Skip to content

GPU graph-capture profiling record

Scope and current evidence status

E8-F7/T7 owns CUDA profiling and machine-bounded optimization recommendations for resident graph capture. E8-F8 is restricted to the example, limitations, and epic closeout. P1--P4 profiling support is shipped, but no reviewed normalized artifact or manifest is checked in. Consequently, results, rankings, and recommendations are unavailable and unshipped.

Do not invent a machine or software identity, result, speedup, raw filename, byte size, digest, or recommendation. An absent artifact is not measured and is not zero.

Frozen matrix and reproduction commands

The fixed matrix uses small (1, 16, 2) and medium (1000, 16, 2) workloads with 100% activity, gas communication, and ordered communication/environment/gas/condensation/coagulation/dilution/wall-loss/ nucleation/diagnostics processes. It records gas/saturation diagnostics, two warmups, three samples, seed 1582, a duration of 0.5 s, and replay counts 1/10/100/1000. Kernel profiling is limited to compatible complete captured-replay small evidence.

The four required measurement rows use that frozen workload, warmup, sample, duration, and replay-count control matrix. Setup, reset, capture construction, and warmups are excluded from every interval. The captured-replay rows use the specified replay counts; those counts do not alter any excluded interval.

The explicit measurement rows are:

  • prepared_uncaptured / host_launch: Time dispatch submission only; exclude in-interval synchronization.
  • prepared_uncaptured / synchronized_elapsed: Time dispatch through one post-dispatch completion boundary.
  • captured_replay / host_launch: Time replay submission only; exclude in-interval synchronization.
  • captured_replay / synchronized_elapsed: Time replay dispatch through one post-dispatch completion boundary.

The first two commands, the repository runner, and MkDocs build are hardware-free:

pytest particula/gpu/tests/profiling_support_test.py -q --no-cov
pytest particula/gpu/tests/benchmark_helpers_test.py -q --no-cov
pytest particula/gpu/tests/benchmark_test.py --benchmark -k resident -v -s --no-cov
pytest particula/gpu/tests/profiling_smoke_test.py --benchmark -q --no-cov
.opencode/tools/run_pytest.py
mkdocs build --strict

Native collection and smoke require native-CUDA qualification. They must report unavailable or skip when qualification fails; CPU and Warp-CPU are not fallbacks.

Future artifact provenance

Reviewed normalized artifacts and their manifest pointer belong in .artifacts/benchmarks/profiling/; ignored local raw reports belong under .artifacts/benchmarks/profiling/raw/. Future rows must include a machine identifier/platform, CUDA device/architecture, driver/runtime, Python, Warp, source revision, selected nsys/ncu versions, command, timestamp, workload, mode, method, and raw artifact reference.

nsys 2026.1.3.425-1 and ncu 2026.2.1.5-1 are qualification requirements, not current evidence. Every reviewed executed row must carry a contained relative raw filename, byte size, and lowercase SHA-256 reference. Never publish raw reports, absolute paths, credentials, usernames, or device pointers.

Methodology and limitations

Reset, warmup, setup, capture construction, and serialization are outside intervals. Host launch has no in-interval synchronization; synchronized elapsed has one post-dispatch completion boundary. Graph launch is not device or kernel duration. Nsight perturbs timing and is not unprofiled throughput. Only a complete compatible attributed nsys duration supports a bounded kernel conclusion.

This record permits no CPU or Warp-CPU fallback, portability or cross-machine/ workload inference, speed thresholds, public runtime profiler API, or auto-tuning. It makes no production, scientific, ownership, or RNG optimization without a separately referenced correctness plan. Missing prerequisites or counters are unavailable evidence, never favorable evidence; no fabricated result, speedup, ranking, or recommendation is current evidence.