[MPlayer-users] -vo gl not working

[Reimar Döffinger]

On 14.04.2013, at 22:54, wm4 <nfxjfg at googlemail.com> wrote:
> On Sun, 14 Apr 2013 22:05:37 +0200
> Reimar Döffinger <Reimar.Doeffinger at gmx.de> wrote:
> 
>> about 10% on a Apple G4 system. On many older systems an additional
>> copy of the video will cost you at least 10% performance. Nothing
> 
> I'm not quite sure if direct rendering (the whole chain from decoder to
> VO) does not slow down modern systems. I _think_ I got slightly slower
> speed when trying with -dr, which could be due to the fact that the
> decoder is made to access slow GPU memory directly (isn't this also
> what MP_IMGFLAG_READABLE is about?), so an additional copy might
> actually speed up things - but I'm not too sure about this.

It's not GPU memory, it is memory DMA-able to the GPU - ideally.
However it does not always work/is not necessary, which is why -dr is not enabled by default.
As to multithreading: I already said that some tuning/adjusting the defaults might make sense.

>> against optimizing also for more modern systems, but please stop the
>> "I benchmarked on my system, so it must be fine", that should only be
>> a last resort approach.
> 
> I don't really understand the point of optimizing for older systems,
> when the associated code makes optimal decoding on newer systems (i.e.
> using MT) break everything.

Which would be a good argument, except that as I already said it does not break anything currently.
And of course for the fact that I consider the proper solution to bugs to fix them, not just disabling any code that doesn't look useful until it happens to work.
Yes, there is a point where it doesn't make sense to keep code, and for -slices it might be come since FFmpeg breaks it too often.
I don't see that for dr, especially since (except for my experiments) it is not used anyway for the only really problematic codec (H264).
Admittedly that means it is not that useful for many users either.
Lastly there is the question why you would see a point in optimizing for newer systems, they don't have any problems decoding in real-time, plus should probably be using GPU decode anyway.
On the other hand, some of the very newest systems like tablets have similar characteristics to really old systems: slow CPUs, tiny caches, pitiful memory bandwidth, crappy GPU drivers that are slow at copying textures.
So all those "useless features" might become more relevant again. Though multithreading is important and going to stay, so of course it must be working. But I say it again: a bug is not a reason for removing a feature when like here that feature has nothing to do with the bug.
Which is why I tend to get really grumpy when I see people use the words "remove" and "don't understand" together.