Re: [eigen] STL Vectors and Alignment

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

On 26.11.2014 17:39, Gabriel wrote:
> Thanks for the answer!
>
> Sorry again for my unclear phrasing!
> Just to be absolutely sure:
>
> Lets say we have:
>
>      struct A{
>           Eigen::Vector2d t;
>      };
>
> If we use A locally:
>      A a;
>      a.t
> is a.t  aligned on the Stack, (how does alignment work on the stack)

Yes, that is aligned on the stack. Also, if you wrote, e.g.
  struct A{
    char dummy;      // compiler will add filling bytes after this
    Eigen::Vector2d t;
  };
a.t will be aligned on the stack.
The data member of Vector2d has a compiler attribute which tells it to 
always have it aligned.

> If we use A on the heap:
>      A * a = new A;
> then as "a" is not aligned , also a.t is not aligned

It is not guaranteed to be aligned. It is, however, aligned on most 
(all?) 64bit systems.

> If we use EIGEN_MAKE_ALLOCATOR_NEW inside A , then "a" is aligned , and

Yes.

> BUT what happens in this example:
>
>      struct B{
>           A a;
>      };
>      struct C{
>           B b;
>      };
>
>      std::vector<   C  ,   Eigen::aligned_allocator<C>  > vec(1);
>
> Is      vec[0].b.a.t      now  aligned in memory? which does *NOT*
> result in a segfault if Eigen uses SIMD internally?

Yes.

> If we use locally lets say:
>
>      C c;
>      c.b.a.t  do something with t
>
> is c.b.a.t  aligned on the Stack?

Yes.

> Or on the heap:
>
>      C * c = new C;
>
> c might not be aligned => c.b  might not be aligned  => c.b.a => might
> not bealigned  => c.b.a.t might not be aligned  => SEGFAULT for usage of
> c.b.a.t

Correct. c is not guaranteed to be aligned (except on most 64 bit 
systems ...)

> How do I need to correct the above example such that in the cases above,
> vector A::t is always aligned to 16bytes?

To make the only not yet working case work, you need to add an
  EIGEN_MAKE_ALIGNED_OPERATOR_NEW
into struct C. It would not hurt to add it into A and B, as well.


Christoph

--
----------------------------------------------
Dipl.-Inf., Dipl.-Math. Christoph Hertzberg
Cartesium 0.049
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

Tel: +49 (421) 218-64252
----------------------------------------------