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
|
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
//
// 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/.
static int nb_temporaries;
void on_temporary_creation(int size) {
// here's a great place to set a breakpoint when debugging failures in this test!
if(size!=0) nb_temporaries++;
}
#define EIGEN_DENSE_STORAGE_CTOR_PLUGIN { on_temporary_creation(size); }
#include "main.h"
#define VERIFY_EVALUATION_COUNT(XPR,N) {\
nb_temporaries = 0; \
XPR; \
if(nb_temporaries!=N) std::cerr << "nb_temporaries == " << nb_temporaries << "\n"; \
VERIFY( (#XPR) && nb_temporaries==N ); \
}
template<typename MatrixType> void product_notemporary(const MatrixType& m)
{
/* This test checks the number of temporaries created
* during the evaluation of a complex expression */
typedef typename MatrixType::Index Index;
typedef typename MatrixType::Scalar Scalar;
typedef typename MatrixType::RealScalar RealScalar;
typedef Matrix<Scalar, 1, Dynamic> RowVectorType;
typedef Matrix<Scalar, Dynamic, 1> ColVectorType;
typedef Matrix<Scalar, Dynamic, Dynamic, ColMajor> ColMajorMatrixType;
typedef Matrix<Scalar, Dynamic, Dynamic, RowMajor> RowMajorMatrixType;
Index rows = m.rows();
Index cols = m.cols();
ColMajorMatrixType m1 = MatrixType::Random(rows, cols),
m2 = MatrixType::Random(rows, cols),
m3(rows, cols);
RowVectorType rv1 = RowVectorType::Random(rows), rvres(rows);
ColVectorType cv1 = ColVectorType::Random(cols), cvres(cols);
RowMajorMatrixType rm3(rows, cols);
Scalar s1 = internal::random<Scalar>(),
s2 = internal::random<Scalar>(),
s3 = internal::random<Scalar>();
Index c0 = internal::random<Index>(4,cols-8),
c1 = internal::random<Index>(8,cols-c0),
r0 = internal::random<Index>(4,cols-8),
r1 = internal::random<Index>(8,rows-r0);
VERIFY_EVALUATION_COUNT( m3 = (m1 * m2.adjoint()), 1);
VERIFY_EVALUATION_COUNT( m3.noalias() = m1 * m2.adjoint(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * (m1 * m2.transpose()), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * m1 * s2 * m2.adjoint(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * m1 * s2 * (m1*s3+m2*s2).adjoint(), 1);
VERIFY_EVALUATION_COUNT( m3.noalias() = (s1 * m1).adjoint() * s2 * m2, 0);
VERIFY_EVALUATION_COUNT( m3.noalias() += s1 * (-m1*s3).adjoint() * (s2 * m2 * s3), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() -= s1 * (m1.transpose() * m2), 0);
VERIFY_EVALUATION_COUNT(( m3.block(r0,r0,r1,r1).noalias() += -m1.block(r0,c0,r1,c1) * (s2*m2.block(r0,c0,r1,c1)).adjoint() ), 0);
VERIFY_EVALUATION_COUNT(( m3.block(r0,r0,r1,r1).noalias() -= s1 * m1.block(r0,c0,r1,c1) * m2.block(c0,r0,c1,r1) ), 0);
// NOTE this is because the Block expression is not handled yet by our expression analyser
VERIFY_EVALUATION_COUNT(( m3.block(r0,r0,r1,r1).noalias() = s1 * m1.block(r0,c0,r1,c1) * (s1*m2).block(c0,r0,c1,r1) ), 1);
VERIFY_EVALUATION_COUNT( m3.noalias() -= (s1 * m1).template triangularView<Lower>() * m2, 0);
VERIFY_EVALUATION_COUNT( rm3.noalias() = (s1 * m1.adjoint()).template triangularView<Upper>() * (m2+m2), 1);
VERIFY_EVALUATION_COUNT( rm3.noalias() = (s1 * m1.adjoint()).template triangularView<UnitUpper>() * m2.adjoint(), 0);
// NOTE this is because the blas_traits require innerstride==1 to avoid a temporary, but that doesn't seem to be actually needed for the triangular products
VERIFY_EVALUATION_COUNT( rm3.col(c0).noalias() = (s1 * m1.adjoint()).template triangularView<UnitUpper>() * (s2*m2.row(c0)).adjoint(), 1);
VERIFY_EVALUATION_COUNT( m1.template triangularView<Lower>().solveInPlace(m3), 0);
VERIFY_EVALUATION_COUNT( m1.adjoint().template triangularView<Lower>().solveInPlace(m3.transpose()), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() -= (s1 * m1).adjoint().template selfadjointView<Lower>() * (-m2*s3).adjoint(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() = s2 * m2.adjoint() * (s1 * m1.adjoint()).template selfadjointView<Upper>(), 0);
VERIFY_EVALUATION_COUNT( rm3.noalias() = (s1 * m1.adjoint()).template selfadjointView<Lower>() * m2.adjoint(), 0);
// NOTE this is because the blas_traits require innerstride==1 to avoid a temporary, but that doesn't seem to be actually needed for the triangular products
VERIFY_EVALUATION_COUNT( m3.col(c0).noalias() = (s1 * m1).adjoint().template selfadjointView<Lower>() * (-m2.row(c0)*s3).adjoint(), 1);
VERIFY_EVALUATION_COUNT( m3.col(c0).noalias() -= (s1 * m1).adjoint().template selfadjointView<Upper>() * (-m2.row(c0)*s3).adjoint(), 1);
VERIFY_EVALUATION_COUNT( m3.block(r0,c0,r1,c1).noalias() += m1.block(r0,r0,r1,r1).template selfadjointView<Upper>() * (s1*m2.block(r0,c0,r1,c1)), 0);
VERIFY_EVALUATION_COUNT( m3.block(r0,c0,r1,c1).noalias() = m1.block(r0,r0,r1,r1).template selfadjointView<Upper>() * m2.block(r0,c0,r1,c1), 0);
VERIFY_EVALUATION_COUNT( m3.template selfadjointView<Lower>().rankUpdate(m2.adjoint()), 0);
// Here we will get 1 temporary for each resize operation of the lhs operator; resize(r1,c1) would lead to zero temporaries
m3.resize(1,1);
VERIFY_EVALUATION_COUNT( m3.noalias() = m1.block(r0,r0,r1,r1).template selfadjointView<Lower>() * m2.block(r0,c0,r1,c1), 1);
m3.resize(1,1);
VERIFY_EVALUATION_COUNT( m3.noalias() = m1.block(r0,r0,r1,r1).template triangularView<UnitUpper>() * m2.block(r0,c0,r1,c1), 1);
// Zero temporaries for lazy products ...
VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) / (m3.transpose().lazyProduct(m3)).diagonal().sum(), 0 );
// ... and even no temporary for even deeply (>=2) nested products
VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) / (m3.transpose() * m3).diagonal().sum(), 0 );
VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) / (m3.transpose() * m3).diagonal().array().abs().sum(), 0 );
// Zero temporaries for ... CoeffBasedProductMode
// - does not work with GCC because of the <..>, we'ld need variadic macros ...
//VERIFY_EVALUATION_COUNT( m3.col(0).head<5>() * m3.col(0).transpose() + m3.col(0).head<5>() * m3.col(0).transpose(), 0 );
// Check matrix * vectors
VERIFY_EVALUATION_COUNT( cvres.noalias() = m1 * cv1, 0 );
VERIFY_EVALUATION_COUNT( cvres.noalias() -= m1 * cv1, 0 );
VERIFY_EVALUATION_COUNT( cvres.noalias() -= m1 * m2.col(0), 0 );
VERIFY_EVALUATION_COUNT( cvres.noalias() -= m1 * rv1.adjoint(), 0 );
VERIFY_EVALUATION_COUNT( cvres.noalias() -= m1 * m2.row(0).transpose(), 0 );
}
void test_product_notemporary()
{
int s;
for(int i = 0; i < g_repeat; i++) {
s = internal::random<int>(16,EIGEN_TEST_MAX_SIZE);
CALL_SUBTEST_1( product_notemporary(MatrixXf(s, s)) );
s = internal::random<int>(16,EIGEN_TEST_MAX_SIZE);
CALL_SUBTEST_2( product_notemporary(MatrixXd(s, s)) );
s = internal::random<int>(16,EIGEN_TEST_MAX_SIZE/2);
CALL_SUBTEST_3( product_notemporary(MatrixXcf(s,s)) );
s = internal::random<int>(16,EIGEN_TEST_MAX_SIZE/2);
CALL_SUBTEST_4( product_notemporary(MatrixXcd(s,s)) );
}
}
|