giotto-ph

giotto-ph computes Vietoris–Rips persistence and does essentially nothing else, which is the right scope for a backend. What makes it worth a note of its own is where it came from: it is a reimplementation of Morozov and Nigmetov’s lock-free Ripser, carrying the apparent-pairs optimization introduced in Ripser 1.2 and an improved version of GUDHI’s edge-collapse preprocessing.

Three lineages, one binary. ripser-cpp contributed the algorithmic core and the apparent-pairs shortcut; a separate research line contributed the lock-free parallelization; gudhi contributed the collapse. This is the sense in which the field’s engines are not really independent — the implementations profiled in this corpus fork and re-merge, and a benchmark comparing “Ripser versus giotto-ph” is comparing two points on one line of descent rather than two designs.

Where the speed comes from

Two distinct places, and they compose.

Parallelism. Apparent-pair detection and the coboundary matrix reduction are both distributed across threads. That matters because Ripser’s own design is what made this possible: the implicit coboundary representation means threads are not contending over a shared explicit matrix, and the lock-free structures (the vendored junction and turf) handle what contention remains.

Edge collapse. A preprocessing pass that removes edges the Rips filtration’s persistence does not depend on, shrinking the complex before any reduction happens. This is the one that changes what is feasible rather than how fast the feasible thing runs: H₂ on a few thousand points is often the difference between a job that finishes and one that does not, and collapse is usually how it finishes.

It also supports distance-to-measure reweighted filtrations, which is the standard answer to Rips persistence being fragile against outliers — a single stray point can create long spurious bars, and reweighting by local density suppresses them.

Building it is the hard part

The C++ engine arrives as git submodules — pybind11, junction, turf — and an sdist does not carry submodules, so the build must reach for git and therefore for the network. That is the constraint holding giotto-ph-recipe out of a public channel, since channels build offline; it is shared with pyflagser-recipe and rivet-recipe and it is one problem rather than three. Vendoring the submodules into a source archive would solve all of them.

Two smaller build facts, both symptoms of vendored code aging at a different rate than its host: the sources need a force-included <cstdint> because GCC 13 dropped a transitive include they relied on, and cmake has to be pinned below 4 because junction’s cmake_minimum_required predates the floor CMake 4 enforces.

In this corpus

giotto-ph-environment stages giotto-ph-recipe, verified green on linux-64, which is what caps it at L1. AGPL-3.0 is redistributable copyleft, so licensing is not the blocker — publishing the recipe would lift the fixture to L3 once the network dependency is resolved. The network clause that comes with the AGPL is discussed under giotto-tda, which heads this family and consumes this engine for its persistence. The technique is persistent-homology.