CUTILE IR Feature request: smem atomics

[spapiniattestable]

Not sure where is the correct place to ask for a feature in CUTILE IR itself.

When can we expect support for shared memory atomics in CUTILE IR?
We are really excited about CUTILE IR, it solves a lot of problems we had.
smem atomics is the last blocker we have right now for adopting it.

Thanks!

[chelini]

Thanks for reaching out. This is indeed the right place to request new features.

At the moment, we do not have a committed public ETA for shared-memory atomics, and it has not yet been prioritized on our roadmap. That said, I’d be happy to look into this.

Could you share which operations and data types are most important to you? For example, whether you need add, min/max, or CAS (compare-and-swap) operations; whether the data types are integer, floating-point, or both; and whether CTA-scoped shared memory is sufficient or cluster shared memory support is required.

Any additional details about your workload or intended usage would also be helpful :)

[spapiniattestable]

CTA scoped atomic "send and forget" add on int32 in smem.
atomicAdd(), without needing the value
red.cta.add.s32
No need for cluster.
I am basically doing weighted histograms
Thank you!