Can Loopy extract isl-like representation from existing CUDA kernels?

[yuxuan-z19]

I’m looking into analyzing nested for loops in existing CUDA functions (from .cu source files), specifically to extract polyhedral dependencies and tiling information. I know that Loopy can generate CUDA code from an isl-like description.

I’m wondering if the reverse is possible: starting from a CUDA kernel (for example, accessed via PyCUDA), can one recover a Loopy/isl-like representation of its loops and dependencies?

My goal is to analyze loop structures and extract tiling/affinity info for LLM-driven optimization research, rather than just executing the kernels.

Note: My experience with compiler internals is limited, so any guidance on feasible approaches or existing tools would be greatly appreciated.

Thanks!

[inducer]

Nope, it can't. That is its own research subject. Consider https://repo.or.cz/pet.git.

[yuxuan-z19]

Thanks for the lightning response, @inducer!