ChanServ changed the topic of #lima to: Development channel for open source lima driver for ARM Mali4** GPUs - Kernel has landed in mainline, userspace driver is part of mesa - Logs at and - Contact ARM for binary driver support!
jrmuizel has joined #lima
megi has quit [Ping timeout: 240 seconds]
enunes has quit [Ping timeout: 265 seconds]
enunes has joined #lima
kaspter has joined #lima
alyssa_ has joined #lima
<alyssa_> anarsoul: random drive by comment but `texture2D(recttexture, v_coord)` can be lowered to texture2D(recttexture, rect_coord)` where `rect_coord = v_coord * scale_factor` as computed in the VS, hence varying high-precision path?
jrmuizel has quit [Remote host closed the connection]
yuq825 has joined #lima
<anarsoul> alyssa_: that's possible but difficult and can be only done when fragment and vertex shader are linked
<anarsoul> alyssa_: it's way easier to add support for it into cogl
_whitelogger has joined #lima
dddddd has quit [Remote host closed the connection]
nerdboy has joined #lima
nerdboy has joined #lima
Barada has joined #lima
<MoeIcenowy> anarsoul: so pp does fp32 on varying, but fp16 on numbers native to pp ?
<anarsoul> MoeIcenowy: not exactly
<anarsoul> MoeIcenowy: if we load varying directly (not from register) into discard register (reg 15) to use it in sampler then it's high precision
<anarsoul> if we load varying into register and then use register (through another load varying op) in sampler then it's fp16
<anarsoul> basically that's the only exception for pp
<anarsoul> for everything else it uses fp16
monstr has joined #lima
Net147_ has joined #lima
Net147 has quit [Ping timeout: 244 seconds]
* rellla triggered piglit with cubemap branch now
<plaes> o_O
_whitelogger has joined #lima
mardestan has joined #lima
yuq825 has quit [Quit: Leaving.]
<mardestan> with plaes it is easy, he fights for estonia and everything is forgatten, if you do not want to fight for escaped porn heros projecting their stuff to known sportsmen than you fight for Kristjan Palusalu, cause this is what i do.
<mardestan> No fucking foregin chimp comes to tell me that i can not respect my own as well as i do not make sense
<mardestan> as you are black haired, but still white estonian , you can fight for Raimond Valgre or Daniel Mäng
<mardestan> I think we both and all estonian men have to prove what wood are we made of!
<mardestan> chipsies and portuguease do not come to tell things to us!
megi has joined #lima
<mardestan> i was the most potential guy, but along the road everything was taken from me, and i do not have much envy to estonians in contrast to you plaes
<mardestan> i love my father with all my heart, and even though i do not participate in big events like i was suppose to
<mardestan> i love magnus kirt the black haired javelin thrower
<mardestan> as well as ott tänak
<mardestan> before you betraye an estonian, like another one like me, you have to think about it, is it worth it
<mardestan> frankly most of the cases it never is, cause the ones that did it to me, were bad persons, plaes are you a bad person, do you feel like sleeping good when you betray an estonian?
<mardestan> with all that said, i'd like and love to see you doing progress, and I even sometmes advise
<mardestan> no matter how bad i look those days , I as a big as expert as i always was
<mardestan> this lunatic abortion leftover wants to show ones cluelessness again
mardestan was banned on #lima by ChanServ [mardestan!*@*]
mardestan was kicked from #lima by ChanServ [Banned: ban spam]
<Tofe> ah, great, thanks!
Net147_ has quit [Quit: Quit]
Net147 has joined #lima
dddddd has joined #lima
rellla has quit [Quit: see you later]
minecrell has quit [Read error: Connection reset by peer]
minecrell has joined #lima
alyssa_ has left #lima [#lima]
rellla has joined #lima
rellla has left #lima [#lima]
jrmuizel has joined #lima
rellla has joined #lima
jbrown has quit [Ping timeout: 246 seconds]
hoijui has joined #lima
rellla has left #lima [#lima]
jbrown has joined #lima
rellla has joined #lima
enunes has quit [Ping timeout: 240 seconds]
enunes has joined #lima
rellla has quit [Remote host closed the connection]
jrmuizel has quit [Remote host closed the connection]
rellla has joined #lima
<rellla> so here is the piglit result for the cubemaps branch:
Barada has quit [Quit: Barada]
hoijui has quit [Ping timeout: 250 seconds]
jrmuizel has joined #lima
<MoeIcenowy> anarsoul: surfaceless deqp still doesn't run here
<MoeIcenowy> anarsoul: it just acquires EGLDisplay by eglGetDisplay(NULL) ...
<MoeIcenowy> looks ridiculous
monstr has quit [Remote host closed the connection]
<anarsoul> :(
armessia has joined #lima
<rellla> seems there are some bits missing still :)
<rellla> anarsoul: btw ppir disasm is wrong/incomplete...
<anarsoul> why?
<anarsoul> is there anything it fails to disassemble?
<armessia> rellla: hi, tnx a lot for running my branch through piglit!
<armessia> Some bits seem to be missing still indeed :) Will look into it
<rellla> anarsoul: the different perspective cases are missing, when source_type is 2
<anarsoul> oh, OK
<rellla> so everything is reported as gl_FragCoord
<anarsoul> fix it? :)
<rellla> i fixed it already and can prepare a MR later.
<anarsoul> cool, thanks
<MoeIcenowy> anarsoul: BTW I think gl_FragDepth cannot be faked, right?
<anarsoul> nope
<anarsoul> hardware doesn't seem to support it
<MoeIcenowy> then OpenRA won't show Tiberium
<anarsoul> :(
<MoeIcenowy> and I don't know how to fix it
<anarsoul> ask OpenRA guys
<MoeIcenowy> but BTW we should get X to work first
<anarsoul> it does work
<anarsoul> but it's not optimal yet
<MoeIcenowy> btw I'm still thinking how to debug dead gp error 0x400000
<MoeIcenowy> apitrace doesn't work on non-X11, at least not retraceable
<anarsoul> it works on wayland
<MoeIcenowy> but I think no gbm?
<anarsoul> no idea
<anarsoul> anyway, we should extend kernel driver to capture the job that caused error and expose it through debugfs
<anarsoul> something similar to i915_error_state
<MoeIcenowy> interesting
<MoeIcenowy> it has support for waffle
<Tofe> anarsoul: I might be wrong, but it doesn't look like an assert was raised here:
<anarsoul> weird
<anarsoul> you may want to retry with updated branch that Connor pushed today
<anarsoul> there's some fixes
Elpaulo has quit [Quit: Elpaulo]
<Tofe> will do, thanks
<Tofe> nope, same thing
<Tofe> I'll try to debug that a bit further tomorrow; I'll probably be able to catch the moment when the "const" op node is created...
<armessia> rellla: tnx for the hint on the assertion failure
<armessia> Now glsl-fs-texturecube and glsl-fs-texturecube-2 pass :-)
<anarsoul> armessia: btw you also need to handle tiled case for cubemaps
<anarsoul> and mipmapping
<armessia> anarsoul: I already put should_tile to through at some point and it still seemed to work, will double check though, also with piglit
<armessia> Didn't test mipmapping yet
<anarsoul> armessia: do you align cube faces at tile boundary?
<anarsoul> (you probably need it)
<armessia> anarsoul: would have to check if the cube faces are properly aligned, I didn't have to change anything so far in lima_resource, depth handling was already present
<anarsoul> OK
<armessia> Force pushed my branch again. Did some clean-up of debugging messages and got rid of the extra ppir_op
<armessia> What piglit test set do you guys run? If I run the gpu set I have more than 26000 tests, the test set in the reports of rellla seems much smaller..
<armessia> Most of them are skipped and some get stuck
<armessia> anarsoul: tnx!
nerdboy has quit [Ping timeout: 268 seconds]
nerdboy has joined #lima
<enunes> anarsoul: so are you working on debugging the gp error with some xorg applications?
jrmuizel has quit [Remote host closed the connection]
<anarsoul> enunes: nope
<enunes> I tried some of the mesa-demos and I can easily reproduce a gp error with applications that send a too large plbu buffer
<enunes> not sure if it's the same one, "state=" is different
<enunes> it's also different than the vertex limitation as this is due to too many separate gl calls, e.g. several glDrawArrays with small buffers
<anarsoul> i.e. we have to flush job when it reaches certain size?
<anarsoul> I'm not sure if it's the same though
<enunes> yeah I did that and it fixes it
<anarsoul> could be
<enunes> I'll check how the blob does it
<enunes> if it's by inserting flushes or separating buffers
armessia has quit [Quit: Leaving]
<anarsoul> btw, while you're here, cwabbott proposed to talk about gpir compiler (you probably saw the email). How we're going to split the slot?
<anarsoul> 15 mins each?
<enunes> I think 45 includes time for q&a
<enunes> so it's less than that
<enunes> but sure it will be great if cwabbott can speak about gpir
<anarsoul> not sure if we'll get a lot of questions
<enunes> I'll share the slides and reply there
<anarsoul> might as well use the time for talk
<anarsoul> should we do small demo? :)
<anarsoul> after all q3a seems to work with gl1 renderer
<enunes> demo would probably be interesting, I can bring my pinebook, unless you plan to bring yours, which is geographically closer :)
<anarsoul> I can bring my pinebook
<anarsoul> too bad dual screen doesn't work well...
<enunes> I tried it once and it worked, but I guess that was with software renderer
<anarsoul> enunes: and technically it's not closer. I'm 4k km away from Montreal, probably the same for you
<anarsoul> just checked, it is but not by much - 4910km vs 6381km :)
<enunes> although if I understand correctly from the reports, meson has a slightly less buggy experience than allwinner now?
<anarsoul> I'm not sure
<anarsoul> meson usually has mali450
<enunes> the cursor plane stuff
<enunes> I haven't tried myself
<anarsoul> they don't have it
<anarsoul> IIRC megi has a patch to mark one of UI planes as cursor
<anarsoul> should work fine on pinebook
<anarsoul> not on pine64 though since HDMI has only 1 UI plane
<anarsoul> enunes: can you check whether your fix fixes sway?
<enunes> I wouldn't call it fix for now, but I can give it a try
<enunes> apparently doesnt fix it, I can open sway and move mouse around but then it freezes with gp error
<anarsoul> so that's something different :(
<anarsoul> I wish we had documentation on hardware
<anarsoul> it would make it so much easier
<enunes> the gp error is also different, the one I see with the plbu was just "4"
<enunes> which makes sense as it's plbu too large or something
<anarsoul> is it 0x400000?
<enunes> with sway yes
<anarsoul> btw, if you get LIMA_GP_IRQ_PLBU_OUT_OF_MEM the fix would be to increase tile heap size
<anarsoul> 0x400000 is LIMA_GP_IRQ_PTR_ARRAY_OUT_OF_BOUNDS and I have no idea what it means
<anarsoul> I wonder what array it refers to
<enunes> I think for plbu we allocate a dynamic buffer?
<enunes> in user space
<anarsoul> enunes: I'm pretty sure that PLBU_OUT_OF_MEM is used to signal that it ran out of heap. ARM kernel driver uses it to grow heap and resume the job
<anarsoul> and in lima we allocate 1mb for heap
<enunes> unless I misunderstood we use the dynarray stuff to assemble it and then move it to u_upload_alloc which is resizable
<anarsoul> oh, you referring to PTR_ARRAY_OUT_OF_BOUNDS
<anarsoul> then see ARM kernel driver
<anarsoul> also you can easily verify if it ran out of heap, just increase heap size in lima
<anarsoul> change gp_tile_heap_size to 2mb
<enunes> I tried it now, even at 8mb the example triggers the gp error
<enunes> maybe it's a good stress, let me try more
<anarsoul> and it's 0x4?
<anarsoul> that's weird
<enunes> yeah at 64mb I don't trigger it anymore
<anarsoul> that's a lot
<anarsoul> guess we can put this test aside for now?
<anarsoul> since we have no way to predict plbu heap size
<enunes> sure, I was just running those demos to see if something interesting shows up
<anarsoul> unless you're interested in implementing heap bo in kernel driver :)
<anarsoul> IIRC panfrost folks already have it
<anarsoul> probably not merged yet
<enunes> I think the 0x400000 error is more interesting to track down first
<anarsoul> definitely
<anarsoul> it's a show stopper
<enunes> I just wish we had a simpler reproducer for it
<anarsoul> sway is pretty simple
<anarsoul> IIRC it doesn't use any fancy shaders
<enunes> indeed, I got an apitrace with a few frames from it and I can reproduce it
<anarsoul> I wonder if the fix will be another 1-liner :)
jrmuizel has joined #lima