gpuval

Documentation

Overview

Command gpuval validates qc on an NVIDIA GPU and writes a Markdown report to paste back: NVDEC decoding speed and frame identity, CUDA VMAF against CPU VMAF (agreement and speed), NVENC ladders against CPU ladders (time, bitrate at equal VMAF), and a CQ sweep calibrating the NVENC probe CQs.

It only uses synthetic content generated with ffmpeg's lavfi sources, or a public clip given with -source (see docs/gpu.md). It is meant to run in the CUDA image, which ships it:

docker run --rm --gpus all --user "$(id -u):$(id -g)" -v "$PWD/gpu-validation:/work" \
    --entrypoint gpuval ghcr.io/eko/qc:<version>-cuda -dir /work

or make gpu-validate from a checkout. Each step records its failure in the report and the next one runs, so a partial GPU (no AV1 NVENC, no CUDA VMAF) still yields a useful report.