intelex for ARM MacOS

[benritchie]

Hey everyone.
we love using intelex. currently we mostly run training and inference in the cloud - intel machines, but our dev laptops are macs. would be great to run brief experiments on them, using the same models.
Therefore, the question is - seeing as #1771 is now implemented, adding base ARM support - how feasible would it be to get a mac build of intelex working again?
Thanks.

[napetrov]

Hi @benritchie ,
Technically this should be possible but you would need first to compiler oneDAL for Mac and then building scikit should be quite easy on top. Overall this is something that would likely to happen once we would have more robust ARM support in CI and public build
the main challenge with ARM build would be that as it's also switches from MKL to openBLAS there are some things that are not quite polished yet.

MacOS public CI was removed in this PR - uxlfoundation/oneDAL#2865 with migration from FPK lib(subset of MKL) to full MKL. Also there are still MacOS targets available in makefile, though they have to be changed to ARM.

Would be happy to help with enabling this back if you would like to contribute here. But there are no such plans currently at least for this year.
@rakshithgb-fujitsu @theComputeKid - feel free to comment as well

[rakshithgb-fujitsu]

Currently oneDAL can be built on arm platforms with SVE isa, the base neon build is untested. Since macs currently only support neon it might require few build changes which should be straightforward to implement along with a new mac arm target.

[benritchie]

Thanks.
I'll have a play, and see if I can get oneDAL building at least.
For interest - what's the vision of oneDAL now - I know it used to be pretty intel specific, and e.g. deliberately not enabling some of the optimised code paths on AMD. Is it now targetting generic support across manufacturers?

• a quick test suggests that might have changed? I'm now seeing similar perf on intel and AMD - is that the case?
  Thanks.

[napetrov]

Great questions. There are several aspects

1. Project is open for other platforms optimizations but initial Intel team focused on Intel specific optimizations. There was no contributions/interest from AMD, but recently couple of changes in dispatching logic on AMD was contributed and this should be reflected in new release. But overall from optimizations perspective, majority of optimizations still Intel specific and AMD acceleration would depend on external contributions.
   Remove dispatching to SSE2 code path on non-intel x86 CPUs oneDAL#3154
   Add proper CPUID eax checking oneDAL#3026

2. oneDAL is now part of UXL and while Intel is main contributor, it's no longer single contributor, so we now have ARM and RISC-V support. There still lot of work to be done, but initial results are quite promising - https://blog-en.fltech.dev/entry/2024/10/25/onedal-arm-ai-oss-framework-acceleration-fujitsu-monaka-en

[ethanglaser]

oneDAL GitHub issue opened for this: http://github.com/uxlfoundation/oneDAL/issues/3175