Fidget

Fidget is experimental infrastructure for complex closed-form implicit surfaces.

At the moment, it is quietly public: it's available on Github and published to crates.io, but I'd appreciate if you refrain from posting it to Hacker News / Twitter / etc; I'm planning to write an overview blog post and put together a few demo applications before making a larger announcement. If you have an overwhelming urge to talk about it, feel free to reach out directly!

The library contains a variety of data structures and algorithms, e.g.

• Manipulation and deduplication of math expressions
• Conversion from graphs into straight-line code ("tapes") for evaluation
• Tape simplification, based on interval evaluation results
• A very fast JIT compiler, with hand-written aarch64 and x86_64 routines for
  • Point-wise evaluation (f32)
  • Interval evaluation ([lower, upper])
  • SIMD evaluation (f32 x 4 on ARM, f32 x 8 on x86)
  • Gradient evaluation (partial derivatives with respect to x, y, and z)
• Bitmap rendering of implicit surfaces in 2D (with a variety of rendering modes) and 3D (producing heightmaps and normals)
• Meshing (using our own implementation of the Manifold Dual Contouring algorithm)

If this all sounds oddly familiar, it's because you've read Massively Parallel Rendering of Complex Closed-Form Implicit Surfaces. Fidget includes all of the building blocks from that paper, but with an emphasis on (native) evaluation on the CPU, rather than (interpreted) evaluation on the GPU.

The library has extensive documentation, including a high-level overview of the APIs in the crate-level docs; this is a great place to get started!

At the moment, it has strong Lego-kit-without-a-manual energy: there are lots of functions that are individually documented, but putting them together into something useful is left as an exercise to the reader. There may also be some missing pieces, and the API seams may not be in the right places; if you're doing serious work with the library, expect to fork it and make local modifications.

Issues and PRs are welcome, although I'm unlikely to merge anything which adds substantial maintenance burden. This is a personal-scale experimental project, so adjust your expectations accordingly.

Demo applications

In the repository on Github, there are two demo applications:

• demo does bitmap rendering from the command line
• viewer is a minimal GUI for interactive exploration

These are deliberately not published to https://crates.io, because they're demo applications and not complete end-user tools.

Support matrix

At the moment, Fidget supports a limited number of platforms:

Explanation of keys

Support tiers represent whether maintainers will be able to help with platform-specific bugs; for example, if you discover an aarch64-pc-windows-msvc-specific issue, expect to do most of the heavy lifting yourself.

CPU requirements

aarch64 platforms require NEON instructions and x86_64 platforms require AVX2 support; both of these extensions are nearly a decade old and should be widespread.

Disabling the jit feature allows for cross-platform rendering, using an interpreter rather than JIT compilation. This is mandatory for the wasm32-unknown-unknown target, which cannot generate "native" code.

Similar projects

Fidget overlaps with various projects in the implicit modeling space:

• kokopelli: script-based CAD/CAM in Python*
• Antimony: CAD from a parallel universe*
• libfive: Infrastructure for solid modeling*
• Massively Parallel Rendering of Complex Closed-Form Implicit Surfaces (MPR)*
• ImplicitCAD: Powerful, Open-Source, Programmatic CAD
• Ruckus: Procedural CAD For Weirdos
• Curv: a language for making art using mathematics
• sdf: Simple SDF mesh generation in Python
• Forged Thoughts: A Modeling & Rendering Language in Rust
• Raumkuenstler: Node-based modeling with an LLVM-powered JIT compiler
• Probably more; PRs welcome!

*written by the same author

(the MPR paper also cites many references to related academic work)

Compared to these projects, Fidget is unique in having a native JIT and using that JIT while performing tape simplification. Situating it among projects by the same author – which all use roughly the same rendering strategies – it looks something like this:

Fidget's native JIT makes it blazing fast. For example, here are rough benchmarks rasterizing this model across three different implementations:

libfive and Fidget are running on an M1 Max CPU; MPR is running on a GTX 1080 Ti GPU. We see that Fidget's interpreter is slightly better than libfive, and Fidget's JIT is nearly competitive with the GPU-based MPR.

Fidget is missing a bunch of features that are found in more mature projects. For example, it only includes a debug GUI, and its meshing is much less battle-tested than libfive.

License

© 2022-2024 Matthew Keeter
Released under the Mozilla Public License 2.0