| 1 |
// ------------------------------------------------------------------------ |
| 2 |
// |
| 3 |
// SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 |
// Copyright (C) 2012 - 2024 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_matrix_free_dof_info_templates_h |
| 16 |
#define dealii_matrix_free_dof_info_templates_h |
| 17 |
|
| 18 |
|
| 19 |
#include <deal.II/base/config.h> |
| 20 |
|
| 21 |
#include <deal.II/base/floating_point_comparator.h> |
| 22 |
#include <deal.II/base/memory_consumption.h> |
| 23 |
#include <deal.II/base/parallel.h> |
| 24 |
|
| 25 |
#include <deal.II/lac/affine_constraints.h> |
| 26 |
|
| 27 |
#include <deal.II/matrix_free/constraint_info.h> |
| 28 |
#include <deal.II/matrix_free/dof_info.h> |
| 29 |
#include <deal.II/matrix_free/hanging_nodes_internal.h> |
| 30 |
#include <deal.II/matrix_free/task_info.h> |
| 31 |
|
| 32 |
#include <limits> |
| 33 |
|
| 34 |
DEAL_II_NAMESPACE_OPEN |
| 35 |
|
| 36 |
namespace internal |
| 37 |
{ |
| 38 |
namespace MatrixFreeFunctions |
| 39 |
{ |
| 40 |
template <typename number> |
| 41 |
void |
| 42 |
DoFInfo::read_dof_indices( |
| 43 |
const std::vector<types::global_dof_index> &local_indices_resolved, |
| 44 |
const std::vector<types::global_dof_index> &local_indices, |
| 45 |
const bool cell_has_hanging_nodes, |
| 46 |
const dealii::AffineConstraints<number> &constraints, |
| 47 |
const unsigned int cell_number, |
| 48 |
ConstraintValues<double> &constraint_values, |
| 49 |
bool &cell_at_subdomain_boundary) |
| 50 |
{ |
| 51 |
Assert(vector_partitioner.get() != nullptr, ExcInternalError()); |
| 52 |
const types::global_dof_index first_owned = |
| 53 |
vector_partitioner->local_range().first; |
| 54 |
const types::global_dof_index last_owned = |
| 55 |
vector_partitioner->local_range().second; |
| 56 |
Assert(last_owned - first_owned < |
| 57 |
std::numeric_limits<unsigned int>::max(), |
| 58 |
ExcMessage("The size local range of owned indices must not " |
| 59 |
"exceed the size of unsigned int")); |
| 60 |
const unsigned int n_owned = last_owned - first_owned; |
| 61 |
|
| 62 |
Assert(dofs_per_cell.size() == 1 || |
| 63 |
cell_number < cell_active_fe_index.size(), |
| 64 |
ExcInternalError()); |
| 65 |
const unsigned int fe_index = |
| 66 |
dofs_per_cell.size() == 1 ? 0 : cell_active_fe_index[cell_number]; |
| 67 |
const unsigned int dofs_this_cell = dofs_per_cell[fe_index]; |
| 68 |
const unsigned int n_components = start_components.back(); |
| 69 |
unsigned int i = 0; |
| 70 |
for (unsigned int comp = 0; comp < n_components; ++comp) |
| 71 |
{ |
| 72 |
std::pair<unsigned short, unsigned short> constraint_iterator(0, 0); |
| 73 |
const auto next = component_dof_indices_offset[fe_index][comp + 1]; |
| 74 |
for (; i < next; ++i) |
| 75 |
{ |
| 76 |
types::global_dof_index current_dof = local_indices_resolved[i]; |
| 77 |
const auto *entries_ptr = |
| 78 |
constraints.get_constraint_entries(current_dof); |
| 79 |
|
| 80 |
// dof is constrained |
| 81 |
if (entries_ptr != nullptr) |
| 82 |
{ |
| 83 |
// in case we want to access plain indices, we need to know |
| 84 |
// about the location of constrained indices as well (all the |
| 85 |
// other indices are collected by the cases below) |
| 86 |
if (current_dof < first_owned || current_dof >= last_owned) |
| 87 |
{ |
| 88 |
ghost_dofs.push_back(current_dof); |
| 89 |
cell_at_subdomain_boundary = true; |
| 90 |
} |
| 91 |
|
| 92 |
// check whether this dof is identity constrained to another |
| 93 |
// dof. then we can simply insert that dof and there is no |
| 94 |
// need to actually resolve the constraint entries |
| 95 |
const auto &entries = *entries_ptr; |
| 96 |
const types::global_dof_index n_entries = entries.size(); |
| 97 |
if (n_entries == 1 && |
| 98 |
std::abs(entries[0].second - 1.) < |
| 99 |
100 * std::numeric_limits<double>::epsilon()) |
| 100 |
{ |
| 101 |
current_dof = entries[0].first; |
| 102 |
goto no_constraint; |
| 103 |
} |
| 104 |
|
| 105 |
// append a new index to the indicators |
| 106 |
constraint_indicator.push_back(constraint_iterator); |
| 107 |
constraint_indicator.back().second = |
| 108 |
constraint_values.insert_entries(entries); |
| 109 |
|
| 110 |
// reset constraint iterator for next round |
| 111 |
constraint_iterator.first = 0; |
| 112 |
|
| 113 |
// add the local_to_global indices computed in the |
| 114 |
// insert_entries function. transform the index to local index |
| 115 |
// space or mark it as ghost if necessary |
| 116 |
if (n_entries > 0) |
| 117 |
{ |
| 118 |
const std::vector<types::global_dof_index> |
| 119 |
&constraint_indices = |
| 120 |
constraint_values.constraint_indices; |
| 121 |
for (unsigned int j = 0; j < n_entries; ++j) |
| 122 |
{ |
| 123 |
if (constraint_indices[j] < first_owned || |
| 124 |
constraint_indices[j] >= last_owned) |
| 125 |
{ |
| 126 |
dof_indices.push_back(n_owned + |
| 127 |
ghost_dofs.size()); |
| 128 |
|
| 129 |
// collect ghosts so that we can later construct |
| 130 |
// an IndexSet for them. also store whether the |
| 131 |
// current cell is on the boundary |
| 132 |
ghost_dofs.push_back(constraint_indices[j]); |
| 133 |
cell_at_subdomain_boundary = true; |
| 134 |
} |
| 135 |
else |
| 136 |
// not ghost, so transform to the local index space |
| 137 |
// directly |
| 138 |
dof_indices.push_back(static_cast<unsigned int>( |
| 139 |
constraint_indices[j] - first_owned)); |
| 140 |
} |
| 141 |
} |
| 142 |
} |
| 143 |
else |
| 144 |
{ |
| 145 |
no_constraint: |
| 146 |
// Not constrained, we simply have to add the local index to |
| 147 |
// the indices_local_to_global list and increment constraint |
| 148 |
// iterator. transform to local index space/mark as ghost |
| 149 |
if (current_dof < first_owned || current_dof >= last_owned) |
| 150 |
{ |
| 151 |
ghost_dofs.push_back(current_dof); |
| 152 |
current_dof = n_owned + ghost_dofs.size() - 1; |
| 153 |
cell_at_subdomain_boundary = true; |
| 154 |
} |
| 155 |
else |
| 156 |
current_dof -= first_owned; |
| 157 |
|
| 158 |
dof_indices.push_back(static_cast<unsigned int>(current_dof)); |
| 159 |
|
| 160 |
// make sure constraint_iterator.first is always within the |
| 161 |
// bounds of unsigned short |
| 162 |
Assert(constraint_iterator.first < |
| 163 |
(1 << (8 * sizeof(unsigned short))) - 1, |
| 164 |
ExcInternalError()); |
| 165 |
constraint_iterator.first++; |
| 166 |
} |
| 167 |
} |
| 168 |
row_starts[cell_number * n_components + comp + 1].first = |
| 169 |
dof_indices.size(); |
| 170 |
row_starts[cell_number * n_components + comp + 1].second = |
| 171 |
constraint_indicator.size(); |
| 172 |
} |
| 173 |
|
| 174 |
// now to the plain indices: in case we have constraints on this cell, |
| 175 |
// store the indices without the constraints resolve once again |
| 176 |
if (store_plain_indices == true) |
| 177 |
{ |
| 178 |
if (cell_number == 0) |
| 179 |
row_starts_plain_indices.resize( |
| 180 |
(row_starts.size() - 1) / n_components + 1); |
| 181 |
row_starts_plain_indices[cell_number] = plain_dof_indices.size(); |
| 182 |
const bool cell_has_constraints = |
| 183 |
cell_has_hanging_nodes || |
| 184 |
(row_starts[(cell_number + 1) * n_components].second > |
| 185 |
row_starts[cell_number * n_components].second); |
| 186 |
if (cell_has_constraints == true) |
| 187 |
{ |
| 188 |
for (unsigned int i = 0; i < dofs_this_cell; ++i) |
| 189 |
{ |
| 190 |
types::global_dof_index current_dof = local_indices[i]; |
| 191 |
if (current_dof < first_owned || current_dof >= last_owned) |
| 192 |
{ |
| 193 |
ghost_dofs.push_back(current_dof); |
| 194 |
current_dof = n_owned + ghost_dofs.size() - 1; |
| 195 |
cell_at_subdomain_boundary = true; |
| 196 |
} |
| 197 |
else |
| 198 |
current_dof -= first_owned; |
| 199 |
plain_dof_indices.push_back( |
| 200 |
static_cast<unsigned int>(current_dof)); |
| 201 |
} |
| 202 |
} |
| 203 |
} |
| 204 |
} |
| 205 |
|
| 206 |
|
| 207 |
|
| 208 |
template <int dim> |
| 209 |
bool |
| 210 |
DoFInfo::process_hanging_node_constraints( |
| 211 |
const HangingNodes<dim> &hanging_nodes, |
| 212 |
const std::vector<std::vector<unsigned int>> &lexicographic_mapping, |
| 213 |
const unsigned int cell_number, |
| 214 |
const TriaIterator<DoFCellAccessor<dim, dim, false>> &cell, |
| 215 |
std::vector<types::global_dof_index> &dof_indices) |
| 216 |
{ |
| 217 |
if (this->hanging_node_constraint_masks_comp.empty()) |
| 218 |
return false; |
| 219 |
|
| 220 |
// 2) determine the refinement configuration of the cell |
| 221 |
const auto refinement_configuration = |
| 222 |
hanging_nodes.compute_refinement_configuration(cell); |
| 223 |
|
| 224 |
if (refinement_configuration == ConstraintKinds::unconstrained) |
| 225 |
return false; |
| 226 |
|
| 227 |
// 3) update DoF indices of cell for specified components |
| 228 |
hanging_nodes.update_dof_indices(cell, |
| 229 |
{}, |
| 230 |
lexicographic_mapping, |
| 231 |
hanging_node_constraint_masks_comp, |
| 232 |
refinement_configuration, |
| 233 |
dof_indices); |
| 234 |
|
| 235 |
hanging_node_constraint_masks[cell_number] = |
| 236 |
compress(refinement_configuration, dim); |
| 237 |
|
| 238 |
return true; |
| 239 |
} |
| 240 |
|
| 241 |
|
| 242 |
|
| 243 |
template <int length> |
| 244 |
void |
| 245 |
DoFInfo::compute_face_index_compression( |
| 246 |
const std::vector<FaceToCellTopology<length>> &faces, |
| 247 |
const bool hold_all_faces_to_owned_cells) |
| 248 |
{ |
| 249 |
AssertDimension(length, vectorization_length); |
| 250 |
|
| 251 |
index_storage_variants[dof_access_face_interior].resize( |
| 252 |
faces.size(), IndexStorageVariants::full); |
| 253 |
dof_indices_contiguous[dof_access_face_interior].resize( |
| 254 |
faces.size() * length, numbers::invalid_unsigned_int); |
| 255 |
dof_indices_interleave_strides[dof_access_face_interior].resize( |
| 256 |
faces.size() * length, numbers::invalid_unsigned_int); |
| 257 |
n_vectorization_lanes_filled[dof_access_face_interior].resize( |
| 258 |
faces.size()); |
| 259 |
|
| 260 |
// all inner faces come before the boundary faces |
| 261 |
unsigned int n_inner_faces = 0; |
| 262 |
for (; n_inner_faces < faces.size(); ++n_inner_faces) |
| 263 |
if (faces[n_inner_faces].cells_exterior[0] == |
| 264 |
numbers::invalid_unsigned_int) |
| 265 |
break; |
| 266 |
|
| 267 |
// all boundary faces come after the inner faces and before the ghosted |
| 268 |
// inner faces |
| 269 |
unsigned int n_boundary_faces = 0; |
| 270 |
for (; n_inner_faces + n_boundary_faces < faces.size(); |
| 271 |
++n_boundary_faces) |
| 272 |
if (faces[n_inner_faces + n_boundary_faces].cells_exterior[0] != |
| 273 |
numbers::invalid_unsigned_int) |
| 274 |
break; |
| 275 |
|
| 276 |
const unsigned int size_exterior_faces = |
| 277 |
hold_all_faces_to_owned_cells ? faces.size() : n_inner_faces; |
| 278 |
|
| 279 |
index_storage_variants[dof_access_face_exterior].resize( |
| 280 |
size_exterior_faces, IndexStorageVariants::full); |
| 281 |
dof_indices_contiguous[dof_access_face_exterior].resize( |
| 282 |
size_exterior_faces * length, numbers::invalid_unsigned_int); |
| 283 |
dof_indices_interleave_strides[dof_access_face_exterior].resize( |
| 284 |
faces.size() * length, numbers::invalid_unsigned_int); |
| 285 |
n_vectorization_lanes_filled[dof_access_face_exterior].resize( |
| 286 |
size_exterior_faces); |
| 287 |
|
| 288 |
for (unsigned int face = 0; face < faces.size(); ++face) |
| 289 |
{ |
| 290 |
auto face_computation = [&](const DoFAccessIndex face_index, |
| 291 |
const std::array<unsigned int, length> |
| 292 |
&cell_indices_face) { |
| 293 |
bool is_contiguous = false; |
| 294 |
bool is_interleaved = false; |
| 295 |
bool needs_full_storage = false; |
| 296 |
for (unsigned int v = 0; |
| 297 |
v < length && |
| 298 |
cell_indices_face[v] != numbers::invalid_unsigned_int; |
| 299 |
++v) |
| 300 |
{ |
| 301 |
n_vectorization_lanes_filled[face_index][face]++; |
| 302 |
if (index_storage_variants[dof_access_cell] |
| 303 |
[cell_indices_face[v] / length] >= |
| 304 |
IndexStorageVariants::interleaved_contiguous) |
| 305 |
is_interleaved = true; |
| 306 |
if (index_storage_variants[dof_access_cell] |
| 307 |
[cell_indices_face[v] / length] == |
| 308 |
IndexStorageVariants::contiguous) |
| 309 |
is_contiguous = true; |
| 310 |
if (index_storage_variants[dof_access_cell] |
| 311 |
[cell_indices_face[v] / length] >= |
| 312 |
IndexStorageVariants::contiguous) |
| 313 |
dof_indices_interleave_strides[face_index][face * length + |
| 314 |
v] = |
| 315 |
dof_indices_interleave_strides[dof_access_cell] |
| 316 |
[cell_indices_face[v]]; |
| 317 |
if (index_storage_variants[dof_access_cell] |
| 318 |
[cell_indices_face[v] / length] < |
| 319 |
IndexStorageVariants::contiguous) |
| 320 |
needs_full_storage = true; |
| 321 |
} |
| 322 |
Assert(!(is_interleaved && is_contiguous), |
| 323 |
ExcMessage("Unsupported index compression found")); |
| 324 |
|
| 325 |
if (is_interleaved || is_contiguous) |
| 326 |
for (unsigned int v = 0; |
| 327 |
v < n_vectorization_lanes_filled[face_index][face]; |
| 328 |
++v) |
| 329 |
dof_indices_contiguous[face_index][face * length + v] = |
| 330 |
dof_indices_contiguous[dof_access_cell][cell_indices_face[v]]; |
| 331 |
if (is_interleaved) |
| 332 |
{ |
| 333 |
bool is_also_contiguous = |
| 334 |
n_vectorization_lanes_filled[face_index][face] == length; |
| 335 |
for (unsigned int v = 0; |
| 336 |
v < n_vectorization_lanes_filled[face_index][face]; |
| 337 |
++v) |
| 338 |
if (dof_indices_contiguous[face_index][face * length + v] != |
| 339 |
dof_indices_contiguous[face_index][face * length] + v || |
| 340 |
dof_indices_interleave_strides[dof_access_cell] |
| 341 |
[cell_indices_face[v]] != |
| 342 |
length) |
| 343 |
is_also_contiguous = false; |
| 344 |
|
| 345 |
if (is_also_contiguous) |
| 346 |
{ |
| 347 |
index_storage_variants[face_index][face] = |
| 348 |
IndexStorageVariants::interleaved_contiguous; |
| 349 |
} |
| 350 |
else |
| 351 |
{ |
| 352 |
bool all_indices_same_offset = |
| 353 |
n_vectorization_lanes_filled[face_index][face] == length; |
| 354 |
for (unsigned int v = 0; |
| 355 |
v < n_vectorization_lanes_filled[face_index][face]; |
| 356 |
++v) |
| 357 |
if (dof_indices_interleave_strides |
| 358 |
[dof_access_cell][cell_indices_face[v]] != length) |
| 359 |
all_indices_same_offset = false; |
| 360 |
if (all_indices_same_offset) |
| 361 |
index_storage_variants[face_index][face] = |
| 362 |
IndexStorageVariants::interleaved_contiguous_strided; |
| 363 |
else |
| 364 |
index_storage_variants[face_index][face] = |
| 365 |
IndexStorageVariants:: |
| 366 |
interleaved_contiguous_mixed_strides; |
| 367 |
} |
| 368 |
} |
| 369 |
else if (is_contiguous && !needs_full_storage) |
| 370 |
index_storage_variants[face_index][face] = |
| 371 |
IndexStorageVariants::contiguous; |
| 372 |
else |
| 373 |
index_storage_variants[face_index][face] = |
| 374 |
IndexStorageVariants::full; |
| 375 |
}; |
| 376 |
|
| 377 |
face_computation(dof_access_face_interior, |
| 378 |
faces[face].cells_interior); |
| 379 |
if (face < n_inner_faces || |
| 380 |
(hold_all_faces_to_owned_cells && |
| 381 |
face >= (n_inner_faces + n_boundary_faces))) |
| 382 |
face_computation(dof_access_face_exterior, |
| 383 |
faces[face].cells_exterior); |
| 384 |
} |
| 385 |
} |
| 386 |
|
| 387 |
|
| 388 |
|
| 389 |
template <int length> |
| 390 |
void |
| 391 |
DoFInfo::compute_vector_zero_access_pattern( |
| 392 |
const TaskInfo &task_info, |
| 393 |
const std::vector<FaceToCellTopology<length>> &faces) |
| 394 |
{ |
| 395 |
// compute a list that tells us the first time a degree of freedom is |
| 396 |
// touched by a cell |
| 397 |
AssertDimension(length, vectorization_length); |
| 398 |
const unsigned int n_components = start_components.back(); |
| 399 |
const unsigned int n_dofs = vector_partitioner->locally_owned_size() + |
| 400 |
vector_partitioner->n_ghost_indices(); |
| 401 |
std::vector<unsigned int> touched_first_by( |
| 402 |
(n_dofs + chunk_size_zero_vector - 1) / chunk_size_zero_vector, |
| 403 |
numbers::invalid_unsigned_int); |
| 404 |
std::vector<unsigned int> touched_last_by( |
| 405 |
(n_dofs + chunk_size_zero_vector - 1) / chunk_size_zero_vector, |
| 406 |
numbers::invalid_unsigned_int); |
| 407 |
std::vector<unsigned int> cells_in_interval; |
| 408 |
for (unsigned int part = 0; |
| 409 |
part < task_info.partition_row_index.size() - 2; |
| 410 |
++part) |
| 411 |
for (unsigned int chunk = task_info.partition_row_index[part]; |
| 412 |
chunk < task_info.partition_row_index[part + 1]; |
| 413 |
++chunk) |
| 414 |
{ |
| 415 |
cells_in_interval.clear(); |
| 416 |
for (unsigned int cell = task_info.cell_partition_data[chunk]; |
| 417 |
cell < task_info.cell_partition_data[chunk + 1]; |
| 418 |
++cell) |
| 419 |
for (unsigned int v = 0; v < vectorization_length; ++v) |
| 420 |
cells_in_interval.push_back(cell * vectorization_length + v); |
| 421 |
if (faces.size() > 0) |
| 422 |
{ |
| 423 |
for (unsigned int face = task_info.face_partition_data[chunk]; |
| 424 |
face < task_info.face_partition_data[chunk + 1]; |
| 425 |
++face) |
| 426 |
for (unsigned int v = 0; v < vectorization_length; ++v) |
| 427 |
{ |
| 428 |
if (faces[face].cells_interior[v] != |
| 429 |
numbers::invalid_unsigned_int) |
| 430 |
cells_in_interval.push_back( |
| 431 |
faces[face].cells_interior[v]); |
| 432 |
if (faces[face].cells_exterior[v] != |
| 433 |
numbers::invalid_unsigned_int) |
| 434 |
cells_in_interval.push_back( |
| 435 |
faces[face].cells_exterior[v]); |
| 436 |
} |
| 437 |
for (unsigned int face = |
| 438 |
task_info.boundary_partition_data[chunk]; |
| 439 |
face < task_info.boundary_partition_data[chunk + 1]; |
| 440 |
++face) |
| 441 |
for (unsigned int v = 0; v < vectorization_length; ++v) |
| 442 |
if (faces[face].cells_interior[v] != |
| 443 |
numbers::invalid_unsigned_int) |
| 444 |
cells_in_interval.push_back( |
| 445 |
faces[face].cells_interior[v]); |
| 446 |
} |
| 447 |
std::sort(cells_in_interval.begin(), cells_in_interval.end()); |
| 448 |
cells_in_interval.erase(std::unique(cells_in_interval.begin(), |
| 449 |
cells_in_interval.end()), |
| 450 |
cells_in_interval.end()); |
| 451 |
|
| 452 |
for (const unsigned int cell : cells_in_interval) |
| 453 |
{ |
| 454 |
for (unsigned int it = row_starts[cell * n_components].first; |
| 455 |
it != row_starts[(cell + 1) * n_components].first; |
| 456 |
++it) |
| 457 |
{ |
| 458 |
const unsigned int myindex = |
| 459 |
dof_indices[it] / chunk_size_zero_vector; |
| 460 |
if (touched_first_by[myindex] == |
| 461 |
numbers::invalid_unsigned_int) |
| 462 |
touched_first_by[myindex] = chunk; |
| 463 |
touched_last_by[myindex] = chunk; |
| 464 |
} |
| 465 |
} |
| 466 |
} |
| 467 |
|
| 468 |
// ensure that all indices are touched at least during the last round |
| 469 |
for (auto &index : touched_first_by) |
| 470 |
if (index == numbers::invalid_unsigned_int) |
| 471 |
index = |
| 472 |
task_info |
| 473 |
.partition_row_index[task_info.partition_row_index.size() - 2] - |
| 474 |
1; |
| 475 |
|
| 476 |
// lambda to convert from a map, with keys associated to the buckets by |
| 477 |
// which we sliced the index space, length chunk_size_zero_vector, and |
| 478 |
// values equal to the slice index which are touched by the respective |
| 479 |
// partition, to a "vectors-of-vectors" like data structure. Rather than |
| 480 |
// using the vectors, we set up a sparsity-pattern like structure where |
| 481 |
// one index specifies the start index (range_list_index), and the other |
| 482 |
// the actual ranges (range_list). |
| 483 |
auto convert_map_to_range_list = |
| 484 |
[=](const unsigned int n_partitions, |
| 485 |
const std::map<unsigned int, std::vector<unsigned int>> &ranges_in, |
| 486 |
std::vector<unsigned int> &range_list_index, |
| 487 |
std::vector<std::pair<unsigned int, unsigned int>> &range_list, |
| 488 |
const unsigned int max_size) { |
| 489 |
range_list_index.resize(n_partitions + 1); |
| 490 |
range_list_index[0] = 0; |
| 491 |
range_list.clear(); |
| 492 |
for (unsigned int partition = 0; partition < n_partitions; |
| 493 |
++partition) |
| 494 |
{ |
| 495 |
auto it = ranges_in.find(partition); |
| 496 |
if (it != ranges_in.end()) |
| 497 |
{ |
| 498 |
for (unsigned int i = 0; i < it->second.size(); ++i) |
| 499 |
{ |
| 500 |
const unsigned int first_i = i; |
| 501 |
while (i + 1 < it->second.size() && |
| 502 |
it->second[i + 1] == it->second[i] + 1) |
| 503 |
++i; |
| 504 |
range_list.emplace_back( |
| 505 |
std::min(it->second[first_i] * chunk_size_zero_vector, |
| 506 |
max_size), |
| 507 |
std::min((it->second[i] + 1) * chunk_size_zero_vector, |
| 508 |
max_size)); |
| 509 |
} |
| 510 |
range_list_index[partition + 1] = range_list.size(); |
| 511 |
} |
| 512 |
else |
| 513 |
range_list_index[partition + 1] = range_list_index[partition]; |
| 514 |
} |
| 515 |
}; |
| 516 |
|
| 517 |
// first we determine the ranges to zero the vector |
| 518 |
std::map<unsigned int, std::vector<unsigned int>> chunk_must_zero_vector; |
| 519 |
for (unsigned int i = 0; i < touched_first_by.size(); ++i) |
| 520 |
chunk_must_zero_vector[touched_first_by[i]].push_back(i); |
| 521 |
const unsigned int n_partitions = |
| 522 |
task_info.partition_row_index[task_info.partition_row_index.size() - 2]; |
| 523 |
convert_map_to_range_list(n_partitions, |
| 524 |
chunk_must_zero_vector, |
| 525 |
vector_zero_range_list_index, |
| 526 |
vector_zero_range_list, |
| 527 |
vector_partitioner->locally_owned_size()); |
| 528 |
|
| 529 |
// the other two operations only work on the local range (without |
| 530 |
// ghosts), so we skip the latter parts of the vector now |
| 531 |
touched_first_by.resize((vector_partitioner->locally_owned_size() + |
| 532 |
chunk_size_zero_vector - 1) / |
| 533 |
chunk_size_zero_vector); |
| 534 |
|
| 535 |
// set the import indices in the vector partitioner to one index higher |
| 536 |
// to indicate that we want to process it first. This additional index |
| 537 |
// is reflected in the argument 'n_partitions+1' in the |
| 538 |
// convert_map_to_range_list function below. |
| 539 |
for (auto it : vector_partitioner->import_indices()) |
| 540 |
for (unsigned int i = it.first; i < it.second; ++i) |
| 541 |
touched_first_by[i / chunk_size_zero_vector] = n_partitions; |
| 542 |
std::map<unsigned int, std::vector<unsigned int>> chunk_must_do_pre; |
| 543 |
for (unsigned int i = 0; i < touched_first_by.size(); ++i) |
| 544 |
chunk_must_do_pre[touched_first_by[i]].push_back(i); |
| 545 |
convert_map_to_range_list(n_partitions + 1, |
| 546 |
chunk_must_do_pre, |
| 547 |
cell_loop_pre_list_index, |
| 548 |
cell_loop_pre_list, |
| 549 |
vector_partitioner->locally_owned_size()); |
| 550 |
|
| 551 |
touched_last_by.resize((vector_partitioner->locally_owned_size() + |
| 552 |
chunk_size_zero_vector - 1) / |
| 553 |
chunk_size_zero_vector); |
| 554 |
|
| 555 |
// set the indices which were not touched by the cell loop (i.e., |
| 556 |
// constrained indices) to the last valid partition index. Since |
| 557 |
// partition_row_index contains one extra slot for ghosted faces (which |
| 558 |
// are not part of the cell/face loops), we use the second to last entry |
| 559 |
// in the partition list. |
| 560 |
for (auto &index : touched_last_by) |
| 561 |
if (index == numbers::invalid_unsigned_int) |
| 562 |
index = |
| 563 |
task_info |
| 564 |
.partition_row_index[task_info.partition_row_index.size() - 2] - |
| 565 |
1; |
| 566 |
for (auto it : vector_partitioner->import_indices()) |
| 567 |
for (unsigned int i = it.first; i < it.second; ++i) |
| 568 |
touched_last_by[i / chunk_size_zero_vector] = n_partitions; |
| 569 |
std::map<unsigned int, std::vector<unsigned int>> chunk_must_do_post; |
| 570 |
for (unsigned int i = 0; i < touched_last_by.size(); ++i) |
| 571 |
chunk_must_do_post[touched_last_by[i]].push_back(i); |
| 572 |
convert_map_to_range_list(n_partitions + 1, |
| 573 |
chunk_must_do_post, |
| 574 |
cell_loop_post_list_index, |
| 575 |
cell_loop_post_list, |
| 576 |
vector_partitioner->locally_owned_size()); |
| 577 |
} |
| 578 |
|
| 579 |
|
| 580 |
|
| 581 |
namespace internal |
| 582 |
{ |
| 583 |
// rudimentary version of a vector that keeps entries always ordered |
| 584 |
class ordered_vector : public std::vector<types::global_dof_index> |
| 585 |
{ |
| 586 |
public: |
| 587 |
ordered_vector() |
| 588 |
{ |
| 589 |
reserve(2000); |
| 590 |
} |
| 591 |
|
| 592 |
void |
| 593 |
reserve(const std::size_t size) |
| 594 |
{ |
| 595 |
if (size > 0) |
| 596 |
this->std::vector<types::global_dof_index>::reserve(size); |
| 597 |
} |
| 598 |
|
| 599 |
|
| 600 |
// insert a given entry. dat is a pointer within this vector (the user |
| 601 |
// needs to make sure that it really stays there) |
| 602 |
void |
| 603 |
insert(const unsigned int entry, |
| 604 |
std::vector<types::global_dof_index>::iterator &dat) |
| 605 |
{ |
| 606 |
AssertIndexRange(static_cast<std::size_t>(dat - begin()), size() + 1); |
| 607 |
AssertIndexRange(static_cast<std::size_t>(end() - dat), size() + 1); |
| 608 |
AssertIndexRange(size(), capacity()); |
| 609 |
while (dat != end() && *dat < entry) |
| 610 |
++dat; |
| 611 |
|
| 612 |
if (dat == end()) |
| 613 |
{ |
| 614 |
push_back(entry); |
| 615 |
dat = end(); |
| 616 |
} |
| 617 |
else if (*dat > entry) |
| 618 |
{ |
| 619 |
dat = |
| 620 |
this->std::vector<types::global_dof_index>::insert(dat, entry); |
| 621 |
++dat; |
| 622 |
} |
| 623 |
else |
| 624 |
++dat; |
| 625 |
} |
| 626 |
}; |
| 627 |
} // namespace internal |
| 628 |
|
| 629 |
|
| 630 |
|
| 631 |
template <typename StreamType> |
| 632 |
void |
| 633 |
DoFInfo::print_memory_consumption(StreamType &out, |
| 634 |
const TaskInfo &task_info) const |
| 635 |
{ |
| 636 |
out << " Memory row starts indices: "; |
| 637 |
task_info.print_memory_statistics(out, |
| 638 |
(row_starts.capacity() * |
| 639 |
sizeof(*row_starts.begin()))); |
| 640 |
out << " Memory dof indices: "; |
| 641 |
task_info.print_memory_statistics( |
| 642 |
out, MemoryConsumption::memory_consumption(dof_indices)); |
| 643 |
out << " Memory constraint indicators: "; |
| 644 |
task_info.print_memory_statistics( |
| 645 |
out, MemoryConsumption::memory_consumption(constraint_indicator)); |
| 646 |
out << " Memory plain indices: "; |
| 647 |
task_info.print_memory_statistics( |
| 648 |
out, |
| 649 |
MemoryConsumption::memory_consumption(row_starts_plain_indices) + |
| 650 |
MemoryConsumption::memory_consumption(plain_dof_indices)); |
| 651 |
out << " Memory vector partitioner: "; |
| 652 |
task_info.print_memory_statistics( |
| 653 |
out, MemoryConsumption::memory_consumption(*vector_partitioner)); |
| 654 |
} |
| 655 |
|
| 656 |
|
| 657 |
|
| 658 |
template <typename Number> |
| 659 |
void |
| 660 |
DoFInfo::print(const std::vector<Number> &constraint_pool_data, |
| 661 |
const std::vector<unsigned int> &constraint_pool_row_index, |
| 662 |
std::ostream &out) const |
| 663 |
{ |
| 664 |
const unsigned int n_rows = row_starts.size() - 1; |
| 665 |
for (unsigned int row = 0; row < n_rows; ++row) |
| 666 |
{ |
| 667 |
if (row_starts[row].first == row_starts[row + 1].first) |
| 668 |
continue; |
| 669 |
out << "Entries row " << row << ": "; |
| 670 |
const unsigned int *glob_indices = |
| 671 |
&dof_indices[row_starts[row].first], |
| 672 |
*end_row = &dof_indices[row_starts[row + 1].first]; |
| 673 |
unsigned int index = 0; |
| 674 |
const std::pair<unsigned short, unsigned short> |
| 675 |
*con_it = &constraint_indicator[row_starts[row].second], |
| 676 |
*end_con = &constraint_indicator[row_starts[row + 1].second]; |
| 677 |
for (; con_it != end_con; ++con_it) |
| 678 |
{ |
| 679 |
for (unsigned int j = 0; j < con_it->first; ++j, ++index) |
| 680 |
{ |
| 681 |
Assert(glob_indices + index != end_row, ExcInternalError()); |
| 682 |
out << glob_indices[index] << ' '; |
| 683 |
} |
| 684 |
|
| 685 |
out << "[ "; |
| 686 |
for (unsigned int k = constraint_pool_row_index[con_it->second]; |
| 687 |
k < constraint_pool_row_index[con_it->second + 1]; |
| 688 |
k++, index++) |
| 689 |
{ |
| 690 |
Assert(glob_indices + index != end_row, ExcInternalError()); |
| 691 |
out << glob_indices[index] << '/' << constraint_pool_data[k] |
| 692 |
<< ' '; |
| 693 |
} |
| 694 |
out << "] "; |
| 695 |
} |
| 696 |
glob_indices += index; |
| 697 |
for (; glob_indices != end_row; ++glob_indices) |
| 698 |
out << *glob_indices << ' '; |
| 699 |
out << std::endl; |
| 700 |
} |
| 701 |
} |
| 702 |
|
| 703 |
|
| 704 |
} // namespace MatrixFreeFunctions |
| 705 |
} // end of namespace internal |
| 706 |
|
| 707 |
DEAL_II_NAMESPACE_CLOSE |
| 708 |
|
| 709 |
#endif |