00:04 imirkin: fling: oh also, to get accel you need nvidia's firmware, available in the linux-firmware package.
02:08 imirkin: skeggsb: looks like people are still having trouble with the acpi fast read routine =/
02:30 gnurou: mupuf: weird - I am seeing the clk 127 message, but apart from that it boots fine on my end
03:00 imirkin: skeggsb: have a look at my analysis in https://bugs.freedesktop.org/show_bug.cgi?id=98240#c12 - not 100% sure what to do about that one. maybe we should only allow the fast method if the checksum matches?
03:07 skeggsb: fuck me, i hate bios manufacturers
03:08 imirkin: they basically carefully analyze our heuristics and come up with ways to break them
03:08 imirkin: so far, with great success
03:09 skeggsb: wonder if it isn't easier to drop it completely.. though, i'd not enjoy that, it takes in the order of seconds to fetch on my laptop without it
03:09 imirkin: well, your call. acpi checksums tend to check out, so if it fails, we could just hard-fail that method
03:09 skeggsb: true
03:10 imirkin: that methods array is a bit confusing, but how is it working without a correct checksum in the first place?
03:12 imirkin: should only be getting a score of 3 at most for a failed thing, no?
03:13 imirkin: at which point it should retry the slow acpi
03:13 skeggsb: yeah, except the acpi rom has a second image which scores 4 on top of the 2 the first image gets
03:14 imirkin: of course it does.
03:14 skeggsb: i'm going to add a "require_checksum" to acpi and make it bail
03:15 imirkin: worksforme
03:17 skeggsb: there we go...
03:27 imirkin: cool that seems like it should work
05:13 mupuf: gnurou: can you send me your .config?
05:14 gnurou: mupuf: I have tried with both tegra_defconfig and multi_v7_defconfig from vanilla 4.8
05:14 mupuf: ok, I will try this too
05:24 mupuf: gnurou: well, I guess I should not have kept my .config
05:24 mupuf: using the def_config was what was needed
05:24 gnurou: mupuf: good! I was surprised by your report, since we have our own test farm for Tegra boards
05:25 gnurou: any regression should have been caught by it
05:25 mupuf: I see
05:25 mupuf: well, one just needs to know the workflow ;)
05:46 mupuf: gnurou: looks like the issue with the fence is indeed gone!
05:46 gnurou: mupuf: good news!
05:48 mupuf: making a full piglit run
05:48 mupuf: we'll see how it goes
06:07 mupuf: well, looks like there are quite a few regressions
06:07 mupuf: now, if only it did not hang the gpu entirely
06:36 mupuf: spec@arb_arrays_of_arrays@execution@atomic_counters@ tests are busted, it would seem
14:40 karolherbst: pmoreau: mind trying out the nouveau-staging thing over the weekend? I would do an update today. I think it might be good to have it now, cause there won't be any updates for 4.9 afaik and this would make it more painless for users to use a custom kernel with updated nouveau
14:41 karolherbst: ohh wait, there is only module out of tree support in there....
14:41 karolherbst: maybe I will add full kernel support when I'Ve got some time, allthough I've already finished it for travis-ci
14:42 karolherbst: nvm then
15:23 ajax: imirkin: re 98030... huh. iiiiinteresting.
15:24 ajax: at least that's a plausible commit to introduce that kind of bug
15:29 imirkin_: ajax: and sounds like nouveau's not the only one with issues
15:29 imirkin_: which makes me happy, since i'm clearly not well placed to figure this out
15:35 tobijk: imirkin_: ping about the little OP_SPLIT addition patch, wasn't it fine the last time you checked? :D maybe we can commit it before branching (today, tomorrow?)
15:36 imirkin_: tobijk: it's in my tree... i'll try to make sure to remember to push it out tonight
15:36 tobijk: ok
18:19 NanoSector: doesn't nouveau and specifically the nvc0 driver in Mesa use gallium?
18:22 imirkin_: i'd qualify it in the opposite direction
18:22 imirkin_: nvc0 implements the gallium API
18:22 NanoSector: ah right
18:23 NanoSector: i was wondering because everyone speaks about gallium only being profitable for AMD graphics
18:23 imirkin_: "everyone"?
18:23 NanoSector: everyone i've heard :P
18:23 karolherbst: uhhh
18:23 imirkin_: sounds like you're in an echo chamber
18:23 NanoSector: lol
18:23 karolherbst: they should stop being fanboys, seriously
18:24 imirkin_: drivers for adreno, vc4, and vivante gpu's also implement the gallium api's
18:24 karolherbst: I don't think anybody is able to lit all the gallium based drivers we actually have :D
18:24 NanoSector: karolherbst: now that you're here, any way i can feed linux 4.8 the reclocking patches for kepler?
18:24 karolherbst: *list
18:25 karolherbst: NanoSector: I am running 4.7 with current master
18:25 imirkin_: karolherbst: whenever you add a new cap, you have to modify all the drivers
18:25 imirkin_: karolherbst: so i actually know the list fairly well :)
18:25 karolherbst: imirkin_: yeah, but you have the drivers directory
18:25 NanoSector: karolherbst: mmkay, arch updated to 4.8 and your patches don't work with it
18:25 karolherbst: so you don't actually have to remember
18:25 imirkin_: no, but eventually you pick it up
18:25 karolherbst: NanoSector: right, cause I am lazy to move to 4.8
18:25 imirkin_: after you've added enough caps
18:25 karolherbst: I see
18:25 NanoSector: karolherbst: if i give you a cookie?
18:25 Yoshimo: NanoSector: maybe they talk about gallium-nine, which with nouveau isn't yet that good because of reclocking issues with nvidia cards
18:25 karolherbst: a cookie won't help
18:26 NanoSector: karolherbst: two cookies? :P
18:26 karolherbst: ...
18:26 NanoSector: Yoshimo: makes sense
18:26 karolherbst: I already have it installed anyway
18:26 NanoSector: lol
18:26 karolherbst: yeah maybe I rebase my stuff today
18:27 NanoSector: aight let me know if you want me to test anything
18:27 karolherbst: I test by using the driver
18:28 NanoSector: fair enough
18:28 NanoSector: i rather meant in the reclocking area since i was having issues with that
18:29 karolherbst: I think for whatever reason I run the noop disc scheduler and it makes everything slow :/
18:32 karolherbst: mupuf: by the way, we have a SENSORS_INA3221 sensor thing in the kernel now
18:32 karolherbst: actually we have SENSORS_INA2XX as well
18:32 karolherbst: ever looked into those?
18:44 karolherbst: NanoSector: well, what issues did you had again?
18:44 NanoSector: my GT750M freezing seconds after a reclock
18:44 karolherbst: Still have other things on my todo list, so depending on the issue I might be able to do something or not
18:44 karolherbst: mhhh
18:44 karolherbst: no idea really about that
18:45 karolherbst: mind doing a mmiotrace of nvidia?
18:45 karolherbst: maybe it is one of those things we can't handle yet
18:45 karolherbst: skeggsb: you still didn't answer my mail :p
18:45 karolherbst: NanoSector: ohh you already did it
18:45 karolherbst: I guess it is fine then
18:45 NanoSector: yeah
18:45 karolherbst: whenever I get time digging into this
18:46 NanoSector: yeah no rush
18:46 karolherbst: those things are rather hard to debug cause I can't test the hw
18:47 karolherbst: crazy, I get nearly a 2MB/s down link through tor... the hell happend, I thought it was like 100kb/s slow or something
18:49 karolherbst: NanoSector: do you think it might be possible that you could provide access to your machine at some point? Maybe I would like to try things out
18:49 karolherbst: though I can't say when
18:50 NanoSector: karolherbst: yeah, that's fine, though it'll have to be remote I think
18:50 NanoSector: if it's in the weekends i can probably attend
18:51 imirkin_: NanoSector: are you using skeggsb's master branch?
18:51 NanoSector: nope i'm using whatever is in the linux 4.8 kernel
18:51 NanoSector: I could try out the skeggsb branch though
18:51 imirkin_: NanoSector: ok, you need to either use ben's branch or karol's
18:52 imirkin_: that has important fixes for reclocking
18:52 NanoSector: yeah I was using karol's on 4.7 but that didn't fix it
18:52 karolherbst: NanoSector: well not much changed, so if you tried my branch at one point and it had issues, chances are slim something changed
18:52 karolherbst: though
18:52 karolherbst: I did improve things
18:52 karolherbst: so maybe
18:52 karolherbst: ohh right
18:52 karolherbst: yeah, then don't bother
18:52 karolherbst: depends on what branch though
18:52 NanoSector: kepler-improved-reclocking or something
18:52 karolherbst: well
18:52 karolherbst: more interested in the vx part
18:53 NanoSector: vx?
18:53 karolherbst: v5 or v2 or whatever
18:53 NanoSector: oh
18:53 NanoSector: stable_reclocking_kepler_v6
18:53 karolherbst: k
18:53 imirkin_: NanoSector: and that didn't fix it? then you're SOL
18:53 karolherbst: imirkin_: we know that we don't handle some memory bits right
18:53 karolherbst: skeggsb knows which one, but he never told me :(
18:54 imirkin_: karolherbst: sure, but without testing the latest fixes first it's not really worth discussing
18:54 NanoSector: imirkin_: so wontfix?
18:54 imirkin_: NanoSector: no, just wontwork for now ;)
18:54 NanoSector: ah
18:54 karolherbst: right
18:54 karolherbst: I have a trace though, so _maybe_ I find something
18:54 karolherbst: sometimes things are rather obvious, but sometimes not
18:54 tobijk: oh god, my distri patched its mesa with imirkins locking branch...
18:54 imirkin_: ahahahahaha
18:55 imirkin_: that's hilarious.
18:55 karolherbst: NanoSector: do you know if it worked with config=NvMemExec=0 ?
18:55 NanoSector: yeah i provided karol with some logs
18:55 karolherbst: tobijk: :D
18:55 imirkin_: does your nouveau now get reported as noouveau?
18:55 NanoSector: karolherbst: i think we prodded around with multiple config files but none worked
18:55 NanoSector: s/config files/config options/
18:55 tobijk: imirkin_: i'm not stupid enough to use that version, i have my own :)
18:55 karolherbst: NanoSector: sure, but you should try NvMemExec at least once again to verify
18:55 imirkin_: well, did they include that bit in their patchset?
18:56 NanoSector: karolherbst: i can't run your branch on 4.8 though, i'd have to install 4.7 again
18:56 imirkin_: tobijk: which distro, so i know to disregard bug reports?
18:56 NanoSector: oh wait i can just pull it from the arch linux archive :x
18:56 imirkin_: (that branch is, sadly, terminally broken)
18:56 tobijk: suse
18:56 karolherbst: :D
18:56 imirkin_: tobijk: ok. does that propagate to OpenSUSE and whatnot?
18:56 karolherbst: that will be fun
18:56 imirkin_: i haven't used SuSE since ... 1998?
18:56 karolherbst: how do they even thing that this is a good idea?
18:57 tobijk: imirkin_: they have picked some patches, dunno havent look closely enough
18:57 imirkin_: probably coz kde is fubar'd
18:57 karolherbst: mhh
18:57 karolherbst: true
18:57 karolherbst: well
18:57 karolherbst: more testing I guess
18:57 imirkin_: the proper solution, though, is "stop shipping nouveau"
18:57 tobijk: http://hastebin.com/elisuzedod.cs
18:57 imirkin_: not "take patches written by someone who says they should not be used"
18:57 karolherbst: imirkin_: well if something is broken you can just break it differently, to make it actually more usable :p
18:58 imirkin_: it'll deadlock most likely.
18:58 karolherbst: ...
18:58 karolherbst: 65?
18:58 karolherbst: that is insane
18:58 tobijk: there are gaps :P
18:58 karolherbst: right
18:58 karolherbst: commercial distro got those
18:58 karolherbst: :p
18:58 imirkin_: not all those patches are from me
18:58 imirkin_: they have some additional fixes on top
18:59 imirkin_: N_05-Use-nv50_render_condition-in-nv50_blitctx_post_blit.patch
18:59 imirkin_: i wonder wtf that is
18:59 karolherbst: I am sure they tested this in depth and think it is actually better
18:59 imirkin_: lol
18:59 karolherbst: tobijk: you could ask them to apply my patches to the kernel :p
18:59 tobijk: karolherbst: no thanks
18:59 karolherbst: :D
18:59 imirkin_: i guess i should rename that branch to "locking-broken-dont-use" or something
19:00 NanoSector:downgrades Linux
19:00 tobijk: not because they are your patches, but because they are unstable
19:00 karolherbst: imirkin_: good idea
19:00 karolherbst: imirkin_: or maybe you should just add patches on top, which say "imrpove overall performance by 20%" but it simply breaks everything
19:00 tobijk: imirkin_: https://build.opensuse.org/package/view_file/X11:XOrg/Mesa/N_05-Use-nv50_render_condition-in-nv50_blitctx_post_blit.patch?expand=1
19:00 imirkin_: oh, so it is opensuse.
19:01 tobijk: yeah but its the same
19:01 imirkin_: ok cool. i'll be sure to ignore any reports from them.
19:01 tobijk: ^^
19:01 karolherbst: :D
19:01 karolherbst: this will be fun
19:01 imirkin_:hopes he'll remember
19:01 karolherbst: you know what will be also fun?
19:01 Manoa: what about loading a clocking-modified bios into fermi ? would that work by loading the custom bios ?
19:01 karolherbst: if they expose OpenGL 4.5 :D
19:01 imirkin_: that's their problem.
19:01 karolherbst: Manoa: nope
19:02 karolherbst: Manoa: you know what would work? load nvidia, let it clock high and switch to nouveau (maybe)
19:02 imirkin_: Manoa: if you can get a vbios that by default clocks to a higher rate, that should work just fine with nouveau
19:02 karolherbst: imirkin_: how should that work out?
19:02 imirkin_: karolherbst: same way it works for the low rate...
19:02 karolherbst: especially since all those GPIOs are different and so
19:02 karolherbst: yeah well
19:02 tobijk: i can take this, i'll post a removal patch to them
19:02 karolherbst: ...
19:02 karolherbst: tobijk: don't bother
19:03 tobijk: its my distri, i use mostly i want a wotking one
19:03 karolherbst: they should learn not to do silly things
19:03 Manoa: do I have to flash the bios into the card or I can load the modified from linux after I modify the bios file ?
19:03 karolherbst: then report actual issues
19:03 imirkin_: the issue is well-known but has no fix
19:03 karolherbst: Manoa: it ain't that simple
19:03 imirkin_: their recourse is to stop shipping kde (unlikely) or stop shipping nouveau (likely)
19:04 karolherbst: I doubt anybody wants to stop shipping nouveau
19:04 imirkin_: imo it's the only responsible choice
19:04 tobijk: imirkin_: may i add some lines from irc here to the removal request?
19:04 karolherbst: sure
19:04 karolherbst: but nobody would do that
19:04 imirkin_: tobijk: this channel is publicly logged. use whatever you like.
19:05 tobijk: it is logged but nobody will find it anyway :D
19:05 karolherbst: sure
19:05 karolherbst: it is in the message
19:05 imirkin_: my point is that there's no expectation of privacy
19:05 karolherbst: tobijk: actually google indexed the logs
19:05 karolherbst: just saying
19:05 imirkin_: (as opposed to something i might say in a PM, in which case i might not expect it to be sent to the nytimes)
19:05 tobijk: karolherbst: if you know what you are looking for, yes
19:06 imirkin_: to clarify - i mean not ship nouveau_dri.so
19:06 Manoa: karolherbst my card has a bios switch I can have 2 bios in the card and switch between them, whay it not so simple, what the problem ?
19:06 karolherbst: imirkin_: sure
19:07 karolherbst: Manoa: because the vbios is made for your GPU model
19:07 imirkin_: one could also try patching kde to end the insanity... but i don't use kde, so i can't really bring myself to care too much about it
19:07 karolherbst: Manoa: actual model, not "name" or anything that silly
19:07 tobijk: dont know, kde is working fine for me
19:07 karolherbst: imirkin_: it is mostl likely within qt5 anyway
19:07 Manoa: yhe I know, there is a program fermi bios editor which load the bios from the card and can modify him
19:07 imirkin_: tobijk: it's not literally kde
19:07 imirkin_: tobijk: it's some component ... qtwebkit? something like that
19:07 karolherbst: Manoa: mhhh, you could try it if the editor is good enough
19:07 imirkin_: except not webkit
19:07 imirkin_: something web-related
19:08 tobijk: yeah qtwebkit is *stupid*
19:08 tobijk: but thats qt :D
19:08 imirkin_: qtwebengine?
19:08 karolherbst: Manoa: you can boot with nouveau.config=NvBios=...path...
19:08 imirkin_: [and you'd have to force nouveau to re-run the vbios]
19:08 Manoa: that's nice :)
19:08 karolherbst: but the file has to be inside initrd most liekly
19:08 imirkin_: it has to be available when nouveau loads
19:08 karolherbst: imirkin_: NvForcePost?
19:08 imirkin_: yea
19:09 karolherbst: nouveau.config=NvBios=...path...,NvForcePost=1 ;)
19:09 karolherbst: but mhhh
19:09 imirkin_: unless it's a secondary gpu, or for whatever other reason your gpu isn't initialized by the bios
19:09 karolherbst: if the editor is _that_ smart
19:09 karolherbst: it might be actually helpfull for us :D
19:10 karolherbst: well I doubt it will work though
19:11 imirkin_: Manoa: generally those things modify things like the available perf levels
19:11 imirkin_: Manoa: not the actual at-boot-time-initialization
19:11 karolherbst: yeah, that's what I think too
19:11 imirkin_: this works nicely for running on nvidia blob, which reads those tables and does appropriate things
19:12 Manoa: I will test this and see what hapen xD
19:14 NanoSector: karolherbst: i might actually have been testing kepler-stable-reclocking-v5 with you and was trying v6 on 4.8 which obviously didn't work, i'll try v6 now on 4.7 to be sure
19:14 karolherbst: NanoSector: k
19:14 NanoSector: it's building
19:14 karolherbst: NanoSector: also disable memory reclocking, just to be sure
19:14 NanoSector: mkay
19:14 imirkin_: NanoSector: btw, nouveau works fine for you right? just reclocking is broken?
19:15 NanoSector: yeah it does
19:15 imirkin_: kk, just checking
19:15 NanoSector: karolherbst: i had to add nouveau.NvMemExec=0 to the kernel command line right?
19:15 karolherbst: NanoSector: yeah
19:15 NanoSector: aight will do
19:15 karolherbst: I just want to know if it is memory related or engine related
19:16 karolherbst: if you reclock just the engine, you need a good benchmark though
19:16 karolherbst: like pixmark_piano
19:16 karolherbst: it doesn't care about memory clock at all
19:16 karolherbst: and just puts 100% on the engines
19:16 karolherbst: good way of testing actual enginge reclocking stability too
19:16 NanoSector: do you happen to know in which package that is?
19:16 karolherbst: gputest
19:16 NanoSector: ok thanks
19:19 NanoSector: wut there was no nouveau.ko.gz in my modules directory
19:20 pmoreau: karolherbst: I can have a look: I should have some time this week-end to work on Nouveau! :-) There is both drm-next + oot Nouveau, plain drm-next and latest Linux release on the image.
19:20 pmoreau: bbl
19:20 tobijk: oh wow a no-maintainer actually pushed those changes to the Mesa packages ~_~
19:20 karolherbst: "no-maintainer"=
19:20 karolherbst: ?
19:20 tobijk: non
19:20 karolherbst: uhhhh
19:21 karolherbst: guess somebody is loosing his push rights
19:21 tobijk: hehe
19:21 tobijk: susse emplyee *shrug*
19:21 karolherbst: yeah well
19:21 karolherbst: should be happy to keep the job :p
19:25 NanoSector: stable-reclocking-v6 hangs my entire system after modprobing nouveau
19:26 karolherbst: uhhhh
19:26 NanoSector: okt 14 21:24:09 rick-laptop kernel: BUG: unable to handle kernel NULL pointer dereference at 0000000000000018
19:26 karolherbst: right
19:26 NanoSector: okt 14 21:24:09 rick-laptop kernel: IP: [<ffffffffa0611005>] led_classdev_suspend+0x5/0x20 [led_class]
19:26 karolherbst: there is a bug
19:26 karolherbst: I didn't update my branch :(
19:26 karolherbst: you can user the master-4.7 branch though
19:26 NanoSector: mhmm
19:26 NanoSector: does that include your patches?
19:26 karolherbst: yeah
19:26 karolherbst: not all, but the important ones
19:26 NanoSector: will do
19:28 pmoreau: Speaking of KDE, I was supposed to test the combo KDE + Factorio, wasn’t I? I’ll do that later.
19:28 karolherbst: imirkin_: "Fixes (boo#997171) as suggested in (fdo#91632)." ;)
19:29 karolherbst: " N_04-* and N_05-* include untested fixes for nv50." :D cause that is what you do! :D
19:30 tobijk: imirkin_: you should not confude people: https://bugs.freedesktop.org/show_bug.cgi?id=91632#c8 :D
19:30 tobijk: *confuse
19:30 tobijk: karolherbst: i guess those coming from the actual suse repos which are hidden :)
19:31 NanoSector: you're all wizards to me, i'm confused as soon as i enter here
19:32 NanoSector: doesn't freeze now
19:37 NanoSector: 0a pstate still working...
19:37 karolherbst: NanoSector: maybe we actually are
19:37 NanoSector: 0f also still working
19:38 karolherbst: meh, so it is a memory problem
19:38 karolherbst: are you running pixmark?
19:38 NanoSector: yes, pixmark_piano as you said
19:38 karolherbst: k
19:38 karolherbst: well
19:38 NanoSector: it's running in pstate 0f now
19:38 karolherbst: you can leave the memory as it is and get whatever boost you can this way
19:38 karolherbst: but
19:38 karolherbst: besides pixmark, nothing else will be much faster
19:38 NanoSector: :(
19:39 karolherbst: maybe I find something silly
19:39 karolherbst: but you said it doesn't hardlock immediatly?
19:39 karolherbst: mhh
19:39 karolherbst: wondering if the plls are messed up again, wouldn't be the first time
19:39 NanoSector: now it doesn't hardlock at all
19:40 NanoSector: but if I used it with dolphin with memory reclocking it'd lock after like a second or two after I executed the command
19:40 karolherbst: ohh wait
19:40 NanoSector: (dolphin the emulator, not the other 5000 products with the same name)
19:40 karolherbst: NanoSector: is this a ddr3 one?
19:40 NanoSector: yup
19:40 karolherbst: with 2GHz?
19:40 NanoSector: 4096
19:40 NanoSector: MB
19:40 NanoSector: uhhh
19:41 NanoSector: yes
19:41 karolherbst: uhhh
19:41 NanoSector: 0f would reclock to 2002 MHz
19:41 karolherbst: that is quite fast
19:41 karolherbst: and 0a?
19:41 NanoSector: 1600 MHz
19:41 karolherbst: k
19:41 karolherbst: and 0a also hard locks?
19:41 NanoSector: iirc yes
19:41 NanoSector: i'll try again
19:41 karolherbst: did you reclock to 07 once?
19:41 karolherbst: no idea about the state of ddr3 based cards though
19:42 karolherbst: there might be silly bugs
19:42 NanoSector: i always first echo 07 before reclocking higher
19:42 karolherbst: mhh
19:42 karolherbst: if 07 works, we might do something silly like we did for gddr5 once
19:42 karolherbst: maybe, maybe not
19:42 NanoSector: though without your patches it'd hardlock if i even pushed 07
19:42 karolherbst: sure
19:42 karolherbst: due to undervolting
19:42 NanoSector: ah
19:43 karolherbst: mupuf: do you have any ddr3 based kepler gpu?
19:45 NanoSector: yeah it's still running now with pstate 0f
19:51 NanoSector: huhh
19:51 NanoSector: well at least with pixmark piano it doesn't freeze if i reclock to 0a, icnluding memory
19:51 karolherbst: guess why
19:51 NanoSector: it doesn't use memory
19:51 karolherbst: exactly
19:51 NanoSector: but if i cat pstate it says that memory is running at 1600 MHz as well
19:52 karolherbst: right
19:52 karolherbst: run furmark ;)
19:52 karolherbst: will lock immediatly
19:53 NanoSector: i uh
19:53 NanoSector: ran Dolphin
19:53 NanoSector: and
19:53 NanoSector: it works...
19:53 NanoSector: i don't have furmark atm
19:53 karolherbst: it is part of gputest
19:54 NanoSector: gputest -h doesn't list it
19:54 karolherbst: aybe they call it fur internally
19:54 karolherbst: no idea
19:54 NanoSector: but dolphin is an app which would immediately hang when reclocking as well (I used to test with it before)
19:54 karolherbst: there are shell scripts
19:54 karolherbst: NanoSector: it might be, that we do less silly things on 0a
19:54 karolherbst: you could try 0f
19:54 NanoSector:covers head in case of exploding GPU
19:55 NanoSector: you never know what happens now with exploding phones and washing machines
19:55 karolherbst: I know we can't destroy those nvidia gpus...
19:55 karolherbst: no idea why
19:55 karolherbst: but I don't know any nouveau dev was ever able to brick one
19:56 NanoSector: hmm it worked
19:56 NanoSector: then the game loaded a new stage
19:56 NanoSector: and it hung
19:57 NanoSector: [ 506.384455] nouveau: Emuthread - Sta[21294]:00000000:0000a06f: detach gr failed, -16
19:57 NanoSector: [ 506.384457] nouveau: Emuthread - Sta[21294]:beef323f:0000a040: suspend failed with -16
19:58 NanoSector: I guess there's at least progress?
19:59 karolherbst: no
20:00 NanoSector: uh so my cpu cores started locking up one by one
20:00 NanoSector: oh furmark was test=fur indeed
20:00 imirkin_: tobijk: how was i confusing people?
20:01 imirkin_: tobijk: i stand behind my statement in comment 8 on that bug
20:02 NanoSector: karolherbst: yes you were right, instant hang on furmark
20:02 NanoSector: yay for being optimistic though :P
20:02 karolherbst: :D
20:03 karolherbst: why is fuse so crappy by the way?
20:03 karolherbst: ...
20:03 igmp: What makes the furmark test so complex to render?
20:03 karolherbst: nothing
20:03 NanoSector: fur
20:03 karolherbst: it just uses your memory bandwidth for real
20:03 karolherbst: like for real real
20:04 NanoSector: oh good lord running it on intel causes my DE to glitch out like everywhere :P
20:04 karolherbst: jo
20:04 karolherbst: run pixmark_piano
20:04 karolherbst: more fun
20:05 NanoSector: the laaaggg
20:05 NanoSector: it's real
20:05 NanoSector: :P
20:05 karolherbst: really makes you think if running everything on the gpu is a good idea :D
20:05 NanoSector: haha
20:05 NanoSector: fun stuff
20:06 NanoSector: anyway, we pinpointed it to the memory, right?
20:06 karolherbst: yeah
20:06 karolherbst: so 0a was also bad?
20:06 karolherbst: ohh
20:06 NanoSector: 0a on furmark froze it
20:06 karolherbst: and did you test 07?
20:06 NanoSector: 07 worked
20:06 karolherbst: I see
20:06 igmp: karolherbst: I'll try when I'll be on sector
20:07 karolherbst: mhhh
20:07 karolherbst: yeah I have an idea what is wrong
20:07 karolherbst: I just need to try things
20:08 NanoSector: we were poking at voltages before
20:08 NanoSector: that i remember
20:08 karolherbst: NanoSector: do you have envytools installed?
20:08 NanoSector: nope
20:09 karolherbst: you should
20:09 NanoSector: i will
20:09 karolherbst: then, switch to 0a and do nvapeek 0x132024
20:09 igmp: btw I want to try writing a gallium pipe driver for fun. But I lack of information about it. Already tried gallium and mesa website for doc. Any pointers?
20:09 NanoSector: installin'
20:09 karolherbst: igmp: I think #dri-devel would be the better channel to ask though
20:10 NanoSector: i suppose i run nvapeek as root
20:10 NanoSector: ?
20:10 karolherbst: isn't there some form of gallium docs?
20:10 karolherbst: NanoSector: yes
20:10 NanoSector: ok, off we go
20:10 igmp: karolherbst: already asked there and failed to get an answer ^^
20:11 karolherbst: bother them
20:11 karolherbst: they should fell guilty to not write any documentation
20:11 NanoSector: 00132024: 00011d01
20:11 karolherbst: and if there are docs, you should fell guilty not to read them
20:11 karolherbst: NanoSector: on 0a?
20:11 NanoSector: on 0a
20:11 karolherbst: mhh k
20:11 NanoSector: it took a long time to run
20:12 karolherbst: you have runpm=0 set, right?
20:12 NanoSector: noo
20:12 karolherbst: :D
20:12 karolherbst: k, there you go
20:12 NanoSector: lol
20:12 igmp: karolherbst: k, I'll just bother them more
20:12 NanoSector: karolherbst: it's still updating furmark frames, just reaaaallllyyyy slowly
20:12 tobijk: imirkin_: you sound a bit too optimistic there, but i guess you just meant to say ("it works mostly with qtwebengine"), anyway never make people use unstable software or they will ship it ;-)
20:12 karolherbst: NanoSector: ohh well
20:12 NanoSector: like, 1 frame per minute
20:13 karolherbst: you shouldn't run anything while doing that
20:13 karolherbst: tobijk: it's their fault
20:13 NanoSector: but my gpu won't reclock at all when nothing is running on it
20:13 NanoSector: because optimus
20:13 karolherbst: you don't pick patches from random branches except the dev says: this is fine and ready for use
20:13 imirkin_: tobijk: no, it mostly resolves the problem. but ultimately it's the wrong solution.
20:13 karolherbst: NanoSector: runpm=0
20:13 NanoSector: oh, derp
20:13 imirkin_: tobijk: because it can never completely resolve the problem
20:14 NanoSector: how do i chain multiple config options?
20:14 mupuf: karolherbst: eww, I don't think I have a ddr3-baed kepler gpu
20:14 karolherbst: NanoSector: you just do, it isn't part of config though
20:14 karolherbst: mupuf: crap
20:14 karolherbst: :(
20:14 imirkin_: karolherbst: i have a GK208 with DDR3...
20:14 imirkin_: works fine
20:14 NanoSector: oh ok, rebootin'
20:14 karolherbst: imirkin_: memory clock is like 1.4GHz?
20:14 imirkin_: 1.8ghz iirc
20:14 karolherbst: I see
20:14 tobijk: karolherbst: gk107 ddr3, works fine
20:14 imirkin_: i don't have it in front of me though
20:14 karolherbst: k
20:14 karolherbst: well
20:15 karolherbst: maybe it is just a silly random thing we do wrong
20:17 imirkin_: [actually it's a GK208B, i.e. NV106]
20:17 NanoSector: without anything running: 00132024: 00021d01
20:18 karolherbst: right
20:18 karolherbst: you need to go to 0a
20:18 NanoSector: and runpm=0 and NvExecMem=0
20:18 NanoSector: i'm in 0a
20:18 karolherbst: ...
20:18 karolherbst: well
20:18 NanoSector: did i fuck up again
20:18 karolherbst: you need to enable memory reclocking again
20:18 NanoSector: :P
20:18 NanoSector: the dots of fuckup
20:20 karolherbst: I am sure 0x10f824 is wrong
20:21 NanoSector: 00132024: 00011d01
20:21 NanoSector: with memory reclocking
20:21 karolherbst: peek 0x10f824
20:21 karolherbst: uhmmm, this looks wrong? "ram_mask(fuc, 0x1373f4, 0x00000000, 0x00010010);"
20:21 NanoSector: 00132024: 00011d01
20:22 karolherbst: NanoSector: ;) right copy/paste pls
20:22 NanoSector: derp
20:22 NanoSector: 0010f824: 23829fe7
20:22 NanoSector: helps if you actually press copy
20:22 karolherbst: huh
20:22 karolherbst: that looks right though
20:22 karolherbst: nvapeek 0x1373f4
20:23 NanoSector: 001373f4: 00000011
20:23 karolherbst: 0x132030
20:24 karolherbst: 0x132034
20:24 NanoSector: 00132030: 04260003
20:24 NanoSector: 00132034: 0b340426
20:24 karolherbst: okay, 0x132034 looks like total crap :p
20:24 karolherbst: k... let check others
20:24 NanoSector: it all looks random to me
20:24 NanoSector: :P
20:25 karolherbst: 0x10f65c
20:25 karolherbst: 0x10f824
20:25 karolherbst: ohh wait
20:25 karolherbst: ignore the last one
20:25 NanoSector: 0010f65c: 00000022
20:26 NanoSector: i already did that last one right?
20:26 karolherbst: 0x10f210
20:26 NanoSector: 80003838
20:26 karolherbst: k
20:27 karolherbst: nvapoke 0x132034 0xff690426; nvapoke 0x10f210 0x80006565
20:27 karolherbst: both with sudo
20:27 karolherbst: or as root
20:27 karolherbst: or whatever
20:28 karolherbst: and then cat pstate allthough it shouldn't change the memory clock
20:28 karolherbst: and then run furmark
20:29 NanoSector: black screen
20:29 karolherbst: sad :(
20:29 karolherbst: well
20:30 karolherbst: it isn't suppose to work this way anyway, cause you have to do other things
20:30 karolherbst: but meh
20:30 karolherbst: guessing around won't help
20:30 NanoSector: it's doing something as my fans are kicking up very loudly
20:30 NanoSector: though dmesg is not happy
20:30 karolherbst: sure, memory messed up
20:31 NanoSector:warms face
20:32 NanoSector: so any idea what's wrong?
20:33 karolherbst: nope
20:33 NanoSector: lol
20:33 NanoSector: clocking it back to 07 makes stuff work again
20:33 karolherbst: mhhh
20:33 karolherbst: interesting
20:34 karolherbst: usually it doesn't, right?
20:34 NanoSector: usually it hangs completely if i try to reclock it in this state
20:34 NanoSector: *reclock it any more
20:34 karolherbst: mhhh
20:34 NanoSector: like the 'echo xx > pstate' hangs
20:34 karolherbst: maybe those pokes weren't as bad
20:38 NanoSector: i kind of like runpm=0
20:38 NanoSector: makes for a nice lap warmer
20:56 karolherbst: hakzsam, imirkin_: ping on my const folding patches
20:56 imirkin_: what did they do again?
20:56 karolherbst: limm form of fmad for gf100+
20:57 imirkin_: didn't you have some serious issues i pointed out?
20:57 NanoSector: that's rude
20:57 NanoSector: :P
20:57 karolherbst: imirkin_: which ones was it?
20:59 karolherbst: I know my branch elim pass had serious issues
20:59 karolherbst: and there was this one RA thing with the cse of phi nodes thing
20:59 karolherbst: but I don't think you said something serious about the fmad thing
21:18 imirkin_: karolherbst: iirc you were doing crazy things to gm107
21:19 karolherbst: uhhh, let me check
21:20 karolherbst: I know there was the comment from hakzsam about splitting it into two functions, but it isn't done for other instructions as well, where someone might say this
21:20 karolherbst: but besides that? https://github.com/karolherbst/mesa/commit/25088d80106868d64bd66956ee2794c0935a9e33
21:20 karolherbst: looks fine to me
21:21 imirkin_: karolherbst: only that change? that's probably fine, will have to look a little carefully
21:21 imirkin_: probably an assert to make sure src2 == dst
21:21 karolherbst: that's the ohly gm107 related change
21:21 karolherbst: there actually is an assert
21:21 imirkin_: oh, wait, you have that ;)
21:21 karolherbst: same for gk110
21:22 imirkin_: you were making some changes to the target iirc
21:22 imirkin_: trying to redefine when the limm thingw as used, or something like that
21:22 karolherbst: imirkin_: https://patchwork.freedesktop.org/series/13485/
21:22 imirkin_: in a different series apparently
21:22 karolherbst: uhh right
21:22 karolherbst: yeah
21:22 karolherbst: that was a different thing
21:22 imirkin_: i got confused :)
21:22 karolherbst: and I was wrong with that
21:22 karolherbst: right
21:23 imirkin_: + // TODO: gm107 can also do this for S32
21:23 imirkin_: pretty sure all gf100+ can
21:23 imirkin_: not 100% sure though
21:23 karolherbst: might be
21:23 imirkin_: on nv50, there is no IMUL operation
21:23 karolherbst: but envydis only contains the gm107 stuff
21:23 imirkin_: (at least not like that)
21:24 karolherbst: and I wasn't able to find any case where this might be hit anyway
21:24 karolherbst: so I couldn't test it
21:24 imirkin_: + // prior gk110 we can't do that if we have a predicate
21:24 imirkin_: errrr
21:24 imirkin_: what?
21:24 karolherbst: right :D
21:24 karolherbst: gf100 has no predicate for fmad limm
21:24 karolherbst: afaik
21:24 imirkin_: uhhhh
21:24 karolherbst: maybe I am silly though
21:25 imirkin_: pretty sure every op on gf100+ has the same predicate fields reserved
21:25 karolherbst: { 0x2000000000000002ull, 0xf800000000000007ull, N("fma"), T(ftz6), T(fmz7), T(sat5), N("f32"), DST, T(acout3a), T(neg9), SRC1, LIMM, T(neg8), DST },
21:25 imirkin_: sure
21:25 imirkin_: so what?
21:25 karolherbst: or does everything have a predicate for gf100?
21:25 imirkin_: yep.
21:25 karolherbst: ohh I see
21:25 imirkin_: look at tabroot
21:25 imirkin_: (iirc)
21:26 imirkin_: + if (chipset < NVISA_GM107_CHIPSET && i->cc)
21:26 imirkin_: that's the wrong thing
21:26 karolherbst: I guess the same goes for that too
21:26 imirkin_: cc is the condition on the predicate
21:26 imirkin_: you want i->flagsDef >= 0
21:27 imirkin_: (if it's actually true... i'd double-check)
21:27 karolherbst: for gk110: { 0x6000000000000000ull, 0xe000000000000003ull, N("fma"), T(p), T(ftz38), T(fmz39), T(sat3a), N("f32"), DST, T(acout37), T(neg3b), SRC1, LIMM, T(neg3c), DST },
21:27 imirkin_: so that has it
21:27 imirkin_: acout37 == .CC
21:27 karolherbst: ahhh
21:27 karolherbst: I see :)
21:27 karolherbst: ohh right I remember now
21:28 karolherbst: k
21:28 karolherbst: then I remove those predicate checks
21:28 imirkin_: well - have to make sure that the emitter handles it i guess
21:28 karolherbst: yeah
21:28 imirkin_: although tbh i haven't the faintest clue what FFMA R0.CC does
21:28 imirkin_: i.e. when that condition code is set
21:29 imirkin_: for int ops it's a carry flag
21:29 karolherbst: maybe it affects the add part?
21:29 imirkin_: well, i just mean for any floating point op
21:29 karolherbst: mhh true
21:29 imirkin_: perhaps mwk knows
21:33 karolherbst: I hope I am right about the neg mods though
21:34 imirkin_: i don't think you are
21:34 karolherbst: well they aren't listed in envydis
21:34 karolherbst: but what do I know
21:34 imirkin_: i just mean ... no other modifiers can be set
21:34 imirkin_: i guess those checks aren't wrong
21:35 imirkin_: but ultimately they're in the wrong place
21:35 karolherbst: ohh right
21:35 imirkin_: it should be calling something like targ->isModSupported or something
21:35 karolherbst: I see
21:35 karolherbst: yeah right
21:35 karolherbst: but those target functions can't handle that limm fmad anyway
21:35 karolherbst: will be a bit painful to add it there :/
21:36 karolherbst: fun part is
21:36 karolherbst: that for gk110 a fma can have those neg mods
21:36 karolherbst: but not the limm one
21:37 karolherbst: mhh
21:37 imirkin_: the limm one can too
21:37 imirkin_: perhaps not as explicitly as you expect
21:37 karolherbst: well envydis tells me otherwise
21:37 imirkin_: the high bit of the immediate is effectively a neg bit
21:37 karolherbst: true
21:37 karolherbst: but all sources can take negs
21:38 imirkin_: uh huh
21:38 karolherbst: at least the fimm one cann
21:38 imirkin_: but the op is a * imm + b
21:38 karolherbst: { 0x9400000000000001ull, 0xb4c0000000000003ull, N("fma"), T(ftz38), T(fmz39), T(sat35), T(frm36), N("f32"), DST, T(neg33), SRC1, T(neg3b), FIMM, T(neg34), SRC3 },
21:38 imirkin_: one could argue that -a * imm could be rewritten as a * -imm
21:38 karolherbst: mhhh
21:38 karolherbst: tricky though
21:39 karolherbst: ohh mhh odd
21:39 karolherbst: I don't check for gk110 and neg
21:39 imirkin_: and note that envydis is most likely incomplete
21:39 imirkin_: esp for gk110
21:39 karolherbst: mhhh, I am silly
21:39 karolherbst: it actually has those negs for gk110
21:39 karolherbst: looked at the wrong thing
21:40 karolherbst: well, we did the gk110 fma together :p
21:40 karolherbst: and I am sure we did it completly
21:48 karolherbst: imirkin_: core reclocking fails, not memory
21:48 imirkin_: feel free to correct me.
21:48 karolherbst: the real issue is, that nouveau tries to set a cstate with a voltage requiernment of 1.275V, but the card can only do 1.215V or 1.2V and so it bails
21:48 imirkin_: the effect is the same.
21:48 karolherbst: well
21:48 karolherbst: memory is fast though
21:49 karolherbst: but yeah, in the end it won't matter if cores won't reclock or the memory
21:49 karolherbst: but core is easier to fix
21:50 karolherbst: \o/ flashed my phone with the right firmware and now it works again
21:50 karolherbst: ...
21:50 karolherbst: why do they send me a middle east version...
21:50 karolherbst: from a store from the UK
21:51 karolherbst: those guys
21:54 mwk: imirkin_: karolherbst: the CC setting works exactly as it did on Tesla, IIRC
21:54 mwk: as in, CC is really a 4-bit register with the usual overflow/sign/zero/carry flags
21:54 karolherbst: yeah, I don't know such details
21:54 imirkin_: mwk: and how did it work on tesla? :)
21:55 imirkin_: mwk: well, on fermi it's a single-bit thing afaik
21:55 mwk: FP ops set zero and sign flags, carry and overflow are cleared
21:55 mwk: nope
21:55 imirkin_: hm ok
21:55 mwk: on Fermi, it's definitely a 4-bit register
21:55 imirkin_: how can you access the zero/sign flags?
21:55 mwk: there are conditional branches that branch by CC instead of the usual Px
21:56 imirkin_: oh...
21:56 imirkin_: i know there's a CSETP
21:56 mwk: also, there are instructions that read $c to $p or $r
21:56 mwk: yep
21:56 imirkin_: (and presumably CSET)
21:56 mwk: so there are a few instructions that can be effectively predicated Tesla-style, by $c with a condition code
21:57 mwk: oh, and $c along with $p0-$p6 is also available as a special register than you can mov from
21:57 mwk: I called it $flags
21:58 mwk: I mentioned the format somewherein envydis sources
21:58 imirkin_: ah neat
22:48 karolherbst: imirkin_: do you plan to push any RA fix for the CSE patch of mine?
22:49 imirkin_: i already did
22:49 karolherbst: ohh, see the commit now
22:49 karolherbst: but I guess you didn't pushed my patch yet? will do a new run tomorrow then
22:49 imirkin_: it was the least intrusive one
22:49 imirkin_: i did not - it seemed like hakzsam was taking care of it
22:50 imirkin_: https://cgit.freedesktop.org/mesa/mesa/commit/?id=300b5ad023962ee95322e890a9ba57396392407e
22:50 imirkin_: that was the fix, ftr
22:50 karolherbst: yeah, saw it
22:50 karolherbst: completly missed it on the ml
23:33 imirkin: alright. this weekend is the weekend of making xf86-video-nouveau work on GM107
23:34 imirkin: i finally obtained one. should be fun.
23:34 karolherbst: sure :D
23:35 karolherbst: good luck with that though
23:37 karolherbst: well, will go to sleep, I just fail terribly at rooting my one phone... need sleep apperantly