| 1 |
// ------------------------------------------------------------------------ |
| 2 |
// |
| 3 |
// SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 |
// Copyright (C) 2016 - 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_mpi_templates_h |
| 16 |
#define dealii_mpi_templates_h |
| 17 |
|
| 18 |
#include <deal.II/base/config.h> |
| 19 |
|
| 20 |
#include <deal.II/base/exceptions.h> |
| 21 |
#include <deal.II/base/mpi.h> |
| 22 |
#include <deal.II/base/symmetric_tensor.h> |
| 23 |
#include <deal.II/base/tensor.h> |
| 24 |
|
| 25 |
#include <deal.II/lac/sparse_matrix.h> |
| 26 |
#include <deal.II/lac/vector.h> |
| 27 |
|
| 28 |
#include <set> |
| 29 |
#include <vector> |
| 30 |
|
| 31 |
DEAL_II_NAMESPACE_OPEN |
| 32 |
|
| 33 |
namespace Utilities |
| 34 |
{ |
| 35 |
namespace MPI |
| 36 |
{ |
| 37 |
namespace internal |
| 38 |
{ |
| 39 |
template <typename T> |
| 40 |
void |
| 41 |
all_reduce(const MPI_Op &mpi_op, |
| 42 |
const ArrayView<const T> &values, |
| 43 |
const MPI_Comm mpi_communicator, |
| 44 |
const ArrayView<T> &output) |
| 45 |
{ |
| 46 |
AssertDimension(values.size(), output.size()); |
| 47 |
#ifdef DEAL_II_WITH_MPI |
| 48 |
if (job_supports_mpi()) |
| 49 |
{ |
| 50 |
if constexpr (running_in_debug_mode()) |
| 51 |
{ |
| 52 |
{ |
| 53 |
const unsigned int rank = this_mpi_process(mpi_communicator); |
| 54 |
unsigned int size = values.size(); |
| 55 |
unsigned int size_min = 0; |
| 56 |
unsigned int size_max = 0; |
| 57 |
int ierr2 = 0; |
| 58 |
ierr2 = MPI_Reduce(&size, |
| 59 |
&size_min, |
| 60 |
1, |
| 61 |
MPI_UNSIGNED, |
| 62 |
MPI_MIN, |
| 63 |
0, |
| 64 |
mpi_communicator); |
| 65 |
AssertThrowMPI(ierr2); |
| 66 |
ierr2 = MPI_Reduce(&size, |
| 67 |
&size_max, |
| 68 |
1, |
| 69 |
MPI_UNSIGNED, |
| 70 |
MPI_MAX, |
| 71 |
0, |
| 72 |
mpi_communicator); |
| 73 |
AssertThrowMPI(ierr2); |
| 74 |
if (rank == 0) |
| 75 |
Assert( |
| 76 |
size_min == size_max, |
| 77 |
ExcMessage( |
| 78 |
"values has different size across MPI processes.")); |
| 79 |
} |
| 80 |
} |
| 81 |
const int ierr = |
| 82 |
MPI_Allreduce(values != output ? values.data() : MPI_IN_PLACE, |
| 83 |
static_cast<void *>(output.data()), |
| 84 |
static_cast<int>(values.size()), |
| 85 |
mpi_type_id_for_type<decltype(*values.data())>, |
| 86 |
mpi_op, |
| 87 |
mpi_communicator); |
| 88 |
AssertThrowMPI(ierr); |
| 89 |
} |
| 90 |
else |
| 91 |
#endif |
| 92 |
{ |
| 93 |
(void)mpi_op; |
| 94 |
(void)mpi_communicator; |
| 95 |
if (values != output) |
| 96 |
std::copy(values.begin(), values.end(), output.begin()); |
| 97 |
} |
| 98 |
} |
| 99 |
|
| 100 |
|
| 101 |
|
| 102 |
template <typename T> |
| 103 |
void |
| 104 |
all_reduce(const MPI_Op &mpi_op, |
| 105 |
const ArrayView<const std::complex<T>> &values, |
| 106 |
const MPI_Comm mpi_communicator, |
| 107 |
const ArrayView<std::complex<T>> &output) |
| 108 |
{ |
| 109 |
AssertDimension(values.size(), output.size()); |
| 110 |
#ifdef DEAL_II_WITH_MPI |
| 111 |
if (job_supports_mpi()) |
| 112 |
{ |
| 113 |
const int ierr = |
| 114 |
MPI_Allreduce(values != output ? |
| 115 |
// TODO This const_cast is only needed for older |
| 116 |
// (e.g., openMPI 1.6, released in 2012) |
| 117 |
// implementations of MPI-2. It is not needed as |
| 118 |
// of MPI-3 and we should remove it at some |
| 119 |
// point in the future. |
| 120 |
const_cast<void *>( |
| 121 |
static_cast<const void *>(values.data())) : |
| 122 |
MPI_IN_PLACE, |
| 123 |
static_cast<void *>(output.data()), |
| 124 |
static_cast<int>(values.size() * 2), |
| 125 |
mpi_type_id_for_type<T>, |
| 126 |
mpi_op, |
| 127 |
mpi_communicator); |
| 128 |
AssertThrowMPI(ierr); |
| 129 |
} |
| 130 |
else |
| 131 |
#endif |
| 132 |
{ |
| 133 |
(void)mpi_op; |
| 134 |
(void)mpi_communicator; |
| 135 |
if (values != output) |
| 136 |
std::copy(values.begin(), values.end(), output.begin()); |
| 137 |
} |
| 138 |
} |
| 139 |
} // namespace internal |
| 140 |
|
| 141 |
|
| 142 |
|
| 143 |
template <typename T> |
| 144 |
T |
| 145 |
sum(const T &t, const MPI_Comm mpi_communicator) |
| 146 |
{ |
| 147 |
if (mpi_communicator == MPI_COMM_SELF) |
| 148 |
return t; |
| 149 |
else |
| 150 |
{ |
| 151 |
T return_value{}; |
| 152 |
internal::all_reduce(MPI_SUM, |
| 153 |
ArrayView<const T>(&t, 1), |
| 154 |
mpi_communicator, |
| 155 |
ArrayView<T>(&return_value, 1)); |
| 156 |
return return_value; |
| 157 |
} |
| 158 |
} |
| 159 |
|
| 160 |
|
| 161 |
|
| 162 |
template <typename T, typename U> |
| 163 |
void |
| 164 |
sum(const T &values, const MPI_Comm mpi_communicator, U &sums) |
| 165 |
{ |
| 166 |
static_assert(std::is_same_v<std::decay_t<T>, std::decay_t<U>>, |
| 167 |
"Input and output arguments must have the same type!"); |
| 168 |
const auto array_view_values = make_array_view(values); |
| 169 |
using const_type = |
| 170 |
ArrayView<const typename decltype(array_view_values)::value_type>; |
| 171 |
sum(static_cast<const_type>(array_view_values), |
| 172 |
mpi_communicator, |
| 173 |
make_array_view(sums)); |
| 174 |
} |
| 175 |
|
| 176 |
|
| 177 |
|
| 178 |
template <typename T> |
| 179 |
void |
| 180 |
sum(const ArrayView<const T> &values, |
| 181 |
const MPI_Comm mpi_communicator, |
| 182 |
const ArrayView<T> &sums) |
| 183 |
{ |
| 184 |
if (mpi_communicator == MPI_COMM_SELF) |
| 185 |
for (unsigned int i = 0; i < values.size(); ++i) |
| 186 |
sums[i] = values[i]; |
| 187 |
else |
| 188 |
internal::all_reduce(MPI_SUM, values, mpi_communicator, sums); |
| 189 |
} |
| 190 |
|
| 191 |
|
| 192 |
|
| 193 |
template <int rank, int dim, typename Number> |
| 194 |
Tensor<rank, dim, Number> |
| 195 |
sum(const Tensor<rank, dim, Number> &t, const MPI_Comm mpi_communicator) |
| 196 |
{ |
| 197 |
// Copy the tensor into a C-style array with which we can then |
| 198 |
// call the other sum() function. |
| 199 |
Number array[Tensor<rank, dim, Number>::n_independent_components]; |
| 200 |
for (unsigned int i = 0; |
| 201 |
i < Tensor<rank, dim, Number>::n_independent_components; |
| 202 |
++i) |
| 203 |
array[i] = |
| 204 |
t[Tensor<rank, dim, Number>::unrolled_to_component_indices(i)]; |
| 205 |
|
| 206 |
sum(array, mpi_communicator, array); |
| 207 |
|
| 208 |
return Tensor<rank, dim, Number>(make_array_view(array)); |
| 209 |
} |
| 210 |
|
| 211 |
|
| 212 |
|
| 213 |
template <int rank, int dim, typename Number> |
| 214 |
SymmetricTensor<rank, dim, Number> |
| 215 |
sum(const SymmetricTensor<rank, dim, Number> &local, |
| 216 |
const MPI_Comm mpi_communicator) |
| 217 |
{ |
| 218 |
// Copy the tensor into a C-style array with which we can then |
| 219 |
// call the other sum() function. |
| 220 |
const unsigned int n_entries = |
| 221 |
SymmetricTensor<rank, dim, Number>::n_independent_components; |
| 222 |
Number |
| 223 |
entries[SymmetricTensor<rank, dim, Number>::n_independent_components]; |
| 224 |
|
| 225 |
for (unsigned int i = 0; i < n_entries; ++i) |
| 226 |
entries[i] = local[local.unrolled_to_component_indices(i)]; |
| 227 |
|
| 228 |
Number global_entries |
| 229 |
[SymmetricTensor<rank, dim, Number>::n_independent_components]; |
| 230 |
sum(entries, mpi_communicator, global_entries); |
| 231 |
|
| 232 |
SymmetricTensor<rank, dim, Number> global; |
| 233 |
for (unsigned int i = 0; i < n_entries; ++i) |
| 234 |
global[global.unrolled_to_component_indices(i)] = global_entries[i]; |
| 235 |
|
| 236 |
return global; |
| 237 |
} |
| 238 |
|
| 239 |
|
| 240 |
|
| 241 |
template <typename Number> |
| 242 |
void |
| 243 |
sum(const SparseMatrix<Number> &local, |
| 244 |
const MPI_Comm mpi_communicator, |
| 245 |
SparseMatrix<Number> &global) |
| 246 |
{ |
| 247 |
Assert( |
| 248 |
local.get_sparsity_pattern() == global.get_sparsity_pattern(), |
| 249 |
ExcMessage( |
| 250 |
"The sparsity pattern of the local and the global matrices should match.")); |
| 251 |
#ifdef DEAL_II_WITH_MPI |
| 252 |
// makes use of the fact that the matrix stores its data in a |
| 253 |
// contiguous array. |
| 254 |
sum(ArrayView<const Number>(local.val.get(), local.n_nonzero_elements()), |
| 255 |
mpi_communicator, |
| 256 |
ArrayView<Number>(global.val.get(), global.n_nonzero_elements())); |
| 257 |
#else |
| 258 |
(void)mpi_communicator; |
| 259 |
if (!PointerComparison::equal(&local, &global)) |
| 260 |
global = local; |
| 261 |
#endif |
| 262 |
} |
| 263 |
|
| 264 |
|
| 265 |
|
| 266 |
template <typename T> |
| 267 |
T |
| 268 |
max(const T &t, const MPI_Comm mpi_communicator) |
| 269 |
{ |
| 270 |
T return_value{}; |
| 271 |
internal::all_reduce(MPI_MAX, |
| 272 |
ArrayView<const T>(&t, 1), |
| 273 |
mpi_communicator, |
| 274 |
ArrayView<T>(&return_value, 1)); |
| 275 |
return return_value; |
| 276 |
} |
| 277 |
|
| 278 |
|
| 279 |
|
| 280 |
template <typename T, typename U> |
| 281 |
void |
| 282 |
max(const T &values, const MPI_Comm mpi_communicator, U &maxima) |
| 283 |
{ |
| 284 |
static_assert(std::is_same_v<std::decay_t<T>, std::decay_t<U>>, |
| 285 |
"Input and output arguments must have the same type!"); |
| 286 |
const auto array_view_values = make_array_view(values); |
| 287 |
using const_type = |
| 288 |
ArrayView<const typename decltype(array_view_values)::value_type>; |
| 289 |
max(static_cast<const_type>(array_view_values), |
| 290 |
mpi_communicator, |
| 291 |
make_array_view(maxima)); |
| 292 |
} |
| 293 |
|
| 294 |
|
| 295 |
|
| 296 |
template <typename T> |
| 297 |
void |
| 298 |
max(const ArrayView<const T> &values, |
| 299 |
const MPI_Comm mpi_communicator, |
| 300 |
const ArrayView<T> &maxima) |
| 301 |
{ |
| 302 |
internal::all_reduce(MPI_MAX, values, mpi_communicator, maxima); |
| 303 |
} |
| 304 |
|
| 305 |
|
| 306 |
|
| 307 |
template <typename T> |
| 308 |
T |
| 309 |
min(const T &t, const MPI_Comm mpi_communicator) |
| 310 |
{ |
| 311 |
T return_value{}; |
| 312 |
internal::all_reduce(MPI_MIN, |
| 313 |
ArrayView<const T>(&t, 1), |
| 314 |
mpi_communicator, |
| 315 |
ArrayView<T>(&return_value, 1)); |
| 316 |
return return_value; |
| 317 |
} |
| 318 |
|
| 319 |
|
| 320 |
|
| 321 |
template <typename T, typename U> |
| 322 |
void |
| 323 |
min(const T &values, const MPI_Comm mpi_communicator, U &minima) |
| 324 |
{ |
| 325 |
static_assert(std::is_same_v<std::decay_t<T>, std::decay_t<U>>, |
| 326 |
"Input and output arguments must have the same type!"); |
| 327 |
const auto array_view_values = make_array_view(values); |
| 328 |
using const_type = |
| 329 |
ArrayView<const typename decltype(array_view_values)::value_type>; |
| 330 |
min(static_cast<const_type>(array_view_values), |
| 331 |
mpi_communicator, |
| 332 |
make_array_view(minima)); |
| 333 |
} |
| 334 |
|
| 335 |
|
| 336 |
|
| 337 |
template <typename T> |
| 338 |
void |
| 339 |
min(const ArrayView<const T> &values, |
| 340 |
const MPI_Comm mpi_communicator, |
| 341 |
const ArrayView<T> &minima) |
| 342 |
{ |
| 343 |
internal::all_reduce(MPI_MIN, values, mpi_communicator, minima); |
| 344 |
} |
| 345 |
|
| 346 |
|
| 347 |
|
| 348 |
template <typename T> |
| 349 |
T |
| 350 |
logical_or(const T &t, const MPI_Comm mpi_communicator) |
| 351 |
{ |
| 352 |
static_assert(std::is_integral_v<T>, |
| 353 |
"The MPI_LOR operation only allows integral data types."); |
| 354 |
|
| 355 |
T return_value{}; |
| 356 |
internal::all_reduce(MPI_LOR, |
| 357 |
ArrayView<const T>(&t, 1), |
| 358 |
mpi_communicator, |
| 359 |
ArrayView<T>(&return_value, 1)); |
| 360 |
return return_value; |
| 361 |
} |
| 362 |
|
| 363 |
|
| 364 |
|
| 365 |
template <typename T, typename U> |
| 366 |
void |
| 367 |
logical_or(const T &values, const MPI_Comm mpi_communicator, U &results) |
| 368 |
{ |
| 369 |
static_assert(std::is_same_v<std::decay_t<T>, std::decay_t<U>>, |
| 370 |
"Input and output arguments must have the same type!"); |
| 371 |
|
| 372 |
static_assert(std::is_integral_v<typename T::value_type>, |
| 373 |
"The MPI_LOR operation only allows integral data types."); |
| 374 |
|
| 375 |
// Specializations of std containers for the data type bool do not |
| 376 |
// necessarily store its elements as a contiguous array. Thus we will use |
| 377 |
// the make_array_view() function with iterators here, which verifies |
| 378 |
// this. |
| 379 |
const auto array_view_values = |
| 380 |
make_array_view(values.cbegin(), values.cend()); |
| 381 |
const auto array_view_results = |
| 382 |
make_array_view(results.begin(), results.end()); |
| 383 |
|
| 384 |
using const_type = |
| 385 |
ArrayView<const typename decltype(array_view_values)::value_type>; |
| 386 |
logical_or(static_cast<const_type>(array_view_values), |
| 387 |
mpi_communicator, |
| 388 |
array_view_results); |
| 389 |
} |
| 390 |
|
| 391 |
|
| 392 |
|
| 393 |
template <typename T> |
| 394 |
void |
| 395 |
logical_or(const ArrayView<const T> &values, |
| 396 |
const MPI_Comm mpi_communicator, |
| 397 |
const ArrayView<T> &results) |
| 398 |
{ |
| 399 |
static_assert(std::is_integral_v<T>, |
| 400 |
"The MPI_LOR operation only allows integral data types."); |
| 401 |
|
| 402 |
internal::all_reduce(MPI_LOR, values, mpi_communicator, results); |
| 403 |
} |
| 404 |
|
| 405 |
|
| 406 |
|
| 407 |
template <typename T> |
| 408 |
T |
| 409 |
reduce(const T &vec, |
| 410 |
const MPI_Comm comm, |
| 411 |
const std::function<T(const T &, const T &)> &combiner, |
| 412 |
const unsigned int root_process) |
| 413 |
{ |
| 414 |
#ifdef DEAL_II_WITH_MPI |
| 415 |
if (n_mpi_processes(comm) > 1) |
| 416 |
{ |
| 417 |
// 1) perform custom reduction |
| 418 |
T result = vec; |
| 419 |
|
| 420 |
const unsigned int rank = this_mpi_process(comm); |
| 421 |
const unsigned int nproc = n_mpi_processes(comm); |
| 422 |
|
| 423 |
for (unsigned int stride = 1; stride < nproc; stride *= 2) |
| 424 |
{ |
| 425 |
unsigned int rank_recv = |
| 426 |
(2 * stride) * |
| 427 |
((rank + nproc - root_process) % nproc / (2 * stride)) + |
| 428 |
root_process; |
| 429 |
unsigned int rank_send = rank_recv + stride; |
| 430 |
|
| 431 |
if (rank_send >= nproc + root_process) // nothing to do |
| 432 |
continue; |
| 433 |
|
| 434 |
rank_recv = rank_recv % nproc; |
| 435 |
rank_send = rank_send % nproc; |
| 436 |
|
| 437 |
if (rank_recv == rank) // process receives data |
| 438 |
{ |
| 439 |
MPI_Status status; |
| 440 |
const auto ierr_1 = MPI_Probe(rank_send, |
| 441 |
internal::Tags::compute_union, |
| 442 |
comm, |
| 443 |
&status); |
| 444 |
AssertThrowMPI(ierr_1); |
| 445 |
|
| 446 |
int amount; |
| 447 |
const auto ierr_2 = MPI_Get_count(&status, MPI_CHAR, &amount); |
| 448 |
AssertThrowMPI(ierr_2); |
| 449 |
|
| 450 |
std::vector<char> temp(amount); |
| 451 |
|
| 452 |
const auto ierr_3 = MPI_Recv(temp.data(), |
| 453 |
amount, |
| 454 |
MPI_CHAR, |
| 455 |
status.MPI_SOURCE, |
| 456 |
status.MPI_TAG, |
| 457 |
comm, |
| 458 |
MPI_STATUS_IGNORE); |
| 459 |
AssertThrowMPI(ierr_3); |
| 460 |
|
| 461 |
result = combiner(result, Utilities::unpack<T>(temp, false)); |
| 462 |
} |
| 463 |
else if (rank_send == rank) // process sends data |
| 464 |
{ |
| 465 |
const std::vector<char> temp = Utilities::pack(result, false); |
| 466 |
|
| 467 |
const auto ierr = MPI_Send(temp.data(), |
| 468 |
temp.size(), |
| 469 |
MPI_CHAR, |
| 470 |
rank_recv, |
| 471 |
internal::Tags::compute_union, |
| 472 |
comm); |
| 473 |
AssertThrowMPI(ierr); |
| 474 |
} |
| 475 |
} |
| 476 |
|
| 477 |
if (rank == root_process) |
| 478 |
return result; |
| 479 |
else |
| 480 |
return {}; |
| 481 |
} |
| 482 |
#endif |
| 483 |
(void)comm; |
| 484 |
(void)combiner; |
| 485 |
(void)root_process; |
| 486 |
return vec; |
| 487 |
} |
| 488 |
|
| 489 |
|
| 490 |
|
| 491 |
template <typename T> |
| 492 |
T |
| 493 |
all_reduce(const T &vec, |
| 494 |
const MPI_Comm comm, |
| 495 |
const std::function<T(const T &, const T &)> &combiner) |
| 496 |
{ |
| 497 |
if (n_mpi_processes(comm) > 1) |
| 498 |
{ |
| 499 |
// 1) perform reduction |
| 500 |
const auto result = Utilities::MPI::reduce<T>(vec, comm, combiner); |
| 501 |
|
| 502 |
// 2) broadcast result |
| 503 |
return Utilities::MPI::broadcast(comm, result); |
| 504 |
} |
| 505 |
else |
| 506 |
return vec; |
| 507 |
} |
| 508 |
|
| 509 |
|
| 510 |
template <typename T> |
| 511 |
std::vector<T> |
| 512 |
compute_set_union(const std::vector<T> &vec, const MPI_Comm comm) |
| 513 |
{ |
| 514 |
return Utilities::MPI::all_reduce<std::vector<T>>( |
| 515 |
vec, comm, [](const auto &set_1, const auto &set_2) { |
| 516 |
// merge vectors, sort, and eliminate duplicates -> mimic std::set<T> |
| 517 |
auto result = set_1; |
| 518 |
result.insert(result.end(), set_2.begin(), set_2.end()); |
| 519 |
std::sort(result.begin(), result.end()); |
| 520 |
result.erase(std::unique(result.begin(), result.end()), result.end()); |
| 521 |
return result; |
| 522 |
}); |
| 523 |
} |
| 524 |
|
| 525 |
|
| 526 |
|
| 527 |
template <typename T> |
| 528 |
std::set<T> |
| 529 |
compute_set_union(const std::set<T> &set_in, const MPI_Comm comm) |
| 530 |
{ |
| 531 |
// convert vector to set |
| 532 |
std::vector<T> vector_in(set_in.begin(), set_in.end()); |
| 533 |
|
| 534 |
// perform operation to vector |
| 535 |
const std::vector<T> vector_out = compute_set_union(vector_in, comm); |
| 536 |
|
| 537 |
// convert vector to set |
| 538 |
return std::set<T>(vector_out.begin(), vector_out.end()); |
| 539 |
} |
| 540 |
|
| 541 |
} // end of namespace MPI |
| 542 |
} // end of namespace Utilities |
| 543 |
|
| 544 |
|
| 545 |
DEAL_II_NAMESPACE_CLOSE |
| 546 |
|
| 547 |
#endif |