Subject: Re: [ublas] [bindings] New traits system
From: Rutger ter Borg (rutger_at_[hidden])
Date: 2010-01-19 10:14:53
Thomas Klimpel wrote:
> What I typed (in the libs/numeric/bindings/atlas directory) to test your
> reported porting of ublas_symm2 was
> bjam define=BOOST_NUMERIC_BINDINGS_LAPACK_CLAPACK
> define=BOOST_NUMERIC_BINDINGS_BLAS_CBLAS ublas_symm2
> This worked, but this is just me cheating again. What you are interested
> in is probably "svn diff numeric-bindings.jam":
I've managed to get all atlas regressions pass except those for getri (known
issue with workspace / non-workspace clapack/lapack), and ublas_gesv5. The
latter compiles, but doesn't pass due to an assertion which I think is
correct. Could you take a look at ublas_gesv5?
Another question for our project's bjam expert: is it possible to split the
tests into three sets: blas / clapack / lapack, and have each set of tests
run against all available back-ends? I guess this would require some re-
ordering, but hopefully it is possible with bjam.
Another idea would be an automated benchmarking of gemm of various blas-
libraries, and put that in the documentation.