W3cubDocs

/Eigen3

Structures Having Eigen Members

Executive Summary

If you define a structure having members of fixed-size vectorizable Eigen types, you must ensure that calling operator new on it allocates properly aligned buffers. If you're compiling in [c++17] 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.

Otherwise, you have to overload its operator new so that it generates properly aligned pointers (e.g., 32-bytes-aligned for Vector4d and AVX). Fortunately, Eigen provides you with a macro EIGEN_MAKE_ALIGNED_OPERATOR_NEW that does that for you.

What kind of code needs to be changed?

The kind of code that needs to be changed is this:

class Foo
{
  ...
  Eigen::Vector2d v;
  ...
};
 
...
 
Foo *foo = new Foo;

In other words: you have a class that has as a member a fixed-size vectorizable Eigen object, and then you dynamically create an object of that class.

How should such code be modified?

Very easy, you just need to put a EIGEN_MAKE_ALIGNED_OPERATOR_NEW macro in a public part of your class, like this:

class Foo
{
  ...
  Eigen::Vector4d v;
  ...
public:
  EIGEN_MAKE_ALIGNED_OPERATOR_NEW
};
 
...
 
Foo *foo = new Foo;

This macro makes new Foo always return an aligned pointer.

In [c++17], this macro is empty.

If this approach is too intrusive, see also the other solutions.

Why is this needed?

OK let's say that your code looks like this:

class Foo
{
  ...
  Eigen::Vector4d v;
  ...
};
 
...
 
Foo *foo = new Foo;

A Eigen::Vector4d consists of 4 doubles, which is 256 bits. This is exactly the size of an AVX register, which makes it possible to use AVX for all sorts of operations on this vector. But AVX instructions (at least the ones that Eigen uses, which are the fast ones) require 256-bit alignment. Otherwise you get a segmentation fault.

For this reason, Eigen takes care by itself to require 256-bit alignment for Eigen::Vector4d, by doing two things:

  • Eigen requires 256-bit alignment for the Eigen::Vector4d's array (of 4 doubles). With [c++11] this is done with the alignas keyword, or compiler's extensions for c++98/03.
  • Eigen overloads the operator new of Eigen::Vector4d so it will always return 256-bit aligned pointers. (removed in [c++17])

Thus, normally, you don't have to worry about anything, Eigen handles alignment of operator new for you...

... 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.

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!

The solution is to let class Foo have an aligned operator new, as we showed in the previous section.

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).

Should I then put all the members of Eigen types at the beginning of my class?

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:

class Foo
{
  double x;
  Eigen::Vector4d v;
public:
  EIGEN_MAKE_ALIGNED_OPERATOR_NEW
};

That said, as usual, it is recommended to sort the members so that alignment does not waste memory. In the above example, with AVX, the compiler will have to reserve 24 empty bytes between x and v.

What about dynamic-size matrices and vectors?

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 fixed-size vectorizable matrices and vectors.

So is this a bug in Eigen?

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 dynamic memory allocation for over-aligned data.

What if I want to do this conditionally (depending on template parameters) ?

For this situation, we offer the macro EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign). It will generate aligned operators like EIGEN_MAKE_ALIGNED_OPERATOR_NEW if NeedsToAlign is true. It will generate operators with the default alignment if NeedsToAlign is false. In [c++17], this macro is empty.

Example:

template<int n> class Foo
{
  typedef Eigen::Matrix<float,n,1> Vector;
  enum { NeedsToAlign = (sizeof(Vector)%16)==0 };
  ...
  Vector v;
  ...
public:
  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
};
 
...
 
Foo<4> *foo4 = new Foo<4>; // foo4 is guaranteed to be 128bit-aligned
Foo<3> *foo3 = new Foo<3>; // foo3 has only the system default alignment guarantee

Other solutions

In case putting the EIGEN_MAKE_ALIGNED_OPERATOR_NEW macro everywhere is too intrusive, there exists at least two other solutions.

Disabling alignment

The first is to disable alignment requirement for the fixed size members:

class Foo
{
  ...
  Eigen::Matrix<double,4,1,Eigen::DontAlign> v;
  ...
};

This v is fully compatible with aligned Eigen::Vector4d. This has only for effect to make load/stores to v more expensive (usually slightly, but that's hardware dependent).

Private structure

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:

struct Foo_d
{
  EIGEN_MAKE_ALIGNED_OPERATOR_NEW
  Vector4d v;
  ...
};
 
 
struct Foo {
  Foo() { init_d(); }
  ~Foo() { delete d; }
  void bar()
  {
    // use d->v instead of v
    ...
  }
private:
  void init_d() { d = new Foo_d; }
  Foo_d* d;
};

The clear advantage here is that the class Foo remains unchanged regarding alignment issues. The drawback is that an additional heap allocation will be required whatsoever.