8000 GitHub - rdadolf/ml-reentry-mlir: A learning exercise for MLIR infrastructure. · GitHub
[go: up one dir, main page]

Skip to content

About

A learning exercise for MLIR infrastructure.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

GNNC

An experimental MLIR-based compiler for Graph Neural Networks.

This is a learning/proof-of-concept project. There's been a bunch of work on GNN compilers, but the landscape has gradually settled into PyG as a home for most users and researchers, especially for new models, and either bespoke implementations or a gradually-fading DGL.

One of the interesting parts of GNN compilers to me has always been aggressive kernel fusion. Most research efforts have tried it, and DGL certainly leveraged it in the practical space. Ultiamtely, though, most tools (the TVM folks excepted) relied on hand-tuned, hardcoded monolithic GPU kernels to carry the day. PyG, notably though, hasn't really gone down that path very far, and I'm casually interested in the compiler alternative for getting the best of both worlds: how far we can go with IR-level fusion and better codegen while retaining the benefits of PyG (and perhaps more importantly, its ecosystem and users). Despite not really being designed for it, triton has had a fair amount of success with it in more mainstream models, and I don't think the community has really pushed the idea as far as it can go.

In any case, that's what this project is. I took a stab at this problem back in 2020, but a lot of the pieces just weren't ready. Aart and the folks behind the sparse_tensor dialect deserve a lot of the credit for the fact that is no longer true, but Torch itself has made a lot of strides in tracing, compilation, and integration, which have helped tremendously. I'm also looking forward to taking advantage of MLIR's Lighthouse effort, which is still young but shaping up well.

Basic Architecture: PyG → torch-mlir → MLIR (linalg → sparse-tensor) → GPU/CPU.

Common commands

First time after clone (the multi-GB builds; the devcontainer runs uv sync automatically):

git submodule update --init                   # NOT --recursive; we don't use them
uv sync
bash tools/build-llvm.sh                      # ~30 min cold; ccache after
bash tools/build-torch-mlir.sh
bash tools/build-gnnc.sh                       # gnnc's C++ MLIR passes
bash tools/link-stack.sh                       # link mlir/torch-mlir/lighthouse into the venv
bash tools/build-pyg-libs.sh                   # optional: PyG baseline only

The Python stack (mlir / torch-mlir / lighthouse) is linked into the venv by tools/link-stack.sh, so the CLIs, pytest, and the editor import it directly — no PYTHONPATH. Re-run it only after the venv is recreated. source tools/env.sh is optional: it just puts the LLVM tools (mlir-opt, FileCheck, …) on your PATH for interactive use.

Worktrees

A plain git submodule update --init re-clones third_party/llvm-project (several GB) into every worktree. To share one local copy instead, point the reference at an existing checkout's llvm-project, scoped to that one submodule so the rest still clone normally:

git submodule update --init --reference <other-checkout>/third_party/llvm-project third_party/llvm-project
git submodule update --init

If you don't care or it's not working, omitting the --reference link is fine, it'll just take longer to initialize LLVM.

Run / test:

gnnc-bench --model gcn --dataset cora            # compile + run + compare
lit test/                                      # MLIR pipeline + pass lit tests
pre-commit run --all-files                     # ruff + hygiene + lit + pytest (in-container)

About

A learning exercise for MLIR infrastructure.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0