`make bench` printed two reports and left nothing behind but a text file, so a
number could be compared with another number from the same afternoon and not
much else. It now prints three and writes the run's own directory: `report.txt`,
`manifest.json` and `summary.json` beside the `trace.pftrace`.
Section [3/3] is the counters — the render stages, the GPU passes, and the
per-frame work census. A counter nobody sampled prints MISS rather than
`0.00`, and the counters and spans this build declares but this run never
reached are named at the end, so a hole in the instrumentation cannot read as a
fact about the workload. GPU passes overlap and resolve late, so they are never
summed.
`manifest.json` says what produced the number: the revision, the dirty flag,
the toolchain, the profile cargo actually built under, the locked `bevy` and
`wgpu`, the adapter and whether its device can time a pass, the window, the
pools, the env toggles and the cache state. The build half is baked in at
compile time, because asking `git` at runtime names the tree the process is
standing in rather than the build. A fact nobody could read is `null`, never a
plausible default.
Reading that report, `preparing match audio` was the largest stage of every
load — one row holding one number, which cannot tell a thousand cheap byte
loops from five hundred external processes. It is the second. Section [1/3] now
splits the stage by the decoder each clip took, in worker time, and counts what
each one prepared, failed, held and left resident.
The T5 XWMA decode goes through the artifact cache under `xwma_pcm`, keyed on
every input the decode reads, so a clip is decoded once per machine rather than
once per launch. A cold load still had to pay the first time: 515 `ffmpeg`
processes, 18.9 s of worker time, almost none of it decoding. A worker now
takes the jobs already queued and answers up to 64 clips through one process —
every clip its own input and its own output, nothing concatenated — which is
1.3 s for the same bytes. Failure stays per clip: a batch `ffmpeg` refuses is
re-asked one clip at a time, leaving the bad clip as the only failure.
`make bench-overhead` can point at either recorder, `VAR=IW4L_PERF` or
`VAR=IW4L_BENCH`; they are separate costs and each needs its own pairs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`make bench` is one recipe over one run now. `IW4L_BENCH=1` arms an in-process
recorder on the spans that already feed Perfetto, and at exit the run prints
two independent reports and writes them to `iw4l-artifacts/bench/<stamp>.txt`:
the waterfall from the shell command to a playable map, carrying the time each
load stage held the pool alone, and frame time as a span tree whose nesting is
recorded rather than declared. Neither needs a trace file or Trace Processor,
and either can be MISS while the other stands. `bench-demo` and `bench-load`
are gone — a goal after `bench` is a demo, `ZONE=` is a live map.
A bot is no longer minted onto an id a real client already owns. The roster
steps over the ids the authority and the local player hold, and a slot that
still lands on the local player is retired, because `is_bot` would otherwise
claim the player and nothing could drive the slot anyway.
A stretch with no authority ack re-adopts the next snapshot and counts one
stall; three of them still fail the link. Failing on the first turned every
long frame into a permanent disconnect.
The loading overlay comes down once the world is installed rather than waiting
on the class-select hop, which a replay never makes — it was playing the whole
demo behind `UiLayer::Loading`, with map ambience held back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013bHwyKwUjaYmtVG1F3Ttku
vtroc