[Advice Needed] Looking for research directions around Tile IR — any suggestions?

[pengjh0111]

Hi all,

I've been working on a project that uses Triton as a frontend for Tile IR, and I managed to get a full end-to-end path working inside torch.compile (PyTorch ops → Triton kernels → Tile IR). The whole-network compilation is running fine now.

I'd like to use this as a base for some research — either on the performance side or more academically focused on Tile IR itself — but honestly I'm a bit lost on where to start.

So I figured I'd just ask here:

• Are there any known pain points or open problems in Tile IR that are worth looking into?
• Any directions you'd find interesting or impactful (scheduling, memory analysis, op fusion, etc.)?
• Any benchmarks you'd recommend for measuring improvements at the Tile IR level?

Any thoughts welcome, thanks!

[sanjoy]

Any directions you'd find interesting or impactful (scheduling, memory analysis, op fusion, etc.)?

I'd be interested in figuring out how to write performant mega-kernels in CUDA Tile, and if having certain language or IR constructs would make that easier.