Dear EasyBuilders,
Following the recent discussions about BLIS, and Field's excellent
presentation at EUM'21 last week [1], we want to evaluate whether BLIS
is a suitable replacement for OpenBLAS in the foss toolchain, for
foss/2021a onwards.
My gut feeling says it's the proper way forward, but that's of course
not enough... :)
Things we want to look into include:
* Low-level performance evaluation compared to OpenBLAS and MKL, similar
to what Field showed in his presentation, but across a (wider) range of
CPU microarchitectures.
We could use tests & scripts available in the BLIS repo [2] for this,
but perhaps also other microbenchmarks.
* Compilation and performance impact for end-user applications like CP2K
& co, which heavily rely on BLAS-3 (see also [3]), on top of a
BLIS-based toolchain compared to foss & intel (& co).
We could use the new gobff toolchain (foss but with OpenBLAS replaced
with BLIS+libFLAME) for this, but perhaps also create additional
toolchains like iibff (Intel compilers + Intel MPI + BLIS + libFLAME +
FFTW).
The idea for the meeting would be to discuss who would look into what
(and on which systems), which tasks have to be done first, etc.
If you are interesting in actively joining this effort, please fill out
the doodle to find a good time for the meeting:
https://doodle.com/poll/ap6n6c9sa53z9zz2
(times should be local to your time zone, first slot starts at 08:00 UTC)
regards,
Kenneth
[1] https://easybuild.io/eum21/#blis
[2] https://github.com/flame/blis/blob/master/test/3/
[3]
https://github.com/easybuilders/easybuild/wiki/Conference-call-notes-20210203#switching-from-openblas-to-blis-in-foss2021a
