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
zausor has quit []
Daaanct12 has quit [Remote host closed the connection]
nlhowell has quit [Ping timeout: 480 seconds]
digetx has quit [Ping timeout: 480 seconds]
digetx has joined #panfrost
<icecream95>
What's the chance that my shader didn't hit any compiler bugs?.. MESA_SHADER_COMPUTE shader: 279 inst, 128 bundles, 230 quadwords, 11 registers, 1 threads, 4 loops, 0:0 spills:fills
<icecream95>
Now I just need to make sure that there really is a way to do image writes to GL_TEXTURE_EXTERNAL_OES textures
<sekil>
From trace i'm see that drawframes spend many time to drawing - more than 16ms - anyone can help me - I understand it takes a lot of time to wait for the GPU to complete its work...
<sekil>
kernel 5.15.32
Didgy has joined #panfrost
<Didgy>
Hello, it's my understanding that Collabora is working on these Panfrost. I have a strong interest in learning to work on them also. Does anyone here know if Collabora would be interested in an internship for me? I already sent a mail a week ago but got no answer. Should I keep sending emails, or try to get in touch a different way?
<Didgy>
these Panfrost drivers*
sekil has quit [Remote host closed the connection]
<icecream95>
Maybe I shouldn't have stopped disabling NDEBUG for release builds
<icecream95>
Actually I think that assert should be removed, there's no reason that the BO has to be mapped
<icecream95>
Ugh, so the problem is that the texture's internal format is the generic GL_RGBA rather than GL_RGBA16F..
rasterman has joined #panfrost
<icecream95>
And the code which does *that* is in st_bind_egl_image..
rasterman has quit [Quit: Gettin' stinky!]
<icecream95>
It seems to be fixed now.. maybe? Format (v6): RGBA16UI RGBA
rasterman has joined #panfrost
soreau has quit [Ping timeout: 480 seconds]
rasterman has quit []
rasterman has joined #panfrost
<icecream95>
OES_texture_view to the rescue maybe?
<icecream95>
Oh wait, that has to read the internal format as well
<icecream95>
Okay.. that's two MRs created fixing things that likely no-one else will ever need
icecream95 has quit [Ping timeout: 480 seconds]
MajorBiscuit has quit [Ping timeout: 480 seconds]
rkanwal has joined #panfrost
Didgy has quit [Quit: Konversation terminated!]
davidlt has quit [Ping timeout: 480 seconds]
MajorBiscuit has joined #panfrost
Net147 has quit [Quit: Quit]
Net147 has joined #panfrost
Net147 has quit []
davidlt has joined #panfrost
soreau has joined #panfrost
greenjustin has joined #panfrost
Net147 has joined #panfrost
Net147 has quit []
Net147 has joined #panfrost
Net147 has quit [Remote host closed the connection]
<greenjustin>
Hi everyone! I've been experimenting with running panfrost on an MT8183 device and I've been running into some stability problems. Basically my dmesg fills up with errors like this (https://pastebin.com/bQmAxxfu) and then either the screen freezes or the kernel outright panics. Have any of you folks seen something like this before?
<macc24>
greenjustin: are you using cadmium?
<greenjustin>
Cadmium? I'm not totally sure what that is
<macc24>
ok, are you using any kernel patches?
cyrozap has quit [Ping timeout: 480 seconds]
<greenjustin>
Err probably a number of them, yes. This is a frankensteined chromeos build with kernel 5.10. I was able to get the driver to load by editing the MT8183 device tree to more closely match upstream, but I'm sure there's a number of deltas between this kernel and normal 5.10
Net147 has joined #panfrost
cyrozap has joined #panfrost
Net147 has quit []
<macc24>
oh my god
<macc24>
i refuse to watch this happen
macc24 has left #panfrost [NOPE]
karolherbst has quit [Read error: Connection reset by peer]
karolherbst has joined #panfrost
Net147 has joined #panfrost
<greenjustin>
Perhaps I should have explained that to begin with... I'm a chromeos dev experimenting with using panfrost on our older MT8183 based devices to help reduce the maintenance burden. There is an actual purpose to this, I just mostly don't know what I'm doing :P
Net147 has quit []
Net147 has joined #panfrost
Net147 has quit []
macc24 has joined #panfrost
<macc24>
hm
Net147 has joined #panfrost
<macc24>
greenjustin: can you post kernel source? i wanna take a look
<macc24>
didn't kukui ship with 4.19/4.14 kernel?
<greenjustin>
maybe originally? but this is jacuzzi on a 5.10
<greenjustin>
oh interesting! thank you so much, let me try that
<robmur01>
there might also be the question of whether the initial regulator voltages are high enough for whatever the initial clock rate is. devfreq just throws up its hands and runs away on MTK things with more than one regulator, so nothing's properly controlled to ensure a valid OPP
guillaume_g has quit []
<greenjustin>
the "regulator-always-on" change worked like a charm! thanks folks!
<greenjustin>
so for my own understanding, was the problem that the voltage was being improperly regulated and causing the GPU to brown out?
<macc24>
greenjustin: do you expect gpu to work when its SRAM is powered off?
<macc24>
robmur01: boot clocks should Just Work™ with default values... i think
<greenjustin>
macc24: Nope, fair enough. I guess I didn't really comprehend that log line meant disabling the SRAM. I'm shocked it displayed anything on the screen at all in that case. I'm guessing the other mediatek drm driver was still working and chrome was just using that to display images, but hardware acceleration was disabled?
<macc24>
greenjustin: gpu(3d accel) and displaying stuff are completely unrelated on most arm chips. you have to let that sink in
<macc24>
display controller's job is as different from gpu's job, as it's different from usb controller's job
<robmur01>
yup, panfrost is only concerned with rendering 3D prettiness to memory buffers, for the display controller (via its DRM driver) to turn into pixels
<greenjustin>
Yeah it took me a while to figure that out. I accidentally disabled the display controller driver instead of kbase, and I was staring a black screen all yesterday wondering why panfrost was loaded but nothing was happening
<robmur01>
in PC GPU land, the 3D driver and display driver (and video codec driver, and HDMI/DP driver, and audio driver, and etc.) are just all built together into one giant lump, which hides the distinction
<macc24>
robmur01: yea it causes SO MUCH CONFUSION
* macc24
remembers all the times she saw people asking about vaapi here
<greenjustin>
wait does mediatek even support vaapi? I thought most of the ARM stuff was all V4L2?
<macc24>
yea it is
<greenjustin>
oh lol
<macc24>
tho i wouldn't count on mtk hardware decoders being well supported under mainline linux
<greenjustin>
I'm not sure I would call them "well supported" at all, let alone on mainline :P
<macc24>
Just MediaTek Things
<greenjustin>
I'm working on outright circumventing MDP though, so maybe that will change a little bit
<macc24>
greenjustin: reverse engineering mm21?
<greenjustin>
macc24: yup that was me! and also mt21c until legal told me I couldn't lol
<greenjustin>
mm21 should be consumable in tip of tree chrome right about now though
<greenjustin>
oh wow I didn't know they were done with that already!
<greenjustin>
Anyways, thanks everyone for being so patient with me! Looks like webgl aquarium can run on chromeos with mesa/panfrost :-)
<macc24>
greenjustin: uh i suggest you get a proper fix working instead of what i suggested
<greenjustin>
macc24: Oh wait what? What would be the proper fix?
<macc24>
greenjustin: i have no idea
<greenjustin>
apologies, I'm still very new to this stuff
<greenjustin>
oh lol
<greenjustin>
macc24: so wait, what's wrong with the current fix? is it power inefficient because the sram will just always be on?
<macc24>
greenjustin: what i think should be done is that regulator should be used by gpu node, and in panfrost there has to be mt8183-specific devfreq code(including changing voltages on that crazy regulator setup of mt8183 ...)
<macc24>
iirc tomeu might have had some patch relating to that... that i procrastinated testing
<greenjustin>
macc24: that I can probably look into, I'm sure the kbase code has provisions for doing that
<greenjustin>
Would that maybe be related to some of the stuttering I'm seeing in the webgl stress tests?
<macc24>
i don't think so, but i'm not a graphics developer so i might be wrong
<robmur01>
I tried hacking something up for devfreq, but it revealed that the generic OPP API doesn't actually abstract well enough for drivers to hook into as they seem to be intended to. IIRC it needed a bit of redesigning to be able to manage regulator lifetimes correctly, at which point I gave up and went back to what I was supposed to be working on
<robmur01>
if you're effectively stuck at about 60% of your maximum clock rate, I could well imagine stress tests going a bit slow
MajorBiscuit has quit [Ping timeout: 480 seconds]
<greenjustin>
I guess I could test this fairly easily. Is there a simple way to query GPU freq or the active cycles counter in panfrost? I know there's a perfetto producer, but wiring that up seems heavy weight for basic metrics
<macc24>
haha but that only works if you have devfreq working
<greenjustin>
right...
<greenjustin>
debugfs not cooperating either
<macc24>
greenjustin: i promise that boot clock on mt8183 is awfully slow
<robmur01>
without devfreq, in principle nothing should have changed the clock rate from whatever we reported at probe time, but nothing's tracking actual activity
<macc24>
greenjustin: quick boot of my krane tels me that boot colck should be 511mhz
<greenjustin>
dmesg has a line claiming the clock is 511999970. weirdly specific, but that still seems high enough that I wouldn't expect this much stutter
<macc24>
heh
<daniels>
macc24: hey, MT8192 has codec patches on the list …
<macc24>
hint: it begins with "mt8192 chr" and ends with "omebook"
<HdkR>
Ah, that Reminds me that I need to drag a pile of laptops around
<macc24>
HdkR: it gets easier when the pile is only 8 laptops
<HdkR>
You're right, that would be easier if that was teh case
<macc24>
with couple more in spare parts
tanty has quit []
spawacz has quit [Ping timeout: 480 seconds]
<daniels>
macc24: hahaha
<macc24>
daniels: change it ;)
<greenjustin>
macc24: I decided to give cadmium a go to see if it repro'd the panfrost performance issues, but I'm getting a black screen on USB boot with my jacuzzi. Have you happened to see anything like this before?