Subject: Re: [ublas] [bindings][blas] Included CBLAS support
From: Thomas Klimpel (Thomas.Klimpel_at_[hidden])
Date: 2010-01-08 13:45:18
Rutger ter Borg wrote:
> Thanks Thomas. I've worked on the LAPACK regression suite, and over 70%
> passes now using the latest generated bindings and traits.
The regression tests for ublas_hptrf_hptrs, ublas_pptrf_pptrs and ublas_sptrf_sptrs are probably still deactivated. You should reactivate them. You see, I also missed 100% ...
> Any hints on what to do with this lapack band storage requirement for e.g.
> gbtrs/gbsv? It kind of doesn't fit anything, it's like a workspace inside a
It is different from workspace. It's the fact that the factored matrix takes more space than the original matrix. This is also called "fill-in". The factorization of a banded-matrix without pivoting would not produce any "fill-in", but would not be numerically stable. But there are pivoting strategies that are numerically stable and at most double the bandwidth, and this is what lapack uses, I guess.
A simple solution would be let the user directly specify the parameter that tells lapack how many bands of the banded-matrix are actually used/nonzero before the factorization.
> Or do you mean CLAPACK, or some other part?
When I say "and the cblas part is definitively the bulk of them", I refer to the CLAPACK part of ATLAS (which is different from clapack, which is just the polished f2c version of lapack)