Replace TTIR string parsing with structured MLIR walk in Triton kernel mutation analysis (#120476)
Summary: Previously, we relied on the `lark`-based parsing of the string TTIR representation dumped by the Triton compiler. However, this has proven to be brittle in the face of changes both in the user-written Triton kernel code and in the Triton compiler code.
In this PR, we add an alternative way of mining the function information from the TTIR based on walking the tree of structured MLIR entities. To this end, we rely on the MLIR bindings exposed by `libtriton` (related PR in Triton: https://github.com/openai/triton/pull/3191).
For now, we introduce gating based on whether `ttir_module.hasattr("walk")`. This will allow switching to the newly introduced TTIR analysis approach only when the new MLIR bindings (including that of `ModuleOp::walk`) become available in the Triton pin. Before then, we'll keep using the old string TTIR parsing-based approach.
Test Plan: The new functionality was tested locally with the latest Triton version compiled with the added new MLIR bindings: all Triton kernel mutation tests in `test_triton_kernels.py` are passing. Here we rely on the CI for regression testing, but it won't cover the new functionality due to gating.
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120476
Approved by: https://github.com/oulgen
