cheatah-plot

Dead-simple cross-platform plotting for cheatah. Hand it some numbers, get a beautiful plot — as a PNG, or as pixels you read back and stream anywhere.

import plot

let xs = # ... an ndarray of x values
let ys = # ... an ndarray of y values
let fig = plot.line(plot.new_figure(), xs, ys)
fig = plot.title(fig, "my line")
plot.save(fig, "line.png")     # a real PNG, rendered by cheatah-plot

The model layers (plot.scale / plot.color / plot.series / plot.stats / plot.figure) are pure cheatah on ndarray — fully testable with no device. The renderer is cheatah-plot's OWN layer built directly on cheatah-gpu's raw Vulkan/Metal forwarders: compute-shader rasterization into an offscreen framebuffer you save to a file or read back (to stream a live plot to a website, no JavaScript). Anything that is linear algebra rides the stdlib linalg — and, for device-resident arrays, cheatah-gpu-linalg's overloads. Headless first; windowing is roadmap. See docs/DESIGN.md for the architecture.

Install

biome add cheatah-plot        # fetches cheatah-plot + its cheatah-gpu stack

Then import plot and draw. scripts/doctor.sh verifies your machine is ready for the GPU path (Vulkan loader, slangc); the model layers and the CPU render path need no GPU at all.

Modules

| import … | what | status | |------------|------|--------| | plot | the umbrella — the whole plot.* surface in one import | working | | plot.scale | the pure axis geometry: ranges, "nice" linear + log ticks, data→pixel | working | | plot.color | colours, categorical palettes, viridis/magma/coolwarm colormaps | working | | plot.series | the mark data + style value every plot call builds | working | | plot.stats | histogram binning, the linalg.lstsq fit line, error magnitudes | working | | plot.figure | the figure model: subplots, axes, the fluent building API | working | | plot.renderer | the 2D renderer: CPU reference + bit-exact emulated-Metal + Vulkan lanes | working | | plot.window | windowing + presentation | roadmap |

Status

The whole stack works end to end: build figures from cheatah, render them to PNGs — on the CPU reference rasterizer everywhere, and through the GPU lanes (Vulkan, emulated Metal) when a device is present, with render() falling back cleanly. Fifteen example programs draw the gallery (examples/purr_plot/). plot.window (presentation) is the next layer.

<!-- coverage:start --> | Metric | plot package | |--------|--------------| | Lines | 100.00% (394/394) | | Functions | 100.00% (121/121) | | Regions | 97.51% | | Branches | 89.74% | <!-- coverage:end -->

License

MIT — © 2026 BigBrain LLC (Joshua Doucette, on its behalf). See LICENSE and NOTICE.