[FFmpeg-devel] [PATCH] avcodec/internal: improve min_size documentation for ff_alloc_packet2()

[Michael Niedermayer]

On Tue, Aug 04, 2015 at 12:57:38AM +0200, wm4 wrote:
> On Tue, 4 Aug 2015 00:51:17 +0200
> Michael Niedermayer <michael at niedermayer.cc> wrote:
> 
> > On Tue, Aug 04, 2015 at 12:12:24AM +0200, wm4 wrote:
> 
> > > > As well as benchmark results in some commits like
> > > > 4302a92835313000d4bf8030baae32c2b130d8a1
> > > 
> > >     avcodec/pcm: Better min_size for ff_alloc_packet2()
> > >     
> > >     33318 -> 30601 decicycles
> > > 
> > > That doesn't look convincing at all. So you save ~3000 decicycles for
> > > a task that isn't CPU bound, but which might blocks entire milliseconds
> > > on I/O.
> > 
> > There was a bug report of a overall (aka time ffmpeg) speedloss
> > in the % values with high res rawvideo
> 
> Was it fixed?

for rawideo and for several others too yes
iam pretty sure there are other encoders that suffer from speedloss
due to this too though


> 
> > > > i do intend to fine tune this for more codecs but i spend more time
> > > > in such disussions than working on the code and having people tell me
> > > > to do it differently each time i take a step kind of slows it down
> > > > and turns it into a mess
> > > 
> > > We also must take care of not turning the codebase into a mess.
> > 
> > I want to fix the mess
> 
> I don't mind, but...
> 
> > and i would suggest that we first test if any / which encoders benefit
> > from the static buffer.
> > It where quite a few long ago but it seems from a few quick tests that
> > this is not so anymore.
> > Can someone confirm ?
> > if it has become useless then I could drop quite a bit of code
> 
> Do you have to do your experiments on the live codebase? If you just
> want to try something, you could do that in a branch. If it gives good
> results, merge, if not, it's a failed experiment.

noone did "experiments on the live codebase"

also i strongly suggest that this discussion ends now and that we
return to do usefull work.

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

it is not once nor twice but times without number that the same ideas make
their appearance in the world. -- Aristotle
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150804/10dbda6d/attachment.sig>