[Rcpp-devel] RcppArmadillo error

[Douglas Bates]

On Thu, Apr 1, 2010 at 8:57 AM, Dirk Eddelbuettel <edd at debian.org> wrote:
>
> On 30 March 2010 at 16:12, Romain Francois wrote:
> | > It is not certain that those will be the same as the libraries used by
> | > Armadillo.  If they are not shared objects you run the risk of getting
> | > two copies of Lapack, etc.  and possibly different BLAS, leading to
> | > different performance characteristics and perhaps different results.
> |
> | Hmmm. Interesting. I did not think of that problem.
> |
> | On my machine, I have :
> |
> | Rscript -e "RcppArmadillo:::LdFlags()"
> | /Library/Frameworks/R.framework/Resources/library/Rcpp/lib/x86_64/libRcpp.a
> | -L/usr/lib -larmadillo
> |
> | so linking against Rcpp "user" library and armadillo... but when I
> | manually set PKG_LIBS to just Rcpp user library, it still works, so we
> | actually do not need to link against armadillo library. armadillo itself
> | does not __need__ to get compiled into a library because it is all
> | templates.
>
> But "it still works" is no proof. We had talked about this once before (I
> think) and I pointed you to the Armadillo FAQ which states that it _can_ be
> used without LAPACK but _should_ be used with LAPACK (for performance).
>
> When you drop -llapack from the link line, the linker "simply" finds it via
> R's library dependencies. So when Armadillo needs to find a Lapack routine
> for inversion, it finds the one R uses.
>
> But I do not think we can turn that upside down and not link Arma with lapack
> as R can be build in several configurations so I fear we'd end with some that
> are suitable for Arma and some that may not be.
>
> So Doug's warning is a good one.
>
> OTOH this was meant to be a simple demo package and I simply do not want to
> spend a day or so on configure et al to make it more complicated than it
> already is.

Thanks for the clarification, Dirk.  I had sort-of reached the same
conclusion myself but not as well thought-out as yours.