Files
chenlin c528dd268f feat: checkpoint Go2 FL simulation performance v0.4.2
Include callable skill and live-GUI prerequisites, epoch sensor reads, packed snapshots, sampled bit-exact kinematics, and verified-tick rendering without changing the physical/control schedule.

Record bounded analytic S1 GUI/headless qualification, original failures and CPU evidence (1557 passed, 6 skipped). Preserve accepted_skill=false and sub-real-time limitations. Exclude native authorities, caches, large media and local delivery copies; retain sealed artifact digest.
2026-10-16 15:46:01 +08:00

3.3 KiB

0.4.2 application hot-path profiling and optimization

Operator: current request explicitly permits full use of RTX5080 and authorizes artifacts under this repository. No more output-path questions. Existing AGENT.md authorizes bounded autonomous work and read-only specialists.

This is an Isaac Lab application-loop performance task, not USD geometry optimization. The omniverse-usd-performance-tuning entry triages it to application/runtime profiling; no USD restructuring, simplification, SO/AV operation or standalone install is required or claimed. Runtime is the already-used local IsaacLab /home/cen/IsaacLab + Isaac Sim /home/cen/isaacsim-6.0.1 on CUDA0 RTX5080, not a CPU fallback. Runtime metadata/probe is recorded in ../setup-preflight.json. A missing SO/AV blocks those unrequested operations, not this Python hot-path benchmark. All generated performance artifacts are under this operator-selected repository (this subdirectory and scripts/).

Bounded task: at most12Kit instances,120000 reserved/actual target-environment calls,24GiB owned logical storage,6hours from scope.json; zero training/parameter search, no artificial GPU power/clock cap, no foreign process termination. Serial runs only; unchanged strict GPU admission/monitoring/final ownership and exact-SHA shutdown handoff. New session/grant/output each run, whole-task parent accounting in addition to per-session ledgers. Do not reuse completed0.4.1 authority.

Preserve old0.4.0 and0.4.1 packages/archive/checksums and all historical evidence. Parent sole writer, no source/config edits while a run is active. Snapshot each attempt's approved code, including unsuccessful attempts. Frozen candidate and original physics/controller/limits/gates/500Hz physics100Hz controls/render_interval20 stay unchanged. Do not thin500Hz traces, postpone failure stopping, bypass provenance, fake FPS via prerecorded/interpolated output, add dummy GPU load, enlarge the task to multiple legs/envs, or optimize RootXY.

  1. Bind a CPU cProfile/perf-counter observer before PREP, measure a fresh analytic GUI run with requested safe stop after first complete cycle. Keep startup, active loop, control/physics throughput, real render cadence and CPU profiling overhead separate. Sample real GPU util/power/clocks externally without changing settings. No guesses about utilization.
  2. Optimize evidenced CPU/GPU transfers/duplicate read-only computation while keeping fail-before-next-step semantics and exact every-tick values. Prefer batching/gathering to omitting checks. CPU equivalence/negative tests, original source hashes and reviewer before actual optimized run.
  3. Compare same-seed same-protocol real traces and observed native counters. Report unchanged gate results and equality/tolerances honestly; CPU profile comparisons share profiler overhead, final benchmark disables profiling.
  4. Run bounded actual full analytic S1 GUI and headless regression on accepted implementation, plus early-close as needed for affected behavior. Preserve current source/version independently; seal only evidence-qualified artifact, disclose residual bottlenecks and do not promiseGPU100% or1x before measured.

No USD stage mutation chain is planned. No optimized USD/report is claimed from SO/AV. This task produces application profiling JSON/Markdown, raw traces and a versioned callable implementation if accepted.