Metadata Declaration in Core

[DavidPoliakoff]

Currently looking for a review of this mechanism for declaring and printing metadata making sure that a code change like this wouldn't perturb anything, it needs more work. A sample output after the change, if you call print_configuration

Kokkos Version: 3.3.0
Compiler:
KOKKOS_COMPILER_GNU: 830
Architecture:
Default Device: N6Kokkos6SerialE
KOKKOS_ENABLE_ISA_KNC: no
KOKKOS_ENABLE_ISA_POWERPCLE: no
KOKKOS_ENABLE_ISA_X86_64: no
Atomics:
KOKKOS_ENABLE_GNU_ATOMICS: no
KOKKOS_ENABLE_INTEL_ATOMICS: no
KOKKOS_ENABLE_WINDOWS_ATOMICS: no

[DavidPoliakoff]

Confirming that this can play nicely with the Tools infrastructure. So from my side, this is what I need

[DavidPoliakoff]

Retest this please

[dalg24]

I would prefer if you tackled the recording the git commit hash in a separate PR

[DavidPoliakoff]

@masterleinad @dalg24 I think I've addressed the comments. I deferred the Git SHA stuff for a future PR, and almost everything else I just implemented as suggested. The only exception is that I'm not adding the machinery to print MSVC, as we didn't used to print that in our print_configuration machinery

[dalg24]

Why do you need the declaration in Kokkos_Core.hpp? Where else than Kokkos_Core.cpp would we use?

[DavidPoliakoff]

This could be used in backends, the CUDA backend will likely declare configuration metadata as well

[dalg24]

Why is this part of "architecture"?

[DavidPoliakoff]

This is part of architecture because that's where it printed in print_configuration before. We can change that, but the underlying goal in this PR was to not shake up the printing as it happened before

[DavidPoliakoff]

@dalg24, remaining are questions about making the default device part of the architecture (which it was before, but which I'm not opposed to changing), and whether we want to expose the declare_configuration_metadata symbol for use in other places

[DavidPoliakoff]

Retest this please

[dalg24]

Add the comment and remove the unused code.

Also I am curious whether you considered generating the tedious bit

#ifdef KOKKOS_ENABLE_FOO
declare_configuration_metadata("options", "KOKKOS_ENABLE_FOO", "yes");
#else
declare_configuration_metadata("options", "KOKKOS_ENABLE_FOO", "no");
#endif

with CMake. (NOT asking you to change)

[DavidPoliakoff]

@dalg24, I copied that tedious bit from our previous implementation. That implementation doesn't look good to me, but every time we change something to "something that's exactly the same" I find out that we broke something somehow

[DavidPoliakoff]

@dalg24, believe I have addressed the comments

[dalg24]

(Need to clean up history)

[DavidPoliakoff]

@dalg24: done