[Ffmpeg-devel-irc] ffmpeg-devel.log.20180530

Thu May 31 03:05:03 EEST 2018

[00:25:44 CEST] <llogan> look like i broke the CSS with 2080b81a...remind me what extra step we need to do when the css is edited?
[00:30:27 CEST] <llogan> looks like lessc was required to generate the css but it wasn't installed.
[00:32:13 CEST] <llogan> i prefer the days when a site could be made with just a text editor with no frameworks, javascript, and all this other horseshit
[00:32:25 CEST] Action: llogan is officially old now
[00:35:07 CEST] <llogan> can't figure it out. oh well, we now have a new look.
[00:45:33 CEST] <llogan> for future reference i had to log into the server, navigate to the directory with the ffmpeg-web files, then run "make clean; make".
[01:10:50 CEST] <cone-578> ffmpeg 03Carl Eugen Hoyos 07master:53d6d4f266b5: ffplay: Print codec_name if decoder for codec_id could not be found.
[01:30:24 CEST] <jdarnley> I'll try to remember that for you.
[01:33:33 CEST] <llogan> thanks, but i'll just never touch it again.
[02:49:48 CEST] <cone-578> ffmpeg 03Michael Niedermayer 07master:a9dacdeea616: avcodec/qtrle: Do not output duplicated frames on insufficient input
[02:49:49 CEST] <cone-578> ffmpeg 03erankor 07master:500e63871162: qt-faststart - stricter input validations
[02:49:50 CEST] <cone-578> ffmpeg 03Andrey Semashev 07master:d7eb8d847571: lavfi/tests: Fix 16-bit vf_blend test to avoid memory not aligned to 2 bytes
[07:27:27 CEST] <cone-025> ffmpeg 03Alex Converse 07master:4d3dd167dfdf: flvenc: Factorize timestamp writing
[07:27:27 CEST] <cone-025> ffmpeg 03Alex Converse 07master:ea0010bf9cd2: flvenc: Fix sequence header update timestamps
[08:02:06 CEST] <gagandeep> kierank: i think now i have encountered that tag = 71 problem for mountain file and timelapse file you gave me
[08:02:43 CEST] <gagandeep> the output has too much noise but otherwise i can make out that yes the general picture is supposed to be this in mountain file
[08:26:29 CEST] <gagandeep> false alarm
[09:15:03 CEST] <gagandeep_> kierank: mountain file is running now
[09:15:27 CEST] <gagandeep_> man, that had many tags all over the place
[09:16:00 CEST] <gagandeep_> will try to run timelapse file as well now, cause that was also on ip frame
[09:34:55 CEST] <gagandeep_> kierank: now i just need to put the second frame in as well so whats the process in a 'non-succinct' manner :)
[09:36:22 CEST] <kierank> Use av_frame_ref to store a copy of frame in CFHDContext
[09:38:45 CEST] <gagandeep_> cfhd context will remain throughout
[09:38:53 CEST] <gagandeep_> what about free buffers function
[09:39:19 CEST] <gagandeep_> ah, ok i will need to edit header for that data
[09:39:51 CEST] <gagandeep_> and try to reference again when asked for second frame
[09:39:56 CEST] <gagandeep_> ok, let me see
[09:40:35 CEST] <gagandeep_> kierank: also ffmpeg output is good only when raw video and pix fmt yuv is selected, mpeg is producing artefacts
[09:40:38 CEST] <gagandeep_> is that alright
[09:41:04 CEST] <kierank> Probably ok we can investigate later
[09:41:58 CEST] <kierank> When you are done with buffered frame use av_frame_unref but I dunno if you need pixel data or coefficients
[09:42:26 CEST] <gagandeep_> i need only pixel data in the end, i don't think i will need coefficients
[09:43:42 CEST] <gagandeep_> afk, going to eat now
[12:32:23 CEST] <cone-071> ffmpeg 03Paul B Mahol 07master:d0e740b8fb30: avfilter/af_afir: fix picking of IR channel
[12:36:41 CEST] <klaxa> so i'm trying to test my allocation checks by using setrlimit() but i get segfaults in api-calls :/
[12:40:38 CEST] <nevcairiel> send patches :)
[12:41:54 CEST] <JEEB> klaxa: yea I wouldn't be surprised if a lot of stuff didn't handle OOM properly :D
[12:46:20 CEST] <klaxa> ok good i thought i was doing something wrong... i think handling OOM properly is desireable, but sadly not the task at hand, maybe if it bugs me enough i'll dig deeper
[12:52:44 CEST] <JEEB> klaxa: basically just try to make sure your code looks OK
[12:52:55 CEST] <JEEB> if nullptr then GTFO etc
[14:09:47 CEST] <thardin> feeling an urge to experiment with frama-c in libavutil
[14:19:28 CEST] <gagandeep> kierank: what about memory allocation for using av_frame_ref
[14:20:11 CEST] <gagandeep> in cfhd you had directly assigned avframe pointer to void data pointer brought as argument in cfhd_decode function
[14:20:12 CEST] <kierank> It's a reference count 
[14:20:20 CEST] <kierank> So it just stops the frame from being freed
[14:20:33 CEST] <kierank> And when you are done with it and unref it gets freed
[16:10:31 CEST] <J_Darnley> A rather generic question for you all.  Can one de-interleave every other value in an array without temporary storage for the whole array?
[16:11:23 CEST] <J_Darnley> Is there some function already in ffmpeg I can use for it?
[17:05:25 CEST] <thardin> getting somewhere with formally proving qt-faststart.c correct
[17:05:30 CEST] <thardin> or, that it should not crash at least
[17:31:44 CEST] <cone-456> ffmpeg 03Paul B Mahol 07master:1aa519245193: avfilter/af_aiir: draw IR frequency response
[17:59:08 CEST] <cone-456> ffmpeg 03Michael Niedermayer 07master:b3332a182f8b: avcodec/idctdsp: Transmit studio_profile to init instead of using AVCodecContext profile
[17:59:09 CEST] <cone-456> ffmpeg 03Michael Niedermayer 07master:8c50d0cccfe4: avcodec/idctdsp: Clear idct/idct_add for studio profile
[17:59:10 CEST] <cone-456> ffmpeg 03Michael Niedermayer 07master:44a2415a6d94: avcodec/h263dec: Reinitialize idct context if it has not been setup for the active profile
[19:53:33 CEST] <cone-456> ffmpeg 03James Almer 07master:cd62f9d557f6: tools/crypto_bench: add support for mbedcrypto
[19:53:34 CEST] <cone-456> ffmpeg 03James Almer 07master:5420c8f76e70: tools/crypto_bench: add missing RC4 implementation from gcrypt
[19:53:35 CEST] <cone-456> ffmpeg 03James Almer 07master:e3734aa6a33b: tools/crypto_bench: add missing RC4 implementation from tomcrypt
[21:27:13 CEST] <klaxa> while i'm fixing up the last bits of the last review round i've been thinking about how to do hls, i think it would be a bad idea to rewrite it from scratch, i think using the libavformat implementation is a no-brainer, however as far as i have investigated it requires to write files to disk, i'd prefer to keep everything in memory though, how should i go about this?
[21:30:53 CEST] <klaxa> i just thought of something like a virtual filesystem in ram?
[21:31:06 CEST] <klaxa> or should i just write to disk anyway? :S
[22:14:43 CEST] <klaxa> then again if people care about it as much as me they can just put it all in /tmp i guess
[22:28:40 CEST] <atomnuker>  /tmp isn't guaranteed to be in ram
[22:29:02 CEST] <atomnuker> /dev/shm is, however
[22:29:31 CEST] <atomnuker> I prefer memfd though
[22:29:56 CEST] <atomnuker> though that's linux only so you'll still need fallbacks for osx or windows
[22:41:38 CEST] <klaxa> that looks exactly like what i want (i think?) might be tricky to get the hls encoder to use them
[22:42:57 CEST] <klaxa> i think for the time being it will be easier to have it working while writing on disk, people that care (like me) can make sure the files reside in ram instead of disk
[22:43:23 CEST] <klaxa> it's certainly an optimization worth implementing, but i think it should be working before i try to optimize it
[22:48:14 CEST] <jkqxz> atomnuker:  I have the same segfault as hanna did above then uploading to rgb0 on radv.
[22:49:16 CEST] <jkqxz> Mapping from VAAPI in RGB does work, but since allocation is the problem the only thing you can do after that is download immediately.  (E.g. "-hwaccel vaapi -hwaccel_output_format vaapi -hwaccel_device /dev/dri/renderD129 -i in.mp4 -an -vf 'scale_vaapi=format=rgb0,hwmap=derive_device=vulkan,format=vulkan,hwdownload,format=rgb0' ..." works.)
[22:50:02 CEST] <jkqxz> Adding a scale_vulkan into that segfaults again, in the allocation for the scale output.
[22:51:29 CEST] <atomnuker> weird, I'm not sure what's going on
[22:51:53 CEST] <atomnuker> I've been wondering if a memory leak was happening, but I don't think so
[22:54:33 CEST] <BBB> I find it so exciting that some people involved in the actual C language appear to agree that this whole ub debate is utterly useless
[22:54:42 CEST] <BBB> (http://www.yodaiken.com/2018/05/20/depressing-and-faintly-terrifying-days-for-the-c-standard/)
[22:56:21 CEST] <JEEB> that PDF seems to be going the rounds
[22:57:19 CEST] <atomnuker> jkqxz: hmm, I think I know what's going on
[22:58:43 CEST] <atomnuker> can you print "use_ded_mem" after line 1012?
[22:59:24 CEST] <atomnuker> maybe the driver is requesting dedicated alloc and its a path in image creation which I couldn't check since neither anv nor nvidia request it
[23:00:11 CEST] <jkqxz> 1
[23:01:59 CEST] <BBB> JEEB: its pretty well-written
[23:02:14 CEST] <BBB> JEEB: I think Ive objected to ub before in this irc channel, as well as on the mailinglist
[23:03:03 CEST] <jkqxz> "{sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS, pNext = 0x0, prefersDedicatedAllocation = 1, requiresDedicatedAllocation = 1}"
[23:04:00 CEST] <JEEB> BBB: ah I wasn't sure which opinion you were of with your initial comment and thus I did a somewhat muted comment :)
[23:04:18 CEST] <BBB> ah
[23:04:20 CEST] <JEEB> I do agree that UB is, put lightly, "somewhat of a PITA"
[23:04:37 CEST] <BBB> I think too many things are marked ub for no good reason
[23:04:45 CEST] <JEEB> yes
[23:05:32 CEST] <klaxa> hmm most ub i have come across were basically "read after invalid write"
[23:06:15 CEST] <BBB> upshifting a negative signed integer is ub
[23:06:24 CEST] <BBB> signed integer overflow (a+b) is ub
[23:06:27 CEST] <klaxa> oh
[23:06:27 CEST] <BBB> etc.
[23:06:30 CEST] <klaxa> wow
[23:06:35 CEST] <BBB> these things are ridiculous, utterly trash
[23:06:36 CEST] <klaxa> could have just... defined it
[23:06:45 CEST] <BBB> we all agree that derefing a null pointer is bad
[23:07:09 CEST] <BBB> or writing (or reading) uninitialized memory or non-existing memory
[23:07:15 CEST] <klaxa> well types in c are kind of all over the place anyway
[23:07:58 CEST] <nevcairiel> the signed constraints are so stupid because C theoretically also supports running on archs that are not 2s complement where this might be weird
[23:08:05 CEST] <iive> maybe it is time people who write compilers to have a say in what the language should look like
[23:09:08 CEST] <BBB> nevcairiel: the problem is that it means people (..) will do silly stuff like int b = .., a = (int)((usigned) b + 10);
[23:09:37 CEST] <BBB> nevcairiel: that sort of stuff is obviously not going to give the same result as adding 10 and (signed) overflowing in the non-2s-complement machine
[23:09:53 CEST] <BBB> so you get dumb stuff like c can run on 1s complement, but ffmpeg requires 2s complement"
[23:10:11 CEST] <BBB> agian, utter trash, and makes our code ugly also
[23:11:05 CEST] <BBB> (and yes, ffmpeg has code like that all over the place, particularly in idcts)
[23:11:14 CEST] <jkqxz> Does AVS2 have any public test samples anywhere?
[23:11:44 CEST] <jkqxz> Or, like, any samples at all anywhere which might be findable by someone who can't read chinese.
[23:12:22 CEST] <BBB> cant you ask the guy on the ml that posted the wrapper patches?
[23:12:45 CEST] <jkqxz> I will if noone happens to know the answer here.
[23:13:27 CEST] <JEEB> jkqxz: I gave it a search and nothing came up with a set of keywords you probably already tried "avs2 reference samples"
[23:14:28 CEST] <atomnuker> jkqxz: could you replace "use_ded_mem = ded_req.prefersDedi...." with "use_ded_mem = 0;" and give it another try?
[23:14:35 CEST] <cone-456> ffmpeg 03Marton Balint 07master:f932e49aab09: avformat/mxfdec: fix klv_decode_ber_length return value usage
[23:14:36 CEST] <cone-456> ffmpeg 03Marton Balint 07master:e7498792b656: avformat/mxfdec: use int math for sample count calculation
[23:14:37 CEST] <cone-456> ffmpeg 03Marton Balint 07master:1616b1be5a12: avformat/mxfdec: add some missing picture_essence_container_uls
[23:14:49 CEST] <atomnuker> wait, no, it requires dedicated allocation
[23:15:51 CEST] <jkqxz> Still segfaults in the same place <https://0x0.st/s2Ec.txt>.
[23:15:52 CEST] <JEEB> oh wow, the code tree for xavs2 looked like x265 first but then suddenly a configure file under build/linux
[23:15:55 CEST] <JEEB> lol
[23:17:34 CEST] <jkqxz> JEEB:  I did try quite a few variations on that and found nothing useful (excepting some companies which want to sell test streams).  Being able to read chinese might have helped, though.
[23:17:37 CEST] <JEEB> the chinaffmpeg.org guy seems to have a PR open there
[23:17:52 CEST] <JEEB> jkqxz: yea, it's mostly meant for internal usage after all
[23:19:34 CEST] <jkqxz> I built both, they were full of crazy warnings.
[23:20:01 CEST] <jkqxz> A file encoded by libxav2 can be decoded by libdavs2, so that's something at least!
[23:20:23 CEST] <JEEB> I asked in a channel with at least some chinese in it
[23:20:45 CEST] <JEEB> maybe something will come out of it, but more likely the people making the wrappers might have info on the reference samples
[23:22:37 CEST] <jkqxz> Neither <https://www.google.co.uk/search?q=libdavs2> nor <https://www.google.co.uk/search?q=libxavs2> fills me with confidence that these libraries see much use.
[23:24:19 CEST] <JEEB> the IEEE standard doesn't seem to be finalized yet either
[23:24:26 CEST] <JEEB> http://standards.ieee.org/develop/wg/AVS_1857.html
[23:24:33 CEST] <JEEB> .4 is AVS2
[23:25:00 CEST] <JEEB> http://www.ieee1857.org/1857.4.asp their site
[23:26:34 CEST] <jkqxz> Lots of Peking University people in that group, who are the authors of these libraries.
[23:27:54 CEST] <nevcairiel> how badly stolen is that code?
[23:29:56 CEST] <JEEB> got as far as http://www.avs.org.cn/avs2/download.asp , where they have a "reference software", "realtime encoder uAVS2" and a "cloud platform openavs2"
[23:30:33 CEST] <nevcairiel> i wonder why they even bother to make that codec if its just ripped off hevc anyway
[23:30:45 CEST] <nevcairiel> they obviously dont care about patents or whatever
[23:30:49 CEST] <JEEB> "homebrew technology!"
[23:30:52 CEST] <nevcairiel> maybe just to look smart?
[23:31:01 CEST] <JEEB> it's not foreign patents, but local patents
[23:31:55 CEST] <JEEB> www.avs.org.cn actually has links towards AVS2 patent licensing
[23:32:03 CEST] <JEEB> so they do care about patents - their own that is
[23:32:30 CEST] <JEEB> "The AVS member can login to the AVS workgroup via FTP to download 92 standard test streams via the member account and password for product testing. The remaining streams can be contacted by the AVS secretariat to sign the relevant agreement for purchase."
[23:32:43 CEST] <JEEB> for AVS Part II which seems to be AVS2
[23:33:52 CEST] <JEEB> http://www.avs.org.cn/fruits/softList.asp
[23:34:37 CEST] <atomnuker> jkqxz: oh, that's with the patchset from the ML, right?
[23:34:48 CEST] <atomnuker> I fixed that bug in my github branch
[23:34:53 CEST] <atomnuker> https://github.com/atomnuker/FFmpeg/tree/exp_vulkan
[23:40:53 CEST] <JEEB> jkqxz: I think I'll just ask the guy who posted the patch
[23:44:13 CEST] <JEEB> there
[23:50:02 CEST] <jkqxz> libdavs2 doesn't build with gcc 7 because it adds -std=c++11, and gcc enforces that C++11 doesn't support hex float constants.1
[23:55:35 CEST] <jkqxz> (The configure script is copied from libx264, of course.)
[00:00:00 CEST] --- Thu May 31 2018