[FFmpeg-cvslog] random thoughts about refactoring

[compn]

On Sun, 10 Jan 2010 17:42:24 +0100, Michael Niedermayer wrote:
>On Sun, Jan 10, 2010 at 12:44:24PM +0100, Diego Biurrun wrote:
>> On Fri, Jan 08, 2010 at 02:28:16AM +0100, Michael Niedermayer wrote:
>> > On Fri, Jan 08, 2010 at 01:20:03AM +0100, Diego Biurrun wrote:
>> > > On Wed, Jan 06, 2010 at 06:38:14PM +0100, Michael Niedermayer wrote:
>> > > > On Tue, Jan 05, 2010 at 11:08:22PM +0100, Diego Biurrun wrote:
>> > > > > On Tue, Jan 05, 2010 at 08:43:51PM +0100, Michael Niedermayer wrote:
>> > > > > > On Tue, Jan 05, 2010 at 11:29:46AM +0100, Diego Biurrun wrote:
>> > > > > > > On Tue, Jan 05, 2010 at 03:20:47AM +0100, Michael Niedermayer wrote:
>> > > > > > > > On Mon, Jan 04, 2010 at 08:23:23PM +0100, Diego Biurrun wrote:
>> > > > > There are people here who know H.264 well enough to lend
>> > > > > a helping hand in theory, but in practice the implementation in lavc is
>> > > > > just too impenetrable.  h264.c is around 10.000 lines if you count svq3.c
>> > > > > which it directly includes!
>> > > > 
>> > > > > 
>> > > > > Also, our H.264 decoder is not a speed demon.
>> > > > 
>> > > > That seems to depend on CPU and situation, gcc making poor inlining choices
>> > > > and overflowing the available caches is not exactly the codes fault ...
>> > > 
>> > > Carl Eugen was the first person I have ever heard claim that our
>> > > H.264 decoder was faster than anything, ever...
>> > 
>> > Ive never tried CoreAVC, but i can assure you the reference implementation
>> > is much much slower
>> 
>> You are referring to the H.264 reference implementation?
>> 
>> Jason mentioned two new H.264 decoders on IRC that are even faster than
>> CoreAVC...
>
>Why is this information not posted on the mailing list ?

i'll paste some relevent irc chat:

[17:17] <Dark_Shikari> I would say the single most important thing for h264 decoding
[17:17] <Dark_Shikari> is to get ffmpeg-mt mergedc
[17:17] <Dark_Shikari> that will give more speed boost than any other optimizations

[17:40] <Dark_Shikari> people would rather download divx h264 for free
[17:40] <Dark_Shikari> or diavc for free
[17:40] <DonDiego> diavc?
[17:41] <Dark_Shikari> decoder which came out of nowhere
[17:41] <Compn> if there are any new free h264 decoders, i'd like to add them to mplayer , for testing
[17:41] <Dark_Shikari> still very buggy and alpha, but faster than divx and coreavc

h264 decoders:

http://di-avc.com/
http://labs.divx.com/
http://corecodec.com/products/coreavc
http://vsofts.com/products/pc-software-h264-professional-sdk.html


[14:30] <Dark_Shikari> ffmpeg's h264 decoding asm is already mostly good
[14:34] <Dark_Shikari> by far most of the time is spent in c code




some good SoC talk:



[17:49] <DonDiego> Dark_Shikari: we have no students who reach the patch review phase, so this is not a problem
[17:49] <Dark_Shikari> lol
[17:49] <DonDiego> however, we have students who seem to take the money and do hardly anything in return..

[17:50] <Dark_Shikari> 2) students should be on IRC every single day
[17:50] <Dark_Shikari> and talking about their project, every single day
[17:50] <Dark_Shikari> 3) students should be on IRC before applying
[17:50] <Dark_Shikari> 4) students should be on IRC after the end date
[17:50] <DonDiego> what requirements do you have for x264?
[17:50] <Dark_Shikari> requirements for x264:
[17:50] <mru> jai is exempt from that rule ;-)
[17:51] <Dark_Shikari> of course, if you prove yourself otherwise, you can break the rules.
[17:51] <Dark_Shikari> 1) You must complete a qualification task that represents a portion of the actual project you intend to do.
[17:51] <Dark_Shikari> For example, if you were writing an h264 decoder, you would have to write a header + bitstream parser
[17:51] <Dark_Shikari> that printed dct coefficients for cavlc streams
[17:51] <DonDiego> yes, i notice that many people from there apply and i suspect that money could be a motivation, there's $4500 to earn, i can live off that kind of money for 1/2 year in germany..
[17:51] <Dark_Shikari> that kind of thing is _useful for your actual project_
[17:51] <kierank> the x264 qualifications are meatier than ffmpeg ones
[17:51] <DonDiego> mru: jai has just taken money and done nothing in return :-/
[17:52] <Dark_Shikari> 2) You must be on IRC often, and get to know the community
[17:52] <Dark_Shikari> you must be involved
[17:52] <Dark_Shikari> you must respond to pings
[17:52] <Dark_Shikari> If I ping you and say "there's a bug in your code here", you should be able to respond
[17:52] <Dark_Shikari> obviously, I can't control timezones/sleep schedules/etc
[17:52] <mru> DonDiego: jai has done stuff though
[17:52] <Dark_Shikari> but in general, you shouold be available.
[17:52] <kierank> 22:42] <@Dark_Shikari> you must respond to pings --> like janihan ;)
[17:52] <Dark_Shikari> lol
[17:52] <DonDiego> mru: that was last year, this year it was money for nothing..
[17:53] * mru takes the chicks for free...
[17:53] <Dark_Shikari> I don't think the x264 requirements are harder than ffmpeg
[17:53] <Dark_Shikari> the ffmpeg qual tasks are really hard
[17:53] <Dark_Shikari> and don't even relate to the main project often
[17:53] <Dark_Shikari> I think the main thing is that we create a community and force people to be involved
[17:53] <Dark_Shikari> and if they can't stay online and involved, they don't get in
[17:53] <mru> that's how communities work
[17:53] <mru> _saying_ you're in doesn't make it so
[17:54] <Dark_Shikari> yup
[17:54] <Dark_Shikari> having a qual task that leads into the main project is the best idea of mine IMO
[17:54] <Dark_Shikari> it gives people a feeling that they're actually part done with the project
[17:54] <Dark_Shikari> before SOC even starts
[17:54] <Dark_Shikari> and people are less likely to quit in the middle
[17:55] <Dark_Shikari> as opposed to simply bullshitting and quitting after getting nothing done
[17:55] <DonDiego> we need to work on our soc, it's in a sad state and not improving
[17:55] <DonDiego> also, our tasks are *much* too hard
[17:55] <Dark_Shikari> well, here's my advice for improving ffmpeg SOC
[17:55] <DonDiego> basically the only student that ever finished a decoder is kostya
[17:55] <DylanZA> just saw this conversation - I rate that ffmpeg makes you jump through too many hoops to get qualified (from personal experience). you never want to do a patch review just to qualify to maybe do something
[17:55] <Dark_Shikari> 1) Adopt the x264 qualification task method: qualification tasks consist of small, relatively easy subsets of the main project the student wants to apply for.
[17:55] <Dark_Shikari> No patch review is required for qualification tasks.
[17:55] <Dark_Shikari> Well, the mentor should review of course
[17:55] <Dark_Shikari> in order to judge the student's ability
[17:55] <Dark_Shikari> but it shouldn't be committed yet
[17:56] <Dark_Shikari> 2) Encourage interaction on IRC.  People who can't keep up to speed on a daily basis shouldn't be students.
[17:56] <DonDiego> DylanZA: i disagree, we still don't get enough capable students..
[17:56] <Dark_Shikari> DonDiego: yes, because capable ones don't like the hoops
[17:56] <Dark_Shikari> DylanZA was one of our successes last year
[17:56] <Dark_Shikari> he picked x264 over ffmpeg
[17:56] <DonDiego> hmm
[17:57] <DonDiego> qual tasks were started after the first year
[17:57] <DylanZA> most reasonable students will pick a couple of projects anyway
[17:57] <DonDiego> to weed out incompetent students
[17:57] <Dark_Shikari> DonDiego: IMO there are two requirements
[17:57] <Dark_Shikari> 1) competent/self-motivated
[17:57] <Dark_Shikari> 2) involved
[17:57] <DylanZA> but if one is too difficult then they choose another. but the poor students might just stick to one project and try their luck
[17:57] <Dark_Shikari> unless you have both, you will have a failure
[17:58] <DylanZA> Dark_Shikari: yuo also need luck
[17:58] <Dark_Shikari> this is why I'm biased against indian/chinese students, they do not tend to be very involved
[17:58] <DonDiego> you probably have a point there..
[17:58] <DylanZA> obviously any student can just pull out at some point
[17:58] <Dark_Shikari> DylanZA: of course, but it's far less likely that a good one will
[17:58] <Dark_Shikari> especially after completing a large part of their stuff
[17:58] <peloverde> You need somebody with "Excellent English skills"
[17:58] <DonDiego> after looking at our soc results, i share that bias
[17:58] <DonDiego> i guess the money is just too tempting

[18:01] <Dark_Shikari> a student should be willing to come on IRC and talk about what he's doing
[18:01] <Dark_Shikari> and not just willing, but he should do it without asking
[18:01] <Dark_Shikari> if he's stumped, he should come on and discuss what the best way to do X is
[18:01] <Dark_Shikari> and others should be there to help
[18:01] <DonDiego> i think the mentors are failing in that regard as well
[18:02] <Dark_Shikari> that's a problem created due to lack of use of IRC
[18:02] <DonDiego> unless everybody is talking to their students in private
[18:02] <DonDiego> which would kind of be a waste
[18:02] <Dark_Shikari> our students got a lot of help on IRC even when mentors weren't available
[18:02] <Dark_Shikari> e.g. holger wasn't always there to mentor Yuvi
[18:02] <DonDiego> that could happen on the ml as well, it's not tied to irc
[18:02] <Dark_Shikari> I don't think an ML is a suitable replacement for IRC.
[18:02] <Dark_Shikari> for some types of discussion, realtime is much preferable

[18:05] <Dark_Shikari> one thing I notice about ffmpeg
[18:05] <Dark_Shikari> at least half the comments are utterly trivial shit
[18:05] <DonDiego> i think michael pulled it from some javadoc style guide
[18:05] <Dark_Shikari> not even minor optimizations but things like periods in comments
[18:05] <Dark_Shikari> it's all bikeshed stuff.
[18:05] <DonDiego> nonsense
[18:05] <DonDiego> don't blow things out of proportion
[18:05] <DonDiego> it's good to have standards for such stuff
[18:06] <Dark_Shikari> true, but I think too much effort is spent debating and doing such things.
[18:06] <DonDiego> the bikeshedding starts when you back and forth endlessly about it
[18:06] <DonDiego> such things should thus be settled *once* and for all without much debate
[18:06] <Dark_Shikari> this actually reminds me of wikipedia
[18:06] <Dark_Shikari> so, on wikipedia, there were a billion things that needed work
[18:07] <Dark_Shikari> these things ranged from simple spelling correction
[18:07] <Dark_Shikari> to doing the research and writing for high-quality featured articles.
[18:07] <Dark_Shikari> I had an extremely strong feeling that a lot of people worked on things like the former to avoid Hard Stuff like the latter.
[18:07] <Dark_Shikari> people found it easier to run autowikibot on 30,000 articles than to do actual important things