Re: [eigen] Why no LU::solveInPlace()?

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

On 17.08.2010 09:36, Gael Guennebaud wrote:
> We definitely have to uniformize that, but actually we could also
> consider removing all the solveInPlace functions...
>
> Indeed, when you write:
>
> X = A.lu().solve(X);
>
> Eigen automatically performs the solve operation inplace, just like it
> would do with :
>
> A.lu().solveInPlace(X);
>
> Advantages:
> - simpler API: only one function
> - works fine with writable temporary objects, e.g., X.col(j) =
> A.lu().solve(X.col(j));
>
> Drawbacks:
> - need to write the "source" and "destination" object twice
> - requires one runtime test (one pointer comparison which is disabled
> at compile time if the types do not match).

I really hope that this comparison will be optimized away in most cases 
by any recent compiler ...

> Recall that to make the solveInPlace() function to work with writable
> temporary objects we either have to declare the argument as a const
> reference or enforce the user to name the temporary that can be a real
> pain without the C++1x auto keyword.
>
> What other people think about it?

I think the const correctness issues are somewhat related to something I 
complained about last month (2nd issue of that mail):

http://listengine.tuxfamily.org/lists.tuxfamily.org/eigen/2010/07/msg00095.html

I think the clean way would be to pass some wrapper as const reference, 
which internally holds a non-const reference/pointer.
Actually the only objects which can be passed to solveInPlace are 
direct-access types (unless I'm wrong) so technically you only need to 
pass a pointer to the start of the data together with the sizes and 
strides (where the latter can be templated).

Is someone currently working on the 'Direct-access type 
unification'-TODO? (or on my issue from above's mail?)
http://eigen.tuxfamily.org/index.php?title=Todo_for_3.0

I'd volunteer to work out some suggestions/concepts, but currently I am 
quite busy and you couldn't expect something before mid-September from me :(

Christoph


--
----------------------------------------------
Dipl.-Inf. Christoph Hertzberg
Cartesium 0.051
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

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