So why is Rav1e as slow as it is? How can it be sped up to real-time? Threads? Chunk-based encoding?
From what I understood from T.Daede’s presentation when it was new, is that its speed advantage comes from being brutally simple, such as only using the smallest transform.
The reference implementation is purely for accuracy, and doens’t care about speed at all.
They also don’t have --release in their README; I wonder if this number was created without it, if so, doing just that should show anywhere from 2x-100x improvements.
EDIT: sent a PR for the README; they said their test scripts use it, so that must not be the case.