Kokkos Variant Checksum Differences

[MrBurmark]

Some kokkos variants produce a checksum that is significantly different than the other variants.

I compiled with mvapich 2.3.7 gcc 11.2.1 and ran with 36 ranks.

The difference between checksums should be negligible, less than 1e-10. This is printed in the RAJAPerf-checksum.txt output file.

Affected kernels:

• Basic_ARRAY_OF_PTRS
• Basic_PI_ATOMIC
• Basic_REDUCE3_INT
• Apps_NODAL_ACCUMULATION_3D

[MrBurmark]

@crtrott are you seeing these differences as well?

[MrBurmark]

@rhornung67 Once we get these passing, should we enable kokkos in one of the builds?

[rhornung67]

Yes, I was planning to add that after the RAJA release and before the RAJAPerf release.

I think in our tests checksum diff tolerance is 1e-8. However, I haven't looked into this in a while.