Verify the View dimensions before allocating?

[Shihab-Shahriar]

I had a bug which looked like Kokkos failed to allocate an otherwise reasonable amount of space. But turned out I had actually overflowed an int variable that I was using as an extent to instantiate a kokkos view. Like the n variable below:

Kokkos::realloc(Kokkos::view_alloc(space, Kokkos::WithoutInitializing), v, n)

I wonder if you would consider adding a check for this sort of degenrate cases? Currently the size_t argument type for the extents masks this type of error. So maybe something along the line of:

// original function
void realloc(size_t n){
}

// an overload for signed types
void realloc(int n){
  if(n>0) realloc(n)
}

[dalg24]

turned out I had actually overflowed an int variable

Overflow of signed integers yields undefined behavior (UB) so you would need to handle it on your side regardless of whether Kokkos::realloc can detect that a negative integer was provided as argument.

That said, I do agree that it is somewhat unfortunate that realloc takes the extents in each dimension as std::size_t and that it silently wraps around when a negative integer is passed.
The issue is well beyond realloc or resize, you get the same behavior when you construct a View with an extent passed a negative integer.

[crtrott]

We will address this with the mdspan refactor.