cart-elc

Source code for CART-ELC
git clone git://git.laack.co/cart-elc.git
Log | Files | Refs | README | LICENSE

StructHavingEigenMembers.dox (7026B)


      1 namespace Eigen {
      2 
      3 /** \eigenManualPage TopicStructHavingEigenMembers Structures Having Eigen Members
      4 
      5 \eigenAutoToc
      6 
      7 \section StructHavingEigenMembers_summary Executive Summary
      8 
      9 
     10 If you define a structure having members of \ref TopicFixedSizeVectorizable "fixed-size vectorizable Eigen types", you must ensure that calling operator new on it allocates properly aligned buffers.
     11 If you're compiling in \cpp17 mode only with a sufficiently recent compiler (e.g., GCC>=7, clang>=5, MSVC>=19.12), then everything is taken care by the compiler and you can stop reading.
     12 
     13 Otherwise, you have to overload its `operator new` so that it generates properly aligned pointers (e.g., 32-bytes-aligned for Vector4d and AVX).
     14 Fortunately, %Eigen provides you with a macro `EIGEN_MAKE_ALIGNED_OPERATOR_NEW` that does that for you.
     15 
     16 \section StructHavingEigenMembers_what What kind of code needs to be changed?
     17 
     18 The kind of code that needs to be changed is this:
     19 
     20 \code
     21 class Foo
     22 {
     23   ...
     24   Eigen::Vector2d v;
     25   ...
     26 };
     27 
     28 ...
     29 
     30 Foo *foo = new Foo;
     31 \endcode
     32 
     33 In other words: you have a class that has as a member a \ref TopicFixedSizeVectorizable "fixed-size vectorizable Eigen object", and then you dynamically create an object of that class.
     34 
     35 \section StructHavingEigenMembers_how How should such code be modified?
     36 
     37 Very easy, you just need to put a `EIGEN_MAKE_ALIGNED_OPERATOR_NEW` macro in a public part of your class, like this:
     38 
     39 \code
     40 class Foo
     41 {
     42   ...
     43   Eigen::Vector4d v;
     44   ...
     45 public:
     46   EIGEN_MAKE_ALIGNED_OPERATOR_NEW
     47 };
     48 
     49 ...
     50 
     51 Foo *foo = new Foo;
     52 \endcode
     53 
     54 This macro makes `new Foo` always return an aligned pointer.
     55 
     56 In \cpp17, this macro is empty.
     57 
     58 If this approach is too intrusive, see also the \ref StructHavingEigenMembers_othersolutions "other solutions".
     59 
     60 \section StructHavingEigenMembers_why Why is this needed?
     61 
     62 OK let's say that your code looks like this:
     63 
     64 \code
     65 class Foo
     66 {
     67   ...
     68   Eigen::Vector4d v;
     69   ...
     70 };
     71 
     72 ...
     73 
     74 Foo *foo = new Foo;
     75 \endcode
     76 
     77 A Eigen::Vector4d consists of 4 doubles, which is 256 bits.
     78 This is exactly the size of an AVX register, which makes it possible to use AVX for all sorts of operations on this vector.
     79 But AVX instructions (at least the ones that %Eigen uses, which are the fast ones) require 256-bit alignment.
     80 Otherwise you get a segmentation fault.
     81 
     82 For this reason, %Eigen takes care by itself to require 256-bit alignment for Eigen::Vector4d, by doing two things:
     83 \li %Eigen requires 256-bit alignment for the Eigen::Vector4d's array (of 4 doubles). With \cpp11 this is done with the <a href="https://en.cppreference.com/w/cpp/keyword/alignas">alignas</a> keyword, or compiler's extensions for c++98/03.
     84 \li %Eigen overloads the `operator new` of Eigen::Vector4d so it will always return 256-bit aligned pointers. (removed in \cpp17)
     85 
     86 Thus, normally, you don't have to worry about anything, %Eigen handles alignment of operator new for you...
     87 
     88 ... except in one case. When you have a `class Foo` like above, and you dynamically allocate a new `Foo` as above, then, since `Foo` doesn't have aligned `operator new`, the returned pointer foo is not necessarily 256-bit aligned.
     89 
     90 The alignment attribute of the member `v` is then relative to the start of the class `Foo`. If the `foo` pointer wasn't aligned, then `foo->v` won't be aligned either!
     91 
     92 The solution is to let `class Foo` have an aligned `operator new`, as we showed in the previous section.
     93 
     94 This explanation also holds for SSE/NEON/MSA/Altivec/VSX targets, which require 16-bytes alignment, and AVX512 which requires 64-bytes alignment for fixed-size objects multiple of 64 bytes (e.g., Eigen::Matrix4d).
     95 
     96 \section StructHavingEigenMembers_movetotop Should I then put all the members of Eigen types at the beginning of my class?
     97 
     98 That's not required. Since %Eigen takes care of declaring adequate alignment, all members that need it are automatically aligned relatively to the class. So code like this works fine:
     99 
    100 \code
    101 class Foo
    102 {
    103   double x;
    104   Eigen::Vector4d v;
    105 public:
    106   EIGEN_MAKE_ALIGNED_OPERATOR_NEW
    107 };
    108 \endcode
    109 
    110 That said, as usual, it is recommended to sort the members so that alignment does not waste memory.
    111 In the above example, with AVX, the compiler will have to reserve 24 empty bytes between `x` and `v`.
    112 
    113 
    114 \section StructHavingEigenMembers_dynamicsize What about dynamic-size matrices and vectors?
    115 
    116 Dynamic-size matrices and vectors, such as Eigen::VectorXd, allocate dynamically their own array of coefficients, so they take care of requiring absolute alignment automatically. So they don't cause this issue. The issue discussed here is only with \ref TopicFixedSizeVectorizable  "fixed-size vectorizable matrices and vectors".
    117 
    118 
    119 \section StructHavingEigenMembers_bugineigen So is this a bug in Eigen?
    120 
    121 No, it's not our bug. It's more like an inherent problem of the c++ language specification that has been solved in c++17 through the feature known as <a href="http://wg21.link/p0035r4">dynamic memory allocation for over-aligned data</a>.
    122 
    123 
    124 \section StructHavingEigenMembers_conditional What if I want to do this conditionally (depending on template parameters) ?
    125 
    126 For this situation, we offer the macro `EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)`.
    127 It will generate aligned operators like `EIGEN_MAKE_ALIGNED_OPERATOR_NEW` if `NeedsToAlign` is true.
    128 It will generate operators with the default alignment if `NeedsToAlign` is false.
    129 In \cpp17, this macro is empty.
    130 
    131 Example:
    132 
    133 \code
    134 template<int n> class Foo
    135 {
    136   typedef Eigen::Matrix<float,n,1> Vector;
    137   enum { NeedsToAlign = (sizeof(Vector)%16)==0 };
    138   ...
    139   Vector v;
    140   ...
    141 public:
    142   EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
    143 };
    144 
    145 ...
    146 
    147 Foo<4> *foo4 = new Foo<4>; // foo4 is guaranteed to be 128bit-aligned
    148 Foo<3> *foo3 = new Foo<3>; // foo3 has only the system default alignment guarantee
    149 \endcode
    150 
    151 
    152 \section StructHavingEigenMembers_othersolutions Other solutions
    153 
    154 In case putting the `EIGEN_MAKE_ALIGNED_OPERATOR_NEW` macro everywhere is too intrusive, there exists at least two other solutions.
    155 
    156 \subsection othersolutions1 Disabling alignment
    157 
    158 The first is to disable alignment requirement for the fixed size members:
    159 \code
    160 class Foo
    161 {
    162   ...
    163   Eigen::Matrix<double,4,1,Eigen::DontAlign> v;
    164   ...
    165 };
    166 \endcode
    167 This `v` is fully compatible with aligned Eigen::Vector4d.
    168 This has only for effect to make load/stores to `v` more expensive (usually slightly, but that's hardware dependent).
    169 
    170 
    171 \subsection othersolutions2 Private structure
    172 
    173 The second consist in storing the fixed-size objects into a private struct which will be dynamically allocated at the construction time of the main object:
    174 
    175 \code
    176 struct Foo_d
    177 {
    178   EIGEN_MAKE_ALIGNED_OPERATOR_NEW
    179   Vector4d v;
    180   ...
    181 };
    182 
    183 
    184 struct Foo {
    185   Foo() { init_d(); }
    186   ~Foo() { delete d; }
    187   void bar()
    188   {
    189     // use d->v instead of v
    190     ...
    191   }
    192 private:
    193   void init_d() { d = new Foo_d; }
    194   Foo_d* d;
    195 };
    196 \endcode
    197 
    198 The clear advantage here is that the class `Foo` remains unchanged regarding alignment issues.
    199 The drawback is that an additional heap allocation will be required whatsoever.
    200 
    201 */
    202 
    203 }