FrameServer 4.21 released, supporting ffmpeg/Resolve/more..


Howard-Vigorita wrote on 3/23/2023, 8:15 PM

@GJeffrey Thanks for that. V210 makes a significant difference. More like Voukoder 13...

rgb32 ... psnr: 30.6171; ssim: 9474; vmaf: 40.1038
vouk .... psnr: 38.1276; ssim: .9448; vmaf: 51.0709
v210 .... psnr: 38.3352; ssim: .9465; vmaf: 52.5895

But unfortunately when v210 is selected, the avi signpost is not recognized by vp20. Don't know if VRQ can load it without vp20 so I can measure it. Btw, v210 appears to be 4:2:2. Doesn't look like there's a 4:2:0 10-bit option.

Howard-Vigorita wrote on 3/23/2023, 10:26 PM

Just tried QsvEncC64 and that worked like a charm with FrameServer. Used this command line which I think I got right to match the x265 ffmpeg command line I used above with v210:

qsvencc64 -i d:\temp\fs.avs -c hevc --vbr 24000 --max-bitrate 48000 --vbv-bufsize 48000 --output-depth 10 -o d:\temp\fs.mkv

Got these results with Intel Qsv ... psnr: 38.3507; ssim: .9473; vmaf: 53.2081

Btw, the v210 x265 render took 3m 14s while the v210 QsvEnc took 58s.

Howard-Vigorita wrote on 4/1/2023, 1:15 PM

New version 7.36 of qsvenc came out a few days ago so I reran everything with a more explicit command line in both versions:

qsvencc64 -i d:\temp\fs.avs -c hevc --input-csp yuv422p10le --vbr 24000 --max-bitrate 48000 --vbv-bufsize 48000 --output-depth 10 -o d:\temp\fs.mkv

Found that with the new version, leaving out the convertto line in the avs text file ran slightly faster at higher quality. Here's what I got before and after the qsvenc update and then dropping the convertto:

Note that the FrameServer v210 format natively converts my 4:2:0 10-bit input clip to a 4:2:2 10-bit sign-post. Which qsvenc then converts back to 4:2:0 10-bit. I would not be surprised if all that needless back-and-forth might affect both performance and quality.