Add Alder Lake configuration for hwaccel

[charlesmunger]

Fixes #3170

[netlify]

✅ Deploy Preview for frigate-docs ready!

Name	Link
🔨 Latest commit	5de558d
🔍 Latest deploy log	https://app.netlify.com/sites/frigate-docs/deploys/62df1a50a23a1100084d6f55
😎 Deploy Preview	https://deploy-preview-3532--frigate-docs.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify site settings.

[NickM-27]

Is this true for 0.10 or are these tested with 0.11? If the RC then this should point towards release-0.11.0 branch

[charlesmunger]

Is this true for 0.10 or are these tested with 0.11? If the RC then this should point towards release-0.11.0 branch

Done.

[charlesmunger]

So, I think this probably shouldn't be merged as is... or at least more discussion is needed. With a variety of tuning parameters, I've been able to get the camera-to-live-view latency of the QSV decoder down to about 2.5 seconds - and at least for small low framerate video, I did not see any CPU usage improvement vs software decoding.

With vaapi, I was able to get the same excellent latency as the software decoder, but it increased CPU usage by about half. I suspect that unless hardware accelerated scaling filters are used, hardware acceleration for substreams isn't worthwhile on fast CPUs, because of the cost of copying raw frame data out of GPU memory. And even for people who are downscaling their detect streams, I bet that lowres 1 or lowres 2 will probably be quite efficient.

It's easy to imagine a setup where video is decoded in hardware, then motion detection is run via opencv's OpenCL kernels directly, with only frames destined for snapshots or inference going to main memory, but that would be a whole lot more complexity and coupling to hardware, plus you'd likely lose process isolation between ffmpeg and the python process.