Home | History | Annotate | Download | only in doc
      1 namespace Eigen {
      2 
      3 /** \eigenManualPage TopicAliasing Aliasing
      4 
      5 In %Eigen, aliasing refers to assignment statement in which the same matrix (or array or vector) appears on the
      6 left and on the right of the assignment operators. Statements like <tt>mat = 2 * mat;</tt> or <tt>mat =
      7 mat.transpose();</tt> exhibit aliasing. The aliasing in the first example is harmless, but the aliasing in the
      8 second example leads to unexpected results. This page explains what aliasing is, when it is harmful, and what
      9 to do about it.
     10 
     11 \eigenAutoToc
     12 
     13 
     14 \section TopicAliasingExamples Examples
     15 
     16 Here is a simple example exhibiting aliasing:
     17 
     18 <table class="example">
     19 <tr><th>Example</th><th>Output</th></tr>
     20 <tr><td>
     21 \include TopicAliasing_block.cpp
     22 </td>
     23 <td>
     24 \verbinclude TopicAliasing_block.out
     25 </td></tr></table>
     26 
     27 The output is not what one would expect. The problem is the assignment
     28 \code
     29 mat.bottomRightCorner(2,2) = mat.topLeftCorner(2,2);
     30 \endcode
     31 This assignment exhibits aliasing: the coefficient \c mat(1,1) appears both in the block
     32 <tt>mat.bottomRightCorner(2,2)</tt> on the left-hand side of the assignment and the block
     33 <tt>mat.topLeftCorner(2,2)</tt> on the right-hand side. After the assignment, the (2,2) entry in the bottom
     34 right corner should have the value of \c mat(1,1) before the assignment, which is 5. However, the output shows
     35 that \c mat(2,2) is actually 1. The problem is that %Eigen uses lazy evaluation (see 
     36 \ref TopicEigenExpressionTemplates) for <tt>mat.topLeftCorner(2,2)</tt>. The result is similar to
     37 \code
     38 mat(1,1) = mat(0,0);
     39 mat(1,2) = mat(0,1);
     40 mat(2,1) = mat(1,0);
     41 mat(2,2) = mat(1,1);
     42 \endcode
     43 Thus, \c mat(2,2) is assigned the \e new value of \c mat(1,1) instead of the old value. The next section
     44 explains how to solve this problem by calling \link DenseBase::eval() eval()\endlink.
     45 
     46 Aliasing occurs more naturally when trying to shrink a matrix. For example, the expressions <tt>vec =
     47 vec.head(n)</tt> and <tt>mat = mat.block(i,j,r,c)</tt> exhibit aliasing.
     48 
     49 In general, aliasing cannot be detected at compile time: if \c mat in the first example were a bit bigger,
     50 then the blocks would not overlap, and there would be no aliasing problem. However, %Eigen does detect some
     51 instances of aliasing, albeit at run time.  The following example exhibiting aliasing was mentioned in \ref
     52 TutorialMatrixArithmetic :
     53 
     54 <table class="example">
     55 <tr><th>Example</th><th>Output</th></tr>
     56 <tr><td>
     57 \include tut_arithmetic_transpose_aliasing.cpp
     58 </td>
     59 <td>
     60 \verbinclude tut_arithmetic_transpose_aliasing.out
     61 </td></tr></table>
     62 
     63 Again, the output shows the aliasing issue. However, by default %Eigen uses a run-time assertion to detect this
     64 and exits with a message like
     65 
     66 \verbatim
     67 void Eigen::DenseBase<Derived>::checkTransposeAliasing(const OtherDerived&) const 
     68 [with OtherDerived = Eigen::Transpose<Eigen::Matrix<int, 2, 2, 0, 2, 2> >, Derived = Eigen::Matrix<int, 2, 2, 0, 2, 2>]: 
     69 Assertion `(!internal::check_transpose_aliasing_selector<Scalar,internal::blas_traits<Derived>::IsTransposed,OtherDerived>::run(internal::extract_data(derived()), other)) 
     70 && "aliasing detected during transposition, use transposeInPlace() or evaluate the rhs into a temporary using .eval()"' failed.
     71 \endverbatim
     72 
     73 The user can turn %Eigen's run-time assertions like the one to detect this aliasing problem off by defining the
     74 EIGEN_NO_DEBUG macro, and the above program was compiled with this macro turned off in order to illustrate the
     75 aliasing problem. See \ref TopicAssertions for more information about %Eigen's run-time assertions.
     76 
     77 
     78 \section TopicAliasingSolution Resolving aliasing issues
     79 
     80 If you understand the cause of the aliasing issue, then it is obvious what must happen to solve it: %Eigen has
     81 to evaluate the right-hand side fully into a temporary matrix/array and then assign it to the left-hand
     82 side. The function \link DenseBase::eval() eval() \endlink does precisely that.
     83 
     84 For example, here is the corrected version of the first example above:
     85 
     86 <table class="example">
     87 <tr><th>Example</th><th>Output</th></tr>
     88 <tr><td>
     89 \include TopicAliasing_block_correct.cpp
     90 </td>
     91 <td>
     92 \verbinclude TopicAliasing_block_correct.out
     93 </td></tr></table>
     94 
     95 Now, \c mat(2,2) equals 5 after the assignment, as it should be.
     96 
     97 The same solution also works for the second example, with the transpose: simply replace the line 
     98 <tt>a = a.transpose();</tt> with <tt>a = a.transpose().eval();</tt>. However, in this common case there is a
     99 better solution. %Eigen provides the special-purpose function 
    100 \link DenseBase::transposeInPlace() transposeInPlace() \endlink which replaces a matrix by its transpose. 
    101 This is shown below:
    102 
    103 <table class="example">
    104 <tr><th>Example</th><th>Output</th></tr>
    105 <tr><td>
    106 \include tut_arithmetic_transpose_inplace.cpp
    107 </td>
    108 <td>
    109 \verbinclude tut_arithmetic_transpose_inplace.out
    110 </td></tr></table>
    111 
    112 If an xxxInPlace() function is available, then it is best to use it, because it indicates more clearly what you
    113 are doing. This may also allow %Eigen to optimize more aggressively. These are some of the xxxInPlace()
    114 functions provided: 
    115 
    116 <table class="manual">
    117 <tr><th>Original function</th><th>In-place function</th></tr>
    118 <tr> <td> MatrixBase::adjoint() </td> <td> MatrixBase::adjointInPlace() </td> </tr>
    119 <tr class="alt"> <td> DenseBase::reverse() </td> <td> DenseBase::reverseInPlace() </td> </tr>
    120 <tr> <td> LDLT::solve() </td> <td> LDLT::solveInPlace() </td> </tr>
    121 <tr class="alt"> <td> LLT::solve() </td> <td> LLT::solveInPlace() </td> </tr>
    122 <tr> <td> TriangularView::solve() </td> <td> TriangularView::solveInPlace() </td> </tr>
    123 <tr class="alt"> <td> DenseBase::transpose() </td> <td> DenseBase::transposeInPlace() </td> </tr>
    124 </table>
    125 
    126 In the special case where a matrix or vector is shrunk using an expression like <tt>vec = vec.head(n)</tt>,
    127 you can use \link PlainObjectBase::conservativeResize() conservativeResize() \endlink.
    128 
    129 
    130 \section TopicAliasingCwise Aliasing and component-wise operations
    131 
    132 As explained above, it may be dangerous if the same matrix or array occurs on both the left-hand side and the
    133 right-hand side of an assignment operator, and it is then often necessary to evaluate the right-hand side
    134 explicitly. However, applying component-wise operations (such as matrix addition, scalar multiplication and
    135 array multiplication) is safe. 
    136 
    137 The following example has only component-wise operations. Thus, there is no need for \link DenseBase::eval()
    138 eval() \endlink even though the same matrix appears on both sides of the assignments.
    139 
    140 <table class="example">
    141 <tr><th>Example</th><th>Output</th></tr>
    142 <tr><td>
    143 \include TopicAliasing_cwise.cpp
    144 </td>
    145 <td>
    146 \verbinclude TopicAliasing_cwise.out
    147 </td></tr></table>
    148 
    149 In general, an assignment is safe if the (i,j) entry of the expression on the right-hand side depends only on
    150 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
    151 not necessary to evaluate the right-hand side explicitly.
    152 
    153 
    154 \section TopicAliasingMatrixMult Aliasing and matrix multiplication
    155 
    156 Matrix multiplication is the only operation in %Eigen that assumes aliasing by default, <strong>under the
    157 condition that the destination matrix is not resized</strong>.
    158 Thus, if \c matA is a \b squared matrix, then the statement <tt>matA = matA * matA;</tt> is safe.
    159 All other operations in %Eigen assume that there are no aliasing problems,
    160 either because the result is assigned to a different matrix or because it is a component-wise operation.
    161 
    162 <table class="example">
    163 <tr><th>Example</th><th>Output</th></tr>
    164 <tr><td>
    165 \include TopicAliasing_mult1.cpp
    166 </td>
    167 <td>
    168 \verbinclude TopicAliasing_mult1.out
    169 </td></tr></table>
    170 
    171 However, this comes at a price. When executing the expression <tt>matA = matA * matA</tt>, %Eigen evaluates the
    172 product in a temporary matrix which is assigned to \c matA after the computation. This is fine. But %Eigen does
    173 the same when the product is assigned to a different matrix (e.g., <tt>matB = matA * matA</tt>). In that case,
    174 it is more efficient to evaluate the product directly into \c matB instead of evaluating it first into a
    175 temporary matrix and copying that matrix to \c matB.
    176 
    177 The user can indicate with the \link MatrixBase::noalias() noalias()\endlink function that there is no
    178 aliasing, as follows: <tt>matB.noalias() = matA * matA</tt>. This allows %Eigen to evaluate the matrix product
    179 <tt>matA * matA</tt> directly into \c matB.
    180 
    181 <table class="example">
    182 <tr><th>Example</th><th>Output</th></tr>
    183 <tr><td>
    184 \include TopicAliasing_mult2.cpp
    185 </td>
    186 <td>
    187 \verbinclude TopicAliasing_mult2.out
    188 </td></tr></table>
    189 
    190 Of course, you should not use \c noalias() when there is in fact aliasing taking place. If you do, then you
    191 may get wrong results:
    192 
    193 <table class="example">
    194 <tr><th>Example</th><th>Output</th></tr>
    195 <tr><td>
    196 \include TopicAliasing_mult3.cpp
    197 </td>
    198 <td>
    199 \verbinclude TopicAliasing_mult3.out
    200 </td></tr></table>
    201 
    202 Moreover, starting in Eigen 3.3, aliasing is \b not assumed if the destination matrix is resized and the product is not directly assigned to the destination.
    203 Therefore, the following example is also wrong:
    204 
    205 <table class="example">
    206 <tr><th>Example</th><th>Output</th></tr>
    207 <tr><td>
    208 \include TopicAliasing_mult4.cpp
    209 </td>
    210 <td>
    211 \verbinclude TopicAliasing_mult4.out
    212 </td></tr></table>
    213 
    214 As for any aliasing issue, you can resolve it by explicitly evaluating the expression prior to assignment:
    215 <table class="example">
    216 <tr><th>Example</th><th>Output</th></tr>
    217 <tr><td>
    218 \include TopicAliasing_mult5.cpp
    219 </td>
    220 <td>
    221 \verbinclude TopicAliasing_mult5.out
    222 </td></tr></table>
    223 
    224 \section TopicAliasingSummary Summary
    225 
    226 Aliasing occurs when the same matrix or array coefficients appear both on the left- and the right-hand side of
    227 an assignment operator.
    228  - Aliasing is harmless with coefficient-wise computations; this includes scalar multiplication and matrix or
    229    array addition.
    230  - When you multiply two matrices, %Eigen assumes that aliasing occurs. If you know that there is no aliasing,
    231    then you can use \link MatrixBase::noalias() noalias()\endlink.
    232  - In all other situations, %Eigen assumes that there is no aliasing issue and thus gives the wrong result if
    233    aliasing does in fact occur. To prevent this, you have to use \link DenseBase::eval() eval() \endlink or
    234    one of the xxxInPlace() functions.
    235 
    236 */
    237 }
    238