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
|
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_PARALLELIZER_H
#define EIGEN_PARALLELIZER_H
namespace Eigen {
namespace internal {
/** \internal */
inline void manage_multi_threading(Action action, int* v)
{
static EIGEN_UNUSED int m_maxThreads = -1;
if(action==SetAction)
{
eigen_internal_assert(v!=0);
m_maxThreads = *v;
}
else if(action==GetAction)
{
eigen_internal_assert(v!=0);
#ifdef EIGEN_HAS_OPENMP
if(m_maxThreads>0)
*v = m_maxThreads;
else
*v = omp_get_max_threads();
#else
*v = 1;
#endif
}
else
{
eigen_internal_assert(false);
}
}
}
/** Must be call first when calling Eigen from multiple threads */
inline void initParallel()
{
int nbt;
internal::manage_multi_threading(GetAction, &nbt);
std::ptrdiff_t l1, l2, l3;
internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
}
/** \returns the max number of threads reserved for Eigen
* \sa setNbThreads */
inline int nbThreads()
{
int ret;
internal::manage_multi_threading(GetAction, &ret);
return ret;
}
/** Sets the max number of threads reserved for Eigen
* \sa nbThreads */
inline void setNbThreads(int v)
{
internal::manage_multi_threading(SetAction, &v);
}
namespace internal {
template<typename Index> struct GemmParallelInfo
{
GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
int volatile sync;
int volatile users;
Index lhs_start;
Index lhs_length;
};
template<bool Condition, typename Functor, typename Index>
void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpose)
{
// TODO when EIGEN_USE_BLAS is defined,
// we should still enable OMP for other scalar types
#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
// FIXME the transpose variable is only needed to properly split
// the matrix product when multithreading is enabled. This is a temporary
// fix to support row-major destination matrices. This whole
// parallelizer mechanism has to be redisigned anyway.
EIGEN_UNUSED_VARIABLE(transpose);
func(0,rows, 0,cols);
#else
// Dynamically check whether we should enable or disable OpenMP.
// The conditions are:
// - the max number of threads we can create is greater than 1
// - we are not already in a parallel code
// - the sizes are large enough
// 1- are we already in a parallel session?
// FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
if((!Condition) || (omp_get_num_threads()>1))
return func(0,rows, 0,cols);
Index size = transpose ? rows : cols;
// 2- compute the maximal number of threads from the size of the product:
// FIXME this has to be fine tuned
Index max_threads = std::max<Index>(1,size / 32);
// 3 - compute the number of threads we are going to use
Index threads = std::min<Index>(nbThreads(), max_threads);
if(threads==1)
return func(0,rows, 0,cols);
Eigen::initParallel();
func.initParallelSession();
if(transpose)
std::swap(rows,cols);
Index blockCols = (cols / threads) & ~Index(0x3);
Index blockRows = (rows / threads);
blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
GemmParallelInfo<Index>* info = new GemmParallelInfo<Index>[threads];
#pragma omp parallel num_threads(threads)
{
Index i = omp_get_thread_num();
Index r0 = i*blockRows;
Index actualBlockRows = (i+1==threads) ? rows-r0 : blockRows;
Index c0 = i*blockCols;
Index actualBlockCols = (i+1==threads) ? cols-c0 : blockCols;
info[i].lhs_start = r0;
info[i].lhs_length = actualBlockRows;
if(transpose) func(c0, actualBlockCols, 0, rows, info);
else func(0, rows, c0, actualBlockCols, info);
}
delete[] info;
#endif
}
} // end namespace internal
} // end namespace Eigen
#endif // EIGEN_PARALLELIZER_H
|
