RGP Inspect

Use the shared rgp-inspect CLI to extract evidence from an existing .rgp capture. Do not recapture a workload unless the trace lacks the required data.

Workflow

1. Run rgp-inspect doctor to locate the full Radeon Developer Tool Suite, amdgpu-dis, and the patched decoder.
2. If the native decoder is absent, read references/build.md and build the pinned decoder visibly.
3. Run rgp-inspect decode CAPTURE.rgp --out REPORT.json --csv REPORT.csv.
4. Check resolution_ratio, decoder_info, and stream validation before interpreting hot instructions.
5. Reject traces with impossible CU/SIMD identifiers or less than 95% ISA resolution.
6. Compare traces only when capture mode, workload shape, dispatch selection, and clock policy match.
7. Report the capture type and distinguish wave-aggregated shader cycles from kernel wall time.

Commands

rgp-inspect doctor
rgp-inspect decode trace.rgp --out trace.json --csv trace.csv

The decoder build and gfx12 token format are revision-sensitive. Do not silently substitute another rocprof-trace-decoder revision. Read references/format.md before changing RDF extraction, address resolution, or report semantics.

Interpretation rules

• Treat instruction-trace values as shader cycles unless trustworthy clock calibration is present.
• Do not convert summed per-wave instruction cycles into kernel wall time.
• Report instruction duration, stall, and scheduler idle separately.
• Use RGP event timing or benchmark events for kernel wall time.
• State whether evidence came from instruction tracing, counter capture, event timing, or an ordinary benchmark.
• Do not infer named GUI counter formulas from raw SPM identifiers without a verified mapping.

If no suitable trace exists, use the rgp-capture skill rather than embedding capture lifecycle logic here.