Design, run, and interpret production-shaped GPUI Criterion benchmarks.
npx skills add https://github.com/zed-industries/zed --skill gpui-bench
Please help me install this Agent Skill. Skill: gpui-bench Source: https://github.com/zed-industries/zed/tree/main/.agents/skills/gpui-bench Command: npx skills add https://github.com/zed-industries/zed --skill gpui-bench
Benchmarking GPUI code is difficult because naive Criterion setups often enable test-only features, use fake executors, or skip the rendering pipeline, producing numbers that do not reflect real UI responsiveness. This Skill enforces production-faithful benchmarking so that frame drops, hangs, and foreground stalls can be reproduced and measured reliably.
test-support occurrences and uses narrow bench-support seams.bench_renderer benchmark that drives a concurrent producer past queue capacity, captures p99 foreground latency and frame overruns on baseline and candidate commits, and reports whether the fix actually improves responsiveness.Use the gpui-bench skill to write a production-shaped Criterion benchmark that reproduces the reported scrolling hitch and reports before/after foreground latency and frame-budget overruns.
High-intent search queries and answers about installing and using this skill.
Use `#[gpui::bench]` with `bench_batched_task` or `bench_renderer`, drive more work than the queue capacity, keep a concurrent producer refilling it, and include a responsiveness signal such as a spinner or scrolling viewport. Assert final work counts and ordering outside the timed region.
`bench_iter` measures synchronous compute work, `bench_task` measures an async GPUI Task to completion including foreground polls and draws, and `bench_renderer` updates an entity in a window and submits a frame through the headless renderer. Choose based on whether the operation is sync, async, or render-driven.
GPUI performs CPU-side window work on all platforms, but only macOS has a headless Metal renderer that submits real GPU frames. On Linux and Windows the present step discards the scene, so results must state this limitation and supplement with app automation or platform profilers.
Enabling `test-support` brings in fake executors, deterministic clocks, and test-only CRDT settings that change scheduling behavior and invalidate responsiveness measurements. Use the `bench-support` feature with narrow production-faithful seams instead, and verify with `cargo tree -e features`.
Keep the benchmark code identical between baseline and candidate, build both with the same optimized profile and lockfile, run on the same machine, alternate runs to reduce thermal bias, and report raw percentiles plus confidence intervals rather than only percentage changes.
cnwzhuDiagnose GPUI development errors and derive actionable fixes for Rust applications.
kitcafJudge experimental results against intended claims and route pivot, supplement, or confirm actions.
BumpyClockProfiles native macOS/iOS apps using Instruments/xctrace Time Profiler to identify bottlenecks and hot paths.
benchflow-aiDetect statistically significant trends in time series data using linear regression and Mann-Kendall tests.