-
Notifications
You must be signed in to change notification settings - Fork 740
/
la_parallel_vector.templates.h
2198 lines (1806 loc) · 74.1 KB
/
la_parallel_vector.templates.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// ---------------------------------------------------------------------
//
// Copyright (C) 2011 - 2021 by the deal.II authors
//
// This file is part of the deal.II library.
//
// The deal.II library is free software; you can use it, redistribute
// it, and/or modify it under the terms of the GNU Lesser General
// Public License as published by the Free Software Foundation; either
// version 2.1 of the License, or (at your option) any later version.
// The full text of the license can be found in the file LICENSE.md at
// the top level directory of deal.II.
//
// ---------------------------------------------------------------------
#ifndef dealii_la_parallel_vector_templates_h
#define dealii_la_parallel_vector_templates_h
#include <deal.II/base/config.h>
#include <deal.II/base/cuda.h>
#include <deal.II/base/cuda_size.h>
#include <deal.II/lac/exceptions.h>
#include <deal.II/lac/la_parallel_vector.h>
#include <deal.II/lac/petsc_vector.h>
#include <deal.II/lac/read_write_vector.h>
#include <deal.II/lac/trilinos_vector.h>
#include <deal.II/lac/vector_operations_internal.h>
#include <memory>
DEAL_II_NAMESPACE_OPEN
namespace LinearAlgebra
{
namespace distributed
{
namespace internal
{
// In the import_from_ghosted_array_finish we might need to calculate the
// maximal and minimal value for the given number type, which is not
// straightforward for complex numbers. Therefore, comparison of complex
// numbers is prohibited and throws an exception.
template <typename Number>
Number
get_min(const Number a, const Number b)
{
return std::min(a, b);
}
template <typename Number>
std::complex<Number>
get_min(const std::complex<Number> a, const std::complex<Number>)
{
AssertThrow(false,
ExcMessage("VectorOperation::min not "
"implemented for complex numbers"));
return a;
}
template <typename Number>
Number
get_max(const Number a, const Number b)
{
return std::max(a, b);
}
template <typename Number>
std::complex<Number>
get_max(const std::complex<Number> a, const std::complex<Number>)
{
AssertThrow(false,
ExcMessage("VectorOperation::max not "
"implemented for complex numbers"));
return a;
}
// Resize the underlying array on the host or on the device
template <typename Number, typename MemorySpaceType>
struct la_parallel_vector_templates_functions
{
static_assert(std::is_same<MemorySpaceType, MemorySpace::Host>::value ||
std::is_same<MemorySpaceType, MemorySpace::CUDA>::value,
"MemorySpace should be Host or CUDA");
static void
resize_val(
const types::global_dof_index /*new_alloc_size*/,
types::global_dof_index & /*allocated_size*/,
::dealii::MemorySpace::MemorySpaceData<Number, MemorySpaceType>
& /*data*/,
const MPI_Comm & /*comm_sm*/)
{}
static void
import_elements(
const ::dealii::LinearAlgebra::ReadWriteVector<Number> & /*V*/,
::dealii::VectorOperation::values /*operation*/,
const std::shared_ptr<const ::dealii::Utilities::MPI::Partitioner> &
/*communication_pattern*/,
const IndexSet & /*locally_owned_elem*/,
::dealii::MemorySpace::MemorySpaceData<Number, MemorySpaceType>
& /*data*/)
{}
template <typename RealType>
static void
linfty_norm_local(
const ::dealii::MemorySpace::MemorySpaceData<Number, MemorySpaceType>
& /*data*/,
const unsigned int /*size*/,
RealType & /*max*/)
{}
};
template <typename Number>
struct la_parallel_vector_templates_functions<Number,
::dealii::MemorySpace::Host>
{
using size_type = types::global_dof_index;
static void
resize_val(const types::global_dof_index new_alloc_size,
types::global_dof_index & allocated_size,
::dealii::MemorySpace::
MemorySpaceData<Number, ::dealii::MemorySpace::Host> &data,
const MPI_Comm &comm_shared)
{
if (comm_shared == MPI_COMM_SELF)
{
Number *new_val;
Utilities::System::posix_memalign(
reinterpret_cast<void **>(&new_val),
64,
sizeof(Number) * new_alloc_size);
data.values = {new_val, [](Number *data) { std::free(data); }};
allocated_size = new_alloc_size;
data.values_sm = {
ArrayView<const Number>(data.values.get(), new_alloc_size)};
}
else
{
#ifdef DEAL_II_WITH_MPI
# if DEAL_II_MPI_VERSION_GTE(3, 0)
allocated_size = new_alloc_size;
const unsigned int size_sm =
Utilities::MPI::n_mpi_processes(comm_shared);
const unsigned int rank_sm =
Utilities::MPI::this_mpi_process(comm_shared);
MPI_Win mpi_window;
Number *data_this;
std::vector<Number *> others(size_sm);
MPI_Info info;
int ierr = MPI_Info_create(&info);
AssertThrowMPI(ierr);
ierr = MPI_Info_set(info, "alloc_shared_noncontig", "true");
AssertThrowMPI(ierr);
const std::size_t align_by = 64;
std::size_t s =
((new_alloc_size * sizeof(Number) + align_by - 1) /
sizeof(Number)) *
sizeof(Number);
ierr = MPI_Win_allocate_shared(
s, sizeof(Number), info, comm_shared, &data_this, &mpi_window);
AssertThrowMPI(ierr);
for (unsigned int i = 0; i < size_sm; i++)
{
int disp_unit;
MPI_Aint ssize;
const auto ierr = MPI_Win_shared_query(
mpi_window, i, &ssize, &disp_unit, &others[i]);
AssertThrowMPI(ierr);
}
Number *ptr_unaligned = others[rank_sm];
Number *ptr_aligned = ptr_unaligned;
AssertThrow(std::align(align_by,
new_alloc_size * sizeof(Number),
reinterpret_cast<void *&>(ptr_aligned),
s) != nullptr,
ExcNotImplemented());
unsigned int n_align_local = ptr_aligned - ptr_unaligned;
std::vector<unsigned int> n_align_sm(size_sm);
ierr = MPI_Allgather(&n_align_local,
1,
MPI_UNSIGNED,
n_align_sm.data(),
1,
MPI_UNSIGNED,
comm_shared);
AssertThrowMPI(ierr);
for (unsigned int i = 0; i < size_sm; i++)
others[i] += n_align_sm[i];
std::vector<unsigned int> new_alloc_sizes(size_sm);
ierr = MPI_Allgather(&new_alloc_size,
1,
MPI_UNSIGNED,
new_alloc_sizes.data(),
1,
MPI_UNSIGNED,
comm_shared);
AssertThrowMPI(ierr);
data.values_sm.resize(size_sm);
for (unsigned int i = 0; i < size_sm; i++)
data.values_sm[i] =
ArrayView<const Number>(others[i], new_alloc_sizes[i]);
data.values = {ptr_aligned, [mpi_window](Number *) mutable {
// note: we are creating here a copy of the
// window other approaches led to segmentation
// faults
const auto ierr = MPI_Win_free(&mpi_window);
AssertThrowMPI(ierr);
}};
# else
AssertThrow(false,
ExcMessage(
"Sorry, this feature requires MPI 3.0 support"));
# endif
#else
Assert(false, ExcInternalError());
#endif
}
}
static void
import_elements(
const ::dealii::LinearAlgebra::ReadWriteVector<Number> &V,
::dealii::VectorOperation::values operation,
const std::shared_ptr<const ::dealii::Utilities::MPI::Partitioner>
& communication_pattern,
const IndexSet &locally_owned_elem,
::dealii::MemorySpace::MemorySpaceData<Number,
::dealii::MemorySpace::Host>
&data)
{
Assert(
(operation == ::dealii::VectorOperation::add) ||
(operation == ::dealii::VectorOperation::insert),
ExcMessage(
"Only VectorOperation::add and VectorOperation::insert are allowed"));
::dealii::LinearAlgebra::distributed::
Vector<Number, ::dealii::MemorySpace::Host>
tmp_vector(communication_pattern);
// fill entries from ReadWriteVector into the distributed vector,
// including ghost entries. this is not really efficient right now
// because indices are translated twice, once by nth_index_in_set(i)
// and once for operator() of tmp_vector
const IndexSet &v_stored = V.get_stored_elements();
for (size_type i = 0; i < v_stored.n_elements(); ++i)
tmp_vector(v_stored.nth_index_in_set(i)) = V.local_element(i);
tmp_vector.compress(operation);
// Copy the local elements of tmp_vector to the right place in val
IndexSet tmp_index_set = tmp_vector.locally_owned_elements();
if (operation == VectorOperation::add)
{
for (size_type i = 0; i < tmp_index_set.n_elements(); ++i)
{
data.values[locally_owned_elem.index_within_set(
tmp_index_set.nth_index_in_set(i))] +=
tmp_vector.local_element(i);
}
}
else
{
for (size_type i = 0; i < tmp_index_set.n_elements(); ++i)
{
data.values[locally_owned_elem.index_within_set(
tmp_index_set.nth_index_in_set(i))] =
tmp_vector.local_element(i);
}
}
}
template <typename RealType>
static void
linfty_norm_local(const ::dealii::MemorySpace::MemorySpaceData<
Number,
::dealii::MemorySpace::Host> &data,
const unsigned int size,
RealType & max)
{
for (size_type i = 0; i < size; ++i)
max =
std::max(numbers::NumberTraits<Number>::abs(data.values[i]), max);
}
};
#ifdef DEAL_II_COMPILER_CUDA_AWARE
template <typename Number>
struct la_parallel_vector_templates_functions<Number,
::dealii::MemorySpace::CUDA>
{
using size_type = types::global_dof_index;
static void
resize_val(const types::global_dof_index new_alloc_size,
types::global_dof_index & allocated_size,
::dealii::MemorySpace::
MemorySpaceData<Number, ::dealii::MemorySpace::CUDA> &data,
const MPI_Comm &comm_sm)
{
(void)comm_sm;
static_assert(
std::is_same<Number, float>::value ||
std::is_same<Number, double>::value,
"Number should be float or double for CUDA memory space");
if (new_alloc_size > allocated_size)
{
Assert(((allocated_size > 0 && data.values_dev != nullptr) ||
data.values_dev == nullptr),
ExcInternalError());
Number *new_val_dev;
Utilities::CUDA::malloc(new_val_dev, new_alloc_size);
data.values_dev.reset(new_val_dev);
allocated_size = new_alloc_size;
}
else if (new_alloc_size == 0)
{
data.values_dev.reset();
allocated_size = 0;
}
}
static void
import_elements(
const ReadWriteVector<Number> &V,
VectorOperation::values operation,
std::shared_ptr<const Utilities::MPI::Partitioner>
communication_pattern,
const IndexSet &locally_owned_elem,
::dealii::MemorySpace::MemorySpaceData<Number,
::dealii::MemorySpace::CUDA>
&data)
{
Assert(
(operation == ::dealii::VectorOperation::add) ||
(operation == ::dealii::VectorOperation::insert),
ExcMessage(
"Only VectorOperation::add and VectorOperation::insert are allowed"));
::dealii::LinearAlgebra::distributed::
Vector<Number, ::dealii::MemorySpace::CUDA>
tmp_vector(communication_pattern);
// fill entries from ReadWriteVector into the distributed vector,
// including ghost entries. this is not really efficient right now
// because indices are translated twice, once by nth_index_in_set(i)
// and once for operator() of tmp_vector
const IndexSet & v_stored = V.get_stored_elements();
const size_type n_elements = v_stored.n_elements();
std::vector<size_type> indices(n_elements);
for (size_type i = 0; i < n_elements; ++i)
indices[i] = communication_pattern->global_to_local(
v_stored.nth_index_in_set(i));
// Move the indices to the device
size_type *indices_dev;
::dealii::Utilities::CUDA::malloc(indices_dev, n_elements);
::dealii::Utilities::CUDA::copy_to_dev(indices, indices_dev);
// Move the data to the device
Number *V_dev;
::dealii::Utilities::CUDA::malloc(V_dev, n_elements);
cudaError_t cuda_error_code = cudaMemcpy(V_dev,
V.begin(),
n_elements * sizeof(Number),
cudaMemcpyHostToDevice);
AssertCuda(cuda_error_code);
// Set the values in tmp_vector
const int n_blocks =
1 + n_elements / (::dealii::CUDAWrappers::chunk_size *
::dealii::CUDAWrappers::block_size);
::dealii::LinearAlgebra::CUDAWrappers::kernel::set_permutated<Number>
<<<n_blocks, ::dealii::CUDAWrappers::block_size>>>(
indices_dev, tmp_vector.begin(), V_dev, n_elements);
tmp_vector.compress(operation);
// Copy the local elements of tmp_vector to the right place in val
IndexSet tmp_index_set = tmp_vector.locally_owned_elements();
const size_type tmp_n_elements = tmp_index_set.n_elements();
indices.resize(tmp_n_elements);
for (size_type i = 0; i < tmp_n_elements; ++i)
indices[i] = locally_owned_elem.index_within_set(
tmp_index_set.nth_index_in_set(i));
::dealii::Utilities::CUDA::free(indices_dev);
::dealii::Utilities::CUDA::malloc(indices_dev, tmp_n_elements);
::dealii::Utilities::CUDA::copy_to_dev(indices, indices_dev);
if (operation == VectorOperation::add)
::dealii::LinearAlgebra::CUDAWrappers::kernel::add_permutated<
Number><<<n_blocks, ::dealii::CUDAWrappers::block_size>>>(
indices_dev,
data.values_dev.get(),
tmp_vector.begin(),
tmp_n_elements);
else
::dealii::LinearAlgebra::CUDAWrappers::kernel::set_permutated<
Number><<<n_blocks, ::dealii::CUDAWrappers::block_size>>>(
indices_dev,
data.values_dev.get(),
tmp_vector.begin(),
tmp_n_elements);
::dealii::Utilities::CUDA::free(indices_dev);
::dealii::Utilities::CUDA::free(V_dev);
}
template <typename RealType>
static void
linfty_norm_local(const ::dealii::MemorySpace::MemorySpaceData<
Number,
::dealii::MemorySpace::CUDA> &data,
const unsigned int size,
RealType & result)
{
static_assert(std::is_same<Number, RealType>::value,
"RealType should be the same type as Number");
Number * result_device;
cudaError_t error_code = cudaMalloc(&result_device, sizeof(Number));
AssertCuda(error_code);
error_code = cudaMemset(result_device, 0, sizeof(Number));
const int n_blocks = 1 + size / (::dealii::CUDAWrappers::chunk_size *
::dealii::CUDAWrappers::block_size);
::dealii::LinearAlgebra::CUDAWrappers::kernel::reduction<
Number,
::dealii::LinearAlgebra::CUDAWrappers::kernel::LInfty<Number>>
<<<dim3(n_blocks, 1), dim3(::dealii::CUDAWrappers::block_size)>>>(
result_device, data.values_dev.get(), size);
// Copy the result back to the host
error_code = cudaMemcpy(&result,
result_device,
sizeof(Number),
cudaMemcpyDeviceToHost);
AssertCuda(error_code);
// Free the memory on the device
error_code = cudaFree(result_device);
AssertCuda(error_code);
}
};
#endif
} // namespace internal
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::clear_mpi_requests()
{
#ifdef DEAL_II_WITH_MPI
for (auto &compress_request : compress_requests)
{
const int ierr = MPI_Request_free(&compress_request);
AssertThrowMPI(ierr);
}
compress_requests.clear();
for (auto &update_ghost_values_request : update_ghost_values_requests)
{
const int ierr = MPI_Request_free(&update_ghost_values_request);
AssertThrowMPI(ierr);
}
update_ghost_values_requests.clear();
#endif
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::resize_val(const size_type new_alloc_size,
const MPI_Comm &comm_sm)
{
internal::la_parallel_vector_templates_functions<
Number,
MemorySpaceType>::resize_val(new_alloc_size,
allocated_size,
data,
comm_sm);
thread_loop_partitioner =
std::make_shared<::dealii::parallel::internal::TBBPartitioner>();
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::reinit(const size_type size,
const bool omit_zeroing_entries)
{
clear_mpi_requests();
// check whether we need to reallocate
resize_val(size, comm_sm);
// delete previous content in import data
import_data.values.reset();
import_data.values_dev.reset();
// set partitioner to serial version
partitioner = std::make_shared<Utilities::MPI::Partitioner>(size);
// set entries to zero if so requested
if (omit_zeroing_entries == false)
this->operator=(Number());
else
zero_out_ghost_values();
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::reinit(
const types::global_dof_index local_size,
const types::global_dof_index ghost_size,
const MPI_Comm & comm,
const MPI_Comm & comm_sm)
{
clear_mpi_requests();
this->comm_sm = comm_sm;
// check whether we need to reallocate
resize_val(local_size + ghost_size, comm_sm);
// delete previous content in import data
import_data.values.reset();
import_data.values_dev.reset();
// create partitioner
partitioner = std::make_shared<Utilities::MPI::Partitioner>(local_size,
ghost_size,
comm);
this->operator=(Number());
}
template <typename Number, typename MemorySpaceType>
template <typename Number2>
void
Vector<Number, MemorySpaceType>::reinit(
const Vector<Number2, MemorySpaceType> &v,
const bool omit_zeroing_entries)
{
clear_mpi_requests();
Assert(v.partitioner.get() != nullptr, ExcNotInitialized());
this->comm_sm = v.comm_sm;
// check whether the partitioners are
// different (check only if the are allocated
// differently, not if the actual data is
// different)
if (partitioner.get() != v.partitioner.get())
{
partitioner = v.partitioner;
const size_type new_allocated_size =
partitioner->locally_owned_size() + partitioner->n_ghost_indices();
resize_val(new_allocated_size, this->comm_sm);
}
if (omit_zeroing_entries == false)
this->operator=(Number());
else
zero_out_ghost_values();
// do not reallocate import_data directly, but only upon request. It
// is only used as temporary storage for compress() and
// update_ghost_values, and we might have vectors where we never
// call these methods and hence do not need to have the storage.
import_data.values.reset();
import_data.values_dev.reset();
thread_loop_partitioner = v.thread_loop_partitioner;
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::reinit(
const IndexSet &locally_owned_indices,
const IndexSet &ghost_indices,
const MPI_Comm &communicator)
{
// set up parallel partitioner with index sets and communicator
reinit(std::make_shared<Utilities::MPI::Partitioner>(
locally_owned_indices, ghost_indices, communicator));
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::reinit(
const IndexSet &locally_owned_indices,
const MPI_Comm &communicator)
{
// set up parallel partitioner with index sets and communicator
reinit(
std::make_shared<Utilities::MPI::Partitioner>(locally_owned_indices,
communicator));
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::reinit(
const std::shared_ptr<const Utilities::MPI::Partitioner> &partitioner_in,
const MPI_Comm & comm_sm)
{
clear_mpi_requests();
partitioner = partitioner_in;
this->comm_sm = comm_sm;
// set vector size and allocate memory
const size_type new_allocated_size =
partitioner->locally_owned_size() + partitioner->n_ghost_indices();
resize_val(new_allocated_size, comm_sm);
// initialize to zero
*this = Number();
// do not reallocate import_data directly, but only upon request. It
// is only used as temporary storage for compress() and
// update_ghost_values, and we might have vectors where we never
// call these methods and hence do not need to have the storage.
import_data.values.reset();
import_data.values_dev.reset();
vector_is_ghosted = false;
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector()
: partitioner(std::make_shared<Utilities::MPI::Partitioner>())
, allocated_size(0)
, comm_sm(MPI_COMM_SELF)
{
reinit(0);
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector(
const Vector<Number, MemorySpaceType> &v)
: Subscriptor()
, allocated_size(0)
, vector_is_ghosted(false)
, comm_sm(MPI_COMM_SELF)
{
reinit(v, true);
thread_loop_partitioner = v.thread_loop_partitioner;
const size_type this_size = locally_owned_size();
if (this_size > 0)
{
dealii::internal::VectorOperations::
functions<Number, Number, MemorySpaceType>::copy(
thread_loop_partitioner,
partitioner->locally_owned_size(),
v.data,
data);
}
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector(Vector<Number, MemorySpaceType> &&v)
: Vector()
{
static_cast<Subscriptor &>(*this) = static_cast<Subscriptor &&>(v);
this->swap(v);
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector(const IndexSet &local_range,
const IndexSet &ghost_indices,
const MPI_Comm &communicator)
: allocated_size(0)
, vector_is_ghosted(false)
, comm_sm(MPI_COMM_SELF)
{
reinit(local_range, ghost_indices, communicator);
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector(const IndexSet &local_range,
const MPI_Comm &communicator)
: allocated_size(0)
, vector_is_ghosted(false)
, comm_sm(MPI_COMM_SELF)
{
reinit(local_range, communicator);
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector(const size_type size)
: allocated_size(0)
, vector_is_ghosted(false)
, comm_sm(MPI_COMM_SELF)
{
reinit(size, false);
}
template <typename Number, typename MemorySpaceType>
Vector<Number, MemorySpaceType>::Vector(
const std::shared_ptr<const Utilities::MPI::Partitioner> &partitioner)
: allocated_size(0)
, vector_is_ghosted(false)
, comm_sm(MPI_COMM_SELF)
{
reinit(partitioner);
}
template <typename Number, typename MemorySpaceType>
inline Vector<Number, MemorySpaceType>::~Vector()
{
try
{
clear_mpi_requests();
}
catch (...)
{}
}
template <typename Number, typename MemorySpaceType>
inline Vector<Number, MemorySpaceType> &
Vector<Number, MemorySpaceType>::
operator=(const Vector<Number, MemorySpaceType> &c)
{
#ifdef _MSC_VER
return this->operator=<Number>(c);
#else
return this->template operator=<Number>(c);
#endif
}
template <typename Number, typename MemorySpaceType>
template <typename Number2>
inline Vector<Number, MemorySpaceType> &
Vector<Number, MemorySpaceType>::
operator=(const Vector<Number2, MemorySpaceType> &c)
{
Assert(c.partitioner.get() != nullptr, ExcNotInitialized());
// we update ghost values whenever one of the input or output vector
// already held ghost values or when we import data from a vector with
// the same local range but different ghost layout
bool must_update_ghost_values = c.vector_is_ghosted;
this->comm_sm = c.comm_sm;
// check whether the two vectors use the same parallel partitioner. if
// not, check if all local ranges are the same (that way, we can
// exchange data between different parallel layouts). One variant which
// is included here and necessary for compatibility with the other
// distributed vector classes (Trilinos, PETSc) is the case when vector
// c does not have any ghosts (constructed without ghost elements given)
// but the current vector does: In that case, we need to exchange data
// also when none of the two vector had updated its ghost values before.
if (partitioner.get() == nullptr)
reinit(c, true);
else if (partitioner.get() != c.partitioner.get())
{
// local ranges are also the same if both partitioners are empty
// (even if they happen to define the empty range as [0,0) or [c,c)
// for some c!=0 in a different way).
int local_ranges_are_identical =
(partitioner->local_range() == c.partitioner->local_range() ||
(partitioner->local_range().second ==
partitioner->local_range().first &&
c.partitioner->local_range().second ==
c.partitioner->local_range().first));
if ((c.partitioner->n_mpi_processes() > 1 &&
Utilities::MPI::min(local_ranges_are_identical,
c.partitioner->get_mpi_communicator()) ==
0) ||
!local_ranges_are_identical)
reinit(c, true);
else
must_update_ghost_values |= vector_is_ghosted;
must_update_ghost_values |=
(c.partitioner->ghost_indices_initialized() == false &&
partitioner->ghost_indices_initialized() == true);
}
else
must_update_ghost_values |= vector_is_ghosted;
thread_loop_partitioner = c.thread_loop_partitioner;
const size_type this_size = partitioner->locally_owned_size();
if (this_size > 0)
{
dealii::internal::VectorOperations::
functions<Number, Number2, MemorySpaceType>::copy(
thread_loop_partitioner, this_size, c.data, data);
}
if (must_update_ghost_values)
update_ghost_values();
else
zero_out_ghost_values();
return *this;
}
template <typename Number, typename MemorySpaceType>
template <typename Number2>
void
Vector<Number, MemorySpaceType>::copy_locally_owned_data_from(
const Vector<Number2, MemorySpaceType> &src)
{
AssertDimension(partitioner->locally_owned_size(),
src.partitioner->locally_owned_size());
if (partitioner->locally_owned_size() > 0)
{
dealii::internal::VectorOperations::
functions<Number, Number2, MemorySpaceType>::copy(
thread_loop_partitioner,
partitioner->locally_owned_size(),
src.data,
data);
}
}
template <typename Number, typename MemorySpaceType>
template <typename MemorySpaceType2>
void
Vector<Number, MemorySpaceType>::import(
const Vector<Number, MemorySpaceType2> &src,
VectorOperation::values operation)
{
Assert(src.partitioner.get() != nullptr, ExcNotInitialized());
Assert(partitioner->locally_owned_range() ==
src.partitioner->locally_owned_range(),
ExcMessage("Locally owned indices should be identical."));
Assert(partitioner->ghost_indices() == src.partitioner->ghost_indices(),
ExcMessage("Ghost indices should be identical."));
::dealii::internal::VectorOperations::
functions<Number, Number, MemorySpaceType>::import_elements(
thread_loop_partitioner, allocated_size, operation, src.data, data);
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::compress(
::dealii::VectorOperation::values operation)
{
compress_start(0, operation);
compress_finish(operation);
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::update_ghost_values() const
{
update_ghost_values_start();
update_ghost_values_finish();
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::zero_out_ghosts() const
{
this->zero_out_ghost_values();
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::zero_out_ghost_values() const
{
if (data.values != nullptr)
std::fill_n(data.values.get() + partitioner->locally_owned_size(),
partitioner->n_ghost_indices(),
Number());
#ifdef DEAL_II_COMPILER_CUDA_AWARE
if (data.values_dev != nullptr)
{
const cudaError_t cuda_error_code =
cudaMemset(data.values_dev.get() +
partitioner->locally_owned_size(),
0,
partitioner->n_ghost_indices() * sizeof(Number));
AssertCuda(cuda_error_code);
}
#endif
vector_is_ghosted = false;
}
template <typename Number, typename MemorySpaceType>
void
Vector<Number, MemorySpaceType>::compress_start(
const unsigned int communication_channel,
::dealii::VectorOperation::values operation)
{
AssertIndexRange(communication_channel, 200);
Assert(vector_is_ghosted == false,
ExcMessage("Cannot call compress() on a ghosted vector"));
#ifdef DEAL_II_WITH_MPI
// make this function thread safe
std::lock_guard<std::mutex> lock(mutex);
// allocate import_data in case it is not set up yet
if (partitioner->n_import_indices() > 0)
{
# if defined(DEAL_II_COMPILER_CUDA_AWARE) && \
defined(DEAL_II_MPI_WITH_CUDA_SUPPORT)
if (std::is_same<MemorySpaceType, dealii::MemorySpace::CUDA>::value)
{
if (import_data.values_dev == nullptr)
import_data.values_dev.reset(
Utilities::CUDA::allocate_device_data<Number>(
partitioner->n_import_indices()));
}
else
# endif
{
# if !defined(DEAL_II_COMPILER_CUDA_AWARE) && \
defined(DEAL_II_MPI_WITH_CUDA_SUPPORT)
static_assert(
std::is_same<MemorySpaceType, dealii::MemorySpace::Host>::value,
"This code path should only be compiled for CUDA-aware-MPI for MemorySpace::Host!");
# endif