00:08 imirkin: wifwj: because XvMC is done.
00:08 imirkin: should work fine
00:08 imirkin: VP1 is pointless.
00:09 imirkin: wifwj: my NV42's fan control definitely *used* to work
00:09 imirkin: it's a Quadro FX 3450
00:10 wifwj: i have a quadro fx3450 here. Its 100% not working. Kernel 4.13
00:10 wifwj: your bios is same as mine? https://bugs.freedesktop.org/show_bug.cgi?id=102352#c4
00:11 imirkin: well, i probably last tested it more like kernel 3.13 than 4.13 ;)
00:14 wifwj: when you add nouveau.debug=debug you can see "nouveau 0000:08:00.0: therm: parsing the fan table failed"
00:15 imirkin: yeah, that's a recent fail
00:15 imirkin:blames karolherbst
00:16 wifwj: the problem was not there on 4.7 or something like that. I remember that this patch was enough for getting fan working fine when the vbios does not contain fantables: https://lists.freedesktop.org/archives/nouveau/2016-September/026082.html
00:17 wifwj: the patch was then included by skeggsb in 4.11-rc1. I tested 4.12 and 4.13 but fan was whole time at 100%. Something have changed in between
00:17 imirkin: what about 4.11?
00:18 wifwj: you write, that you blame karolherbst for that. you know what change affects that? In what kernel version have this been merged?
00:19 imirkin: i just remember he was messing around with bios parsing stuff
00:19 imirkin: i think it was more clock-related, but could have spilled over into fan stuff
00:19 imirkin: either way, if you can bisect it to a specific change, that greatly increases the odds of it being fixed
00:21 wifwj: i think mupuf is way better then me in understanding such things
00:23 imirkin: yes... and if you just leave it to others, then you have to work around their schedules
00:23 imirkin: in mupuf's case, i don't think he's had any time for nouveau in ... quite a while
00:24 gnarface: also probably worth mentioning here that this information is not known to be correct, accurate, or even generally always parseable under the rules the standard dictates in every BIOS
00:25 gnarface: there's actually evidence microsoft collaborated secretly with multiple big hardware vendors to sabotage compliance with their own standard here in secret, specifically to make it harder for linux support to happen later
00:25 gnarface: so if you even flashed your motherboard bios, it could have changed the behavior of your tests
00:26 gnarface: (regardles sof whether they were working right before or not)
00:46 imirkin: wifwj: this is my nv42's vbios: https://people.freedesktop.org/~imirkin/traces/nv42-vbios.rom
00:48 wifwj: they differ. I dont know where/why. just checked the checksums
00:49 wifwj: now i see also the size differ. your bios file is smaller
00:50 gnarface: ah, yes different vendors won't even always use the reference one. sometimes they hide their own weird crap in there too.
00:51 imirkin: INFO: BIOS version 0x05. from 05/04/05 for NV41 [NV41/NV42]
00:51 imirkin: [from mine]
00:53 imirkin: [output from the nvbios tool]
00:53 imirkin: it's a single-slot board with 2x DVI and a stereo connector
00:53 imirkin: and a fan that sounds like it powered the concord in a former life
01:01 wifwj: i think every quadro fx3450 looks like that: https://www.techpowerup.com/gpudb/1368/quadro-fx-3450
01:02 imirkin: yeah, looks right
01:02 wifwj: have you checked how the bios here differs from yours? https://bugs.freedesktop.org/show_bug.cgi?id=102352#c4
01:02 imirkin: "We recommend the NVIDIA Quadro FX 3450 for gaming with highest details at resolutions up to, and including, 1024x768."
01:02 imirkin: heh
01:16 wifwj: imirkin: i am glad you have now at least two nv4x cards. one geforce for pci and one quadro for pcie. btw, you can use the quadro to reproduce exact that here: https://bugs.freedesktop.org/show_bug.cgi?id=102430
01:20 imirkin: i have a bunch...
01:20 imirkin: nv42, nv44 pcie, and nv44 pci
01:20 imirkin: i have something like a dozen nvidia gpu's overall
01:24 wifwj: maybe you could play a bit with them. you would find many things as you probably expect now ;) . Booting rebeccablackos live linux iso would probably also result in challenging issues
01:29 imirkin: yeah, i don't like challenging issues.
01:30 imirkin: it's very hard to make something go from non-working to working
01:32 wifwj: you already did on the issue you had send a you named it "dirty patch". it went to a better situation afterwards :)
01:39 imirkin: well, fixing crashes is usually straightforward
01:42 imirkin: i need to properly understand it though... we'll see.
01:42 imirkin: unfortunately i have a NV5 and NV17 in my PCI slots right now
02:04 wifwj: thanks so far. i am happy to maybe see your not-ditry patches next days
05:03 imirkin: karolherbst: can you get a mmt trace of tests/spec/arb_compute_shader/execution/basic-local-index.shader_test ? i want to see if the blob has a better way of computing it
05:15 imirkin: hm, interesting. the "tid" special reg appears to be tidx,y,z with x/y/z taking up 16/10/6 bits respectively.
05:16 imirkin: should be faster to pull than the 3 separate SR's... but ... doesn't seem worth it
05:17 imirkin: should probably have an opt that coalesces them if multiple ones are fetched in a single BB
06:07 karanagarwal17: Hey everyone. I am using nouveau to connect 2 monitors to my laptop, and having some issues.
06:07 karanagarwal17: Can anyone help ?
08:18 karolherbst: imirkin: I highly doubt the fan change was due to my changes.. anyhow, I will do the mmt traces in around 8 hours
17:29 karolherbst: imirkin: https://filebin.ca/3euVJvgGk1k0/log.mmt.xz
17:29 imirkin_: karolherbst: mind pastebinning the shader?
17:30 karolherbst: uhh, yeah, wait
17:31 karolherbst: imirkin_: https://gist.github.com/karolherbst/86a9efffcd1f28a6d52856d110627287
17:31 karolherbst: or do you mean extracted from the mmt?
17:31 imirkin_: hehe yeah
17:32 imirkin_: sorry, i thought that was implied :p
17:32 karolherbst: imirkin_: https://gist.githubusercontent.com/karolherbst/86a9efffcd1f28a6d52856d110627287/raw/3f52ea69459b343a08a8f072bfe61eaf3753756e/shader
17:33 imirkin_: huh, ok
17:33 imirkin_: that's doubly odd
17:33 imirkin_: i would have thought they'd hae the opt
17:34 imirkin_: ok, thanks for checking
17:40 pmoreau: imirkin_: Would it be fine with you to have a legalize pass to split 64-bit loads and stores into 32-bit chunks, rather than having that in each frontend?
17:43 imirkin_: pmoreau: yeah
17:44 imirkin_: pmoreau: i think you can actually do 64-bit loads/stores on fermi+
17:44 pmoreau: I would guess so for Fermi, definitely the case on Kepler+, which can even do 128-bit loads and stores.
17:44 imirkin_: mmmm ... don't trust the flags
17:44 imirkin_: sometimes they lie
17:45 imirkin_: i.e. there's an instruction, but it doesn't work ;)
17:45 imirkin_: the other alternative is to always split it into smaller chunks and then let the MemoryOpt pass handle it. but for loads/stores, it seems fine to split after.
17:46 imirkin_: e.g. 128-bit const buf loads don't work on kepler+
17:47 pmoreau: I haven’t tried on const buffers, but I have definitely seen them on global memory loads/stores from compiling CUDA kernels. And it is advertised as such in different CUDA talks.
17:47 pmoreau: Maybe they just hack things behind the scene though
17:49 imirkin_: ok
17:49 imirkin_: well i've never tried it with gmem
17:49 imirkin_: glsl is dumb
17:49 imirkin_: if you have an ssbo, you have to check for every goddamn 32-bit piece
17:49 imirkin_: like if you have a ssbo { vec2 foo; }
17:49 imirkin_: and then you write foo = vec2(a, b);
17:49 imirkin_: you can't do that as one write
17:49 imirkin_: because the underlying buffer may only be 4 bytes
17:49 imirkin_: and so the latter 4 bytes have to be ignored
17:50 imirkin_: i should do it as a shader variant
17:50 imirkin_: i.e. have a global check for the min being > something
17:50 imirkin_: and if it is, stop doing the stupid checks
17:50 karolherbst: ohh, before I forget it again: last chances to complain, otherwise I add it most likely today or tomorrow (EVoC/GSoC stuff): https://gist.githubusercontent.com/karolherbst/f80890aad3983cd37a502888229d0978/raw/a7809a6c8846c60134a95a2b645f92602a7da9c0/gistfile1.txt
17:51 imirkin_: karolherbst: haven't read it, but do you have a proposed mentor for each one of those?
17:52 imirkin_: preferably a proposed mentor who isn't me
17:53 karolherbst: imirkin_: no
17:54 imirkin_: ok, that's a pretty serious problem
17:54 karolherbst: yeah sure
17:54 karolherbst: but I think I could mentor most of those
17:54 karolherbst: except the video accel things
17:54 karolherbst: and maybe nir, but this could be also done by somebody not doing nouveau things
17:54 karolherbst: and vulkan
17:54 karolherbst: yeah...
17:54 pmoreau: Dang, SSBO looks annoying :-/
17:55 karolherbst: so yeah, we have a mentor problem with those already published
17:55 karolherbst: the others aren
17:55 karolherbst: t the issue
17:56 karolherbst: imirkin_: but mupuf and me are on the Xorg EVoC mentor list
17:59 imirkin_: karolherbst: ok, and you or mupuf would be willing to mentor any of those projects?
17:59 karolherbst: basically yes
17:59 imirkin_: excellent.
17:59 karolherbst: I would just feal unsure about the video stuff, because lack of knowledge
17:59 karolherbst: but we didn't get many students for this anyhow
18:04 karolherbst: imirkin_: would it be okay if getUniqueInsn() also returns a random element?
18:04 karolherbst: or should it assert that there is really just one unique one?
18:13 karolherbst: imirkin_: okay, returning any instruction in getInsn isn't a good idea
18:22 karolherbst: imirkin_: https://mini.karolherbst.de/nouveau/piglit/nve6-cts/regressions.html see this "weird" run? this was built with O0
18:24 karolherbst: not even caring much about the fails, but the passes
18:32 ylwghst: Hi
18:33 ylwghst: this always brick my xserver https://gist.github.com/ylwghst/ae64db7dcf5348ee6bd5020e1f437ce1
18:34 karolherbst: ylwghst: plasma5?
18:34 karolherbst: ylwghst: or any other application which might trigger this? or does it just happen randomly?
18:35 ylwghst: karolherbst: i use only i3 WM and slim display manager
18:36 ylwghst: karolherbst: propably when i start atom editor after some time of use but im not sure
18:36 karolherbst: atom == chrome
18:36 karolherbst: and chrome is complex enough to trigger some odd things inside nouveau
18:36 ylwghst: but it could be firefox too
18:36 karolherbst: well, nouveau doesn't do well with multithreaded stuff and many OpenGL contexts
18:36 karolherbst: allthough it should be better
18:37 karolherbst: ylwghst: what kernel are you on?
18:37 ylwghst: karolh$ uname -vrs
18:37 ylwghst: Linux 4.9.58 #1-NixOS SMP Sat Oct 21 15:21:39 UTC 2017
18:37 karolherbst: ylwghst: you could try 4.13
18:37 karolherbst: it should have some changes to improve the situation afaik
18:37 ylwghst: i would like but i cant get it work
18:38 karolherbst: not quite sure
18:38 karolherbst: mhh
18:38 karolherbst: what's the issue with a newer kernel? Nouveau related or some random stuff?
18:38 ylwghst: it cant bot it due to error alocating crypto tfm error
18:38 ylwghst: boot*
18:38 ylwghst: even if i enable crypto modules
18:38 karolherbst: meh :/
18:39 karolherbst: well, I can just suggest you to use a newer kernel version and see if it changes something
18:39 karolherbst: if not, we could dig deeper
18:43 ylwghst: there isnt another kernel available in nixos repository currently so i have to stuck with 4.9. or older
18:44 karolherbst: :/
19:36 ylwghst: karolherbst:
19:36 ylwghst: so
19:36 ylwghst: im in 4.13 already
19:36 ylwghst: solved the problem with encryption
19:36 imirkin_: ylwghst: what gpu are you on?
19:36 ylwghst: karolherbst: i just started firefox and atom and there is
19:37 imirkin_: looks like a tesla of some sort?
19:37 ylwghst: https://gist.github.com/ylwghst/f933c8584d04f4abd2c87b1a6553854b
19:37 ylwghst: not at all :-)
19:37 ylwghst: nvidia geforce 320m
19:37 imirkin_: yeah, that's a tesla
19:37 ylwghst: i se
19:38 ylwghst: ok
19:38 imirkin_: GT215 or MCP89?
19:38 ylwghst: MCP89
19:38 imirkin_: (that whole family is called 'tesla', unrelated to the marketing designation for more recent high-end boards)
19:38 ylwghst: didnt know
19:38 ylwghst: ok
19:39 imirkin_: anyways, i'm aware of some extra-special problems on MCP89 which were never diagnosed =/
19:39 karolherbst: ylwghst: but does it at least recover from the error?
19:39 karolherbst: or did it so before as well?
19:39 imirkin_: the whole tesla family is plagued with issues of fifo desync issues, but it's more rare
19:39 imirkin_: whereas the people with MCP89's who hit the issues, hit it all the time
19:40 karolherbst: :/
19:40 karolherbst: messy
19:40 imirkin_: (i call it "fifo desync" but in practice i have no clue what the issue is)
19:40 karolherbst: sadly I only have a MCP79 myself :/
19:41 imirkin_: and this is only *some* MCP89's
19:41 karolherbst: the best kind of issues!
19:41 imirkin_: my guess is that something similar to what was fixed for MCP79's needs to happen for MCP89's
19:41 karolherbst:thinks we should start to blame hardware for random faults, just that nvidia workarounds stuff like that
19:41 karolherbst: ohhh wait
19:41 karolherbst: imirkin_: MCP89 are those weirdo GPUs with hybrid dedicated/sys memory thing, right?
19:42 imirkin_: not sure actually.
19:42 imirkin_: i forget if they steal vram or not
19:42 karolherbst: well, according to wikipedia, they are
19:42 karolherbst: well
19:42 karolherbst: they do both...
19:42 karolherbst: 256MB dedicated, + 1.5GB stolen
19:42 imirkin_: aka GART
19:43 imirkin_: i think they come up with just the 256MB of VRAM
19:43 karolherbst: ylwghst: is this a macbook by any chance?
19:43 karolherbst: ahh, k
19:43 ylwghst: karolherbst: it is
19:43 ylwghst: mid 2010 white
19:43 imirkin_: they're only in macbooks and mac minis. not even sure about the latter.
19:43 karolherbst: meh
19:43 karolherbst: I would blame apple
19:43 imirkin_: that's really the key - come up with someone to blame
19:43 imirkin_: then our work is done :)
19:43 karolherbst: yep
19:44 karolherbst: but I assume nvidia does the right think on those as well
19:44 karolherbst: but mehh
19:44 imirkin_: of course it does
19:44 imirkin_: also knowing what the list of "right things" is is convenient.
19:44 karolherbst: I don't think highly of those apple vbios, because they look super weird usually
19:44 hanetzer: ofc, its apple :P
19:44 karolherbst: ylwghst: maybe a mmiotrace might help
19:44 karolherbst: let me check something
19:45 imirkin_: i think we have traces
19:45 imirkin_: people have come in with issues on those macbooks since the dawn of time
19:45 karolherbst: well
19:45 karolherbst: we have exactly 0 traces
19:45 karolherbst: and one vbios
19:45 imirkin_: lost in the annals of time then
19:45 karolherbst: most likely
19:46 karolherbst: I know I took every tesla+ trace from that mmiotrace gmail account
19:46 ylwghst: what should i buy if i get rid of this?
19:46 imirkin_: something with an AMD and/or Intel GPU :)
19:46 karolherbst: ylwghst: you could help out by creating an mmiotrace of the nvidia driver
19:47 imirkin_: i also suspect blob driver would work on that hw, if that's a palatable option for you
19:47 karolherbst: 304.x of the driver, right?
19:47 imirkin_: 340.x
19:48 karolherbst: I thought that was EOL for fermi
19:48 ylwghst: 340 from nvidia?
19:48 imirkin_: no, for tesla
19:48 imirkin_: 304 was EOL for nv4x
19:48 ylwghst: by
19:48 karolherbst: ohhh
19:48 karolherbst: k
19:48 ylwghst: are you talking about nvidias propertiary driver?
19:48 imirkin_: pretty sure they still support fermi in their latest drivers. just no vk support, i think.
19:48 karolherbst: ylwghst: yeah
19:48 imirkin_: ylwghst: yes.
19:49 ylwghst: i have got working in on centos7
19:49 ylwghst: but its very old kernel
19:49 ylwghst: no luck on 4.9
19:49 karolherbst: ylwghst: newest nvidia drivers work on newest kernel
19:49 ylwghst: ends on black screen when display manager is started
19:49 karolherbst: even 340.104
19:49 karolherbst: mhhh
19:49 karolherbst: intersting
19:49 ylwghst: not on mbp at all
19:49 ylwghst: there is some kind of bug or what
19:49 imirkin_: yeah, we can't really help with the blob driver.
19:50 ylwghst: w8
19:50 ylwghst: on centos7 i have to change pci registers
19:50 karolherbst: *sigh*, well if you get nvidia to run, we could actually get a trace and try to figure out what the hell is going on
19:50 ylwghst: karolherbst: i can
19:50 karolherbst: nice
19:52 ylwghst: viz. https://imgur.com/a/YiBXi
19:52 ylwghst: but
19:52 ylwghst: there was the same problem
19:52 ylwghst: after i compiled nvidia drivers i wasnt able to start x actually it was ending on blanked screen
19:53 ylwghst: i get it work when i loaded
19:53 karolherbst: ohh nice
19:53 karolherbst: well
19:53 ylwghst: setpci module into grub
19:53 karolherbst: getting an mmiotrace on that system should be good enough
19:53 ylwghst: and in grub changed these registers
19:53 ylwghst: insmod setpci
19:53 ylwghst: setpci -s "00:17.0" 3e.b=8
19:53 ylwghst: setpci -s "02:00.0" 04.b=7
19:53 karolherbst: ohh right
19:53 karolherbst: this macbook magic thing...
19:54 ylwghst: but i cant get it work on 4.9
19:54 karolherbst: well, the kernel doesn't really matter
19:54 karolherbst: if you can get an mmiotrace on your old installation, this should help us already
19:54 ylwghst: maybe its work but nix is compiling it by itself
19:54 ylwghst: and i do think there are some patches
19:55 ylwghst: because these registers are already correctly set on nixos
19:55 ylwghst: but it doesnt work
19:55 ylwghst: karolherbst: i can fire it up centos from usb
19:55 ylwghst: fire up
19:56 ylwghst: i actually got rid of centos a long time ago
19:56 pmoreau: karolherbst: Weird Apple VBIOSes? Never heard of those… :-D
19:56 ylwghst: karolherbst: how this could help?
19:56 karolherbst: pmoreau: ....
19:56 karolherbst: ylwghst: an mmiotrace records what nvidia does with the GPU
19:57 karolherbst: and usually this is the best way to increase our knowledge about nvidia gpus
19:57 karolherbst: and if we know more, we can write a better driver
19:57 ylwghst: im tired of it but ill do it if this can help
19:57 ylwghst: ok
19:58 ylwghst: inst nouveau better already ? haha
21:15 Guepi: hi everybody !
21:17 karolherbst: imirkin: .... nvidia doesn't pass spec@arb_shader_image_load_store@layer
21:27 ylwghst: karolherbst: so these errors comes definetely when atom is running
21:28 karolherbst: yeah, as I expected
21:28 karolherbst: bad thing is, issues like that are super hard to track down
21:30 ylwghst: you said it..
21:31 ylwghst: going to boot centos for now
21:37 Guepi: see you later
22:29 scientes: I'm having thermal problems with an optimus laptop
22:29 scientes: is there a way I can throttle down?
22:30 imirkin_: the gpu should auto-suspend if you load nouveau (and don't use the nvidia gpu)
22:31 scientes: i'm using nouveau
22:31 scientes: with wine
22:31 scientes: and the computer panics due to overheat
22:31 imirkin_: reclocking is manual. most recent gpu's come up in the lowest power state
22:31 imirkin_: you can check in pstate
22:31 karolherbst: scientes: sounds like a tesla gpu... or broken fan
22:32 imirkin_: optimus with tesla was rare
22:32 scientes: 01:00.0 VGA compatible controller: NVIDIA Corporation GT218M [NVS 3100M] (rev a2)
22:34 karolherbst: scientes: what does sensors say?
22:35 karolherbst: something around 80°C?
22:35 karolherbst: scientes: also you could try to clock the GPU down through the pstate file
22:35 karolherbst: this should work any maybe even reduce the heat generated
22:36 scientes: between 80 an 100
22:36 scientes: 75 and 100
22:36 karolherbst: scientes: cat /sys/kernel/debug/dri/1/pstate
22:36 karolherbst: as root
22:37 scientes: 03: core 135 MHz shader 270 MHz memory 135 MHz
22:37 scientes: 07: core 405 MHz shader 810 MHz memory 405 MHz
22:37 scientes: 0f: core 606 MHz shader 1468 MHz memory 790 MHz
22:37 scientes: AC: core 108 MHz shader 108 MHz memory 108 MHz
22:37 scientes: nvidia is 0, i915 is 1
22:37 karolherbst: ohh, okay
22:37 karolherbst: the clocks are a little odd though
22:37 karolherbst: scientes: does sensors report a voltage?
22:37 karolherbst: like "GPU core: +0.98 V (min = +0.60 V, max = +1.20 V)" ?
22:37 scientes: no
22:37 karolherbst: meh
22:38 scientes: GPU core: -0.02 V (min = +0.80 V, max = +1.10 V)
22:38 karolherbst: oh wow, this sounds broken
22:38 RSpliet: 108MHz DDR3 I don't think is stable... there's sth very fishy about those values
22:39 karolherbst: yeah, gpu basically suspended
22:39 karolherbst: except it isn't
22:39 scientes: graphics work
22:39 karolherbst: scientes: can you run something with DRI_PRIME on the nvidia gpu and check sensors again?
22:39 scientes: I get this using DRI_PRIME too
22:39 scientes: i'm running wine
22:39 scientes: with 3D game
22:40 karolherbst: mhh, super weird
22:40 imirkin_: so it's actually semi-common for those tesla's to have a phantom VGA output that keeps nouveau thinking that the GPU should be alive
22:40 scientes: I get the same with native 3D game
22:40 karolherbst: mhh, okay
22:40 scientes: wine must set DRI_PRIME
22:41 karolherbst: well the painful part about those Tesla GPUs is, that they run pretty hot
22:42 karolherbst: for whatever reason, but I think nvidia liked the idea of GPUs being potentially able to boil water
22:42 scientes: GT218M [NVS 3100M] <- this is tesla?
22:42 karolherbst: yeah
22:42 scientes: it rebooted once with nouveau disable, just compiling all cores at once
22:42 scientes: this laptop runs hot
22:43 karolherbst: scientes: could you upload the vbios.rom file somewhere?
22:43 karolherbst: it should be near the pstate file
22:46 scientes: can i email it to you, I don't have a vpn right now
22:48 scientes: back
22:49 karolherbst: scientes: you can send it to karolherbst@gmail.com if you want
22:54 karolherbst: okay, so that GPU _does_ have voltage control
22:55 karolherbst: scientes: as root, echo 03 > pstate. This _might_ hang your GPU, but maybe it will be alright. There is something odd going on, but not as odd as it actually look like. Most likely
22:59 karolherbst: RSpliet: the -0.02V is -EINVAL, because its vbios only defines 0x0, 0x2, 0x4, 0x6... VIDs voltages... and I am super sure a weird VID is selected. The clocking state? whatever.. tesla is weird
23:01 karolherbst: or maybe we indeed don't implement reclocking exactly right on nva8
23:01 karolherbst: or old kernel and I fixed those things? dunno
23:19 ylwghst-centos: Hey all