Backends¶

pivoterpy dispatches clique counting to one of two supported backends selected by backend= on pivoter: rust (default) or python.

Rust (backend="rust")¶

• What it is: A PyO3 extension module (pivoter_rust) built from the crate in this repository, using Rayon for parallelism when procs > 1.
• When to use it: Default choice for speed on non-trivial graphs; same resolutions as Python (global, vertex, edge).
• Requirement: A wheel or a local maturin develop / maturin build install so import pivoter_rust succeeds.

If the extension is missing, you will see an ImportError pointing at pivoter_rust. Fix by installing a wheel or building from source (Installation).

Implementation notes (high level): degeneracy ordering, neighborhood representations tuned for the Rust code paths, and parallel work over the search structure as described in the upstream Pivoter line of work—not identical line-by-line to the Python kernel, but intended to match counts for the same graph and parameters.

Python (backend="python")¶

• What it is: Pure Python kernel: degeneracy ordering, Succinct Clique Tree (SCT), and optional multiprocessing when procs > 1.
• When to use it: No Rust toolchain, debugging, teaching, or double-checking results on small instances.
• Tradeoff: Much slower on large or dense graphs than Rust, but no compiled extension.

Choosing a backend¶

Outputs are meant to be consistent between backends for the same graph and resolution (up to the usual integer arithmetic and any min_k / max_k truncation rules).

CUDA¶

The repository contains experimental CUDA wrapper code (CUDAKernel / _cu_backend), but the public pivoter constructor only accepts python and rust. There is no supported backend="cuda" in the current API. Treat CUDA-related modules as internal or future work.

Troubleshooting¶