What you've been able to do is already quite a feat for lower-end Intel hardware. Squeezing out any more performance will be extra icing on the cake. Just tested beta2 on my N150 and confirm no more drop-outs.
My go-to will likely be AMD 5800H for DV, HDR-to-DV conversions, and full 3D when that's implemented. Mostly goodbye Amlogic at that point 😉😁.
I'm working to improve HDR10 to DV conversion with AI, the current pipeline is far too computationally expensive. AI itself isn't the issue, as inference takes around 1 ms. The real bottleneck is the signal analysis required to feed the model.
My algorithm works in two stages:
- Analyze the image and generate an initial set of L1/L3 metadata.
- Use an AI model trained on paired HDR10/DV masters to correct and refine the initial analysis.
The algorithm itself works very well, I tried more complicated one, this one is simple, lightweight, stable, and produces excellent results (better than the conversion with VS10 Dolby engine). However, implementing the entire pipeline properly and efficiently is more challenging than expected. In the one is beta2 can only fit if you have a strong GPU, on beta3 the pipeline will be more accessible for GPU.