[SYCL][CUDA] Implement root group barrier

[keyradical]

This PR adds an algorithm for doing a GPU wide barrier in CUDA backend.

Rough outline of the algorithm:

• Every 0th thread from each workgroup performs atomic.add(1)
• The same thread checks the atomic result with ld.acquire in a loop until it's equal to total amount of workgroups.
• All threads call group-wide barrier.sync

One caveat to this is that there is no initialization of the atomic start value. So if we call this barrier several times in a kernel, on the second iteration, the start value will already contain the result from previous barrier. That's why we actually spin the while loop while current value % totalWgroups != 0.

[maarquitos14]

I don't see that we ever check neither X nor Y. Should we? Otherwise, why we need this? How do we make sure it ran?

[keyradical]

I wanted to explicitly delay some of the workgroups by adding them more work to do, because I've seen this test passing if insufficient barrier was used. For instance on CUDA backend, doing work-group wide barrier would be enough for it to pass and that is not correct. I think this test should perform some work-group divergence to actually check that we actually perform gpu-wide barrier.

How do we make sure it ran?

The X and Y are declared as volatile and my understanding was that this would prevent compiler from removing them with some optimization.

[maarquitos14]

Right, can we have a comment explaining it? Otherwise we risk that this code will just be removed in the future thinking it's not required.

[maarquitos14]

LGTM.

[MartinWehking]

LGTM

[keyradical]

@intel/llvm-gatekeepers can you merge this, please?