10:02mupuf: Seems like we received a Titan :) I will test it in the coming days. Karol will be happy as it will definitely be a challenging one to keep in the power budget
10:09pq: hehe, a donation? Congrats! :-)
11:06mupuf: pq: yes, generous donation
11:07mupuf: we needed a NVF0 and we got ... an insane GPU
11:07mupuf: apparently, this is closer to what the user cares about, and nouveau should work on it (maybe not at an amazing speed, the display side is what he is interested in)
11:08mupuf: the user = the donator
11:11pq: hope it doesn't fry your box :-D
11:17mupuf: that would be unfortunate
11:17mupuf: I still had to pay for the customs (VAT). Was an interesting time to figure out everything when it is in Finnish
11:22mlankhorst: good learning experience!
11:37karolherbst: yay an answered question
11:48karolherbst: github has a caching problem somewhat
11:53pmoreau: mupuf: IIRC, last time I tried (i.e. last year) it worked. I don’t remember at all for reclocking.
11:56karolherbst: yay, titan
11:57karolherbst: mupuf: mind running pixmark_julia_fp64 on it to check if nouveau can already provide usefull f64 performance?
12:02mupuf: karolherbst: will check it out, yes
12:02mupuf: f32 vs fp64
12:03mupuf: the thing is that f32 may be faster too if the card is able to split its floating point unit
12:03mupuf: no idea if it is possible
12:04karolherbst: yeah maybe
12:04karolherbst: but the difference shouldn't be too big
12:08mupuf: we'll see
12:42karolherbst: but do you have the titan now or do you still need to pick it up or something?
12:53mupuf: karolherbst: it is at my place
12:54karolherbst: nice :)
12:55mupuf: if it works, then, I will thank the generous donato
12:55mupuf: if it does not ... I am in for a treat with the customs to get reimbursed
12:55mupuf: but let's not think about that
12:56karolherbst: this therm table is stupid by the way :/ there are some bytes behaving pretty similiar and I don't really get what the entry tells nvidia to do. I mean yeah it controls the downclocking policy somewhat, but besides that :/
13:00karolherbst: fun, maxwells have 4 entries there
13:00karolherbst: and there are at least 3 temperatures in each
13:04mupuf: all the fun
13:05karolherbst: it wouldn't be so annyoing if those entries wouldn't be 24 bytes long
13:12bozhan: karolherbst: hi, you've asked me if i have nvdia driver to help to make some dump? for my GK107 . I have it now :)
13:14karolherbst: bozhan: do you know in which context I asked you?
13:16bozhan: karolherbst: because when i use your tree and put 0f in pstate - i have kernel lockups ...
13:16karolherbst: ahh okay
13:16karolherbst: even with the increased voltage thing?
13:19bozhan: i think we didn't tried that because i have debian and was almost imposible for me to run nvidia binary, so we tried just with your tree.... and after lockups ... you tell me that maybe something is messed in kernel...
13:21bozhan: now i have kernel 4.6 with nvidia blob... and if you want me to test something. or make some dump?
13:22karolherbst: mhh tracing nvidia should be fine
13:22karolherbst: just a few reclocking things
13:23bozhan: i've made one dump when just run optirun glxgears ... do you want me to send it to you, what else?
13:24karolherbst: well you should start the trace and do "optirun -b none nvidia-settings -c :8"
13:24karolherbst: then check the performance section and wait until it clocks down
13:24karolherbst: then bring it to clock up again
13:24karolherbst: like running something in optirun in another shell
13:24bozhan: where is that performance section?
13:25karolherbst: where the gpu is
13:28bozhan: oo i found it :)
13:44bozhan: karolherbst: where i have to send dump?
13:46karolherbst: bozhan: either upload it somewhere or firstname.lastname@example.org
13:46bozhan: do i have to set some fancy name to it ?
13:51bozhan: karolherbst: just sent it
15:04n-wInd0Wz: hi.. my nouveau.ko is spitting DRM:: evicting console.. and other messages repeatedly into dmesg
15:07imirkin: that's good
15:07n-wInd0Wz: did I cause it's DRM to blow up when I was walking /dev/mem and looking at pci configuration space, and all Option ROMs that were in memory (searching for their headers and the PCIR header for PCI Data Structure to ensure they are real, I pulled 53 option roms)... I also specifically read the VGA (Quadro2) Option ROM memory at 0xb1000000-0xb1007fff
15:08imirkin: touching PCI config space is not a generally safe operation
15:08n-wInd0Wz: I've been doing a lot of pci/pci-express work lately, and analyzing configuration space (MMIO 0x80000000) is part of it
15:09imirkin: btw, when you say "Quadro2", are you talking about the NV15 quadro2?
15:09n-wInd0Wz: well... it's supposed to be safe to query the memory mapped config space for PCI/PCIE.. AFAIK, but probably best if you're the OS. I've been doing read-only operations
15:09n-wInd0Wz: i think most x86 pci config space queries use PIO to ports 0xcf8-0xcff which may be a bit different than just reading the mmio-memory at 0x80000000-0x90000000
15:09n-wInd0Wz: so did i possibly blow up the nouveau.ko driver you think,? heh
15:10n-wInd0Wz: and how would i fix it
15:10imirkin: "reboot" :)
15:10n-wInd0Wz: tried it.. will again.
15:10imirkin: power off?
15:10imirkin: you could have wedged the card bigtime
15:11imirkin: and yes, it *should* be safe to read mmio stuff, but... reality occasionally differs.
15:26n-wInd0Wz: imirkin: hrm.. alright. well, can someone explain what nouveau is? it's a drm driver for nvidia gpu's? i've read online that people are using nvidia drivers instead, in some cases.. what is going on with the driver situation for certain graphics chipsets?
15:27n-wInd0Wz: also, that drm detection shit was lame, i'm gonna throw nouveau.ko into ida-6.9 and hexrays and decompile it, i'm curious what the hell set it off.
15:27Tom^: why decompile and not just browse the source?
15:31n-wInd0Wz: wait, there's sourc efor tihs
15:31n-wInd0Wz: but it's got drm in the binary?
15:32n-wInd0Wz: why don't people recompile and rip that out, because that just disabled my graphics capabilities after I decided to walk /dev/mem with an LKM and tool I wrote
15:32Tom^: Direct Rendering Manager, not Digital Rights Management
15:32n-wInd0Wz: Oh, shit
15:33n-wInd0Wz: So, the other thing I did was uninstall dbus, I hate dbus and systemd... I read that it's possible to do
15:33n-wInd0Wz: but after I did it, nouveau stopped working
15:33n-wInd0Wz: i thought it was my pci memory stuff
15:33n-wInd0Wz: now i'm thinking there's some relation to dbus...
15:43karolherbst: n-wInd0Wz: is udev still installed?
15:45karolherbst: Tom^: by the way, you didn't try out my newest branch, did you?
15:45Tom^: karolherbst: nope
15:46Tom^: karolherbst: still stuck on windows because the division and upcoming mirrors edge and upcoming mass effect and and ...
15:46karolherbst: well wine has proper d3d10 and d3d11 support now
15:46Tom^: it does?
15:46karolherbst: yeah, somewhat
15:46karolherbst: TYPELESS teaxtures are supported now
15:46Tom^: O_o hm
15:46karolherbst: which is used by like 100% of all d3d10+ games
15:47karolherbst: sind 1.9.8 I think
15:47karolherbst: now issues can be actually found in those games
15:47karolherbst: Tom^: https://bugs.winehq.org/show_bug.cgi?id=22762
15:47karolherbst: Tom^: and more
15:47Tom^: heh nice
15:53pmoreau: Tom^: +1 for ME and ME!
15:54Tom^: pmoreau: =D
15:55pmoreau: But Mass Effect has been postponed to next year, dunno about Mirrors Edge.
15:55Tom^: it has? i missed that. damn
15:55Tom^: mirrors edge is supposed to be out 7th june
15:56pmoreau: Oh, pretty soon then :-)
15:57pmoreau: There will first be Stellaris on 9th May and Hearts of Iron IV on 6th June from Paradox Interactive.
15:58pmoreau: And since they will both be released on Linux on day 1, I’ll have to do some testing with Nouveau
16:07Tom^: oh and no mans sky looks cool
16:08Tom^: which is also soon
16:08karolherbst: pmoreau: :)
16:28Lemmata: hello, I'm poking around and trying to get my GTX 965M videocard working on Debian and I am getting an error that the firmware has failed to load: "firmware: failed to load nvidia/gm206/gr/sw_nonctx.bin (-2)". I am running the 4.6.0-rc3-amd64 kernel from the Debian experimental repo. How do I approach solving this problem?
16:29imirkin_: Lemmata: install the firmware :)
16:29karolherbst: Lemmata: install a recent linux-firmware package
16:29imirkin_: unfortunately GM20x is a very locked-down gpu, which requires signed firmware. as a result, nouveau can't supply its own.
16:30Lemmata: imirkin_: that's a shame, I was hoping to avoid non-free components
16:30imirkin_: Lemmata: well, all you have to do is break their key, and then we can upload all the firmware we want.
16:31karolherbst: well you need the firmware only for acceleration anyway.
16:31karolherbst: Lemmata: if you don't use OpenGL and are fine with sofware rendering you won't need them :D
16:31imirkin_: karolherbst: it's clearly a laptop, this is an accelerator gpu
16:32Lemmata: imirkin_: I assume breaking the key is an involved process?
16:32karolherbst: more like time consuming
16:33karolherbst: it won't matter if we start this year or next, the result is the same :)
16:33Hoolootwo: what size key?
16:33karolherbst: 128 bit AES I think
16:33karolherbst: not sure though
16:33imirkin_: not sure, but i assume big... it's a signature process
16:33imirkin_: i doubt it's AES - that's for encryption, not signining iirc
16:34karolherbst: ohh right, silly me
16:35Hoolootwo: if it's a <1024 bit RSA it could probably be cracked in a couple months/weeks
16:35mwk: it's 128-bit AES, used with CMAC as a sig
16:35mwk: iow. forget it
16:35Riastradh: More important than `what size key' is `what signature scheme'.
16:35Hoolootwo: ah okay
16:35Riastradh: RSASSA-PSS with a 768-bit key is breakable on a GPU cluster; Ed25519 with a 256-bit key is not.
16:35imirkin_: unfortunately they're not idiots
16:35karolherbst: Riastradh: well it is AES though
16:36Riastradh: If it's AES-CMAC, then it's symmetric, so the `signing key' and `verification key' are one and the same, and must be embedded in the device somewhere.
16:36karolherbst: Riastradh: try it and report back in 10 years :)
16:36karolherbst: Riastradh: XD
16:36mwk: Riastradh: it is, but it's not exactly easy to extract it
16:36karolherbst: Riastradh: right, we already had this thought
16:36imirkin_: Riastradh: yeah, somewhere in those 8 billion transistors...
16:36mwk: if you have a proper lab, do it
16:37Riastradh: Sure. Not saying that it's easy! But it's within the realm of possibility to do that, unlike anything cryptanalytic against AES itself.
16:37karolherbst: mwk: do you think nvidia can replace the key?
16:37mwk: karolherbst: not in an already-manufactured gpu, no
16:37mwk: but if you break GM20x, GP10x will ship with a different key
16:37karolherbst: mwk: like 100% sure?
16:37mwk: and you get to do it all over again
16:38mwk: 100% sure
16:38karolherbst: mwk: yeah of course, that's why I was thinking if they have means to update it
16:38imirkin_: mwk: i'm sure after a few times, one would get pretty good at it :)
16:38mwk: and GP20x may ship with RSA, not AES-CMAC
16:38mwk: and then it's good night
16:38Riastradh: imirkin_: Yes, but they would probably learn to use an asymmetric signature scheme.
16:38karolherbst: then we wait for GP10x and then get the key :)
16:42Yoshimo: are you tired of waiting for maxwell reclock support for gen2? ;)
16:52Lemmata: so is the approach from here on out to use non-free firmware if nvidia continues to require signed firmware?
16:52Lemmata: Or are there any alternatives?
16:52imirkin_: Lemmata: i hear AMD makes gpu's
16:53Lemmata: imirkin_: and they aren't locking their GPUs down?
16:53imirkin_: i'm unaware of any firmware signature stuff. however there also isn't open-source command processor firmware, so... not great.
16:54Lemmata: sooo time to invest in some FPGAs and publish some open GPU designs to opencores? :p
16:55loonycyborg: iirc there already are some
16:57loonycyborg: now only need to somehow secure a fab to manufacture them :P
16:58Lemmata: hit up bunnie, get it to the fab and put it into the novena
17:50hakzsam: GL 4.2 on Kepler is now upstream! bbl
17:53imirkin_: hakzsam: give it a test on mupuf's TK1
17:53imirkin_: hakzsam: i'm guessing we select some things slightly wrong there
17:53hakzsam: imirkin_, will do, but later ;)
18:05karolherbst: hakzsam: did you update docs/GL3.txt? :D
18:08mupuf: hakzsam: and where is the enable for image_size?
18:08imirkin_: that's a package deal with the st/mesa impl
18:08imirkin_: no reason to break them out
18:08mupuf: imirkin: ack
18:08mupuf: it is not in meta :)
18:09imirkin_: not sure what meta has to do with it
18:09imirkin_: i don't think it wraps any images stuff at all
18:09mupuf: ok, for intel, it was not broken up
18:09imirkin_: right, coz it's a mesa driver
18:09imirkin_: just like st/mesa enables both. but no reason to pass up idiocy through gallium.
18:09hakzsam: karolherbst, no, because it's still not exposed on fermi/maxwell
18:10mupuf: hakzsam: that is not an acceptable reason, as I said before, but it is OK
18:10hakzsam: it is
18:10hakzsam: same for compute shaders btw
18:10mupuf: hakzsam: still eating, so you can make a quick run on the titan if you want
18:10mupuf: I know :)
18:10imirkin_: mupuf: GL3.txt is whatever we want it to be
18:10mupuf: imirkin: sure thing
18:11hakzsam: mupuf, time to eat too :)
18:11imirkin_: unless it's generating oodles of confusion amongst the COUNTLESS contributors to nouveau, i don't really care.
18:11mupuf: so, the titan has this lovely LED
18:11mupuf: behind Geforce GTX
18:11imirkin_: i think there's a gpio that controls it, fyi
18:11mupuf: and IIRC, it is attached to a PWM controller
18:12karolherbst: hakzsam: I don't think this is a requiernment for docs/GL3.txt because i965 is also listed everywhere :)
18:12mupuf: so... morse code anyone? :D
18:12mupuf: with fade in/out to make it look good
18:12karolherbst: mupuf: :)
18:12karolherbst: mupuf: on windows with geforce experience you can configure those LEDs :D
18:12mupuf: oh, so I can change the color too?
18:12karolherbst: I think so, yes
18:13mupuf: well, that will make it hard for humans to get the message from Linus :D
18:13imirkin_: we could expose something to the LEDs subsystem
18:14imirkin_: which allows userspace control, as well as various triggers
18:14imirkin_: (like network activity etc)
18:14mupuf: ah ah ah
18:14mupuf: yeah :D
18:15hakzsam: imirkin_, "The NVC0 driver still has ARB_compute_shader, ARB_robust_buffer_access_behavior, and ARB_shader_image_size before it will be through with the OpenGL 4.3 changes. "
18:15hakzsam: I'll reply on the forums later :D
18:15imirkin_: hakzsam: ?
18:16hakzsam: imirkin_, phoronix news
18:16hakzsam: ARB_shader_image_size and ARB_compute_shader are already implemented
18:16imirkin_: and should have been enabled as a result of images coming on
18:16hakzsam: right, it is
18:18mupuf: ok, time to check out this titan
18:19mupuf: idle power, 12W, cool!
18:19mupuf: 12.4W to be precise
18:20imirkin_: peak 12.4MW :)
18:21mupuf: ah ah
18:21mupuf: well, at boot clocks, xonotic maxes out the card at 19W
18:22mupuf: so... definitaly can do better
18:22karolherbst: mupuf: 12W on 07?
18:22mupuf: let's push the clock when the benchmark is done
18:24mupuf: oopsie, I had vsync on
18:25karolherbst: mupuf: does "nvapoke 0x20200 0x60 27722455" reduce the idle power consumption significantly? :D
18:26karolherbst: so there seems to be some regs needed for desktop chips
18:27mupuf: yes, told you it was not that easy for my nve6
18:27mupuf: ok, I need to update your kernel
18:28mupuf: reclocking failed
18:28karolherbst: no surprise
18:29karolherbst: voltage error?
18:29mupuf: well, no idea
18:29mupuf: but I am using an old version of your tree
18:29mupuf: so, I need to update it
18:30mupuf: https://github.com/karolherbst/nouveau/commits/stable_reclocking_kepler_v5 ?
18:38mupuf: karolherbst: hmm, does not seem compatible with the kernel that I deployed
18:38mupuf: which is the 4.5-rc7
18:39hakzsam: mupuf, are you using the titan?
18:39karolherbst: mupuf: drm-next
18:40karolherbst: mupuf: rebase on my master branch or bens master or whatever
18:40karolherbst: I don't think you need to remove any commits :/
18:40karolherbst: let me check
18:41karolherbst: mupuf: nope, it should reabse cleanly
18:41mupuf: karolherbst: very good
18:41mupuf: hakzsam: well, you can make a short piglit run if you want, but let me reboot
18:41mupuf: I crashed the GPU
18:42mupuf: hakzsam: go for it
18:42hakzsam: mupuf, thanks
18:43karolherbst: mupuf: don't forget to add the vbios
18:43mupuf: yop, I am doing this
18:48hakzsam: well, the titan has some strange issues :)
18:49mupuf: karolherbst: pushed
18:50karolherbst: hakzsam: strange as in like kepler2 strange or something completly different?
18:50hakzsam: seems like different
18:56hakzsam: imirkin_, ahah, your reply is fun :)
18:57imirkin_: i try.
18:57hakzsam: mupuf, this titan uses SM35 or SM37?
18:57imirkin_: is there a diff?
18:57hakzsam: no clue
18:57mupuf: hakzsam: you know as well as I do
18:57imirkin_: as far as ISA goes, it should be SM35
18:57hakzsam: imirkin_, but some CAS tests fail on this titan...
18:58hakzsam: and not with that gk208
18:58hakzsam: deqp-gles31/functional/image_load_store/2d/atomic/comp_swap_r32i_return_value: fail
18:58hakzsam: deqp-gles31/functional/image_load_store/3d/atomic/comp_swap_r32i_return_value: fail
18:58imirkin_: probably due to factors unrelated to the ISA
18:58hakzsam: (for example)
18:58imirkin_: but rather that we don't configure something properly that happens to work out on GK208
18:59hakzsam: yeah, probably
18:59imirkin_: i'd carefully look at the code
18:59imirkin_: with nvdisasm
18:59imirkin_: and comapre it to what nouveau *thinks* it's emitting
18:59hakzsam: I'm going to do
18:59hakzsam: but maybe mupuf wants to use the titan? or karolherbst ?
19:00mupuf: hakzsam: I am doing stuff that should not interfere with you
19:01mupuf: like, this useless thing: GPIO 19: line 19 tag 0x84 [LOGO_LED_PWM] IN DEF 0 param 1 gpio: normal SPEC_OUT 0x84 [SOR1_PANEL_BACKLIGHT_LEVEL]
19:01mupuf: and I have to add the chipset too to nvbios
19:01mupuf: so, you can continue
19:09mupuf: RRRrrr, I hate that there can be more than one temperature sensor :s
19:10mupuf: but it has an INA3221 for the power meter
19:11mupuf: GPIO voltage management
19:12mupuf: seems like everything nvbios is happy with this vbios. Next!
19:15karolherbst: mupuf: LED!! :D
19:17RSpliet: mupuf: can we please implement that LED support with a bit of troll? like make it blink some message in morse O:-)
19:17hakzsam: mupuf, well, I'm done with the titan
19:17hakzsam: mupuf, will check on the tegra
19:20mupuf: RSpliet: that was my proposal!
19:20mupuf: we could relay the message of Linus :D
19:20RSpliet: why so hateful
19:23RSpliet: if we can just make it stream tr(ol)*
19:23Yoshimo: now that we got firmware i don't think it is appropriate
19:23mupuf: Yoshimo: do we?
19:23karolherbst: the titan is kepler
19:23mupuf: I am sorry, but we we still have almost nothing
19:23Yoshimo: mupuf: maxwell reclocking is missing i know
19:23mupuf: and remember, we can't even do fan management
19:24Yoshimo: the problem is, if you use swear words now, it might ruin the little progress there is
19:24RSpliet: mupuf: did they stop requiring firmware for the copy engines?
19:24mupuf: no worries, have you seen what gnurou uses for his official presentations?
19:24mupuf: the photo of Linus giving the finger :D
19:25karolherbst: "it's cold out here" :)
19:25karolherbst: or is kepler meant is in the coold guy from way before? :D
19:25mupuf: we all work together, but there is no point in denying that the situation pure sucks :s
19:25mupuf: karolherbst: ?
19:25RSpliet: mupuf: oh I wish it did, now that'd be useful
19:26karolherbst: mupuf: after which Kepler was Kepler named?
19:26RSpliet: the scientist
19:26imirkin_: how many are there?
19:26karolherbst: well there is also the spacecraft :)
19:26imirkin_: which is named after...
19:26karolherbst: right :D
19:26RSpliet: imirkin_: the GPU?
19:27imirkin_: i meant (famous) people named Kepler
19:27RSpliet: imirkin_: no I jokingly implied the spacecraft is named after... the GPU
19:27imirkin_: ah i see.
19:30karolherbst: mupuf: you really have to show of in the repository, don't you? :D
19:30mupuf: karolherbst: ?
19:30karolherbst: you named the folder mupuf_titan
19:31mupuf: right, hmm
19:31mupuf: usually, it would have just been mupuf
19:31karolherbst: right :D
19:31mupuf: well, at least, it is easy to find
19:31karolherbst: yeah, because you will have that many nvf0 soon
19:32karolherbst: mupuf: but is there only one LED?
19:32karolherbst: ohh right
19:32karolherbst: it is like the voltage PWM and the reg(s) for it are somewhere else
19:33karolherbst: uhh nvidia begins to downclock at 78°C
19:33karolherbst: (well 80°C most likely)
19:34mupuf: 07: 121 fps
19:34mupuf: 0a: 191 fps
19:34mupuf: seems like it survived the reclock to f!
19:34karolherbst: ofcourse :p
19:34karolherbst: run furmark and watch the pwoer consumption
19:35mupuf: 34W for now
19:35mupuf: more 35W in average
19:35mupuf: f: 193 fps
19:35karolherbst: seems low?
19:35mupuf:smells the CPU-limited case
19:36mupuf: it is only an i3
19:36mupuf: and a cheap one
19:36mupuf: yep, fully cpu-limited
19:39mupuf: 65W for furmark
19:40mupuf: it is very very fast though :D
19:40karolherbst: tessmark_x16 then
19:40mupuf: 81 FPS :D
19:40karolherbst: well 81 fps for furmark isn't "fast"
19:40karolherbst: or is it full hd?
19:40mupuf: full hd, yes
19:40karolherbst: do pixmark_piano
19:41karolherbst: this has like 0 cpu usage
19:41mupuf: yeah, minimal cpu usage
19:42karolherbst: mhh looking at those nnumbers, I would suggest to increase the boost level :D
19:42mupuf: 13 W
19:42mupuf: err, FPS
19:42mupuf: yeah, I could do that
19:43karolherbst: 850MHz isn't really much
19:43karolherbst: boost 1 isn't much more though
19:43mupuf: karolherbst: what is the parameter I need to set again?
19:43mupuf: boost 2?
19:43karolherbst: near the pstate file
19:43karolherbst: ohh yeah
19:43mupuf: oh right
19:43karolherbst: you can cat the file
19:44mupuf: 2 = 1071 MHz
19:45mupuf: 15 FPS
19:46mupuf: 59W with volplosion
19:46mupuf: but I do not understand, why are we so far from the budget :o?
19:46karolherbst: maybe there is something odd with the reading
19:47karolherbst: Tom^: you had higher numbers, right? .D
19:47mupuf: well, I doubt it
19:47mupuf: 45 FPS on volplosion
19:47mupuf: well, I can test heavn
19:47karolherbst: mupuf: maybe the unk0 field means something else
19:47karolherbst: mupuf: you could try something
19:48karolherbst: mupuf: drm/nouveau/nvkm/subdev/iccsense/base.c:280
19:48karolherbst: remove the !r->mode check
19:48karolherbst: this shouldn't cause the subdev to load on the titan
19:48karolherbst: ohh wait
19:48karolherbst: it doesn't change anything
19:49Tom^: karolherbst: in?
19:49karolherbst: Tom^: power consumption
19:49Tom^: i stopped looking at 290W
19:50karolherbst: mupuf: yeah, I would say the power reading does something bad :/
19:50karolherbst: mupuf: maybe I take a look tomorrow
19:51mupuf: karolherbst: yeah, please do
19:51mupuf: it is at 73°C
19:51mupuf: but 60W of power usage
19:51mupuf: so, there is something funky
19:51karolherbst: somethign isn't right there
19:51karolherbst: there are three rails with 5mohm
19:51mupuf: I will start the blob and see what it has to say
19:58mupuf: yes, 1920*1200, extreme, tesselation normal
19:58karolherbst: and msaa?
19:59mupuf: oh, I do not think I set any MSAA
19:59mupuf: will check
19:59karolherbst: ahh okay
19:59mupuf: checking in the blob now
19:59karolherbst: well I get like 15 fps with mine on extreme, normal tess and 4xmsaa :D at 1920x1080, so the titan is just a bit faster than 300%
20:00mupuf: the fan is way more silent
20:00mupuf: so, there is definitely something we need to do
20:00mupuf: a new table to RE, likely
20:00karolherbst: yeah, cuase the yparse the new fan table
20:01karolherbst: ohh wait
20:02karolherbst: there is no fan or fan_mgmt table
20:05mupuf: 70.1 FPS on the blob
20:05mupuf: 57%, there is definitely work to be done there :D
20:07mupuf: let's see piano on the blob
20:07karolherbst: piano should be like 25% faster
20:08karolherbst: more would surprise me
20:08mupuf: 17 FPS
20:09karolherbst: 13 FPs with nouveau right?
20:10karolherbst: okay, so that means it isn't worse then all the other keplers before that :)
20:12karolherbst: ohh with boost 2
20:12karolherbst: maybe nouveau got better :)
20:12karolherbst: well I have some opts to increase nouveau perf in pixmark_piano by 5%
20:12karolherbst: just saying :D
20:12mupuf: so... there is an option for using double precision
20:13mupuf: so, there is likely something to do in HW to enable it
20:14karolherbst: well I read something, that there are seperated engines or something and you can use either of them
20:14karolherbst: but no idea if that is true
20:14karolherbst: but maybe there are just some specilized f64 units which suck at f32, but that would be somewhat weird
20:15karolherbst: or maybe not and the blob generates full f64 binaries
20:15karolherbst: mupuf: did nouveau suck at the fp64 julia bechmark compared to nvidia?
20:16mupuf: I did not check
20:16karolherbst: does nvidia-smi tell you anything?
20:18mupuf: shit! it does export the power usage!
20:18mupuf: and power budget
20:18mupuf: fuck yeah!!!!!
20:18mupuf: finally we have a gpu where we can do REing on!
20:18karolherbst: power budget is 250W?
20:18karolherbst: ohh wait
20:18karolherbst: let me guess again
20:19karolherbst: yeah, 250W
20:19karolherbst: or 265W
20:19karolherbst: maybe there is a "boosted pwoer budget" which is 265W
20:19karolherbst: and the normal one is 250W
20:21karolherbst: imirkin_: :D your response
20:22mgoodwin: What's the best way to get the latest NVE6 vdpau firmware
20:22mgoodwin: The script is going after 325.15 which is quite old.
20:22imirkin_: mgoodwin: use that script.
20:22imirkin_: there's nothing old about it
20:22imirkin_: that firmware basically never gets updated
20:24mgoodwin: I have an issue with nouveau freezing the entire system so I was trying to make sure everything was as updated as possible
20:24mgoodwin: I come back every few months to test again
20:24karolherbst: mgoodwin: depends on the issue
20:25mgoodwin: https://bugs.freedesktop.org/show_bug.cgi?id=91988 this one in its many forms
20:26mgoodwin: I'll try to find the ones ive commented on, sec.
20:26karolherbst: it is really hard to track those down though :/ I really would like to fix it, but well
20:26karolherbst: I have a clue, but I have no access to a GPU which hangs at random thing
20:26karolherbst: or rather an idea what it might be
20:27mgoodwin: https://bugs.freedesktop.org/show_bug.cgi?id=72180 and https://bugs.freedesktop.org/show_bug.cgi?id=92863
20:27imirkin_: mgoodwin: some people with GK106's have more success with blob ctxsw fw
20:27mgoodwin: see the latter ticket :|
20:27imirkin_: even though ben will deny this :)
20:27mgoodwin: That's mine
20:28mgoodwin: ah looks like you worked that a bit yourself
20:31mgoodwin: I would blame the hardware but it's 100% stable with proprietary drivers
20:31karolherbst: in doubt it is never the hardware :D
20:32imirkin_: determining blame does seem key
20:32imirkin_: i recommend blaming canada
20:33karolherbst: imirkin_: though I found cards which have by default the same voltage than with my patches clocked to 07 and nvidia on lowest perf level :/
20:33karolherbst: imirkin_: maybe there are really some cards which boot at too low voltages
20:34mgoodwin: Have you made any correlation with EVGA cards
20:34mupuf: karolherbst: well, at least, we can check what is going on
20:34mgoodwin: Mine is an EVGA "SuperClocked"
20:34mupuf: and see if the values can be correlated between nouveau and nvidia
20:34karolherbst: mupuf: right, but how much is the power budget ?
20:35karolherbst: mupuf: or was it now related to the nve6
20:35karolherbst: mgoodwin: you never gave us your vbios, right?
20:35mupuf: the power budget reported by nvidia is 250W for the titan
20:35karolherbst: mgoodwin: usually those superclocked cards have a pretty tight volting :/
20:35mgoodwin: I think I did?
20:35karolherbst: mgoodwin: k
20:35karolherbst: mupuf: l
20:36karolherbst: mgoodwin: here in irc?
20:36mgoodwin: https://bugs.freedesktop.org/show_bug.cgi?id=92863 tell me if the attachment on here is adequate
20:36mgoodwin: It has been a *while*, sorry
20:36karolherbst: no, that should be fine
20:38karolherbst: sometimes those guys are a bit crazy
20:38karolherbst: 2561MHz in the vbios
20:38karolherbst: yeah.. course
20:39karolherbst: 1045MHz effective rate at boost 0
20:39mgoodwin: reading from mine?
20:39mgoodwin: Is there a safe way to undo that? i.e. flash it back I guess
20:39mgoodwin: I don't use this for gaming anymore so I just want a stable card
20:40mgoodwin: Or is it too tightly coupled with the third party manufg.
20:40karolherbst: buying a SC card isn't the right choice then :D
20:40karolherbst: nouveau should get it working though
20:40mgoodwin: Things change, had this since 2012 I think
20:40karolherbst: mgoodwin: k, you could try out something
20:40karolherbst: mgoodwin: did you cloned my repository?
20:41karolherbst: mgoodwin: we could try to rule at least the voltage out as the point of failure
20:41mgoodwin: User level points of contentions seem to be vsync and vdpau
20:41mgoodwin: It always crashes eventually when I use vdpau
20:42mgoodwin: and I just had to turn off Vsync in Plasma
20:42mgoodwin: (Just crashed 3 times in a row)
20:42karolherbst: mgoodwin: but you also should run nouveau, but you don't need to login, I would like to know the voltage at boot
20:43karolherbst: sure about vsync?
20:43mgoodwin: I don't have a ton of time presently but I can get back in here with my laptop and then do some testing
20:43karolherbst: well it won't take long
20:43mgoodwin: Not totally?
20:43karolherbst: 20 minutes max
20:43mgoodwin: Are you familiar with KDE compositor settings?
20:43karolherbst: mgoodwin: yeah, I use plasma5 myself
20:44mgoodwin: It feels like a crap shoot between rendering backend, EGL, and tearing prevention
20:44karolherbst: mgoodwin: well for the things I need: 1. run a tool while booted with nvidia for some time
20:44karolherbst: mgoodwin: 2. boot with nouveau and run sensors and note down the voltage
20:44mgoodwin: Right now I'm running OpenGL 3.1, EGL, and Automatic and it's not crashing, but I do have tearing
20:44karolherbst: with nvidia?
20:44mgoodwin: nouveau atm
20:44karolherbst: ahh okay
20:45mgoodwin: nvidia doesn't crash at all
20:45karolherbst: did you change the pstate?
20:45karolherbst: then run sensors
20:45karolherbst: and tell me the voltage
20:45karolherbst: mgoodwin: you get tearing with "automatic" as long as you use DRI2
20:46karolherbst: "Full scene repaints" is the option which kind of never tears
20:46mgoodwin: sec, getting laptop - the display doesn't resume from suspend (intel, not your deal) :(( I have bad luck with linux graphics lol
20:46mgoodwin: Right. That's what i was referring to as 'vsync'
20:46mgoodwin: I usually use fill scene repaints, but turned it off
20:46karolherbst: ohh so nouveau is unstbale on your laptop or your desktop?
20:46mgoodwin: crap it's doing it again
20:46mgoodwin: and I didn't touch anything
20:47mgoodwin: windows are flickering and the background is disappearing
20:47mgoodwin: I didn't mean to confuse you, I'm having separate intel issues on my laptop and was just making fun of it
20:47karolherbst: ahh okay
20:47mgoodwin: Im getting my laptop so that I can continue to talk
20:48karolherbst: ohh okay
20:48karolherbst: well the stuff you need to do won't change much though on your desktop
20:48karolherbst: running sensors really doesn't mess anyhting up :D
20:50mgoodwin: oh I captured a screenshot
20:50mgoodwin: It used to just lock up, seems to only do that with OpenGL 2 now
20:52mgoodwin: alright, reading up to do what you asked
20:56mgoodwin: karolherbst: not sure how to get voltage
20:56mgoodwin: seeing fan speed and temp
20:56karolherbst: mhh which kernel are you on?
20:56mgoodwin: cmd: Linux saturn.net.overtmind.com 4.4.8-300.fc23.x86_64 #1 SMP Wed Apr 20 16:59:27 UTC 2016 x86_64 x86_64 x86_64 GNU/Linux
20:56karolherbst: the hwmon voltage interfaces was added in 4.5 :/
20:58karolherbst: well we have tools to read it out though
20:58karolherbst: mgoodwin: do you have git and the normal build tools installed?
20:58mgoodwin: up to you, i can probably run 4.5+ on here depending on what's in rawhide
20:58mgoodwin: i can
20:59karolherbst: then clone this: https://github.com/karolherbst/nouveau.git
20:59karolherbst: and git checkout stable_reclocking_kepler_v5
20:59karolherbst: then run make in the top dir
20:59mgoodwin: I can safely go up to 4.6 [y/n]
21:00karolherbst: thought 4.6 is RC
21:00karolherbst: 4.5 should be fine
21:02mgoodwin: I can grab 4.5.2-301.fc24 from --releasever=24
21:08mgoodwin: That locked on boot
21:08mgoodwin: I'll have to start in multi-user.target
21:09karolherbst: does this sometimes happen that it locks on boot?
21:10mgoodwin: sometimes but not usually
21:10mgoodwin: normally I just get going doing something and then it locks
21:10karolherbst: well you can run sensors from a tty anyway
21:10karolherbst: or through ssh
21:10mgoodwin: Something that I reproduced more than once was a lock up while mousing over panel icons real fast
21:11karolherbst: this sounds like a general thing though
21:11karolherbst: like whenever there is some more load on the gpu
21:12mgoodwin: GPU core: +0.86v
21:12mgoodwin: but that's in non-graphical mode
21:12karolherbst: doesn't matter
21:12karolherbst: so it boots with 0.86v :/
21:12karolherbst: that's... low
21:12RSpliet: karolherbst: not uncommon I think
21:12karolherbst: RSpliet: this is low
21:13karolherbst: RSpliet: the right entry is: voltage_min = 825000, voltage_max = 925000
21:13RSpliet: I've seen my GTX 780 boot with a similar voltage I think
21:13karolherbst: and usually the effective voltage is more close to max
21:13karolherbst: RSpliet: yeah,but it also depends on the clocks set
21:13mgoodwin: what's the other way to verify voltage
21:14karolherbst: mgoodwin: do you have envytools installed?
21:14mgoodwin: now i do
21:15karolherbst: I have to think on how to read out the speedo again
21:16karolherbst: mgoodwin: nvapoke 0x21000 40040001 && nvapoke 0x122634 0 && nvapeek 0x0214a8 && nvapoke 0x122634 0x41 && nvapoke 0x21000 40040000
21:16mgoodwin: nice, now I want to ssh
21:16mgoodwin: :P sec
21:17karolherbst: RSpliet: but those cards have not that high clocks in general, too
21:18karolherbst: RSpliet: GTX 770 has like 20% higher clocks than a 780 ti
21:18mezo: is it btw. normal that my system freeze alot while chaning pstate? especially if something like a game uns in background?
21:18karolherbst: mezo: yes, because my patches didn't land yet :p
21:19karolherbst: mezo: either compile nouveau from my branch or wait til 4.8
21:19mezo: and how far is autoreclock away?
21:19karolherbst: mezo: stable reclocking is the requiernemnt for this
21:19karolherbst: mezo: otherwise it could be implemented in a week or something
21:19mezo: https://github.com/karolherbst/nouveau/tree/stable_reclocking_kepler_v5 <-- u talk about this? im on this branch
21:19karolherbst: mezo: right
21:20karolherbst: mezo: well there are also otherissues sadly
21:20karolherbst: mezo: but is it only unstable while changing
21:20karolherbst: mezo: or does it also crash later on?
21:20mgoodwin: returned with 000214a8: 000006bb
21:20karolherbst: mgoodwin: k...
21:21mezo: puh. good question. im sure i had some crashes here and there later on
21:21karolherbst: mgoodwin: so thiswould lead to 0.8625V :D
21:21mezo: but not 100% sure
21:21mezo: but mostly while chaning clock
21:21mezo: and something is running in background
21:21karolherbst: mezo: well the question is rather, if you don't touch the clock at full load, and you are able to run at full load for several days
21:21karolherbst: mezo: then this is already better than stock nouveau
21:22karolherbst: mezo: there are some display related things not working right
21:22karolherbst: mezo: sadly I have only a optimus setup and was able to reclock over millions of time while something was running at full load :D
21:24mezo: because there is no autoreclock i only change clock for gaming. so most of the time im on lowest clock anyway
21:24karolherbst: mezo: yeah I understand
21:25karolherbst: mezo: but maybe you could just run it for two or three days on max clock and see how well it goes?
21:25mezo: nonono ;) the fan goes crazy on max clock
21:25mezo: cant handle this
21:25karolherbst: ahh okay
21:25karolherbst: mupuf: your call :p
21:32mupuf: karolherbst: my call for what?
21:32mupuf: sorry, my internet just dropped
21:32karolherbst: fan management :)
21:37mgoodwin: karolherbst: i built your project and installed it
21:38karolherbst: mgoodwin: you run make in top dir I guess?
21:40karolherbst: mgoodwin: LD_LIBRARY_PATH=lib bin/nv_cmp_volt
21:40karolherbst: as root
21:42mupuf: karolherbst: I am done with the titan, want me to turn off the machine?
21:42mupuf: hakzsam: are you ok with me leaving the titan there?
21:42karolherbst: mupuf: mhh I think I could take a loot at the power stuff
21:42mgoodwin: That didn't build :\
21:42mupuf: or shall I plug another low power card?
21:42karolherbst: mgoodwin: did you change the branch?
21:42hakzsam: mupuf, yeah sure
21:43mupuf: hakzsam: sure for what? Want another gpu or one is fine?
21:43hakzsam: mupuf, it's fine
21:43mgoodwin: nope, which one
21:43mupuf: like, the gk208 could be plugged
21:43karolherbst: mgoodwin: stable_reclocking_kepler_v5
21:43hakzsam: mupuf, I will probably ask you for a fermi or a maxwell this week but not today :)
21:44karolherbst: mupuf: now we can RE those sense stuff the right way, yay!
21:44mgoodwin: ok works
21:44mupuf: karolherbst: yep!
21:45mgoodwin: Where to go from here
21:45mupuf:will send an email to the donator to say we got it and it works
21:45karolherbst: mupuf: tell him also that we can do stuff with it, we weren't able to do before .D
21:45karolherbst: though maybe a cheap Quadro would also do
21:45karolherbst: but hey, it was his decision
21:48mgoodwin: 862500, 852447, -10053, 98.834435, 7, 0, 40
21:49karolherbst: okay, as I thought, nouveua would volt to the same
21:49karolherbst: but maybe the reclocking process in itself would make it stable
21:49karolherbst: mgoodwin: mind trying out nouveau from my branch and check if that's more stbale?
21:49mgoodwin: I'm supposed to be running nouveau though right
21:49mgoodwin: Or did you want me to do that with nvidia too
21:49karolherbst: doesn't matter much now
21:51mgoodwin: i'll try
21:51karolherbst: for that go into drm
21:51karolherbst: and run make there
21:52karolherbst: and you might need to install the source/header package of your kernel
21:52mgoodwin: ok done
21:53karolherbst: building or installing?
21:53karolherbst: there should be a file called nouveau.ko inside nouveau
21:54karolherbst: and there sould be a nouveau.ko(.xz) file inside /lib/modules/$kernel_version/kernel/drivers/gpu/drm/nouveau/
21:55mgoodwin: compress and overwrite that?
21:56mgoodwin: I've built kernels but I don't usually replace individual modules, that's newish to me
21:56mgoodwin: So whatever you recommend
21:56mgoodwin: obv need to redo initramfs too
21:58karolherbst: mgoodwin: you don't need to compress it
21:58karolherbst: mgoodwin: just remove the system one
21:58karolherbst: place the new one there
21:58karolherbst: and regenerate initramfs
21:59mgoodwin: holy hell
21:59mgoodwin: is it normal for that kmod to be 113M
21:59karolherbst: if debug symbols aren't split out
22:01karolherbst: mgoodwin: CONFIG_DEBUG_INFO_SPLIT helps a lot for this .D
22:04mgoodwin: ok 15M better
22:08mgoodwin: lockup on graphical boot
22:08mgoodwin: but I didn't get to verify im using yours yet
22:10mgoodwin: looks like it
22:17karolherbst: mgoodwin: well try to reclock once
22:17karolherbst: mgoodwin: /sys/kernel/debug/dri/0
22:17karolherbst: mgoodwin: there should be two files
22:17karolherbst: pstate and boost
22:17karolherbst: try to echo 07 into pstate
22:17karolherbst: and see if that makes it stable somewhat
22:21mgoodwin: it was locked up when i did it and it made teh screen flicker
22:21mgoodwin: guess i'll have to try it again from console then start x
22:22mgoodwin: 07: core 324 MHz memory 648 MHz
22:23karolherbst: mgoodwin: or maybe nouveau just forgets something which is liker super important on your gpu
22:23mgoodwin: what's weird is that I'm not getting any dmesg output
22:24karolherbst: mgoodwin: wait some time
22:24karolherbst: mgoodwin: or mhh
22:24karolherbst: mgoodwin: reclock to 0a and 07 :D
22:25karolherbst: I am serious though
22:26mgoodwin: the 07 pstate is identical to boot
22:26mgoodwin: based on nv_cmp_volt
22:26mgoodwin: 0a is a little higher
22:29mgoodwin: NvClkMode=10 ?
22:29karolherbst: mgoodwin: well try NvClkMode=7
22:29karolherbst: this should also be stable and if not, we miss something
22:30mgoodwin: question though
22:30mgoodwin: 07 is the same voltage
22:30mgoodwin: aren't we targeting a higher voltage?
22:31karolherbst: well it is the same nvidia might use too, except I did something terribly wrong
22:31karolherbst: mgoodwin: but I had also somebody who had the same issues
22:31karolherbst: but after a reclock it was better
22:31karolherbst: even after the voltage stayed the same
22:31karolherbst: we don't reclock on boot
22:31karolherbst: nvidia does
22:31mgoodwin: I see, that's weird.
22:31mgoodwin: (why reclock on boot?)
22:32mgoodwin: I mean why would *they* do that
22:32mgoodwin: So then: options nouveau pstate=1 NvClkMode=07
22:33karolherbst: you don't need pstate anymore with 4.5
22:34karolherbst: it is always in debugfs now
22:35mgoodwin: So traditionally it stayed at the same clock that bios/uefi booted with?
22:35karolherbst: kind of yes
22:37mgoodwin: locked at boot
22:37mgoodwin: ssh'd in, used nv_cmp_volt
22:37karolherbst: there is one thing you could try though
22:37mgoodwin: it output ONCE then halted
22:37mgoodwin: gpu is really stuck
22:37karolherbst: ohh wait
22:37karolherbst: you did the NvClkMode thing wrong by the way
22:38karolherbst: options nouveau config=NvClkMode=07
22:38mgoodwin: ah right
22:38karolherbst: and you have to regenerate initramfs
22:38karolherbst: but after that there is something else we can try
22:38mgoodwin: yeah, that I have down
22:40mgoodwin: lock at boot is probably a misnomer too
22:40mgoodwin: It locks when it gets into the desktop
22:40mgoodwin: to be precise
22:42mgoodwin: [ 2.347554] nouveau: unknown parameter 'NvClkMode' ignored
22:42mgoodwin: oh wait did I not fix that
22:42mgoodwin: I did, and i know I regen'd
22:43karolherbst: mgoodwin: /sys/module/nouveau/parameters/config
22:44mgoodwin: going to inspect initramfs and make sure before I reboot again, sec.
22:45mgoodwin: strangely it didn't lock up that time despite apparently not getting the right options
22:45mgoodwin: go figure
22:45mgoodwin: locked up that time
22:46mgoodwin: and NvClkMode=07 is in the sys parameters
22:49karolherbst: then do this:
22:49karolherbst: options nouveau config=NvClkMode=07,NvVoltOffsetmV=+
22:49karolherbst: options nouveau config=NvClkMode=07,NvVoltOffsetmV=+150
22:50mgoodwin: I was trying =10 while waiting and it hasn't crashed yet. Trying your changes next...
22:53mgoodwin: Never thought to reclock at boot though
22:53mgoodwin: Perhaps PGRAPH + reclocking will be stable
22:54karolherbst: mgoodwin: maybe
22:54mgoodwin: [ 2.394470] nouveau 0000:01:00.0: volt: Volt Offset applied: 150
22:54mgoodwin: 1012500, 852447, -160053, 84.192296, 7, 0, 42
22:55mgoodwin: What I find somewhat amusing is that before back at the end of 2015, it could take an hour or so while browsing or something to lock up
22:55mgoodwin: Whatever changes have been made aggravate it sooner it seems
22:55mgoodwin: Stable so far...
22:57mgoodwin: lol :|
22:59mgoodwin: [ 242.923162] nouveau 0000:01:00.0: fifo: read fault at 97dc410000 engine 00 [GR] client 0d [GPC1/GCC] reason 00 [PDE] on channel 9 [023f569000 plasmashell]
22:59mgoodwin: [ 242.923169] nouveau 0000:01:00.0: fifo: gr engine fault on channel 9, recoverin
22:59mgoodwin: this time
23:00mgoodwin: @ karolherbst
23:15karolherbst: mgoodwin: me :/
23:16karolherbst: mgoodwin: it would be weird if that one is reclocking or voltage related at all
23:43mgoodwin: for whatever reason this hasn't crashed yet: config=NvClkMode=10,NvVoltOffsetmV=+120,NvForcePost=1
23:43mgoodwin: I also extracted firmware for VDPAU but am not loading it yet
23:51karolherbst: mupuf: k, out sensor code is good
23:51karolherbst: mupuf: I removed all but the rail we parse, and I still get the same power reading
23:53karolherbst: mgoodwin: mhhh
23:53karolherbst: mgoodwin: at least something
23:53karolherbst: mgoodwin: the voltoffsetmv option is a safe one though
23:54karolherbst: mgoodwin: it just increases the voltage inside special ranges for each clock
23:54karolherbst: mgoodwin: you still get the same max voltage, so that the max clock just drops