03:19 hary_poter: tired of jews, lies, poverty and bad sound? add me to qox. my id is EEEDE72B6187425700FCA07D1B3FC2D49C6C5B6BDAEFA819D35976E8860B2E7DFDF6FC49EAF5
03:26 pq: hmm, chanserv has an akick command, should've tried that
03:29 mupuf: pq: thanks for the quick response time!
03:31 pq: just happened to be around :-)
07:14 pmoreau: imirkin: Ping, if you have some time to look at the new commit message: https://lists.freedesktop.org/archives/mesa-dev/2016-March/110398.html
07:18 imirkin: pmoreau: how do you feel about http://hastebin.com/oyofenijag.pas ?
07:20 pmoreau: imirkin: "if a function had arguments." in second paragraph -> "if a function had unused arguments.", apart from that, sounds good :-)
07:21 pmoreau: Thanks!
07:21 pmoreau: I'll go back to fix the MUL/MAD patch next week, and get as far as I can on the out-of-SSA pass.
07:22 pmoreau: Once I have the out-of-SSA , I'll have a look at how OpenCL 2.1 deals with loading SPIR-V binaries, and add that to clover.
07:22 imirkin: pmoreau: are you sure about that?
07:22 imirkin: pretty sure it's the case for *any* arguments...
07:22 pmoreau: That way, I can send my initial stuff for merging, without depending on any LLVM version.
07:23 pmoreau: imirkin: Ok, not completely. I always forget that *using* a register do not necessarily mean adding a def on it.
07:24 imirkin: :)
07:24 imirkin: this isn't about registers btw, but rather about values
07:24 imirkin: values can be defined and used
07:24 pmoreau: True
07:24 imirkin: using adds... uses :)
07:24 rotp_aryh: tired of jews, lies, poverty and bad sound? add me to qox. my id is EEEDE72B6187425700FCA07D1B3FC2D49C6C5B6BDAEFA819D35976E8860B2E7DFDF6FC49EAF5
07:25 imirkin: tired of fucking idiots is more like it =/
07:25 pmoreau: +1 :-/
07:31 imirkin: pmoreau: pushed
07:32 pmoreau: imirkin: Thanks a lot! :-)
07:34 imirkin: hakzsam_: we're falling behind... airlied sent out his series for shader images on r600
07:35 hakzsam_: imirkin, is this really a problem? I don't think :)
07:35 imirkin: what, falling behind?
07:35 imirkin: no, of course not
07:35 hakzsam_: they have images but no compute shaders :p
07:35 imirkin: just a motivating factor i suppose
07:36 hakzsam_: sure
07:36 hakzsam_: compute shaders are ready to go, I'll send the v2 today
07:36 imirkin: ok awesome
07:36 imirkin: you figured out that thing with ubo's?
07:36 hakzsam_: yes
07:36 imirkin: excellent
07:37 hakzsam_: right now, I'm looking at images for kepler
07:37 imirkin: hopefully calim did a lot of the painful work already
07:37 hakzsam_: seems like yeah
07:37 imirkin: and you'll be able to reuse the majority of the lowering passes
07:37 hakzsam_: yes
07:38 imirkin: i would just assume he got that stuff right - except he was trying to be all clever and wasn't assuming that you knew the format
07:38 pmoreau: hakzsam_: What happened to your rework of the validation pass on NV50? Was it merged?
07:38 imirkin: and so he had that crazy jump table thing
07:38 imirkin: i don't think that's necessary
07:38 hakzsam_: pmoreau, yep
07:38 imirkin: definitely not for GL, but i think we could get away with not having it even for CL
07:39 hakzsam_: imirkin, we will see
07:39 imirkin: you can reuse my logic i added for nvc0
07:39 imirkin: to unpack the bytes
07:39 hakzsam_: okay
07:39 pmoreau: hakzsam_: Cool :-)
07:41 imirkin: btw, if someone has time, could you spend some time testing 11.2.0-rc4 to make sure it's not totally broken on nv50 and nvc0?
07:41 imirkin: a piglit run would be ideal, but also just running some games on top would be nice
07:41 imirkin: i think xexaxo wants to cut the release tomorrow, so i dunno if i'll have time to look at it
07:44 hakzsam_: it's a bit late, but I'll try :)
07:45 pmoreau: I can launch a Piglit run on Tesla in a couple of hours
07:47 imirkin: pmoreau: that'd be awesome
07:47 imirkin: i can try to do some stuff tonight too... tbd
07:48 pmoreau: I'll try to remember to do it, but you might want to ping me just to be sure I don't forget ;-)
08:08 karolherbst: lol..
08:08 karolherbst: nouveau is faster than nvidia on team fortress
08:09 imirkin: but we more than make up for it on everything else...
08:09 karolherbst: right
08:10 karolherbst: I am sure team fortress doesn't need much gpu power anyway
08:14 hakzsam_: imirkin, I have just fixed image buffers on kepler :)
08:14 imirkin: awesome
08:17 hakzsam_: imirkin, 2d atomic operations almost work too
08:17 imirkin: =]
08:17 imirkin: calim's 4-year-old code still works
08:18 hakzsam_: yep :)
08:18 hakzsam_: imirkin, Passed: 72/159 (45.3%) --> with 3d images :D
08:19 imirkin: "with 3d images"?
08:19 hakzsam_: 3d images almost work too
08:19 imirkin: oh, but that's by chance
08:19 hakzsam_: maybe, but it works
08:20 imirkin: either it's not depth-tiled, or it's picking coordinates in the first depth tile
08:22 hakzsam_: imirkin, http://hastebin.com/cadoyosuho.coffee
08:22 hakzsam_: I still have a bunch of things to fix but it's a very good start
08:22 hakzsam_: thanks to calim
08:24 imirkin: very nice indeed
09:09 hakzsam_: imirkin, happy reviewing :)
09:11 hakzsam_: pmoreau, you have a GM20x, right?
09:11 pmoreau: hakzsam_: Yep! And even a box to put it in! :-)
09:12 hakzsam_: nice
09:12 hakzsam_: pmoreau, do you have time to test compute shaders on it?
09:12 pmoreau: Hum… not sure. Depends how much is needed to test it
09:13 hakzsam_: pmoreau, run deqp
09:13 hakzsam_: should take ~5 minutes or so
09:15 pmoreau: In theory, but in practice I need to clone the kernel/Nouveau/Mesa, and compile them, which adds some more time.
09:16 pmoreau: I have just done a basic Linux install on that box for now, didn't had enough time to proceed further.
09:16 hakzsam_: oh okay
09:16 hakzsam_: np
09:20 Yoshimo: i might offer my 980 later, where to get that software?
09:22 hakzsam_: Yoshimo, clone https://android.googlesource.com/platform/external/deqp
09:23 hakzsam_: Yoshimo, or maybe if you have piglit, this could be enough
09:42 Yoshimo: pmoreau: so i cloned that folder
09:44 Yoshimo: what exactly do i have to run after?
09:45 hakzsam_: Yoshimo, which one? piglit or deqp?
09:45 hakzsam_: and do you have mesa master?
09:46 Yoshimo: deqp
09:47 Yoshimo: well i have oibafs mesa ppa, so it should be close enough
09:47 hakzsam_: you need to build from the git repo because you will need my branch to test
09:48 hakzsam_: if it's too painful don't worry, I'll wait for pmoreau :)
09:49 Yoshimo: anyway i would try to get this deqp thing running first
09:49 Yoshimo: well only because my attempt to test the nvc0 fix for oblivion ended in a blackscreen that doesn't mean i am not trying
09:50 hakzsam_: you might need this patch for deqp to build it
09:50 hakzsam_: http://hastebin.com/axobudusuy
10:10 Yoshimo: hakzsam_: once it is built, how would i execute that tool?
11:01 hakzsam_: Yoshimo, go to deqp/modules/gesl31 and run MESA_GLES_VERSION_OVERRIDE=3.1 ./deqp_gles31
11:02 hakzsam_: which version of mesa do you have?
11:03 Yoshimo: that command results in memory access error and crashes, let me see the mesa version
11:04 hakzsam_: which chipset?
11:04 Yoshimo: Mesa 11.3.0-devel (git-a94d8d5 2016-03-31 trusty-sarnex-ppa)
11:04 hakzsam_: if you want to set up piglit https://cgit.freedesktop.org/piglit
11:04 Yoshimo: NVIDIA GM204
11:05 hakzsam_: your mesa is not too old, ~1 month
11:05 hakzsam_: did you already try to build your own mesa?
11:06 Yoshimo: yes last week
11:07 hakzsam_: pmoreau, 11.2.0-rc4 is totally broken GF119
11:07 hakzsam_: Yoshimo, nice, any chance you could update and rebuild mesa master?
11:08 hakzsam_: pmoreau, err, *not*
11:08 hakzsam_: pmoreau, :D
11:08 Yoshimo: yea sure, let me just finish cloning piglit
11:09 hakzsam_: ok
11:38 pmoreau: hakzsam_: So what's the status in the end: 11.2.0-rc4 is completely fine on GF119?
11:41 stardancer: tired of jews, lies, poverty and bad sound? add me to qox. my id is EEEDE72B6187425700FCA07D1B3FC2D49C6C5B6BDAEFA819D35976E8860B2E7DFDF6FC49EAF5
11:43 RSpliet: klines are probably useless, guessing they're using tor exit nodes
11:45 hakzsam_: pmoreau, it's fine
11:45 pmoreau: Ok :-)
11:51 Yoshimo: mhmm which ubuntu package is missing if piglit says OPENCL_opencl_LIBRARY-NOTFOUND ?
11:57 karolherbst: could be ocl-icd
11:57 karolherbst: no idea though
11:58 hakzsam_: Yoshimo, you can disable CL support IIRC
12:03 Tom^: karolherbst: did you see the phoronix benchmark of your awesome work? you just beat nvidia blob in tf2.
12:03 Tom^: :D
12:03 karolherbst: that*s not my work
12:03 karolherbst: this is with stock nouveau
12:03 karolherbst: ohh well besides the gddr5 thing
12:04 Tom^: well did the kepler really reclock without your patches?
12:04 karolherbst: bug I did only the last 1% to get it stable
12:04 Tom^: i dont recall it did :o
12:04 karolherbst: yeah the 680 did
12:04 karolherbst: the 780ti didn't
12:06 Tom^: or well to be fair everyones work at mesa and nouveau, you brought the awesomeness at higher clocks on the 780
12:07 karolherbst: yeah and it will still take at least two months to get it upstreamed
12:07 karolherbst: but there is nothing big anymore, just some minor refactorings and a little feature
12:08 karolherbst: we need a "daemon" which checks the temperature every second and adjusts the clocks in the kernel module, the code itself will be small, but requires some rework
12:14 SaveTheRobots: karolherbst: just fyi, 0f isn't completely stable for my with my 780 Ti after all ;(
12:15 SaveTheRobots: it locked up after 40 mins of gaming
12:15 SaveTheRobots: (boost 0)
12:15 karolherbst: mhh okay, so maybe the PLL settings indeed matters
12:16 karolherbst: SaveTheRobots: we could try to just use the blob configuration
12:16 SaveTheRobots: how would i go about doing that?
12:17 karolherbst: SaveTheRobots: with nouveau set 0f and boost 1
12:18 karolherbst: nvapeek 00137004 && nvapeek 0013701c
12:18 SaveTheRobots: ok, i don't have time atm but i'll give you a shout when i'm free
12:19 karolherbst: okay
13:04 fucrktheyhw: tired of jews, lies, poverty and bad sound? add me to qox. my id is EEEDE72B6187425700FCA07D1B3FC2D49C6C5B6BDAEFA819D35976E8860B2E7DFDF6FC49EAF5
13:28 pmoreau: hakzsam_: What do you think I should run for the piglit tests?
13:29 hakzsam_: a full run
13:29 hakzsam_: here's the instructions https://people.freedesktop.org/~imirkin/
13:32 pmoreau: Ah right, thanks!
13:33 hakzsam_: pmoreau, once it's done, please send us the report
13:33 pmoreau: Nah, I'll keep it for myself! :-p
13:34 hakzsam_: :p
13:36 Walakea: i dont know where to find out whether the xorg-video-nvidia ... nouveau oackage in debian stable supports my card (940M)
13:38 hercha: Salut
13:38 pmoreau: Walakea: You need to use xorg-video-modesetting, not xorg-video-nouveau for your card
13:39 Walakea: for now i dont care much as i am using the backported intel driver (i have both cards) but i would like to make the Nvidia card functioning in future
13:40 Walakea: but i dont really know how it works with graphics cards in G/L
13:40 pmoreau: G/L?
13:40 Walakea: GNU/Linux
13:40 pmoreau: K
13:42 Walakea: and what is the modesetiing package about?
13:42 pmoreau: It should be included with xorg-server since version 1.18 IIRC
13:43 pmoreau: But it's a generic driver for X.
13:44 Walakea: xorg server is the thing that will in future be replaced with Wayland?
13:45 pmoreau: Most likely, you will be able to switch between both cards using vgaswitcheroo, and offload some rendering to the non main card using PRIME (see https://nouveau.freedesktop.org/wiki/Optimus/)
13:45 pmoreau: Yes
13:46 pmoreau: Though, Canonical might argue that Mir is the future, not Wayland, but…
13:48 pmoreau: Walakea: Hum, 940M… is that a GM108 chipset?
13:48 pmoreau: I thought it was one of the GM20x ones
13:49 Walakea: where do i fing out?
13:49 Walakea: *find
13:49 pmoreau: `lspc -d 10de:` should tell you that
13:51 Walakea: i dont have the command lspc
13:52 pmoreau: lspci, sorry
13:53 Walakea: oh yes, i thought it was lspci but then i forgot the :
13:53 Walakea: ":"
13:53 Walakea: 04:00.0 3D controller: NVIDIA Corporation Device 1347 (rev ff)
13:53 Walakea: Asus laptop
13:55 pmoreau: Weird naming…
13:55 pmoreau: Could you please link your dmesg then?
13:55 dontkikme: tired of jews, lies, poverty and bad sound? add me to qox. my id is EEEDE72B6187425700FCA07D1B3FC2D49C6C5B6BDAEFA819D35976E8860B2E7DFDF6FC49EAF5
13:58 Walakea: there are lots of things in dmesg
13:59 pmoreau: Just paste it on any website please, so I can have a look.
14:00 Walakea: https://paste.debian.net/423083/
14:01 Walakea: [ 0.044473] Your BIOS is broken and requested that x2apic be disabled.
14:01 Walakea: wow
14:02 pmoreau: Hum, you do not seem to have Nouveau loaded
14:03 pmoreau: But, if it is a GM108, you'll hit this bug: https://bugs.freedesktop.org/show_bug.cgi?id=89558
14:04 pmoreau: The chipset is apparently still not recognised by the kernel, but there is a proposed patch attached to make it work.
14:05 Walakea: will the future kernel work properly?
14:06 pmoreau: A future kernel, most likely, the next version, probably not.
14:06 pmoreau: You should consider subscribing to the bug report to get informed when support gets merged for that card.
14:07 Walakea: bad thing is i still dont know how to make my wifi work properly, i downloaded some software targeted directly at my card and it seems to work but only with kernel 3.1X
14:08 pmoreau: :-/
14:10 pmoreau: It should be possible to backport the patch enabling basic support to 3.1x, but I bringing acceleration support will most likely be a different story…
14:12 Walakea: for future, are the wifi and graphics cards the most troublesome devices to properly run on G/L?
14:13 pmoreau: I would say yes, from my experience at least
15:32 karolherbst: can I convince somebody with a gk110+ gpu to try something out for me?
15:58 AndrewR: hi. It seems new vdpau stuff in mesa master broke it for presentation part at least on nv50 (nv92): http://fpaste.org/348027/
18:02 imirkin: ugh... vdpau breakage... i should have seen that coming =/
18:02 imirkin: hopefully i can repro
18:44 imirkin: ok yeah. repro'd. easy fix.
19:06 karolherbst: nice
19:11 imirkin: oh, and also, i grabbed a GK208 i wasn't using from work, might replace my GF108 with it...
19:32 karolherbst: imirkin: ohh, could you check something on that?
19:33 imirkin: not plugged in
19:33 imirkin: maybe this weekend
19:33 karolherbst: k
19:43 AndrewR: imirkin, thanks for fixing this vdpau breakage
19:43 imirkin: np
19:43 imirkin: can you double-check on nv50? i only tested on nvc0
19:45 AndrewR: imirkin, it works for me (tested with two videos)
19:45 imirkin: ok excellent
19:45 imirkin: it's fallout from the recent vdpau changes, as you surmised
19:47 AndrewR: imirkin, I tend to update mesa from time to time, and mplayer's default tend to be vdpau - so if it broke I see it quite fast (on first video watch)
19:47 imirkin: cool
19:47 imirkin: i don't update my system install very often
19:47 imirkin: i like to keep a stable base :)
19:50 AndrewR: I tested new kernel (4.6-rc1) and it broke my usb audio :}
19:50 imirkin: hopefully not nouveau's doing :)
19:51 AndrewR: imirkin, no, just some alsa bug ...
19:51 karolherbst: well it is rc1
19:51 karolherbst: all kind of things break
19:51 karolherbst: I got unlucky once and it didn't even boot
19:57 AndrewR: karolherbst, I run into this: https://bugzilla.kernel.org/show_bug.cgi?id=115301 BUT after patching it I still has no audio..but at least attempt at reclocking card (after small change for allowing reclock on nv92) worked ... h264 still hangs.
19:57 karolherbst: yeah well
19:57 imirkin: yeah... it's very sad i never worked out the issue with h264 on G92
19:57 karolherbst: most of the reclocking stuff is in a pretty "hanging" state
19:57 karolherbst: like it kind of works
19:57 karolherbst: but it could also not work
19:58 imirkin: karolherbst: i think it's moderately reliable on the older chips (where it works)
19:58 karolherbst: well
19:58 karolherbst: it isn't enabled for g8X
19:58 imirkin: right
19:58 imirkin: i mean esp like GT21x
19:58 imirkin: RSpliet did a pretty good job on those
19:59 karolherbst: we could also just enable it on fermi without memory reclocking, that's what I meant. It kind of works, but not enough
19:59 imirkin: largely pointless, since memory clocks are so low by default
19:59 karolherbst: it still gives like 25% more pref on mine
20:00 imirkin: otoh on my gpu:
20:00 imirkin: 0f: core 710 MHz memory 500 MHz
20:00 imirkin: AC: core 405 MHz memory 324 MHz
20:00 karolherbst: and somebody here got like 10000% more perf in a use case, but I think the application was just silly when the cores are too slow
20:00 karolherbst: ...
20:00 imirkin: [i assume that's missing some sort of doubling? otherwise that's pretty sad]
20:00 karolherbst: no, sounds right
20:00 karolherbst: ddr3?
20:00 imirkin: ya
20:00 karolherbst: like 420 or someting? :D
20:00 imirkin: 610
20:01 karolherbst: uhh eah, that one sucks
20:01 imirkin: i go for the high-end gpu's :p
20:01 imirkin: i think it was like $20 on ebay or something... i splurged
20:01 karolherbst: mhh
20:01 karolherbst: somethign was odd with the 610
20:02 imirkin: i just put down a $10 bid on a PCI FX5200 ... that'd allow me to test every driver without swapping boards
20:02 imirkin: since i should be able to convince the vieux driver to run on it
20:02 karolherbst: okay, I think your clocks have to be doubled, but this is still sad :D
20:03 imirkin: karolherbst: now you know why i get 3fps in heaven :)
20:03 karolherbst: 8 GB/s
20:03 karolherbst: there is a faster 610 with 1800Mhz though
20:03 karolherbst: and this is also what nvidia says
20:03 imirkin: perhaps... but it's not this one
20:04 karolherbst: yeah, wikipedia has 1000/1800, so I think both exists
20:04 imirkin: it comes up as pcie v1 too
20:04 imirkin: even my GT215 comes up as pcie v2
20:04 karolherbst: yeah well
20:04 karolherbst: there is a PCI version of the 610
20:04 imirkin: heh, perhaps. this is pcie though.
20:04 karolherbst: imirkin: also 1x width?
20:04 imirkin: no, 16x
20:05 karolherbst: at least :D
20:05 imirkin: at least the connector is
20:05 imirkin: LnkSta: Speed 2.5GT/s, Width x16
20:05 imirkin: oh my bad. GT 620
20:05 karolherbst: gf119 or gf108?
20:05 imirkin: GF108
20:05 imirkin: VGA compatible controller [0300]: NVIDIA Corporation GF108 [GeForce GT 620] [10de:0f01] (rev a1)
20:06 karolherbst: k
20:06 karolherbst: that one is somewhat usable
20:06 karolherbst: doubled cores than 610 :D
20:06 karolherbst: but it should be able to do 2.0
20:06 imirkin: so like 2 instead of 1? :)
20:06 imirkin: of course. it just doesn't come up with v2
20:06 karolherbst: is that with nvidia or nouveau?
20:07 karolherbst: ahhh
20:07 imirkin: nouveau
20:07 karolherbst: yeah
20:07 karolherbst: that happens
20:07 karolherbst: some gpu boot with v1
20:07 imirkin: and this is one of them
20:07 karolherbst: yeah, but nouveau should put it into v2 though
20:07 imirkin: eventually
20:07 karolherbst: well it was merged for 4.5 I think, so it should happen
20:07 imirkin: i'm on 4.3.0
20:07 karolherbst: otherwise it is a bug
20:08 karolherbst: ohhh
20:08 karolherbst: yeah, okay
20:08 imirkin: i have a bunch of spinning hd's attached to this box, so i try not to reboot it too often
20:10 karolherbst: eah, but 4.3
20:11 imirkin: uptime: 55 days...
20:12 karolherbst: yeah well
20:15 karolherbst: gnurou: you need some stupid ACPI stuff
20:15 karolherbst: gnurou: and mostly you only get this on mobile gpus
20:16 gnurou: karolherbst: ah, talking about bug 94725?
20:16 karolherbst: I don't think any desktop gpu can be suspended safely?
20:16 karolherbst: and with suspended I mean put into d3cold
20:16 gnurou: eh, that'd be too bad :/
20:16 imirkin: gnurou: runpm is done using ACPI calls
20:17 karolherbst: gnurou: and those change pwoer state messages are useless
20:17 imirkin: gnurou: most desktop motherboards don't have the circuitry for it in the first place
20:17 karolherbst: I get them too
20:17 gnurou: mmm so this means I need a mobile GPU in order to reproduce this?
20:17 imirkin: gnurou: yep
20:17 imirkin: gnurou: or a million-dollar motherboard
20:17 imirkin: laptop might be the cheaper option :)
20:17 karolherbst: gnurou: I always wanted to get runpm right even on my chip there are issues, but...
20:17 karolherbst: it is rather anoying
20:18 gnurou: damn, guess I cannot help then :/
20:18 karolherbst: well
20:18 karolherbst: there is one idea though
20:18 karolherbst: what does nouveau do, when the main gpu comes from the intel cpu
20:19 karolherbst: gnurou: in any way, the card only suspens when it isn't in use
20:20 karolherbst: and when the gpu is the only gpu, it may be suspended when no display is attached, but usually it isn't powered off
20:22 airlied: that might be the Windows 10 does stuff different bug
20:22 karolherbst: uhhh
20:22 karolherbst: right
20:22 karolherbst: ...
20:22 karolherbst: damn you windows :D
20:23 airlied: gnurou: it might be good to have a patch to read the register that only gets reset on power down on resume
20:23 airlied: to see if the GPU has actually powered down or not
20:24 karolherbst: there are some race condititons in the code anyway
20:25 karolherbst: some kworkers and timers you don't get canceled in fini and the they do stuff exactly when the gpu goes into sleep
20:26 gnurou: airlied: mmm, so you mean that maybe devinit should not be run in that case?
20:28 karolherbst: gnurou: I tink it has to, because a suspended gpu is like a powered off gpu and when it powers on, it has to be fully posted and everthing
20:28 airlied: karolherbst: in that case the register should be correct
20:28 airlied: and devinit can run
20:28 airlied: but running devinit when something has gone wrong and the BIOS hasn't shut the power off causes a lot of fun
20:29 karolherbst: gnurou: you said something about a mmio reg which could be used to check if a card is posted or not?
20:29 karolherbst: this could be used in that case
20:29 karolherbst: if we resume, but the card is posted, then it wasn't powered off
20:29 karolherbst: or something like that
20:30 karolherbst: I doubt the firmware is smart enough to ever POST the gpu by itself on a GPU resume
20:30 karolherbst: at least for me it doesn't
20:30 gnurou: right, it seems like this register is only checked during initial boot, and not PM
20:31 gnurou: maybe we should also check it in that case... let me confirm that it is correct in the S3 case, and if so we can try this
20:33 karolherbst: uhh "ACPI: \_SB_.PCI0.PEG0.PEGP: failed to evaluate _DSM"
20:33 karolherbst: yeah, the gpu might be not powered off at all
20:34 karolherbst: usually something like this should be in dmesg: VGA switcheroo: detected Optimus DSM method \_SB_.PCI0.PEG0.PEGP handle
20:34 karolherbst: or whatever type of DSM method is getting used
20:35 gnurou: and base->post will be set by devinit_fini, so devinit will be run fully when we resume
20:36 karolherbst: gnurou: there are two while loops inside gm200_devinit_post
20:36 karolherbst: there nouveau could get stuck
20:36 gnurou: yeah, these would deserve having a timeout too...
20:36 karolherbst: :)
20:37 karolherbst: ohh wait
20:37 karolherbst: but I think it hangs somewhere else
20:37 karolherbst: maybe pmu_load is the bad thing
20:39 gnurou: what makes you think it is hanging elsewhere?
20:39 karolherbst: just my gdb doing crazy stuff
20:40 gnurou: karolherbst: oh, so you can repro the issue?
20:40 karolherbst: no idea, I just loaded my compiled version and checked what gdb say where it is
20:41 gnurou: not sure it is reliable :)
20:41 karolherbst: right
20:42 karolherbst: I am sure it is the while loop before pmu_load
20:42 karolherbst: RIP: 0010:[<ffffffff81415940>] [<ffffffff81415940>] ioread32+0x30/0x40
20:42 karolherbst: this kind of tells us
20:42 karolherbst: and post is true on resume
20:42 karolherbst: so it fits
20:43 karolherbst: so it got stuck doing this: while (nvkm_rd32(device, 0x10a10c) & 0x00000006)
20:44 gnurou: sounds plausible
20:44 karolherbst: ohh regarding that lspci issue. I should upstream this: https://github.com/karolherbst/linux/commit/cb918e4c926990dfcfce92e1ecd905e0896de605
20:44 karolherbst: and then libpciaccess won't wakeup gpus anymore
20:44 karolherbst: if it uses this interface
20:44 gnurou: and IIRC this is what is supposed to happen if devinit is run twice
20:45 karolherbst: yeah, it makes sense
20:45 karolherbst: because the ACPI DSM method is faulty
20:45 karolherbst: and never executed
20:45 karolherbst: and then the gpu is just put into d3cold stay without the DSM call
20:45 karolherbst: this could lead to all kind of issues
20:45 karolherbst: or not suspend at all
20:46 gnurou: so wouldn't a proper fix be to not suspend in that case?
20:46 karolherbst: mhh
20:46 gnurou: I mean even if we fix devinit, there are plenty of other things that could be done wrong
20:46 karolherbst: no, that is actually fine
20:47 karolherbst: it will free ressources and stuff
20:47 karolherbst: we just don't need to post
20:47 karolherbst: gnurou: on suspend, all fini calls are called
20:47 karolherbst: all subdevs are destroyed and everything
20:47 imirkin: karolherbst: feel like tracing a deqp test for me?
20:48 karolherbst: ohh
20:48 karolherbst: I still didn't compile it cause of the stupid python thing
20:48 imirkin: python thing?
20:48 karolherbst: it calls python3 on a python2 script
20:48 imirkin: oh, coz your distro is all broken?
20:48 karolherbst: :D
20:48 karolherbst: hell no
20:48 karolherbst: the build system is
20:49 karolherbst: if it wants python2, it should tell me so
20:49 karolherbst: well
20:49 karolherbst: I just hack up the CMakeCache file :D
20:49 karolherbst: nice, it compiles now
20:50 karolherbst: imirkin: so, what shall I run?
20:52 imirkin: it does tell you - it runs 'python'. which means python2.
20:52 imirkin: karolherbst: ./deqp-gles31 --deqp-visibility=hidden --deqp-case='dEQP-GLES31.functional.shaders.sample_variables.sample_mask_in.bit_count_per_sample.multisample_texture_8'
20:52 imirkin: from the modules/gles31 dir
20:52 karolherbst: why is python python2?
20:53 imirkin: because if you grab software from 15 years, it'll just use "python"
20:53 karolherbst: :D
20:53 imirkin: and python3 is not backwards-compatible with python2
20:53 karolherbst: not python1?
20:53 imirkin: therefore it has to be a separate binary name
20:53 imirkin: python1 was never a real thing
20:53 imirkin: and it wasn't sufficiently popular for it to matter
20:53 karolherbst: ahh I see
20:53 imirkin: python2, on the other hand, was.
20:55 karolherbst: mhhhh
20:55 karolherbst: how can I convince it to use my nvidia gpu?
20:56 imirkin: ohhhh
20:56 karolherbst: right, I have to use GLX
20:56 imirkin: hrm.
20:56 imirkin: yeah
20:56 imirkin: -DDEQP_TARGET=x11_glx
20:56 karolherbst: well I compiled with glx and egl
21:00 karolherbst: mhhh
21:01 karolherbst: ohh okay, I know
21:02 karolherbst: okay, it works
21:03 karolherbst: imirkin: http://filebin.ca/2cDX3l0JpZVk/deqp_stuff.xz
21:04 karolherbst: it worked, just passing the arguments is a bit broken with my script
21:04 imirkin: ah ok
21:04 imirkin: thanks
21:07 imirkin: ohhhh i see.....
21:07 imirkin: covmask is the full coverage mask, even with per-sample shading
21:08 imirkin: useless.
21:08 imirkin: GRRRR!
21:08 imirkin: this will want a variant thing
21:08 imirkin: i'll have to defeat it somehow
21:16 karolherbst: uhhh
21:16 karolherbst: the nvidia windows driver supports panel self refresh...
21:16 karolherbst: with desktop gpu suspending (kind of)
21:16 karolherbst: ohh wait
21:16 karolherbst: wrong
21:16 karolherbst: memory goes into self-refresh
21:17 karolherbst: when the display is in suspend
21:17 karolherbst: "Power Saving: The driver now puts the GPU in the OFF state, with video memory in self-refresh mode, when the monitor is blanked or inactive. (Windows 10)"
21:19 karolherbst: gnurou: there might be a way :D
21:21 imirkin: i disapprove of how the pixld instruction works.
21:21 imirkin: it is not convenient =/
22:01 imirkin: hakzsam_: i briefly looked at that sampler issue with shadow textures... very odd. the one thing i did notice is that unknown TSC bit 0x300 is set on the second word for the samplers, but... i doubt it matters.
22:01 imirkin: hakzsam_: make me an mmt of nouveau for the same test and i can try to see if there are obvious differences...
22:01 imirkin: there were also some unknown bits set in the launch desc
23:36 gnurou: aha, I just understood why devinit does not use a timeout when waiting on the PMU... that's because at this time the timer subdev is potentially not initialized
23:36 gnurou: or rather, is *absolutely* not initialized