alyssa 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
<cphealy>
If AFBC_FORMAT_MOD_TILED is not set and AFBC_FORMAT_MOD_USM is set, do I get an AFBC buffer that has linear/uncompressed contents?
<icecream95>
AFBC_FORMAT_MOD_TILED is only for inter-superblock tiling, the superblocks will always be tiles and have 4x4 tiles inside them
<cphealy>
icecream95: got it, tnx.
Tk55 has joined #panfrost
Tk55 has quit [Remote host closed the connection]
vstehle has quit [Ping timeout: 480 seconds]
pauper has joined #panfrost
pauper has quit [autokilled: Suspected spammer. Mail support@oftc.net with questions (2021-06-04 00:46:02)]
atler is now known as Guest794
atler has joined #panfrost
Guest794 has quit [Ping timeout: 480 seconds]
adjtm has joined #panfrost
djancak has joined #panfrost
djancak has quit [autokilled: Suspected spammer. Mail support@oftc.net with questions (2021-06-04 01:51:41)]
zeiris16 has joined #panfrost
zeiris16 has quit [Remote host closed the connection]
DanDare has joined #panfrost
DanDare has quit [autokilled: Possible spambot. Mail support@oftc.net if you think this is in error. (2021-06-04 02:31:26)]
<robclark>
cphealy: a bandwidth compressed but not tiled format seems unlikely.. I've not come across one.. I think generally you assume that if a hw block grok's compressed it also grok's tiled
<robclark>
and if it doesn't, then simple dumb linear
<alyssa>
compressed non-tiled doesn't even make much sense, just spatially
<alyssa>
iirc even JPEG is sorta tiled?
* alyssa
wikipedias JPEG since she might be making things up
<alyssa>
yeah, 8x8 tiles for DCT
<alyssa>
Aside - when did the term "swizzled" go out of fashion in favour of "tiled"?
<alyssa>
I would say it's better since swizzled is overloaded (texture swizzles, format swizzles, shader swizzles), but then again, so is tiled (tiled rendering, tiler, tiling)
<HdkR>
We need more quilted formats
<cyrozap>
What about gerrymandered formats?
<alyssa>
cyrozap: My vote is for quilted but that doesn't matter.
<cyrozap>
We could use "gerrymandered" for formats where the data is partitioned in an arbitrary way so as to produce a particular result :P
<alyssa>
cyrozap: Hmm, instead of "specially crafted input"?
pglife_bot has joined #panfrost
pglife_bot has quit [autokilled: Suspected spammer. Mail support@oftc.net with questions (2021-06-04 03:16:56)]
<robclark>
yeah, compression metadata is block oriented, so it would be kinda dumb to no have "pixel" data be in a block layout
Danct12 has joined #panfrost
Payhn has joined #panfrost
Payhn has quit [Remote host closed the connection]
vstehle has joined #panfrost
MrCircuitMatt has joined #panfrost
MrCircuitMatt has quit [Remote host closed the connection]
macc24 has joined #panfrost
SexiestManAlive has joined #panfrost
SexiestManAlive has quit [Remote host closed the connection]
<Dylanger>
It's not permissions as far as I can tell
<Dylanger>
Thanks for helping anyway
oskie1 has joined #panfrost
oskie1 has quit [Remote host closed the connection]
nroberts has joined #panfrost
<icecream95>
Dylanger: Have you tried setting the EGL_PLATFORM=surfaceless environment variable?
shadowpaste has joined #panfrost
shadowpaste has quit [Remote host closed the connection]
<Dylanger>
Hm just tried that, same result
<HdkR>
Yea. If it isn't a permission issue then more likely is that it couldn't guess which platform you were wanting to use correctly
<HdkR>
or your drivers aren't compiled with every platform
<Dylanger>
I was thinking about trying panfrost, see if that just magically fixes it
<Dylanger>
Over from `/vendor/lib64/egl/libGLES_mali.so`
<HdkR>
oh
<HdkR>
Mali blob...
<Dylanger>
That's what panfrost is right?
<HdkR>
depending on your device that will only support fbdev, wayland, or x11 on Linux
<Dylanger>
It's a replacement for that blob?
<HdkR>
panfrost is a replacement for the blob yes
<Dylanger>
That blob is damn huge, it's like 47MB
<Dylanger>
<HdkR "depending on your device that wi"> Hm, Android?
adjtm has joined #panfrost
<macc24>
Dylanger: what gpu do you have
<Dylanger>
Mali-G57 MC4
<macc24>
oh panfrost won't run on it. it's valhall
<Dylanger>
valhall too new?
<macc24>
yea
<macc24>
and there are no devices to reverse it on that run linux
<Dylanger>
I think it could be possible my driver doesn't support SURFACELESS platform
<Dylanger>
I would not want to reverse that lol, that thing would crash IDA
<HdkR>
If you're on Android then you can still use surfaceless. Just has to be from command line
<Dylanger>
hm
<HdkR>
Inside of a java/jni app you have to do shenanigans
Yendred has joined #panfrost
Yendred has quit [Remote host closed the connection]
adjtm has quit [Quit: Leaving]
Oddmonger11 has joined #panfrost
Oddmonger11 has quit [Remote host closed the connection]
<tomeu>
bbrezillon: could be, some tests create lots of cmdbuffers
<bbrezillon>
tomeu: it has to be in a lot of commands in a single command buffer, with no batch split
<macc24>
bbrezillon: how's panvk?
<bbrezillon>
still buggy :P
<bbrezillon>
and incomplete
<macc24>
so on g72 it's super buggy?
<tomeu>
super untested, probably? :)
<macc24>
idk i can test it
<macc24>
dunno if results would of any use
<tomeu>
it's probably a bit early
<macc24>
ok
* macc24
goes back to sleep for another week
<alyssa>
macc24: otoh if you want to test GLES3.1 ... ;-)
<macc24>
alyssa: there is a CI, you know?
<alyssa>
macc24: The CI is that we push stuff and then you play video games automatically and complain if it breaks. Continuous integration.
<alyssa>
;-P
<macc24>
i switched over from gaming on my duet to gaming on my lazor ;)
<alyssa>
:-P
<macc24>
unless someone will get fex running somewhat well on armv8.0
<macc24>
gets*
Cyb3r-Assassin has joined #panfrost
Cyb3r-Assassin has quit [Remote host closed the connection]
robmur01 has quit [Ping timeout: 480 seconds]
_whitelogger has joined #panfrost
karolherbst has quit [Quit: Konversation terminated!]
karolherbst has joined #panfrost
karolherbst has quit []
karolherbst has joined #panfrost
Ekkie has joined #panfrost
Ekkie has quit [Remote host closed the connection]
camus has quit []
Tomey-M has joined #panfrost
Tomey-M has quit [Remote host closed the connection]
sreeharsha[m|gr] has joined #panfrost
sreeharsha[m|gr] has quit [Remote host closed the connection]
Blunt has joined #panfrost
Blunt has quit [autokilled: Suspected spammer. Mail support@oftc.net with questions (2021-06-04 16:25:47)]
robmur01 has joined #panfrost
<alyssa>
italove: I'm doing some Midgard debug. I fixed a scheduler/RA issue that was causing OOMs on dEQP-GLES31
<alyssa>
with a bit more elbow grease we should finally be able to do full gles3.1 CI on t860
<macc24>
:eyes:
<alyssa>
macc24: italove did a -ton- of work getting Midgard into shape this past year :)
<macc24>
alyssa: now for the hardest issue to solve
<macc24>
does firefox run on kevin at 60 frames per second on about:support page?
<alyssa>
uhhh
<macc24>
and also
<macc24>
i today fired up kevin, didn't notice anything wrong
<alyssa>
I think so
<macc24>
until i opened some supertuxkart do do testing
<macc24>
to do*
<alyssa>
if you use the OpenGL renderer in firefox
<alyssa>
(not software "Basic", also not WebRender)
<macc24>
and then i realized that i used llvmpipe as the renderer
<macc24>
it wasn't any bit slower than using mali t860 to drive display
<alyssa>
Speaking of I broke gallium_hud again didn't I
<macc24>
in regular ui
<alyssa>
oh no phew
<macc24>
alyssa: ummm
<alyssa>
macc24: with opengl about:support is scrolling by at like 45fps ish
<macc24>
is GALLIUM_HUD=fps,cpu supposed to work?
<alyssa>
uh no lies
<alyssa>
hold down an arrow key it'll scroll at 60fps
<macc24>
ups it also froze my kevin when closing glxgears
<macc24>
or sway
<alyssa>
macc24: update your mesa, you need b131f96aa83c115 for gallium_hud to work again on midgard
<macc24>
alyssa: lemme recompile it
<macc24>
damn i forgot how slow rk3399 is xD
<alyssa>
heh
<alyssa>
it's hard for me to forgot, moonlighting with the m1 :-p
* macc24
has been spoiled by sc7180
<macc24>
using m1 is cheating
<alyssa>
look the policy is I have to dogfood my own gfx drivers
<alyssa>
if I want a fast computer, well...
<macc24>
start developing freedreno >:D
<alyssa>
oh, I'm sure robclark doesn't want my help ;-)
<macc24>
tbh i didn't notice much wrong with freedreno
<macc24>
anything i threw at it ran fine
<macc24>
except firefox but it might be firefox bug
<macc24>
someone's waiting for an apitrace of that
<macc24>
alyssa: ummm
<macc24>
yea, so, on main branch gallium hud is still broken
<alyssa>
:|
<alyssa>
it.. works for me../
<macc24>
oops, was cloning master branch
<alyssa>
ah yeah that would do it
<alyssa>
made the same mistake this morning lol
<macc24>
wait
<macc24>
my scripts are written so the master branch is cloned
<alyssa>
ahh
<macc24>
that'd be a lot of laptops to compile mesa on
<macc24>
oh shit my c100pa has 80% of battery left
<macc24>
and last time i touched it was on 5.12-rc8
<macc24>
that piece of shit laptop is a tank
<macc24>
i love it
<alyssa>
thonk
<alyssa>
and fixed an annoying regression
<macc24>
\o/
<alyssa>
I /think/ it's finally stable enough for CI but more testing needed
<alyssa>
Still some fails of course but nothing so.. er... dramatic
<macc24>
what a coincidence
<macc24>
i'm planning on actually caring about kevin support this week in cadmium
<alyssa>
Still lots of tests to run but it's looking like right around 99% pass rate
<alyssa>
admittedly I'd feel a lot more comfortable if it were 99.9%
<macc24>
now how does this translate to "real" world?
<alyssa>
hard to say
<macc24>
will panfrost shit itself when i launch a program slightly wrong?
<alyssa>
errr
<alyssa>
for context dEQP-GLES31 on mali g52 is 99.99% pass
<alyssa>
with the last 2 fails being RA bugs I need to sort out but have been procrastinating on since I still am considering SSA-based RA for bifrost
<macc24>
with or without OsO scheduler?
<macc24>
OoO
<alyssa>
i think without but that won't land if there are deqp regressions
<macc24>
yep gallium hud is working
<alyssa>
ok maybe more like 98%
<macc24>
did you mean 0.05%?
<alyssa>
dEQP-GLES31 on Mali T860 i mean
<alyssa>
though some of these are probably "easy" bugs since nobody has done full dEQP-GLES31 runs on here before
karolherbst_ has joined #panfrost
<alyssa>
Ok, with my current branch it's at 98.77% pass rate.
<macc24>
\o/
<alyssa>
Test case 'dEQP-GLES31.functional.state_query.integer.max_compute_shared_memory_size_getfloat'.. Fail (Expected greater or equal to 16384, got -9.68205e+07)
<alyssa>
I have so many questions
karolherbst is now known as Guest881
karolherbst_ is now known as karolherbst
* macc24
smells int -> float conversion
Guest881 has quit [Ping timeout: 480 seconds]
<alyssa>
why does supertuxkart performance suck on panfrost so much
<alyssa>
kevin should be able to do more than 3fps on this track c'mon
<macc24>
alyssa: turn off advanced rendering pipeline or something
<alyssa>
Up to 40fps but that's still kinda low for the hw specs :|
<alyssa>
uhm, one thing at a time. back to deqp fails
<macc24>
alyssa: if you teach me how to perf i can procrastinate benchmarking everything around me
<alyssa>
i don't know how to perf
<alyssa>
zmike tried to teach me, though
<macc24>
zmike: how to perf
<alyssa>
ok, silly issue (affecting bifrost too I was just running deqp wrong) fixed, fails should be down 208->199
<alyssa>
ARB_framebuffer_no_attachment is totally broken on midgard
<alyssa>
seeing as that's done entirely as common code it's probably a 1 line bug
<alyssa>
tbh not sure why that's not failing on bifrost too, this is a common bug
pastly-antispam has quit [Quit: time for a tune up]