axpby introduced deep_copies when alpha,beta are scalars

[brian-kelley]

@eeprude

Tpetra in Trilinos has tests that count the number of Kokkos::deep_copy calls (generally, they want to minimize that number). These tests are failing with KokkosKernels develop branch due to new deep_copies in axpby, probably introduced in #1895 or #2039 . I think it's one of those because they weren't in the latest release, 4.2, which doesn't fail the Tpetra test.

The specific case in Tpetra is where x,y are rank 2, and alpha,beta are scalars. These scalar values are being deep_copied to device views called "managed_a" and "managed_b". But this case should involve no deep_copies, and instead the axpby functor should just have the alpha/beta as members.

The only time there should be a deep_copy for alpha and/or beta is when the user gives them as rank-1 views on host.

It's possible to check that no deep copies are happening by running this case of axpby inside a Kokkos profiling region, and running the kernel logger from Kokkos tools. Here's a good example of this: https://github.com/kokkos/kokkos-tools/wiki/KernelLogger

[eeprude]

@brian-kelley Thank you for all the hints. I have taken a look at the code and I will work on it during the winter break.

[brian-kelley]

@eeprude No, you don't have to do it during break! This isn't even a configuration that people use, it's just for catching issues before we release to Trilinos. It just needs to be happen before we ship 4.3