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
|
// 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/.
#define TEST_ENABLE_TEMPORARY_TRACKING
#include "main.h"
template<typename Dst, typename Lhs, typename Rhs>
void check_scalar_multiple3(Dst &dst, const Lhs& A, const Rhs& B)
{
VERIFY_EVALUATION_COUNT( (dst.noalias() = A * B), 0);
VERIFY_IS_APPROX( dst, (A.eval() * B.eval()).eval() );
VERIFY_EVALUATION_COUNT( (dst.noalias() += A * B), 0);
VERIFY_IS_APPROX( dst, 2*(A.eval() * B.eval()).eval() );
VERIFY_EVALUATION_COUNT( (dst.noalias() -= A * B), 0);
VERIFY_IS_APPROX( dst, (A.eval() * B.eval()).eval() );
}
template<typename Dst, typename Lhs, typename Rhs, typename S2>
void check_scalar_multiple2(Dst &dst, const Lhs& A, const Rhs& B, S2 s2)
{
CALL_SUBTEST( check_scalar_multiple3(dst, A, B) );
CALL_SUBTEST( check_scalar_multiple3(dst, A, -B) );
CALL_SUBTEST( check_scalar_multiple3(dst, A, s2*B) );
CALL_SUBTEST( check_scalar_multiple3(dst, A, B*s2) );
CALL_SUBTEST( check_scalar_multiple3(dst, A, (B*s2).conjugate()) );
}
template<typename Dst, typename Lhs, typename Rhs, typename S1, typename S2>
void check_scalar_multiple1(Dst &dst, const Lhs& A, const Rhs& B, S1 s1, S2 s2)
{
CALL_SUBTEST( check_scalar_multiple2(dst, A, B, s2) );
CALL_SUBTEST( check_scalar_multiple2(dst, -A, B, s2) );
CALL_SUBTEST( check_scalar_multiple2(dst, s1*A, B, s2) );
CALL_SUBTEST( check_scalar_multiple2(dst, A*s1, B, s2) );
CALL_SUBTEST( check_scalar_multiple2(dst, (A*s1).conjugate(), B, s2) );
}
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::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 = (m1 * m2.adjoint()).transpose(), 1);
VERIFY_EVALUATION_COUNT( m3.noalias() = m1 * m2.adjoint(), 0);
VERIFY_EVALUATION_COUNT( m3 = s1 * (m1 * m2.transpose()), 1);
// VERIFY_EVALUATION_COUNT( m3 = m3 + s1 * (m1 * m2.transpose()), 1);
VERIFY_EVALUATION_COUNT( m3.noalias() = s1 * (m1 * m2.transpose()), 0);
VERIFY_EVALUATION_COUNT( m3 = m3 + (m1 * m2.adjoint()), 1);
VERIFY_EVALUATION_COUNT( m3 = m3 - (m1 * m2.adjoint()), 1);
VERIFY_EVALUATION_COUNT( m3 = m3 + (m1 * m2.adjoint()).transpose(), 1);
VERIFY_EVALUATION_COUNT( m3.noalias() = m3 + m1 * m2.transpose(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() += m3 + m1 * m2.transpose(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() -= m3 + m1 * m2.transpose(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() = m3 - m1 * m2.transpose(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() += m3 - m1 * m2.transpose(), 0);
VERIFY_EVALUATION_COUNT( m3.noalias() -= m3 - 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);
VERIFY_EVALUATION_COUNT( m3.template triangularView<Upper>() = (m1 * m2.adjoint()), 0);
VERIFY_EVALUATION_COUNT( m3.template triangularView<Upper>() -= (m1 * 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 ...
m3.setRandom(rows,cols);
VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) / (m3.transpose().lazyProduct(m3)).diagonal().sum(), 0 );
VERIFY_EVALUATION_COUNT( m3.noalias() = m1.conjugate().lazyProduct(m2.conjugate()), 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
VERIFY_EVALUATION_COUNT( m3.col(0).template head<5>() * m3.col(0).transpose() + m3.col(0).template 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 );
VERIFY_EVALUATION_COUNT( cvres.noalias() = (m1+m1) * cv1, 0 );
VERIFY_EVALUATION_COUNT( cvres.noalias() = (rm3+rm3) * cv1, 0 );
VERIFY_EVALUATION_COUNT( cvres.noalias() = (m1+m1) * (m1*cv1), 1 );
VERIFY_EVALUATION_COUNT( cvres.noalias() = (rm3+rm3) * (m1*cv1), 1 );
// Check outer products
#ifdef EIGEN_ALLOCA
bool temp_via_alloca = m3.rows()*sizeof(Scalar) <= EIGEN_STACK_ALLOCATION_LIMIT;
#else
bool temp_via_alloca = false;
#endif
m3 = cv1 * rv1;
VERIFY_EVALUATION_COUNT( m3.noalias() = cv1 * rv1, 0 );
VERIFY_EVALUATION_COUNT( m3.noalias() = (cv1+cv1) * (rv1+rv1), temp_via_alloca ? 0 : 1 );
VERIFY_EVALUATION_COUNT( m3.noalias() = (m1*cv1) * (rv1), 1 );
VERIFY_EVALUATION_COUNT( m3.noalias() += (m1*cv1) * (rv1), 1 );
rm3 = cv1 * rv1;
VERIFY_EVALUATION_COUNT( rm3.noalias() = cv1 * rv1, 0 );
VERIFY_EVALUATION_COUNT( rm3.noalias() = (cv1+cv1) * (rv1+rv1), temp_via_alloca ? 0 : 1 );
VERIFY_EVALUATION_COUNT( rm3.noalias() = (cv1) * (rv1 * m1), 1 );
VERIFY_EVALUATION_COUNT( rm3.noalias() -= (cv1) * (rv1 * m1), 1 );
VERIFY_EVALUATION_COUNT( rm3.noalias() = (m1*cv1) * (rv1 * m1), 2 );
VERIFY_EVALUATION_COUNT( rm3.noalias() += (m1*cv1) * (rv1 * m1), 2 );
// Check nested products
VERIFY_EVALUATION_COUNT( cvres.noalias() = m1.adjoint() * m1 * cv1, 1 );
VERIFY_EVALUATION_COUNT( rvres.noalias() = rv1 * (m1 * m2.adjoint()), 1 );
// exhaustively check all scalar multiple combinations:
{
// Generic path:
check_scalar_multiple1(m3, m1, m2, s1, s2);
// Force fall back to coeff-based:
typename ColMajorMatrixType::BlockXpr m3_blck = m3.block(r0,r0,1,1);
check_scalar_multiple1(m3_blck, m1.block(r0,c0,1,1), m2.block(c0,r0,1,1), s1, s2);
}
}
EIGEN_DECLARE_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)) );
CALL_SUBTEST_2( product_notemporary(MatrixXd(s, s)) );
TEST_SET_BUT_UNUSED_VARIABLE(s)
s = internal::random<int>(16,EIGEN_TEST_MAX_SIZE/2);
CALL_SUBTEST_3( product_notemporary(MatrixXcf(s,s)) );
CALL_SUBTEST_4( product_notemporary(MatrixXcd(s,s)) );
TEST_SET_BUT_UNUSED_VARIABLE(s)
}
}
|
