🏎️ Implement GPU acceleration

[Gamer92000]

LectureCut needs to support GPU acceleration

This includes

• figuring out what kind of acceleration is best, and
• applying the acceleration if wanted.

Reference

On the feat/nvidia branch you can see my general idea of this and a helpful function to see what flags Ffmpeg was compiled with.

Figuring out what is best

I don't really have a good idea of how to do this. Obviously, LectureCut should use the fastest possible accelerator. So you need to figure out if Ffmpeg supports it and whether the user has the required hardware. There should be just one flag, like --enable-acceleration to use the previously determined best possible accelerator. This step can take a few seconds to figure things out if you want to go the route of brute force testing against Ffmpeg.

Applying acceleration

The acceleration does not need to be applied to all Ffmpeg calls. Only when video is involved and actual de-, en-, or transcoding takes place.

💡 NOTE

If this turns out to work really well and reliably then you can also turn it on by default and have a --force-disable-accelerator option to do everything on the CPU again.

[Gamer92000]

So it turns out that Nvidia limits the number of concurrent NvENC sessions to 3. We could either

• expect the user to patch their Nvidia driver,
• create two threadpools, one for non-transcoding and one for transcoding, or
• somehow abstract the transcoding away from Cuda / NvENC and use something more generic.

Pure GPU decoding showed no performance improvement on my setup (R9 5950X + RTX 2070S). However, I could see up to 1.6 times the performance on the transcoding step with a patched driver.