Sébastien Fabbro wrote: > Hi all, > > I'm in the process of updating the blas/lapack system in Gentoo. > I took the opportunity to upgrade and cleanup all blas/cblas/lapack > packages. The changes are a switch to the new style virtuals, > installation of pkg-config files, but mainly a lot of testing and fixing. > > Following an earlier thread [1] on merging blas and cblas, I decided not > to do it to allow combinations such as acml(blas) + cblas-reference > (cblas) or even gsl(cblas) which does not even depend on blas. > > Since there are many changes and blas/lapack are essential to a lot of > Gentoo packages, they are in the science overlay [2] for pre-commit. > > I would be grateful if many people test them before I commit them to > portage and ask for stabilization. Report by replying to this thread, or > emailing directly to me. > > Some early draft documentation can be found in [3]. > > Comments welcome! > > Thanks, > > Sébastien Fabbro > > [1] http://thread.gmane.org/gmane.linux.gentoo.science/757 > [2] http://gentooscience.org/ or "layman -a science" > [3] http://dev.gentoo.org/~bicatali/blas-lapack.txt
I pretty much said my piece in [1], but some recent notes: 1. ATLAS is very close to 3.8.0. Clint is wringing out the remaining bugs on things like Sparc and PPC -- as far as I know x86 and amd64 are already rock solid. 2. As far as the Intel C and Fortran compilers are concerned, since I'm now an Intel-free zone I pretty much don't care. :) But they are "less than free", Clint Whaley from ATLAS doesn't think they're worth the hassles of renewing eval licenses, and I've seen a few other performance-related sites that claim they are inferior to GCC in a number of respects. Check out http://www.agner.org/optimize for an in-depth comparison of (C) compilers. 3. Again, since I'm an Intel-free zone, I don't know about the Intel math libraries. I'd be surprised if anything beat ATLAS on dense linear algebra, but for other types of number crunching, like FFTs, there is probably a good reason to have the chip-specific libraries around. So I'll test what's there over the weekend and see what happens. I'm going to see if I can get the High-Performance Linpack benchmark working over the weekend. I think *that's* in the overlay too, but it's masked at the moment and I haven't attempted to either unmask it or download the original. I'm actually planning to extract the "structure" of it from the raw code and use it as a test of concurrency primitives in a number of languages, so it's great timing for it to show up in an overlay. :) -- [EMAIL PROTECTED] mailing list
