[Ffmpeg-devel-irc] ffmpeg-devel.log.20151230

burek burek021 at gmail.com
Thu Dec 31 02:05:03 CET 2015


[02:27:12 CET] <kierank> so I need to output a frame that's already been output (essentially a skip frame)
[02:27:22 CET] <kierank> I can do that but just storing the avframe right?
[03:27:26 CET] <J_Darnley> Hm.  So close.  Just some incorrect chroma to go.
[03:39:01 CET] <atomnuker> I never thought people with so much paranoia about integer overflow existed
[03:41:00 CET] <J_Darnley> Where did I see that the other day?  (Some shitpost on HN no doubt.)  Undefined behaviour should start nethack.
[03:41:41 CET] <atomnuker> "It's undefined behavior, this is insane, anything can happen!"
[03:42:00 CET] <J_Darnley> Ah the daala thread?  I've been keeping it collapsed in Thunderbird.
[03:49:22 CET] <BBB> atomnuker: sorry about being so persistent in that thread
[03:49:32 CET] <BBB> I just really dont want us to fall in the trap of appeasing ubsan
[03:49:36 CET] <BBB> its not useful
[03:51:19 CET] <jamrial_> atomnuker: leaving the whole thing about overflow with invalid input on dsp code not being important aside, the decoder should definitely not crash in a normal run with invalid/fuzzed input
[03:51:30 CET] <jamrial_> no idea if Andreas' crashes were all ubsan related, though
[03:54:35 CET] <atomnuker> BBB: I agree with you, please keep being persistant
[03:55:19 CET] <atomnuker> jamrial_: It's an RFC and the first time the decoder's been sent to the ML, let alone fuzzed
[03:58:31 CET] <jamrial_> i know, just pointing it out
[04:15:28 CET] <BBB> crashes should be fixed, and Im not arguing against fixing overflows in non-dsp code
[04:15:59 CET] <BBB> Im just saying that dsp code overflows are probably ok, its quite likely we clip at the end of the dsp code anyway (simd code always clips when packing)
[04:16:17 CET] <BBB> so although overflows occur, the output range is still defined
[04:16:34 CET] <BBB> atomnuker: is it OK if I leave a full review for over the weekend?
[04:16:47 CET] <BBB> (you can commit before that if you wish, but I dont think Ill have time during the week before that)
[04:18:41 CET] <atomnuker> BBB: Yeah, sure, I'll have time to fix whatever I can by then and send off a v3
[09:14:59 CET] <fritsch> nevcairiel: https://github.com/mpc-hc/FFmpeg/commit/ecc953ccbc2370c8177deb4077fde1007ec0bb3b <- mind PRing that to ffmpeg master?
[09:16:19 CET] <fritsch> nevcairiel: and the corresponding dxva2_hwaccel, too, please?
[09:55:59 CET] <nevcairiel> there are reasons that hasn't happened yet, just unlocking 10-bit hevc dxva2 will break stuff
[10:16:47 CET] <xintox> why do i have to pass \r\n manually to headers? that seems like ffmpeg should just insert those
[10:25:10 CET] <fritsch> nevcairiel: can you be more concrete? does that mean the branch I linked ships broken stuff?
[10:26:08 CET] <nevcairiel> that branch is generally designed to work in exactly one use-case
[10:26:32 CET] <fritsch> okay - any hints to what it actually breaks?
[13:06:39 CET] <J_Darnley> Is there a black and white or greyscale vsynth test in fate?
[13:39:31 CET] <J_Darnley> nevermind, I'll just use testsrc and lut
[13:50:53 CET] <durandal_1707> ubitux: is overlap patch for showspectrum ok?
[14:02:37 CET] <J_Darnley> Can anyone point out the error in my inderstanding of the assembly code here http://pastebin.com/uUqWwSw6
[14:03:42 CET] Action: J_Darnley facepalms
[14:03:50 CET] <J_Darnley> (why do I see it only now?)
[14:04:25 CET] <J_Darnley> the high dqword starts two bytes further in the broken version
[14:04:37 CET] <BBB> J_Darnley: try pshufb :D
[14:04:59 CET] <J_Darnley> that doesn't cross lanes
[14:05:09 CET] <BBB> owh crap
[14:05:16 CET] <J_Darnley> indeed
[14:05:22 CET] <nevcairiel> evil lane business is evil
[14:05:27 CET] <J_Darnley> I think I grok them now.
[14:06:51 CET] <J_Darnley> Something like conjoined independant registers
[14:07:16 CET] <BBB> sorry boss, my video encoder can only encode videos up until 352x288 resolution
[14:07:38 CET] <nevcairiel> 20 years ago you might have won something with that
[14:07:43 CET] <BBB> oh, you want bigger video, HD you say? hm & sounds like PR bullshit to me, why dont you cut it up in many small videos and feed it to my encoder independently
[14:07:55 CET] <BBB> < $intel
[14:08:14 CET] <J_Darnley> Didn't ffmpeg+x264 do a BBB presentation like that at some conference?
[14:08:35 CET] <J_Darnley> (that would be big buck bunny)
[14:08:58 CET] <nevcairiel> now on to philosophy, did the big bunny or the BBB exist first
[14:09:13 CET] <BBB> BBB stems from the 90s, I think
[14:09:17 CET] <BBB> so probably me
[14:09:25 CET] <BBB> (BBB as in my nickname)
[14:13:38 CET] <J_Darnley> kierank: what profiling tools are available on avdev?
[14:15:25 CET] <J_Darnley> ah gprof's there
[14:32:59 CET] <ubitux> durandal_1707: why not use AVRational?
[14:33:05 CET] <ubitux> durandal_1707: how does it affect output?
[14:58:53 CET] <durandal_1707> ubitux: it gives higher framerate and more resolution
[14:59:22 CET] <ubitux> can you show 2 screenshots at the same t?
[15:00:18 CET] <durandal_1707> with and without it?
[15:00:54 CET] <durandal_1707> overlap 0 default, gives same output
[15:02:04 CET] <durandal_1707> 0.5 enlarges width by 100%
[16:47:19 CET] <kierank> J_Darnley: should be able to start-timer
[16:53:26 CET] Action: J_Darnley goes to look that up
[16:54:13 CET] <J_Darnley> Do you mean ffmpeg's timer.h tool?
[17:09:23 CET] <kierank> yes
[17:29:33 CET] <Daemon404> kierank, do you still need cfhd
[17:29:38 CET] <kierank> yes 
[17:29:42 CET] <kierank> I pinged koda dabout it
[17:29:44 CET] <kierank> but he is not alive
[17:30:05 CET] <Daemon404> food coma probably
[17:30:32 CET] <Daemon404> let me see if we have any in the 'shitty files' archive
[17:36:57 CET] <Daemon404> cant find one, might be blacklisted during ul
[17:47:49 CET] <cone-234> ffmpeg 03Ganesh Ajjanagadde 07master:29bc293297a6: configure: make gcc 2 detection more robust
[17:47:50 CET] <cone-234> ffmpeg 03Ganesh Ajjanagadde 07master:b272c3a5aabe: lavc/dsd_tablegen: speed up table generation
[17:54:36 CET] <cone-234> ffmpeg 03Ganesh Ajjanagadde 07master:05434b0eea3f: lavc/cook: get rid of wasteful pow in init_pow2table
[17:57:06 CET] <kierank> I wonder if we can convince ganesh to make thread safe table inits
[17:57:14 CET] <kierank> he seems interested in working on tables
[18:04:23 CET] <Daemon404> doubtful kierank 
[18:04:33 CET] <Daemon404> he seems mostly interested in 'interesting' 'optimizations'
[18:04:39 CET] <Daemon404> table init is a slog
[18:05:22 CET] <Daemon404> i started going through and marking some files here, but i got bored: https://github.com/dwbuiten/FFmpeg/commit/3d2fc03955348f49f2a6b67a66cc038d19ba62b1
[18:05:48 CET] <Daemon404> i dont know how elenril and co do it
[18:54:02 CET] <cone-234> ffmpeg 03Paul B Mahol 07master:2893dbfe2c31: avfilter/avf_showspectrum: add moreland color map
[19:08:28 CET] <durandal_170> ubitux: 0 overlap: http://i67.tinypic.com/2q3vgx4.png
[19:10:02 CET] <durandal_170> ubitux: 0.5 overlap: http://i65.tinypic.com/e1fbqa.png
[20:17:52 CET] <llogan> kierank: there was an actual award to collect?
[20:18:24 CET] <kierank> llogan: probably
[20:18:45 CET] <durandal_1707> for what?
[20:20:11 CET] <llogan> https://twitter.com/obencoder/status/682048485596737536
[20:20:15 CET] <cone-234> ffmpeg 03Michael Niedermayer 07master:389ce22b5f09: avformat/movenc: Use strict_std_compliance from the muxer layer instead of the encoder layer
[20:22:14 CET] <llogan> kierank: is streamingmedia legit?
[20:22:22 CET] <kierank> yes
[20:22:32 CET] Action: llogan is ignorant as usual
[20:23:00 CET] <kierank> well tbh they are legit enough to give us an award
[20:23:07 CET] <kierank> usually these things you pay money and get an award
[20:24:22 CET] <JEEB> wow
[20:24:29 CET] <llogan> so, who is in California?
[20:26:17 CET] <llogan> "Dear, please send monies to Dr. Elvis Guru, esteemed representative of FFmepg".
[20:27:41 CET] <llogan> looks like i have to clean up after roger again
[21:00:14 CET] <llogan> durandal_1707: ahistogram docs: "scroll from to bottom." looks like forgotten "top"?
[21:05:45 CET] <cone-234> ffmpeg 03Vicente Olivert Riera 07master:6a7e94cb14f3: mips: add mips32r6 architecture variant
[21:05:46 CET] <cone-234> ffmpeg 03Vicente Olivert Riera 07master:dde994751b16: mips: put "disable mipsfpu" in a better place for loongson
[21:19:08 CET] <J_Darnley> kierank: I've got a pair of avx2 commits you can have a look at here https://gitlab.com/J_Darnley/ffmpeg/commits/v210_asm
[21:19:18 CET] <J_Darnley> I'll be honest: they're pretty pathetic in terms of speedup.
[21:19:58 CET] <kierank> probably still worth it
[21:20:12 CET] <kierank> Might need to give you access to one of those intel boxes that downclock with avx2
[21:20:34 CET] <kierank> because the avx2 might completely negate itself
[21:20:51 CET] <J_Darnley> What!?
[21:21:24 CET] <J_Darnley> lol do they need an avx2slow flag?
[21:21:53 CET] <jamrial> no
[21:22:42 CET] <nevcairiel> "one of"? dont all do that?
[21:22:48 CET] <nevcairiel> well all that have a turbo mode
[21:22:50 CET] <jamrial> J_Darnley: use vpermq
[21:23:09 CET] <jamrial> instead of movq/movhps + vinserti128
[21:23:48 CET] <kierank> nevcairiel: only some SKUs
[21:23:53 CET] <jamrial> wait, nevermind. the mem operands you need don't allow that
[21:24:05 CET] Action: llogan remembers the "turbo" button on old 486
[21:24:32 CET] <jamrial> kierank: all haswell models?
[21:24:42 CET] <kierank> no afaik
[21:25:05 CET] <kierank> I think e5- range only
[21:25:19 CET] <kierank> it was in the x265 presentation at vdd
[21:25:20 CET] <nevcairiel> and the desktops?
[21:26:05 CET] <J_Darnley> jamrial: yes.  I had vpermq before and struggled to see the source of my problems.  Which was the +6 (basically).
[21:26:15 CET] <jamrial> J_Darnley: you can however use vextracti128 with a mem operand as dest
[21:26:29 CET] <J_Darnley> unaligned too?
[21:26:36 CET] <jamrial> yeah
[21:27:10 CET] <jamrial> every avx+ instruction except those that specify alignment (like movdqa and movaps) work with unaligned memory
[21:27:32 CET] <J_Darnley> What was I looking at earlier?
[21:27:47 CET] <jamrial> what do you mean?
[21:28:23 CET] <J_Darnley> Oh yes.  The Intel manual's large table of exception types.
[21:28:46 CET] <kierank> J_Darnley: feel free to submit to the ml
[21:28:58 CET] <kierank> maybe as an rfc
[21:29:01 CET] <cone-234> ffmpeg 03Paul B Mahol 07master:0f387f2b6265: avfilter/avf_showspectrum: add nebulae color map
[21:35:34 CET] <J_Darnley> OMFG!  That makes it look much better!
[22:20:16 CET] <durandal_1707> anybody likes any specific palette?
[22:29:54 CET] <J_Darnley> durandal_1707: have you got a "heat map" one?
[22:30:53 CET] <durandal_1707> shouldn't that be intensity, already available?
[22:31:19 CET] <J_Darnley> dark blue through red onto yellow?
[22:37:11 CET] <J_Darnley> Drat.  I need to wait for the next nightly build.
[22:40:02 CET] <J_Darnley> Anyway.  Here's what I meant by heat map (an excerpt from some old code of mine) http://pastebin.com/xNA9cXWV
[22:41:02 CET] <J_Darnley> Input 0..1 and it would return a nice thermal camera like image
[22:41:12 CET] <J_Darnley> well colours for one
[22:41:42 CET] <J_Darnley> ha ha.  I do still have a rendered image from it
[22:42:59 CET] <J_Darnley> http://users.telenet.be/darnley/wp-keyboard-heat.png
[22:43:58 CET] <wm4> this image proves that space bar heating is overrated, and one should use the shift keys instead?
[22:46:54 CET] <TD-Linux> space bar heating?
[22:49:06 CET] <wm4> just an unfunny joke https://xkcd.com/1172/
[23:01:02 CET] <durandal_170> intensity option is same as heat
[23:28:53 CET] <cone-234> ffmpeg 03yukari yakumo 07master:f87e29a61587: configure: support --cpu=host and Clang
[00:00:00 CET] --- Thu Dec 31 2015


More information about the Ffmpeg-devel-irc mailing list