chore: extract PerfSampler from AutoPilot for AltTester reuse - #8826
Conversation
Move per-frame CPU/GPU sampling, CSV writer, 8-line summary writer, and PercentWorst out of AutoPilot.RunAsync into a new static PerfSampler (Configure/Begin/End). AutoPilot now brackets StandAtSpawnAsync with PerfSampler.Begin/End and keeps Application.Quit in finally. ProfilingPlugin calls PerfSampler.Configure(profiler) unconditionally, so AltTester-driven InWorld fixtures can drive Begin/End via AltDriver.CallStaticMethod without --autopilot. CSV columns, line endings (CRLF), encoding (UTF-8 no BOM), and the eight summary lines are byte-identical to the pre-refactor output.
|
Windows and Mac build successful in Unity Cloud! You can find a link to the downloadable artifact below. |
|
New build in progress, come back later! |
popuz
added
force-build
PercentWorst computed k as (int)(times.Count * fraction), which floors to 0 whenever Count * fraction < 1 — e.g. the 0.1% bucket on any window shorter than 1000 samples. Take(0).Average() then threw InvalidOperationException, killing the whole summary write. Standalone AutoPilot never hit this because StandAtSpawnAsync yields ~90s of samples on a dev machine. Fixture-level InWorld windows on the T4 chassis (~15-20s, low FPS) regularly land below 1000 samples and crashed the entire perf-summary writer. Floor k to 1 so 0.1% worst on small windows simply reports the worst single frame instead of crashing. Output is byte-identical for any window with >= 1000 samples (k = (int)(N * 0.001f) >= 1 already). Also guard the empty-list case defensively.
Bump cloud_build_url default to pr-22430-3a0b8d5, which carries the PerfSampler.PercentWorst guard against tiny sample windows (decentraland/unity-explorer#8826). Improve AttachPerf diagnostics: log ex.ToString() instead of just ex.Message so AltTester's wrapped TargetInvocationException reveals the real Player-side stack in the Allure report. Best-effort attach perf.csv on the End-failure path — PerfSampler.End disposes the CSV before WriteSummary runs, so a summary-side crash still leaves a complete CSV on disk that can be inspected post-mortem.
popuz
added
the
no QA needed
popuz
changed the title
popuz
changed the title
popuz
removed
the
no QA needed
popuz
changed the title
popuz
changed the title
…lttester' into feat/test-automation/autopilot-alttester
This comment has been minimized.
This comment has been minimized.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
PR #8826, run #26192309600 Builds: Windows change, Windows baseline, macOS change, macOS baseline Framework 13 i7
|
popuz
added
the
no QA needed
This comment has been minimized.
This comment has been minimized.
github-actions
Bot
removed
the
no QA needed
This comment has been minimized.
This comment has been minimized.
|
PR #8826, run #26217949913 Builds: Windows change, Windows baseline, macOS change, macOS baseline Framework 13 i7
|
pravusjif
left a comment
pravusjif
left a comment
There was a problem hiding this comment.
If we extend some virtual input simulation for Alttester tests in the future, we can have automatized waypoint-based paths in-world, measuring performance thanks to this initiative, great job!
This comment has been minimized.
This comment has been minimized.
popuz
added
the
no QA needed
|
since we have proper performance report - it means that behavior wasn't broken and it acts as a simple smoke test for run of explorer. Thus - |
|
Claude finished @popuz's task in 1m 59s —— View job PR Re-review — chore: extract PerfSampler from AutoPilot for AltTester reuse
Root-cause checkThis PR extracts per-frame CPU/GPU sampling out of Progress since last review
Blocking IssueLINQ violation —
On the "it only runs in CI" argument: the code lives in REVIEW_RESULT: FAIL ❌ |
github-actions
Bot
removed
the
no QA needed
|
PR #8826, run #26233457658 Builds: Windows change, Windows baseline, macOS change, macOS baseline Framework 13 i7
|
3 participants
What does this PR change?
Extracts per-frame CPU/GPU sampling out of
AutoPilotinto a reusablePerfSamplerso the same code path can produce perf summaries for both:--autopilot-csv/--autopilot-summaryCLI flow (byte-identical output).explorer-automation, which callPerfSampler.Begin/Endover the AltTester bridge to bracket a test fixture instead of a 90 sStandAtSpawnAsyncloop.Changes
PerformanceAndDiagnostics/AutoPilot/PerfSampler.cs—Configure(IProfiler)/Begin(csvPath, summaryPath)/End()static API. Owns the per-framePlayerLoophook, the CSV writer, the 8-line summary writer, and thePercentWorstaggregation. SameIProfiler.LastFrameTimeValueNs/LastGpuFrameTimeValueNssignal source as before — no newProfilerRecorderinstances, no analytics-pipeline dependency.AutoPilot.cs—RunAsyncnow bracketsStandAtSpawnAsync(90)withPerfSampler.Begin/End.Application.Quitstays infinallyand is only reachable from the standalone path; AltTester-driven sampling never quits the player.ProfilingPlugin.cs— callsPerfSampler.Configure(profiler)unconditionally (outside theAUTOPILOTflag gate) so the InWorld path works without--autopilot.Compatibility
CSV columns, line endings (CRLF), encoding (UTF-8 no BOM), and the eight summary lines (
CPU average,CPU 1% worst,CPU 0.1% worst,CPU worst,GPU average,GPU 1% worst,GPU 0.1% worst,GPU worst) are byte-identical to the pre-refactor output. The Part B consumer in explorer-automation parses the same file with one regex regardless of which path produced it.Test Instructions
Standalone path — same as before this PR:
Wait ~3 min (load + 90 s stand). Player exits.
perf-summary.txtshould contain the same 8 lines as a pre-refactor build.AltTester-driven path is exercised by the companion PR — no manual steps needed here:
Quality Checklist
--autopilot-summaryoutput verified byte-identical to pre-refactor.PerfSampler.Configureplacement reviewed — outside theAUTOPILOTgate so the InWorld path works.usingdirectives for analytics namespaces inPerfSampler.cs(signal source isIProfileronly).