alyssa changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - Logs https://freenode.irclog.whitequark.org/panfrost - <daniels> avoiding X is a huge feature
maciejjo has quit [Ping timeout: 246 seconds]
maciejjo has joined #panfrost
vstehle has quit [Ping timeout: 265 seconds]
warpme_ has quit [Quit: Connection closed for inactivity]
atler has quit [Ping timeout: 252 seconds]
atler has joined #panfrost
chewitt has quit [Quit: Adios!]
pendingchaos has quit [Quit: No Ping reply in 180 seconds.]
pendingchaos has joined #panfrost
WoC has quit [Ping timeout: 250 seconds]
davidlt has joined #panfrost
WoC has joined #panfrost
nlhowell has quit [Ping timeout: 246 seconds]
vstehle has joined #panfrost
zkrx has quit [Ping timeout: 252 seconds]
zkrx has joined #panfrost
unoccupied has joined #panfrost
unoccupied is now known as leah
guillaume_g has joined #panfrost
yann has quit [Ping timeout: 252 seconds]
vstehle has quit [Quit: WeeChat 3.0]
vstehle has joined #panfrost
stikonas has joined #panfrost
stikonas has quit [Remote host closed the connection]
raster has joined #panfrost
warpme_ has joined #panfrost
wwilly has quit [Quit: Leaving]
amonakov has quit [Ping timeout: 260 seconds]
chewitt has joined #panfrost
yann has joined #panfrost
amonakov has joined #panfrost
adjtm_ has quit [Quit: Leaving]
ente has quit [Quit: Caught SIGSEGV, exiting...]
AreaScout_ has joined #panfrost
<warpme_> guys: i'm trying to get llvmpipe working on current 21.1.0-rc3 cross-compiled mesa. cross-compiled llvm is 12.0.0; hw is aarch64 soc (h6). App segfaults like this (in JIT i think): https://pastebin.com/5MNKiuU3 Maybe somebody have some pointers here?
yann has quit [Ping timeout: 260 seconds]
<alyssa> italove: Just took your disassembler for a spin
<alyssa> I added some minor cosmetic adjustments, overall I'm super hthrilled with this
<alyssa> and a lot less scared to debug this midgard regression I hit
<alyssa> kudos!
<amonakov> better disasm sounds lovely, where should I look?
<alyssa> amonakov: this is for midgard
<amonakov> ah
guillaume_g has quit [Quit: Konversation terminated!]
<macc24> icecream95: by fixing display on mainline-mainline 5.13 on duet it looks like i have accidentally fixed resuming too
yann has joined #panfrost
catfella_ has quit [Ping timeout: 240 seconds]
Stenzek has quit [Ping timeout: 276 seconds]
Stenzek has joined #panfrost
archetech has joined #panfrost
<daniels> italove: <3
<italove> alyssa: daniels: I'm glad you liked it :)
<raster> alyssa: hey hey... do you feel like compiler optimizing?
<raster> ddk is about 2.1x as fast (ie does it in half the time)
<raster> it's a wayland client again... but it does a 2 pass 7 tap sample blur with 2x downa nd upscale
<raster> it'll just report time taken with some glreadpixels to force the pipeline
<raster> but this is probably the core of why "kde plasma is slow" :)
<raster> time for real life though...
davidlt has quit [Ping timeout: 240 seconds]
<alyssa> eyes
<alyssa> i like compiler optimizing
archetech has quit [Quit: Konversation terminated!]
<icecream95> alyssa: Why does the gl_FragColor lowering do a load from the tilebuffer to get the colour to write to the other RTs?
<icecream95> (My guess at an answer: Because someone put NIR passes in the wrong order?)
<alyssa> This is a good question..
<icecream95> nir_lower_fragcolor does nir_ssa_def *frag_color = nir_load_var(b, out);
<alyssa> - nir_ssa_def *frag_color = nir_load_var(b, out);
<alyssa> + nir_ssa_def *frag_color = instr->src[1];
<alyssa> does something like that do it?
<icecream95> Well, nir_ssa_for_src(b, instr->src[1], 4)
megi has quit [Quit: WeeChat 3.1]
megi has joined #panfrost
stikonas has joined #panfrost
adjtm has joined #panfrost
<alyssa> raster: ..this doesn't run surfaceless does it.
<raster> alyssa: no. wayland client with real surface
<raster> just run weston :)
<raster> or something
<italove> ooc, is it bifrost or midgard?
<alyssa> the board isn't connected to a monitor
<raster> you dont need to see it
<raster> it'll spew out timing info to stdout
<alyssa> raster: Eyeballing the shader, this hits our lack of parallelism in the compiler
<raster> it probably does
<raster> i was testing on midgard right now
<raster> but this was a result of that plasma report
<raster> i thought i'd cook up a more isolated case
<alyssa> currently all of the tex ops will depend on the each other in sequence
<alyssa> (on panfrost)
<raster> (thats easier to inspect and debug)
<raster> it isnt the same blur algo tho
<alyssa> whereas the DDK can parallelize that
<raster> kwin uses double-kawase...
<raster> which i was going ot implement later
<raster> but i first did the most common 2 pass gaussian-approximation blur that is common enough
<raster> well thught i'd help as its not something glmark2 does :)
<raster> so some other workloads isolated out to play with
<raster> either way i do know ddk pretty much comes in at 2x the speed
<raster> a bit more ... so it's definitely doing a better job.
<alyssa> glmark2 does have a blur shader in there somewhere
<alyssa> i think
<alyssa> not nice and isolated like this tho :+1:
<raster> eg? it does do blur? where?
<raster> i never rememebr seeing it... hmmm
<raster> oh wait
<raster> it was the shadows in the "compositing" test
<raster> right?
* alyssa can never remember how to start weston from ssh
<raster> aaaah
<icecream95> alyssa: WLR_BACKENDS=headless sway
<alyssa> or that
<icecream95> And that even hits an assertion in Panfrost on exit...
<raster> i set up one of my vt's to auto-log in
<raster> /etc/systemd/system/getty@tty1.service.d/override.conf
<icecream95> alyssa: On the topic of texture instructions, what does the skip modifier do on Bifrost?
<alyssa> Oh boy.
<raster> and a [service]
<raster> ExecStart=-/sbin/agetty --skip-login --noissue --login-options "-f raster" %I linux
<alyssa> icecream95: .skip on Bifrost does the same thing as .skip on Midgard.
<raster> or whatever your user is
<alyssa> ("Not helpful.")
<raster> and i have my ~/.profile run screen
<raster> so i can screen -rd
<raster> to sonnect to a tty screen session and thus run any compositor as if it were on the tty
<raster> :)
<alyssa> icecream95: .skip means "Skip the instruction if the thread is discarded [including helper invocations]"
<raster> once u have this set up to auto-log in and run screen this way life is easy
<raster> so /etc/systemd/system/getty@tty1.service.d/override.conf is
<alyssa> icecream95: midgard_helper_invocations.c does this on midgard, where .skip is !ctx->helper_execute
<raster> and my ~/.profile is
<raster> from there on in just ssh in, screen -rd
<raster> and bingo
<raster> console cmdline shell
<raster> :)
<alyssa> ugh weston
<alyssa> [18:16:08.722] logind: failed to get session seat
* alyssa digs through chat logs
<alyssa> I know for a fact daniels explained this to me before..
<alyssa> [18:16:52.742] fatal: environment variable XDG_RUNTIME_DIR is not set.
<alyssa> ^ as root
<raster> useing the screen method?
<raster> did you run westojn as the same user that is logged into the tty?
<alyssa> Oh panfrost bug. lovely.
<raster> panfrost? bugs? NEVVVAAAAH!
<raster> :)
<alyssa> shh :p
<daniels> weston —backend=headless-backend.so !
<raster> btw ... i was amazed at how fast a desktop gpu is at this blurring fun
<raster> i had to like do i htink 20,000 or maybe 50,000 iterations on a 6800xt to match the time it took for my midgard gpu to do 500 :)
<alyssa> raster: will have to try on the m1 heh
<alyssa> daniels: that worked! wasn't in the man page tho
<raster> i wonder how the m1 will do
<alyssa> raster: libEGL warning: MESA-LOADER: failed to open swrast: /home/alyssa/lib/dri/swrast_dri.so: cannot open shared object file: No such file or directory (search paths /home/alyssa/lib/dri)
<raster> its worth comparing ot midg/bifrost as it's integrated and sharing the same mem
<alyssa> it does not seem to want to use hw?
<raster> dedicated gpu's have the dvantage of mountains of cache and dedicate fast ram
<raster> well the 6800xt i think has like 128m of cache?
<raster> it's silly amounts
<HdkR> "Infinity Cache" is quite silly :)
<raster> thats probably whyt he blur is so fast on the 6800ct
<raster> all the dataset fits in cache
<daniels> alyssa: it’s just eglGetDisplay
<daniels> (ok it’s eglGetPlatformDisplayKHR(EGL_PLATFORM_SURFACELESS_MESA) with a null display and attribs for all you smarties in the back)
<daniels> but the point remains, we’re not doing anything to steer Mesa away from (or, tbf, towards) hw
<alyssa> weston is fine, that was raster's app
<raster> yeah. my test app is a wl client
<daniels> oh
* daniels coughs
<daniels> weston --backend=headless-backend.so --use-gl
<alyssa> thanks
<raster> i just use tty auto-login + screen
<alyssa> blur-perf: ../src/gallium/drivers/panfrost/pan_blit.c:79: panfrost_u_blitter_blit: Assertion `!"Unsupported blit\n"' failed.
<raster> then all compositors work
<alyssa> oh come on now what
<raster> :)
<daniels> alyssa: crafty of bbrezillon to set a trap like that
<alyssa> gotta coax me into review duty somehow
<alyssa> my chromebook feels so slow after m1'ing
<alyssa> yeesh
<macc24> alyssa: just get a qualcomm laptop and start developing freedreno >:D
<HdkR> Nah, that has Google backing, M1 is a better target since Apple won't support OSS there :)
<macc24> muh not doing anything
stikonas has quit [Ping timeout: 260 seconds]
raster has quit [Quit: Gettin' stinky!]