01:07pabs3: epsilon_0: what kernel command-line parameter did you add?
01:08 < mia\> epsilon_0: also, depending on how your system is configured, it might be necessary to add that EDID file to your initramfs for it to work
01:09pabs3: indeed, I had to do that
01:09pabs3: also, this page might be a useful alternative: https://wiki.debian.org/RepairEDID
01:44epsilon_0: Yes I added it to initramfs
01:52epsilon_0: also even if edid is not provided its not even giving a low resolution display
01:52epsilon_0: just ignoring it
01:52 < mia\> defective cable?
01:54epsilon_0: that same monitor works on second os with same cable
01:54epsilon_0: (with nvidia drivers)
01:55epsilon_0: when I look at my initrd with 'lsinitrd /boot/<initram-file>'
01:55epsilon_0: i see the following two lines
01:55epsilon_0: drwxr-xr-x 2 root root 0 May 15 14:24 lib/firmware/edid
01:55epsilon_0: -rw-r--r-- 1 root root 256 May 15 14:01 lib/firmware/edid/1920x1080.bin
01:55epsilon_0: my kernel boot commandline is -
01:55epsilon_0: linux /vmlinuz-5.10.33-gentoo-dist root=UUID=f6073ec2-4d9a-4758-adb4-5b5db6ec94f1 drm.edid_firmware=/lib/firmware/edid/1920x1080.bin drm.edid_firmware=eDP-1:/lib/firmware/edid/1920x1080.bin drm.edid_firmware=DP-1:/lib/firmware/edid/1920x1080.bin drm.edid_firmware=DP-4:/lib/firmware/edid/1920x1080.bin ro
01:56epsilon_0: i basically just added as many damn edid's as i could think of
01:56epsilon_0: cuz the syntax documentaion is hard to find
01:56pabs3: I think drop the "/lib/firmware/" from each of the command-line arguments
01:57epsilon_0: ok, trying that out
01:57pabs3: DP-1 is the correct one according to the message you posted earlier
01:58pabs3: epsilon_0: I think drop all of them except drm.edid_firmware=DP-1:edid/1920x1080.bin
02:01epsilon_0: Still stuck at boot screen
02:05epsilon_0: did not work unfortunately
03:04pabs3: anything interesting in dmesg?
04:48imirkin: gyah!! adding debug_printfs in nv50_query_hw "fixes" stuff
04:48imirkin:is going to cry
05:30imirkin: pmoreau: can you check if dEQP-GLES3.functional.transform_feedback.position.points_separate passes for you on your two gpu's?
05:30imirkin: i have a workaround, but it ain't pretty
05:44imirkin: mwk: can you think of any reason that a nva3 board would see a "fence" write (done using semaphore release) but not a previously-issued query result? both in gart.
05:45imirkin: the fence write is done with the "UNK4" bit which i believe ensures previous writes to be completed
08:25ccr: heisenbugs :/
11:36mwk: imirkin: what is the exact 1b0c method word value?
11:37mwk: for both the query and the fence
15:09imirkin: mwk: fence: https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/nv50/nv50_screen.c#n632
15:10imirkin: mwk: query: https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/nv50/nv50_query_hw.c#n225
15:11mwk: imirkin: the obvious problem here is that the units don't match
15:12imirkin: why should the units match?
15:13mwk: or hm
15:14imirkin: also fwiw we use a semi-identical process on nvc0 where it works fine. of course the specific bits are different there though.
15:15mwk: I don't know why
15:15mwk: but please check it
15:15mwk: my intuition would be that it could possibly race
15:16mwk: because the fence and query go different directions in the pipeline
15:16imirkin: i thoguht that's what the "UNK4" bit was for
15:16mwk: like, try using a fence with unit set to STRMOUT
15:16imirkin: ok, will do
15:17imirkin: but that's obviously impractical
15:17mwk: you know how the query already includes its own sync mechanism, right?
15:17imirkin: sort of
16:46Lyude: imirkin: yeah seeing as it was a DDC error that makes sense
16:46Lyude: the hpd pin for the connector was probably making contact before the actual ddc pins were
16:47imirkin: Lyude: well, hpd pin is *supposed* to make contact
16:47imirkin: or ... dunno
16:47Lyude: supposedly :). I've seen this happen with DVI though
16:47imirkin: but i was just unplugging the hdmi cable
16:47imirkin: whereas the adapter was secured onto the board
16:47Lyude: huh, weird
16:48Lyude: hdmi usually isn't as bad with that
16:48imirkin: monitor -> hdmi output -> hdmi cable -> hdmi->dvi adapter -> dvi-i connector
16:48imirkin: however it was flickering while i was using it too
16:48imirkin: so i suspect just something was loose
16:48imirkin: curious that 'xrandr' would force a flicker on it too though
16:49imirkin: but after i fiddled with it, no more random flicker, and no more flicker on xrandr
16:49imirkin: so i'm going to leave this in the "meat-space sucks" land
16:55RSpliet: Lyude: mind taking the honours of bringing xorg-x11-drv-nouveau(.fc32) from testing to stable? Looks like the Bodhi powers that be have deemed it good enough :-)
16:56Lyude: RSpliet: did it not auto-push itself?
16:57RSpliet: I don't think it has. Nothing about auto-pushing on bodhi, and it still is listed as in "testing"
16:57RSpliet: There's a line "This update can be pushed to stable now if the maintainer wishes". I presume that means it'll take some intervention :-)
16:58Lyude: RSpliet: I can push the f34 one but the other two need more time before bodhi will allow me to push them
16:59RSpliet: I figured. There's not that many users of it in the first place, certainly not who also stuck around on f33. I guess nobody's going to test it
17:08imirkin: karolherbst: care to glance at https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10820 ?
17:10imirkin: it's basically a copy of the nv50 logic, but there are some slight differences to what the hw wants, and also we get to use EXTBF :)
17:26karolherbst: yeah.. I already took a look, but my biggest concern is mostly around hurting the 2D use case where 3D matters very little
17:27imirkin: karolherbst: yeah, i mean ... we have to do something dumb like that on maxwell too
17:27karolherbst: well.. we could also not tile on the z axis :p
17:28karolherbst: but yeah.. for pre maxwell the situation is.. more annoying
17:28imirkin: karolherbst: the way i look at it is that the extra constbuf/etc fetches will pale in comparison to the image access time
17:28karolherbst: we could just get rid of some of the calculations
17:28imirkin: not tiling along the z axis is not an option
17:28imirkin: even if there's 1 z tile
17:28imirkin: it still doesn't "just work out"
17:28imirkin: and you can't have "linear" 3d textures
17:29karolherbst: well.. things get cached at some point, but yeah. I think for pre maxwell this is matters probably not much
17:29imirkin: this is for fermi only
17:29imirkin: kepler has the sueau which Just Works (tm)
17:29karolherbst: yeah.. right
17:30imirkin: and maxwell uses those image accessors which unfortunately don't allow you to use a "2d" access type on a "3d" image descriptor
17:30karolherbst: checked what nvidia is doing?
17:30imirkin: way back in the day, iirc they were doing something like this
17:30imirkin: but i don't know specifically
17:30imirkin: i expect that they have shader variants
17:31karolherbst: yeah.. the only thing I know is, that nvidia disables tiling on the GPUs I checked, but I also wasn't looking really deeply to check out when they are doing it and when not
17:31karolherbst: ohh, maybe
17:31imirkin: karolherbst: how can you disable tiling?
17:31karolherbst: well, on the z axis
17:31imirkin: how can you disable tiling on the z axis?
17:31karolherbst: you set it to 1
17:31imirkin: set what to 1?
17:31karolherbst: the tiling parameter
17:31imirkin: that means each tile has a depth of 4
17:32karolherbst: well, you specify the sizes of the tiles and there you can say it's just x*y*1
17:32imirkin: let's say you do that.
17:32karolherbst: and then it all just works out if you addjust the offset
17:32karolherbst: I mean.. I had that working locally
17:33imirkin: oh right. if each z tile is depth 1
17:33imirkin: then it's essentially a "tall" texture
17:33karolherbst: sure.. it might hurt perf of 3d images, but I wouldn't care much about that
17:33imirkin: i.e. x = x, y = y * z
17:33imirkin: in any case ...
17:34imirkin: there's no real reason to resort to this
17:34karolherbst: at least that's what nvidia did as well :)
17:34imirkin: well, the thing i did works fine
17:34karolherbst: yeah, it does
17:34karolherbst: it just makes shaders annoying to look at :D
17:34imirkin: eh, very slightly
17:34imirkin: it's not a ton of extra code
17:34karolherbst: well.. you also have to deal witht he predication stuff
17:34imirkin: what predication stuff?
17:35karolherbst: well, you predicate the tex instructions
17:35imirkin: we already do that.
17:35karolherbst: because you emit 2D and 3D
17:35imirkin: what i do in no way changes that
17:35imirkin: on maxwell, yes
17:35imirkin: on fermi, no
17:35imirkin: i do the 3d logic irregardless
17:35karolherbst: I think for fermi this is fine as the hw is very limited anyway
17:35karolherbst: and if we would care, we would do shader variants
17:36imirkin: (i could make it conditional, i suppose, but that would have very little end effect i think)
17:36imirkin: and makes the pass more annoying to write :)
17:36karolherbst: yeah.. maybe.. dunno
17:36karolherbst: nvidia does crazy things to squeeze even 0.5% perf out of something
17:36imirkin: thing is ... image access will eclipse any sort of ALU ops we do to fix up the coords
17:36imirkin: at least that's how i think abou tit
17:37karolherbst: depends on what's in the cache
17:37imirkin: if there's an image-heavy thing, i can try running it and see if there's any effect
17:37karolherbst: I doubt you will see much though honestly.. and writing benchmarks is also annoying
17:37karolherbst: the only thing I can think of is CL benchmarks doing image stuff
17:38karolherbst: and then you still have crappy perf anyway
17:38karolherbst: because fermi
17:38imirkin: also because Quadro 400 ;)
17:38imirkin: not exactly the king of the fermi series
17:39karolherbst: anyway.. today I will collect the GPUs from mia\ :)
17:40 < mia\> speaking of fermi, i already sold that one on ebay but i used to have a gtx 560 ti (the one with 448 cores). put an aftermarket heatsink on it, ended up being some 4-slot monstrosity (but it was slightly faster at cycles raytracing than the gtx 960!)
17:41karolherbst: what people do for 5% more perf
17:41 < mia\> i did get the thing for free back then
17:42 < mia\> but holy hell people were not kidding about fermi running hot
17:42karolherbst: fermi was cool in comparison to tesla
17:42imirkin: you think gtx 560 was hot?
17:42imirkin: try gtx 480 :)
17:42karolherbst: I have this mac mini here with a 8600m or something
17:42karolherbst: 80º idling
17:43karolherbst: might also be a different tesla
17:43karolherbst: but those were hot :D
17:43 < mia\> the 9800 gt i sent you ran pretty cool with the stock heatsink though
17:43 < mia\> i only replaced that because the fan died
17:43karolherbst: using nouveau will probably change that :D
17:44karolherbst: although not sure how advanced power management was back then
17:44 < mia\> not very
17:44karolherbst: you choose from 2 perf profiles and maybe enable some addiditonal features to reduce idle power consumption
17:44karolherbst: the latter part is what we generally suck at if we don't suck with the first part
17:44 < mia\> (also i mean under load of course)
17:44karolherbst: run furmark
17:45karolherbst: furmark is this crazy benchmark which will bring the GPU to max heat _and_ reducing clocks by 60%
17:45 < mia\> yeah i did, i don’t think i’ve seen temps much higher than 75°C with that card
17:45karolherbst: ohh, cool
17:45 < mia\> unless the sensors were lying
17:45karolherbst: probably not
17:46karolherbst: there are also ways of disabling the power capping
17:46karolherbst: but... that's a bit.. dangerous
17:49 < mia\> my current GPU is a vega 56 and i found out the hard way that apparently asus put bad fan speed tables in the firmware :)
17:49 < mia\> it got very toasty when i ran some badly written unity game on it, thermal shutdown
17:50 < mia\> easy to fix, but still
17:50 < mia\> i don’t get why fan control is so difficult to get right for all those vendors. my mainboard is no exception; fan control is buggy
17:51 < mia\> put together my own with an stm32 :/
17:51karolherbst: the only devices which proper fan control are 0.1% of all laptops and that's it
17:52karolherbst: either they spin up for every little load, they annoy you for 10 minutes even though everything is idle or it's just broken and devices emergency shutdown :D
17:52 < mia\> miracles do happen though; last yeart asus fixed a firmware issue that fixed the acpi stuff on my 2015 laptop
17:52 < mia\> -t
17:52karolherbst: uhh nice
17:53 < mia\> fan would be stuck at 100% after suspend to ram, required hard power off
17:53karolherbst: I am quite happy with the dell laptop I have now. You can put it into silent fan mode and fans only turn on ~5 seconds after heavy load
17:53karolherbst: so spikes from web browser are irrelevant even if the CPU reaches 100º
17:54 < mia\> yeah. it is otherwise very quiet
17:54karolherbst: yeah.. mine is silent :D
17:54karolherbst: the only time the fans spin up is if the discrete GPU is on or when I am compiling stuff
17:54 < mia\> garbage specs (low-power haswell) but i’ll use it until it dies and cannot be fixed because i can’t find a decent replacement nowadays
17:55karolherbst: yeah.. good laptops are expensive
17:55 < mia\> i cannot find good ones in any price range
17:55karolherbst: well, the M1 macbooks are supposed to be very good :p
17:56karolherbst: are beating intel chips in literally every benchmarks and are still less power hungry
17:56karolherbst: intel is a joke anyway..
17:57 < mia\> i don’t get why they insist on bandwidth starving their igpus so much
17:57 < mia\> this laptop can barely run a browser at 1080p60
17:57karolherbst: when I was replacing my dell dock with a fan by a silent one I noticed that the CPU alone is able to draw 110W (even though it's TDP is 45W), because of "performance"
17:57karolherbst: mia\: RAM isn't very fast
17:58karolherbst: it got better with DDR4, but compared to GPU memory it's slow
17:58karolherbst: what do high end nvidia GPUs do? 1 TB/s?
17:58karolherbst: try reaching that with DDr4
17:59karolherbst: I am not sure if you can even reach 100 MB/s with dual channel RAM
17:59 < mia\> well, AMD integrated seems to have a lot more bandwidth though (not that i’ve measured, but i get that impression looking at benchmarks)
17:59karolherbst: * 100 GB/s
17:59karolherbst: AMD tends to enable higher memory speeds by default
18:00karolherbst: I think
18:00karolherbst: but yeah.. without dedicated GPU RAM you are kind of screwed and intel might even mess up something
18:00karolherbst: there is a reason they added this L4 cache :D
18:01 < mia\> heh
18:02 < mia\> also seems like on linux you get to choose between a broken DDX, broken vsync, or massive input lag (i’ve seen up to 5 frames)
18:03 < mia\> not happy with that
18:03imirkin: why choose?
18:03imirkin: you can have them all!
18:03karolherbst: honestly, I am quite fed up with the garbage x86 CPUs we get these days... and if we won't have fanless high perf x86 laptops in 2022, apple will clean up the entire laptop market just like that I figure
18:03karolherbst: mia\: it got better with DRI3, but ultimately wayland is the solution
18:03 < mia\> good riddance tbh
18:03 < mia\> karolherbst: i was seeing MORE lag with wayland!
18:03karolherbst: uhh.. annoying
18:03karolherbst: but I guess it depends on the compositor
18:04 < mia\> which is the opposite of my eperience with AMD
18:04karolherbst: I think fixing that would be cool and if somebody would look into improving that
18:05 < mia\> these are all first world problems anyway
18:05karolherbst: I am actually searching for fan less home servers.. but in the end you have to buy this 150W TDP fanless case and buy the parts yourself to get that :D
18:07 < mia\> our “home server” is also an htpc. not fanless, but the UPS humming is much louder than it anyway
18:08 < mia\> one of the amplifiers has fans, too (but mercifully they are temperature-controlled)
18:08karolherbst: I was thinking of getting this one as you can out tons of storage into it as well: https://streacom.com/products/db4-fanless-mini-itx-case/ and there is a kit to get to 105W TDP CPU cooling
18:09karolherbst: the only problem is.. they only have a 240W PSU you can install internally :D
18:09karolherbst: which.. should be enough I guess
18:10karolherbst: but I am more surprised that all those "NAS" things out there are either super expensive or annoying and buying single slot USB cases is usually cheaper
18:10 < mia\> https://0x0.st/-bZG.jpg well
18:10karolherbst: yeah well.. mhh
18:11karolherbst: I didn't plan on getting a rack, but...
18:11 < mia\> well in that case https://nazrin.moe/system/cache/media_attachments/files/106/248/099/557/604/636/original/446daea444e81948.jpg
18:12karolherbst: but I need it for food, so..
18:12karolherbst: but this home server thing is on my todo list for 5 years already...
18:13 < mia\> taken a liking to these studio racks. built one for my desktop PC as well and it’s much nicer than a tower
18:14karolherbst: the company doing that fanless cubes also has htpc cases: https://streacom.com/products/fc10-fanless-atx-case/ :D
18:15karolherbst: oh well
18:16karolherbst: I will continue searching and in 5 years when we finally have 2nm this won't matter anymore
19:19karolherbst: mia\: thanks :D
19:19 < mia\> :)
19:22karolherbst: and next month I might even have a shelf where I can sort all my GPUs :D
19:25 < mia\> https://www.amazon.com/LINKUP-Extreme-Shielded-Twin-axial-Extension/dp/B07CS54VHS/keywords=pcie+8x+to+16x+cable get a bunch of these cursed riser cables for easy swapping :D
19:27karolherbst: the problem is rather, how can I choose which GPU to actually use
19:29karolherbst: ohh you mean one per GPU..
19:29karolherbst: I think I have like 20 GPUs here :D
19:33 < mia\> i was mostly kidding anyway because *just look at it*, 400cm pcie extension, i don’t even
19:36ccr: sounds totally feasible and stable
21:41imirkin: mwk: congrats, you were right (as if i doubted it)
21:41imirkin: mwk: changing the fence to unit = strmout fixes those queries
21:41imirkin: obviously i can't just do that, but i'll work it out a different way