| name | perf-check |
| description | Run the rocDecode performance test and compare to baseline to catch regressions |
| allowed-tools | ["Bash(cmake *)","Bash(test/perf_regression.py *)","Bash(python3 test/perf_regression.py *)","Bash(nproc)"] |
Run the rocDecode performance regression check: measure decode FPS and compare it against
the GPU-specific baseline. Execute the steps in order, stopping only if a step fails. Run
from the rocDecode project root (the directory containing build/, samples/, and
test/).
IMPORTANT: Run each command as a SEPARATE Bash tool call. Do not chain commands with && or |.
The GPU is detected automatically (via amd-smi/rocm-smi/KFD) and mapped to the matching
column in the baseline (MI250X, MI300X, MI300A, MI350, MI355, Navi31, Navi48).
Override with ROCDECODE_PERF_GPU=<column> if detection is wrong.
Perf streams and the baseline file are located via ROCDECODE_PERF_DIR (default
$HOME/rocDecodePerformance). That directory must contain the per-codec stream
subdirectories AvcPerformance, Av1Performance, HevcPerformance, Vp9Performance, and
the baseline rocDecode_perf_baseline.html (download it from SharePoint first). The
regression threshold is 5% Avg FPS drop, overridable via ROCDECODE_PERF_TOLERANCE.
Step 1 — Verify the ROCm toolchain
Run: test/perf_regression.py --check-rocm
Confirms ROCM_PATH points at a usable ROCm install (default /opt/rocm). If it fails,
set ROCM_PATH — in ~/.profile, since skills run cmake in a non-interactive shell that
does not source ~/.bashrc — and re-run before continuing.
Step 2 — Build the performance sample
Each sample builds in its own directory. Run these two commands separately:
cmake -S samples/videoDecodePerf -B samples/videoDecodePerf/buildcmake --build samples/videoDecodePerf/build
This requires rocDecode to be installed already (make install); run the /validate skill
first if you have not built and installed the library this session.
Step 3 — Run the regression check
Run: test/perf_regression.py
If the skill was invoked with a quick argument (e.g. /perf-check quick), run the fast
variant instead — test/perf_regression.py --quick — which measures one stream per leaf
subfolder capped at ≤4K (the 8K streams dominate runtime, so they are skipped) instead of
every stream. Use it for a fast sanity check; use the full run before finalizing. Each run
prints its own elapsed time.
Streams within tolerance pass on a single run; streams that appear to regress are
re-measured (3-run average) to rule out noise before being reported.
Report the final summary box from perf_regression.py to the user, and list any streams
marked REGRESSED (with their measured vs. baseline FPS and delta%).
