<anarsoul>
narmstrong: hm, isn't it supposed not to start a job if board is in use?
<narmstrong>
anarsoul: no, the runner is not tied to lava, it simply submits the job and waits
<narmstrong>
which is quite wrong
<narmstrong>
I'd like to use docker-machine to run a docker instance on a board when it's ready
<narmstrong>
but it needs quite some work
<anarsoul>
OK, so what do I do for now? :)
<narmstrong>
wait and pray :-p
yann|work has quit [Ping timeout: 268 seconds]
<icecream95>
narmstrong: ERROR: Job failed: execution took longer than 1h0m0s seconds
warpme_ has joined #panfrost
<bbrezillon>
tomeu, robher, robmur01: would any of you be able to take over patch 7 and 8 of my 5.4-fixes series (or propose something else to fix the problem)?
<bbrezillon>
I don't think I'll have time to work on a v2, unfortunately
<bbrezillon>
(at least not before beginning of next year)
yann|work has joined #panfrost
icecream95 has quit [Ping timeout: 265 seconds]
pH5 has joined #panfrost
<narmstrong>
device must be stuck
<narmstrong>
hmm, the board is Idle
<narmstrong>
anarsoul: your jobs never started :-/
_whitelogger has joined #panfrost
anarsoul|2 has joined #panfrost
anarsoul has quit [Ping timeout: 268 seconds]
EmilKarlson has quit [Quit: killed]
TheCycoONE1 has quit [Quit: killed]
thefloweringash has quit [Quit: killed]
flacks has quit [Quit: killed]
eballetbo[m] has quit [Quit: killed]
rhyskidd has joined #panfrost
abordado_ has joined #panfrost
megi has joined #panfrost
abordado_ has quit [Quit: Leaving]
abordado has joined #panfrost
thefloweringash has joined #panfrost
EmilKarlson has joined #panfrost
flacks has joined #panfrost
eballetbo[m] has joined #panfrost
TheCycoONE1 has joined #panfrost
<alyssa>
I uh need to get a trace of register spilling on SFBD..
<HdkR>
:>
<alyssa>
I should also uh stop lowering tmp arrays now that I understand how the stack works ...
abordado has quit [Read error: Connection reset by peer]
abordado_ has joined #panfrost
* alyssa
trying to figure out when the right time to upload an FBD is ..
tgall_foo has joined #panfrost
abordado__ has joined #panfrost
abordado_ has quit [Ping timeout: 250 seconds]
tlwoerner has quit [Quit: Leaving]
tlwoerner has joined #panfrost
<alyssa>
Okay, I have a branch that correctly calculcates/sizes everything
<alyssa>
Let's see what CI thinks
<HdkR>
TLS spilling woo?
<alyssa>
Weeeee.
<HdkR>
I'm over here adding phi nodes to my IR and improving RA to support that :D
yann|work has quit [Ping timeout: 250 seconds]
pH5 has quit [Quit: bye]
unoccupied has joined #panfrost
unoccupied has quit [Client Quit]
guillaume_g has quit [Quit: Konversation terminated!]
<alyssa>
Hm
<alyssa>
So the branch passes CI on T760/T820, but fails on T720/T860
<alyssa>
What an .. odd combination
<alyssa>
Worse, I can't repro the T860 fail
jschwart has joined #panfrost
nerdboy has joined #panfrost
* alyssa
spills
raster has quit [Quit: Gettin' stinky!]
<alyssa>
Hmm.
<alyssa>
Oh, interesting. The spilling in Neverball disappears if I promote 8 uniforms instead of all 16
<alyssa>
The extra bit of register file is enough to eliminate hundreds of ld/st instruction
<alyssa>
So that heuristic needs to be adjusted for sure.
<alyssa>
Looks like glamor's shader might be the same way.
<alyssa>
Alternatively I could "demote" uniforms but that would require a lot more roundtrips and stuff then I'd want
<jschwart>
icecream95: I hope you still find this in the logs or it's helpful in general, but glmark2 works, I have fancy 3D objects spinning on my screen in turn now
<jschwart>
so only X.org is having problems for me
<jschwart>
maybe it is possible to run SDL2 based programs without X11, then I could play some more :)
<alyssa>
In other news, with the spilling fixes branch, fishgl now works..
<alyssa>
....at 2fps :|
abordado has joined #panfrost
<jschwart>
alyssa: I'll build your rev branch now & try with that
<jschwart>
thanks!
abordado__ has quit [Ping timeout: 246 seconds]
<alyssa>
it should be merged Soon
<alyssa>
With the uniform register threshold changed, the spilling in fishgl goes away and we go up to.. 3fps :|
Stary has joined #panfrost
abordado has quit [Ping timeout: 268 seconds]
<jschwart>
alyssa: it's working with your branch !!!
<jschwart>
glxgears running great now :)
<anarsoul|2>
alyssa: up to 3 fps from how many fpses?
<steev>
0
nerdboy has quit [Ping timeout: 268 seconds]
icecream95 has joined #panfrost
<jschwart>
ri-li works great
<jschwart>
supertux too but is kind of slow, especially the overworld
<jschwart>
supertuxkart says OpenGL 3.1 is needed and says my OpenGL version is too old
davidlt has quit [Ping timeout: 240 seconds]
<icecream95>
jschwart: Supertuxkart does work without OpenGL 3.1, just some of the fancy effects don't work
<HdkR>
Does supertuxkart not support runtime selection of ES or GL?
<icecream95>
It automatically uses ES2 (or maybe just desktop GL 2) if it can't create a GL3 context
<HdkR>
I presume that's auto dropping to GL 2
<HdkR>
Not many things support runtime selection between GL and ES
<icecream95>
jschwart: Supertux probably isn't using GL.
<alyssa>
jschwart: Oh, nice :)
stikonas has joined #panfrost
<alyssa>
anarsoul|2: steev is correct.
<icecream95>
HdkR: Supertuxkart prints "OpenGL version string: OpenGL ES 2.0" on startup
<alyssa>
The desktop GL 2.1 renderer would probably be faster but :shrug:
<HdkR>
ah, neat
<alyssa>
what distro?
<icecream95>
I compiled stk from source
<alyssa>
Oh
<alyssa>
FWIW, the desktop GL 2.1 and the embedded GLES 2.0 renderers are slightly different
<alyssa>
On panfrost at least, the desktop one will be faster since the desktop->embedded lowering in Gallium is more efficient than the lowering STK does for GLES2
<alyssa>
IIRC, maybe that's been changed since
<icecream95>
STK also prints "OpenGL version: 2.1", so I'm not sure what it's actually using...
<alyssa>
Oops
<icecream95>
Is there a way to tell whether something is using normal GL or ES?
<icecream95>
I tried apitrace and it gave "error: unavailable function glXGetProcAddressARB"
<HdkR>
Didn't build glx libraries, or is this potentially apitrace not understanding an environment without glx?
<alyssa>
icecream95: I just pushed the cleaned up spilling series
<icecream95>
apitrace works with STK if I pass it `-a egl`
<icecream95>
You do need some slightly edited shaders to get it to work
<HdkR>
Why would anyone use gl4es when the driver supports desktop GL :/
<icecream95>
Otherwise it fails with `Unexpected exception: GLVideoSystem: GlewError: Missing GL version`
<icecream95>
HdkR: Because it works a lot of the time. :)
<alyssa>
So does native desktop GL ...
<HdkR>
So does MESA_GL_VERSION_OVERRIDE :P
<icecream95>
With GL, the bottleneck for supertux is C++ RTTI
<icecream95>
MESA_GL_VERSION_OVERRIDE=3.3 works!
nerdboy has joined #panfrost
<icecream95>
Is there a MESA_GL_VERSION_OVERRIDE for GLSL? Because otherwise shader compilation fails without my manually converted shaders
<alyssa>
MESA_GLES_VERSION_OVERRIDE=3.0
<alyssa>
the right version override also does the right glsl
<HdkR>
MESA_GLSL_VERSION_OVERRIDE
<HdkR>
If you only want to override the shading language
<alyssa>
but I don't understand why you want gl4es when we have native GL?
<HdkR>
^
<alyssa>
If an app that runs with GL 2.1 doesn't work in Panfrost, please file a bug and attach an apitrace.
<icecream95>
I think at the moment there are only a few games which I still need gl4es for, but I'll file bugs next time I try them
<anarsoul|2>
alyssa: btw, do you support importing/exporting non-linear BOs in panfrost?
<alyssa>
anarsoul|2: Not yet, but I think someone at collabora has patches floating around for it
<jschwart>
I'm trying Ultrastar Deluxe now
<icecream95>
A few weeks ago (it's fixed now), there was a weird bug where I tried drawing three sprites with SDL_gpu and they ended up in the wrong place, but gl4es worked fine
<anarsoul|2>
I'm really not sure what I'm supposed to do if I get DRM_FORMAT_MOD_INVALID
<anarsoul|2>
other drivers do some guess work
<anarsoul|2>
but meh
<jschwart>
it works a lot faster now, but I think it needs the VPU driver to fully work
<alyssa>
A video game needs a VPU?
<alyssa>
Kids these days ;)
<HdkR>
Got to play back those FMVs somehow :P
<icecream95>
supertux actually runs faster with gl4es...
<jschwart>
alyssa: it's karaoke
<icecream95>
By ~15%
<jschwart>
the clips are playing in the background
<jschwart>
for songs where there's no clip or when the video quality is not so high, it works fine now it seems :)
<icecream95>
Does anyone know how to make meson do LTO in parallel (i.e pass -flto=4 to the linker instead of just -flto)? Otherwise it takes ages to link...
<alyssa>
icecream95: BTW, looking at the OpenLara trace now
<alyssa>
Looks to be yet another spilling failure :|
<alyssa>
again if I tweak the uniform threshold, the spilling goes away, and then it renders fine.
<alyssa>
I.. think
* alyssa
should probably figure out the uniform threshold thing too
<icecream95>
alyssa: I'm still compiling the first set of commits you pushed and now you push some more?
yann|work has quit [Ping timeout: 276 seconds]
<jschwart>
I'm off to bed here, I've been enjoying panfrost a lot meanwhile!
<jschwart>
thanks a lot all!!!
<jschwart>
have a nice day
yann has joined #panfrost
<icecream95>
alyssa: neverball still has some rendering issues, e.g. Neverball Easy level 15
jschwart has quit [Ping timeout: 250 seconds]
<daniels>
anarsoul|2: if you get invalid, assume linear
<anarsoul|2>
daniels: but what if it's tiled? :)
<daniels>
anarsoul|2: it shouldn't be
<anarsoul|2>
OK
<anarsoul|2>
thanks
<daniels>
anarsoul|2: i mean, if you're using GBM or Wayland or X11 then you explicitly use modifiers to allocate buffers when the relevant EGL extensions are exposed
<anarsoul|2>
(that's not what vc4 nor v3d do btw)
<daniels>
anarsoul|2: most other relevant people have facilities for passing modifiers around as well
<anarsoul|2>
daniels: yeah, but some software is stuck with older API
<daniels>
so in that case, if you're asked to explicitly allocate or import with MOD_INVALID, only ever allocate linear; then you only allocate tiled when someone explicitly requests it through the modifier API (in which case it's their problem to deal with if they later drop the modifier), or for internal allocations like FBOs (in which case they'll not get exported)
<daniels>
which software?
<anarsoul|2>
unity from ubports
<daniels>
anarsoul|2: that's a shame
<anarsoul|2>
they'll fix it
anarsoul|2 is now known as anarsoul
<anarsoul>
btw they have the same issue with etnaviv
<daniels>
and they'll be slower on intel
<alyssa>
icecream95: It's funny that you think I'm good enough at Neverball to get to 15 ;)