[dXnSix]

A 50-iteration improvement loop on playback performance, memory and allocation cost, and time to first frame: 41 changes kept, 9 tried and reverted. Each experiment is recorded under .agentx/experiments/2026-10-27-performance-improvements/.

Results

vector-animation-check bench artifacts-temp/corpus (the 53 corpus files, release build, median of 3):

Measure	Before	After	Change
Time to first frame (sum)	253.5 ms	93.8 ms	−63%
— read	73.9 ms	62.6 ms	−15%
— compile	179.6 ms	31.2 ms	−83%
Allocations at load	2.87 M	0.77 M	−73%
Bytes allocated at load	1,168 MB	382 MB	−67%
Held after load	99.7 MB	67.4 MB	−32%
frame_at a frame (sum)	354 µs	201 µs	−43%
Allocations a frame	7,789	2,908	−63%
Engine-thread work a frame	833 µs	476 µs	−43%
Graph breakpoints	115,936	53,982	−53%
Graph nodes	5,026	4,972	−1%
Canvas ops (save/restore aside)	45,331	44,936	−1%

An app's imported compiled scene (Source::Json) reads 30% faster: it is no longer written back to text and parsed again.

What changed

• Player load: the plan is compiled once on load, for the view, and the first drawing takes it (it was compiled twice). The loaded scene is shared with what plays it instead of cloned.
• Plan compile: moving layers' matrices are multiplied directly instead of sampling probe frames. Outline switches are bisected on the easing alone. Plan images use the fast PNG encoder. Sampled parts are joined without copying them, which was quadratic. A still scene is recognized without sampling its frames. Graph nodes use a fast hash.
• Graph size: a repeating track (looping CSS animations over their common timeline) is one repetition read modulo its period. Sampled values keep only the samples a line does not stand for; a gradient's reach is sampled within 0.1%. Single-valued tracks are constants.
• Canvases: saves that keep nothing are dropped, and consecutive still translations are combined.
• Core frame_at: fill and stroke take the shape's path rather than copies, there is no merged copy of a single outline, and rectangles, ellipses and stars allocate their outlines once.
• Renderer: box blurs run over pixel chunks and blur columns a row at a time; blurs are written back in a flat pass. Layers composite only where they draw, and normal-blend layers use a direct source-over loop that computes exactly what tiny-skia's pipeline computes.
• Importers:
  • SVG: keyframe values are parsed once, CSS rules are indexed by id, class or tag, class matching allocates nothing, and cascaded declarations are shared.
  • Lottie: pointer keys are escaped only when needed, baked paths are moved instead of cloned, merge buffers are presized, and a JSON value is copied only where expressions are baked into it.
  • Compiled scenes: parsed in one pass. New APIs compiled::from_value and dxn_vector_animation_lottie::read_value read JSON values directly.
• Measurement: new vector-animation-check bench <dir> [--runs n] [--files] [--values] with a counting allocator.

For review

• Output unchanged:
  • The reference renderer's 4,629 frames of the corpus (4 fps, 160 px) are byte-identical to main.
  • Each corpus file's compiled scene JSON and warnings are byte-identical to main.
  • The corpus plan-versus-reference test (DXN_CORPUS) gives the same per-file results.
  • The direct compositing loop has a unit test against tiny-skia's draw_pixmap over all 8.4 million (channel, alpha, destination) combinations, at four opacities.
• Behaviour changes:
  • The player's choice between the native graph and sampling the whole scene now comes from the compile for the view (as measured, or at the scene's own size) at the display's pixel ratio. It used to come from a separate compile at the scene's size and 1 pixel a point.
  • JSON-value sources are not checked against Limits::max_bytes, since the value is already in memory. Node, keyframe, depth and point limits still apply through Scene::check and the converters.
• Tests: new test groups cover each new mapping: repeating_tracks, sampled_values, constant_tracks, canvas_saves, still_translations, still_parts and json_values in the player, plus new tests in css_selectors and compiled_checks.
• Validation: all of AGENTS.md's commands pass: fmt, clippy for the core and the player, cargo test --all-features, the player's tests, and the wasm check.
• Not measured: NSH's own evaluation of the graph on platform UI threads. Fewer breakpoints and ops should help there.