ChanServ changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - Logs https://oftc.irclog.whitequark.org/panfrost - <macc24> i have been here before it was popular
rasterman has quit [Quit: Gettin' stinky!]
JulianGro has quit [Remote host closed the connection]
bbrezillon has quit [Read error: Connection reset by peer]
bbrezillon has joined #panfrost
hyrc_ has joined #panfrost
hyrc has quit [Ping timeout: 480 seconds]
JulianGro has joined #panfrost
* WoC G'Night
MajorBiscuit has joined #panfrost
italove3 has joined #panfrost
Major_Biscuit has joined #panfrost
MajorBiscuit has quit [Ping timeout: 480 seconds]
rasterman has joined #panfrost
pbc has joined #panfrost
hyrc_ has quit []
chewitt has joined #panfrost
pbc has quit [Ping timeout: 480 seconds]
pbc has joined #panfrost
chewitt has quit [Quit: Zzz..]
wens has left #panfrost [#panfrost]
tolszak has joined #panfrost
tolszak has quit []
tolszak has joined #panfrost
<tolszak> Is there anyone so kind to test simple qml script on any of Mali-G52 devices with panfrost except odroid n2+?
<Rathann> tolszak: I only have a PineBook Pro around, with Mali-T860
<tolszak> Rathann: What OS?
<tolszak> PErhaps it is worth to check anything with panfrost
<tolszak> Rathann: Ive uploaded simple script to test it on any OS that has qml application installed: https://github.com/tolszak/panfrost_issue/tree/master
<Rathann> tolszak: Fedora Linux 35
<Rathann> tolszak: installing qml and its dependencies now
<Rathann> ok, it's running
<Rathann> it looks like I can change the scale using left and right arrows
<Rathann> what should I expect it to do?
<Rathann> tolszak: ^
<Rathann> ah, scale 16
<Rathann> looks fine to me
<Rathann> I have Qt 5.15.2 installed
<Rathann> kernel 5.15.16
<Rathann> lunchtime, bbl
<Rathann> ok, I'm back
<tolszak> Rathann: on odroid n2 when scale is 16 or more
<tolszak> the image gets distorted
<tolszak> You can try bigger scale, there is something wrong with interpolating vertex coordinates I think. I wonder if it is only n2+ or something else too.
<Rathann> I went as high as 300, no artifacts
<tolszak> Rathann: Thanks!
<Rathann> added a comment in the issue, as well
<Rathann> hope that helps
<Rathann> only left half of the image is visible on my screen though
<Rathann> or, rather, top left quarter
<robmur01> yup, seems to work OK on T830 too, lemme try sticking a Bifrost on the FPGA...
* robmur01 is obviously supposed to be doing Powerpoint today
jolan has quit [Quit: leaving]
jolan has joined #panfrost
Rathann has quit [Quit: Leaving]
<robmur01> Confirmed on G52r1! At scale 16 the view is suddenly centred just below box 761, at 31 it jumps to pure white, and at 34 it's back to the middle of the blue boxes again
<robmur01> also the menus in LXDE aren't rendering properly...
<robmur01> FWIW Mesa here seems to be git-7c01c9003f
<alyssa> robmur01: there's a patch for the LXDE issue, I think
<robmur01> ...which of course isn't a mainline commit. Based on main from 24th November, let's update that
<tolszak> robmur01: 31 pure white is probably cause the coordinate hits limit somwhere, when you increase scale should be alright again
<robmur01> tolszak: yup, at 34 it comes back
<robmur01> boo, X doesn't want to start on a G71 :(
<alyssa> robmur01: G71 isn't on the allowlist since it's probably broken have you *met* G71?
<alyssa> though straight up crashing, umm, shouldn't happen on main
<alyssa> I think rasterman fixed the latest such crash
<robmur01> what allowlist? :P
<alyssa> List of known Mali GPU IDs in Mesa, we refuse to probe on unsupported GPUs
<alyssa> QA measure
<alyssa> I think nowadays every Bifrost is allowlisted except G71 and G51
<alyssa> and G51 is probably fine, but I can't test that
<rasterman> robmur01: what do you have that has a g71?
<robmur01> I have two patches, one the make the HDLCD work, one to make my GPUs "work"
<alyssa> rasterman: an FPGA :-p
<rasterman> oh ha
<robmur01> oh, actually it's because the G71 broke at probe by failing to reset. I forgot that bitfile can be a bit flaky. May as well try a G51 then :D
<rasterman> robmur01: well print the gpuid - at least it doesnt seg anymore :)
<daniels> doesn't the hisi960 have a G71?
<rasterman> see my recent patch to add the g76 gpuid
<rasterman> its like 4 places you need to add itin
<robmur01> daniels: indeed, but what it lacks is an upstream display driver
<daniels> robmur01: it also lacks the ability for human beings to purchase it
<daniels> last I looked it was about as easy to source as your FPGA's bitstream
<robmur01> lol
<daniels> but yeah, the upstream story is really sad
<rasterman> hahaha
<rasterman> yeah... i gave up on the hikey960...
<rasterman> mine broke... and no new ones available
<alyssa> I have one sitting in a box somehwere.. um
<daniels> alyssa: tell me you didn't try to cross the border with that :P
<alyssa> daniels: no comment
<rasterman> hehe
<rasterman> thats how to add a gpuid if you need to.
<rasterman> it's not hard. it COULD be nicer and be in one spot with some array of structs but hey - the set of gpuid's is pretty finite and most worth talking about are there now
<alyssa> rasterman: usually the hard part is supporting a new architecture
<alyssa> the gpu id's are noise
<robmur01> rasterman: nah, all my local patch does is fall through the default case in panfrost_create_screen() and replace the unreachable() with a return in panfrost_model_name() :P
<alyssa> Crikey
<rasterman> alyssa: oh no one makes new gpu architectures... that's crazy talk! they're all the same re-cycled ones ... :)
<alyssa> new major versions
<alyssa> but yes
<rasterman> robmur01: hahah ok - very lazy "make it not fail" :)
<alyssa> CSF Malis are still basically midgard ;-p
<rasterman> my midgard is also still a midgard...
<rasterman> i guess i need a proper bifrost system orthet than my odroid go advance...
<rasterman> it's not super fun to work with
<alyssa> also Valhall's tiling scheme is the same as Utgard's for fun
<rasterman> if it ain't broke...
<robmur01> *shakes fist* damn you, panfrost_get_quirks(), how dare you not understand G51 already!
<robmur01> OK so there may sometimes be more places to hack...
<alyssa> robmur01: add to the same as G31/G52 i guess
nlhowell has quit [Remote host closed the connection]
nlhowell has joined #panfrost
<robmur01> looks like it, although the datasheet says it also has a version dependency for aniso like G72
<alyssa> Womp womp.
<robmur01> happy glmark, that's better :)
<robmur01> menus are back too, maybe because G51, maybe because new Mesa build - so scientific!
<robmur01> shock horror, tolszak's test is *not* broken on G51 :O
<rasterman> hahaha
<tolszak> robmur01: I wonder who is the smartest person in this room and can fix it on G52 in reasonable amount of time. I'm only sure that it is not me :)
<alyssa> I nominate jekstrand
<alyssa> :-p
<robmur01> nor me, since it's not a kernel issue :)
<alyssa> It's certainly not me because have you *seen* my todo list
<tolszak> robmur01: I've also noticed something that might be related to the issue. I have 18MP camera that provides MJPEG output which is decoded to YUYV. Which in turn is rendered using simple shader (inside Qt)
<tolszak> And sometimes colors are broken on screen, The jpeg from camera is definitely correct.
<robmur01> for completeness, back to G52 with new Mesa, LXDE good, testcase still bad, phew! Science!
<tolszak> Probably coordinates which are to pick parts of yuyv plane break too (like vertex coordinate in example I shared). But that's only a guess
<alyssa> robmur01: hold on so I can allowlist G51?
<alyssa> what GPU ID and quirks?
<alyssa> if you send the MR i can assign to marge and be lazy but IDK if the sign offs are too much work for you guys
<alyssa> it'd be cool to have all bifrost allowlisted..
<robmur01> 0x7090, and I used G31 quirks since mine's r1p1 which is when it gained aniso support according to the datasheet. No idea about swizzles
tolszak has quit [Ping timeout: 480 seconds]
<alyssa> ack
<alyssa> HAS_SWIZZLES predates tomeu and I understanding Bifrost, it shouldn't exist
<alyssa> is probably unused too
<robmur01> AFAICS it selects between pipe_format_v6 and pipe_format_v7, which smells like it might be a candidate for the arch major version anyway
Major_Biscuit has quit [Ping timeout: 480 seconds]
<alyssa> robmur01: Correct, it should just be a dev->arch check directly
<alyssa> In hindsight, the vast majority of issues/features should just be an arch major check (and those get inlined / constant folded in the hot paths of the driver!)
<alyssa> and quirks.h should only be for model-specific differences (like aniso filtering)
Stary has joined #panfrost
mriesch has joined #panfrost
Stary_ has joined #panfrost
Stary is now known as Guest577
Stary_ is now known as Stary
Guest577 has quit [Read error: Connection reset by peer]
ndufresne7 has joined #panfrost
ndufresne7 is now known as ndufresne
kinkinkijkin has joined #panfrost
macc24 has joined #panfrost
<macc24> so i finally connected my kevin to a 2560x1440p monitor
<macc24> and on 1920x1080 that mali t860 really feels fast...
<macc24> :/ sucks that the resolution is the biggest perf limit on kevin
<macc24> i wonder if i can drive internal output at 1200x800
<macc24> is it me or are supertuxkart karts too dark on panfrost on mesa 21.3.4
<macc24> on mali t860
<macc24> nvm it's fine, forgot to reboot after update
Bennett has joined #panfrost
<alyssa> Midgard is so cute, I just want to squish it!
<HdkR> I have a mallet that could help
<alyssa> rofl
<HdkR> Big squish
<alyssa> robmur01: 7 files changed, 31 insertions(+), 116 deletions(-)
<alyssa> this is your fault :-p
<robmur01> my favourite kind of diffstat!
<alyssa> I started by deleting HAS_SWIZZLES and then proceeded to delete half the rest of the file too! :-p
pbc has quit [Ping timeout: 480 seconds]
<alyssa> rasterman: Trying out the central model list
<alyssa> bikeshedding hard the cosmetics..