Draft: Add portable query for exec space

[mauneyc-LANL]

PR Summary

Adds portableExecIsHost() function that returns true if the execution space is on host, false if on device.

For PORTABILITY_STRATEGY_KOKKOS, compares Kokkos::DefaultExecutionSpace to Kokkos::HostSpace::execution_space. This can differentiate between openmp and cuda, for instance

For the explicit PORTABILITY_STRATEGY_*, just return true or false as expected.

This maybe useful e.g. https://github.com/lanl/spiner/pull/64

PR Checklist

• [x] Any changes to code are appropriately documented.
• [x] Code is formatted.
• [x] Install test passes.
• [x] Docs build.
• [x] If preparing for a new release, update the version in cmake.

[mauneyc-LANL]

I wasn't quite sure, save doing a lot of #ifdef KOKKOS_ENABLE_OPENMP ect. on the most correct way to implement a test. Ideas welcome

[dholladay00]

@mauneyc-LANL I think using something similar to what I had is a bit more robust, for example does that do the right thing if OpenMP is the default memory space, I think this code below gets at the heart of the matter (do I need to copy memory somewhere else), or possibly we need this and other variables for different use cases.

     using HS = Kokkos::HostSpace;
     using DMS = Kokkos::DefaultExecutionSpace::memory_space;
     constexpr const bool execution_is_host{
         Kokkos::SpaceAccessibility<DMS, HS>::accessible};

[mauneyc-LANL]

@dholladay00 I've redone with your snippet.

If there are more flags like this that should be included, maybe best to close this PR and make a new one that's more comprehensive? Since this was originally spawned from an issue you had, you can probably judge how significant these changes are to get merged

[mauneyc-LANL]

@Yurlungur My initial commit was a function, but that was changed to a constant; I have no preference, though the constant should be namespaced - just let me know what namespace you want here.

The OpenMP specification has stuff to query the memory space, and Kokkos probably has some code committed to do this.

[Yurlungur]

@Yurlungur My initial commit was a function, but that was changed to a constant; I have no preference, though the constant should be namespaced - just let me know what namespace you want here.

Let's do PortsOfCall that's used elsewhere as well. It's inconsistent with the macros, but it feels more correct from a C++ style perspective.

The OpenMP specification has stuff to query the memory space, and Kokkos probably has some code committed to do this.

Ah good point. So for OpenMP Target since Kokkos knows at compile time, we should be able to query at compile time too.

[chadmeyer]

I agree with the sentiment of @Yurlungur that this should be compile time known so it should be a compile-time-evaluatable quantity -- whether a constexpr variable or constexpr function or whatnot. I'm not sure I know enough to weigh in as to which option is going to be better behaved.

[mauneyc-LANL]

@Yurlungur see 8204a93 for suggested changes

[Yurlungur]

LGTM thanks @mauneyc-LANL . One last request: Can you add one line to the docs mentioning this constant exists?

[mauneyc-LANL]

@Yurlungur see e82175b

[dholladay00]

@mauneyc-LANL did you run the test with none and kokkos portability strategies?

[mauneyc-LANL]

Testing passes with kokkos

[dholladay00]

I'm going to merge this in.