Update: benchmark_rounds.sh adds Host/Device, renames Elapsed to Total

[hw-native-sys-bot]

Summary

Restore Host / Device per-round columns that #828 dropped (the RunTiming walls from #790) alongside the device-log Total / Sched / Orch, and rename Elapsed → Total since "elapsed" reads ambiguously next to two distinct wall clocks. The output is now a 5-column table:

Column	Source	Meaning
Host (us)	RunTiming.host_wall_us (framework)	steady_clock wall around dispatch
Device (us)	RunTiming.device_wall_us (framework)	AICPU mailbox orch_start → orch_end
Total (us)	device log	max(end) − min(start) across sched_* / orch_* events
Sched (us)	device log	sched_start → sched_end
Orch (us)	device log	orch_start → orch_end

Key changes

• parse_timing takes two inputs (framework stdout + device log) and merges per-round timings into the 5-column table. awk handles file 1 via FNR == NR, file 2 with the existing sched/orch parser.
• Framework header is matched anywhere on the line, not anchored at column 0 — the SceneTestCase runner prints the case name with end="" so _log_round_timings's header gets concatenated to the same line as Case1 ....
• Avg / Trimmed Avg lines now cover all five metrics. The /benchmark skill is updated to parse Total Trimmed Avg: (was Trimmed Avg:) and its sample tables rename Elapsed (us) → Total (us).
• Performance Summary table grows to 5 columns; the per-column avg extraction is set -e + pipefail-safe (uses parameter expansion instead of grep | awk pipelines that could silently die on no-match).

Sample output

  Round      Host (us)   Device (us)    Total (us)    Sched (us)     Orch (us)
  ------  ------------  ------------  ------------  ------------  ------------
  0           460759.8       29346.9        1178.9        1178.9         792.1
  ...
  Host Avg: 455769.2 us  |  Device Avg: 8619.4 us  |  Total Avg: 1158.0 us  |  Sched Avg: 1158.0 us  |  Orch Avg: 708.6 us  (25 rounds)
  Host Trimmed Avg: 370587.1 us  (dropped 10 low + 10 high, 5 rounds used)
  Device Trimmed Avg: 6820.6 us  ...
  Total Trimmed Avg: 1156.1 us  ...
  Sched Trimmed Avg: 1156.0 us  ...
  Orch Trimmed Avg: 704.6 us   ...

Test plan

• bash -n tools/benchmark_rounds.sh
• ./tools/benchmark_rounds.sh -n 25 -d 1 on paged_attention_unroll → 5 columns populated, Performance Summary + Benchmark complete sections all render
• /benchmark skill grep targets (Total Trimmed Avg:, Orch Trimmed Avg:) both present in output
• Full suite (-n 100 over all examples) on a free device

[gemini-code-assist]

Code Review

This pull request enhances the benchmarking tool by expanding the reported metrics from three to five: Host, Device, Total, Sched, and Orch. Key changes include updating tools/benchmark_rounds.sh to capture framework stdout for Host and Device timings, refactoring the parse_timing function to process multiple data sources, and updating the summary table to dynamically display available metrics. Documentation in SKILL.md has also been updated to reflect these changes. I have no feedback to provide as there were no review comments to evaluate.