Newer RAJA support request

[chapman39]

Hello,

The LiDO team is seeking to use a GPU-enabled hiop. However, we use a much newer version of RAJA: @2024.07.0. I noticed that hiop uses @0.14 and lower.

This is also preventing us from linking hiop with Axom on GPU machines. This is because we build Axom with GPU support, and as a result it is incompatible with a GPU-less hiop.

Would it possible to support newer RAJA/ Umpire versions in the hiop project + Spack package? If you need any additional information let me know. Thanks.

[nychiang]

@chapman39 we do provide the spack formula to build hiop with the latest RAJA/Umpire: see here
We plan to update the formula on spack repository with the next hiop release.

[cnpetra]

@chapman39 I think LiDO uses the quasi-Newton solver of HiOp, which is currently being ported to GPUs. I suggest we have a meeting to understand what are the needs on LiDO's end to prioritize GPU-related developments in HiOp.

[chapman39]

Sounds good. It might be more relevant to talk to @tepperly about priorities, since he is currently experimenting with GPU capabilities in lido.

[chapman39]

I reached out to Tom, and right now LiDO is testing GPU capabilities on the physics (MFEM?) side, rather than solvers (e.g. hiop). Testing solvers on the GPU is not a priority at this time.

With that being said, it would still be useful to allow hiop to build with the latest version of RAJA in the Spack package, maybe not using it to the fullest extent - but this would allow LiDO to build hiop with +axom support.

[chapman39]

On a related note, I have a PR in Spack adding +axom to hiop. spack/spack#49817