[FFmpeg-devel] Prores encoding optmizations

[Martin Vignali]

2016-05-23 13:44 GMT+02:00 Håvard Espeland <espeland at mixedrealities.no>:

>
> > The SIMD won't be accepted if it's intrinsics. The codeword encoding is
> not
> > SIMD, is it? So that may be worth upstreaming.
>
> All optimizations we’ve done are SIMD so it does not apply. Basically what
> we do for codewords is to process the shifting/masking for eight codewords
> at a time. The put_bits calls are of course executed sequentially as before.
>
>
> > When you say "no significant PSNR or file size differences", though, it
> > sounds like the output changed? Did it? Can you explain why?
>
> For the tractor sequence we get the following results.
> time ./ffmpeg -y -i ~/Downloads/tractor_1080p25.y4m -codec:v prores_aw
> -an  -profile:v 2 /tmp/video.mov
>
> ffmpeg-master:
> frame=  690 fps= 22 q=-0.0 Lsize=  611897kB time=00:00:27.56
> bitrate=181881.2kbits/s speed=0.876x
> video:611891kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB
> muxing overhead: 0.000996%
>
> real    0m31.491s
> user    2m3.452s
> sys     0m1.168s
> [Parsed_psnr_0 @ 0x3ab6da0] PSNR y:49.596357 u:54.113644 v:54.561940
> average:51.331354 min:49.684050 max:53.658792
>
> AVX2 optimizations:
> frame=  690 fps= 32 q=-0.0 Lsize=  611896kB time=00:00:27.56
> bitrate=181881.0kbits/s speed=1.27x
> video:611890kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB
> muxing overhead: 0.000996%
>
> real    0m23.396s
> user    1m25.092s
> sys     0m1.252s
> [Parsed_psnr_0 @ 0x2838da0] PSNR y:49.595523 u:54.114016 v:54.562335
> average:51.330827 min:49.683836 max:53.658654
>
> As you can see, there is 1kb size difference in the output file and some
> minor variations in psnr. I believe the difference might be due to rounding
> since we are using float point math and round towards zero, while ffmpeg
> master uses fixed points for dct. If we pursue to clean up the code and
> send it upstream we might also look into improving this so we get identical
> results.
>
> Have you test your optimizations in the other prores encoder (prores
kostya) (who i think have more features (interlaced encoding and 444
versions)) ?


Martin