[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: RFC: Changing the debian system default BLAS/LAPACK implementation


Generally in favour as I usually opt for faster OpenBLAS as defaults as well.

But allow me to mention one cautionary tale that is very current.  Besides
the Debian work, I am also upstream for a few R packages. One or two have
reasonably become popular and widely ysed, and I run what we call
'reverse-depends' in R/CRAN land: check all packages that depend on it.
Which, for Rcpp, is close to 2000 now :-/

To help a little, I have access to an old VM in a data center in Europe where
I run plain Debian testing.  I had a few packages reliably and repeatedly
fail (small parts of) their 'R CMD check ...' tests until a few days ago when
I, very reluctantly, replaced libopenblas* with the fallback slow reference
blas. Now tests pass. Of course, this was not pervasive.  In the most recent
case it affect three out of 700+ packages for RcppArmadillo.

This is not an indictment of OpenBLAS. It is maybe simply a reflection of
some tests being too tight. I really don't know. Maybe it really is just a
reiteration of 'what every computer scientist needs to know about floating
point' ...  But _in practice_ these libraries are never as _perfectly_
interchangeable as we think they should better _in theory_.

That said, more choices are better than fewer so I'd love for you to continue
your work so that I can also try Blis :)


http://dirk.eddelbuettel.com | @eddelbuettel | edd@debian.org

Reply to: