findRelOffset test assumes UVM

[ibaned]

According to @crtrott , KokkosKernels is not allowed to assume UVM.

[mndevec]

I am rolling back on this. As I notice that not only the test, but also the implementation itself is sequential, as a result, the implementation itself is assuming UVM on GPU side.

@mhoemmen, is this the only intended use of findRelOffset? Does this require a GPU implementation, or does it make sense to have a GPU implementation?

[mhoemmen]

@mndevec findRelOffset should be callable on the GPU too. It's meant to be called by a single thread, but could be called on device. Thus, I don't think the function assumes UVM, but it does look like the test assumes UVM.

[mndevec]

Thanks Mark. So, is it a function that should only be called in the parallel region?

[mhoemmen]

@mndevec wrote:

Thanks Mark. So, is it a function that should only be called in the parallel region?

No, not necessarily. It may be called on device, but it may also be called on host. KOKKOS_FUNCTION means __host__ __device__, not just __device.

[mndevec]

Okay thanks. In this case,
Unit tests should call this function
1- in a parallel region for all execution spaces.
2- directly from host, only if the view is not in Device Memory.

[mhoemmen]

@mndevec Yes, that's correct. Thanks!

[crtrott]

May want add new unit test to call in CUDA.