Skip to main content
Open Candle work benchmark from the demo home screen (/demo/candle-work). The two panels receive the same moving viewport, live candle and changing price range. Switch between sharp/rounded bodies and 40/200 candles to compare the current rendering with a one-pass worklet prototype. Both panels keep four batched paths: up/down bodies and up/down wicks. This experiment does not replace batches with individual candle draw calls.

​
Run the measurement

Press Run paired benchmark. The preview pauses while a UI worklet measures three versions with real Skia builders:
  1. Main: build geometry objects, then scan bodies and wicks for each color.
  2. One pass: calculate each candle and emit directly into the four builders, reusing rectangle input objects.
  3. Production: use the same body-path helper as useCandlePaths. It keeps the geometry and four scans, but reuses one plain rectangle within each body-path rebuild instead of calling Skia.XYWHRect for every sharp body.
The benchmark uses 6,000 history candles and windows containing approximately 40, 200 and 1,000 candles, plus partially visible/live candles. It warms up each version, rotates their order over seven rounds, and reports median milliseconds per rebuild. Each round rebuilds 40 times. Show raw samples exposes every round for comparison. Native path commands are compared outside timed loops. The times include geometry, builder calls and four detach() operations. They exclude the engine, rendering, GPU, React and mapper scheduling. The preview is a visual comparison; it is not a display-FPS measurement.

​
What this experiment shows

On the dedicated iOS simulator development build, reusing rectangle inputs alone provided essentially the same improvement as the full one-pass rewrite for sharp candles. Rounded candle timings were essentially unchanged in the production-helper rerun. The larger rewrite is therefore not justified by timing alone. The smaller rectangle-reuse change is now integrated in the production candle hook. The one-pass rewrite remains a demo-only prototype. Rectangle inputs are local to each worklet invocation; mutable scratch is not shared between React and the UI runtime. No new public prop or native helper is introduced. Physical device/release validation remains outstanding. These results do not establish CPU, battery, heap retention or whole-chart FPS improvements.