What does triton_gpu.async_commit_group do?

[prathams417]

I'm an intern working at Intel on the triton compiler, so forgive me if I missed something.

I was wondering if someone could give me a some insight as it what the triton_gpu.async_commit_group operation does. I thought it was a wait-like operation that waits for the async operations to complete, but a separate triton_gpu.async_wait exists. There is minimal to no documentation on this operation.

If we look at the PTX it maps to cp.async.commit_group , but I don't quite understand the explanation behind it (https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#data-movement-and-conversion-instructions-cp-async-commit-group).

The current idea on Intel hardware is to map that triton operation to the OpGroupWaitEvents spir-v operation, but I'm not sure if it is an equivalent operation. Any guidance would be greatly appreciated.