mirror of
https://github.com/zed-industries/zed.git
synced 2026-10-19 11:04:06 +00:00
4f6d97b9de
Frame and hang telemetry judged every frame against a fixed 24 ms budget. That's too lax on 120–360 Hz displays and too strict at 30 Hz, so the numbers couldn't be compared across users or catch regressions on fast displays. This judges each frame against its window's display refresh interval (from #65059), as measurement version 3. **Frame timing.** A frame's work runs from the later of its frame signal and its first invalidation until its submission to the platform begins, on every platform, so time spent waiting for a late frame signal or inside the submission doesn't count. A frame is late when its work exceeds one refresh interval, but never less than 8.33 ms (one refresh at 120 Hz), since frames rarely fit a single refresh on faster displays yet; on displays whose refresh interval is unknown (X11, web) the budget is 16.7 ms. Missed refreshes still count against the real interval. **Frame Duration Report**, per window and refresh interval: on-time, late, and unmeasured frame counts and missed refreshes, overall and for frames that responded to input; work (as a percentage of the refresh interval) and draw-time buckets, computed from hdrhistogram histograms; GPU and CPU. **Hang Incidents.** Threshold incidents (one event of 100 ms or more) are unchanged. Budget incidents are split by trigger: - `late_frame`: a presented frame over its budget, listing the events that overlap that frame's work; - `blocking_work`: in an interval with no measured frame, a single event of at least the frame budget (16.7 ms in release builds), long enough to delay a frame. Total foreground time across an interval no longer counts, since gaps between short work don't block frames. Draw contributors carry request-layout, prepaint, and paint times, and any entity view phase whose own work (excluding nested views) takes 0.5 ms or more becomes a contributor named by its view type. Events also report counts per trigger, incidents during active use (within a minute of a key press, click, or scroll) with active seconds for rates per active hour, stall buckets, and contributor totals summed across every incident (`view:<type>:<phase>`, `action:…`, `task:…`). **Verification.** With a synthetic render delay on a 240 Hz display, on-time frames went from 99.9% at no delay to 0% at 6 ms. In a real Zed on Linux at 360 Hz, scrolling a markdown preview with #63138's per-frame code-block highlighting reintroduced roughly doubled missed refreshes per frame (3.2 to 6.0) and attributed the extra time to `MarkdownPreviewView`'s request-layout phase. Profiler overhead was +2.3% of main-thread time in a 1,500-view GPUI stress test and about 0.3% in the Zed scroll run. Not yet verified: Windows, and input-driven frames from real use. Time spent inside the platform submission (GPU, compositor) isn't measured on any platform. Release Notes: - [GPUI] Improved frame and hang telemetry to measure frames against each window's display refresh rate |
||
|---|---|---|
| .. | ||
| src | ||
| Cargo.toml | ||
| LICENSE-GPL | ||