Skip to content
ONDA

Backends

ONDA ships two rendering backends. Both consume the same scene graph; they differ in capability and determinism. Choose with --backend auto|vello|cpu.

Vello (GPU) — the vector backend

Section titled “Vello (GPU) — the vector backend”

onda-vello is ONDA’s GPU-native vector backend. It maps an ONDA scene onto a vello::Scene and rasterizes it on the GPU (compute) to an RGBA framebuffer. It does true vector rendering:

  • Anti-aliased fills and strokes
  • Real rounded rectangles (cornerRadius)
  • Arbitrary Bézier paths (<Path> / ShapeGeometry::Path)
  • Linear & radial gradients (gradient paint)
  • Clip regions (the clip prop — any geometry as a mask)
  • Native per-glyph vector text — glyph outlines drawn through Vello’s glyph runs, resolution-independent and ready for per-character/kinetic animation

It renders offscreen and reads the frame back into a framebuffer, so the existing GIF/MP4 encoders apply unchanged.

Requires a GPU adapter. --backend vello errors if none is found; --backend auto falls back to CPU.

CPU — the deterministic reference

Section titled “CPU — the deterministic reference”

onda-renderer is the CPU reference rasterizer, built on tiny-skia (the Skia raster pipeline) + kurbo — no GPU, deliberately dependency-light, so the scene-graph → pixels contract can be pinned down and tested. It is the correctness oracle the GPU backend is checked against, and it is far from a toy: across the bulk of the surface it is byte-identical to Vello.

The CPU backend draws, matching Vello pixel-for-pixel:

  • Anti-aliased fills and strokes (cap / join / dash)
  • Real rounded rectangles (cornerRadius)
  • Arbitrary Bézier paths (<Path>)
  • Linear & radial gradients
  • Native text and images
  • The full per-pixel effect chain — blur, directional blur, bloom, color-grade, grain, duotone, posterize, vignette, chromatic aberration, goo, chroma-key, backdrop-blur, and mattes

A narrow set is GPU (Vello) only — the CPU reference skips it or degrades:

  • Rotation (rotation / rotation3d) — ignored
  • Clipping (clip) — ignored; use a matte instead, which the CPU honors
  • Blend modes (blendMode) — composite as Normal
  • Per-run rich text and letterSpacing — drawn in the node’s base style
  • 3D out-of-plane tilt and extrude — degrade to a flat 2.5D composite
  • light-wrap, the cinematic finish / linear and motion blur — export/native only on either backend

Which to use

Section titled “Which to use”
You want…Use
Rotation, clipping, blend modes, true 3D / extrudevello
Bit-identical, reproducible output across machinescpu (bundled font)
A sensible default that picks the GPU when presentauto (the default)

The determinism caveat

Section titled “The determinism caveat”

The CPU backend is bit-identical across runs and machines when using the bundled default font — which is why it’s the reference oracle and the right pick for reproducible pipelines. Passing --system-fonts uses the host’s fonts and makes CPU output machine-dependent.

The GPU (Vello) backend is not guaranteed bit-identical across different GPUs and drivers. If you need reproducibility, render on the CPU; if you need rotation/clip/blend/3D or have a GPU, render with Vello.