Community project: pure Rust port of Box2D v3 (box2d-rust)

[larsbrubaker]

Hi Erin and team,

Just wanted to share that we've completed a pure Rust port of Box2D v3: box2d-rust.

• GitHub: https://github.com/larsbrubaker/box2d-rust
• crates.io: https://crates.io/crates/box2d-rust
• Live WebAssembly demo (mirrors the samples app, all 13 categories): https://larsbrubaker.github.io/box2d-rust/

A few notes on the approach:

• It's a strict behavioral port of the pinned v3.1.1 source (56edae7) — same algorithms, same f32 arithmetic, no reordered floating-point operations.
• The hand-rolled b2Atan2 / b2ComputeCosSin were ported bit-for-bit, and the FallingHinges determinism test produces bit-identical results against the C build.
• The C test suite was ported alongside the code — 129 tests, green in both standard and BOX2D_DOUBLE_PRECISION (large world) modes.
• Not ported by design: the threading/task system (the port is serial), the global world registry, and the arena allocator (replaced with Rust Vec-indexed pools).

Thank you for Box2D and for keeping the C source so readable and well-commented — Erin's explanatory comments made an exact port feasible, and we preserved them throughout. MIT licensed, credit to the original throughout.

Feel free to close this issue — it's just an FYI/thank-you. Happy to add a link wherever it's useful, or answer any questions.

— Lars Brubaker

[erincatto]

Cool project!

The browser demo is not showing any shapes in Chrome on Windows.

Did you compare the performance of the Rust version versus the C version?

[larsbrubaker]

Thanks for taking a look — you caught a real bug. v1.2.0 added the b2Profile timings, and std::time::Instant::now() panics on wasm32-unknown-unknown, so every world.step trapped before anything drew. Fixed (the timer uses performance.now() on wasm now) and the demos are back up: https://larsbrubaker.github.io/box2d-rust/

On performance: I ported your benchmark app 1:1 (examples/benchmark in the repo) and ran both engines single-threaded on the same machine (C with -w=1, MSVC /O2 vs rustc release, interleaved per scene, min of runs). The port started the day at ~1.9x slower geomean; two fixes later it's at ~1.25x (range 1.13–1.47 across your 10 scenes). The two finds: I was accidentally running your B2_VALIDATE-only island validation in release builds (quadratic on island-churning scenes — spinner went 2.73x → 1.21x), and I hadn't ported the 4-wide b2FloatW contact solver, so graph colors were solving scalar. Both are fixed; the wide kernels are lane-exact ports, so determinism held bit-for-bit throughout (FallingHinges still hashes to the same value as your build). Full table and remaining roadmap are in the README's Performance section.

The port is serial by design for now — no task system — so the numbers above are your serial fallback vs my serial build. Multithreading is the obvious next frontier.