Le mercredi 18 août 2010 à 12:13 +0100, Simon McVittie a écrit : > On Tue, 17 Aug 2010 at 23:09:08 +0200, Sylvestre Ledru wrote: > > Quick remember, Atlas is a linear algebra library implementing the BLAS > > API/ABI. It is widely used in the scientific computing world but also by > > some spreadsheets (openoffice). > > This is an highly optimized library. The optimisation is done at build > > time against the hardware it is building on. > > It seems to me that there are two major use cases: > > For use on smallish data sets (most OpenOffice users, and I suspect many > scientific users too), all that matters is that it's correct and not > ridiculously slow; conservative assumptions (libatlas3gf-base), or indeed > the unoptimized BLAS, would probably be fine. > > Prompting OpenOffice users with a debconf prompt that, as far as they're > concerned, says "[things you don't understand] will be slow unless you > [things you don't understand]" seems very undesirable. I don't want to > have to care about scientific computing stuff on large data sets, because > I don't *have* large data sets; I just want a spreadsheet :-) > > (Is there anything that specifically needs Atlas and couldn't use BLAS?) BLAS or Atlas base are fine in these cases.
I am more concern about advanced users of scientific computing software (Scilab, R, Octave...) which are familiar with such tools but not familiar enough with the internals. They just see these software as a whole and would not guess that changing RefBLAS => Atlas could improve the performances to a 40 factor... Here, at Scilab (and other people at the DebConf reported me the same experiences), it is not rare that people are complaining about the speed of the software because RefBLAS is used as the linear algebra library. > The other use case is serious HPC where it's worth rebuilding your packages > for the particular host/cluster, in which case an easy-ish way to build > a libatlas3gf-local package (/usr/share/doc/libatlas*/examples?) seems > sufficient? It is already possible to do it easily. apt-get source atlas && cd atlas* && fakeroot debian/rules custom Sylvestre -- To UNSUBSCRIBE, email to debian-devel-requ...@lists.debian.org with a subject of "unsubscribe". Trouble? Contact listmas...@lists.debian.org Archive: http://lists.debian.org/1282132799.11038.8112.ca...@korcula.inria.fr
