Re: rocBLAS Reference Library
I happen to be the co-maintainer of BLAS/LAPACK debian packages.
Netlib reference BLAS is slow. That's true. And in libraries which seriously
need high computation performance, we can add the following line in
debian/control:
Recommends: libopenblas0 | libblis3 | libatlas3-base | libmkl-rt | libblas3,
So that openblas will be installed by default and delivers acceptable performance.
We have three libopenblas0 variants: openmp, pthread, and serial.
Just pick one that uses a compatible threading library.
libopenblas0-openmp/unstable 0.3.23+ds-3 amd64
libopenblas0-pthread/unstable,now 0.3.23+ds-3 amd64 [installed,automatic]
libopenblas0-serial/unstable 0.3.23+ds-3 amd64
libopenblas0/unstable,now 0.3.23+ds-3 amd64 [installed,automatic]
If the threading library does not matter at all. Then a formal way
to specify the libopenblas dependency can look like this:
libopenblas0-pthread | libopenblas0
In this case we will not force the user to install pthread version when
the user has installed the openmp version manually.
Actually, patching autopkgtest is not the best solution here.
Because it only improves performance for testing phase.
Just add the above recommend line to rocBLAS, following pytorch.
Then you don't have to change the autopkgtest dependency.
Recommended package will be automatically pulled, by default.
On Tue, 2023-08-29 at 11:46 -0600, Cordell Bloor wrote:
Hello,
I was preparing to report the issue with OpenBLAS that was preventing its use with rocblas-test, but it seems the problem has already been resolved. The autopkgtest suite for the library should be much faster if we moved from Netlib BLAS to OpenBLAS. Am I understanding correctly that the way to specify this would be to update d/tests/control from Depends: librocblas0-tests to Depends: librocblas0-tests libopenblas0-pthread?
On my Ryzen 3950X, the rocblas test suite completes in ~131 minutes with OpenBLAS as the reference library.
Sincerely,
Cory Bloor
Reply to: