Build oneDAL from source

[isuruf]

Since oneDAL is an open source project under the UXL foundation (a sub organization of the Linux foundation),
it would be good to build oneDAL from source instead of re-packaging the binary.

Thoughts on this @napetrov ?

[napetrov]

@isuruf Yes, this was a longerterm plan to go with source build, starting with ARM, RISC-V builds. First we need to cover valdiation for those- now ARM validation is verry limited and need to be extended first through the upcoming year.
The first question would be on any potential issues of 2 feedstocks populating same package to different arches?

On x86 build - this is more interesting. While x86 portion is more stable, oneDAL also covers SYCL based stuff which is less stable on one hand, have issues with public validation and we do have limitation on minimal requirements for SYCL there.

And also we currently have intel driven cycles vs external to Intel release process - there are plans for establishing proper governance and decoupling this more from Intel.

[isuruf]

On x86 build - this is more interesting. While x86 portion is more stable, oneDAL also covers SYCL based stuff which is less stable on one hand, have issues with public validation and we do have limitation on minimal requirements for SYCL there.

Is this about DPC++ version needed? I thought that we have all the dependencies in conda-forge to build oneDAL from source.

And also we currently have intel driven cycles vs external to Intel release process - there are plans for establishing proper governance and decoupling this more from Intel.

Yep, given that it's governed by uxlfoundation, I'm hoping we don't have to rely on builds by Intel.

[napetrov]

Is this about DPC++ version needed? I thought that we have all the dependencies in conda-forge to build oneDAL from source.

DPC++ requirements on glibc, but honestly we didn't look on current state for last 6 months.

Yep, given that it's governed by uxlfoundation, I'm hoping we don't have to rely on builds by Intel.

Yes, but it's not a immediate thing - even now intel doing 99.8% contribution, which would be changing overtime and we would get more contribution, governance body. but this would take time.

So i think we would start with non x86 builds to start new feedstock and then move more things there. Might be one of the things that might work is decoupling of cpu and sycl implementaitons - this is somethign that we would discuss within oneDAL governance body.
Also arm target feedstock would depend on non Intel contributors support.