`fastcore` (Rust)¶

The core crate. Everything the Python and R bindings do is implemented here; they are thin adapters that translate each language's idioms into the core's index-based API.

Modules¶

Module	What it does
`dag`	Traversal and geometry on rooted trees: geodesic distances, linear segments, Strahler index, twig pruning, node classification, connected components, synapse flow centrality, cycle detection.
`topo`	Repairing fragmented skeletons: `stitch_fragments` finds the minimal-length edges that reconnect the pieces (optionally preferring fragments of similar calibre), `reroot_rewire` re-derives the parent vector afterwards.
`mesh`	Triangle meshes as vertex graphs: `mesh_connected_components`, plus a parallel Dijkstra/BFS behind `geodesic_matrix_mesh`, `geodesic_nearest_mesh`, `geodesic_farthest_mesh` and — for arbitrary (cyclic) graphs given as an edge list — `geodesic_matrix_graph`.
`nblast`	The NBLAST pipeline — `build_index`, `score_pair`, `nblast_query_target`, `nblast_allbyall`, `nblast_pairs`, plus the `Smat` scoring matrix and `Opts`.
`synblast`	Synapse-based NBLAST: `synblast_query_target`, `synblast_allbyall`.
`matches`	Pulling the top matches back out of a score matrix — `top_matches` (top-N), `matches_above` (absolute threshold or a percentage band around each group's best), `count_matches` — without copying or transposing a matrix that may be tens of GB.
`cmtk`	CMTK spatial transforms: `Registration::from_path` reads a `*.list` registration (12-DOF affine + cubic B-spline warp), `transform_points` / `inverse_transform_points` apply it. Matches CMTK's `streamxform` to ~4e-7 without needing CMTK installed.
`elastix`	Elastix spatial transforms: `ElastixTransform::from_path` reads a `TransformParameters` file and the initial-transform chain hanging off it (affine / Euler / similarity / translation, plus cubic B-spline warps), `transform_points` / `inverse_transform_points` apply it. Matches `transformix` to 5e-7 without needing Elastix installed — and adds an inverse, which Elastix itself cannot compute. `probe_invertible` answers whether a file inverts without reading its coefficients, ~20x faster than a full parse.

See Concepts › Rooted trees and Concepts › NBLAST for the ideas behind them, and the capability matrix for how each module maps onto the Python and R functions.

Using the crate¶

fastcore is not published to crates.io, so depend on it via git:

[dependencies]
fastcore = { git = "https://github.com/schlegelp/fastcore-rs" }

API reference¶

There is no docs.rs page (see above). Build the reference locally:

git clone https://github.com/schlegelp/fastcore-rs
cd fastcore-rs
cargo doc -p fastcore --open

Shape of the API¶

The core is index-based and ndarray-typed. Where the Python bindings accept node_ids and parent_ids with arbitrary IDs and map them for you, fastcore expects the mapping to have happened already:

A tree is an ArrayView1<i32> of parent indices, with roots encoded as negative values.
Edge weights, coordinates and masks are passed as separate arrays.
NBLAST takes prepared point clouds (build_index) rather than dotprop objects.

That mapping step is exactly what navis_fastcore's internal _ids_to_indices and nat.fastcore's public node_indices exist to do.

Parallelism¶

dag and the NBLAST modules parallelise with rayon. Opts::threads caps the pool for NBLAST; the Python bindings surface this as n_cores and release the GIL around the call.