09:12karolherbst[d]: phomes_[d]: how are you running the linux version of xcom2?
09:22karolherbst[d]: ohh or is it the windows version and it has a vulkan renderer?
09:38phomes_[d]: xcom2 is the windows version and running with dxvk
09:38karolherbst[d]: okay..
11:53karolherbst[d]: phomes_[d]: was my UGPR branch ever seeing benefits in the games you are tracking?
11:55phomes_[d]: it did tip lego builder's journey to 33 fps before. But other changes landed that also got it there. So it does not show up in the sheet
11:56karolherbst[d]: ahh
11:56karolherbst[d]: okay
11:56karolherbst[d]: let me check if any game stands out in the stats, but it always looked more like 1-2% across the board kinda change
11:56karolherbst[d]: ohh control has a -10% instruction count hit...
11:56karolherbst[d]: deathloop as well
11:57karolherbst[d]: god of war ragnarok
11:57karolherbst[d]: seems like there are a few games with high hits
11:57karolherbst[d]: FEAR as well
11:58karolherbst[d]: I ahve some of them.. maybe I should check it out after the VK CTS run is complete
11:58asdqueerfromeu[d]: phomes_[d]: ~~How well does LEGO Island run though?~~
12:00karolherbst[d]: phomes_[d]: you should add `-nostartupmovies` to your xcom2 start up options 😄
12:02phomes_[d]: thanks 🙂
12:02phomes_[d]: I am not familiar with lego island sorry
12:06marysaka[d]: `-nostartupmovies` causes a lockup with gfxrecon btw
12:06marysaka[d]: so if you want to capture do not use it but otherwise it boots instantly to menu
12:07karolherbst[d]: most games have an option like this, because the devs of those games are just as annoyed as us 🙃
13:22karolherbst[d]: anyway.. would be good to get reviews on https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39384 because I kinda want to also clean up patches that depend on it
13:25karolherbst[d]: ohh there is also https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38319
13:25karolherbst[d]: Mel started it, but I've added some patches on top
13:57karolherbst[d]: Sooooo
13:57karolherbst[d]: I tested Control first on my Turing
13:57karolherbst[d]: and I was like: 7-8 fps.. yeah that ain't great
13:58karolherbst[d]: switched to my ampere GPU: 165 fps 😂
14:01karolherbst[d]: so that's the compute WFI stuff prolly because control does compute stuff quite a bit
14:02karolherbst[d]: the turing one is a quadro RTX 6000 and the ampere one is a quadro RTX A6000, so the difference shouldn't be _that_ huge
15:05mohamexiety[d]: the a6000 should be 50%ish or so ahead of the turing 6000 at the absolute best case
15:18karolherbst[d]: yeah.. but like turing doesn't have compute MME
15:18karolherbst[d]: and we use 3D MME to write compute invocations
15:18karolherbst[d]: and that kinda sucks for perf
15:21mohamexiety[d]: yeah i know
15:21mohamexiety[d]: was just comparing to how the difference should normally be to highlight it's not a turing/ampere thing
15:23karolherbst[d]: I'll try to hack something up, but we could maybe allocate 8 bytes and just semaphore it...
16:23mhenning[d]: Yeah, I was considering using semaphores for it
16:25mhenning[d]: at first I thought you could use the semaphore to atomic inc directly to the query pool address but I think there's a complication in that secondary command bufs don't know what the query pool address is but still need to do the increment
16:26karolherbst[d]: yeah.. not sure how that will look like, but anything is going to be better than what we have today
17:30karolherbst[d]: phomes_[d]: if you got some time, mind verifying this patch to the MR? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38682#note_3454140
17:39phomes_[d]: still same issue in lego builders journey
18:16karolherbst[d]: mhh
18:16karolherbst[d]: thanks for checking tho!
18:19karolherbst[d]: phomes_[d]: but can you check if it's really https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38682/diffs?commit_id=9178ab75b10005e0d707996cf93513ed4f0b23a3 that breaks this or the other commit?
18:22mhenning[d]: yes, it's the mrt bit that makes the difference
18:22mhenning[d]: see also https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38682#note_3279047
18:22karolherbst[d]: mhhh
18:23karolherbst[d]: yeah.. I don't really get why tbh.. unless the hw does something funky in some cases...
18:24karolherbst[d]: but what we do now is, that the hw apparently enables MRT if the sph bit _and_ `SET_CT_MRT_ENABLE` is TRUE
18:24mhenning[d]: Yeah, it's odd. I spent a little time debugging it, but it's not obvious what's going on
18:24karolherbst[d]: soo...
18:24karolherbst[d]: MRT disabled says it simply broadcasts each color to all outputs
18:25karolherbst[d]: ehh
18:25karolherbst[d]: color0 to all outputs
18:26karolherbst[d]: mhenning[d]: can you have like a single output in the FP, but multiple render targets attached to the invocation or something?
18:27karolherbst[d]: and they just don't get touched?
18:28karolherbst[d]: so I was wondering if we need to track "only writes color0" vs "broadcasts color0", and then flip `SET_CT_MRT_ENABLE` instead of the SPH bit depending on the draw?
18:29karolherbst[d]: maybe I should write something up for that...
18:29karolherbst[d]: but I don't have access to the game
18:34mhenning[d]: Not really sure. I think unwritten outputs are supposed to be undefined, which means we're allowed to do this opt. It's possible the game is doing something out of spec.
18:35karolherbst[d]: yeah....
18:35karolherbst[d]: but this is what that flag is changing, so I'd say it's likely something like that?
18:37karolherbst[d]: looks like that game goes on sale for 80% regularly, maybe I should buy it one day
18:52mhenning[d]: karolherbst[d]: I'd really rather not bother with that, especially if the game is out of spec. For https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39247 at least, we have the option of only disabling the mrt bit if the optimizaiton actually turns on. I remember fewer details from https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38682
19:06karolherbst[d]: mhh but that won't allow us to enable constant color rendering for shaders that only write to a single color
20:12glehmann: surprised to see https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39247 have an effect, does nvk not remove fragment output writes if no attachment is used for the slot?
20:14glehmann: otherwise, I'm kind of suprised that anything actually wants to write the same value to multiple attachments
20:14mhenning[d]: no, we have no knowledge of which attachments are used at compile time - shaders aren't specialized that way right now
20:37glehmann: that sounds pretty bad, I've seen games that use literally the same fragment shader for a depth pre pass as the later color draw
20:37glehmann: and if discard is used, the hw will still have to execute the FS
20:42glehmann: another reason why ESO is a bad idea, even for hardware that doesn't need state to compile shaders
20:57mhenning[d]: do you have an example of a game that does that? I'd be curious if we could measure a perf difference
21:37tiagotorres: cyclic redundancy check isn't much superior to the TCP checksum either. You know dudes wire is wire, and a chip is a chip, same wise as lie is a lie and traitor is a traitor, so wire can not be a cascaded bus for tcp or crossbar bus it's just a wire, so it counts only digital bit streams where all lows and highs are encoded through two voltages. hence tcp is just that bad because all the checksum algorithms are ignoring the order of words. There
21:37tiagotorres: isn't that much else to it. Prime is a very simple concept in cryptography too, 10 toes ten fingers debian fellows are wrong, the solution comes through eulers imaginary numbers, though i do not know how blind man was so famous mathematician, prime is a number that divides with itself or 1 only all other dividers yield a number which is floating point number. Only then that number is considered prime. But i got into troubles because animals from
21:37tiagotorres: estonia conspired against me, there is not much more to it, no one wants to talk this over with me, cause they were wrong to begin with in doing that through abuse and against laws. And TCP is a same when you offer details about any communication to a abuser not man in the middle attacker, it no matter how big birth leftover or trash one is, can possibly lie about everything and intrude into a system which was used to provide service to others.
22:03mariopocelli: no matter if that was sport, kids, girlfriends, sex or programming, their job was to betray me. So as they did that through the years on every opportunity they of course do not want to speak until i am alive. I can say it's only a miracle if i ever recover from the set of betrayals since i got very injured, and it is likely whenever i recover, if i am forced to continue the same, they injure me again. And really no one goes to conspire against
22:03mariopocelli: another person if ones could handle such by other means, so hence it's theater, i know who i played 10years with i know the guy was crying every day just to build a moment of betrayals at any stage in the future. But i never did much to him bad, it just he could not beat me in legal sports.
23:12user8473: Hi, switching to any kernel of the 7.0 series causes unusably frequent mmu faults. The simplest way I found to reproduce it was snapping windows, that seems to have about 10% chance to cause one. After a git bisect, the first bad commit seems to be https://gitlab.freedesktop.org/drm/kernel/-/commit/d19512f5abb198daf29da877f6a02c667a95c03d . I'm running KDE Plasma on Gentoo on a
23:12user8473: TU104, my Mesa version is 26.0.5 with this (almost certainly irrelevant) asahi bugfix backport: https://paste.gentoo.zip/aGhIRYPD
23:40mhenning[d]: hmm that kernel patch also bumps the version number which has the effect of telling userspace that compression is okay to turn on
23:41mhenning[d]: might be the same issue as https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15358
23:48mhenning[d]: user8473: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41332 might have a workaround