Docs menu: Rendering backends Documentation home
On this page Support matrix Notes on each renderer SVG CPU GPU Choosing a renderer

Rendering backends

What the SVG, CPU and GPU renderers each support today.

MSX has three renderers. They share the scene model, so one file can go to any of them, but they do not draw every feature. This page lists what each one does today.

Renderer Crate Output Needs
SVG msx-render-svg SVG text nothing
CPU msx-render-cpu PNG, animated GIF nothing
GPU msx-render-gpu PNG, animated GIF the gpu Cargo feature and a GPU adapter, which can be a software one

The CPU renderer rasterizes with tiny-skia, and evaluates signed distance fields and splats per pixel. The GPU renderer tessellates shapes with lyon and draws through wgpu.

Support matrix

Feature SVG CPU GPU
Rect, circle, ellipse, line, polyline, polygon, path yes yes yes
Text yes yes, DejaVu Sans only yes, DejaVu Sans only
Image yes yes yes
Group, use yes yes yes, but a use of an sdf node, splat or image draws nothing
Group style reaching its children yes no no
Linear and radial gradients interpolated interpolated interpolated
Conic gradient not drawn yes yes
Shader fills fallback_color fallback_color runs the WGSL
Signed distance fields omitted yes yes
Gaussian splats approximated yes yes
Layer opacity and blend modes yes, see below yes yes
Layer effects all five all five blur only
Layer clip not enforced not documented not documented
Layer clip_path clipPath element antialiased mask mask texture, hard edges
Layer mask mask element buffer scaled by luminance or alpha texture multiplied by luminance or alpha
Viewbox yes no no
Keyframe animation static frames GIF GIF

Where a gradient is not interpolated, the paint is the average of its stops. That applies to gradient fills of sdf nodes and splats on the CPU and GPU renderers.

Notes on each renderer

SVG

  • Every style key is written into the output, which makes SVG the reference for what a style means.
  • Layers become groups with mix-blend-mode and a generated filter chain for effects. The blend modes subtract and divide have no CSS counterpart and are written as normal.
  • An sdf element is replaced by a comment. Splats become a three stop radial gradient that reaches out to 2.4 sigma.
  • A conic gradient has no SVG form. A comment is written and the fill is not drawn.

CPU

  • The rasterizer applies stroke width, dashes, caps, joins, the miter limit, the fill rule, fill_opacity and stroke_opacity.
  • Compositing for layers uses the shared blend functions in msx-render-core, which cover all twelve modes.
  • Text is drawn as glyph outlines from an embedded DejaVu Sans face, through the same path pipeline as other shapes. font_family is ignored and there is no OpenType shaping. See Text.
  • Linear and radial gradients are tiny-skia shaders mapped through the bounding box. Conic gradients have no tiny-skia shader, so the renderer rasterizes the shape coverage and colors each covered pixel itself.

GPU

  • Elements draw in document order. Neighboring elements of one kind (vector shape, shader fill, sdf node, splat, image) share one draw, and a new draw starts when the kind changes.
  • Shapes, text and gradients share one vertex batch. A gradient is evaluated per fragment from its bounding box, so it interpolates on fills and strokes, and a gradient holds at most 32 stops. A longer list is resampled to 32 evenly spaced stops, which softens any hard stop past that limit.
  • Text is drawn from the same embedded DejaVu Sans outlines as on the CPU renderer. Shape and text edges are not antialiased.
  • A shader fill runs for real on shapes, SDF nodes, splats and shapes inside layers. Shader strokes fall back to the flat color, and a shape's opacity is not multiplied into the shader output.
  • When the shader file cannot be read, the GPU renderer paints fallback_color and prints a line on the standard error stream.
  • Layers composite in an isolated buffer with all twelve blend modes, and blur is the one effect implemented.
  • A machine with no adapter gets a clear error from the CLI and not a crash.

Choosing a renderer

  • Use SVG for output that has to stay vector, and for text that should use the viewer's own fonts.
  • Use the CPU rasterizer for PNG and GIF output that needs no special hardware. It covers text, gradients including conic, SDF shapes, splats, layers and all five effects.
  • Use the GPU renderer when a scene has shader fills, or when blur on large layers needs speed.

The samples page shows the same set of files through these renderers, and the build report lists the result of each CI run.