alyssa changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard + Bifrost + Valhall - Logs https://oftc.irclog.whitequark.org/panfrost
Leopold__ has quit [Remote host closed the connection]
Leopold_ has joined #panfrost
Leopold__ has joined #panfrost
Leopold_ has quit [Ping timeout: 480 seconds]
stipa is now known as Guest6815
stipa has joined #panfrost
Guest6815 has quit [Ping timeout: 480 seconds]
camus has quit []
camus has joined #panfrost
PaulFertser has joined #panfrost
<PaulFertser> Hi there! Probably not important but very puzzling for me: I have RK3399 with Mali-T860 (in Pinebook Pro) running Mesa 22.3.6. With vblank_mode=0 es2gears_x11 on Xorg I get about 140 FPS, with es2gears_wayland under Sway I get about 750. Is this normal? Why is the difference so huge?
<PaulFertser> (tests were made full-screen 1920x1080; on Wayland the results improve considerably when es2gears window is smaller, on Xorg they stay the same)
Cyrinux947 has quit []
Cyrinux947 has joined #panfrost
rasterman has joined #panfrost
underpantsgnome[m] has quit []
Cyrinux947 has quit []
Cyrinux947 has joined #panfrost
atler has joined #panfrost
nlhowell has joined #panfrost
nlhowell has quit [Ping timeout: 480 seconds]
nlhowell has joined #panfrost
nlhowell has quit [Ping timeout: 480 seconds]
nlhowell has joined #panfrost
nashpa has joined #panfrost
dliviu has quit [Ping timeout: 480 seconds]
pbrobinson has quit [Ping timeout: 480 seconds]
pbrobinson has joined #panfrost
nashpa has quit []
dliviu has joined #panfrost
simon-perretta-img has quit [Quit: Leaving]
simon-perretta-img has joined #panfrost
kinkinkijkin has joined #panfrost
cphealy has joined #panfrost
rasterman- has joined #panfrost
rasterman- has quit []
rasterman- has joined #panfrost
rasterman- has quit [Remote host closed the connection]
simon-perretta-img_ has joined #panfrost
rasterman has quit [Ping timeout: 480 seconds]
simon-perretta-img has quit [Ping timeout: 480 seconds]
simon-perretta-img_ has quit [Ping timeout: 480 seconds]
rasterman has joined #panfrost
<CounterPillow> panfrost on xorg is not fully accelerated iirc, but that's okay because nobody should be using xorg anymore
alyssa has quit [Quit: alyssa]
<PaulFertser> CounterPillow: I'm not sure I understand, isn't DRI making it irrelevant? The DDX driver is modeset, it shouldn't be related at all.
Leopold__ has quit [Remote host closed the connection]
thaytan has quit [Remote host closed the connection]
thaytan has joined #panfrost
dliviu has quit [Ping timeout: 480 seconds]
stepri01 has quit [Quit: leaving]
dliviu has joined #panfrost
<robclark> maybe something about glamor is suboptimal compared to gl compositor? You could compare, for ex, glmark2 with wayland vs x11
<PaulFertser> I tried doing perf record for both es2gears runs and the results were pretty much the same afaict.
<PaulFertser> Is that worth pasting?
<PaulFertser> Yes, glmark2 gives extremely different results too, will show full logs a bit later. I suspect something is wrong with my Xorg config but it's all default Debian and moreover I'd like to understand where the difference might come from.
<rasterman> realistically you will not measure anything useful between the 2
<rasterman> 99% of your pipeline is either entirely in compositor code or in client code
<rasterman> the code is essentially the same
<rasterman> you do 99% of the same and simple the path to request a buffer swap is marginally different
<rasterman> both will send a request over a unix socket
<rasterman> ther are no copies going down any pipes
<rasterman> its "please take buffer X and swap it please"
<rasterman> on the other side from the client (x+compositor vs wayland compositor) it's a bit different
<rasterman> there is an extra bounce through the xorg server on the x side
<rasterman> ie xserver implements the swap then sends unix socket msgs to compositor (damage events) ...
<PaulFertser> Wayland glmark2: https://termbin.com/ne4m ; Xorg glmark2: https://termbin.com/ewk3
<rasterman> the compositor (unless it's using hw overlays) is going to be essentially the same code
<rasterman> ie do the same rendering work which tbh is 99% of the workload anyway
<rasterman> ie submit the same shader programs with the same triangles and same sized and formatted src buffers and dest buffers
<rasterman> and then compositor will bounce back to xserver to say "swap this buffer" to do the actual display
<rasterman> those extra bounces (which are async) from xorg -> compositor -> xorg are really mostly what differs
<rasterman> for glmark it'll be the same in fact the vast vast vast majority is the workload not the ipc bits
<rasterman> which again are all local to client or compositor
<PaulFertser> I do not think I'm running any compositor with Xorg though?
<PaulFertser> Am I supposed to?
<rasterman> you SHOULDNT have vastly differing numbers
<rasterman> they will have pretty small differences unless something is seriously wrong somewhere
<rasterman> oh
<rasterman> you arent running a compositor?
<rasterman> i think it may also depend on the compositor
<rasterman> i only test mine... :)
<PaulFertser> I never understood why I might need a compositor at all.
<PaulFertser> So yes, I'm running just Xorg.
<rasterman> but then the wl pipeline and x11 pipeling i know will be incredibly similar...
<rasterman> pipeline
<rasterman> aaaah
<rasterman> if it splain xorg then you will have the xserver doing memcyp's to the framebuffer
<rasterman> and thats single buffering
<PaulFertser> You can see from my perf report that it's about the same I think.
<rasterman> and that vastly changes the pipelin e on the x side
<rasterman> its actually stupidly inefficient to not use a compositor for things like this
<PaulFertser> Plain Xorg with DRI2/3, glxinfo saying direct rendering: yes etc.
desktoplover has joined #panfrost
<PaulFertser> So it's rendering directly but then it gets results and sends it to X server to copy to framebuffer?
<rasterman> gimme a bit
<rasterman> yur glmark scorews are vastly off
<PaulFertser> Wikipedia tells me compositing is all about weird effects none of which I need, so I never understood the utility.
<rasterman> but i think thats due to the lack of compositor oon the x side...
<rasterman> ummm no it's not all about thaf
<rasterman> it happens to allow those to be implemented
<rasterman> it allows a lot of things...
<rasterman> its about indirecting rendering
<PaulFertser> Indirect rendering faster than direct?
<PaulFertser> I was reading https://en.wikipedia.org/wiki/Compositing_window_manager but I couldn't figure it might affect performance of regular OpenGL apps just showing their things.
<rasterman> well indirect == rendering of an app goes to some buffer and is stored there
<rasterman> then a compositor takes over and that buffer gets displayed in some way by the compositor
<rasterman> it may just blast it to the screen
<rasterman> it may wrap it around a 3d spinning bunny rabbit
<rasterman> it may fade it in and out and zoom it in and out
<rasterman> add dropshadows or whatever it likes
<rasterman> this is simply a choice of what the compositor wants to do display-wise to give the user some kind fo experience
<rasterman> wtf is up with my rockpro64...
<PaulFertser> And without compositor with direct rendering isn't it supposed to just blast it on screen right away?
<rasterman> oh dang
<PaulFertser> My Mesa isn't the latest, it's what Debian stable ships, I can compile HEAD if needed.
<rasterman> thats what
Cyrinux947 has quit []
<PaulFertser> I understand Xorg is dead etc. But I like my xmonad with xmobar, I like synaptics touchpad driver (libinput doesn't have the options I need), I like setxkbmap working right away...
<rasterman> thats a problem then
<rasterman> no compositor
<rasterman> and wm and compositor are separate
Cyrinux947 has joined #panfrost
<PaulFertser> So if I just add xcompmgr to .xinitrc then it somehow improves glmark2 automagically?
<rasterman> that will have a likely worse pipeline as it'll use xrender to composite
<rasterman> i can go on about why ... but no.. that wnt really help
<rasterman> you need a gl based compositor
<rasterman> you probabkly need it to indirect just the client window and nothing more - not the frame + border a wm might add
<PaulFertser> What would be an example of such a compositor?
<rasterman> there was once one
<rasterman> i dont remember the name
<rasterman> glcompmgr?
<rasterman> i didnt pay much attention as my wm is a compositor and thats all i cared about
<rasterman> and it does things in a way to allow for optimal pipelines
<rasterman> gimme a bit
<rasterman> i'm collecting benchmarks
<rasterman> also after some updates trying to get xorg + panfrost to work again...
<rasterman> grrr X .. why u so blank?
<PaulFertser> It used to show that cool background by default. And supported Ctrl-Alt-Backspace...
<rasterman> what cool bg?
<rasterman> ummm
<rasterman> to each their own...
<rasterman> but that imho was not cool
<rasterman> :)
<PaulFertser> At least it wasn't all blank black, so one could see X started right away.
<rasterman> black is better
<PaulFertser> And then you have xsetroot in your session with whatever you really want (I want solid black).
<rasterman> as it at least starts grom an empty state and leaves it to your wm/compositor/desktop to start filling it with something
<PaulFertser> But it's good to have to see X started when you run it without session.
<rasterman> it's not
<rasterman> its glitchy
<PaulFertser> What wasn't glitchy in XFree86 times? :)
<rasterman> it makes the whole experience feel hacked togther without any polish at all
<rasterman> the fact that today it defaults to "empty with black and no pointer" is good™
<rasterman> hell i go to great effort to even have my desktop fade in from black at the very starts
<rasterman> so my machine boots
<rasterman> all i get is the firmware logo
<rasterman> screen goes black
<rasterman> it is black for afew seconds
<rasterman> then my desktop fades in...
<rasterman> no flickering, glitching, blinking etc.
<rasterman> just smooood
<rasterman> professional. i's dotted, t's crossed.
<rasterman> it plays my little login jungle to know it worked if i wasnt looking too :)
* PaulFertser prefers Linux messages from the drivers as they probe hardware...
<rasterman> on shutdown it fades to black and then just silently powers off
<rasterman> nicely
<rasterman> suspendresume do the same
<rasterman> fade in and out
<PaulFertser> Sounds cool indeed
<rasterman> screen blank/unblank from going udle fades in and out
<rasterman> hell it even fades your backlight in and out too
<rasterman> so blakc means black
<rasterman> not some dim gray because the backlight is fully on just with black pixels :)
<rasterman> if i have an issue i can always get my console and poke around :)
<rasterman> i just don't have issues 99.99% of the time so why make it look bad?
<rasterman> :)
<PaulFertser> And you always had the knack \ Fade to black
<mntirc> i always do Xorg -retro
<PaulFertser> (not-so-obscure Dire Straits reference to probably motivate someone to listen to those songs again)
<mntirc> :D
<mntirc> (gives a nice pattern or moire)
pbrobinson has quit [Ping timeout: 480 seconds]
pbrobinson has joined #panfrost
<rasterman> weird now it just doesnt boot anymore
<rasterman> sigh...
<rasterman> and there it goes
<rasterman> these numbers have changed a lot ... its now much worse in x than i remember
rasterman has quit [Quit: Gettin' stinky!]
Lucretia has quit []
desktoplover has quit [Remote host closed the connection]
desktoplover has joined #panfrost
Lucretia has joined #panfrost
Net147 has quit [Ping timeout: 480 seconds]
Net147 has joined #panfrost
stipa is now known as Guest6879
stipa has joined #panfrost
Guest6879 has quit [Ping timeout: 480 seconds]
desktoplover has quit []
dliviu has quit []
dliviu has joined #panfrost