<anarsoul>
alyssa: does midgard have any varying location requirements? i.e. (start address + len) should not be crossing 1mb boundary?
<leinax>
Are there any known memory leaks atm using panfrost with the mali t720? I'm seeing leaks when running glmark (it's worse when running its wayland backend over the drm backend).
<icecream95>
leinax: Does reverting 4fee7b30 have any affect?
<leinax>
icecream95: I'll give it a shot
pH5 has quit [Ping timeout: 265 seconds]
megi has quit [Ping timeout: 240 seconds]
vstehle has quit [Ping timeout: 268 seconds]
NeuroScr has quit [Quit: NeuroScr]
chewitt has joined #panfrost
NeuroScr has joined #panfrost
chewitt has quit [Quit: Zzz..]
vstehle has joined #panfrost
davidlt has joined #panfrost
<tomeu>
leinax: how do the leaks manifest?
stikonas has joined #panfrost
stikonas has quit [Remote host closed the connection]
suihkulokki has quit [Read error: error:1408F119:SSL routines:ssl3_get_record:decryption failed or bad record mac]
suihkulokki has joined #panfrost
pH5 has joined #panfrost
davidlt has quit [Ping timeout: 268 seconds]
abordado has joined #panfrost
abordado has quit [Remote host closed the connection]
<tomeu>
it happens when running dEQP-GLES3.functional.samplers.single_tex_3d.diff_mag_filter, and I also get a bunch of [13877.128412] panfrost 1800000.gpu: js fault, js=0, status=INSTR_TYPE_MISMATCH, head=0x21da880, tail=0x21da880
<leinax>
tomeu: There's a portion of the slabtop output and pmap output for the process. That was from running the glmark refract benchmark overnight with the drm backend.
<leinax>
icecream95: I don't see a noticeable difference after reverting that patch.
<alyssa>
anarsoul: Not that I know of but I believe it.
<alyssa>
leinax: I know there are leaks, they are not known leaks ;)
<alyssa>
tomeu: :eyes:
<alyssa>
Oh, joy...
<alyssa>
tomeu: well, the "XXX: tag error" tells you pretty explicitly that you'll get an INSTR_TYPE_MISMATCH when running
<alyssa>
tomeu: Oddly, despite huge numbers of faults, that test passes for me.
<alyssa>
I'm not able to reproduce that segfault etc
<alyssa>
Possibly that was fixed specially in my branch
<alyssa>
Faults look like 3D texture issues which I thought you had fixed already but maybe no.
<alyssa>
Okay, we should be passing all of dEQP-GLES3.functional.texture.format.*
<alyssa>
(with a small patch)
<alyssa>
Which is pretty cromulent, since it removes a major point of variability when debugging framebuffer format tests
<alyssa>
Passed: 538/538 (100.0%)
<alyssa>
:)
<alyssa>
So why then are blend shaders so, so problematic.
<alyssa>
If only I could get R11G11B10 blending working, I might be unblocked on STK ... Ferb, I know what we're going to do today!
<anarsoul>
alyssa: ping
<anarsoul>
alyssa: out of curiosity, does midgard also do indexed draw inefficiently if index array is much smaller than vertex array?
<anarsoul>
i.e. you have index array [0, 1000, 2], do you have to invoke vertex shader 1001 times?
<alyssa>
anarsoul: I believe so
<alyssa>
(hence why min/max_index are specified, right?)
* alyssa
tries to understand where blending goes wrong
<anarsoul>
*sigh*
<alyssa>
anarsoul: I thought Gallium had a heuristic to deal with this though
<alyssa>
(unrolling the index array if it thinks that'd help?)
<anarsoul>
utgard blob does some optimizations, i.e. it does VS_DRAW_CMD(start = 0, count = 2); VS_DRAW_CMD(start = 1000, count=1) for this case
<alyssa>
Clever
<anarsoul>
but CPU caches
<alyssa>
A swizzle away keeps the doctor away~
<alyssa>
HMm
<alyssa>
Yeah, here bad swizzle on line 391
<alyssa>
It should read hr1.zzzw
<alyssa>
but it reads hr1.zzwy
<alyssa>
which means we're getting back garbage at least partially
<alyssa>
This will be just a *joy* to debug :|
<alyssa>
(It's correct in the optimized NIR, so it's somewhere in our backend compiler)
<alyssa>
It's correct in the unoptimized MIR
<alyssa>
It's correct in the optimized scheduled MIR too
<alyssa>
Wait a minute, I was misreading this whole time, this swizzle isn't the bug. Ugh.
<alyssa>
----The problem is one line up.
<alyssa>
reading r1.yyzw incorrectly
<alyssa>
(when r1.w is undefiend at this point)
<alyssa>
Wrong after RA at any rate
<alyssa>
And right before RA
<alyssa>
So RA looks to be the culprit anyhow
raster has quit [Quit: Gettin' stinky!]
buzzmarshall has quit [Remote host closed the connection]
pH5 has quit [Quit: bye]
chewitt has joined #panfrost
raster has joined #panfrost
pH5 has joined #panfrost
chewitt has quit [Quit: Zzz..]
pak0st has joined #panfrost
pak0st has quit [Remote host closed the connection]
B1oHazard has joined #panfrost
<B1oHazard>
hi, can somebody say current status of bifrost gpus in panfrost? thanks
<alyssa>
B1oHazard: Not supported at this time :-(
<alyssa>
We know enough to write a driver but there hasn't been the interest so we've been focusing on advancing Midgard
<B1oHazard>
any chance to get bifrost support it this year?
<alyssa>
Chance? Anything can happen, look where Panfrost was a year ago ;)
grw has joined #panfrost
<alyssa>
If I nop out a lot of RA cleverness, R11 blending works and STK is in slightly less terrible shape
<alyssa>
Total hack but that tells us where to focus for that. and that there are way more bugs creeping beneath the surface.
<B1oHazard>
your work for panfrost is great :) will waiting when u focus on bifrost development, thanks
<alyssa>
Thank you :)
<alyssa>
My focus has been ES3 on midgard
B1oHazard has quit [Ping timeout: 252 seconds]
leinax has quit [Remote host closed the connection]
leinax has joined #panfrost
NeuroScr has joined #panfrost
adjtm_ has joined #panfrost
rhyskidd has quit [Ping timeout: 260 seconds]
nerdboy has quit [Ping timeout: 265 seconds]
adjtm has quit [Ping timeout: 240 seconds]
rhyskidd has joined #panfrost
afaerber has quit [Ping timeout: 250 seconds]
afaerber has joined #panfrost
afaerber has quit [Quit: Leaving]
afaerber has joined #panfrost
nerdboy has joined #panfrost
NeuroScr has quit [Quit: NeuroScr]
raster has quit [Quit: Gettin' stinky!]
BenG83 has joined #panfrost
raster has joined #panfrost
BenG83 has quit [Ping timeout: 258 seconds]
fysa has joined #panfrost
davidlt has quit [Ping timeout: 260 seconds]
raster has quit [Read error: Connection reset by peer]