SVE Backend

[jeremylt]

No description provided.

[jeremylt]

This now compiles and passes the t3 tests on Ookami.

ToDo:

• Performance comparisons
• Makefile flag fix (not sure how the AVX one even works)
• Improved vectorization instructions?
• Add SVE backend to README
• Do we want manual unrolling for opt? The unrolling is pretty straightforward

[LeilaGhaffari]

I remember from Ookami's talk that performance with GCC was the worst among all compilers. I did a brief experiment before losing access to Ookami.
sve/blocked with armclang: DoFs/Sec in CG: 1.43289 (1.43289) million
opt/blocked with gcc: DoFs/Sec in CG: 0.729637 (0.729637) million

I am not sure why these numbers are so small compared to what we had Friday but I think the poor performance with sve might partly have something to do with compiler, maybe?! I have to apply for an account to do more experiments though.

[jedbrown]

Is this ready for review? Should we include it in v0.11?

[jeremylt]

The two big todos are fixing the makefile magic and seeing if this actually does anything different than OPT in terms of performance.

[jedbrown]

I noticed that libxsmm contains aarch64/SVE code and it's announced as supported for the next release.

[jedbrown]

Do we have a place where we can measure performance? There is a machine at Sandia that you can access if you put in a Sarape request and AWS c7g also has SVE. JLSE also has a system that I could try requesting.

[sebastiangrimberg]

Hi @jedbrown @jeremylt, if there is any way I can help here let me know. Happy to take the time to test on AWS's c6g/c7g for you if there's a specific case you're interested in, also against whatever libxsmm has for Arm optimizations as well.

[jeremylt]

I haven't had time to close the loop on this one. The big two todos are fixing the makefile detection of SVE support and running some basic performance comparisons to the /CPU/self/opt backends. I'm not sure the best way to do the first but we have a script to run our PETSc examples for the second.

[sebastiangrimberg]

Sounds good, no rush. I'd be very happy to run the PETSc benchmarks on whatever instance types you want whenever it is ready, if that would be helpful to you.