Integration with CubeCL

[nathanielsimard]

We'd love to explore running CuTile kernels through CubeCL. CubeCL is a multi-platform GPU compute language for Rust that powers Burn, and integrating CuTile would let users write tile-optimized NVIDIA kernels within that ecosystem.

We've opened a tracking issue on our side with more details and two proposed integration paths: tracel-ai/cubecl#1282

Would be great to hear your thoughts on feasibility and whether there are any API considerations on the CuTile side that could facilitate this.

[elibol]

Hey Nathaniel, thanks for opening this.

The standalone launch path (option 1 in tracel-ai/cubecl#1282) is the direction we'd be able to support in the near term. The compilation API and ABI documentation still need to be prioritized, but the core interop work (borrow_raw constructors on cuda-core types, #69) is already on our near-term TODO.

Happy to coordinate on details as we get to it 🙂

[elibol]

@nathanielsimard cudarc interop has been merged into main. Can you check it out here and let me know if that will work for you all?

There is some DSL API cleanup I need to do before cutting the next release, but after that I can prioritize the remaining tasks to get interop working :)