KokkosBlas::gemv and gemm no longer compile with mixed CudaSpace and CudaUVMSpace views.

[vbrunini]

We have previously used KokkosBlas::gemv and gemm with A being a matrix in CudaSpace and x/y (or b/c for gemm) being in CudaUVMSpace. As of the most recent snapshot to Trilinos this now fails to compile due to the static_asserts that check SpaceAccessibility<...>::assignable.

@lucbv

[lucbv]

Thanks for reporting this @vbrunini I would have thought that CudaUVMSpace is assignable from the Cuda execution space?
I will talk with the Kokkos Core team and clarify that, in general the new assertion is there to make sure that users do not pass something that is stored in HostSpace with a device execution space.

[lucbv]

Ah, actually the issue is with assignable vs accessible that makes sense, my bad for misunderstanding, will get that fixed.

[lucbv]

Okay, should be addressed in PR #1914

[vbrunini]

Thanks!

[lucbv]

Can you check that everything works fine now that #1914 was merged?
If so we can close this issue.

[vbrunini]

I think the change looks good, but it will be a bit before I get a chance to test it in sierra. Could you add a kokkos-kernels test for the mixed memory space usage? I think if you do that it's safe to close this before I get a chance to test in the app.

[lucbv]

I can add such a test but we will have to be a bit careful with it as it requires the flag KOKKOSKERNELS_TEST_ETI_ONLY=OFF and this is only tested in one of our CI builds and it is running on CPU...
I'll discuss with others to see if we can add such a build on GPU but we might simply not have the testing capacity for it, we need to check.