ChanServ changed the topic of #dri-devel to: <ajax> nothing involved with X should ever be unable to find a bar
<mattst88> displayport monitors flicker to no end when using DP1.2 (in the on-screen monitor settings), but are fine with DP1.1
co1umbarius has joined #dri-devel
<mattst88> is this... a thing people have encountered before?'
<mattst88> I definitely bought too many displayport cables in an attempt to figure out what was wrong before just messing with monitor settings
<imirkin> mattst88: i've definitely seen flicker, usually a sign of bad drivers
<imirkin> mattst88: is the computer-side hardware capable of driving DP 1.2?
<imirkin> (sometimes bad/finicky firmware on the monitor, lots of good options)
<mattst88> imirkin: I'd assume so: 01:00.0 VGA compatible controller: NVIDIA Corporation GM107GLM [Quadro M2000M] (rev a2)
<imirkin> using the definition-of-perfection nouveau driver, i'm sure?
<mattst88> no, actually have the proprietary driver loaded atm
<imirkin> phew
<mattst88> wanted to rule-out nouveau :)
<imirkin> GM107 should support DP 1.2 just fine
<imirkin> not HDMI 2.0 though
<imirkin> DP 1.2 is Kepler+, which is the generation before this one
<imirkin> in _general_ DP 1.2 MST works fine on nouveau - lots of people using it, not a lot of complaints
<imirkin> could be the monitor being annoying
<imirkin> is it an early-firmware dell by any chance?
<imirkin> those are well-known to have DP 1.2 problems
ybogdano has quit [Ping timeout: 480 seconds]
<LaserEyess> z/35
<LaserEyess> fduck
<LaserEyess> ignore me
rasterman has quit [Quit: Gettin' stinky!]
gruetzkopf has quit [Read error: Connection reset by peer]
<mattst88> imirkin: no, lenovo P27H-20
alatiera5 has joined #dri-devel
<imirkin> mattst88: not familiar with that one ... any USB-C fun going on that you forgot to mention?
<mattst88> imirkin: nope, just displayport (and a displayport KVM, which is likely somehow partially responsible)
<imirkin> yes, that tends to be related
<imirkin> do things work fine if you connect directly?
<imirkin> (stupid question ... does the KVM advertise DP 1.2 support?)
<mattst88> yeah, it does advertise 1.2 support -- it's this one: https://smile.amazon.com/dp/B07Y5YBPXT
<mattst88> I don't recall whether it works without the KVM in the mix. I think I tested that months ago, but I've just forgotten at this point
alatiera has quit [Ping timeout: 480 seconds]
<imirkin> mattst88: based on the questions in the amazon page, the thing does sound a bit dodgy
<imirkin> mattst88: "does it work with 4k monitors" -> "nope", despite the marketing materials explicitly saying yes?
<imirkin> and they somehow manage not to support adapters, which is another indicator of dodginess
<imirkin> "we do not support docking stations"
<imirkin> you name it, they don't support it :)
shashank1202_ has joined #dri-devel
Lucretia has quit []
fcarrijo has joined #dri-devel
fcarrijo has quit []
nchery has quit [Quit: Leaving]
camus1 has joined #dri-devel
<mattst88> imirkin: yeah :|
<mattst88> seems to work fine after setting the monitors to DP1.1, so *shrug*
<mattst88> KVMs are just generally awful though
camus has quit [Ping timeout: 480 seconds]
<mattst88> my best guess is that the KVM attenuates the signal somewhat, and the bandwidth requirements of DP1.2 are greater than DP1.1
<mattst88> and maybe the attenuation only sufficiently messes up the DP1.2 signal
<HdkR> I have some KVMs which are just passthrough, so you have to use some /really/ good cables
<HdkR> Otherwise flicker mess
<mattst88> yeah, that's kind of what I was thinking as well, so I got some short (1m) cables that are VESA certified. unfortunately, still flickered
<HdkR> I had to switch over to some fiber cables to make them less likely to flicker myself
Bhawan has quit []
jernej_ has quit []
jernej has joined #dri-devel
milek7 has quit [Remote host closed the connection]
milek7 has joined #dri-devel
bbrezillon has quit [Ping timeout: 480 seconds]
bbrezillon has joined #dri-devel
camus has joined #dri-devel
camus1 has quit [Read error: Connection reset by peer]
<imirkin> mattst88: DP 1.2 bandwidth requirement isn't any higher, and there's link training to ensure you use a stable freq
<imirkin> DP 1.2 *can* do higher rates, but it doesn't have to
<mareko> zmike: I think the way to go with a mega draw is to just let the driver parse tc_batch directly instead of inventing a complicated mega draw interface
boistordu has joined #dri-devel
Company has quit [Quit: Leaving]
mclasen has quit [Ping timeout: 480 seconds]
shashank1202_ has quit [Quit: Connection closed for inactivity]
camus1 has joined #dri-devel
camus has quit [Ping timeout: 480 seconds]
camus has joined #dri-devel
camus1 has quit [Read error: Connection reset by peer]
lemonzest has joined #dri-devel
Daanct12 has joined #dri-devel
gpoo has quit [Ping timeout: 480 seconds]
Danct12 has quit [Ping timeout: 480 seconds]
Danct12 has joined #dri-devel
Duke`` has joined #dri-devel
macromorgan is now known as Guest4076
macromorgan has joined #dri-devel
Guest4076 has quit [Read error: Connection reset by peer]
kenjigashu has joined #dri-devel
Daanct12 has quit [Ping timeout: 480 seconds]
kenjigashu has quit []
aravind has joined #dri-devel
kenjigashu has joined #dri-devel
kenjigashu has quit []
xlei_ has joined #dri-devel
xlei has quit [Ping timeout: 480 seconds]
camus1 has joined #dri-devel
camus has quit [Ping timeout: 480 seconds]
i-garrison has joined #dri-devel
agx has quit [Read error: Connection reset by peer]
agx has joined #dri-devel
camus1 has quit [Remote host closed the connection]
camus has joined #dri-devel
kenjigashu has joined #dri-devel
itoral has joined #dri-devel
kenjigashu has quit []
Duke`` has quit [Ping timeout: 480 seconds]
zf has quit [Read error: Connection reset by peer]
<mareko> do we have a NIR pass that groups independent texture opcodes such that there are no other instructions between them?
zf has joined #dri-devel
i-garrison has quit []
<mareko> instead of: "load; use; load; use;", the idea is to do "load; load; use; use;"
zf` has joined #dri-devel
i-garrison has joined #dri-devel
Hi-Angel has joined #dri-devel
zf has quit [Ping timeout: 480 seconds]
danvet has joined #dri-devel
unsolo_ has joined #dri-devel
zf has joined #dri-devel
unsolo has quit [Ping timeout: 480 seconds]
sdutt has quit [Ping timeout: 480 seconds]
zf` has quit [Ping timeout: 480 seconds]
unsolo_ has quit [Ping timeout: 480 seconds]
mlankhorst has joined #dri-devel
zf has quit [Remote host closed the connection]
zf has joined #dri-devel
zf has quit []
zf has joined #dri-devel
alanc has quit [Remote host closed the connection]
alanc has joined #dri-devel
camus1 has joined #dri-devel
pnowack has joined #dri-devel
camus has quit [Ping timeout: 480 seconds]
zf has quit []
zf has joined #dri-devel
unsolo has joined #dri-devel
elongbug has joined #dri-devel
rasterman has joined #dri-devel
Venemo__ is now known as Venemo
<Venemo> Just realized I wasn't authenticated...
<Venemo> mareko: there exists a NIR scheduler, but we don't use it. I think pendingchaos tried it but it wasn't too good. The LLVM and ACO schedulers should do a good job at this
fxkamd has quit [Remote host closed the connection]
jkrzyszt has joined #dri-devel
pjakobsson has joined #dri-devel
rgallaispou has joined #dri-devel
i-garrison has quit []
i-garrison has joined #dri-devel
<mareko> LLVM doesn't do that
<mareko> is there a helper that can tell whether nir_instr has no side effects? (is movable)
<mareko> I guess I can copy instr_is_invariant
dongwonk has quit [Remote host closed the connection]
camus has joined #dri-devel
camus1 has quit [Read error: Connection reset by peer]
pcercuei has joined #dri-devel
Lucretia has joined #dri-devel
zf` has joined #dri-devel
zf has quit [Ping timeout: 480 seconds]
mlankhorst has quit [Ping timeout: 480 seconds]
camus1 has joined #dri-devel
camus has quit [Ping timeout: 480 seconds]
<cwabbott> mareko: at least the ACO pre-RA scheduler will do that for you
<cwabbott> something like that is the domain of the scheduler
<cwabbott> if LLVM doesn't do that, ask one of your LLVM guys or help with the switch to ACO ;)
hansg has joined #dri-devel
mlankhorst has joined #dri-devel
camus has joined #dri-devel
mclasen has joined #dri-devel
camus1 has quit [Ping timeout: 480 seconds]
unsolo_ has joined #dri-devel
mairacanal has joined #dri-devel
unsolo has quit [Ping timeout: 480 seconds]
flacks has quit [Quit: Quitter]
mairacanal has quit []
flacks has joined #dri-devel
JohnnyonFlame has quit [Ping timeout: 480 seconds]
gpoo has joined #dri-devel
Company has joined #dri-devel
tango_ has quit [Read error: No route to host]
tango_ has joined #dri-devel
itoral has quit []
camus1 has joined #dri-devel
camus has quit [Ping timeout: 480 seconds]
thellstrom has joined #dri-devel
vivijim has joined #dri-devel
camus has joined #dri-devel
camus1 has quit [Read error: Connection reset by peer]
unsolo_ has quit [Ping timeout: 480 seconds]
sdutt has joined #dri-devel
Peste_Bubonica has joined #dri-devel
sdutt has quit []
sdutt has joined #dri-devel
Net147 has quit [Ping timeout: 480 seconds]
mairacanal has joined #dri-devel
mairacanal has quit [Remote host closed the connection]
camus1 has joined #dri-devel
camus has quit [Ping timeout: 480 seconds]
<zmike> mareko: I was actually considering that something like that might be nice for various other things, like inlining multisample resolve blits for me
f11f12 has joined #dri-devel
ppascher has quit [Quit: Gateway shutdown]
nchery has joined #dri-devel
ppascher has joined #dri-devel
camus has joined #dri-devel
camus1 has quit [Ping timeout: 480 seconds]
fxkamd has joined #dri-devel
Bhawan has joined #dri-devel
aravind has quit [Ping timeout: 480 seconds]
Lucretia has quit []
Lucretia has joined #dri-devel
Duke`` has joined #dri-devel
jkrzyszt has quit [Ping timeout: 480 seconds]
mbrost has joined #dri-devel
Bhawan39 has joined #dri-devel
hansg has quit [Remote host closed the connection]
Bhawan has quit [Ping timeout: 480 seconds]
dongwonk has joined #dri-devel
<anholt> folks with drivers in CI, want to weigh in on https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13517 ?
* zmike steps onto the scale
jkrzyszt has joined #dri-devel
X-Scale` has joined #dri-devel
mlankhorst has quit []
X-Scale has quit [Ping timeout: 480 seconds]
tobiasjakobi has joined #dri-devel
tobiasjakobi has quit [Remote host closed the connection]
jkrzyszt has quit [Ping timeout: 480 seconds]
rgallaispou has quit [Read error: Connection reset by peer]
Bhawan39 has quit []
unsolo has joined #dri-devel
<zmike> I need a GL adult
<zmike> according to https://www.khronos.org/registry/OpenGL/extensions/ARB/ARB_vertex_attrib_64bit.txt, trying to specify a GL_DOUBLE vertex attrib using a non-VertexAttribL function is undefined behavior
<imirkin> zmike: it's definitely possible to have GL_DOUBLE vertex attribs in GL 3.0
<imirkin> or rather
<imirkin> to feed GL_DOUBLE data to float vertex attribs
<zmike> I see
<imirkin> almost no hw supports this, so double -> float happens on cpu somewhere
<imirkin> i guess an argument could be made to feed it in as a plain double to the gpu and do the -> float conversion there. but you could end up with too many attribs, since it still only counts as a single slot
<zmike> but if you advertise support for a 64bit vertex attrib format in gallium, does gallium assume your hw has that support or does it just enable the conversion
<imirkin> 64-bit attribs are different than what i just said
<imirkin> that's where the double (or int64 or whatever) actually makes it all the way through to the shader unmolested
<imirkin> in practice this gets lowered to 2x32 in st/mesa
<zmike> I'm talking about 32bit shader inputs with 64bit attribs
<imirkin> ok, that has nothing to do with the ext you're quoting then
<zmike> well you started before I finished
<imirkin> hehe ok. gave you 30s :p
<imirkin> anyways, in gallium iirc that comes through as PIPE_FORMAT_R64*
<zmike> the main spec allows passing 64bit through non *L* functions according to the charts in chapter 10
<zmike> which seems to go entirely against the 64bit attrib spec
<imirkin> mmmm
<imirkin> let's have a read then
<zmike> gallium does pass 64bit attribs through, except it's usually rewritten to 32bit attribs
<imirkin> what you're saying about the main spec matches my understanding
<zmike> it's very confusing
<imirkin> you're confusing the format of the source data with the format of the attrib in the shader
<imirkin> 64-bit data + 64-bit attrib is a separate case from 64-bit data and 32-bit attrib
<imirkin> in the first case, the data and attrib becomes 2x32
<imirkin> in the second case, you get one attrib with PIPE_FORMAT_R64*
<zmike> so if you don't use a *L* function then you get the actual 64bit attrib format
<imirkin> yes, but it's a 32-bit shader var
<zmike> I see
<imirkin> so you lose the double precision
<imirkin> it's just a way for the application to let GL worry about it
dviola has quit [Quit: WeeChat 3.3]
<imirkin> where do you see that you can only feed GL_DOUBLE's into *L*?
<zmike> so for such cases, when you get a 64bit format like that, you should be reading it as 32bit (i.e., only reading those bits) ?
<imirkin> no
<imirkin> you should be doing fp64 -> fp32
<zmike> ok
<imirkin> nouveau uses the translate module to do that.
zf` is now known as zf
<imirkin> (on the cpu)
<zmike> https://www.khronos.org/registry/OpenGL/extensions/ARB/ARB_vertex_attrib_64bit.txt clearly states "If a vertex attribute is specified using the wrong attribute function, the values of the corresponding shader input are undefined."
iive has joined #dri-devel
<imirkin> sounds accurate
<zmike> thus if L is not used, GL_DOUBLE cannot be used
<imirkin> where do you see that?
<zmike> it's that whole paragraph
gouchi has joined #dri-devel
<zmike> right at the top of overview
<imirkin> look at issue #1
<imirkin> overview is non-nomrative.
<imirkin> check why the "L" variants were added in the first place.
<zmike> this makes my head hurt
<zmike> why would they say that in the spec if it's not actually true
<imirkin> i don't see where in the overview they say that doubles must go through *L*
<imirkin> but i'm only skimming
dviola has joined #dri-devel
<anholt> yeah, I'm not seeing it either.
<anholt> the overview also talks about the implicit 64-32 prior to the extension.
<imirkin> they say that they're adding *L* variants to allow specifying 64-bit *attributes*, to distinguish from e.g. 64-bit data going into 32-bit attributes
<zmike> maybe I'm just reading it wrong then
<imirkin> anyways, it helps knowing what all exists when trying to understand this stuff
<imirkin> one can easily jump to a wrong conclusion
<zmike> every time I try to read GL specs I just miss vulkan that much more
<imirkin> this is why i always hated doing piglits + core/driver impl of something -- i was just making sure it was workign the way i thought was right, which might not be the way it was intended
<imirkin> zmike: it all takes some orientation. i can never figure anything out from vk specs
camus1 has joined #dri-devel
dongwonk has quit [Remote host closed the connection]
dongwonk has joined #dri-devel
<zmike> I guess we're a matched set then
<imirkin> zmike: btw, piglit for e.g. the GL 3.0 + GL_DOUBLE attrib thing is "draw-vertices -fbo -auto"
<zmike> yep, that's what I'm looking at
<imirkin> idiotic feature, if you ask me
<zmike> doesn't seem that smart
<imirkin> probably catering to some CAD programs? dunno
camus has quit [Ping timeout: 480 seconds]
ybogdano has joined #dri-devel
<imirkin> (whenever i see a weird GL feature, i just assume some CAD program is to blame)
* zmike has yet to run a CAD program
<imirkin> (like 2-sided fill modes, edge flags, not to mention everyone's favorite -- render modes)
<ajax> i will not rest until we have all of 1.5 + ARB_imaging accelerated with compute shaders or better
<imirkin> [on the off chance you're not intimately familiar with render modes, http://docs.gl/gl3/glRenderMode ]
<imirkin> ajax: maybe finally we'll have accum buffers
<ajax> they're just a funny name for sint16, right?
ybogdano has quit [Remote host closed the connection]
<imirkin> ajax: sounds not-wrong. i know quite little about them.
<imirkin> except that there's literally 0 applications that use them
* ajax sees a path including mesa-demos/src/redbook and stops himself
<imirkin> there are piglit tests too
<imirkin> i meant real applications
<ajax> probably they exist for irix
<imirkin> yeah
<imirkin> and probably SGI hardware supported them
<imirkin> like in Indy's or O2's or whatever
<ajax> i think by the time that became a req'd render target in d3d we already had FBOs so why would you need them a) from the window system b) sized to the window
<imirkin> why would you need them at all
<ajax> dynamic range
<ajax> same reason you'd want fp16, just, cheaper
Daanct12 has joined #dri-devel
<HdkR> Does anything actually using ARB_imaging?
<ajax> again, no
<HdkR> I thought it was just the joke you get to bring up when GL is considered a good API :P
<HdkR> Compat GL is my flavour of pain and suffering
<ajax> this isn't about fixing some specific app, this is merely going for 100% completion as a point of, like, pride
dviola has quit [Quit: WeeChat 3.3]
Danct12 has quit [Ping timeout: 480 seconds]
tango_ has quit [Ping timeout: 480 seconds]
<HdkR> Feels like one of those things that would be ultra low on the priority list. Just behind GL_NV_path_rendering
<imirkin> ahaha
Daaanct12 has joined #dri-devel
<imirkin> i think "stick eye with hot poker" is higher on that list than GL_NV_path_rendering
<ajax> wow. i had managed to avoid reading NV_path_rendering somehow
Daaanct12 has quit []
<imirkin> basically logo instead of glsl ;)
Danct12 has joined #dri-devel
<milek7> SVG in GL?
Daanct12 has quit [Ping timeout: 480 seconds]
<imirkin> and PS :)
<ajax> and implicitly whatever the system font loader can load
rgallaispou has joined #dri-devel
rgallaispou has left #dri-devel [#dri-devel]
<imirkin> i mean, it's a neat concept ... render SVG's on the GPU. but perhaps the GL driver is the wrong place to implement that? :)
<jenatali> We implemented SVG in D2D for hardware-accelerated rendering back in Win8 or 8.1 I think
Hi-Angel has quit [Read error: No route to host]
xlei_ has quit [Ping timeout: 480 seconds]
ybogdano has joined #dri-devel
xlei has joined #dri-devel
<imirkin> jenatali: hopefully in the D2D library rather than the vendor-supplied portion?
<jenatali> Yep
gpoo has quit [Ping timeout: 480 seconds]
<jenatali> Implemented on top of D3D, rather than by a driver
<imirkin> right. that makes sense
<imirkin> which is what this NV_path_rendering should have been -- a small lib that makes use of GL to do things
dongwonk has quit [Ping timeout: 480 seconds]
dongwonk has joined #dri-devel
halfline[m] has joined #dri-devel
halfline is now known as enilflah
halfline[m] has quit []
halfline[m] has joined #dri-devel
halfline[m] has quit []
halfline[m] has joined #dri-devel
halfline[m] is now known as halfline
halfline has quit []
halfline has joined #dri-devel
<FLHerne> to be fair, from the app-developer side NV_path_rendering is pretty convenient
<HdkR> It would be convenient as a library as well ;)
<HdkR> Obviously Nvidia wouldn't want to open that though
tango_ has joined #dri-devel
gpoo has joined #dri-devel
<milek7> I think they said it was in driver due to performance?
<milek7> "The 8-bit memory transactions during the stencil and cover steps can often run at memory bus saturating rates. Our OpenGL driver implementation makes use of a configurable front-end processor within the GPU—not otherwise accessible to applications—to transition quickly between the stencil step and cover step and back."
<HdkR> If the performance couldn't be achieved as a library then there are now concerns about deficiencies in GL that should also have been rectified :D
X-Scale has joined #dri-devel
X-Scale` has quit [Ping timeout: 480 seconds]
FireBurn has joined #dri-devel
camus has joined #dri-devel
camus1 has quit [Ping timeout: 480 seconds]
<dcbaker> HdkR: you mean Vulkan? :D
<HdkR> dcbaker: I don't see VK_NV_path_rendering yet :P
<imirkin> don't give 'em ideas
<dcbaker> Oh, i thought you meant, implement it as a library on top of Vulkan :)
<dcbaker> Since GL is too slow
<agd5f> imirkin, IIRC there were patches for render mode, but I don't think they ever got applied.
<imirkin> agd5f: if you say so. i don't remember that. mareko talked about some ideas for doing it.
lemonzest has quit [Quit: WeeChat 3.2]
<agd5f> imirkin, in mesa there was a hw_gl_select branch prior to the migration to gitlab
<imirkin> oh, that was a while ago
<mareko> cwabbott: I'm reinventing that instruction scheduler in NIR, right now I just group all tex opcodes by moving everything else after the last tex or before the first tex in a block, but having something that handles indirections between tex would be nice
<Venemo> mareko LLVM's AMDGPU backend runs a scheduler both before and after register allocation
<Venemo> that said, there is a NIR scheduler which AFAIK isn't very good yet
<agd5f> pepp, was involved as well it looks like
<mareko> Venemo: LLVM doesn't do anything wrt grouping tex
<mareko> I kinda need it today, and somethiing like that isn't gonna happen today in the LLVM world
gawin has joined #dri-devel
<Venemo> Oh, I see
<Venemo> Well, then the NIR scheduler may be what you want
<Venemo> Or if you are looking to moving tex accross blocks, nir_opt_gcm is the hot stuff
<mareko> the idea I have is to rank tex opcodes but the number of tex indirections within a block, and then group individual ranks, and move ALUs forward
<mareko> *but=by
<mareko> I have no clue how to use the NIR scheduler
<mareko> anholt: ^^
<Venemo> I think pendingchaos has been looking into it on our side
<Venemo> Maybe he can offer some advice
JohnnyonFlame has joined #dri-devel
adjtm has quit [Quit: Leaving]
<gawin> mareko: using chance that you're online, should be this code changed to separate cases with and without alpha?
<mareko> gawin: why?
<gawin> is it ok to always pass alpha?
<mareko> why not?
<mareko> only bpp matters for perf
<gawin> maybe some garbage? the issue with textures (which is happening on r300g) is that some of them are red or transparent
sdutt has quit [Ping timeout: 480 seconds]
<gawin> I can validate the "layer" of compiler, but "layer" with switches for hardware is kinda difficult topic for me (so sorry for some trivial questions)
<jenatali> Hooray, I'm finally allowed to talk about Android stuff. Been super annoying to be flailing around with no idea what I'm doing and not being able to ask people
<FLHerne> jenatali: Is there some announcement?
<jenatali> FLHerne: Android apps on Win11 went out to Windows insiders last week
<FLHerne> ^yes
<jenatali> Yep
camus1 has joined #dri-devel
<Sachiel> finally, decent games for windows
<FLHerne> Layer all the things
<robclark> jenatali: so, do you find android cts-runner as annoying as I do? ;-)
<jenatali> robclark: Actually haven't tried it yet, but I'm sure I will
<robclark> ahh, lucky you
<FLHerne> There've been so many years of people inventing layers to run MS APIs on Linux (ndiswrapper, samba, WINE, nine, case-insensitive filesystem hacks, ...)
camus has quit [Ping timeout: 480 seconds]
<FLHerne> and now you guys suddenly want to emulate everyone else's
<HdkR> It's a fancy new world we live in
pendingchaos has quit [Ping timeout: 480 seconds]
<FLHerne> next up, can you run Android apps in your Win11 translation layer in WINE?
<FLHerne> I bet you can compile WINE for Android and make it loop
<HdkR> Integrating Android applications is definitely one way to bolster the number of "native" ARM apps running on Windows ARM :D
alyssa has joined #dri-devel
<alyssa> is this a flake or a regression?
gouchi has quit [Remote host closed the connection]
pendingchaos has joined #dri-devel
<robclark> alyssa: looks like .trace file was truncated
<robclark> so nfs issue or something like that
<robclark> I saw that happen once before
<mareko> gawin: if you want to disable alpha, that would be colormasking, which destroys perf
<alyssa> robclark: so reassign to marge?
<robclark> yeah
<robclark> and hope that whatever network issue happened has resolved itself, I guess
<alyssa> Woof. alright
<gawin> mareko: so probably this is not it, do you perhaps know tools like renderdoc but which is working with gl 2.x?
<robclark> alyssa: tbf, it is an infra flake I've only seen one, and that was maybe a month or two ago.. but would be nice if the trace runner thing detected this case better
<mareko> gawin: apitrace?
dongwonk has quit [Remote host closed the connection]
dongwonk has joined #dri-devel
<alyssa> robclark: sure
<alyssa> I don't merge enough code these days to have CI related meltdowns.
<mareko> Venemo: nir_schedule doesn't work with SSA :(
<alyssa> mareko: that could presumably be fixed
<alyssa> Though I guess backends that are willing to ingest SSA themselves are also doing their own scheduling.
<robclark> mareko: fwiw, trying to schedule things around texture fetches (and other instructions whose consumers would stall) is a thing we handle in driver backend instruction scheduler in ir3
<robclark> (albeit sometimes with better results than others.. there are some pedantic cases where register pressure can increase)
<mareko> increasing register pressure is kinda the point
<Venemo> mareko :(
<robclark> yeah..to a point.. I'd kinda like to make scheduler aware of register pressure so it can make the tradeoff about register pressure vs stalling better
<alyssa> a Hard problem
<Venemo> mareko: ACO on RadeonSI?
<pendingchaos> I think nir_schedule probably could easily
<pendingchaos> I don't know why it doesn't
<pendingchaos> (work with ssa)
<alyssa> pendingchaos: because it's only used by a single driver AFAIK and that driver uses it after lowering SSA
<Venemo> pendingchaos: if it doesn't handle SSA, how did you try it?
<pendingchaos> from-ssa -> schedule -> to-ssa
<alyssa> awful :-P
<Venemo> oh, okay
<pendingchaos> just need to make phis unmovable, I think
<mareko> block-level scheduling would be sufficient
<mareko> no phis
<anholt> ok, back and could maybe talk nir_schedule.
<anholt> alyssa: we've seen intermittent fails where traces are truncated. I checked and we have lots of free space. so my guess is that the replay's internal downloading stuff just lost its connection and needed to retry, or something
<anholt> but it doesn't generate logs that I've seen
<anholt> unfortunately, the trace replay stuff got moved to piglit, so it's hard to work on now so I haven't been motivated to see what we can do about it.
vivijim has quit [Ping timeout: 480 seconds]
<anholt> mareko: I agree that we can probably make nir_schedule work with SSA easily. my concern with the pass as I left it is that it's easy for it to make things worse. what I wanted to do to it next was the paper "A randomized heuristic approach to register allocation"
<pendingchaos> if it's used to group texture instructions and otherwise preserve the order of instructions, maybe it won't be too bad?
<anholt> pendingchaos: the other piece that's an issue is that different drivers would probably want different scheduling heuristics, like "just group texture instructions (but not too many)?")
<anholt> so there's some shared work of making the dag and making a schedule to consider using a heuristic, but then some driver specific work of what the heuristic is and how to evaluate the proposed schedule
Duke`` has quit [Ping timeout: 480 seconds]
f11f12 has quit [Remote host closed the connection]
f11f12 has joined #dri-devel
camus has joined #dri-devel
thellstrom has quit [Ping timeout: 480 seconds]
rasterman has quit [Quit: Gettin' stinky!]
camus1 has quit [Ping timeout: 480 seconds]
adjtm has joined #dri-devel
adjtm is now known as Guest4136
adjtm has joined #dri-devel
<alyssa> right now the bifrost GLES stack is too memory bound for me to pay much attention to compiler stuff :(
<alyssa> otoh maybe if I weren't bailing on instruction scheduling the memory latencies could be better hidden.
Guest4136 has quit [Ping timeout: 480 seconds]
<jekstrand> boo memory
<alyssa> jekstrand: what?
<alyssa> did I say something?
<alyssa> ...I can't quite remember 🙃
<jekstrand> Got paged out?
<alyssa> mm... paging...
danvet has quit [Ping timeout: 480 seconds]
gawin has quit [Remote host closed the connection]
sdutt has joined #dri-devel
iive has quit []
adjtm is now known as Guest4139
adjtm has joined #dri-devel
<karolherbst> I don't know if I should be proud, but here is my patch to make marge-bot to behave more like we need it in the kernel space: https://gist.github.com/karolherbst/d3589dedc76759e18cdf406dd1d71572
gpoo has quit [Ping timeout: 480 seconds]
<alyssa> karolherbst: I don't know if i should be impressed
<karolherbst> me neither
<karolherbst> but it does work!
<karolherbst> will be fun to get that stuff upstreamed :D
<karolherbst> maybe others have better ideas
<alyssa> karolherbst: why are existing trailers getting stripped..?
<karolherbst> s-o-b handling is painful though
<alyssa> ^^
<karolherbst> alyssa: ehh.. stale comment
<karolherbst> soo..
<alyssa> oh no
<karolherbst> the old trailer code just removed _all_ occurences with a key
<karolherbst> so.. if you want to add s-b-o it removed all others
<karolherbst> and just adds the new one
Guest4139 has quit [Ping timeout: 480 seconds]
<karolherbst> I just didn't remove the comment when copying the code
<karolherbst> but the code to figure out who assigned to the bot :D
sdutt has quit [Remote host closed the connection]
<alyssa> yeah uh won't marge crash if someone assigns other than you/Ben/Lyude?
<alyssa> :-p
<karolherbst> the plan is to configure that and bail if somebody else does it :p
<karolherbst> upstream is a bit picky about emails used so I thought being explicit here might make sense
<karolherbst> also.. fetching emails from accounts on gitlab requires the bot having admin access to gitlab :)
<karolherbst> _maybe_ we could use public email addresses, but then it requires those to fill that information in and make it public or somtehing
<alyssa> I mean it's kernel... emails aren't exactly private...
<karolherbst> yeah..
<alyssa> $ grep '[a-z]+@[a-z]+\.[a-z]' MAINTAINERS
<karolherbst> I need to figure out if I can access that info from the API
<alyssa> here, spam bot authors take note ^^ :-p
<karolherbst> alyssa: I mean.. you might not knowm but in the past people were convinced that writing name at provider dot com is a good way of preventing spambots to find your email and I am currently wondering if there was a discussion on the git mailing list about this as well at some point in time :D
pcercuei has quit [Quit: dodo]
boistordu has quit [Remote host closed the connection]
f11f13 has joined #dri-devel
f11f12 has quit [Ping timeout: 480 seconds]
camus1 has joined #dri-devel