Re: [Lapackpp-devel] Interesting crash in QR solve

[Christian S. <sti...@tu...>]

Hi Michael,

Am Montag, 17. Juli 2006 19:55 schrieb Michael Price:
> Thanks for responding so quickly and thoughtfully.  As for negative
> submatrix increments, I'm not using them, but I also may have reported
> the wrong failed assertion; it may have been "assert(cI.end() <
> mat.size(0))" on line 273.

yes, might have happened.

> For now I've only been doing overdetermined systems (for example: 50 to
> 200 equations, 3 unknowns).  I think the change of linslv.cc:220 makes
> sense to correctly initialize Xtmp for either m >= n or m < n.  I wonder
> why that was changed to Xtmp = B?

That change was probably "not-so-well-thought-out". You can see in the CVS 
logs that it wasn't me. Obviously at the time I should have had a closer look 
and/or have insisted on a test case. Well, you never stop learning.

> Anyway, changing the work vector size for the QR solve routine to agree
> with the manual (as you suggested) eliminated the crash I was
> experiencing.  I will take a closer look at it later, since we will also
> be using the sparse LU solver and eigenvalue functions (which I
> understand we may need to debug somewhat).  I'll keep you posted.

Are you familiar with using CVS? If yes, then I'd be happy to add you as a 
developer to the lapackpp project on sourceforge. That way, you have the 
faster CVS access as available for developers, and potentially you could 
contribute your own patches as well. Surely I'll have a close look at any 
submitted code, but CVS ensures nothing will get lost. If you like to use 
developer CVS, you would have to get an account at sourceforge.net and tell 
me the account name.

Regards,

Christian