Monday, 2023-03-27

*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC05:24
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc05:25
*** Ritish <Ritish!~Ritish@> has joined #libre-soc06:18
*** Ritish <Ritish!~Ritish@> has quit IRC06:19
*** Ritish <Ritish!~Ritish@> has joined #libre-soc06:20
*** Ritish <Ritish!~Ritish@> has joined #libre-soc06:43
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:5533:3ca0:acee:7057> has joined #libre-soc06:45
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC06:47
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc06:47
*** Ritish <Ritish!~Ritish@> has quit IRC06:48
*** ritty <ritty!~Ritish@> has joined #libre-soc06:50
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:5533:3ca0:acee:7057> has quit IRC06:52
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:afc2:4db:7336:3b7a> has joined #libre-soc06:53
*** ritty <ritty!~Ritish@> has quit IRC06:54
*** psydroid <psydroid!~psydroid@user/psydroid> has quit IRC07:00
*** yambo <yambo!> has quit IRC07:00
*** sauce <sauce!> has quit IRC07:00
*** sauce <sauce!> has joined #libre-soc07:00
*** yambo <yambo!> has joined #libre-soc07:01
*** psydroid <psydroid!~psydroid@user/psydroid> has joined #libre-soc07:05
*** ritty <ritty!~Ritish@> has joined #libre-soc08:28
*** ritty <ritty!~Ritish@> has joined #libre-soc08:31
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:afc2:4db:7336:3b7a> has quit IRC08:31
*** ritty <ritty!~Ritish@> has quit IRC08:32
*** ritty <ritty!~Ritish@> has joined #libre-soc08:32
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:ab6c:530f:9209:3dcc> has joined #libre-soc08:51
*** ritty <ritty!~Ritish@> has quit IRC08:54
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:ab6c:530f:9209:3dcc> has quit IRC10:08
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:39f7:3fb0:87d8:b07d> has joined #libre-soc10:08
*** ritish_ <ritish_!~Ritish@2401:4900:1f2d:bbee:39f7:3fb0:87d8:b07d> has quit IRC10:11
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC11:26
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc11:26
*** alethkit is now known as nobody11:58
*** nobody is now known as alethkit11:59
lkclyes there exists (surpriiise) a debian package for that12:18
lkclapt-mirror or something12:18
lkclwhere you can then set the entries in /etc/apt/sources.list to point at the apt-mirror instead12:19
lkcland yes, you reaallly don't want to be slowing down things by not using apt-mirror12:19
lkclthere exist several such packages btw, i don't know their differences, i used them over 10 years ago. i just remember they exist12:20
markosapt-mirror is great and very easy to setup indeed12:27
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC12:28
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc12:28
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC13:41
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc13:42
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC16:16
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc16:16
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC18:51
*** gnucode2 <gnucode2!~gnucode2@user/jab> has joined #libre-soc19:17
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc19:21
sadoon[m]<lkcl> "there exist several such..." <- I've used apt mirror before but got throttled / semi banned because I did parallel downloads lol21:05
sadoon[m]Using debmirror currently which seems to do the same thing21:05
programmerjakelkcl, can you move the budget from #1041 to #972? thx22:09
markoslkcl, I thoght libvpx was bad22:23
markosbut libaom is worse22:23
markosin the sense that each function can have multiple versions due to particular coefficients that require totally different SIMD algorithms and unrolling22:24
markoswell not each function22:24
markosbut some quite used ones22:24
markossome convolution functions in particualr22:25
programmerjakewell, it's designed for experimentation, not as much for high performance22:25
markosoriginal function 15 lines, SSE/NEON ~45022:26
markostrue, but both Intel and Arm are keen on getting it as optimized as possible22:26
markosit's like a contest22:26
markosI don't have final benchmarks for libvpx but with latest patches we managed to get higher speed in 4k highbd encoding on graviton3 than the other intel graviton offerings22:27
markosI guess Arm wants to repeat that with libaom22:28
programmerjakeimho if you have the choice, work on a av1 decoder/encoder actually designed from the beginning for high performance22:28
markosanyway, the point is that it would take a fraction of the time to finish the port to SVP6422:28
programmerjakewait, intel has graviton offerings?22:28
markossorry s/intel graviton offerings/aws x86 offerings/22:29
programmerjakei thought that was purely amazon arm22:29
markosbut I'm actually getting paid by Arm to do NEON optimizations on libvpx/libaom :)22:29
programmerjakeoh, fun...22:30
programmerjakewell, good luck :)22:30
markoswell, it's not that bad, it's quite rewarding to see your code upstreamed in a project like that22:30
markosand tbh I'm glad it's so hard because they hire me to do it :)22:31
programmerjakeyeah, i was pretty happy to get my code into wine and libsdl22:31
markosbut having worked on most major SIMD platforms out there -except MIPS MSA- I'm just amazed at how much simpler it is to stuff on SVP64, even on assembly22:32
programmerjakethough it wasn't nearly as difficult22:32
markoss/to stuff/to do stuff/22:32
markosbtw regarding the av1 decoder part, they're already doing the dav1d decoder themselves22:33
markoswhen av1 becomes completely stable, they're going to fork that to av2 and want to have most optimizations for NEON already in there before that happens22:34
markoslibaom av1 that is22:34
programmerjakethey haven't already started on av2?22:35
markosin specs yes, code not afaik22:35
programmerjakei'd expect them to start right after releasing av122:35
markoswell, perhaps they have some older fork, but the goal is to sync and propagate all the optimizations to the av2 fork22:36
markosI'm not following av2 tbh22:36
markosso there is some pressure to finish everything before the summer22:36
programmerjakemaybe av3 will use machine learning for the decoder/encoder instead of dcts/arithmetic encoding22:37
markosif we had svp64 hardware to play with, I would probably finish a complete port in much less time22:37
markosyes, and require a 4090 Ti and 900Watts to do the decoding22:38
programmerjakei tried to build a machine learning lossless audio codec but never finished.22:38
programmerjakenot necessarily, some simpler algorithms can run on much weaker hw, e.g. a cellphone22:39
markosyou just do different kinds of LA calculations, ML is basically tons of LA and graph theory22:40
markosperhaps it would be quicker in finding common patterns in most videos and encode them faster22:41
markosah sorry you said audio22:41
programmerjakein case you're curious:
markosthis is a huge patch :)22:42
programmerjakeyeah, the sdl maintainer squashed a few weeks of work into 1 commit22:43
markosstill undecided whether squashing huge commits is a good thing in the end, sure it makes things easier to rollback in a single move22:44
markosI've worked in some projects with squashing and it did make some things easier22:45
programmerjakehere's the full unsquashed history:
programmerjakemy basic idea for a ml audio codec is to build a 100% reproducible neural net that predicts the next audio sample and store the difference between the prediction and the actual value22:48
programmerjakekinda like FLAC except with a different predictor22:49
programmerjakewe specifically rebased all our sdl commits so they would properly separate who did what so i was a little disappointed they all got squashed without even a Coauthored-By tag22:52
markosyeah that sucks22:56

Generated by 2.17.1 by Marius Gedminas - find it at!