How can I verify that torchcodec is using the GPU/NVDEC?

[FredrikNoren]

I saw the ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i test/resources/nasa_13013.mp4 -f null - command in the doc, which I can use to verify that it's working on the ffmpeg side. Is there any way to verify that it also works on the torchcodec side?

[traversaro]

Probably not a perfect method, but I typically check for GPU usage via nvidia-smi or nvtop, if there is GPU usage while using torchcodec, I assume that everything is actually running on the GPU.

[NicolasHug]

nvidia-smi dmon also has some interesting stats.

For TorchCodec-specific usage, you are guaranteed that decoding will happen on GPU as long as you pass a cuda device as the device parameter. There is no automatic CPU fallback, i.e. if you ask for a CUDA device and the decoding somehow doesn't work on GPU, you'd get a loud error.

Hope this helps!

[FredrikNoren]

@NicolasHug Thanks! Yup that's super helpful