00:00 CounterPillow: have you considered actually banning the hostname and not just the nick
00:00 karolherbst: that's pointless
00:01 karolherbst: we are talking about somebody who is willing and able to get different VPNs
00:01 CounterPillow: they'd need to re-register each time which wastes more of their time
00:02 karolherbst: they'll do exactly that
00:03 karolherbst: we are talking about an ill person here, no point in making them waste money if the outcome for us is exactly the same
02:54 balrog: I'm running into what looks like a deadlock when trying to run this software under llvmpipe for vulkan. This happens even if I reduce thre LP_NUM_THREADS to 0. How do I debug?
02:54 balrog: https://github.com/ngscopeclient/scopehal-apps
02:54 balrog: this software heavily uses compute shaders
03:01 airlied: compute shader in llvmpipe aren't threaded so I can't imagine a deadlock
03:02 airlied: does it still consume CPU?
03:05 balrog: it pegs the CPU
03:05 balrog: and the program hangs
03:05 balrog: it seems to be stuck waiting for a fence
03:06 airlied: are the threads still all going, sounds like it's still computing
03:06 balrog: is there a way to tell which thread the shader is running in? I tried in gdb and it isn't obvious
03:07 airlied: it should be running in all the helper threads
03:07 airlied: though I don't think the compute ones get good names
03:08 balrog: and that's even with LP_NUM_THREADS=0?
03:08 airlied: oh with that option it should all be on the main thread
03:08 balrog: ahhh
03:08 airlied: depending on how the application submits the work
03:08 balrog: the main thread is stuck in lvp_pipe_sync_wait_locked() in lvp_pipe_sync.c
03:08 balrog: from vk_fence
03:09 airlied: whatever thread does the gl dispatch should have it usually
03:10 balrog: it submits — and is stuck — here
03:10 balrog: https://github.com/ngscopeclient/scopehal/blob/c598d7f8b698440ed3c0f46ba06c69d9cff5ee43/scopehal/QueueManager.cpp#L121
03:11 balrog: (this does not get stuck on any hardware GPUs that I know of, and does not trip up any Vulkan validation layers)
03:11 airlied: oh it's vulkan based, so then it will be on "gpu" thread
03:11 balrog: know of / tested with*
03:11 airlied: which will exist no matter what LP_NUM_THREADS is set to
03:11 balrog: there seems to be no "gpu" thread that I can see
03:11 balrog: at least from gdb's `info thread`
03:12 airlied: yeah I don't think it has a name
03:12 balrog: hmm. I checked each thread and it wasn't obvious that way either
03:13 airlied: it will have lvp_execute_cmds in the backtrace
03:13 airlied: but it might not backtrace at all
03:13 airlied: if it's inside the shader code
03:13 balrog: there is one thread that will not backtrace
03:13 balrog: and is unnamed
03:13 airlied: that's probably the jit one
03:14 airlied: so it's likely executing code
03:14 balrog: It's possible the shader doesn't have sufficient barriers
03:14 balrog: we already ran into this problem once :/
03:16 balrog: is there a way to debug the shader?
03:16 airlied: the only thing I can think off is for some reason the compute shaders goes into an infinite loop
03:17 airlied: not really without rebuilding llvmpipe to stick printf equivalents into the shader code
03:17 airlied: or you can sometimes dump the assembly and try to line it up with the gdb backtrace
03:17 airlied: LP_DEBUG=ir,asm might help
03:18 balrog: so the previous would that require me to rebuild mesa in debug?
03:24 balrog: I'll go ahead and do that
03:25 airlied: yeah even the LP_DEBUG might need debug builds, not sure if we ever changed that
03:25 balrog: the release notes claim it was changed but the source code seems to still require it
03:27 balrog: yeah being able to see the ir or the asm probably would help
03:32 roughdestiny: rust is pretty good, i little bit looked and it turns out they do not deallocate this allocation until the end of the program, where heap gets allocated automatically but it is user responsibility , they just lack test for this, it's analogues to new and free in c++, scala is also nice language with clean syntax, but the situation it requires jvm is new for me, but that happens cause signal collect infers or uses some java
03:32 roughdestiny: types. So we come to the spot where issues start to get closed, central europe has what i suggested, they always have everything, they are in possession of latest nuclear science too, those people are crazy smart scientists. In other words subset is already named so, it's possible to smash more sets into a returned var, in their case their hold it alive a bit longer. They do everything that third world war would not happen,
03:32 roughdestiny: but this time rus started to get back in business, much emphasis was put on the transgender tolerance politics to castrate germans out. in those location in the central azis it's like in estonia very smart people live there, our local ones are also lot more skilled and smarter than me in many areas.
03:34 balrog: uh spammer?
03:36 DemiMarie: Does virtio-GPU’s Wayland support require Wayland support, or can I use it without it?
03:37 DemiMarie: Sorry, I meant GPU acceleration 😆
03:37 DemiMarie: Also, how much more work would be needed to support Xen?
03:38 balrog: > When run inside Linux perf, LLVMpipe will create a /tmp/perf-XXXXX.map file with symbol address table
03:38 balrog: hmm... how do I make it do this?
03:42 airlied: it doesn't work very well unfortunately
03:43 airlied: there was a time llvm did debuginfo gen automagically, that time went away
03:44 balrog: is it what's discussed here? https://llvm.org/docs/DebuggingJITedCode.html
03:45 balrog: or is this a different sort of jit?
03:45 airlied: that is part of it, but only once you emit all the debug info
03:48 airlied: https://gitlab.freedesktop.org/airlied/mesa/-/commits/gallivm-debuginfo-hacks-wip has some hacks in the area
04:05 caramelcandy: You do not break me anyways, when half of your line up gets continuously killed off, cause of assaulting me persistently that should remind you that such retarded puppets as karolherbst should in the world hold their mouth shut, cause that is what you are next to me, a puppet, or any serious estonian sport or science star or a war hero. You are a little toxic worm where instead of brain cells you have green gas in the head.
04:06 balrog: how are these spammers "identified and verified with services"?
04:13 airlied: bmodem: it's not really a spammer, it's a person with a mental instability who just continues to create new accounts and vpns
04:14 airlied: oops wrong person
04:14 airlied: balrog: ^
04:25 DemiMarie: Could there be a requirement to talk to a person before joining? Something like, “Send a message to XYZ with why you want to join.”
04:26 airlied: not that I know off
04:27 airlied: and I'm not sure how you think that would stop them
04:27 airlied: they are perfectly capable of saying sane shit in order to get past any door staff
05:00 Company: DemiMarie: who would you want to do that job?
05:00 Company: Turing teting spam accounts sounds like a really boring task that nobody wants to do
05:02 CounterPillow: it's not a bot
05:02 CounterPillow: any hoop a human can jump through, they can jump through
05:04 Company: I said "Turing testing" becuse that's roughly what it amounts to for 99% of cases
05:04 Company: this one would actually be interesting
05:06 airlied: we have no spam problem though so it would be a waste of time
05:06 Company: besides, you want people to join, so making them do tests is very counterproductive
05:06 Company: even requiring accounts is
08:09 airlied: karolherbst: I think I just got radeonsi/llvm to do a function call
09:22 pq: DRM_MODE_TV_MODE_MONOCHROME is proposed new UAPI, FYI
09:23 pq: I'm a bit puzzled why UAPI patches are not required to say "UAPI" in their title.
09:24 pq: particularly since it's often impossible to determine whether it's UAPI or not by looking at the files modified
09:25 CounterPillow: Is Linux getting black-and-white TV support?
09:25 pq: it seems so, yes
09:25 pq: vc4 particularly
11:05 caramelcandy: it's not that i am spamming nonsense anyhow, i got so much luck first time on death penalty that the staff got me out to one hour walk, and never validated my condition when i went back to mental institution, because staff severly liked me, i was able to add 3l vodka worth of alcohol every day to the mix of 15mg abilify and 400mg seroquel in combination, all the case as to what happened was irritating from bombs to abuse in medical
11:05 caramelcandy: facilities by intention, i just had luck, but there is old saying that luck favors the stronger ones, but thinking back to it i noticed how british are so abusive and i do not feel like having had strong persons luck neither in estonia nor at overseas to escape death , i have thought it's a matter of time when they succeed, and i think russians think the same, so they organized an attack, but the point where united states gives guns or
11:05 caramelcandy: provides a lot of help to ukraine i do not believe in, ukraine has to pay for those military solutions and packages, through the lendlease or similar leasecargo company, so there is not that much of hope for the future with my background of getting tremendously abused. Not sure how russians are sided, but a lot of those scammers live happily and it could be war soon as it likely is with russian federation, which might as well still
11:06 caramelcandy: execute me. Tragic stuff overall, my life did not go well so that my engines worked, dad abused me with most others in the start and made clever manipulation to turn things against the odds. But this did not mean that i did rest alt i work every day to riddle things, however i am taking a break after this year finally, so your dreams come true. After mission accomplished i no longer be present here, even if i live.
12:51 Lynne: was it ever investigated whether ngg culling helps on gfx11+?
13:13 mareko: Lynne: yes and yes
13:13 mareko: jenatali: FYI https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27589#note_2288693
13:15 Lynne: mareko: should it get enabled on
13:15 Lynne: gfx11 by default
13:26 jenatali: mareko: I'm out of town for a week. If it can wait, I'll look when I'm back. If it can't you can add an xfail for now
13:42 mareko: Lynne: radeonsi enables it default
13:42 mareko: radv should probably too
13:56 karolherbst: airlied: nice
14:06 karolherbst: and that we have chatter about how to prevent joss from joining: there are no technical solutions for social problems. We might want to move to a chat platform where you have to authenticate via gitlab or social logins, but even that will probably not "solve" this and cause other issues, like gitlab spam protection being a pain.
14:12 balrog: [22:16:05] <airlied> the only thing I can think off is for some reason the compute shaders goes into an infinite loop
14:12 balrog: the shader has a while(true) loop with a break on ending condition
14:12 balrog: so yes, probably this
14:14 pac85: I wonder if someone could take a look at https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26393 which fixes a problem where some textures can get reallocated but their contents don't get copied. My doubt is how it interacts with the locking that https://gitlab.freedesktop.org/mesa/mesa/-/commit/8d20c660a9831c367d98ed2fea25e5276e6466f2 talks about
14:34 ondracka: Hi, anyone knows what is a test failure like this about? Error: Failed to invoke dEQP for dEQP-GLES31.info.renderer,dEQP-GLES31.info.version,dEQP-GLES31.info.extensions
14:34 ondracka: https://gitlab.freedesktop.org/mesa/mesa/-/jobs/55163462#L919
14:34 ondracka: I would understand a plain fail, but what is "Failed to invoke" about?
14:53 mareko: ondracka: this may be useful: git shortlog -ns --grep=ci.*llvmpipe
15:04 marex: hey, I just got a gitlab notification that access to drm/misc was granted to me, I do have push access to drm-misc, so is this expected or did something odd happen ?
15:18 mripard: marex: it's expected, I've been working on it this afternoon and will send everyone a mail about it by tomorrow
15:18 marex: ok
16:18 Hazematman: Hey could someone add a lavapipe tag to my MR? Also I think I'm over 25 commits merged into mesa now, is it possible to get a developer Developer role on Mesa now? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27686
16:27 daniels: airlied: I'm temporarily adding mripard as an owner for drm so he can do some repo setup
16:29 jenatali: Hazematman: Labels get added when the draft status is removed
16:34 Hazematman: jenatali: Thanks for letting me know. Are you talking about "MR Label Maker" bot, or just a general policy?
16:34 jenatali: Oh I'm just talking about the bot
16:38 HdkR: jenatali: quick question if you know. Does dzn have a weird interaction with calibrated_timestamps and arm64ec timestamp scaling?
16:38 HdkR: `TIME_DOMAIN_QUERY_PERFORMANCE_COUNTER` is a weird construct to have living in Vulkan
16:39 HdkR: (Or maybe XTA timestamp scaling, not sure where the scaling happens here
16:40 jenatali: HdkR: I'm not familiar with that construct
16:41 HdkR: Ah cool. It's something I was worrying about until I realized that time domain was Windows only
16:45 jenatali: HdkR: dzn just implements D3D12 timestamps which are based on QPC
16:46 HdkR: jenatali: Cool, so QPC probably scales internally and dzn will just pick up that scaling automatically
16:47 jenatali: I don't know what kind of scaling you're talking about
16:47 jenatali: You have to QueryPerformanceFrequency and scale it yourself as an app
16:48 HdkR: jenatali: ARM based devices have too slow of a cycle counter so XTA needs to scale it from 19.2Mhz to something more reasonable that x86 applications expect. Something closer to 1Ghz to 3Ghz range. Something that XTA does but I haven't wanted to do in FEX
16:48 jenatali: Oh that must happen in the Windows kernel then
16:49 jenatali: For calibration specifically the CPU timestamp sample happens in the kernel so it's as close as possible to the GPU timestamp
16:49 jenatali: So if that's getting scaled it's not happening in the XTA layer in usermode
16:49 HdkR: ah interesting
16:50 HdkR: GPU timestamp is the same cycle counter on Snapdragon, so that's also 19.2Mhz. So they likely want to scale the same amount
16:52 HdkR: Theoretically in GL/Vulkan there is no real way to directly compare CPU and GPU time domains
16:52 HdkR: Not sure if D3D gives that option
16:56 jenatali: HdkR: ID3D12CommandQueue::GetClockCalibration
16:57 HdkR: Hopefully vkd3d-proton is fine with timestamp scaling :BlobSweat:
17:40 abhinav__: jani vsyrjala Hi, gentle ping for reviews on https://patchwork.freedesktop.org/series/129748/ . now CI is green (IGT CI should be unrelated as this patch only adds an API, the usage is only within msm).
18:02 vsyrjala: mripard: commit 1fd4a5a36f9f ("drm/connector: Rename legacy TV property") broke i915
18:02 vsyrjala: i guess you didn't do a proper build test for that series since just after that commit it wouldn't even build afaics
18:09 mripard: vsyrjala: could you be more specific? do you have a broken defconfig?
18:10 vsyrjala: no. the tv mode property is broken now
18:12 vsyrjala: 'git reset --hard <^ mentioned commit>; make ...' should tell you why
18:33 mripard: thanks, that's helpful
18:37 mripard: 6.3 onwards compiles though
18:37 vsyrjala: yeah. bisect could hit the build breakage though. but nothing we can do about that now
18:38 mripard: I'm not sure what you want me to do then?
18:39 vsyrjala: needs to be fixed, because it's now broken. the build breakage is just helpful to show why it's broken
18:42 u-amarsh04: still bisecting my issue: https://gitlab.freedesktop.org/mesa/mesa/-/issues/10228 (8 good test builds, 8 bad test builds and 194 skipped test builds)
18:47 mripard: vsyrjala: so there's nothing we can do about it now, but it's broken and we must fix it? I'm confused
18:49 vsyrjala: nothing we can do about the bisect problem. but the actual functionality can be fixed
19:10 jenatali: mareko: Ok I got a few minutes to take a look to unblock you - I threw a diff in a comment. Feel free to commit that if it unblocks you; I don't expect it regress anything else
20:19 saneloops: I was not sure why some engineers slapped me with fast fourier transform, maybe the decimal point buses use that somehow, but me i think 1024 digits all in 1024 possible combinations that is 1024*1024*1024 so its billion samples indeed, i am unsure where those 4.2 billion values come from, but this is hw land, but i think that is why they searched me with lot's of fft links, that this connects to this theory somehow, and understanding
20:19 saneloops: that might give some benefits in the industry say per mental stability or sanity. It's simply over my head and skillset.
20:41 vsyrjala: fft gave him butterflies
21:01 saneloops: nah just not in the mood and correct timeline yet to understand , i know my theory works in higher level well, but how hardware does things was yeah never known to me, the register bus is in two configurations unsigned or signed this is known and everything adds up there with possible combinations but you know that's about it currently. Higher priority is to get my hardware to modern code paradigm, than understand how hw works. I already
21:01 saneloops: lack eyes for this, vision has suffered ontop.
23:40 Here2Learn: hello
23:42 Here2Learn: are all iGPU's UMA or is it just mostly modern ones? has there ever been a case where an iGPU isn't capable of UMA? I used to think it was all the same but then I came across this:
23:42 Here2Learn: "Shared graphics-memory is not the same as unified memory. In a shared graphics-memory system, a part of the system RAM is reserved for the GPU. The amount of RAM reported to the OS is the total RAM minus the chunk reserved the GPU uses. The OS only can use it's own part, and cannot access the GPU memory (and vice versa). If you want to make something available to the GPU, it still has to be copied to the reserved GPU part.
23:43 Here2Learn: In unified memory both the OS and the GPU can access the entire range of memory, no need for a (slow) copy. Consoles use the same strategy (using fast GDDR for both system and GPU), and it's one of the reasons consoles punch above their weight graphically." from: https://news.ycombinator.com/item?id=27182715
23:43 Here2Learn: So now, I'm not so sure anymore. I guess it isn't really the same then?
23:43 mareko: it's half true
23:44 mareko: there is a piece of RAM that's carved out and private to the iGPU, but it's small and mostly for driver-internal stuff
23:45 mareko: apps don't usually use it, they just use RAM owned by the OS if the carved-out part is too small
23:48 mareko: you can set how much RAM you want to reserve for the iGPU in the BIOS, usually you can choose between 128 MB and 1 GB, the driver will just use the rest of system memory if it needs more
23:49 mareko: it's the same as a dedicated card, which can also use system memory; while dGPUs try not to use it due to latencies and bandwidth, iGPUs use it freely
23:51 Here2Learn: I see
23:52 mareko: neither CPU nor GPU can use all RAM, the carve-out is only for the GPU, while the GPU is limited how much regular sysmem it can use, which we can change, right now it's something like 1/2 of RAM
23:53 mareko: that's iGPU