Fix allocating large Views

[masterleinad]

Fixes #4906. For Kokkos::Views with dynamic View greater or equal to 3, we use unsigned int as type for storing the dynamic rank. Hence, we need to explicitly cast to size_t in some places where we multiply the ranks and return size_t.
Still needs the test.

[fnrizzi]

@masterleinad is this ready or still draft? ah i see, the test is missing

[masterleinad]

@masterleinad is this ready or still draft? ah i see, the test is missing

I just need to add a test and make sure it works on all backends. Otherwise, it's ready from my side.

[mzuzek]

Hi @masterleinad!
Do you think operator() in ViewOffset could also be affected and should have size_type casts in index multiplication ?

[masterleinad]

Do you think operator() in ViewOffset could also be affected and should have size_type casts in index multiplication ?

It's probably a good idea.

[PhilMiller]

Should this potentially get cherry picked for 3.6 as well?

[PhilMiller]

In operator() and other spots with nested expressions, I think the casts may need to be applied pretty much ubiquitously, since otherwise subexpressions may overflow before reaching the cast

[masterleinad]

In operator() and other spots with nested expressions, I think the casts may need to be applied pretty much ubiquitously, since otherwise subexpressions may overflow before reaching the cast

Thinking about it, we should discuss this. I'm pretty sure that the type for the variables was chosen here due to performance implications which might matter more for View access than calling span() or the like. In the end, this is something that can be worked around in user code by providing the indices as size_t.

[masterleinad]

I marked this as ready for review and would consider making the discussion about the call operator separate.

[PhilMiller]

Could this assign a non-zero value foo, and then assert on the host that v.access(idx, ...) == foo? As written, this is a weak smoke-test that the particular case doesn't happen to crash in our CI environments

[masterleinad]

4a657fd

[PhilMiller]

Up to you whether to make the test more demanding.