Re: [eigen] nesting

[ Thread Index | Date Index | More lists.tuxfamily.org/eigen Archives ]


On Fri, Feb 5, 2010 at 10:22 AM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:
> Let's first consider dynamic sized matrices only. For them we could create a
> Temporary class inheriting DenseMatrixBase which would behave like a shared
> pointer with automatic deletion. So instead of copying the whole matrix
> data, you would only copy the sizes, the pointer and increment a counter.
> Get it?

Funny since I've also been considering shared_ptrs for the product
base members...

> Now, of course this concept does not work with fixed size objetcs. For them
> you would have to nest them in a kind of shared-pointer-expression, and so
> create the matrix objects using operator new... that is even worse :(

.... but I did not see this problem.

> So unless someone see how to solve this latter issue, what Hauke is trying
> to do could work. But instead of specialzing ei_ref_selector for all
> expressions I would suggest a different approach using new bit Flags. The
> rules would be as follow:
> - if the expr has the NestByReference flag, then nest by reference,
> otherwise nest by value
> - by default, only Matrix and Array would have the NestByReference flag
> - the NestByReference flag is not inherited
> - an expression nesting a temporary would have the NestParentByReference
> flag
> - the NestParentByReference flag is not inherited, but if one of the
> children of an expression has the NestParentByReference flag, then the
> current expression will have the NestByReference flag. If we make sure that
> NestParentByReferenceBit == NestByReference>>1, then we simply have to do:
> Flags = ... | ((Lhs::Flags|Rhs::Flags)&NestParentByReference) << 1), or we
> could also use a macro: Flags = ... |
> EIGEN_PROPAGATE_NESTING_BIT(Lhs::Flags|Rhs::Flags).
>
> As an implication, we could get rid of ei_ref_selector. Everything will be
> done by ei_nesting and the ei_traits, so fewer template instanciations.
>
> To detect if a child generates a temporary, ei_nesting could expose an
> Evaluate boolean.

To me this sounds very good. If I understand this correctly, you don't
even need to touch every expression but only those nesting temporaries
and as you said, it will even reduce compile time.

- Hauke



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/