[Ffmpeg-devel-irc] ffmpeg-devel.log.20180323

[burek]

[00:51:21 CET] <cone-954> ffmpeg 03Mark Thompson 07master:2a1542d105dc: lavfi/opencl: Derive global work size from plane image sizes
[00:51:22 CET] <cone-954> ffmpeg 03Mark Thompson 07master:0568af521ef3: lavfi/opencl: Avoid deprecation warnings when built with post-1.2 headers
[00:51:23 CET] <cone-954> ffmpeg 03Mark Thompson 07master:9313422dfc8d: hwcontext_opencl: Avoid deprecation warnings when built with post-1.2 headers
[00:51:24 CET] <cone-954> ffmpeg 03drfer3 07master:6135c958cd3c: lavfi: Add OpenCL avgblur filter
[00:51:25 CET] <cone-954> ffmpeg 03Mark Thompson 07master:f86e8c91d020: hwcontext_drm: Clarify value for unknown format modifiers
[00:51:26 CET] <cone-954> ffmpeg 03Mark Thompson 07master:c6bbb8cca7fa: hwcontext_vaapi: Add support for legacy DRM mapping
[00:51:27 CET] <cone-954> ffmpeg 03Mark Thompson 07master:ca9f13bbce48: hwcontext_vaapi: Pass correct read/write flags when exporting surfaces
[00:51:28 CET] <cone-954> ffmpeg 03Mark Thompson 07master:ac993e73deca: hwcontext_opencl: Add support for mapping DRM objects to Beignet
[00:51:29 CET] <cone-954> ffmpeg 03Mark Thompson 07master:abf35afb6f70: vf_scale_vaapi: Apply cropping rectangle to input
[00:51:38 CET] <atomnuker> was just about to ask, thanks
[00:52:19 CET] <jkqxz> What shall I do about the vf_crop bit?
[01:09:14 CET] <Chloe> wm4: I guess I should write a set which merges lavd into lavf otherwise no one is even going to comment on that as an actually viable possibility
[01:10:24 CET] <Chloe> It doesn't look like 'doing it fast' is even a thing anymore, or I would have gotten more comments on the other patches I expect--so I'm just going to prioritise 'doing it right'
[01:16:52 CET] <wm4> dunno
[01:25:12 CET] <Chloe> wm4: nice. No one else with give input either, so I'll just make the set. People only give input after something happens anyway (if they do at all)
[01:25:19 CET] <Chloe> s/with/will/
[01:28:08 CET] <atomnuker> jkqxz: avgblur_opencl doesn't work on nvidia
[01:28:11 CET] <atomnuker> "Failed to finish command queue: -36."
[01:29:09 CET] <atomnuker> also lavfi always warns if drm mapping is unavailable, maybe you should make it a debug message only
[01:30:08 CET] <jkqxz> INVALID_COMMAND_QUEUE?  That doesn't make any sense.
[01:30:43 CET] <jkqxz> Which message?
[01:31:22 CET] <jkqxz> (It works on Intel, AMD and Mali.)
[01:33:10 CET] <wm4> Chloe: true
[01:34:47 CET] <wm4> Chloe: how would it merge it?
[01:35:58 CET] <atomnuker> jkqxz: what do you mean?
[01:37:03 CET] <atomnuker> oh, I mean "[AVHWDeviceContext @ 0x56275e9a77c0] Beignet DRM to OpenCL mapping not usable."
[01:38:06 CET] <jkqxz> Well, you built it in and it's not there at runtime.
[01:39:58 CET] <jkqxz> Should the (existing) convolution filter be ignoring the div/bias options if the convolution is the identity?
[01:42:34 CET] <atomnuker> jkqxz: it already tells you if it failed to map the frame
[01:42:58 CET] <atomnuker> it shouldn't warn you always, especially when you don't specify you want to map something
[01:44:52 CET] <Chloe> wm4: Not sure: mv libavdevice libavformat/device > fix broken stuff, is how I'm going to try it
[01:45:32 CET] <wm4> that's a big change, I don't think it'd have much chance of success/acceptance
[01:45:43 CET] <wm4> unless everyone somehow agrees on it in advance
[01:46:56 CET] <Chloe> wm4: nicolas might agree, then you only need one iteration function for formats and 'devices'
[01:48:49 CET] <Chloe> and yes, it's a big change, but that's why I started the discussion thread
[01:48:57 CET] <Chloe> except there's zero discussion and no one cares
[02:02:44 CET] <Chloe> With having a static list and no initialisation, you cant really have inter-library dependencies like there was, as you'd always either have to link lavf + lavd or just lavf (lavf was either built with references to lavd or it wasnt)
[02:03:41 CET] <Chloe> as far as I can see, it's not actually possible to use the same iteration function across libraries without some intialisation function which is called outside of both libraries (something which defeats the purpose of the entire thing0
[02:05:45 CET] <Chloe> atomnuker: any ideas?
[02:11:44 CET] <wm4> we could still have a lavd register function, which would call an avpriv_ function that sets a second array of formats
[02:22:51 CET] <rcombs> that's what I do in lavc
[06:56:06 CET] <gagandeep> guys i am getting make error f
[06:56:25 CET] <gagandeep> libavfilter/allfilters
[06:57:43 CET] <gagandeep> REGISTER_FILTER(AVGBLUR_OPENCL, avgblur_opencl, vf);
[07:10:19 CET] <gagandeep> vlibavfilter/allfilters.c:30:13: error: CONFIG_AVGBLUR_OPENCL_FILTER undeclared
[08:23:45 CET] <nevcairiel> re-run configure
[08:29:45 CET] <gagandeep> nevcairiel: yeah i think it is working now
[10:17:41 CET] <cone-493> ffmpeg 03Paul B Mahol 07master:4e1307c0f76d: avfilter/af_channelsplit: add channels option
[11:31:08 CET] <Chloe> wm4: I guess we we still have avdevice_register_all() then everyone is happy
[11:31:27 CET] <Chloe> lavd is still a separate library (for god knows what reason but whatever)
[11:31:35 CET] <Chloe> only the lavf iteration function is used
[11:31:55 CET] <Chloe> and we get (mostly) static library initialisation
[11:32:03 CET] <Chloe> s/we we/if we/
[11:36:29 CET] <wm4> yeah
[11:37:04 CET] <JEEB> I don't remember the exact reasons for lavd separation but IIRC it was something along the lines of not having the devices in lavf
[11:37:06 CET] <durandal_1707> atomnuker: https://github.com/xiph/rnnoise ---> is this what you meant to use?
[11:37:10 CET] <JEEB> (as many didn't want them)
[11:43:54 CET] <durandal_1707> atomnuker: one can use lv2 plugin: noise-repellent
[11:53:45 CET] <Chloe> JEEB: it happened in version 0.5
[11:58:03 CET] <JEEB> yes it was quite a while ago
[12:32:39 CET] <jdarnley> I think I've had a conceptual "eureka moment".
[12:33:22 CET] <jdarnley> I can have the h-filter filter all lines available without condition.
[12:34:06 CET] <jdarnley> That will leave whichever buffer I end-up using in its "half-way state".
[12:35:05 CET] <jdarnley> Then I think I can just v-filter as many lines as I can, storing the progress somewhere.
[12:39:30 CET] <jdarnley> ... so "odd" lines require the unmodified, half-way values
[12:39:57 CET] <jdarnley> ... even lines require the final odd lines.
[12:41:15 CET] <jdarnley> ... I can't modify any even line that is required for an odd line.
[12:42:57 CET] <jdarnley> ... lines 0-15: I can filter line 13 but I cannot filter line 14
[12:45:01 CET] <jdarnley> 14 must be unmodified for line 15 which also requires line 16 (which isn't yet available)
[12:47:01 CET] <Chloe> h-filter filter?
[12:47:44 CET] <jdarnley> h-filter: noun, filter: verb
[12:48:01 CET] <jdarnley> or rather, h-filter: adverb
[12:48:22 CET] <Chloe> I mean, what are you doing? writing a new filter?
[12:48:39 CET] <jdarnley> No this is still the wavelets in the vc2 encoder.
[12:53:09 CET] <jdarnley> On the next call to the function I can run the h-filter on all the new lines.  So I need to indicate where that starts.
[12:53:58 CET] <jdarnley> After that I can continue on the odd lines so I need to indicate where that starts.
[12:54:12 CET] <Chloe> Ah I dont know anything about wavelets, they sound interesting though. I've only ever looked at h.264/h.265/jpeg stuff
[12:55:05 CET] <jdarnley> I can't comment on their technical merits but I will say that AFAIK they no format that uses them has widespread adoption.
[12:55:07 CET] <thardin> wavelets tend to optimize for psnr, blur things
[12:55:24 CET] <thardin> but that might depend on what kernel you use
[12:55:26 CET] <jdarnley> s/they//
[12:56:06 CET] <thardin> some wavelet formats deal with bitrate slicing quite well. but not all
[12:56:50 CET] <jdarnley> I don't like them because they are hard to write, in my opinion.
[12:57:53 CET] <jdarnley> But then I've only worked on them in vc2, a format I don't like.
[12:58:26 CET] <thardin> I used to write some wavelet stuff back in the days (~2005). not terribly difficult. didn't grasp how fft/fast dct worked back then, wavelets are more straightforward
[13:17:03 CET] <CounterPillow> RELEASE OR RIOT
[13:17:51 CET] <jdarnley> Screw releases.  Use git HEAD.
[13:18:00 CET] <Chloe> CounterPillow: you can riot all you want, but we have an unstable API atm
[13:18:14 CET] <Chloe> CounterPillow: you can shout at me for that, but it's getting sorted
[13:18:28 CET] <CounterPillow> I'm not going to shout at a child, that would be very rude.
[13:20:25 CET] <Chloe> CounterPillow: I'm an adult now!
[13:21:34 CET] <CounterPillow> send nudes
[13:22:12 CET] <jdarnley> thelonelyisland.mp4
[13:22:40 CET] <Chloe> /ban CounterPillow 
[13:23:15 CET] <CounterPillow> "nudes" as in patches, and "send" as in "send to the mailing list"
[13:23:39 CET] <Chloe> CounterPillow: if you check the mailing list, there is an on-going discussion
[13:23:51 CET] <Chloe> and there is a set currently on there which 'completes' it
[13:24:05 CET] <Chloe> it's just if it's the best solution
[13:24:34 CET] <CounterPillow> Gonna need to get in the middle of this with some good internet comment etiquette
[13:26:11 CET] <Chloe> CounterPillow: feel free to give your opinion on the ML
[13:26:28 CET] <CounterPillow> nah I don't want to get banned from yet another project
[13:26:45 CET] <CounterPillow> Last time I gave my opinion, I was temporarily suspended from github for bullying
[13:27:35 CET] <durandal_1707> wm4: segafilm muxer is nice addition
[13:41:16 CET] <Chloe> wm4: I also don't think it's too bad, the only reason I was wondering about it was that I couldn't see a use-case, but I think the one he gave is pretty valid
[13:47:03 CET] <atomnuker> jdarnley: is that for simd or do you still plan to rewrite the encoder's transforms?
[15:46:42 CET] <jkqxz> durandal_1707:  Should the convolution filter be ignoring the rdiv/bias values for a plane where the convolution is the identity?
[15:57:47 CET] <durandal_1707> jkqxz: that is very useless scenario, it shouldn't
[15:58:31 CET] <durandal_1707> also very inefficient way to decrease/increase overall brightness
[16:04:00 CET] <jkqxz> Yeah, it's pretty degenerate.
[16:08:52 CET] <jkqxz> Care to offer an opinion in reply to <http://ffmpeg.org/pipermail/ffmpeg-devel/2018-March/227192.html>?
[16:10:11 CET] <jkqxz> It's probably easier to just do it anyway there, but it does mean that it produces different output to the existing filter.
[16:24:39 CET] <cone-910> ffmpeg 03Paul B Mahol 07master:fe0fdc51b56e: avfilter/vf_convolution: disable fast path if rdiv/bias are different from defaults
[16:31:03 CET] <jkqxz> durandal_1707:  That works, thank you!
[17:02:13 CET] <kierank> michaelni: what are the limits on skip_bits_long?
[17:05:46 CET] <atomnuker> INT32_MAX I think
[19:01:30 CET] <atomnuker> jkqxz: mapping drm to opencl gives me black screen (works when I map to vaapi)
[19:17:40 CET] <cone-910> ffmpeg 03Paul B Mahol 07master:a8c2d375ca68: avfilter/avf_showwaves: add draw mode which controls how single sample is drawn
[19:24:56 CET] <klaxa> atomnuker: my uni seems to be even messier than me and they have misplaced a form i submitted half a year ago which is delaying my immatriculation and they are unwilling to sign anything else, i'll talk to them again on monday (office closed fr-so) and if i'm lucky i will get some sort of proof of enrollment, but given their shittyness in almost every semester i had to deal with them i'm not too confident
[19:27:18 CET] <klaxa> what i find pretty funny is that i sent them multiple copies of said form and they still insist they never received it (´4@)
[19:27:39 CET] <atomnuker> yeah, fuck bureaucracy well
[19:27:59 CET] <atomnuker> my favorite one is when you have to move
[19:28:29 CET] <atomnuker> try to rent a place -> need a bank account -> try to open a bank account -> need a proof of address -> try to rent a place
[19:28:32 CET] <cone-910> ffmpeg 03Paul B Mahol 07master:74c6a6d3735f: doc/filters: itemize channelmap examples
[20:20:51 CET] <atomnuker> jamrial: "2018-03-21 - xxxxxxx - lavc 58.15.100 - avcodec.h"
[20:21:01 CET] <atomnuker> but lavu minor is 11 right now
[20:21:22 CET] <jamrial> and?
[20:22:02 CET] <atomnuker> well, the doc/apichanges for av_packet_make_writeable point to a version of lavu that doesn't exist yet
[20:22:10 CET] <jamrial> re read that line
[20:22:29 CET] <atomnuker> ah, nevermind
[20:23:25 CET] <atomnuker> was just rebasing and had to update apichanges and I thought I was missing a conflict in lavu
[20:46:05 CET] <jamrial> durandal_1707: commit 10542491 should have also changed the prototype
[20:47:30 CET] <durandal_1707> if it ain't broken do not fix it
[21:48:46 CET] <Chloe> hopefully that's the last set I have to send about this API
[21:58:04 CET] <jkqxz> atomnuker:  Mapping DRM from what?  Vulkan?
[21:58:11 CET] <atomnuker> no, opencl
[21:58:23 CET] <jkqxz> OpenCL to OpenCL?
[21:58:32 CET] <atomnuker> no, drm to opencl
[21:58:43 CET] <jkqxz> DRM from from?
[21:58:45 CET] <atomnuker> I don't get it, when I tested the patchset it worked
[21:58:47 CET] <jkqxz> *from what
[21:58:52 CET] <atomnuker> drm from kmsgrab
[21:59:51 CET] <atomnuker> so "sudo ffmpeg -init_hw_device "opencl=ocl:1" -f kmsgrab -format bgra -framerate 60 -crtc_id 0 -plane_id 0 -device /dev/dri/card1 -i /dev/null -filter_hw_device ocl -vf hwmap,unsharp_opencl,hwdownload,format=bgra"
[22:07:31 CET] <jkqxz> Works for me.
[22:08:14 CET] <jkqxz> (On Coffee Lake, just tried again.  I tested on amdgpu and rockchip as well before.)
[22:09:49 CET] <jkqxz> Sure that card1 is the right device, and you aren't capturing an inactive screen?
[22:53:11 CET] <jkqxz> atomnuker:  Does <https://0x0.st/sB2V.diff> fix your avgblur_opencl + Nvidia problem?
[22:56:07 CET] <cone-878> ffmpeg 03Paul B Mahol 07master:3e3d56776166: avfilter/avf_showvolume: implement basic rms metering mode
[23:02:41 CET] <atomnuker> jkqxz: nope
[23:19:42 CET] <atomnuker> jkqxz: btw what's up with VA_FOURCC_ABGR in hwcontext_vaapi
[23:22:31 CET] <durandal_1707> jkqxz: what performance boost one get with opencl avgblur implementation?
[23:27:11 CET] <atomnuker> durandal_1707: in my case hwaccel + opencl is 3x slower than cpu decode/hwaccel + avgblur
[23:27:13 CET] <jkqxz> atomnuker:  What about it?
[23:27:42 CET] <jkqxz> durandal_1707:  It uses much less CPU.  Depending on what you're doing it might use more or less total time.
[23:28:13 CET] <atomnuker> jkqxz: if its there you logically map it to what's in vaapi, but if it isn't there you map it to something that doesn't seem to match at all
[23:28:34 CET] <atomnuker> rgba vs bgra
[23:30:14 CET] <jkqxz> I don't follow.  ABGR isn't mapped at all if the headers don't have it.
[23:31:09 CET] <atomnuker> didn't notice that isn't an #else, nvm
[23:35:09 CET] <durandal_1707> i get useless output with rnnoise
[23:45:51 CET] <atomnuker> what do you mean?
[23:46:24 CET] <durandal_1707> atomnuker: https://github.com/xiph/rnnoise
[23:46:38 CET] <atomnuker> yes
[23:46:40 CET] <atomnuker> useless?
[23:47:55 CET] <durandal_1707> it uses floats, and without any parameters is initialized
[23:48:41 CET] <atomnuker> there's a newer version of that in the opus repo
[00:00:00 CET] --- Sat Mar 24 2018