(row, n_values, cols.data(), vals.data(), false, true); |
| 4613 |
} |
| 4614 |
} |
| 4615 |
|
| 4616 |
|
| 4617 |
|
| 4618 |
template <typename number> |
| 4619 |
void |
| 4620 |
AffineConstraints<number>::add_entries_local_to_global( |
| 4621 |
const std::vector<size_type> &local_dof_indices, |
| 4622 |
SparsityPatternBase &sparsity_pattern, |
| 4623 |
const bool keep_constrained_entries, |
| 4624 |
const Table<2, bool> &dof_mask) const |
| 4625 |
{ |
| 4626 |
Assert(sparsity_pattern.n_rows() == sparsity_pattern.n_cols(), |
| 4627 |
ExcNotQuadratic()); |
| 4628 |
|
| 4629 |
const size_type n_local_dofs = local_dof_indices.size(); |
| 4630 |
typename internal::AffineConstraints::ScratchDataAccessor<number> |
| 4631 |
scratch_data(this->scratch_data); |
| 4632 |
|
| 4633 |
const bool dof_mask_is_active = (dof_mask.n_rows() == n_local_dofs); |
| 4634 |
if (dof_mask_is_active == true) |
| 4635 |
{ |
| 4636 |
AssertDimension(dof_mask.n_cols(), n_local_dofs); |
| 4637 |
} |
| 4638 |
else |
| 4639 |
{ |
| 4640 |
// if the dof mask is not active, all we have to do is to add some indices |
| 4641 |
// in a matrix format. To do this, we first create an array of all the |
| 4642 |
// indices that are to be added. these indices are the local dof indices |
| 4643 |
// plus some indices that come from constraints. |
| 4644 |
std::vector<size_type> &actual_dof_indices = scratch_data->columns; |
| 4645 |
actual_dof_indices.resize(n_local_dofs); |
| 4646 |
make_sorted_row_list(local_dof_indices, actual_dof_indices); |
| 4647 |
const size_type n_actual_dofs = actual_dof_indices.size(); |
| 4648 |
|
| 4649 |
// now add the indices we collected above to the sparsity pattern. Very |
| 4650 |
// easy here - just add the same array to all the rows... |
| 4651 |
for (size_type i = 0; i < n_actual_dofs; ++i) |
| 4652 |
sparsity_pattern.add_row_entries(actual_dof_indices[i], |
| 4653 |
make_array_view(actual_dof_indices), |
| 4654 |
true); |
| 4655 |
|
| 4656 |
// need to add the whole row and column structure in case we keep |
| 4657 |
// constrained entries. |
| 4658 |
std::vector<std::pair<size_type, size_type>> &cell_entries = |
| 4659 |
scratch_data->new_entries; |
| 4660 |
cell_entries.resize(0); |
| 4661 |
cell_entries.reserve(n_local_dofs); |
| 4662 |
for (size_type i = 0; i < n_local_dofs; ++i) |
| 4663 |
if (is_constrained(local_dof_indices[i])) |
| 4664 |
{ |
| 4665 |
if (keep_constrained_entries == true) |
| 4666 |
for (size_type j = 0; j < n_local_dofs; ++j) |
| 4667 |
{ |
| 4668 |
cell_entries.emplace_back(local_dof_indices[i], |
| 4669 |
local_dof_indices[j]); |
| 4670 |
cell_entries.emplace_back(local_dof_indices[j], |
| 4671 |
local_dof_indices[i]); |
| 4672 |
} |
| 4673 |
else |
| 4674 |
{ |
| 4675 |
cell_entries.emplace_back(local_dof_indices[i], |
| 4676 |
local_dof_indices[i]); |
| 4677 |
} |
| 4678 |
} |
| 4679 |
sparsity_pattern.add_entries(make_array_view(cell_entries)); |
| 4680 |
|
| 4681 |
return; |
| 4682 |
} |
| 4683 |
|
| 4684 |
// complicated case: we need to filter out some indices. then the function |
| 4685 |
// gets similar to the function for distributing matrix entries, see there |
| 4686 |
// for additional comments. |
| 4687 |
internal::AffineConstraints::GlobalRowsFromLocal<number> &global_rows = |
| 4688 |
scratch_data->global_rows; |
| 4689 |
global_rows.reinit(n_local_dofs); |
| 4690 |
make_sorted_row_list(local_dof_indices, global_rows); |
| 4691 |
const size_type n_actual_dofs = global_rows.size(); |
| 4692 |
|
| 4693 |
// create arrays for the column indices that will then be written into the |
| 4694 |
// sparsity pattern. |
| 4695 |
std::vector<size_type> &cols = scratch_data->columns; |
| 4696 |
cols.resize(n_actual_dofs); |
| 4697 |
|
| 4698 |
for (size_type i = 0; i < n_actual_dofs; ++i) |
| 4699 |
{ |
| 4700 |
std::vector<size_type>::iterator col_ptr = cols.begin(); |
| 4701 |
const size_type row = global_rows.global_row(i); |
| 4702 |
internal::AffineConstraints::resolve_matrix_row( |
| 4703 |
global_rows, i, 0, n_actual_dofs, dof_mask, col_ptr); |
| 4704 |
|
| 4705 |
// finally, write all the information that accumulated under the given |
| 4706 |
// process into the global matrix row and into the vector |
| 4707 |
if (col_ptr != cols.begin()) |
| 4708 |
sparsity_pattern.add_row_entries(row, |
| 4709 |
make_array_view(cols.begin(), col_ptr), |
| 4710 |
true); |
| 4711 |
} |
| 4712 |
internal::AffineConstraints::set_sparsity_diagonals(global_rows, |
| 4713 |
local_dof_indices, |
| 4714 |
dof_mask, |
| 4715 |
keep_constrained_entries, |
| 4716 |
*scratch_data, |
| 4717 |
sparsity_pattern); |
| 4718 |
} |
| 4719 |
|
| 4720 |
|
| 4721 |
|
| 4722 |
template <typename number> |
| 4723 |
void |
| 4724 |
AffineConstraints<number>::add_entries_local_to_global( |
| 4725 |
const std::vector<size_type> &row_indices, |
| 4726 |
const AffineConstraints<number> &col_constraints, |
| 4727 |
const std::vector<size_type> &col_indices, |
| 4728 |
SparsityPatternBase &sparsity_pattern, |
| 4729 |
const bool keep_constrained_entries, |
| 4730 |
const Table<2, bool> &dof_mask) const |
| 4731 |
{ |
| 4732 |
const size_type n_local_rows = row_indices.size(); |
| 4733 |
const size_type n_local_cols = col_indices.size(); |
| 4734 |
|
| 4735 |
// Early return if the length of row and column indices is zero, relevant |
| 4736 |
// for the usage with FE_Nothing. |
| 4737 |
if (n_local_cols == 0 && n_local_rows == 0) |
| 4738 |
return; |
| 4739 |
|
| 4740 |
typename internal::AffineConstraints::ScratchDataAccessor<number> |
| 4741 |
scratch_data(this->scratch_data); |
| 4742 |
std::vector<std::pair<size_type, size_type>> &cell_entries = |
| 4743 |
scratch_data->new_entries; |
| 4744 |
cell_entries.resize(0); |
| 4745 |
cell_entries.reserve(n_local_rows * n_local_cols); |
| 4746 |
|
| 4747 |
// if constrained entries should be kept, need to add rows and columns of |
| 4748 |
// those to the sparsity pattern |
| 4749 |
if (keep_constrained_entries == true) |
| 4750 |
{ |
| 4751 |
for (const size_type row_index : row_indices) |
| 4752 |
if (is_constrained(row_index)) |
| 4753 |
for (const size_type col_index : col_indices) |
| 4754 |
cell_entries.emplace_back(row_index, col_index); |
| 4755 |
for (const size_type col_index : col_indices) |
| 4756 |
if (col_constraints.is_constrained(col_index)) |
| 4757 |
for (const size_type row_index : row_indices) |
| 4758 |
cell_entries.emplace_back(row_index, col_index); |
| 4759 |
sparsity_pattern.add_entries(make_array_view(cell_entries)); |
| 4760 |
} |
| 4761 |
|
| 4762 |
// if the dof mask is not active, all we have to do is to add some indices |
| 4763 |
// in a matrix format. To do this, we first create an array of all the |
| 4764 |
// indices that are to be added. these indices are the local dof indices |
| 4765 |
// plus some indices that come from constraints. |
| 4766 |
const bool dof_mask_is_active = |
| 4767 |
dof_mask.n_rows() == n_local_rows && dof_mask.n_cols() == n_local_cols; |
| 4768 |
if (dof_mask_is_active == false) |
| 4769 |
{ |
| 4770 |
std::vector<size_type> &rows = scratch_data->rows; |
| 4771 |
std::vector<size_type> &cols = scratch_data->columns; |
| 4772 |
rows.resize(n_local_rows); |
| 4773 |
cols.resize(n_local_cols); |
| 4774 |
// TODO these fills may not be necessary: previously we assumed all zeros |
| 4775 |
// which seems incorrect. At least this way things will crash |
| 4776 |
std::fill(rows.begin(), |
| 4777 |
rows.end(), |
| 4778 |
std::numeric_limits<size_type>::max()); |
| 4779 |
std::fill(cols.begin(), |
| 4780 |
cols.end(), |
| 4781 |
std::numeric_limits<size_type>::max()); |
| 4782 |
make_sorted_row_list(row_indices, rows); |
| 4783 |
col_constraints.make_sorted_row_list(col_indices, cols); |
| 4784 |
const size_type n_actual_rows = rows.size(); |
| 4785 |
|
| 4786 |
// now add the indices we collected above to the sparsity pattern. Very |
| 4787 |
// easy here - just add the same array to all the rows... |
| 4788 |
for (size_type i = 0; i < n_actual_rows; ++i) |
| 4789 |
sparsity_pattern.add_row_entries(rows[i], make_array_view(cols), true); |
| 4790 |
return; |
| 4791 |
} |
| 4792 |
|
| 4793 |
// TODO: implement this |
| 4794 |
DEAL_II_NOT_IMPLEMENTED(); |
| 4795 |
} |
| 4796 |
|
| 4797 |
|
| 4798 |
|
| 4799 |
template <typename number> |
| 4800 |
void |
| 4801 |
AffineConstraints<number>::add_entries_local_to_global( |
| 4802 |
const std::vector<size_type> &row_indices, |
| 4803 |
const std::vector<size_type> &col_indices, |
| 4804 |
SparsityPatternBase &sparsity_pattern, |
| 4805 |
const bool keep_constrained_entries, |
| 4806 |
const Table<2, bool> &dof_mask) const |
| 4807 |
{ |
| 4808 |
// Call the function with the same name that takes a column constraint as well |
| 4809 |
add_entries_local_to_global(row_indices, |
| 4810 |
*this, |
| 4811 |
col_indices, |
| 4812 |
sparsity_pattern, |
| 4813 |
keep_constrained_entries, |
| 4814 |
dof_mask); |
| 4815 |
} |
| 4816 |
|
| 4817 |
DEAL_II_NAMESPACE_CLOSE |
| 4818 |
|
| 4819 |
#endif |