[Ffmpeg-devel-irc] ffmpeg-devel.log.20180409

[burek]

[03:48:19 CEST] <cone-387> ffmpeg 03Michael Niedermayer 07master:249aca8f98ff: avcodec/movtextdec: Check style_start/end
[03:48:19 CEST] <cone-387> ffmpeg 03Meng Wang 07master:3b2fd960481d: avcodec/arm/hevcdsp_sao : add NEON optimization for sao
[04:14:54 CEST] <jamrial> blegh, checkasm for hevc_mc. less chances now to see the conflicting mc asm between ffmpeg and libav dealt with
[10:09:01 CEST] <CounterPillow> Do I smell a 4.0? Of course I don't, it'll never be released
[10:16:15 CEST] <nevcairiel> everytime someone asks, we delay by a week
[10:23:25 CEST] <CounterPillow> Pretty sure ffmpeg developers have no concept of time, so I doubt this
[14:04:45 CEST] <durandal_1707> cehoyos is very *high*
[18:02:24 CEST] <BBB> jamrial: cant you get that hevc asm person to actually implement the x86 parts (or port them, from libav) also?
[18:02:32 CEST] <BBB> jamrial: you dont have to do everything yourself, y'know
[18:07:05 CEST] <jamrial> i know
[18:08:20 CEST] <jamrial> thing is we don't even know if it's faster. we'd only know after they are ported, and if they are slower then it would result in a lot of work wasted
[18:08:51 CEST] <jamrial> it's not just the asm, it's also the entire mc dsp. libav rewrote it from scratch
[18:10:18 CEST] <iive> can't we benchmark their decoder vs out decoder?
[18:11:28 CEST] <BBB> jamrial: I think at some level maybe not everything has to be measured. what we do by default is nothing, either having someone rewrite it because (s)he wants to or take it from libav or both better
[18:11:31 CEST] <jamrial> apples to oranges. they don't even support rext, so the entire decoder is pretty different
[18:11:34 CEST] <BBB> so if someone wants to do either, lets do it
[18:15:07 CEST] <iive> jamrial, what is rext and how does this make things different?
[18:16:16 CEST] <iive> if both decoders decode same stream and produce same output, then we can benchmark. well, if our decoder is faster and more complete... then there is nothing to do more
[18:16:23 CEST] <jamrial> iive: format range extensions. things like 12bit and such. dsp modules of course are adapted to handle it
[18:16:52 CEST] <jamrial> overall decoding speed doesn't tell you if their mc implementation is better than ours, you know
[18:17:10 CEST] <iive> so it won't produce same output
[18:18:43 CEST] <iive> why is idct /quant/ etc also different ?
[18:37:31 CEST] <iive> at the end, it is the over performance that matters
[18:41:08 CEST] <Chloe> I looked at the libav/hevc asm in the past
[18:41:23 CEST] <Chloe> libav/ffmpeg hevc*
[18:42:26 CEST] <Chloe> I remember a lot of their asm still being just older ffmpeg versions. I ported a bit of our asm to libav
[18:46:09 CEST] <jamrial> they rewrote the mc dsp and then wrote asm for it. whereas we just took openhevc development as is
[18:52:38 CEST] <iive> was their old mc dsp similar to ours/openhevc?
[18:56:42 CEST] <jamrial> the initial decoder was the same in both projects. we then took openhevc's changes whereas they didn't
[18:57:24 CEST] <iive> so... benchmark their decoder before mc dsp asm changes and after.
[18:57:39 CEST] <iive> but i guess they have put a ton of other changes around too.
[19:14:16 CEST] <BBB> iive: I think youre missing a major point here - we all know what needs to be done, its rather finding someone to actually do it
[19:16:40 CEST] <daddesio> is all of the code you're talking about contained in this file? https://github.com/libav/libav/blob/master/libavcodec/x86/hevc_mc.asm
[19:19:23 CEST] <BBB> yes
[19:19:57 CEST] <daddesio> interesting, FFmpeg seems to already have that file, but the implementation is totally different (and a different author).
[19:20:19 CEST] <kurosu> if there was an incentive for ffhevc besides arm vendors, it has likely passed
[19:20:24 CEST] <jamrial> no, also outside the x86 folder
[19:20:30 CEST] <jamrial> it's the entire mc module
[19:20:45 CEST] <kurosu> daddesio, libav never picked up the openhevc code (or openhevc authors never submitted them patches)
[19:21:33 CEST] <kurosu> another point, object code size, last I checked it was like 14x kB vs 5x kB
[19:22:40 CEST] <kurosu> (the later being libav)
[19:29:13 CEST] <BBB> hm?
[19:29:27 CEST] <BBB> kurosu: why do you say that? broadcast could use ffhevc also as a base for x86 implementations?
[19:29:48 CEST] <BBB> my impression was that hevc is fairly popular in broadcast
[19:30:16 CEST] <BBB> hi btw :) long time no see
[19:30:26 CEST] <kurosu> I don't know, just observing symptoms as effectively, nobody pushing its development in ffmpeg
[19:30:40 CEST] <kurosu> I suspect h/w decoders have become mainstream faster also
[19:30:48 CEST] <BBB> true
[19:31:21 CEST] <kurosu> BBB: hi too - yes, paying a visit here and there, and today's one
[19:31:53 CEST] <BBB> are you excited about av1 yet?
[19:32:51 CEST] <kurosu> midly - as for any codec there will be stuff broken, and as long as I don't have to deal with the broken stuff... :)
[19:33:09 CEST] <BBB> hehe :)
[19:33:23 CEST] <kurosu> for pure coding performance, no
[19:33:38 CEST] <kurosu> for pure deployment, at least I want to see how ffav1 comes to be
[19:34:29 CEST] <kurosu> I won't comment more on it, but overall, it's mild also on this side :(
[19:35:42 CEST] <jamrial> my excitement is slightly toned down by the fact libaom, at least at first, will be absurdly slow at encoding anything
[19:36:47 CEST] <kurosu> I expect all codecs from now on to likely be, or at least not reach the advertised performance
[19:37:45 CEST] <kurosu> some people in here are attending NAB I guess
[19:38:09 CEST] <BBB> libaom at speed=1 isnt too terrible
[19:38:19 CEST] <BBB> I mean, its still a minute per frame or so, but for short testing clips, thats ok
[19:38:20 CEST] <jamrial> kurosu: you mean things like all the "50%!!!one" marketing?
[19:38:33 CEST] <BBB> kurosu: Ill be at NAB tomorrow, yes
[19:39:14 CEST] <kurosu> jamrial, yes, but physically impossible - like x265 is 10-20% below the advertised hevc just because hm is meant to produce the best numbers, not realistic ones
[19:39:44 CEST] <jamrial> BBB: consider using 2 instead. it recently got some tweaks and it's considerably faster than 1 at little quality cost
[19:39:51 CEST] <kurosu> BBB, cool, have you ever been there? It seemed like it wouldn't be your type
[19:39:54 CEST] <jamrial> sort of like placebo vs veryslow with x264
[19:40:49 CEST] <BBB> kurosu: never been, first time
[19:41:17 CEST] <kurosu> I think in no occasion would I enjoy Las Vegas
[19:41:17 CEST] <BBB> kurosu: its probably not what I like, but you know how these things go, sometimes you gotta do what you gotta do
[19:41:45 CEST] <kurosu> yeah, that's a great opportunity for you
[19:42:43 CEST] <kurosu> well, looking forward for ave (caesar, fitting for LV)
[19:42:54 CEST] <kurosu> s/for/to
[19:44:29 CEST] <BBB> :)
[19:44:51 CEST] <BBB> in general its good to be at a venue where people are excited about the thing youre trying to make a business out of
[19:45:13 CEST] <kurosu> so, jamrial, still not voted as the new ffmpeg leader, btw? maybe that concept should be deprecated
[19:45:49 CEST] <jamrial> why would i even be voted for that?
[19:46:11 CEST] <kurosu> nor for wanting, just for general involvement and expertise
[19:49:33 CEST] <BBB> jamrial: he has a point
[19:49:59 CEST] <kurosu> also, sometimes I think someone should tell both sides of an argument to shut up - but that's maybe my hubris talking
[19:50:17 CEST] <BBB> no, youre right
[19:50:49 CEST] <jamrial> well, thanks, but that's not something i want
[19:51:07 CEST] <kurosu> I bet - nobody actually wants to deal and be subjected to this
[19:51:59 CEST] <kurosu> at least, I expect a lot of people to value your opinion and be amenable to reason through you
[19:52:33 CEST] <kurosu> anyway, just idle chat which is starting to be awkward
[19:56:21 CEST] <BBB> maybe we value him more because hes not acting like a power-hungry maniac
[19:56:34 CEST] <BBB> <3
[19:56:45 CEST] <BBB> benevolent dictator
[20:00:54 CEST] <kurosu> there must be some latin or greek parabola about power on this exact topic
[20:12:27 CEST] <daddesio> parabola, eh? *googles* today I learned about the greek roots of the conic sections: https://english.stackexchange.com/q/175756
[20:12:58 CEST] <wm4> yeah jamrial would make a good leader
[20:13:30 CEST] <wm4> he's about the only one on ffmpeg-devel who shows at least some skill (or even just attempt) to mediate conflicts
[20:14:06 CEST] <durandal_1707> *sigh*
[20:37:15 CEST] <atomnuker> kurosu: are you that worried that the broadcast people you work for will actively spy on IRC for any ill comments you make?
[20:39:46 CEST] <kurosu> atomnuker, no, I'm just trying not to enter any such game
[20:42:21 CEST] <iive> they don't have to spy, just google his nick...
[00:00:00 CEST] --- Tue Apr 10 2018