Bug#569436: [Pkg-scicomp-devel] Bug#569436: Bug#569436: lapack: zgesvd seems to give incorrect results

Sylvestre Ledru sylvestre at debian.org
Sun Feb 21 10:49:29 UTC 2010


Le mardi 16 février 2010 à 08:40 -0600, Kumar Appaiah a écrit :
> Dear Sylvestre,
> 
> On Tue, Feb 16, 2010 at 03:35:28PM +0100, Sylvestre Ledru wrote:
> > 
> > > > Since SVD is something I consider very important, I would request you
> > > > to please re-upload Blas with an optimization level of -O2, so that
> > > > things remain sane with all the reverse dependencies.
> > > OK, thank for your investigation. I will do that then.
> > I had a look and the problem is that blas is compiled by default with
> > -03. I am afraid of switching to -02 decreases the performance on a key
> > component...
> > However, I have a few leads:
> > * disable optimisation for just this file
> 
> But which file? From what I can see, zgesvd in Lapack seems to
> indirectly call several Blas functions. So, which file would you
> compile with -O2?
We would have to dig to find out...

> > * see if forcing gcc-4.3 brings back the initial behavior
> 
> True, but this is not a permanent solution, since it is unlikely that
> Squeeze will be shipped with two gcc's.
Probably not.

> > * report upstream (which I will probably have to do anyway)
> 
> That is true. There is a fourth option:
> 
> * Use -O2 on amd64, and -O3 on the rest: this is a sane option, since
>   Atlas is available anyway on amd64, and Atlas provides better
>   performance than the corresponding Blas/Lapack libraries
>   anyway. Would you be agreeable to this?
I am not very happy about this neither. amd64 represents now a lot of
computers...


For the record, I just tried with Scilab and an old version of atlas and
I get:
    0.0288316 - 0.1225715i    0.2941664 + 0.0112984i  - 0.7517416 -
0.0002798i  
  - 0.0345902 - 0.5162818i  - 0.3585172 - 0.2455863i    0.6758972 -
0.3549573i  
  - 0.0702840 - 0.8216660i  - 0.7627657 + 0.4766231i  - 0.0113633 +
0.1333745i  
 

Sylvestre







More information about the debian-science-maintainers mailing list