[Ffmpeg-devel-irc] ffmpeg-devel.log.20161223

[burek]

[02:56:46 CET] <slavanap> Hello. I want to use ffmpeg to retrieve picture for every frame as (byte array + image format). Is there examples in C/C++? Does ffmpeg support seeking to specific frame natively? How ffmpeg handle 3D videos (if implemented already)? Thanks!
[02:58:57 CET] <cone-599> ffmpeg 03Thomas Turner 07master:8dcb28cf6dd1: avutil/tests: Improved code coverage for random_seed
[03:44:10 CET] <Compn> depends on the format of the 3d video
[03:44:14 CET] <Compn> e.g. 3d bluray not supported
[03:44:27 CET] <Compn> i think most / all other 3d formats are supported via filters
[03:44:55 CET] <Compn> oh shit he left :\
[04:07:51 CET] <cone-599> ffmpeg 03Nicolas George 07master:8156b5ac9436: avfilter/af_amerge: detect EOF immediately
[08:38:42 CET] <atomnuker> lossless audio encoding shootout:
[08:39:02 CET] <atomnuker> 5th place with 43016 bytes, the ffmpeg flac encoder at the highest compression_level
[08:39:56 CET] <atomnuker> 4th place with 42440 bytes, flacenc at the highest encoding settings (32 order LPC, all exaustive search options, took 10 minutes to encode)
[08:40:36 CET] <atomnuker> 3rd place with 42144 bytes, the ffmpeg tta encoder, took less than 10 seconds to encode
[08:41:00 CET] <atomnuker> 2nd place with 41868 bytes, the libwavpack encoder
[08:41:34 CET] <atomnuker> 1st place with 41788 bytes, the ffmpeg wavpack encoder (highest settings, took 10 minutes to encode)
[08:42:36 CET] <atomnuker> durandal_1707, you take a bronze
[08:43:33 CET] <atomnuker> (I'd never have thought TTA was advanced enough to beat FLAC)
[09:25:17 CET] <durandal_1707> atomnuker: how long is input audio?
[09:25:42 CET] <atomnuker> 6 minutes 30 seconds, rock
[09:25:53 CET] <durandal_1707> you forgot sonicls
[09:26:58 CET] <durandal_1707> michaelni doesn't want to remove that obscure piece of shit
[09:27:38 CET] <atomnuker> 45100 bytes isn't bad, certainly better than MLP (60 megs)
[09:28:08 CET] <atomnuker> crap, I mean KILObytes
[09:29:02 CET] <atomnuker> for everything because du likes to not append the units
[09:30:37 CET] <j-b> atomnuker: nice comparison
[09:45:56 CET] <cone-363> ffmpeg 03James Almer 07master:41de965e31ef: avcodec/pixlet: remove unnecessary double to float conversion
[09:45:56 CET] <cone-363> ffmpeg 03Paul B Mahol 07master:51ed1a7dd40a: avcodec/pixlet: check out of bounds pfx value
[11:19:29 CET] <cone-363> ffmpeg 03Paul B Mahol 07master:0ed4993ad0e2: avcodec/asvdec: return correct value in case of error
[13:06:36 CET] <durandal_1707> anyone wants to write utvideo assembly?
[13:44:50 CET] <wm4> atomnuker: tta takes 10secs, ffwavpack 10 minutes? wtf
[13:46:01 CET] <nevcairiel> should also try the alien magic mode in  ffflacenc
[13:46:06 CET] <nevcairiel> i hear its super slow
[13:46:06 CET] <nevcairiel> :D
[13:51:17 CET] <wm4> more importantly, why does ffmpeg even have a tta encoder
[13:52:56 CET] <nevcairiel> because paul was bored one day and wrote one?
[13:53:26 CET] <nevcairiel> its only 200-something lines including  comments and whatnot, its not l ike its a complex format
[13:53:36 CET] <nevcairiel> works surprisingly well for that
[15:28:05 CET] <cone-363> ffmpeg 03Nicolas George 07master:373f21a91126: lavfi: avfilter_graph_request_oldest: request a frame again before returning.
[17:19:31 CET] <BBB> durandal_1707: utvideo & I mean
[17:20:19 CET] <durandal_1707> BBB: ?
[17:20:27 CET] <BBB> utvideo assembly
[17:20:35 CET] <BBB> that suggests utvideo is worth writing assembly for :-p
[17:20:37 CET] <BBB> is it?
[17:21:17 CET] <durandal_1707> well I see that there is already assembly for it in huffyuv
[17:22:26 CET] <durandal_1707> only median was slow, and I improved it by 10%
[17:22:50 CET] <durandal_1707> left prediction did not made difference
[17:26:05 CET] <durandal_1707> thing is that internally rgb is stored as planar
[17:26:26 CET] <durandal_1707> and we decode it as packed
[17:26:46 CET] <durandal_1707> this is swscale job
[17:27:02 CET] <durandal_1707> similar is in encoder
[17:44:23 CET] <cone-363> ffmpeg 03Paul B Mahol 07master:ea93052db359: avcodec/utvideodec: add SIMD support for median prediction for planar formats
[18:04:41 CET] <kierank> durandal_1707: ok christmas now, I can finish mpeg-4 sstp decoder...
[18:07:00 CET] <durandal_1707> kierank: you have samples or encoder?
[18:07:40 CET] <kierank> there are samples on trac ticket
[18:07:54 CET] <kierank> spec is not hard, integrating 10/12-bit into mpegvideo context is hard
[18:35:30 CET] <durandal_1707> is mmxext slow?
[18:38:15 CET] <kierank> it's a bit old
[18:39:26 CET] <durandal_1707> but can mmxext assembly be made faster?
[18:41:28 CET] <kierank> why not rewrite it in sse
[18:41:31 CET] <jamrial> if you make it use xmm regs sure
[18:43:08 CET] <kierank> yeah you've written it directly in mmx
[18:54:49 CET] <cone-363> ffmpeg 03Paul B Mahol 07master:9b26bf7e2a39: avfilter/vf_deband: do not use uninitialized value
[19:01:01 CET] <durandal_1707> jamrial: please see if median prediction can be made faster
[21:03:27 CET] <atomnuker> lol, the lavc fft supports framesizes up to 131072
[21:04:21 CET] <atomnuker> too bad I have to rewrite the entire fucking opus fft
[21:05:26 CET] <atomnuker> it's simply not usable as a forward mdct because someone thought it would be clever to optimize the twiddles
[21:05:57 CET] <atomnuker> rather than, I dunno, not use an incredibly inefficient recursive fft based on multiples of 15
[21:06:53 CET] <atomnuker> worst part is there's some useless aarch64 assembly so I'll still have to keep the old twiddle generating code in that init function
[21:07:34 CET] <atomnuker> (which is barely 1.4 times faster, which makes me think I can beat it with a better C code and delete that shit)
[21:14:53 CET] <atomnuker> libav merges messed the shit out of all the ffts and dcts
[21:18:15 CET] <durandal_1707> atomnuker: what exactly?
[21:50:57 CET] <cone-363> ffmpeg 03Michael Niedermayer 07master:c193132ce590: avutil/tests/random_seed: eliminate goto
[21:50:58 CET] <cone-363> ffmpeg 03Tobias Rapp 07master:6d579d7c1bdc: avformat/avidec: skip odml master index chunks in avi_sync
[21:50:59 CET] <cone-363> ffmpeg 03Michael Niedermayer 07master:03ce71e4a118: swscale/swscale: Fix dereference of stride array before null check
[22:45:04 CET] <TD-Linux> atomnuker, how is the fdct perf compared to libopus? any reason not to just steal it directly?
[22:46:55 CET] <atomnuker> TD-Linux: yah, it integrates windowing in the transform which I'm not a fan of since it's not reusable for other codecs
[22:47:30 CET] <atomnuker> I *could* just copy and paste kiss-fft and do my own pre/post shuffles but nah
[22:48:00 CET] <atomnuker> I wanna see if I can do better than both since you can implement the FFT as an in-place transform which should be faster
[00:00:00 CET] --- Sat Dec 24 2016