aboutsummaryrefslogtreecommitdiffhomepage
path: root/doc/I11_Aliasing.dox
blob: 3f7e849bde13ec77b49c4f3dc6d70561264d58df (plain)
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
namespace Eigen {

/** \page TopicAliasing Aliasing

In Eigen, aliasing refers to assignment statement in which the same matrix (or array or vector) appears on the
left and on the right of the assignment operators. Statements like <tt>mat = 2 * mat;</tt> or <tt>mat =
mat.transpose();</tt> exhibit aliasing. The aliasing in the first example is harmless, but the aliasing in the
second example leads to unexpected results. This page explains what aliasing is, when it is harmful, and what
to do about it.

\tableofcontents


\section TopicAliasingExamples Examples

Here is a simple example exhibiting aliasing:

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include TopicAliasing_block.cpp
</td>
<td>
\verbinclude TopicAliasing_block.out
</td></tr></table>

The output is not what one would expect. The problem is the assignment
\code
mat.bottomRightCorner(2,2) = mat.topLeftCorner(2,2);
\endcode
This assignment exhibits aliasing: the coefficient \c mat(1,1) appears both in the block
<tt>mat.bottomRightCorner(2,2)</tt> on the left-hand side of the assignment and the block
<tt>mat.topLeftCorner(2,2)</tt> on the right-hand side. After the assignment, the (2,2) entry in the bottom
right corner should have the value of \c mat(1,1) before the assignment, which is 5. However, the output shows
that \c mat(2,2) is actually 1. The problem is that Eigen uses lazy evaluation (see 
\ref TopicEigenExpressionTemplates) for <tt>mat.topLeftCorner(2,2)</tt>. The result is similar to
\code
mat(1,1) = mat(0,0);
mat(1,2) = mat(0,1);
mat(2,1) = mat(1,0);
mat(2,2) = mat(1,1);
\endcode
Thus, \c mat(2,2) is assigned the \e new value of \c mat(1,1) instead of the old value. The next section
explains how to solve this problem by calling \link DenseBase::eval() eval()\endlink.

Note that if \c mat were a bigger, then the blocks would not overlap, and there would be no aliasing
problem. This means that in general aliasing cannot be detected at compile time. However, Eigen does detect
some instances of aliasing, albeit at run time.  The following example exhibiting aliasing was mentioned in
\ref TutorialMatrixArithmetic :

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include tut_arithmetic_transpose_aliasing.cpp
</td>
<td>
\verbinclude tut_arithmetic_transpose_aliasing.out
</td></tr></table>

Again, the output shows the aliasing issue. However, by default Eigen uses a run-time assertion to detect this
and exits with a message like

\verbatim
void Eigen::DenseBase<Derived>::checkTransposeAliasing(const OtherDerived&) const 
[with OtherDerived = Eigen::Transpose<Eigen::Matrix<int, 2, 2, 0, 2, 2> >, Derived = Eigen::Matrix<int, 2, 2, 0, 2, 2>]: 
Assertion `(!internal::check_transpose_aliasing_selector<Scalar,internal::blas_traits<Derived>::IsTransposed,OtherDerived>::run(internal::extract_data(derived()), other)) 
&& "aliasing detected during tranposition, use transposeInPlace() or evaluate the rhs into a temporary using .eval()"' failed.
\endverbatim

The user can turn Eigen's run-time assertions like the one to detect this aliasing problem off by defining the
EIGEN_NO_DEBUG macro, and the above program was compiled with this macro turned off in order to illustrate the
aliasing problem. See \ref TopicAssertions for more information about Eigen's run-time assertions.


\section TopicAliasingSolution Resolving aliasing issues

If you understand the cause of the aliasing issue, then it is obvious what must happen to solve it: Eigen has
to evaluate the right-hand side fully into a temporary matrix/array and then assign it to the left-hand
side. The function \link DenseBase::eval() eval() \endlink does precisely that.

For example, here is the corrected version of the first example above:

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include TopicAliasing_block_correct.cpp
</td>
<td>
\verbinclude TopicAliasing_block_correct.out
</td></tr></table>

Now, \c mat(2,2) equals 5 after the assignment, as it should be.

The same solution also works for the second example, with the transpose: simply replace the line 
<tt>a = a.transpose();</tt> with <tt>a = a.transpose().eval();</tt>. However, in this common case there is a
better solution. Eigen provides the special-purpose function 
\link DenseBase::transposeInPlace() transposeInPlace() \endlink which replaces a matrix by its transpose. 
This is shown below:

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include tut_arithmetic_transpose_inplace.cpp
</td>
<td>
\verbinclude tut_arithmetic_transpose_inplace.out
</td></tr></table>

If an xxxInPlace() function is available, then it is best to use it, because it indicates more clearly what you
are doing. This may also allow Eigen to optimize more aggressively. These are some of the xxxInPlace()
functions provided: 

<table class="manual">
<tr><th>Original function</th><th>In-place function</th></tr>
<tr> <td> MatrixBase::adjoint() </td> <td> MatrixBase::adjointInPlace() </td> </tr>
<tr class="alt"> <td> DenseBase::reverse() </td> <td> DenseBase::reverseInPlace() </td> </tr>
<tr> <td> LDLT::solve() </td> <td> LDLT::solveInPlace() </td> </tr>
<tr class="alt"> <td> LLT::solve() </td> <td> LLT::solveInPlace() </td> </tr>
<tr> <td> TriangularView::solve() </td> <td> TriangularView::solveInPlace() </td> </tr>
<tr class="alt"> <td> DenseBase::transpose() </td> <td> DenseBase::transposeInPlace() </td> </tr>
</table>


\section TopicAliasingCwise Aliasing and component-wise operations

As explained above, it may be dangerous if the same matrix or array occurs on both the left-hand side and the
right-hand side of an assignment operator, and it is then often necessary to evaluate the right-hand side
explicitly. However, applying component-wise operations (such as matrix addition, scalar multiplication and
array multiplication) is safe. 

The following example has only component-wise operations. Thus, there is no need for .eval() even though
the same matrix appears on both sides of the assignments.

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include TopicAliasing_cwise.cpp
</td>
<td>
\verbinclude TopicAliasing_cwise.out
</td></tr></table>

In general, an assignment is safe if the (i,j) entry of the expression on the right-hand side depends only on
the (i,j) entry of the matrix or array on the left-hand side and not on any other entries. In that case it is
not necessary to evaluate the right-hand side explicitly.


\section TopicAliasingMatrixMult Aliasing and matrix multiplication

Matrix multiplication is the only operation in Eigen that assumes aliasing by default. Thus, if \c matA is a
matrix, then the statement <tt>matA = matA * matA;</tt> is safe. All other operations in Eigen assume that
there are no aliasing problems, either because the result is assigned to a different matrix or because it is a
component-wise operation.

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include TopicAliasing_mult1.cpp
</td>
<td>
\verbinclude TopicAliasing_mult1.out
</td></tr></table>

However, this comes at a price. When executing the expression <tt>matA = matA * matA</tt>, Eigen evaluates the
product in a temporary matrix which is assigned to \c matA after the computation. This is fine. But Eigen does
the same when the product is assigned to a different matrix (e.g., <tt>matB = matA * matA</tt>). In that case,
it is more efficient to evaluate the product directly into \c matB instead of evaluating it first into a
temporary matrix and copying that matrix to \c matB.

The user can indicate with the \link MatrixBase::noalias() noalias()\endlink function that there is no
aliasing, as follows: <tt>matB.noalias() = matA * matA</tt>. This allows Eigen to evaluate the matrix product
<tt>matA * matA</tt> directly into \c matB.

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include TopicAliasing_mult2.cpp
</td>
<td>
\verbinclude TopicAliasing_mult2.out
</td></tr></table>

Of course, you should not use \c noalias() when there is in fact aliasing taking place. If you do, then you
may get wrong results:

<table class="example">
<tr><th>Example</th><th>Output</th></tr>
<tr><td>
\include TopicAliasing_mult3.cpp
</td>
<td>
\verbinclude TopicAliasing_mult3.out
</td></tr></table>


\section TopicAliasingSummary Summary

Aliasing occurs when the same matrix or array coefficients appear both on the left- and the right-hand side of
an assignment operator.
 - Aliasing is harmless with coefficient-wise computations; this includes scalar multiplication and matrix or
   array addition.
 - When you multiply two matrices, Eigen assumes that aliasing occurs. If you know that there is no aliasing,
   then you can use \link MatrixBase::noalias() noalias()\endlink.
 - In all other situations, Eigen assumes that there is no aliasing issue and thus gives the wrong result if
   aliasing does in fact occur. To prevent this, you have to use \link DenseBase::eval() eval() \endlink or
   one of the xxxInPlace() functions.

*/
}