Members
countreadonly
maxMsreadonly
meanMsreadonly
namereadonly
p95Msreadonly
shareOfFramereadonly
This span's total as a fraction of the frame track's total, or null where
that fraction is not a share of anything.
Read per track, never summed across them. The Engine phases tile the
engine step and the Terrain phases tile the Engine's terrain phase, so
one sum over everything counts the streamer twice — which is exactly the
arithmetic that makes a flame chart lie. frame is the denominator and not
a member of either set: it is the wall-clock period the whole tiling sits
inside, and it is much the larger of the two — 17.9 ms against a 4.4 ms
engine on the summit recording, so the phases account for a quarter of it
and the rest is the GPU and the compositor, which nothing here measures.
Null when the span's own occurrences overlap each other. A worker job's queue wait runs concurrently with every other job's and with the frames it spans, so its total divided by the frame total is a real ratio and a meaningless share — it came out at 58,767% on a saturated pool, which is a number that reads as a bug rather than as a fact about concurrency. The overlap test is what separates the tracks that tile a serial timeline from the ones that describe other threads, and it is empirical rather than a list of track names, so a track added later is classified correctly without anybody remembering to.
Null, too, when the total exceeds all the frame time in the window. A
span that occurs once has nothing to overlap, so the test above passes it
— and navigation to first light is one entry covering the whole boot,
which printed 320%. A contained serial span cannot exceed 100% by
construction, so the ceiling is the definition rather than a heuristic.