ChanServ changed the topic of #lima to: Development channel for open source lima driver for ARM Mali4** GPUs - Kernel driver has landed in mainline, userspace driver is part of mesa - Logs at https://oftc.irclog.whitequark.org/lima/
<anarsoul> heh, if ladders are *very* expensive with the way we currently compile them
<anarsoul> we end up with a long chain of branches
<anarsoul> i.e. if unrolled loop had a single iteration, it'll end up branching multiple times. ouch!
ity1 has quit [coherence.oftc.net resistance.oftc.net]
cwabbott has quit [coherence.oftc.net resistance.oftc.net]
narmstrong has quit [coherence.oftc.net resistance.oftc.net]
daniels has quit [coherence.oftc.net resistance.oftc.net]
stilbruch has quit [coherence.oftc.net resistance.oftc.net]
tlwoerner has quit [coherence.oftc.net resistance.oftc.net]
anarsoul has quit [coherence.oftc.net resistance.oftc.net]
robher has quit [coherence.oftc.net resistance.oftc.net]
austriancoder has quit [coherence.oftc.net resistance.oftc.net]
dri-logg1r has quit [coherence.oftc.net resistance.oftc.net]
cyrozap has quit [coherence.oftc.net resistance.oftc.net]
anarsoul[m] has quit [coherence.oftc.net resistance.oftc.net]
gamiee has quit [coherence.oftc.net resistance.oftc.net]
linkmauve has quit [coherence.oftc.net resistance.oftc.net]
freemangordon1 has quit [coherence.oftc.net resistance.oftc.net]
uis has quit [coherence.oftc.net resistance.oftc.net]
jernej has quit [coherence.oftc.net resistance.oftc.net]
jelly has quit [coherence.oftc.net resistance.oftc.net]
mripard has quit [coherence.oftc.net resistance.oftc.net]
mmind00 has quit [coherence.oftc.net resistance.oftc.net]
enunes has quit [coherence.oftc.net resistance.oftc.net]
xdarklight has quit [coherence.oftc.net resistance.oftc.net]
Danct12 has quit [coherence.oftc.net resistance.oftc.net]
tanty has quit [coherence.oftc.net resistance.oftc.net]
rellla has quit [coherence.oftc.net resistance.oftc.net]
gamiee has joined #lima
<anarsoul[m]> Not for a single iteration though, but every extra iteration would result in unnecessary branch
<anarsoul> enunes: there is 3 more MRs for you to review :) but I'm almost done
<anarsoul> just write coalescing is left, and I'll probably try to tackle it from NIR side
<anarsoul> or rather mov coalescing
<anarsoul> btw, this particular deqp shader: https://gist.github.com/anarsoul/2549ae5bbca365c31a36bf32491885a9 is now compiled into just 6 instructions: https://gist.github.com/anarsoul/43f83abe9c93c9694a7c7841c34e22ab
<anarsoul> blob compiles it into 4, so there is still some room for improvement
<anarsoul> yet 12 -> 6 is massive improvement
<anarsoul> !33754 improves glmark2 "-b loop:fragment-steps=5:fragment-uniform=true:vertex-steps=5" by 10% (50 -> 55 fps)
fomys has joined #lima
dri-logger has joined #lima
dri-logg1r has joined #lima
dri-logger has quit [Ping timeout: 480 seconds]
<anarsoul> somehow mpv with "--gpu-context=x11egl" is way smoother for me than with wayland. Weird :)
<anarsoul> e.g. no drops playing sw-decoded 720p video on pinebook with x11egl, and significant drops on wayland
<linkmauve> anarsoul, both on the same compositor, with the only variable being using Xwayland or not?
<anarsoul> yeah
<anarsoul> both on sway
<anarsoul> tbh I don't want to dig into mpv internals again
<linkmauve> Maybe just open them an issue then?
<anarsoul> I have a strong suspicion that mpv does something wrong again
<anarsoul> they haven't fixed previous issue yet :)
<linkmauve> I wish the patches to use V4L2 requests in ffmpeg would get merged someday this decade.
<anarsoul> jernej: ^^ :)
<linkmauve> Otherwise reproducing this error is hard, as the latest branch from Kwiboo only applies on 6.something and the API changed enough that current mpv master doesn’t build against it.
<anarsoul> perfectionism is what kills a lot of opensource projects
<anarsoul> it will *never* be perfect. Merge it as is, fix it later if necessary
<linkmauve> Which reminds me I still have a JPEG driver to rebase and ship in Linux.
<anarsoul> I guess unless some big crop starts pushing v4l2-request into ffmpeg it'll never happen
<anarsoul> anyway, I tested my latest lima changes on a pinebook and I don't see any issues
<linkmauve> Kwiboo did a lot of work to fix it last year, I still believe~
<linkmauve> Do you have a branch with everything applied? I could test it on my PinePhone and various Olimex Lime boards.
<anarsoul> supertuxkart seems to be a bit smoother (mostly sits in 30-ish FPS for levels with non-complex geometry)
<linkmauve> Ta.
<anarsoul> what's interesting, if I run glmark2 on drm instead of x11 (glmark2 doesn't work on wayland for me) I get double the score in some benchmarks
<anarsoul> while on x11 it's way more humble
<anarsoul> (it's actually Xwayland on weston, not plain x11)
<anarsoul> and it makes sense for drm I guess, since some fragment shaders are now half the size with the latest changes
<anarsoul> linkmauve: btw you don't really need V4L2 request to reproduce the issue. I can easily reproduce it with libva if I force mpv to use GLES2
<anarsoul> it's basically just the gl2/gles2 code in mpv bit rotting since it's hard to find the hardware that doesn't support GL3 or at least ARB_framebuffer_object
<anarsoul> and I don't really understand why mpv wants ARB_framebuffer_object, they don't use depth buffer as far as I can tell
* anarsoul shrugs
<jernej> linkmauve, anarsoul: we all do :)
<jernej> there was little to no response on patches and latest kwiboo work is miles ahead of previous
<jernej> any corp backing is with gstreamer, which has proper V4L2 request support for years now
<anarsoul> I guess I am old fashioned. I use mpv
<anarsoul> (and I don't like gst, it has terrible api)