[Ffmpeg-devel-irc] ffmpeg-devel.log.20180826

burek burek021 at gmail.com
Mon Aug 27 03:05:04 EEST 2018


[01:22:16 CEST] <jamrial> tmm1: would be neat if you could test the configure patches to see if they don't break detection/deps of apple framework stuff
[02:49:44 CEST] <bofh_> atomnuker: so if you keep around 2 different copies of fft5 you can do the rest via reindexing (also, seriously, hardcode the twiddles *into* the fft5 part of that code instead of regenerating them at runtime, it's like 4 values). I don't know if you can do better, I'm trying to get a general result right now on feasibility of forward/inverse FFTs via just reindexing.
[02:49:59 CEST] <bofh_> (also sorry for the delay, I've been at a superconductivity conference in Beijing for the past few days)
[02:53:29 CEST] <atomnuker> wish I was there
[02:54:58 CEST] <atomnuker> yeah, the 5-point fft is small enough to template, I assume all you need to do is change the sign
[02:55:36 CEST] <atomnuker> can you tell me what changes to make or submit a diff maybe?
[12:34:31 CEST] <cone-765> ffmpeg 03Nicolas George 07master:962c9313af31: lavfi/avf_concat: switch to activate.
[16:48:49 CEST] <durandal_1707> Compn: you have commit rights for mplayer, so commit fixed imm4 patch
[18:05:58 CEST] <cone-306> ffmpeg 03Paul B Mahol 07master:c84a57d1a218: avcodec/mscc: fix several bugs
[18:08:38 CEST] <cone-306> ffmpeg 03Paul B Mahol 07master:3aacb0d196d1: avcodec/proresdec2: add frame threading support
[18:09:19 CEST] <BradleyS> ^ ooo
[18:09:40 CEST] <BradleyS> is there a ML thread or other information with benchmarks?
[18:10:19 CEST] <durandal_1707> BradleyS: why you ask?
[18:10:39 CEST] <BradleyS> because i use prores a lot and run it through handbrake a lot
[18:11:16 CEST] <durandal_1707> then test latest ffmpeg master and report your findings
[18:11:33 CEST] <BradleyS> of course :)
[18:15:54 CEST] Action: BradleyS wonders whether durandal_1707 is the author
[18:16:15 CEST] <BradleyS> i haven't put together everyone's name/handle combination quite yet
[18:18:02 CEST] <kierank> he is author of thread patch yes
[18:18:29 CEST] <BradleyS> cool o/
[18:18:48 CEST] <BradleyS> adding it to my long todo list, will do some testing here
[18:22:16 CEST] <BradleyS> when i asked about benchmarks i of course assumed the patch was a bit more involved, heh
[18:23:52 CEST] <durandal_1707> it depends how many CPUs one have
[18:27:33 CEST] <durandal_1707> here on my CPU, with 4 cores, i get speed increased from 693 fps (slice mt) to 721 fps (frame mt) for static hd720 video
[18:29:24 CEST] <BradleyS> i can test 720/1080 on my 6 core westmere, and probably 4k when new camera stuffs arrive soonish (or using test media)
[19:52:48 CEST] <durandal_1707> Dav1d is written in Rust
[19:53:14 CEST] <JEEB> funky
[19:54:09 CEST] <atomnuker> pretty sure it isn't
[19:54:22 CEST] <JEEB> that's what I would have expected
[19:54:41 CEST] <JEEB> that said I also don't know what durandal_1707 based that sentence on
[19:55:51 CEST] <durandal_1707> it is decoder for rav1e
[19:58:16 CEST] <kurosu> Speak of the devil
[20:51:22 CEST] <durandal_1707> BBB: is Dav1d going to be sold like that vp9 encoder?
[20:52:10 CEST] <atomnuker> no, it'll be released as a bsd 2 clause code
[20:52:25 CEST] <BBB> BBB is short for beelzebub, which is actually one of the fallen angels, yes :-p
[20:52:29 CEST] <BBB> so I guess I am a devil
[20:53:11 CEST] <BBB> Im not in charge of license details, ask, j-b
[20:53:27 CEST] <durandal_1707> thanks
[20:58:06 CEST] <durandal_1707> ubitux: is lut1d patch fine, should i use LUT3DContext instead?
[21:02:02 CEST] <kierank> durandal_1707: how do I make fate test
[21:02:08 CEST] <kierank> for mpeg4video
[21:02:33 CEST] <JEEB> I think the latest thing is the md5sum based stuff?
[21:02:53 CEST] <kierank> mpeg4 is framecrc
[21:03:00 CEST] <JEEB> yea because it's older
[21:03:05 CEST] <JEEB> (than the md5 stuff)
[21:03:38 CEST] <JEEB> also an alternative is to make the test a separate C app
[21:03:41 CEST] <JEEB> like with movenc
[21:03:47 CEST] <kierank> I don't need that afaik
[21:03:49 CEST] <kierank> it's one frame
[21:04:06 CEST] <JEEB> yeh, framemd5 or something sounds good enough then.
[21:04:44 CEST] <JEEB> for something like a single frame test it might even be nice if we had per-plane and possibly even per-line hashes or something
[21:04:50 CEST] <JEEB> but I don't think we have that :)
[21:06:52 CEST] <durandal_1707> per pixel component sha1024 hashes are not enough?
[21:07:23 CEST] <JEEB> heh
[21:10:07 CEST] <kierank> atomnuker: ping
[21:10:32 CEST] <durandal_1707> someone please reply to MWSC and WCMV decoders patches with just LGTM, so I can push them and move on with *SECRET*
[21:13:59 CEST] <kierank> https://obe.tv/clang/scan-build-2018-08-26-195358-16089-1/index.html
[21:14:09 CEST] <kierank> most are junk but some are real
[21:19:06 CEST] <atomnuker> yeah, that's true, I'll fix some in my code
[21:19:31 CEST] <kierank> atomnuker: do you have any comment on the vc2enc one?
[21:19:42 CEST] <kierank> is it actually real
[21:19:58 CEST] <JEEB> yea, scan-build has a lot of misdetections, but it's a nice OSS analyzer
[21:20:05 CEST] <JEEB> which reminds me, clang 7 is at rc2
[21:24:57 CEST] <atomnuker> kierank: actually both the one in opusenc and vc2enc are bogus
[21:25:12 CEST] <atomnuker> in vc2enc  if (quant_buf[1] == quant) { is never going to be true for the first slice
[21:25:18 CEST] <atomnuker> as quant_buf[1] = -1
[21:25:30 CEST] <atomnuker> and quant is clipped like av_clip(quant + signed_step, 0, s->q_ceil-1);
[21:25:33 CEST] <kierank> you can add an assert I think
[21:25:38 CEST] <kierank> and it will go away
[21:58:01 CEST] <atomnuker> "Twitch is building a VP9 live-encoding solution that can deliver broadcast-level video quality with 25+% bitrate savings"
[21:58:07 CEST] <atomnuker> yah, another encoder
[21:58:17 CEST] <atomnuker> "Based on our feasibility study, FPGA is the only available solution in the market right now"
[21:58:23 CEST] <atomnuker> nevermind
[21:58:30 CEST] <kierank> hahahaha
[21:58:31 CEST] <kierank> they are idiots
[21:58:35 CEST] <kierank> going to cost them soo much
[21:58:36 CEST] <JEEB> ah, oppan FPGA style
[21:58:51 CEST] <atomnuker> https://www.meetup.com/SF-Video-Technology/events/wlnthpyxlbnc/
[21:59:15 CEST] <kierank> it cannot be cost effective for them to do this
[21:59:32 CEST] <atomnuker> "Yueshi will explain the thought process of selecting VP9 as Twitch's next-gen video codec format" <- couldn't they wait a bit for av1?
[21:59:55 CEST] <jamrial> atomnuker: no hardware until like 2020
[22:00:13 CEST] <JEEB> yea, the decoders for VP9 have been on the market since 2016 at the very least
[22:00:14 CEST] <jamrial> nvidia is about to release their new cards, obviously without av1
[22:00:18 CEST] <JEEB> (my 2016 TV has support for it)
[22:00:27 CEST] <JEEB> and nvidia encoders, yea
[22:00:38 CEST] <atomnuker> can't they implement a partial cuda decoder like they've done before?
[22:00:48 CEST] <JEEB> although did they have VPx encoders in their cards yet?
[22:00:55 CEST] <JEEB> I only saw AVC and HEVC generally
[22:00:57 CEST] <jamrial> so you'd have to wait until the following gen
[22:00:57 CEST] <kierank> jamrial: why do they need hardware
[22:01:03 CEST] <kierank> tvs can be upgraded
[22:01:04 CEST] <kierank> browsers can
[22:01:08 CEST] <kierank> even phones
[22:01:22 CEST] <kierank> atomnuker: probably nobody is using the AWS fpgas so they are getting them for free
[22:01:32 CEST] <JEEB> :D
[22:01:50 CEST] <jamrial> kierank: hardware decoders. you can't add transistors to a phone or set top box already in the market
[22:02:01 CEST] <JEEB> reminds me of another live streaming service advertising some hardware encoding solution they picked
[22:02:04 CEST] <jamrial> software decoder sure. just update libavcodec in your browser
[22:02:05 CEST] <kierank> phones are very powerful these days
[22:02:24 CEST] <kierank> probably the saved rf bandwidth will save more power than the cpu cost
[22:02:44 CEST] <kierank> twitch is unrelated to set top boxes
[22:02:50 CEST] <JEEB> true
[22:02:55 CEST] <JEEB> it's mostly mobile and browsers
[22:02:59 CEST] <kierank> and smart tvs
[22:03:13 CEST] <JEEB> and on those VP9 is already
[22:06:55 CEST] <durandal_1707> multimedia is no longer relevant
[22:07:09 CEST] <JEEB> right. we can then all go home
[22:08:08 CEST] <durandal_1707> i hate object based languages and javascript
[22:08:51 CEST] <kierank> durandal_1707: no job at google for you then
[22:08:55 CEST] <kierank> it's c++ minimum there
[22:11:54 CEST] <durandal_1707> kierank: i got contacted by them again, (not counting cases prior to joining FFmpeg)
[22:12:49 CEST] <JEEB> golang, javascript, C++ seem to be the usual cases at GOOG
[22:53:47 CEST] <durandal_1707> what os is OS/2 ?
[22:56:02 CEST] <JEEB> OS/2 is Ossi, from the 1990s. Koreans still seem to be maintaining it
[22:56:10 CEST] <JEEB> which is why it was surprising that the guy who tested it was !Korean
[22:56:19 CEST] <JEEB> because all of the OS/2 patches for VLC and FFmpeg seemed to come from Korea
[23:02:00 CEST] <Shiz> there's a commercial OS/2 distri again these days
[23:02:03 CEST] <Shiz> it's crazy
[23:10:16 CEST] <durandal_1707> looks like nobody here cares for Dolby Atmos and Dolby Vision?
[23:12:02 CEST] <nevcairiel> some of my users care and occasionally annoy me
[23:26:44 CEST] <TD-Linux> someone make a bsf that make the dolby atmos light go on but otherwise does nothing
[23:27:12 CEST] <JEEB> :D
[23:27:24 CEST] <JEEB> yea the atmos light seems to be the only thing people care about
[23:30:01 CEST] <nevcairiel> i had to do some work on bitstreaming to fit the higher bitrate truehd streams into there properly, but that was about all
[00:00:00 CEST] --- Mon Aug 27 2018


More information about the Ffmpeg-devel-irc mailing list