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
<alyssa>
I must say I am getting pretty confused with this Valhall not working thing
<bifrostest>
the huawei mate 10 pro runs kirin 970 also, so maybe linux can be installed on it
<bifrostest>
and i could get one for 40 euro
<bifrostest>
nevermind it doesn't benchmark better than g52 mp6
cphealy has quit [Ping timeout: 480 seconds]
macc24 has joined #panfrost
chewitt has joined #panfrost
nlhowell has quit [Ping timeout: 480 seconds]
vstehle has joined #panfrost
tpefreedom has quit [Ping timeout: 480 seconds]
camus has joined #panfrost
camus1 has quit [Read error: Connection reset by peer]
indy has quit [Ping timeout: 480 seconds]
camus1 has joined #panfrost
camus has quit [Remote host closed the connection]
rasterman has joined #panfrost
camus has joined #panfrost
camus1 has quit [Ping timeout: 480 seconds]
<robmur01>
FWIW if a Hikey doesn't outperform just about *any* G52, that's probably indicative of bad benchmarking or issues with other bits of the software stack (DVFS, memory tuning, etc.) - the amount of processing resources in a G52 MP6 config should be roughly comparable to a G72 MP4
* robmur01
still has a Hikey960 knocking around here, but still hasn't felt masochistic enough to revisit the massive headache of trying to update it...
<robmur01>
also it's an early one that doesn't even have a heatsink so it's only really useful for testing thermal throttling :D
camus1 has joined #panfrost
camus has quit [Read error: Connection reset by peer]
nlhowell has joined #panfrost
nlhowell is now known as Guest7143
nlhowell has joined #panfrost
Guest7143 has quit [Ping timeout: 480 seconds]
cphealy has joined #panfrost
indy has joined #panfrost
<bifrostest>
kirin 970 only seems to have kernel 4.9, so no panfrost afaik
<robmur01>
sadly there's doesn't seem to be any more recent movement, so those patches would probably take a bit of effort now to forward-port to newer than middle-of-last-year mainline :(
<CounterPillow>
I guess I should probably just file an issue, unless I'm doing something horribly wrong here
<alyssa>
CounterPillow: which gpu?
<alyssa>
I just clicked the shadertoy link (in firefox) and it renders ok on my t860 laptop with an admittedly oldmesa
<CounterPillow>
RK3566 Mali G52r1
<alyssa>
(and.. admittedly slow)
<alyssa>
ack
<CounterPillow>
I'll file an issue with all the deets and the patch
soreau has quit [Remote host closed the connection]
soreau has joined #panfrost
<robclark>
CounterPillow, alyssa: fwiw, that shader doesn't appear to work w/ freedreno/a618 (with patched kmscube, that is.. in chrome(ium) it is fine).. so I guess I wouldn't rule out a kmscube bug (possibly shader is using unsupported input uniform or something like that?)
<CounterPillow>
Works on amdgpu though, so :shrug:
<robclark>
same version of mesa?
<CounterPillow>
Yep
<CounterPillow>
I would try it in firefox on the board but debian seemingly only ships firefox-esr, and it is very unhappy when run and stops updating page content after loading a page
<CounterPillow>
(I'll blame that one on a combination of ancient firefox and wayland though)
<robclark>
disabling fp16 lowering seems to "fix" it.. (although it works fine in CrOS in browser where we are using fp16/int16 lowering)
<robclark>
possibly amdgpu is not using mediump lowering.. not sure offhand
<alyssa>
robclark: we do fp16 more aggressively on bifrost than midgard so might be related
<alyssa>
CounterPillow: Try setting `PAN_MESA_DEBUG=nofp16` env var
<robclark>
well, failing at figuring out what the resulting shader is on real web page.. but pretty sure issue is that somewhere in that shader, fp16 isn't enough precision.. rather than this being an actual driver bug
<CounterPillow>
alyssa: !!!! that works
<alyssa>
so definitely a precision issue..
<CounterPillow>
So I guess I should close the bug report, since it's a problem with the shader expecting more than 16 bits of precision out of fp16
<robclark>
changing shadertoy_fs_tmpl to default to highp also "fixes" it.. the question is, what does the shader look like when you run it in browser (on CrOS where chrome is using gles)..
<robclark>
could well be a kmscube bug, if the "real" shadertoy uses highp
<alyssa>
robclark: I don't see any precision lines on the "real" shadertoy
<robclark>
right.. but the actual shader is the thing you can edit inserted into some glue
<robclark>
but possibly that differs from what the web site uses
<alyssa>
right..
<bifrostest>
PAN_MESA_DEBUG=nofp16 PPSSPPSDL still yield blend error.
<kisak>
ppssppss ... the cat call emulator?
<kisak>
(sorry, offtopic)
<bifrostest>
i have given the info to ppsspp devs also
<alyssa>
it's a panfrost bug, not a ppsspp one
<alyssa>
not their fault
<bifrostest>
they now have confirmation it's not their fault, fwiw
<alyssa>
anarsoul: "all the RTs must be in the same buffer"
<alyssa>
ever heard of... mali t6xx? :-p
<alyssa>
the fact they advertise es3 on that thing is an abomination
<alyssa>
wait till you hear what they did...
<anarsoul>
extra copy?
<alyssa>
nope
<alyssa>
...kernel level MMU trickery to alias multiple buffers into contigous gpu virtual memory
<anarsoul>
ouch
<alyssa>
I don't know details and don't want to
<alyssa>
but it's all in kbase (the gpl kernel downstream)
<anarsoul>
I'm not going to implement MRT in lima anyway
<anarsoul>
other than using it for MSAA
<alyssa>
i'm not entirely sure how this relates to MSAA
<anarsoul>
alyssa: directly :)
<alyssa>
a single fragment shader invocation only decides the colour of a single sample
<anarsoul>
you need to enable mrt to store each sample
<alyssa>
per-sample shading requires (in the worst case) N times the number of invocaiton
<anarsoul>
alyssa: basically I've never seen blob not using the same reg number for all 4 MRTs
<alyssa>
OK
<anarsoul>
but it enables MRT for MSAA if it doesn't need to use in-place resolve
<anarsoul>
also wallpapering for multisample buffers is a pain...
<alyssa>
yep.
<anarsoul>
Utgard PP ISA knows nothing about MRT
<anarsoul>
at least it seems so - blob just uses each RT as an individual texture and does wallpapering 4 times with different sample_mask
camus1 has quit [Remote host closed the connection]
camus has joined #panfrost
nlhowell has quit [Ping timeout: 480 seconds]
nlhowell has joined #panfrost
rcf has quit [Quit: WeeChat 3.2.1]
rcf has joined #panfrost
<spawacz>
Hello, i've plugged my rockpro64 to TV but the display is cropped a bit from the sides. May that be panfrost issue? I dont know how to diagnose it
<spawacz>
with normal pc monitor evrything is fine
<daniels>
does your TV have overscan control? this can sometimes be worded as selecting 'PC mode' or similar
<spawacz>
oof never did anything more that switching channels ;p But im going to give it a shot
<alyssa>
wouldn't be panfrost
<spawacz>
is it possible that the TV got some nonstandard resolution?
<alyssa>
yes but this isn't a panfrost issue
<daniels>
right, the display is handled by an SoC-specific driver, e.g. Rockchip/Amlogic/MediaTek
<HdkR>
Jeez I really loathe overscan. Should have went away with the move to digital inputs :|
<spawacz>
can i ask xrandr to leave some blank space around the edges?
<HdkR>
According to Google search results, potentially
<anarsoul>
lima uses primitive type 15 which is rectangle (but not quad) for wallpapering, but it passes only 3 vertices
<alyssa>
sounds like t could be a parallelogram to me?
<alyssa>
the fourth vertex is implied
<anarsoul>
yeah, but which 3 vertices do I need to pass? :)
<anarsoul>
any?
<macc24>
what if you pass random vertices
<spawacz>
No it's still cropped
<spawacz>
tty is also cropped
<spawacz>
forgot to mention
<HdkR>
spawacz: That's how overscan works. If you don't solve it on the television end then you'll always get cropped weirdness
<alyssa>
anarsoul: mathematically, I mean, from any 3 you can infer the 4th
<alyssa>
I have no idea what Mali does
<spawacz>
ah damn
<anarsoul>
alyssa: OK, I guess I'll need to figure that out by trial and error :)
<alyssa>
anarsoul: only 4! combinations :)
<alyssa>
*possibilities
nlhowell has quit [Ping timeout: 480 seconds]
<CounterPillow>
Woo, found an actual panfrost bifrost bug now: indefinite hang during shader compilation in inference. Currently building a more recent mesa and then reporting it.
kenzie has joined #panfrost
<spawacz>
My TV has underscan, overscan, auto under hdmi settings
<spawacz>
which one should i choose?
<HdkR>
I would hope for a none setting
<spawacz>
nvm
<spawacz>
Underscan worked :D
<alyssa>
CounterPillow: You sound quite excited to be finding panfrost bugs? :-p
<CounterPillow>
Look, I thought I ran out of stuff to tinker with, but now I've discovered a whole new source of tinkering
<alyssa>
Hehe fair
<bifrostest>
who is billionaire here. let's start a company making 10 watt gpus for m.2 slots.
Bennett has joined #panfrost
macc24 has quit [Ping timeout: 480 seconds]
tpefreedom has joined #panfrost
<alyssa>
I am not a billionaire
<alyssa>
CounterPillow: your hangy shader runs fine on t860, curiously