alyssa changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - https://gitlab.freedesktop.org/panfrost - Logs https://freenode.irclog.whitequark.org/panfrost - <daniels> avoiding X is a huge feature
<alyssa> You have schooled me well.
NeuroScr has quit [Ping timeout: 246 seconds]
NeuroScr has joined #panfrost
NeuroScr has quit [Ping timeout: 245 seconds]
vstehle has quit [Ping timeout: 245 seconds]
NeuroScr has joined #panfrost
_whitelogger has joined #panfrost
davidlt has joined #panfrost
<daniels> \o/
<forkbomb>
<HdkR> \_/()\_/
<HdkR> Hm, nonsense characters
robert_ancell has quit [Quit: Leaving]
vstehle has joined #panfrost
NeuroScr has quit [Ping timeout: 258 seconds]
NeuroScr has joined #panfrost
<Depau> The mouse pointer looks a bit weird on sway, though
<Depau> I'll try supertuxkart now ;)
NeuroScr has quit [Ping timeout: 248 seconds]
<daniels> Depau: i hate to burst your bubble, but llvmpipe (seen in your kmscube debug output) is the Mesa software renderer
<Depau> Yep, that's before I patched mesa
<Depau> Kmscube crashes with llvmpipe
<daniels> ah, I see
<daniels> cool!
<Depau> (sorry for the pictures of the screen, I haven't installed the program to take screenshots yet) https://i.postimg.cc/j5N0RbQX/IMG-20190725-081510.jpg
<Depau> Here you can see both that it's using panfrost and the weird pointer I was talking about
<tomeu> very cool!
<Depau> BTW glxgears segfaults, I'll check with gdb later
guillaume_g has joined #panfrost
<rtp> Depau: you may want to try with the 4 patches from the sfbd branch in tomeu's mesa repo on gitlab
yann has quit [Ping timeout: 268 seconds]
<rtp> tomeu: btw, do you know if alyssa plans to merge it or does it need more work/test/... ?
<tomeu> rtp: don't know, do you think ti has been tested enough?
<rtp> tomeu: I guess I'm so far the only one having tested it so not sure it's enough. I'm only wondering the status
<tomeu> rtp: well, we know it most likely doesn't regress rk3399 and rk3288 as per the CI
<tomeu> what I'm not sure yet is if we want to whitelist that GPU
<tomeu> but given we have quite some people interested in the snow...
<guillaume_g> tomeu: before whitelist, I think modifier invalid/linear is a blocker, no?
<tomeu> yeah, what can be done about it?
<rtp> tomeu: well, I'm not using snow, I'm using peach pit. For now, my snow is collecting dust as I've yet to create a root fs for tests :)
<rtp> guillaume_g: yeah ...
<tomeu> ah, ok, though both have 76xx, which is the problematic part
<tomeu> kbase has a very long list of errata...
<tomeu> s/76xx/t6xx
<rtp> they're both t6xx. One is 604 and the other 628
<tomeu> right, that's what I meant
<guillaume_g> yeah, t604 in snow is very early rev with some erratas
<tomeu> let's see what alyssa says, but I think it's very cool that people are working on t6xx
<tomeu> I could trivially schedule CI jobs to snow and peach pi once the modifiers issue is solved
<daniels> guillaume_g: the modifier thing is a kmscube bug
<daniels> just comment out the line which adds the MODIFIERS flag to the drmModeAddFB2WithModifiers call
<rtp> tomeu: fwiw, I've always wanted to see it working on t6xx (even opened a bug on gitlab at months ago) but stopped due to -enotime and failing to get logs from the blob
<daniels> (the correct fix is to check for the drm ADDFB2_MODIFIERS client cap or whatever it's called, and to not pass any modifiers to KMS if it isn't present)
* rtp notes to try again without the linear change something else like glmark2 then
<guillaume_g> daniels: oh really, just a kmscube issue? Cool. :)
<daniels> yeah
sravn has quit [Quit: WeeChat 2.4]
adjtm has quit [Quit: Leaving]
sravn has joined #panfrost
_whitelogger has joined #panfrost
davidlt_ has joined #panfrost
davidlt has quit [Ping timeout: 248 seconds]
somy has joined #panfrost
davidlt__ has joined #panfrost
raster has joined #panfrost
davidlt_ has quit [Read error: Connection reset by peer]
guillaume_g has quit [Ping timeout: 244 seconds]
adjtm has joined #panfrost
guillaume_g has joined #panfrost
cwabbott has quit [Quit: cwabbott]
cwabbott has joined #panfrost
yann has joined #panfrost
cwabbott has quit [Client Quit]
cwabbott has joined #panfrost
<alyssa> tomeu: I'll merge the sfbd patches this morning, I was just waiting on confirmation they did the expected thing since I did it w/o hw or anything
<tomeu> who needs hw...
<alyssa> However, T600 in general must remain blacklisted for the near future
<rtp> alyssa: too early / too many things to solve ?
<Depau> I tried the sfbd branch on snow, I see no visible changes though I haven't checked what the patches actually do
davidlt__ is now known as davidlt
<rtp> Depau: fixes segfaults with (at least) glmark2-es2-drm
<Depau> It doesn't fix it with glxgears, it appears to get EACCES with ioctl DRM_IOCTL_MODE_CREATE_DUMP, then it segfaults when doing DRM_IOCTL_PANFROST_SUBMIT
<Depau> I'll collect more information and see if I can patch it myself, I don't want to add extra burden on you guys if you're not planning to support t604 for the time being
<alyssa> rtp: Pretty much, yeah
<alyssa> Whitelisting means we're ready for end-users who've never heard of Panfrost to use the driver
<alyssa> Given glxgears is segfaulting.... definitely too early :)
<guillaume_g> glxgears runs fine here. :)
<Depau> guillaume_g: Yeah my snow is the unlucky one haha
<Depau> guillaume_g: do you get the permission denied? https://hastebin.com/simoqehapi.cs
<guillaume_g> Depau: No, I did not
BenG83 has joined #panfrost
<alyssa> Depau: # chmod 777 /dev/dri/*
<Depau> alyssa: I chmoded it 666 but it didn't work
<Depau> Why is that needed, though?
<alyssa> Depau: It shouldn't be but thought maybe permission denied?
<Depau> alyssa: IDK. I rebuilt Mesa and glxgears without stripping the debugging symbols, I'll check with gdb later. I'll also test glmark2
<alyssa> Meanwhile, I have about 100 regressions to sort. Weee.
* alyssa broke txf
yann has quit [Ping timeout: 245 seconds]
* alyssa broke bias
<alyssa> You know Midgard, you're starting to getting a little annoynig
<alyssa> *annoying
<alyssa> (Needing to put bias in upper half of a reg..)
<alyssa> Nope, that's not true, just saw an example with .y as well
<alyssa> So I've seen x/y/z
<alyssa> er
<alyssa> y/z/w
<alyssa> Still need to see X.
<alyssa> ohhhh I do have something flipped
<alyssa> something decodd as hr56.w, should be r28.z
<alyssa> Yup, flipped, fixed nice
<alyssa> (Bias/lod should work now)
yann has joined #panfrost
<alyssa> Fixed more things, next set of regressions will actually be txf this time ummm
<alyssa> txf is.. annoying on midgard
<alyssa> The strategy I guess is emitting an imov 0 alongside
herbmilleriw has quit [Ping timeout: 244 seconds]
herbmilleriw has joined #panfrost
<alyssa> ...does my copyprop respect masks?
raster has quit [Read error: Connection reset by peer]
stikonas has joined #panfrost
adjtm has quit [Ping timeout: 245 seconds]
<alyssa> "g.sqtp.wyzx.b" dEQP shaders totally make sense :V
<alyssa> Ugh, ok, this regression is..... subtle
<alyssa> I see the issue, just trying to wrap my head around what it really means
<alyssa> Involving CF shenanigans
<alyssa> Okay, I see
* alyssa adds moar validation
<alyssa> So now the new validation check I add is failed for this regressing case
<alyssa> Hence, up to the prepass to deal with
<alyssa> Wait but that doesn't make any sense :V
<alyssa> The issue is reading the TEXW when we w-- no, I guess it is right, it's just really weiiiiiird
<alyssa> There we are
<alyssa> The code gen is pretty bad in that case but it's a purely CTS issue, I think (hope)
<alyssa> A new regression
<alyssa> You've got to be kidding me
adjtm has joined #panfrost
<alyssa> Oh ffs
<alyssa> Oh ffs!
<alyssa> Somewhere in there we regressed half of glmark :blink:
<anarsoul> :(
<alyssa> anarsoul: Fixed maybe
<alyssa> I suppose my IRC logs are entertaining to read retroactively, huh :P
<mifritscher> alyssa: to be honest - yes :-) I'm doing this for months now ;-) But this is also a good way to learn a lot :-)
<alyssa> mifritscher: Hehe :) I appreciate the honesty ;)
<mifritscher> and I can understand this style of working well - because I catch myself often working this way as well *g*
<alyssa> mifritscher: Style, being cursing on IRC?
<mifritscher> the bad thing is: I curse in real a lot - I got even sometimes troubles with my girlfriend *g*
<mifritscher> at my buero there is the joke "Michael does close the door? ok, he must work with powerpoint or word the next time" ^^
<mifritscher> but it helps to keep cool inwardly
<mifritscher> (I'm forced to work with MS products at work sometimes :-/ )
<mifritscher> Sadly I don't get any compensation despite I already asked for ...
guillaume_g has quit [Quit: Konversation terminated!]
<alyssa> :(
<alyssa> So, functional regressions aside, time to look at some perf regressions in individual shaders
<alyssa> Oh, yeah, that's a bug.
* alyssa <3 shaderdb
<alyssa> One perf bug fixed
<alyssa> Wat. The disassemblies are the same.
<alyssa> Oh, wait, no
<alyssa> Ok, yeah, this is also a regression
somy has quit [Ping timeout: 276 seconds]
davidlt has quit [Ping timeout: 245 seconds]
stikonas has quit [Remote host closed the connection]
<alyssa> There, sorted with a dumb opt pass :p
<alyssa> Aaand the last perf-related bug was unrelated to this pass, a scheduler bug actually
<alyssa> total bundles in shared programs: 2519 -> 2472 (-1.87%)
<alyssa> bundles in affected programs: 791 -> 744 (-5.94%)
<alyssa> So that was a nice win to find and fix :)
rhyskidd has quit [Quit: rhyskidd]
BenG83 has quit [Quit: Leaving]
stikonas has joined #panfrost
stikonas has quit [Remote host closed the connection]
<alyssa> Unfortunately that regressed a few dEQP cases
<alyssa> Staring at the disassembly, I don't see why
<alyssa> Maybe there's some scheduling req I don't know about
<alyssa> Nope, here's a shader from the blob that uses the pattern of both smul and sadd moving to diff components of the same reg
<alyssa> Well, close enough anyway
<alyssa> Oh, I see it now.
<alyssa> RA bug maybe
* alyssa glances at clock
<alyssa> To be sorted out tomorrow, it looks like :)
raster has joined #panfrost
raster has quit [Remote host closed the connection]