| 1 |
// ------------------------------------------------------------------------ |
| 2 |
// |
| 3 |
// SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 |
// Copyright (C) 2020 - 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_noncontiguous_partitioner_h |
| 16 |
#define dealii_mpi_noncontiguous_partitioner_h |
| 17 |
|
| 18 |
#include <deal.II/base/config.h> |
| 19 |
|
| 20 |
#include <deal.II/base/array_view.h> |
| 21 |
#include <deal.II/base/communication_pattern_base.h> |
| 22 |
#include <deal.II/base/mpi_stub.h> |
| 23 |
#include <deal.II/base/types.h> |
| 24 |
|
| 25 |
#include <deal.II/lac/vector_operation.h> |
| 26 |
|
| 27 |
|
| 28 |
DEAL_II_NAMESPACE_OPEN |
| 29 |
|
| 30 |
namespace Utilities |
| 31 |
{ |
| 32 |
namespace MPI |
| 33 |
{ |
| 34 |
/** |
| 35 |
* A flexible Partitioner class, which does not impose restrictions |
| 36 |
* regarding the order of the underlying index sets. In other words, |
| 37 |
* this class implements the interface of the |
| 38 |
* Utilities::MPI::CommunicationPatternBase base class with no |
| 39 |
* assumption that every process stores a contiguous part of the |
| 40 |
* array of objects, but that indeed the locally owned indices |
| 41 |
* can be an arbitrary subset of all indices of elements of the array |
| 42 |
* to which they refer. |
| 43 |
* |
| 44 |
* If you want to store only contiguous parts of these arrays on |
| 45 |
* each process, take a look at Utilities::MPI::Partitioner. |
| 46 |
*/ |
| 47 |
class NoncontiguousPartitioner |
| 48 |
: public Utilities::MPI::CommunicationPatternBase |
| 49 |
{ |
| 50 |
public: |
| 51 |
/** |
| 52 |
* Default constructor. Requires calling one of the reinit() functions |
| 53 |
* to create a valid object. |
| 54 |
*/ |
| 55 |
NoncontiguousPartitioner() = default; |
| 56 |
|
| 57 |
/** |
| 58 |
* Constructor. Set up point-to-point communication pattern based on the |
| 59 |
* IndexSets arguments @p indexset_locally_owned and @p indexset_ghost for the MPI |
| 60 |
* communicator @p communicator. |
| 61 |
*/ |
| 62 |
NoncontiguousPartitioner(const IndexSet &indexset_locally_owned, |
| 63 |
const IndexSet &indexset_ghost, |
| 64 |
const MPI_Comm communicator); |
| 65 |
|
| 66 |
/** |
| 67 |
* Constructor. Same as above but for vectors of indices @p indices_locally_owned |
| 68 |
* and @p indices_ghost. This allows the indices to not be sorted and the |
| 69 |
* values are read and written automatically at the right position of |
| 70 |
* the vector during update_values(), update_values_start(), and |
| 71 |
* update_values_finish(). It is allowed to include entries with the |
| 72 |
* value numbers::invalid_dof_index which do not take part of the index |
| 73 |
* exchange but are present in the data vectors as padding. |
| 74 |
*/ |
| 75 |
NoncontiguousPartitioner( |
| 76 |
const std::vector<types::global_dof_index> &indices_locally_owned, |
| 77 |
const std::vector<types::global_dof_index> &indices_ghost, |
| 78 |
const MPI_Comm communicator); |
| 79 |
|
| 80 |
/** |
| 81 |
* Fill the vector @p ghost_array according to the precomputed communication |
| 82 |
* pattern with values from @p locally_owned_array. |
| 83 |
* |
| 84 |
* In the default case, only one object is communicated per entry |
| 85 |
* (`n_components_templated == 1'). If you want to communicate more |
| 86 |
* entries, you can increase the value of @p n_components_templated in the |
| 87 |
* case that you know the size at compile time. If you want to set the |
| 88 |
* size during runtime, you can set @p n_components. However, |
| 89 |
* @p n_components_templated has to be set to `0` in this case. Either |
| 90 |
* @p n_components_templated or @p n_components can be set. |
| 91 |
* |
| 92 |
* @pre The vectors only have to provide a method begin(), which allows |
| 93 |
* to access their raw data. |
| 94 |
* |
| 95 |
* @pre The size of both vectors must be at least as large as the number |
| 96 |
* of entries in the index sets passed to the constructors or the |
| 97 |
* reinit() functions. |
| 98 |
* |
| 99 |
* @note This function calls the methods update_values_start() and |
| 100 |
* update_values_finish() in sequence. Users can call these two |
| 101 |
* functions separately and hereby overlap communication and |
| 102 |
* computation. |
| 103 |
*/ |
| 104 |
template <typename Number, unsigned int n_components_templated = 1> |
| 105 |
void |
| 106 |
export_to_ghosted_array( |
| 107 |
const ArrayView<const Number> &locally_owned_array, |
| 108 |
const ArrayView<Number> &ghost_array, |
| 109 |
const unsigned int n_components = 0) const; |
| 110 |
|
| 111 |
/** |
| 112 |
* Same as above but with an interface similar to |
| 113 |
* Utilities::MPI::Partitioner::export_to_ghosted_array_start and |
| 114 |
* Utilities::MPI::Partitioner::export_to_ghosted_array_finish. In this |
| 115 |
* function, the user can provide the temporary data structures to be |
| 116 |
* used. |
| 117 |
* |
| 118 |
* @pre The size of the @p temporary_storage vector has to be at least |
| 119 |
* temporary_storage_size. The reason for this is that this vector is |
| 120 |
* used as buffer for both sending and receiving data. |
| 121 |
* |
| 122 |
* @note Any value less than 10 is a valid value of |
| 123 |
* @p communication_channel. |
| 124 |
*/ |
| 125 |
template <typename Number, unsigned int n_components_templated = 1> |
| 126 |
void |
| 127 |
export_to_ghosted_array( |
| 128 |
const unsigned int communication_channel, |
| 129 |
const ArrayView<const Number> &locally_owned_array, |
| 130 |
const ArrayView<Number> &temporary_storage, |
| 131 |
const ArrayView<Number> &ghost_array, |
| 132 |
std::vector<MPI_Request> &requests, |
| 133 |
const unsigned int n_components = 0) const; |
| 134 |
|
| 135 |
/** |
| 136 |
* Start update: Data is packed, non-blocking send and receives |
| 137 |
* are started. |
| 138 |
* |
| 139 |
* @note In contrast to the function |
| 140 |
* Utilities::MPI::Partitioner::export_to_ghosted_array_start, the user |
| 141 |
* does not pass a reference to the destination vector, since the data |
| 142 |
* is received into a designated part of the buffer @p temporary_storage. This |
| 143 |
* allows for padding and other post-processing of the received data. |
| 144 |
* |
| 145 |
* @pre The required size of the vectors are the same as in the functions |
| 146 |
* above. |
| 147 |
* |
| 148 |
* @note Any value less than 10 is a valid value of |
| 149 |
* @p communication_channel. |
| 150 |
*/ |
| 151 |
template <typename Number, unsigned int n_components_templated = 1> |
| 152 |
void |
| 153 |
export_to_ghosted_array_start( |
| 154 |
const unsigned int communication_channel, |
| 155 |
const ArrayView<const Number> &locally_owned_array, |
| 156 |
const ArrayView<Number> &temporary_storage, |
| 157 |
std::vector<MPI_Request> &requests, |
| 158 |
const unsigned int n_components = 0) const; |
| 159 |
|
| 160 |
/** |
| 161 |
* Finish update. The method waits until all data has been sent and |
| 162 |
* received. Once data from any process is received it is processed and |
| 163 |
* placed at the right position of the vector @p dst. |
| 164 |
* |
| 165 |
* @note In contrast to the function |
| 166 |
* Utilities::MPI::Partitioner::export_to_ghosted_array_finish, the user |
| 167 |
* also has to pass a reference to the buffer @p temporary_storage, |
| 168 |
* since the data has been received into the buffer and not into the |
| 169 |
* destination vector. |
| 170 |
* |
| 171 |
* @pre The required size of the vectors are the same as in the functions |
| 172 |
* above. |
| 173 |
*/ |
| 174 |
template <typename Number, unsigned int n_components_templated = 1> |
| 175 |
void |
| 176 |
export_to_ghosted_array_finish( |
| 177 |
const ArrayView<const Number> &temporary_storage, |
| 178 |
const ArrayView<Number> &ghost_array, |
| 179 |
std::vector<MPI_Request> &requests, |
| 180 |
const unsigned int n_components = 0) const; |
| 181 |
|
| 182 |
/** |
| 183 |
* Similar to the above functions but for importing vector entries |
| 184 |
* from @p ghost_array to @p locally_owned_storage. |
| 185 |
* |
| 186 |
* @note In contrast to the functions in |
| 187 |
* Utilities::MPI::Partitioner, this function expects that |
| 188 |
* locally_owned_storage is empty. |
| 189 |
*/ |
| 190 |
template <typename Number> |
| 191 |
void |
| 192 |
import_from_ghosted_array( |
| 193 |
const VectorOperation::values vector_operation, |
| 194 |
const ArrayView<Number> &ghost_array, |
| 195 |
const ArrayView<Number> &locally_owned_storage) const; |
| 196 |
|
| 197 |
/** |
| 198 |
* Similar to the above function with the difference that |
| 199 |
* users can provide temporaty arrays. This function calls |
| 200 |
* import_from_ghosted_array_start() and |
| 201 |
* import_from_ghosted_array_finish() in sequence. |
| 202 |
*/ |
| 203 |
template <typename Number> |
| 204 |
void |
| 205 |
import_from_ghosted_array(const VectorOperation::values vector_operation, |
| 206 |
const unsigned int communication_channel, |
| 207 |
const ArrayView<Number> &ghost_array, |
| 208 |
const ArrayView<Number> &temporary_storage, |
| 209 |
const ArrayView<Number> &locally_owned_storage, |
| 210 |
std::vector<MPI_Request> &requests) const; |
| 211 |
|
| 212 |
/** |
| 213 |
* Start update for importig values: Data is packed, non-blocking send |
| 214 |
* and receives are started. |
| 215 |
*/ |
| 216 |
template <typename Number> |
| 217 |
void |
| 218 |
import_from_ghosted_array_start( |
| 219 |
const VectorOperation::values vector_operation, |
| 220 |
const unsigned int communication_channel, |
| 221 |
const ArrayView<Number> &ghost_array, |
| 222 |
const ArrayView<Number> &temporary_storage, |
| 223 |
std::vector<MPI_Request> &requests) const; |
| 224 |
|
| 225 |
/** |
| 226 |
* Finish update for importing values. The method waits until all data has |
| 227 |
* been sent and received. Once data from any process is received it is |
| 228 |
* processed and placed at the right position of the vector |
| 229 |
* @p locally_owned_storage. |
| 230 |
*/ |
| 231 |
template <typename Number> |
| 232 |
void |
| 233 |
import_from_ghosted_array_finish( |
| 234 |
const VectorOperation::values vector_operation, |
| 235 |
const ArrayView<const Number> &temporary_storage, |
| 236 |
const ArrayView<Number> &locally_owned_storage, |
| 237 |
std::vector<MPI_Request> &requests) const; |
| 238 |
|
| 239 |
/** |
| 240 |
* Returns the number of processes this process sends data to and the |
| 241 |
* number of processes this process receives data from. |
| 242 |
*/ |
| 243 |
std::pair<unsigned int, unsigned int> |
| 244 |
n_targets() const; |
| 245 |
|
| 246 |
/** |
| 247 |
* Return the size of the temporary storage needed by the |
| 248 |
* export_to_ghosted_array() functions, if the temporary storage is |
| 249 |
* handled by the user code. |
| 250 |
*/ |
| 251 |
unsigned int |
| 252 |
temporary_storage_size() const; |
| 253 |
|
| 254 |
/** |
| 255 |
* Return memory consumption in Byte. |
| 256 |
*/ |
| 257 |
types::global_dof_index |
| 258 |
memory_consumption(); |
| 259 |
|
| 260 |
/** |
| 261 |
* Return the underlying MPI communicator. |
| 262 |
*/ |
| 263 |
MPI_Comm |
| 264 |
get_mpi_communicator() const override; |
| 265 |
|
| 266 |
void |
| 267 |
reinit(const IndexSet &locally_owned_indices, |
| 268 |
const IndexSet &ghost_indices, |
| 269 |
const MPI_Comm communicator) override; |
| 270 |
|
| 271 |
/** |
| 272 |
* Initialize the inner data structures using explicit sets of |
| 273 |
* indices. See the documentation of the other reinit() function for |
| 274 |
* what the function does. |
| 275 |
*/ |
| 276 |
void |
| 277 |
reinit(const std::vector<types::global_dof_index> &locally_owned_indices, |
| 278 |
const std::vector<types::global_dof_index> &ghost_indices, |
| 279 |
const MPI_Comm communicator); |
| 280 |
|
| 281 |
private: |
| 282 |
/** |
| 283 |
* MPI communicator. |
| 284 |
*/ |
| 285 |
MPI_Comm communicator; |
| 286 |
|
| 287 |
/** |
| 288 |
* The ranks this process sends data to. |
| 289 |
*/ |
| 290 |
std::vector<unsigned int> send_ranks; |
| 291 |
|
| 292 |
/** |
| 293 |
* Offset of each process within send_buffer. |
| 294 |
* |
| 295 |
* @note Together with `send_indices` this forms a CRS data structure. |
| 296 |
*/ |
| 297 |
std::vector<types::global_dof_index> send_ptr; |
| 298 |
|
| 299 |
/** |
| 300 |
* Local index of each entry in send_buffer within the destination |
| 301 |
* vector. |
| 302 |
* |
| 303 |
* @note Together with `send_ptr` this forms a CRS data structure. |
| 304 |
*/ |
| 305 |
std::vector<types::global_dof_index> send_indices; |
| 306 |
|
| 307 |
/** |
| 308 |
* The ranks this process receives data from. |
| 309 |
*/ |
| 310 |
std::vector<unsigned int> recv_ranks; |
| 311 |
|
| 312 |
/** |
| 313 |
* Offset of each process within recv_buffer. |
| 314 |
* |
| 315 |
* @note Together with `recv_indices` this forms a CRS data structure. |
| 316 |
*/ |
| 317 |
std::vector<types::global_dof_index> recv_ptr; |
| 318 |
|
| 319 |
/** |
| 320 |
* Local index of each entry in recv_buffer within the destination |
| 321 |
* vector. |
| 322 |
* |
| 323 |
* In the case that an entry is more than once requested by the |
| 324 |
* same rank, local index of each entry in recv_buffer within |
| 325 |
* recv_indices_duplicates_ptr. |
| 326 |
* |
| 327 |
* @note Together with `recv_ptr` this forms a CRS data structure. |
| 328 |
*/ |
| 329 |
std::vector<types::global_dof_index> recv_indices; |
| 330 |
|
| 331 |
/** |
| 332 |
* In the case that an entry is more than once requested by the |
| 333 |
* same rank, offset of each index in recv_indices_duplicates. |
| 334 |
*/ |
| 335 |
std::vector<unsigned int> recv_indices_duplicates_ptr; |
| 336 |
|
| 337 |
/** |
| 338 |
* Local index of each entry within the destination vector in the |
| 339 |
* case that an entry is more than once requested by the same rank. |
| 340 |
* |
| 341 |
* @note Together with `recv_indices_duplicates_ptr` |
| 342 |
* this forms a CRS data structure. |
| 343 |
*/ |
| 344 |
std::vector<unsigned int> recv_indices_duplicates; |
| 345 |
|
| 346 |
/** |
| 347 |
* Buffer containing the values sorted by rank for sending and receiving. |
| 348 |
* |
| 349 |
* @note Only allocated if not provided externally by user. |
| 350 |
* |
| 351 |
* @note At this place we do not know the type of the data to be sent. So |
| 352 |
* we use an arbitrary type of size 1 byte. The type is cast to the |
| 353 |
* requested type in the relevant functions. |
| 354 |
*/ |
| 355 |
mutable std::vector<std::uint8_t> buffers; |
| 356 |
|
| 357 |
/** |
| 358 |
* MPI requests for sending and receiving. |
| 359 |
* |
| 360 |
* @note Only allocated if not provided externally by user. |
| 361 |
*/ |
| 362 |
mutable std::vector<MPI_Request> requests; |
| 363 |
}; |
| 364 |
|
| 365 |
} // namespace MPI |
| 366 |
} // namespace Utilities |
| 367 |
|
| 368 |
DEAL_II_NAMESPACE_CLOSE |
| 369 |
|
| 370 |
#endif |