[Pkg-scicomp-devel] Bug#578052: Bug#578052: Bug#578052: [atlas] providing >20 separate optimized library packages is insane
Sylvestre Ledru
sylvestre at debian.org
Sat Apr 24 23:44:18 UTC 2010
Le lundi 19 avril 2010 à 10:21 +0200, Fabian Greffrath a écrit :
> Am 18.04.2010 11:16, schrieb Sylvestre Ledru:
> > By the way, since it has been this way for a decade, would you mind
> > if I lower the importance of this bug ?
>
> Please set the severity to what you consider appropriate.
OK, thanks.
> > Sorry but I don't know yet.
>
> I've spent some though about this issue over the weekend. Please allow
> me to give the following suggestion. (Disclaimer: (1) I do not really
> care about these libraries and do not even consider myself a typical
> user of them; I have just become curious about the Debian packaging.
Thanks for the precision.
> (2) I don't know how far these libraries are interconnected and if
> they are ABI-compatible in a way that allows for all possible
> combinations of atlas, lapack and blas libraries.)
They are. If it fails, you can consider it as a bug.
> - I would split up the shared atlas libraries into 3 packages:
> * libatlas3gf (contains the non-blas, non-atlas libraries)
> * liblapack-atlas3gf (contains the atlas-variant of lapack)
> * libblas-atlas3gf (contains the atlas-variants of blas)
After some investigation (ie, I don't know if it is possible), I could do that.
> - Pros:
> * Reduced complexity
> * No more update-alternatives (i.e. symlink hell)
> * No more manual selection of optimized library flavours required
> * No more redundant installation of two or more library packages
> providing the same ABI
>
> Cons:
> * No more on-the-fly switching between flavours
> * Only one combination of atlas/blas/lapack library packages can be
> installed
> * Switching between library flavours means installing one package
> and removing the other
>
> What do you think about it?
Well, for now, I would not prefer implementing your proposal. The two
first cons seems to me quite important. It is a common thing in
numerical computation software to switch between implementations (the
various atlas or refblas)...
One of the goal of the update-alternatives implementation was to tackle
it.
However, you have a point when you say that it is important that the
system should managed transparently it.
Sylvestre
More information about the Pkg-scicomp-devel
mailing list