Expand description
Two recordings compared (ADR 0058, decision 12): each span’s and
phase’s spread, the burns, the gaps and the counters, before and after
— kui-trace diff A B, with --fail-over a CI’s threshold.
Spans are matched by their name’s text and origin, never by id: ids and name ids differ between two recordings of one session. What a span or a gap costs is a frame’s share (its total over the frames), so two recordings of one script compare even when one ran a few frames longer.
Report::worst_growth reads only the rows that say what a frame
costs and that were big enough before to say it (Unit::counts):
a time of at least NOISE_US (below it, a timer’s jitter is a large
share of nothing), and a rate of at least one a frame. A maximum, a
count, a share and vsync’s wait are shown and never counted: one
outlier frame, a recording’s length and a faster render (which waits
longer) are not regressions.
Structs§
- Report
- What changed from
atob.
Enums§
- Unit
- What a row counts.
Constants§
- MARK
- The growth
Report::textmarks. - NOISE_
RATE - Below this many a frame before, a rate row’s growth is noise.
- NOISE_
US - Below this many µs before, a time row’s growth is noise.
Functions§
- diff
aagainstb.