[FFmpeg-user] minterpolate performance & alternative

[pdr0]

Mark Filipak (ffmpeg) wrote
> On 01/28/2021 07:42 PM, pdr0 wrote:
>> Mark Filipak (ffmpeg) wrote
>>> But perhaps by "process in parallel" you mean something else, eh?
>>> ...something I'm unaware of. Can
>>> you expand on that?
>> 
>> 
>> I mean "divide and conquer" to use all resources. If you're at 20% CPU
>> usage, you can run 4-5 processes
>> 
>> eg. Split video in to 4-5 segments. Process each simultaneously, each to
>> a
>> lossless intermediate, so you're at 100% CPU usage. Then reassemble and
>> encode to your final format
> 
> I don't think that will work very well, even if I carefully cut on key
> frames. The reason is that 
> the minterpolate filter drops 10 frames and that means that at the join of
> each section there'll be 
> a 1/6 second jump (or maybe worse).


This has nothing to do with keyframes; I'm not talking about stream copy.

eg. -vf trim splits in the uncompressed domain (data is decoded to
uncompressed frames) . If you need  to you can split stages as lossless
intermediates

You cut on cadence boundaries of 24 frame cycles. 24p is evenly divisible
into 120p (5x) . 60p takes every 2nd frame from that 120p result this is
what you are doing with optical flow retiming, so every frame is evenly
spaced in time. 





> Note: Whether the minterpolate filter drops 10 frames or 5 frames is the
> subject of [FFmpeg-user] 
> minterpolate PTS v frame count weirdness.

Not sure what this is referring to ? Any more details ?

If you're getting PTS , frame count weirdness, split it out as lossless
intermediates



--
Sent from: http://ffmpeg-users.933282.n4.nabble.com/