Document Multi-GPU and lock-based atomic operations under "Known Issues"

[dalg24]

We need to add an entry to the known issues with CUDA and HIP, that dispatching a kernel that calls an atomic function that is not lock-free (e.g atomic operation on a Kokkos::complex<double>) will fail**. Also, see desul/desul#110.

I actually believe the same goes for SYCL when building a queue to a device that is not the one that Kokkos was initialized with.

** what does actually happen there? Do we know?

cc @tcclevenger @Rombur @masterleinad

[masterleinad]

I actually believe the same goes for SYCL when building a queue to a device that is not the one that Kokkos was initialized with.

I don't think this is any different from accessing global memory allocated by other queues. As long as you are using the same context, you are good.

[dalg24]

Oh, because for SYCL we placed the lock arrays in device memory and there is no such thing as constant memory?

As long as you are using the same context, you are good.

What does that mean?
Can you access the lock tables from a queue on a different GPU?

[masterleinad]

Can you access the lock tables from a queue on a different GPU?

If they share the same context, I would assume so. I'm not sure if we want to rely on that, though, since we don't enforce it in Kokkos and I'm not sure we even can. On the other hand, we are explicitly sharing the same context in our interoperability tests.
Ultimately, we just need to test since device_global variables are still experimental.