Andrew Ho

It looks like on some platforms (Windows) UNKNOWN is used since they can have the same extension for shared and static libraries, and the function `mfem_get_target_options` uses different logic to...

Yes, it is used on all platforms, and I agree we should come up with a proper fix before this is merged. I'll look into that.

@cjvogl I think the UNKNOWN_LIBRARY issue should be fixed now? I can see cudart at least being in the linker flags in the generated cmake file. The CMake targets change...

@jandrej @camierjs @cjvogl @v-dobrev any chance you could take another look at this? I've been maintaining these changes as a git stash so I can apply these to other branches...

I looked through the HIP and CUDA parts again, and I believe both should be good now @pazner

I don't seem to be able to push to this PR, but here's a patch which fixes deprecation warnings internal to cusparse on some compiler platforms. @adam-sim-dev @pazner @tzanio can...

I think that was a copy-paste error. Yes, you're right we should use `CUDA_DIR` instead.

Are there any tests/examples in this branch which use these new functions?

@pazner I didn't dive into the details of why I got an order of magnitude slower with mod/div on CPUs, it could be that I'm benchmarking this single threaded/single rank...

I only ran a few examples for testing on lassen and tioga, as well as relied on the auto test. I just did a `make test` with CUDA+HYPRE and CUDA+METIS+MPI+MFEM...