git:20260629.b025b10 to git:20260814.317a19c

---

name: perf-example-device

description: Benchmark the hardware performance of a single example ($ARGUMENTS). Use when the user asks to benchmark or measure the latency of one example on hardware.

---

# Benchmark the hardware performance of a single example at $ARGUMENTS

Reference `tools/benchmark_rounds.sh` for the full implementation pattern

(device log resolution, timing parsing, reporting format) — this skill runs the

same logic for a single example. Detection / isolation procedures live in

[`../../lib/onboard-detection.md`](../../lib/onboard-detection.md).

1. Locate the test file under `$ARGUMENTS/`: pick the single `test_*.py` that

lives directly in that directory. If none exists, tell the user the

directory is not a scene test and stop.

2. If `command -v npu-smi` is not found, tell the user this requires hardware

and stop.

3. **Precheck + detect platform** (§A) — derive the arch from `$ARGUMENTS`,

gate on real silicon, then read the detected arch into `PLATFORM`.

4. **Select a single idle device** (§C) — or let `task-submit` pick via

`--device auto --device-num 1`.

5. **Run through `task-submit`** (§E, `--device-num 1`) following the same

pattern as `run_bench()` in `tools/benchmark_rounds.sh`: snapshot logs, run

`python $ARGUMENTS/test_<name>.py -p <platform> -d $TASK_DEVICE --rounds 10 --skip-golden`,

find the new log, parse timing, report results.