| 1 |
// ------------------------------------------------------------------------ |
| 2 |
// |
| 3 |
// SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 |
// Copyright (C) 2017 - 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_partitioner_templates_h |
| 16 |
#define dealii_partitioner_templates_h |
| 17 |
|
| 18 |
#include <deal.II/base/config.h> |
| 19 |
|
| 20 |
#include <deal.II/base/mpi_tags.h> |
| 21 |
#include <deal.II/base/partitioner.h> |
| 22 |
|
| 23 |
#include <deal.II/lac/la_parallel_vector.h> |
| 24 |
|
| 25 |
#include <limits> |
| 26 |
#include <type_traits> |
| 27 |
|
| 28 |
|
| 29 |
DEAL_II_NAMESPACE_OPEN |
| 30 |
|
| 31 |
namespace Utilities |
| 32 |
{ |
| 33 |
namespace MPI |
| 34 |
{ |
| 35 |
#ifndef DOXYGEN |
| 36 |
|
| 37 |
# ifdef DEAL_II_WITH_MPI |
| 38 |
|
| 39 |
template <typename Number, typename MemorySpaceType> |
| 40 |
void |
| 41 |
Partitioner::export_to_ghosted_array_start( |
| 42 |
const unsigned int communication_channel, |
| 43 |
const ArrayView<const Number, MemorySpaceType> &locally_owned_array, |
| 44 |
const ArrayView<Number, MemorySpaceType> &temporary_storage, |
| 45 |
const ArrayView<Number, MemorySpaceType> &ghost_array, |
| 46 |
std::vector<MPI_Request> &requests) const |
| 47 |
{ |
| 48 |
AssertDimension(temporary_storage.size(), n_import_indices()); |
| 49 |
AssertIndexRange(communication_channel, 200); |
| 50 |
Assert(ghost_array.size() == n_ghost_indices() || |
| 51 |
ghost_array.size() == n_ghost_indices_in_larger_set, |
| 52 |
ExcGhostIndexArrayHasWrongSize(ghost_array.size(), |
| 53 |
n_ghost_indices(), |
| 54 |
n_ghost_indices_in_larger_set)); |
| 55 |
|
| 56 |
const unsigned int n_import_targets = import_targets_data.size(); |
| 57 |
const unsigned int n_ghost_targets = ghost_targets_data.size(); |
| 58 |
|
| 59 |
if (n_import_targets > 0) |
| 60 |
AssertDimension(locally_owned_array.size(), locally_owned_size()); |
| 61 |
|
| 62 |
Assert(requests.empty(), |
| 63 |
ExcMessage("Another operation seems to still be running. " |
| 64 |
"Call update_ghost_values_finish() first.")); |
| 65 |
|
| 66 |
const unsigned int mpi_tag = |
| 67 |
Utilities::MPI::internal::Tags::partitioner_export_start + |
| 68 |
communication_channel; |
| 69 |
Assert(mpi_tag <= Utilities::MPI::internal::Tags::partitioner_export_end, |
| 70 |
ExcInternalError()); |
| 71 |
|
| 72 |
// Need to send and receive the data. Use non-blocking communication, |
| 73 |
// where it is usually less overhead to first initiate the receive and |
| 74 |
// then actually send the data |
| 75 |
requests.resize(n_import_targets + n_ghost_targets); |
| 76 |
|
| 77 |
// as a ghost array pointer, put the data at the end of the given ghost |
| 78 |
// array in case we want to fill only a subset of the ghosts so that we |
| 79 |
// can move data to the right position in a forward loop in the _finish |
| 80 |
// function. |
| 81 |
AssertIndexRange(n_ghost_indices(), n_ghost_indices_in_larger_set + 1); |
| 82 |
const bool use_larger_set = |
| 83 |
(n_ghost_indices_in_larger_set > n_ghost_indices() && |
| 84 |
ghost_array.size() == n_ghost_indices_in_larger_set); |
| 85 |
Number *ghost_array_ptr = |
| 86 |
use_larger_set ? ghost_array.data() + n_ghost_indices_in_larger_set - |
| 87 |
n_ghost_indices() : |
| 88 |
ghost_array.data(); |
| 89 |
|
| 90 |
for (unsigned int i = 0; i < n_ghost_targets; ++i) |
| 91 |
{ |
| 92 |
// allow writing into ghost indices even though we are in a |
| 93 |
// const function |
| 94 |
const int ierr = |
| 95 |
MPI_Irecv(ghost_array_ptr, |
| 96 |
ghost_targets_data[i].second * sizeof(Number), |
| 97 |
MPI_BYTE, |
| 98 |
ghost_targets_data[i].first, |
| 99 |
mpi_tag, |
| 100 |
communicator, |
| 101 |
&requests[i]); |
| 102 |
AssertThrowMPI(ierr); |
| 103 |
ghost_array_ptr += ghost_targets_data[i].second; |
| 104 |
} |
| 105 |
|
| 106 |
Number *temp_array_ptr = temporary_storage.data(); |
| 107 |
# if defined(DEAL_II_MPI_WITH_DEVICE_SUPPORT) |
| 108 |
// When using device-aware MPI, the set of local indices that are ghosts |
| 109 |
// indices on other processors is expanded in arrays. This is for |
| 110 |
// performance reasons as this can significantly decrease the number of |
| 111 |
// kernel launched. The indices are expanded the first time the function |
| 112 |
// is called. |
| 113 |
if ((std::is_same_v<MemorySpaceType, MemorySpace::Default>)&&( |
| 114 |
import_indices_plain_dev.empty())) |
| 115 |
initialize_import_indices_plain_dev(); |
| 116 |
# endif |
| 117 |
|
| 118 |
for (unsigned int i = 0; i < n_import_targets; ++i) |
| 119 |
{ |
| 120 |
# if defined(DEAL_II_MPI_WITH_DEVICE_SUPPORT) |
| 121 |
if constexpr (std::is_same_v<MemorySpaceType, MemorySpace::Default>) |
| 122 |
{ |
| 123 |
const auto chunk_size = import_indices_plain_dev[i].size(); |
| 124 |
using IndexType = decltype(chunk_size); |
| 125 |
|
| 126 |
auto import_indices = import_indices_plain_dev[i]; |
| 127 |
auto locally_owned_array_data = locally_owned_array.data(); |
| 128 |
MemorySpace::Default::kokkos_space::execution_space exec; |
| 129 |
Kokkos::parallel_for( |
| 130 |
"dealii::fill temp_array_ptr", |
| 131 |
Kokkos::RangePolicy< |
| 132 |
MemorySpace::Default::kokkos_space::execution_space>( |
| 133 |
exec, 0, chunk_size), |
| 134 |
KOKKOS_LAMBDA(IndexType idx) { |
| 135 |
temp_array_ptr[idx] = |
| 136 |
locally_owned_array_data[import_indices[idx]]; |
| 137 |
}); |
| 138 |
exec.fence(); |
| 139 |
} |
| 140 |
else |
| 141 |
# endif |
| 142 |
{ |
| 143 |
// copy the data to be sent to the import_data field |
| 144 |
std::vector<std::pair<unsigned int, unsigned int>>::const_iterator |
| 145 |
my_imports = import_indices_data.begin() + |
| 146 |
import_indices_chunks_by_rank_data[i], |
| 147 |
end_my_imports = import_indices_data.begin() + |
| 148 |
import_indices_chunks_by_rank_data[i + 1]; |
| 149 |
unsigned int index = 0; |
| 150 |
for (; my_imports != end_my_imports; ++my_imports) |
| 151 |
{ |
| 152 |
const unsigned int chunk_size = |
| 153 |
my_imports->second - my_imports->first; |
| 154 |
{ |
| 155 |
std::memcpy(temp_array_ptr + index, |
| 156 |
locally_owned_array.data() + my_imports->first, |
| 157 |
chunk_size * sizeof(Number)); |
| 158 |
} |
| 159 |
index += chunk_size; |
| 160 |
} |
| 161 |
|
| 162 |
AssertDimension(index, import_targets_data[i].second); |
| 163 |
} |
| 164 |
|
| 165 |
// start the send operations |
| 166 |
const int ierr = |
| 167 |
MPI_Isend(temp_array_ptr, |
| 168 |
import_targets_data[i].second * sizeof(Number), |
| 169 |
MPI_BYTE, |
| 170 |
import_targets_data[i].first, |
| 171 |
mpi_tag, |
| 172 |
communicator, |
| 173 |
&requests[n_ghost_targets + i]); |
| 174 |
AssertThrowMPI(ierr); |
| 175 |
temp_array_ptr += import_targets_data[i].second; |
| 176 |
} |
| 177 |
} |
| 178 |
|
| 179 |
|
| 180 |
|
| 181 |
template <typename Number, typename MemorySpaceType> |
| 182 |
void |
| 183 |
Partitioner::export_to_ghosted_array_finish( |
| 184 |
const ArrayView<Number, MemorySpaceType> &ghost_array, |
| 185 |
std::vector<MPI_Request> &requests) const |
| 186 |
{ |
| 187 |
Assert(ghost_array.size() == n_ghost_indices() || |
| 188 |
ghost_array.size() == n_ghost_indices_in_larger_set, |
| 189 |
ExcGhostIndexArrayHasWrongSize(ghost_array.size(), |
| 190 |
n_ghost_indices(), |
| 191 |
n_ghost_indices_in_larger_set)); |
| 192 |
|
| 193 |
// wait for both sends and receives to complete, even though only |
| 194 |
// receives are really necessary. this gives (much) better performance |
| 195 |
AssertDimension(ghost_targets().size() + import_targets().size(), |
| 196 |
requests.size()); |
| 197 |
if (requests.size() > 0) |
| 198 |
{ |
| 199 |
const int ierr = |
| 200 |
MPI_Waitall(requests.size(), requests.data(), MPI_STATUSES_IGNORE); |
| 201 |
AssertThrowMPI(ierr); |
| 202 |
} |
| 203 |
requests.resize(0); |
| 204 |
|
| 205 |
// in case we only sent a subset of indices, we now need to move the data |
| 206 |
// to the correct positions and delete the old content |
| 207 |
if (n_ghost_indices_in_larger_set > n_ghost_indices() && |
| 208 |
ghost_array.size() == n_ghost_indices_in_larger_set) |
| 209 |
{ |
| 210 |
unsigned int offset = |
| 211 |
n_ghost_indices_in_larger_set - n_ghost_indices(); |
| 212 |
// must copy ghost data into extended ghost array |
| 213 |
for (const auto &ghost_range : ghost_indices_subset_data) |
| 214 |
{ |
| 215 |
if (offset > ghost_range.first) |
| 216 |
{ |
| 217 |
const unsigned int chunk_size = |
| 218 |
ghost_range.second - ghost_range.first; |
| 219 |
if constexpr (std::is_same_v<MemorySpaceType, |
| 220 |
MemorySpace::Host>) |
| 221 |
{ |
| 222 |
// If source and destination are overlapping, we must be |
| 223 |
// careful to use an appropriate copy function. |
| 224 |
if (ghost_range.first > offset) |
| 225 |
std::copy_backward(ghost_array.data() + offset, |
| 226 |
ghost_array.data() + offset + |
| 227 |
chunk_size, |
| 228 |
ghost_array.data() + |
| 229 |
ghost_range.first); |
| 230 |
else |
| 231 |
std::copy(ghost_array.data() + offset, |
| 232 |
ghost_array.data() + offset + chunk_size, |
| 233 |
ghost_array.data() + ghost_range.first); |
| 234 |
std::fill(ghost_array.data() + |
| 235 |
std::max(ghost_range.second, offset), |
| 236 |
ghost_array.data() + offset + chunk_size, |
| 237 |
Number{}); |
| 238 |
} |
| 239 |
else |
| 240 |
{ |
| 241 |
Kokkos::View<const Number *, |
| 242 |
MemorySpace::Default::kokkos_space> |
| 243 |
ghost_src_view(ghost_array.data() + offset, chunk_size); |
| 244 |
Kokkos::View<Number *, MemorySpace::Default::kokkos_space> |
| 245 |
ghost_dst_view(ghost_array.data() + ghost_range.first, |
| 246 |
chunk_size); |
| 247 |
|
| 248 |
// If source and destination are overlapping, we can't |
| 249 |
// just call deep_copy but must copy the data to a buffer |
| 250 |
// first. |
| 251 |
if ((offset < ghost_range.first && |
| 252 |
ghost_range.first < offset + chunk_size) || |
| 253 |
(ghost_range.first < offset && |
| 254 |
offset < ghost_range.first + chunk_size)) |
| 255 |
{ |
| 256 |
Kokkos::View<Number *, |
| 257 |
MemorySpace::Default::kokkos_space> |
| 258 |
copy(Kokkos::view_alloc( |
| 259 |
"copy", Kokkos::WithoutInitializing), |
| 260 |
chunk_size); |
| 261 |
Kokkos::deep_copy(copy, ghost_src_view); |
| 262 |
Kokkos::deep_copy(ghost_dst_view, copy); |
| 263 |
} |
| 264 |
else |
| 265 |
{ |
| 266 |
Kokkos::deep_copy(ghost_dst_view, ghost_src_view); |
| 267 |
} |
| 268 |
Kokkos::deep_copy( |
| 269 |
Kokkos::View<Number *, |
| 270 |
MemorySpace::Default::kokkos_space>( |
| 271 |
ghost_array.data() + |
| 272 |
std::max(ghost_range.second, offset), |
| 273 |
(offset + chunk_size - |
| 274 |
std::max(ghost_range.second, offset))), |
| 275 |
0); |
| 276 |
} |
| 277 |
offset += chunk_size; |
| 278 |
} |
| 279 |
else |
| 280 |
{ |
| 281 |
AssertDimension(offset, ghost_range.first); |
| 282 |
break; |
| 283 |
} |
| 284 |
} |
| 285 |
} |
| 286 |
} |
| 287 |
|
| 288 |
|
| 289 |
|
| 290 |
template <typename Number, typename MemorySpaceType> |
| 291 |
void |
| 292 |
Partitioner::import_from_ghosted_array_start( |
| 293 |
const VectorOperation::values vector_operation, |
| 294 |
const unsigned int communication_channel, |
| 295 |
const ArrayView<Number, MemorySpaceType> &ghost_array, |
| 296 |
const ArrayView<Number, MemorySpaceType> &temporary_storage, |
| 297 |
std::vector<MPI_Request> &requests) const |
| 298 |
{ |
| 299 |
AssertDimension(temporary_storage.size(), n_import_indices()); |
| 300 |
AssertIndexRange(communication_channel, 200); |
| 301 |
Assert(ghost_array.size() == n_ghost_indices() || |
| 302 |
ghost_array.size() == n_ghost_indices_in_larger_set, |
| 303 |
ExcGhostIndexArrayHasWrongSize(ghost_array.size(), |
| 304 |
n_ghost_indices(), |
| 305 |
n_ghost_indices_in_larger_set)); |
| 306 |
|
| 307 |
(void)vector_operation; |
| 308 |
|
| 309 |
// nothing to do for insert (only need to zero ghost entries in |
| 310 |
// compress_finish()). in debug mode we want to check consistency of the |
| 311 |
// inserted data, therefore the communication is still initialized. |
| 312 |
// Having different code in debug and optimized mode is somewhat |
| 313 |
// dangerous, but it really saves communication so do it anyway |
| 314 |
if constexpr (running_in_debug_mode() == false) |
| 315 |
if (vector_operation == VectorOperation::insert) |
| 316 |
return; |
| 317 |
|
| 318 |
// nothing to do when we neither have import |
| 319 |
// nor ghost indices. |
| 320 |
if (n_ghost_indices() == 0 && n_import_indices() == 0) |
| 321 |
return; |
| 322 |
|
| 323 |
const unsigned int n_import_targets = import_targets_data.size(); |
| 324 |
const unsigned int n_ghost_targets = ghost_targets_data.size(); |
| 325 |
|
| 326 |
Assert(requests.empty(), |
| 327 |
ExcMessage("Another compress operation seems to still be running. " |
| 328 |
"Call compress_finish() first.")); |
| 329 |
|
| 330 |
// Need to send and receive the data. Use non-blocking communication, |
| 331 |
// where it is generally less overhead to first initiate the receive and |
| 332 |
// then actually send the data |
| 333 |
|
| 334 |
const unsigned int mpi_tag = |
| 335 |
Utilities::MPI::internal::Tags::partitioner_import_start + |
| 336 |
communication_channel; |
| 337 |
Assert(mpi_tag <= Utilities::MPI::internal::Tags::partitioner_import_end, |
| 338 |
ExcInternalError()); |
| 339 |
requests.resize(n_import_targets + n_ghost_targets); |
| 340 |
|
| 341 |
// initiate the receive operations |
| 342 |
Number *temp_array_ptr = temporary_storage.data(); |
| 343 |
for (unsigned int i = 0; i < n_import_targets; ++i) |
| 344 |
{ |
| 345 |
AssertThrow( |
| 346 |
static_cast<std::size_t>(import_targets_data[i].second) * |
| 347 |
sizeof(Number) < |
| 348 |
static_cast<std::size_t>(std::numeric_limits<int>::max()), |
| 349 |
ExcMessage("Index overflow: Maximum message size in MPI is 2GB. " |
| 350 |
"The number of ghost entries times the size of 'Number' " |
| 351 |
"exceeds this value. This is not supported.")); |
| 352 |
const int ierr = |
| 353 |
MPI_Irecv(temp_array_ptr, |
| 354 |
import_targets_data[i].second * sizeof(Number), |
| 355 |
MPI_BYTE, |
| 356 |
import_targets_data[i].first, |
| 357 |
mpi_tag, |
| 358 |
communicator, |
| 359 |
&requests[i]); |
| 360 |
AssertThrowMPI(ierr); |
| 361 |
temp_array_ptr += import_targets_data[i].second; |
| 362 |
} |
| 363 |
|
| 364 |
// initiate the send operations |
| 365 |
|
| 366 |
// in case we want to import only from a subset of the ghosts we want to |
| 367 |
// move the data to send to the front of the array |
| 368 |
AssertIndexRange(n_ghost_indices(), n_ghost_indices_in_larger_set + 1); |
| 369 |
Number *ghost_array_ptr = ghost_array.data(); |
| 370 |
for (unsigned int i = 0; i < n_ghost_targets; ++i) |
| 371 |
{ |
| 372 |
// in case we only sent a subset of indices, we now need to move the |
| 373 |
// data to the correct positions and delete the old content |
| 374 |
if (n_ghost_indices_in_larger_set > n_ghost_indices() && |
| 375 |
ghost_array.size() == n_ghost_indices_in_larger_set) |
| 376 |
{ |
| 377 |
std::vector<std::pair<unsigned int, unsigned int>>::const_iterator |
| 378 |
my_ghosts = ghost_indices_subset_data.begin() + |
| 379 |
ghost_indices_subset_chunks_by_rank_data[i], |
| 380 |
end_my_ghosts = ghost_indices_subset_data.begin() + |
| 381 |
ghost_indices_subset_chunks_by_rank_data[i + 1]; |
| 382 |
unsigned int offset = 0; |
| 383 |
for (; my_ghosts != end_my_ghosts; ++my_ghosts) |
| 384 |
{ |
| 385 |
const unsigned int chunk_size = |
| 386 |
my_ghosts->second - my_ghosts->first; |
| 387 |
if (ghost_array_ptr + offset != |
| 388 |
ghost_array.data() + my_ghosts->first) |
| 389 |
{ |
| 390 |
if constexpr (std::is_same_v<MemorySpaceType, |
| 391 |
MemorySpace::Host>) |
| 392 |
{ |
| 393 |
if (offset > my_ghosts->first) |
| 394 |
std::copy_backward(ghost_array.data() + |
| 395 |
my_ghosts->first, |
| 396 |
ghost_array_ptr + |
| 397 |
my_ghosts->second, |
| 398 |
ghost_array.data() + offset); |
| 399 |
else |
| 400 |
std::copy(ghost_array.data() + my_ghosts->first, |
| 401 |
ghost_array.data() + my_ghosts->second, |
| 402 |
ghost_array_ptr + offset); |
| 403 |
std::fill( |
| 404 |
std::max(ghost_array.data() + my_ghosts->first, |
| 405 |
ghost_array_ptr + offset + chunk_size), |
| 406 |
ghost_array.data() + my_ghosts->second, |
| 407 |
Number{}); |
| 408 |
} |
| 409 |
else |
| 410 |
{ |
| 411 |
Kokkos::View<Number *, |
| 412 |
MemorySpace::Default::kokkos_space> |
| 413 |
copy("copy", chunk_size); |
| 414 |
Kokkos::deep_copy( |
| 415 |
copy, |
| 416 |
Kokkos::View<Number *, |
| 417 |
MemorySpace::Default::kokkos_space>( |
| 418 |
ghost_array.data() + my_ghosts->first, |
| 419 |
chunk_size)); |
| 420 |
Kokkos::deep_copy( |
| 421 |
Kokkos::View<Number *, |
| 422 |
MemorySpace::Default::kokkos_space>( |
| 423 |
ghost_array_ptr + offset, chunk_size), |
| 424 |
copy); |
| 425 |
Kokkos::deep_copy( |
| 426 |
Kokkos::View<Number *, |
| 427 |
MemorySpace::Default::kokkos_space>( |
| 428 |
std::max(ghost_array.data() + my_ghosts->first, |
| 429 |
ghost_array_ptr + offset + chunk_size), |
| 430 |
(ghost_array.data() + my_ghosts->second - |
| 431 |
std::max(ghost_array.data() + my_ghosts->first, |
| 432 |
ghost_array_ptr + offset + |
| 433 |
chunk_size))), |
| 434 |
0); |
| 435 |
} |
| 436 |
} |
| 437 |
offset += chunk_size; |
| 438 |
} |
| 439 |
AssertDimension(offset, ghost_targets_data[i].second); |
| 440 |
} |
| 441 |
|
| 442 |
AssertThrow( |
| 443 |
static_cast<std::size_t>(ghost_targets_data[i].second) * |
| 444 |
sizeof(Number) < |
| 445 |
static_cast<std::size_t>(std::numeric_limits<int>::max()), |
| 446 |
ExcMessage("Index overflow: Maximum message size in MPI is 2GB. " |
| 447 |
"The number of ghost entries times the size of 'Number' " |
| 448 |
"exceeds this value. This is not supported.")); |
| 449 |
if (std::is_same_v<MemorySpaceType, MemorySpace::Default>) |
| 450 |
Kokkos::fence(); |
| 451 |
const int ierr = |
| 452 |
MPI_Isend(ghost_array_ptr, |
| 453 |
ghost_targets_data[i].second * sizeof(Number), |
| 454 |
MPI_BYTE, |
| 455 |
ghost_targets_data[i].first, |
| 456 |
mpi_tag, |
| 457 |
communicator, |
| 458 |
&requests[n_import_targets + i]); |
| 459 |
AssertThrowMPI(ierr); |
| 460 |
|
| 461 |
ghost_array_ptr += ghost_targets_data[i].second; |
| 462 |
} |
| 463 |
} |
| 464 |
|
| 465 |
|
| 466 |
|
| 467 |
namespace internal |
| 468 |
{ |
| 469 |
// In the import_from_ghosted_array_finish we need to invoke abs() also |
| 470 |
// on unsigned data types, which is ill-formed on newer C++ |
| 471 |
// standards. To avoid this, we use std::abs on default types but |
| 472 |
// simply return the number on unsigned types |
| 473 |
template <typename Number> |
| 474 |
std::enable_if_t<!std::is_unsigned_v<Number>, |
| 475 |
typename numbers::NumberTraits<Number>::real_type> |
| 476 |
get_abs(const Number a) |
| 477 |
{ |
| 478 |
return std::abs(a); |
| 479 |
} |
| 480 |
|
| 481 |
template <typename Number> |
| 482 |
std::enable_if_t<std::is_unsigned_v<Number>, Number> |
| 483 |
get_abs(const Number a) |
| 484 |
{ |
| 485 |
return a; |
| 486 |
} |
| 487 |
|
| 488 |
// In the import_from_ghosted_array_finish we might need to calculate the |
| 489 |
// maximal and minimal value for the given number type, which is not |
| 490 |
// straight forward for complex numbers. Therefore, comparison of complex |
| 491 |
// numbers is prohibited and throws an exception. |
| 492 |
template <typename Number> |
| 493 |
DEAL_II_HOST_DEVICE Number |
| 494 |
get_min(const Number a, const Number b) |
| 495 |
{ |
| 496 |
return std::min(a, b); |
| 497 |
} |
| 498 |
|
| 499 |
template <typename Number> |
| 500 |
std::complex<Number> |
| 501 |
get_min(const std::complex<Number> a, const std::complex<Number>) |
| 502 |
{ |
| 503 |
AssertThrow(false, |
| 504 |
ExcMessage("VectorOperation::min not " |
| 505 |
"implemented for complex numbers")); |
| 506 |
return a; |
| 507 |
} |
| 508 |
|
| 509 |
template <typename Number> |
| 510 |
DEAL_II_HOST_DEVICE Number |
| 511 |
get_max(const Number a, const Number b) |
| 512 |
{ |
| 513 |
return std::max(a, b); |
| 514 |
} |
| 515 |
|
| 516 |
template <typename Number> |
| 517 |
std::complex<Number> |
| 518 |
get_max(const std::complex<Number> a, const std::complex<Number>) |
| 519 |
{ |
| 520 |
AssertThrow(false, |
| 521 |
ExcMessage("VectorOperation::max not " |
| 522 |
"implemented for complex numbers")); |
| 523 |
return a; |
| 524 |
} |
| 525 |
} // namespace internal |
| 526 |
|
| 527 |
|
| 528 |
|
| 529 |
template <typename Number, typename MemorySpaceType> |
| 530 |
void |
| 531 |
Partitioner::import_from_ghosted_array_finish( |
| 532 |
const VectorOperation::values vector_operation, |
| 533 |
const ArrayView<const Number, MemorySpaceType> &temporary_storage, |
| 534 |
const ArrayView<Number, MemorySpaceType> &locally_owned_array, |
| 535 |
const ArrayView<Number, MemorySpaceType> &ghost_array, |
| 536 |
std::vector<MPI_Request> &requests) const |
| 537 |
{ |
| 538 |
AssertDimension(temporary_storage.size(), n_import_indices()); |
| 539 |
Assert(ghost_array.size() == n_ghost_indices() || |
| 540 |
ghost_array.size() == n_ghost_indices_in_larger_set, |
| 541 |
ExcGhostIndexArrayHasWrongSize(ghost_array.size(), |
| 542 |
n_ghost_indices(), |
| 543 |
n_ghost_indices_in_larger_set)); |
| 544 |
|
| 545 |
// in optimized mode, no communication was started, so leave the |
| 546 |
// function directly (and only clear ghosts) |
| 547 |
if constexpr (running_in_debug_mode() == false) |
| 548 |
if (vector_operation == VectorOperation::insert) |
| 549 |
{ |
| 550 |
Assert(requests.empty(), |
| 551 |
ExcInternalError( |
| 552 |
"Did not expect a non-empty communication " |
| 553 |
"request when inserting. Check that the same " |
| 554 |
"vector_operation argument was passed to " |
| 555 |
"import_from_ghosted_array_start as is passed " |
| 556 |
"to import_from_ghosted_array_finish.")); |
| 557 |
|
| 558 |
Kokkos::deep_copy( |
| 559 |
Kokkos::View<Number *, typename MemorySpaceType::kokkos_space>( |
| 560 |
ghost_array.data(), ghost_array.size()), |
| 561 |
0); |
| 562 |
return; |
| 563 |
} |
| 564 |
|
| 565 |
// nothing to do when we neither have import nor ghost indices. |
| 566 |
if (n_ghost_indices() == 0 && n_import_indices() == 0) |
| 567 |
return; |
| 568 |
|
| 569 |
const unsigned int n_import_targets = import_targets_data.size(); |
| 570 |
const unsigned int n_ghost_targets = ghost_targets_data.size(); |
| 571 |
|
| 572 |
# if defined(DEAL_II_MPI_WITH_DEVICE_SUPPORT) |
| 573 |
// When using device-aware MPI, the set of local indices that are ghosts |
| 574 |
// indices on other processors is expanded in arrays. This is for |
| 575 |
// performance reasons as this can significantly decrease the number of |
| 576 |
// kernel launched. The indices are expanded the first time the function |
| 577 |
// is called. |
| 578 |
if ((std::is_same_v<MemorySpaceType, MemorySpace::Default>)&&( |
| 579 |
import_indices_plain_dev.empty())) |
| 580 |
initialize_import_indices_plain_dev(); |
| 581 |
# endif |
| 582 |
|
| 583 |
if (vector_operation != VectorOperation::insert) |
| 584 |
AssertDimension(n_ghost_targets + n_import_targets, requests.size()); |
| 585 |
// first wait for the receive to complete |
| 586 |
if (requests.size() > 0 && n_import_targets > 0) |
| 587 |
{ |
| 588 |
AssertDimension(locally_owned_array.size(), locally_owned_size()); |
| 589 |
const int ierr = |
| 590 |
MPI_Waitall(n_import_targets, requests.data(), MPI_STATUSES_IGNORE); |
| 591 |
AssertThrowMPI(ierr); |
| 592 |
|
| 593 |
const Number *read_position = temporary_storage.data(); |
| 594 |
# if defined(DEAL_II_MPI_WITH_DEVICE_SUPPORT) |
| 595 |
if constexpr (std::is_same_v<MemorySpaceType, MemorySpace::Default>) |
| 596 |
{ |
| 597 |
if (vector_operation == VectorOperation::add) |
| 598 |
{ |
| 599 |
for (const auto &import_indices_plain : |
| 600 |
import_indices_plain_dev) |
| 601 |
{ |
| 602 |
const auto chunk_size = import_indices_plain.size(); |
| 603 |
|
| 604 |
using IndexType = decltype(chunk_size); |
| 605 |
MemorySpace::Default::kokkos_space::execution_space exec; |
| 606 |
Kokkos::parallel_for( |
| 607 |
"dealii::fill locally_owned_array, add", |
| 608 |
Kokkos::RangePolicy< |
| 609 |
MemorySpace::Default::kokkos_space::execution_space>( |
| 610 |
exec, 0, chunk_size), |
| 611 |
KOKKOS_LAMBDA(IndexType idx) { |
| 612 |
locally_owned_array |
| 613 |
.data()[import_indices_plain(idx)] += |
| 614 |
read_position[idx]; |
| 615 |
}); |
| 616 |
exec.fence(); |
| 617 |
|
| 618 |
read_position += chunk_size; |
| 619 |
} |
| 620 |
} |
| 621 |
else if (vector_operation == VectorOperation::min) |
| 622 |
{ |
| 623 |
for (const auto &import_indices_plain : |
| 624 |
import_indices_plain_dev) |
| 625 |
{ |
| 626 |
const auto chunk_size = import_indices_plain.size(); |
| 627 |
|
| 628 |
using IndexType = decltype(chunk_size); |
| 629 |
MemorySpace::Default::kokkos_space::execution_space exec; |
| 630 |
Kokkos::parallel_for( |
| 631 |
"dealii::fill locally_owned_array, min", |
| 632 |
Kokkos::RangePolicy< |
| 633 |
MemorySpace::Default::kokkos_space::execution_space>( |
| 634 |
exec, 0, chunk_size), |
| 635 |
KOKKOS_LAMBDA(IndexType idx) { |
| 636 |
locally_owned_array |
| 637 |
.data()[import_indices_plain(idx)] = |
| 638 |
internal::get_min( |
| 639 |
locally_owned_array |
| 640 |
.data()[import_indices_plain(idx)], |
| 641 |
read_position[idx]); |
| 642 |
}); |
| 643 |
exec.fence(); |
| 644 |
|
| 645 |
read_position += chunk_size; |
| 646 |
} |
| 647 |
} |
| 648 |
else if (vector_operation == VectorOperation::max) |
| 649 |
{ |
| 650 |
for (const auto &import_indices_plain : |
| 651 |
import_indices_plain_dev) |
| 652 |
{ |
| 653 |
const auto chunk_size = import_indices_plain.size(); |
| 654 |
|
| 655 |
using IndexType = decltype(chunk_size); |
| 656 |
MemorySpace::Default::kokkos_space::execution_space exec; |
| 657 |
Kokkos::parallel_for( |
| 658 |
"dealii::fill locally_owned_array, max", |
| 659 |
Kokkos::RangePolicy< |
| 660 |
MemorySpace::Default::kokkos_space::execution_space>( |
| 661 |
exec, 0, chunk_size), |
| 662 |
KOKKOS_LAMBDA(IndexType idx) { |
| 663 |
locally_owned_array |
| 664 |
.data()[import_indices_plain(idx)] = |
| 665 |
internal::get_max( |
| 666 |
locally_owned_array |
| 667 |
.data()[import_indices_plain(idx)], |
| 668 |
read_position[idx]); |
| 669 |
}); |
| 670 |
exec.fence(); |
| 671 |
|
| 672 |
read_position += chunk_size; |
| 673 |
} |
| 674 |
} |
| 675 |
else |
| 676 |
{ |
| 677 |
for (const auto &import_indices_plain : |
| 678 |
import_indices_plain_dev) |
| 679 |
{ |
| 680 |
// We can't easily assert here, so we just move the |
| 681 |
// pointer matching the host code. |
| 682 |
const auto chunk_size = import_indices_plain.size(); |
| 683 |
read_position += chunk_size; |
| 684 |
} |
| 685 |
} |
| 686 |
} |
| 687 |
else |
| 688 |
# endif |
| 689 |
{ |
| 690 |
// If the operation is no insertion, add the imported data to the |
| 691 |
// local values. For insert, nothing is done here (but in debug |
| 692 |
// mode we assert that the specified value is either zero or |
| 693 |
// matches with the ones already present |
| 694 |
if (vector_operation == VectorOperation::add) |
| 695 |
for (const auto &import_range : import_indices_data) |
| 696 |
for (unsigned int j = import_range.first; |
| 697 |
j < import_range.second; |
| 698 |
j++) |
| 699 |
locally_owned_array[j] += *read_position++; |
| 700 |
else if (vector_operation == VectorOperation::min) |
| 701 |
for (const auto &import_range : import_indices_data) |
| 702 |
for (unsigned int j = import_range.first; |
| 703 |
j < import_range.second; |
| 704 |
j++) |
| 705 |
{ |
| 706 |
locally_owned_array[j] = |
| 707 |
internal::get_min(*read_position, |
| 708 |
locally_owned_array[j]); |
| 709 |
++read_position; |
| 710 |
} |
| 711 |
else if (vector_operation == VectorOperation::max) |
| 712 |
for (const auto &import_range : import_indices_data) |
| 713 |
for (unsigned int j = import_range.first; |
| 714 |
j < import_range.second; |
| 715 |
j++) |
| 716 |
{ |
| 717 |
locally_owned_array[j] = |
| 718 |
internal::get_max(*read_position, |
| 719 |
locally_owned_array[j]); |
| 720 |
++read_position; |
| 721 |
} |
| 722 |
else |
| 723 |
for (const auto &import_range : import_indices_data) |
| 724 |
for (unsigned int j = import_range.first; |
| 725 |
j < import_range.second; |
| 726 |
j++, read_position++) |
| 727 |
// Below we use relatively large precision in units in the |
| 728 |
// last place (ULP) as this Assert can be easily triggered |
| 729 |
// in p::d::SolutionTransfer. The rationale is that during |
| 730 |
// interpolation on two elements sharing the face, values on |
| 731 |
// this face obtained from each side might be different due |
| 732 |
// to additions being done in different order. If the local |
| 733 |
// value is zero, it indicates that the local process has |
| 734 |
// not set the value during the cell loop and its value can |
| 735 |
// be safely overridden. |
| 736 |
Assert( |
| 737 |
*read_position == Number() || |
| 738 |
internal::get_abs(locally_owned_array[j] - |
| 739 |
*read_position) <= |
| 740 |
internal::get_abs(locally_owned_array[j] + |
| 741 |
*read_position) * |
| 742 |
100000. * |
| 743 |
std::numeric_limits<typename numbers::NumberTraits< |
| 744 |
Number>::real_type>::epsilon(), |
| 745 |
typename dealii::LinearAlgebra::distributed::Vector< |
| 746 |
Number>::ExcNonMatchingElements(*read_position, |
| 747 |
locally_owned_array[j], |
| 748 |
my_pid)); |
| 749 |
} |
| 750 |
|
| 751 |
AssertDimension(read_position - temporary_storage.data(), |
| 752 |
n_import_indices()); |
| 753 |
} |
| 754 |
|
| 755 |
// wait for the send operations to complete |
| 756 |
if (requests.size() > 0 && n_ghost_targets > 0) |
| 757 |
{ |
| 758 |
const int ierr = MPI_Waitall(n_ghost_targets, |
| 759 |
&requests[n_import_targets], |
| 760 |
MPI_STATUSES_IGNORE); |
| 761 |
AssertThrowMPI(ierr); |
| 762 |
} |
| 763 |
else |
| 764 |
AssertDimension(n_ghost_indices(), 0); |
| 765 |
|
| 766 |
// clear the ghost array in case we did not yet do that in the _start |
| 767 |
// function |
| 768 |
if (ghost_array.size() > 0) |
| 769 |
{ |
| 770 |
Assert(ghost_array.begin() != nullptr, ExcInternalError()); |
| 771 |
|
| 772 |
# if defined(DEAL_II_MPI_WITH_DEVICE_SUPPORT) |
| 773 |
if constexpr (std::is_same_v<MemorySpaceType, MemorySpace::Default>) |
| 774 |
{ |
| 775 |
Kokkos::deep_copy( |
| 776 |
Kokkos::View<Number *, MemorySpace::Default::kokkos_space>( |
| 777 |
ghost_array.data(), n_ghost_indices()), |
| 778 |
Number(0)); |
| 779 |
} |
| 780 |
else |
| 781 |
# endif |
| 782 |
{ |
| 783 |
std::fill(ghost_array.data(), |
| 784 |
ghost_array.data() + n_ghost_indices(), |
| 785 |
Number(0)); |
| 786 |
} |
| 787 |
} |
| 788 |
|
| 789 |
// clear the compress requests |
| 790 |
requests.resize(0); |
| 791 |
} |
| 792 |
|
| 793 |
|
| 794 |
# endif // ifdef DEAL_II_WITH_MPI |
| 795 |
#endif // ifndef DOXYGEN |
| 796 |
|
| 797 |
} // end of namespace MPI |
| 798 |
|
| 799 |
} // namespace Utilities |
| 800 |
|
| 801 |
|
| 802 |
DEAL_II_NAMESPACE_CLOSE |
| 803 |
|
| 804 |
#endif |