kaspter has quit [Remote host closed the connection]
kaspter has joined #panfrost
camus has joined #panfrost
stikonas_ has quit [Remote host closed the connection]
kaspter has quit [Ping timeout: 268 seconds]
camus is now known as kaspter
vstehle has quit [Ping timeout: 240 seconds]
nerdboy has joined #panfrost
wens has joined #panfrost
nerdboy has quit [Ping timeout: 252 seconds]
megi has quit [Ping timeout: 240 seconds]
kaspter has quit [*.net *.split]
icecream95 has quit [*.net *.split]
adjtm has quit [*.net *.split]
rcf has quit [*.net *.split]
tomeu has quit [*.net *.split]
robmur01 has quit [*.net *.split]
JuJu has quit [*.net *.split]
cphealy has quit [*.net *.split]
shadeslayer has quit [*.net *.split]
lfrb has quit [*.net *.split]
jailbox has quit [*.net *.split]
Depau has quit [*.net *.split]
HdkR has quit [*.net *.split]
rcf has joined #panfrost
nerdboy has joined #panfrost
davidlt has joined #panfrost
vstehle has joined #panfrost
nerdboy has quit [Ping timeout: 250 seconds]
yann has quit [Ping timeout: 246 seconds]
guillaume_g has joined #panfrost
nerdboy has joined #panfrost
yann has joined #panfrost
kaspter has joined #panfrost
Xalius|PBP has quit [Remote host closed the connection]
warpme_ has joined #panfrost
kaspter has quit [Ping timeout: 246 seconds]
kaspter has joined #panfrost
raster has joined #panfrost
paulk-leonov has quit [Ping timeout: 265 seconds]
paulk-leonov has joined #panfrost
kaspter has quit [Ping timeout: 246 seconds]
kaspter has joined #panfrost
Depau has joined #panfrost
robmur01 has joined #panfrost
adjtm has joined #panfrost
robmur01 has quit [Remote host closed the connection]
robmur01 has joined #panfrost
camus has joined #panfrost
kaspter has quit [Ping timeout: 265 seconds]
camus is now known as kaspter
<raster>
fascinating... with linux-next and some patches to stop some kernel hangs... nd performance has dropped by like ~20%
<raster>
i was getting like 850-880 fps now down to about ~670
<raster>
(glmark2)
kaspter has quit [Ping timeout: 268 seconds]
kaspter has joined #panfrost
megi has joined #panfrost
tomeu has joined #panfrost
vstehle has quit [Ping timeout: 268 seconds]
vstehle has joined #panfrost
kaspter has quit [Remote host closed the connection]
kaspter has joined #panfrost
Elpaulo has quit [Quit: Elpaulo]
guillaume_g has quit [Quit: Konversation terminated!]
Elpaulo has joined #panfrost
camus has joined #panfrost
kaspter has quit [Ping timeout: 246 seconds]
camus is now known as kaspter
davidlt has quit [Ping timeout: 268 seconds]
robertfoss has quit [Ping timeout: 246 seconds]
robertfoss has joined #panfrost
guillaume_g has joined #panfrost
stikonas has joined #panfrost
<tomeu>
alyssa: bbrezillon submitted this change to CI by mistake, and I'm surprised that the rk3288 is much slower with it, but none of rk3399, h6 nor s912 are
<daniels>
raster: which kernel version did you come from?
<tomeu>
bbrezillon: robher robmur01: any ideas on when could this happen? 2019-12-16T14:40:19 [ 76.402612] panfrost ffa30000.gpu: gpu sched timeout, js=0, config=0x3306, status=0x8, head=0x24bb200, tail=0x24bb200, sched_job=7c2625a8
<tomeu>
2019-12-16T14:40:19 [ 76.451669] panfrost ffa30000.gpu: AS_ACTIVE bit stuck
<robher>
tomeu: That seems bad...
<tomeu>
we seem to still have something quite wrong with respect to error handling
<tomeu>
when we reset the gpu, etc
<tomeu>
I may be still missing some patches, not sure
<robher>
tomeu: was this after a reset?
<robher>
or a reset didn't fix it?
<tomeu>
robher: well, gpu sched timeout caused a reset, and then we got the AS_ACTIVE bit stuck
<tomeu>
wonder if we got a page fault while the gpu was being reset
<robher>
tomeu: it wouldn't surprise me if the the gpu can hang (either itself or a bus) and it's not recoverable.
<robher>
Simply doing a memory access to a location that doesn't generate a response is enough from what I remember on AXI buses. And it's all async, so you might not hang until some number of transactions later.
<tomeu>
hmm
<tomeu>
I think what happens is that we run a test that is a know failure and that causes a gpu reset that leaves things in a state bad enough to disrupt other tests after
<tomeu>
but things eventually recover
<tomeu>
it's just that a bunch of tests that normally pass started failing
<bbrezillon>
tomeu: memset of the entire buf
<bbrezillon>
but I don't know why it doesn't impact rk3399, h6 and s912
<tomeu>
bbrezillon: you mean that the AS_ACTIVE bit stuck could be caused by the memset?
<tomeu>
that was a separate run without the memset
<tomeu>
and on rk3399
<tomeu>
so probably totally unrelated things
<bbrezillon>
no no, I was replying to the first question => why it's slower on rk3288
<bbrezillon>
don't know what's the problem with the AS_ACTIVE bit
<tomeu>
ah, ok
guillaume_g has quit [Quit: Konversation terminated!]
guillaume_g has joined #panfrost
<alyssa>
urjaman: "we report the actual device in glxinfo" Yeah, I landed that on Friday since I figured it'd be helpful :)
<alyssa>
icecream95: Sounds like snafu to me..
<alyssa>
Firefox WebGL is very low performance on Panfrost right now, since it has a questionable path for copying the 3D results onto the 2D canvas (glReadPixels, which forces flushes and software detiling which is extra slow)
<alyssa>
Chromium has a faster path for that but Chromium has other issues right now, hence why it's disabled
<alyssa>
Xalius|PBP: I have anecdotally heard that the RK RGA isn't suitable for general composinting for X, hence why no such stack exists
<alyssa>
I use glamor myself (on my personal laptop; my dev laptop runs whatever I need at the moment but usually Weston) but YMMV
<alyssa>
tomeu: I mean, that memset is all software ... rk3288 is a 32-bit CPU, the others in CI are 64-bit so maybe have a faster path for that?
<alyssa>
Other than that I got nothing
<tomeu>
oh, right, good point
<raster>
daniels: sorry - was out. linux-next (so 5.5 rc1)
stikonas has quit [Remote host closed the connection]
stikonas has joined #panfrost
raster has quit [Quit: Gettin' stinky!]
yann has quit [Ping timeout: 252 seconds]
raster has joined #panfrost
jailbox has joined #panfrost
lfrb has joined #panfrost
raster has quit [Read error: Connection reset by peer]
davidlt has joined #panfrost
yann has joined #panfrost
megi has quit [Ping timeout: 276 seconds]
megi has joined #panfrost
yann has quit [Ping timeout: 245 seconds]
yann has joined #panfrost
* alyssa
poking at shader-db
<alyssa>
which has quickly led me to discover we expose shadow samplers
<alyssa>
So I guess I should bang out some support for that.
<alyssa>
Nothing to RE, just some code to hookup
<alyssa>
Set the `shadow` flag on the texture and route through the shadow coordinate and that should be it, in theory
<alyssa>
I think there are deqp-gles3 tests for this
<alyssa>
Yeah, here. Let's go.
<alyssa>
Just trying to figure out how to RA this well.
<Lyude>
while I'm not working on anything panfrost related atm, it is very nice to keep seeing panfrost all over in igt's source :)
davidlt has quit [Ping timeout: 248 seconds]
icecream95 has joined #panfrost
<alyssa>
Lyude: \o/
guillaume_g has quit [Quit: Konversation terminated!]
<alyssa>
master with shader-db's GLSL 1.20 collection takes ~34s
<alyssa>
but if I change the uniform/work threshold, 99.95% of spilling disappears and it takes ~17s instead, half that
<alyssa>
In the latter case, about 24% of our time is spent computing liveness info
<alyssa>
6% solving lcra equations
<alyssa>
wait, no, 1.5% solving lcra
<alyssa>
20% of our time validating NIR shaders (this is gone in release builds)
<alyssa>
15% of our time adding lcra interference constraints, though -- this is a tad surprising? but I can look into it
<alyssa>
6% scheduling
<alyssa>
lcra_add_node_intererence and mir_bytemask_of_read_components are surprisingly bottlenecks
jailbox has quit [Ping timeout: 276 seconds]
stikonas has quit [Remote host closed the connection]