Av1an
Av1an copied to clipboard
master-of-zen
Feature Request: Target Quality with ssimulacra2-rs
Kind of self explanatory, use ssimulacra2-rs to probe target quality. I understand this would likely be very slow, but I'm really curious to see how much more efficient it might be. It should also allow reliably targeting higher qualities near visual transparency that VMAF will break down for.
It would work for x264/x265 quite well, and maybe aom-av1-psy, but I'm not sure if it is wise to do so until ssimu2-rs gets speed parity with the mainline ssimu2 algo.
There might be something else we could implement, but that would require using a different metric that does not(yet) exist.
As of this comment, my 5800X gets about 6.1 fps with 8 threads, about 6.7 fps with 12 threads, and about 7.1 fps with 16 threads. It doesn't scale particularly well with higher thread counts, but I think ~6 fps is a fairly reasonable speed as it is for a large portion of the audience of Av1an. Talking with Soichiro earlier, it seems the primary obstacle to ssimulacra2_rs speed at the moment is blur calculation, and the limitation on threading scalability is due to the decode speed of the video.
I'm not sure if it is wise to do so until ssimu2-rs gets speed parity with the mainline ssimu2 algo.
Has anyone done a proper speed comparison between the two? We know where ssimulacra2 takes time (in general, not just the Rust port) and both look pretty similar in where they spend their time.
As of this comment, my 5800X gets about 6.1 fps with 8 threads, ...
I'm surprised it even continues to scale at all beyond 8 threads. I am almost sure that (depending on the CPU, of course) the algorithm hits a memory bottleneck somewhere in the 4-6 thread range. I'd also be interested to see how well the C++ version scales, but I don't think that one currently has threading support.
I too have seen little improvements of going over 8 threads, just fyi. I believe a speed comp was done at some point by either Soichiro, Blue or someone else but it's likely lost into thousands of messages regarding ssimu2.
