What does triton_gpu.async_commit_group do? #3249
Closed
prathams417
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
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.
Beta Was this translation helpful? Give feedback.
All reactions