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
|
// ---------------------------------------------------------------------
//
// Copyright (C) 2009 - 2018 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 at
// the top level of the deal.II distribution.
//
// ---------------------------------------------------------------------
#include <deal.II/base/parallel.h>
DEAL_II_NAMESPACE_OPEN
namespace internal
{
namespace VectorImplementation
{
// set minimum grain size. this value has been determined by experiments
// with the actual dealii::Vector implementation and takes vectorization
// that is done inside most functions of dealii::Vector into account
// (without vectorization, we would land at around 1000). for smaller
// values than this, the scheduling overhead will be significant. if the
// value becomes too large, on the other hand, the work load cannot be
// split into enough chunks and the problem becomes badly balanced. the
// tests are documented at https://github.com/dealii/dealii/issues/2496
unsigned int minimum_parallel_grain_size = 4096;
}
namespace SparseMatrixImplementation
{
// set this value to 1/16 of the value of the minimum grain size of
// vectors (a factor of 4 because we assume that sparse matrix-vector
// products do not vectorize and the other factor of 4 because we expect
// at least four entries per row). this rests on the fact that we have to
// do a lot more work per row of a matrix than per element of a vector. it
// could possibly be reduced even further but that doesn't appear worth it
// any more for anything but very small matrices that we don't care that
// much about anyway.
unsigned int minimum_parallel_grain_size = 256;
}
}
namespace parallel
{
namespace internal
{
#ifdef DEAL_II_WITH_THREADS
TBBPartitioner::TBBPartitioner()
:
my_partitioner(std::make_shared<tbb::affinity_partitioner>()),
in_use(false)
{}
TBBPartitioner::~TBBPartitioner()
{
AssertNothrow(in_use == false,
ExcInternalError("A vector partitioner goes out of scope, but "
"it appears to be still in use."));
}
std::shared_ptr<tbb::affinity_partitioner>
TBBPartitioner::acquire_one_partitioner()
{
dealii::Threads::Mutex::ScopedLock lock(mutex);
if (in_use)
return std::make_shared<tbb::affinity_partitioner>();
in_use = true;
return my_partitioner;
}
void
TBBPartitioner::release_one_partitioner(std::shared_ptr<tbb::affinity_partitioner> &p)
{
if (p.get() == my_partitioner.get())
{
dealii::Threads::Mutex::ScopedLock lock(mutex);
in_use = false;
}
}
#else
TBBPartitioner::TBBPartitioner() = default;
#endif
}
}
DEAL_II_NAMESPACE_CLOSE
|
