01:13Dezponia: Sweet! With the 4.4 git kernel my GTX TITAN Black no longer crashes when I run OpenGL. Havent managed to build a new mesa though so still on 11.0.7, llvm-libs-svn constantly fails when building.
01:13Dezponia: Still its stabel and plays Xonitc on max everything at 40-60FPS :)
01:17Dezponia: Guess whatever that crash was on 4.2.5 was solved later. Will try 4.3 when it hits arch stabel (currently in testing)
01:27koz_: Dezponia: You clearly have all the moneyz for graphics hardware.
01:28Dezponia: koz_: Not really, I'm just trying to bunker up on the last "good" freedom respecting hardware I can find, as the future looks bleak from both intel and amd and ARM is an unknown right now but knowing the embeded industry, it looks very bleak
01:29koz_: Fair enough.
01:29Dezponia: I might be interested in knowing what sort of hardware would be useful for nouveau devs to get their hands on though
01:30Dezponia: RSpliet mentioned he had to borrow a fermi card for his project yesterday. Without access to hardware I can totally understand if development hits a few snags. But thats a problem that can be fixed and development eased (even if manpower/time will always be the bottleneck)
01:30Yoshimo: im always told that they lack time not hardware
01:31Dezponia: Thats harder to fix :) but for specific instances hardware can be fixed :)
01:33Dezponia: I suspect I have to get a newer mesa compiled to really try out Heaven and such but I'm grabbing them regardless to see what does and doesnt work for now :)
01:34Dezponia: Just super psyched the desktop doesnt freeze because something dared try opengl on kernel 4.4 :P
01:36Yoshimo: thinking about yesterday Dezponia, mr. Spliet might make use of a couple of fermi cards if i read that right
01:37Dezponia: Yoshimo: Thats what I was refearing to
01:37Dezponia: Yoshimo: I'd assume it would be most useful to have a cross section of various cards from different "code names"
01:38Dezponia: Like a GF104, GF110 and GF114 seems interesting
01:41Yoshimo: fermi cards are very cheap these days indeed
01:42Dezponia: Seems like it, even the top card, the 580 3GB edition is fairly cheap
01:42Yoshimo: 20€ for a 560ti ;)
01:42Dezponia: Only real question is would a dev prefer the stock reference pcb cards or those fancy dual fan super extreme oc ultra super edition cards?
01:43Dezponia: I would guess that the stock pcb would be the most representative of the average card
01:44Yoshimo: on the other hand the non.standard things might be the more intresting ones
01:44Dezponia: Best to ask from case to case I guess :)
01:44Dezponia: Either way I'm very partial to buying the ones with the most VRAM (surprise surprise)
02:10Dezponia: karolherbst: Hai there. My card is no longer crashing when hitting opengl on the 4.4-git kernel :)
02:10karolherbst: imirkin: if somebody wants to get 0f working, tell them to use the 4.4 kernel :D
02:11karolherbst: none of my branches will compile with 3.16 anymore I think
02:11Dezponia: karolherbst: Have not managed to build mesa because the llvm-libs-svn build wont work but whatever, it works, tesselation also works in heaven. No idea what clock I'm running at but its probably low.
02:11karolherbst: Dezponia: boot with nouveau.pstate=1
02:11Dezponia: oh right, hurr durr
02:11karolherbst: then you get a file named pstate inside /sys/class/drm/card0/device
02:12Dezponia: karolherbst: I have not tried 4.3 but its in Arch's testing repo so should hit stabel in a few days, will test agian then
02:12karolherbst: you will need 4.4 for 0f pstate anyway
02:12Dezponia: Oh right, nevermind then :P
02:14Dezponia: Alrighty, time to see what explodes
02:14karolherbst: well you won't be able to clock to 0f anyway, because it will get a volting error
02:15karolherbst: thing is, we know a way to fix this, but this can lead to heat issues and nouveau doesn't do anything to prevent the gpu from overheating
02:15karolherbst: well the gpu itself does, but currently the thresholds are just too high
02:15karolherbst: maybe not
02:16karolherbst: well there is also a downclock and max fan threshold, but if the temperature would increase further nouveau don't do anyhting itself
02:16Dezponia: Well i see the pstate file, I assume the AC value is the current clock?
02:16Dezponia: So lowest right now, no surprises
02:16karolherbst: your card should reclock to 0a just fine though
02:17Dezponia: What about 0d?
02:17karolherbst: same as 0f I think
02:17Dezponia: Looked like it, cant see the difference
02:17karolherbst: most of the time there is none
02:17karolherbst: I only found one card where there was one
02:17karolherbst: and the memory clock was -1000MHz
02:18karolherbst: you want to run the unigine benchmark
02:18karolherbst: and check the gpu temp with sensors
02:18karolherbst: and check that your fan is working
02:19Dezponia: karolherbst: temp 53C, 810rpm fan and I can see it spinning
02:19Dezponia: No reclock yet though, not automatic?
02:19karolherbst: not yet
02:19karolherbst: I have a working prototype though
02:20karolherbst: but we have to deal with all the other issues first
02:20Dezponia: Alrighty, so what do I do to test the reclock thingy then? :)
02:22karolherbst: echo 0a into the pstate file :D
02:22karolherbst: you can just select the pstates that way
02:23Tom^: echo 0a | sudo tee /sys/class/drm/card0/device/pstate
02:24Dezponia: Seems to have worked, screen flashed for a few frames and fps jumped by 15 or so
02:24Dezponia: temps 61 fan still 780-810 rpm
02:24karolherbst: the screen shouldn't flash though
02:25Tom^: wait Dezponia which kernel version?
02:25karolherbst: yeah, your fan might not work yet, there is a patch though
02:25Dezponia: Tom^: 4.4 git
02:25karolherbst: Dezponia: you have envytools installed right?
02:25Dezponia: karolherbst: yes
02:25Tom^: karolherbst: hm didnt the fan patch land in 4.4 ?
02:25karolherbst: Dezponia: do nvaforcetemp 89 and check sensors again
02:25karolherbst: Tom^: don't think so
02:25Tom^: ah that explains that then
02:26karolherbst: Dezponia: fan still at 810?
02:26Dezponia: karolherbst: Yepp
02:26karolherbst: then do nvaforcetemp 91
02:26karolherbst: this should boost your fan then :D
02:26Dezponia: fan into overdrive
02:27karolherbst: then nvaforcetemp 0
02:27Dezponia: 5000rpm and now its going down
02:27Dezponia: fan back down
02:27karolherbst: yeah, this means you just need that patch Tom^ linked
02:28karolherbst: without this nouveau will have a funny behaviour if you clock to max :D
02:28karolherbst: like it waits until the temperature reaches 90°C
02:28karolherbst: then it boost your fans to max
02:28Dezponia: Pfff, thats finnneeeeee :P
02:28karolherbst: and then after the temperature felt it sets the fan to min speed agian
02:28karolherbst: I think it is just a bit annoying :D
02:28Dezponia: Well yeah :P
02:29karolherbst: Dezponia: then try out 0f, but I think this will just fail
02:29karolherbst: the memory clock should go up though
02:29karolherbst: the core clock will just stay at 0a level
02:29Tom^: and core, but you will make it freeze on load.
02:29Dezponia: karolherbst: No need to adjust the fan in preperation or such first?
02:29karolherbst: Dezponia: well with 90°C the fan will boost
02:29karolherbst: your gpu shouldn't get any hotter than 90°C with that
02:30Dezponia: Works fine for now, still running
02:30Dezponia: Core has not moved though, just memory
02:30karolherbst: yeah as I said :p
02:30karolherbst: and if you check dmesg
02:30karolherbst: there will be some volting issue -22 or something
02:30Tom^: but didnt you change that? or has that also not landend :p
02:31Dezponia: karolherbst: Correct
02:31Tom^: im so spoiled running your branches. ^_^
02:31Dezponia: Temp at 73 now, slowly climbing
02:32karolherbst: Tom^: not yet
02:32Dezponia: This is cool though, looking forward to seeing what 4.5 brings and such, perhaps then things will settle down :)
02:32karolherbst: Tom^: we have to setup some overheating stuff first
02:32karolherbst: yeah maybe
02:33Dezponia: 82C and rising :)
02:33karolherbst: don't expect this to be running perfect before 4.6 though :D
02:33Dezponia: Thats fine. I'm in no rush
02:34Dezponia: Seeing some textures flashing red in heaven
02:34Dezponia: Hit 90 and the fan kicked in
02:36karolherbst: Dezponia: yeah I know, but this happens just like never, does it?
02:37Dezponia: Super rare, seen it 1 time in the flyby mode and 1 time when I drove the camera into a wall and the floatin island sort of pushed it into the roof
02:38Dezponia: Seems to only happen on the floating fortress for some reason
02:41Dezponia: Is there some special resolution or so you guys use for getting good FPS numbers to compare?
02:43karolherbst: Dezponia: full hd fullscreen ultra settings + 4x msaa :D
02:43Dezponia: karolherbst: Tesselation extreme?
02:47karolherbst: Dezponia: why not, thought it looks a bit strange :D
02:48Tom^: tesselation normal :p
02:48Dezponia: Tom^: Thanks, just looking to get some numbers that you can compare with
02:48Tom^: Dezponia: http://i.imgur.com/IIZVfd1.png
02:48Dezponia: Running a benchmark set now with extreme to see what happens
02:50Dezponia: 11.8 average, 5.2 min and 24.7 max with extreme tesselation :P
02:52Dezponia: Normal tesselation certainly does look more... normal :P
03:01Dezponia: 10 min, 25.2 max, 13.7 average with Tom^'s settings
03:01Dezponia: As it turns out core clock is important, who knew? :P
03:01Tom^: yea im running at 0f :p
03:02karolherbst: shouldn't matter that much though
03:02karolherbst: I mean how many MHz are missing? 200?
03:02Tom^: from my 0a to 0f around 800mhz
03:02Dezponia: I was running at 324 core
03:02karolherbst: that many
03:02Tom^: yes that many ;)
03:02karolherbst: 324 core is a bit slow though
03:02Dezponia: Normal "base" is 889 on this card
03:02karolherbst: Dezponia: did you also get 324MHz on 0a?
03:03Dezponia: It never changed
03:03Dezponia: voltage error thingy
03:03karolherbst: that's strange
03:03karolherbst: let me check
03:03karolherbst: ha :D
03:03karolherbst: all your pstate have the same core clock
03:03karolherbst: except 07
03:03karolherbst: that's unusual for high end cards
03:04karolherbst: oh meh, you could compile nouveau yourself with one of my branches and the fan patch
03:04karolherbst: but because I know that this will work, it is just a thing you should do if you are up to :D
03:05Dezponia: karolherbst: I'd be glad to if I had any clue how. Also as I mentioned I'm still stuck on mesa 11.0.7 since I wouldnt not get the deps built to get 11.2
03:05karolherbst: mesa doesn't have anything to do with that
03:05karolherbst: just the nouveau kernel module
03:05Dezponia: But mah fps!
03:06karolherbst: the issue is this basically: the vbios tells us to clock to a cstate, which needs a voltage higher than 1.215V
03:07karolherbst: I don't know why the vbios does this, but it seems to be a common thing
03:08Dezponia: Hmm, broke a fanblade on a case fan while poking around at it. Guess that gives me an excuse to upgrade that fan to a PWM one since it was the only one that was still voltage controllered
03:08Dezponia: Breaking stuff for science!
03:11Dezponia: karolherbst: So about that nouveau kernel module
03:18karolherbst: Dezponia: git clone https://github.com/karolherbst/nouveau.git -b stable_reclocking_kepler
03:25karolherbst: cd drm
03:27Dezponia: That did not work :P modules recepie failed
03:29Tom^: yea comment out that silly tegra include
03:29Tom^: karolherbst: what provides that anyways? tired of having to comment it out. i want it provided. ;)
03:30Tom^: is it some kern config not enabled in arch kernels?
03:30Dezponia: Tom^: Where do I comment that out? :P
03:31Tom^: Dezponia: the place it errors in, dont recall it but its in a .h file
03:31karolherbst: Tom^: no clue
03:33Dezponia: Think it might be a problem with the 4.4 kernel I'm using, you're on 4.3 right Tom^?
03:33Tom^: Dezponia: pastebin the errors
03:36Tom^: i would say you are just missing your kernel headers
03:36karolherbst: most likely
03:36Dezponia: Ah right, never built those. Could be useful for this I guess... :P
03:36Dezponia: I'm not a developer if thats not clear :P
03:37Tom^: dont worry im not that either.
03:37Tom^: we are all in the same swedish #välfärd
03:37Dezponia: Learn ALL the things!
03:38Tom^: you can run LANG=C make , to make the make output more karolherbst friendly. :p
03:38Tom^: i think
03:39Dezponia: As it turns out the headers are quite large. Oh well I'll wander off and do other things while waiting for them to download
04:42karolherbst: Dezponia: and?
04:45Dezponia: karolherbst: Doesnt build
04:45karolherbst: ohh wait
04:45Dezponia: I mean the kernel headers :P
04:45Dezponia: They dont build either :P
04:45karolherbst: is there a makefile inside /lib/modules/4.4.0-rc4-g800f1ac/build?
04:46karolherbst: you need the kernel headers for the kernel you are running
04:47karolherbst: otherwise you could also switch to a 4.3 stable kernel
04:47karolherbst: my branch has most of the 4.5 changes already anyway
04:49Dezponia: We'll see, I'll poke around a bit more at it, see what explodes
05:05karolherbst: Dezponia: well I would suggest you to just use the normal kernel package and the headers
05:05karolherbst: this will work for sure
05:11Dezponia: Pulling in 4.3.2 from arch testing
05:15Dezponia: karolherbst: Seems like 4.3.2 also works, no crash on login with KDE set to opengl mode
05:16Dezponia: And so far its building your thingy
05:23Dezponia: karolherbst: Ok, built from the makefile in the DRM folder
05:23Dezponia: Now what? :P
05:23karolherbst: now you have to install the module
05:23Dezponia: Where is the thing and where do I put the thing?
05:23karolherbst: mabe make install works if you regenerate your initramfs
05:23karolherbst: yeah we could try that out
05:23karolherbst: make install
05:24karolherbst: regenerate initramfs
05:24Dezponia: From drm?
05:24Dezponia: To many makefiles :P
05:26Dezponia: Forgot to set pstate but whatever. Can I see if its properly runnng or reboot again with the pstate line?
05:26karolherbst: you need the pstate line
05:28Dezponia: Alrighty, back again
05:30karolherbst: Dezponia: is there a line like "nouveau 0000:01:00.0: clk: base: 705 MHz, boost: 797 MHz" for you in dmesg?
05:31karolherbst: if so, you can start testing stuff, if not, then you have to wait, cause i am hungry :D
05:31Dezponia: Not even sure I'm running the nouveau thing I built, seemed to install
05:31karolherbst: well then you need to overwrite the nouveau.ko(.xz) file inside /lib/modules/
05:32karolherbst: for the 4.3 kernel
05:32karolherbst: inside kernel/drivers/gpu and so on
05:32Dezponia: "Initialized nouveau 1.3.0" from the dmesg if thats helpful
05:32karolherbst: it should be 1.3.1
05:32karolherbst: or maybe 1.3.2
05:38Dezponia: Done, works
05:38Dezponia: [ 11.802032] nouveau 0000:01:00.0: clk: base: 888 MHz, boost: 980 MHz
05:39Dezponia: [ 12.645525] [drm] Initialized nouveau 1.3.1 20120801 for 0000:01:00.0 on minor 0
05:41Dezponia: karolherbst: Now then? :)
05:42Dezponia: karolherbst: I changed the pstate and the screen still flashed black for a fraction of a second but now the fan kicks up
05:43Dezponia: seems the fan is working as intended right now at any clock level
05:43Dezponia: Core also reclocks properly at 0a
05:44Dezponia: on 0f now, core and memory reclocking properly and heaven running well
05:44Dezponia: Fan around 2300rpm at 61C
05:47Dezponia: Its not turboboosting to 980 or above though, rock solid on 887-888
05:57Dezponia: Not quite Tom^'s levels of good but certainly miles ahead of before. 36.4 frames average, 26.1 min and 67.8 max
05:59karolherbst: Dezponia: now boot with nouveau.config=NvBoost=2
05:59karolherbst: this should clock up to 980MHz then
06:00Dezponia: Neat, will do
06:03pmoreau: karolherbst: Would it be possible/interesting to change NvBoost similarly to pstate?
06:05karolherbst: pmoreau: this is more of a hack
06:05Dezponia: karolherbst: works, on 979 core
06:05Dezponia: 978 even
06:06karolherbst: Dezponia: would be nice to know how hot the gpu will get with unigine
06:07Dezponia: karolherbst: Will check, doesnt seem unresonable so far
06:08Dezponia: karolherbst: Running the benchmark now at fullscreen 1080 4xAA ultra and normal tess
06:21Dezponia: karolherbst: New benchmark run, fan seems to top out around 2910rpm at 67C (never seen it go above this for either RPM or C). Results where 40.2fps average, 27.9 min and 75.0 max
06:22Dezponia: I should probably get an Intel i7 3770 to pair this with, the i7 2700K does not have PCI-E (this computer was not built with this card in mind)
06:23Dezponia: The GTX680 I previously had was right at the edge of what PCI-E 2.0 could handle and a faster card like this can make use of 3.0
06:24karolherbst: pcie doesn't make much of a difference
06:24karolherbst: nouveau sets your pcie card speed to 2.5 only anyway
06:24Dezponia: karolherbst: I know :P
06:25karolherbst: now you could boot with nouveau.config=NvBoost=3
06:25Dezponia: Thats the 1000ish core mark?
06:25Dezponia: 1050* ?
06:25karolherbst: something like that
06:25karolherbst: not quite sure
06:25karolherbst: didn't checked carefully enough for that
06:26Dezponia: Weill do, I'm happy already :)
06:27karolherbst: pmoreau: NvBoost=3 will be the default later anyway, the only problems are heat management
06:27karolherbst: and maybe sane reclocking
06:27karolherbst: but also the blob uses the highest clock possible as long as the temperature stays inside valid ranges(and power usage)
06:27Dezponia: temps seem good though, I knew Nvidia had improved their stock cooler with the 600 series and much more with the 700 series and that seems to be true
06:28karolherbst: well 70°C is a bit high already
06:29karolherbst: I think nvidia tries to target 75 for desktop cardS? maybe 80
06:29karolherbst: ohh wait
06:29karolherbst: I can check with your vbios :D
06:29Dezponia: I would guess 80
06:30Dezponia: 0f is now 1084mhz
06:30karolherbst: 82.5 °C is the target
06:31karolherbst: and 95°C is the "max"
06:31karolherbst: and max means: downclock to lowest clock
06:31Dezponia: Going to check what 0f brings me just sitting on the desktop first with these clocks
06:31karolherbst: and between 82.5 and 95 downclock a little, but currently we don't know how exactly
06:31karolherbst: a lot of heat
06:32Dezponia: karolherbst: Look I totally NEED 1000fps on my wallpaper or it wont feel smooth when I move the mouse
06:33Dezponia: Seems to calm down at 58C with 2040rpm fan on the desktop
06:34Dezponia: Slowly going to 59c
06:35karolherbst: we can reduce the powr consumption a bit on idle though, but I didn't care enough yet to write a mainline patch for that
06:36Dezponia: Running another benchmark again to see how the temps look after it and how it performs (if stable)
06:37Dezponia: karolherbst: BTW which results am I looking for here? nouveau+shaderpatch? http://i.imgur.com/IIZVfd1.png
06:37Dezponia: Or just regular?
06:38Dezponia: With the nouveau I built from you
06:38karolherbst: the left one I think?
06:38karolherbst: no idea what those shader patches are
06:38karolherbst: but you are pretty close already to those :D
06:38karolherbst: with 1080MHz you should go easily over them
06:39karolherbst: Tom^: sorry :D
06:39Dezponia: So I assume boost 4 would be that 1202 state or whatever?
06:39karolherbst: Tom^: seems like a Titan Black card will beat yours :p
06:39karolherbst: Dezponia: there is no boost 4
06:39Dezponia: Ah, hurr durr
06:39karolherbst: 1 means base clock, 2 means use avg boost clock, 3 means use highest clock supported
06:39Dezponia: I just remember you mentioned the vbios for this card had a cstate for 1202 which is weird :P
06:40karolherbst: nouveau can't use higher clocks because that would mean using a higher voltage than the card is able to handle
06:40Dezponia: So I guess 1084 is the highest this card will boost
06:40karolherbst: you could modify your vbios though :D
06:40karolherbst: yeah something like that
06:40karolherbst: though the nvidia driver won't use this clock
06:41Dezponia: karolherbst: I might do that. Gigabyte already sells this card in a gigahertz edition with only the difference being the base and boost clocks going up 13%. From 889/980 to 1006/1111
06:41Dezponia: They bundled it with a windforce triplefan cooler since Nvidia did not allow any custom cards for the titan black
06:42Dezponia: All titan blacks use the same pcb and cooler. So they got creative and sold it as the a bundle :P
06:42Dezponia: Thats why I thought perhaps using the vbios from the gigaherts edition would work
06:42karolherbst: I bet the blob only uses a clock near 1000MHz though
06:42Dezponia: Same card but cherry picked
06:42karolherbst: could be
06:42karolherbst: but you can easily modify your vbios yourself
06:42karolherbst: there is really no magic to it
06:43Dezponia: As far as I know the only partner that were allowed to change ANYTHING about the titan black was EVGA which is basically Nvidias best friend. They sold a "hydrocooler" edition that came pre-fitted for watercooling but thats it. It also ran at the gigahertz speed
06:43Dezponia: karolherbst: I'll have to learn that some day then :)
06:44Dezponia: I always get a weird drop to 8fps minimum the first run in heaven, doing run 1 now since I guess its some shader caching or whatever
06:44karolherbst: well you can always just OC through the nvidia-driver anyway
06:44karolherbst: +135MHz through coolbits
06:44Dezponia: Never done that :)
06:44Dezponia: Will try, if something explodes I blame Nvidia for shoddy engineering
06:45Tom^: karolherbst: never.
06:45Tom^: im still awaiting results. :P
06:46Dezponia: Tom^: Is your 780Ti a reference card?
06:46Tom^: a what now?
06:46Dezponia: Tom^: Does it use the stock cooler or one of those fancy dual fan thingies?
06:46pmoreau: karolherbst: Sounds good :-)
06:46Tom^: fancy dual things, vbios OC'ed by MSI
06:46Dezponia: Yeah, I figured :)
06:46pmoreau: karolherbst: Don't you have patches to let Dezponia change the PCIe speed with Nouveau?
06:47Tom^: Dezponia: does my world record of unigine and nouvea still stand or did you beat it?
06:48Dezponia: Tom^: Still running but think it might stand :) You are a messurably better CPU and such as well
06:48Tom^: good, good.
06:48Dezponia: 29.9 min, 81.1 max, 43.5 average 1097 score :)
06:48Dezponia: 3 score away :P
06:48Tom^: not even close
06:49Dezponia: Not sure if mesa changes performance any but this is on 11.0.7
06:49Dezponia: Couldnt build 11.2 for some reason
06:49karolherbst: 11.2 will be better for sure
06:49karolherbst: Dezponia: isn't there any newer package available?
06:50karolherbst: pmoreau: yeah well, we can just add 5% with that, but that's all
06:51Dezponia: karolherbst: there is in the AUR but I cant build llvm-libs-svn which is required to build it :P
06:51Dezponia: Need it for llvm-svn which is needed by clang-svn which is needed by mesa-git 11.2 :P
06:51Dezponia: Tom^: How did you build mesa 11.2? Just AUR?
06:51Tom^: Dezponia: mesa-git from aur indeed.
06:52karolherbst: Dezponia: llvm 3.7 is enough
06:52Dezponia: Tom^: Weird, guess it might've broken between your build and now
06:52pmoreau: Dezponia: You will get troubles with building llvm due to an update of gcc
06:52pmoreau: Still trying to solve that… --"
06:52Tom^: that could be that then
06:52Tom^: ABI c++ change eh?
06:52Dezponia: karolherbst pmoreau: I can just just change the aur package to accept 3.7 then?
06:53Tom^: yea i built it before that landed
06:53pmoreau: Dezponia: Yeah, should work
06:54Dezponia: pmoreau: Alrighty, will give it a try, brb with questions about why it wont build :P
06:54pmoreau: There are two patches for LLVM to get around that issue, but… didn't solved the problem after applying them.
06:54pmoreau: Let's see if downgrading to GCC 5.2 helps
06:54Tom^: rewrite LLVM in C instead.
06:55Dezponia: Also relicense it as GPLv3 while at it because thats what I like
06:55pmoreau: Eh! I want to continue hacking on Mesa/Nouveau, not drown myself into rewriting LLMV. :-p
06:55Dezponia: Going to do a quick check what temps I get before I move to building mesa
06:56Dezponia: pmoreau: Rewrite the kernel in rust
06:56Dezponia: And bring the grsec guys with you, those guys are cool
06:58pmoreau: Dezponia: No OpenCL for you! :-p
06:59Dezponia: Seems the card runs at 1084 core 6999 memory with 3390RPM fanspeed and 71C temps in heaven
06:59Dezponia: 1080p 4x AA normal tess ultra, the usual
06:59Tom^: sounds about exactly the same as mine
06:59Dezponia: Neat, except I have a regular Nvidia cooler and no fancy dual fan thing :P
06:59Tom^: or well my fanspeed was quite lower
07:00Dezponia: Tom^: Whats your max clock though? Since it has a factory OC on it I would assume it could go a bit higher than regular cards?
07:00Tom^: 1097 on blob when it boosts
07:01karolherbst: pmoreau: just use the old c++ ABI ;)
07:01Dezponia: Tom^: And nouveau with boost=3?
07:01Tom^: that option didnt exist when i tested
07:02karolherbst: Dezponia: for your interest: mesa is also not GPL, nor is the nouveau kernel module :p
07:02Tom^: iirc i was on 1084 because it was the highest i was allowed by the cstate and volt limits
07:02Dezponia: karolherbst: I know :)
07:02Dezponia: Ah, same as me then I guess
07:06pmoreau: karolherbst: How do you tell that to Mesa?
07:06karolherbst: pmoreau: you have to tell that to gcc
07:07pmoreau: If anyone wants to try running an OpenCL hello_world: https://phabricator.pmoreau.org/F15369 (I haven't had time to check the generated code for Fermi+, and now I can't, so it might not work for those cards)
07:10Dezponia: Hmmm, still no joy building mesa with llvm 3.7
07:10Dezponia: Complains it cant find the llvm libraries but llvm-libs is installed
07:11pmoreau: It complains about not finding the shared llvm libraries I guess?
07:12Dezponia: Found where it wants the llvm-libs-svn in the package build, dident change that first time
07:13karolherbst: Dezponia: that's an pkgbuild issue isn't it?
07:13karolherbst: just don't depend on llvm-libs-svn, because that's stupid
07:14Dezponia: karolherbst: Changed it to normal llvm-libs and we'll see if it explodes
07:14Dezponia: Still explodes it seems
07:14pmoreau: Add --disable-llvm-shared-libs when building Mesa
07:22Dezponia: pmoreau: Where do I stick that part in the packagebuild? :P
07:23Dezponia: Under line 42?
07:24Dezponia: There I see it actually, line 48
07:24pmoreau: Oh right :-)
07:29Dezponia: Now then, lets see if this tihng will work for once :P
07:31karolherbst: Dezponia, Tom^: seems like your cards are basically the same anyway
07:31Tom^: it should be, just his has more VRAM
07:31karolherbst: yeah and more DP perf
07:31Dezponia: And the double float thingy
07:33karolherbst: would be interresting to know if there is a way to just enable it on other gpus :D
07:35Dezponia: Not sure how they lock it, its the same die on both the 780ti and black
07:36Dezponia: vbios lock?
07:39karolherbst: no idea
07:39Tom^: nah doubt that
07:39Dezponia: Magic then
07:39Tom^: ever since radeon did so with their 6950 which could just be vbios unlocked to an 6970 i think ever vendor started doing HW locks.
07:39Tom^: or simply not including the proper pins
07:43Dezponia: Yay! Finally got mesa 11.2 :P
07:43Dezponia: Now to see if it runs :P
07:55Dezponia: Tom^: Score of 1101 ;)
07:55Dezponia: Totally worth all that extra cash now
07:55Tom^: no. :p
07:55Dezponia: Hush you
07:56Dezponia: I'll have the last laught once SuperTuxKart upgrades to mega resolution textures and require 6GB of VRAM!
07:57Tom^: you mean when you actualy have an 8k monitor?
07:57Dezponia: Nope :P
07:57Dezponia: I did run 1920x1080 in triple surround before though
07:58Dezponia: That sort of counts, right?
08:00Dezponia: karolherbst: BTW with mesa 11.2 I have not seen any render issues like the red textures in heaven yet
08:04karolherbst: seems like we really have to fix all those issues now for kepler :D
08:04Tom^: yes. :p
08:05karolherbst: meh :D
08:06Dezponia: What issues? :P So far on 11.2 I have not seen anything weird (yet)
08:07Tom^: dynamic reclocking, proper volt handling. power sensors so it doesnt go above power budget etc.
08:08Dezponia: Ah, right, the things I've been doing manually now :)
08:42chillfan: hey :)
08:51chillfan: I sent over a v-bios file for the gtx 780ti, wondering if that on it's own is helpful for the reclocking support, when I have some time I could try to help a little more with that
08:51imirkin: karolherbst: --^
08:52chillfan: imirkin: cheers :)
08:54chillfan: have some stuff that needs doing here, but i can try to install nouveau from git today for sure
09:00karolherbst: yeah I know :D
09:00karolherbst: chillfan: where did you send the vbios to?
09:00chillfan: the mmio dumps email
09:00karolherbst: usually with newest nouveau and newest mesa the 780 ti should run at least on default clocks
09:00karolherbst: yeah, I have no access to that :D
09:01chillfan: oh okay, somewhere else it should go?
09:02karolherbst: well I won't have much time to look into the issue today anyway, so it would be a good idea to test with a newer software stack first
09:02karolherbst: otherwise Tom^ or Dezponia can also help you, because they have practiacally the same gpu ;)
09:02imirkin: karolherbst: http://filebin.ca/2PpOToo6k4gd/gtx780ti-vbios.rom
09:03karolherbst: imirkin: do you want to add it to the repository? with the strap peek? I really have no time today for this :/ maybe in 4 or 5 hours
09:03chillfan: will do, working on getting this system on the 'edge', should I clone a specific repo or just the one mentioned in InstallNouveau?
09:03imirkin: karolherbst: not particularly... i'm way behind on that repo, updating would be a huge annoying download
09:08chillfan: ok back in a few :)
09:19pmoreau: karolherbst: I could put it in the repo
09:33Dezponia: Ahahahaha, found a way to cheese the heaven benchmark max fps score :P
09:34Dezponia: Bet I could beat the Nvidia windows drivers max fps with this method if I just wanted to
09:53karolherbst: pmoreau: k, thanks :D
09:56Dezponia: Tom^: Heaven benchmark results: https://seafile.dezponia.se/d/d4ff2e0114/files/?p=/Unigine_Heaven_Benchmark_4.0_20151213_1805.html&dl=1 :P
09:57Dezponia: Tom^: You still hold the highest max FPS at least, I've yet to beat it :)
09:58karolherbst: Dezponia: pro tip: skip through every scene with enter unitl you get to the first one again, then start the benchmark ;)
09:59Dezponia: karolherbst: Just starting and stopping the benchmark with F9 and Esc works just as well, restarts at that point
09:59imirkin: karolherbst: shouldn't matter... we precompile all shaders now
09:59Dezponia: karolherbst: Or just take free-cam and point it into the clouds, then you'll start with 100+ max fps and it'll count towards your score
10:00Dezponia: imirkin: For me I did see a difference on run 2 compared to run 1 in the minfps
10:00Dezponia: imirkin: Every run 1 I got 8'ish FPS, every run after that I never went below 29 basically
10:01imirkin: hm ok
10:01imirkin: also, Tom^, you should probably redo yours... i've pushed some stuff wrt correctness that may slow you down :)
10:01Dezponia: imirkin: There are certain spots that tank the FPS for a split second
10:02Dezponia: On the nvidia blob heaven still tanks to 10 fps in some spots on the first run, also runs with a lower clock and higher temp
10:05Dezponia: Interesting, sensors wont pick up the temp on the nvidia blob
10:28karolherbst: imirkin: yeah I know, but unigine adds some shaders later
12:08mupuf: imirkin: hey, for nouveau's shaderdb, wouldn't it make sense to try to track the number of register used? Or you can divide them in pools of PoT numbers that can be used by a thread
12:09mupuf: The latter is of course better but may be unpractical
12:42RSpliet: mupuf: I take it "gpr" is "general purpose register"
12:42mupuf: that would be one way of putting it, I guess :D
12:43karolherbst: mupuf: any idea how I can set the divider caused by the first threshold on the FSRM?
12:44imirkin: mupuf: i do track the number of registers used
12:44mupuf: karolherbst: 20074 does not allow you to do that?
12:44karolherbst: only for the second threshold
12:44karolherbst: didn't matter what I put it in there, the first one always caused 2
12:44mupuf: ah, so it changed since fermi then
12:45sarnex: have you guys heard of nouveau causing cpu core stalls?
12:45mupuf: karolherbst: may the fuzz be with you!
12:45karolherbst: yeah fun
12:45mupuf: that's how I found where the reg was
12:45mupuf: force it to the first level
12:45mupuf: and start fuzzing
12:46mupuf: but start with 20074,
12:46karolherbst: yeah, this will be fun :D
12:46mupuf: it shouldn't take too long
12:46karolherbst: also we have to figure out wich thresholds to use actually
12:46RSpliet: sarnex: pipeline stalls? well, only if it's an in-order CPU that waits for the bus on every reg read/write
12:46karolherbst: yeah I know, not that many regs in there
12:47mupuf: karolherbst: the values? Yes, we need to either find them in the vbios or hardcode them, like the blob does
12:47karolherbst: mupuf: also I didn't find any way to trigger the THRESHOLD_4_DIV
12:47sarnex: RSpliet: im not sure what that means really, basically if i have nouveau loaded(my main gpu is amd, nouveau is used with PRIME sometimes), i get weird behavior then like 10 seconds latter i see cpu stall detected in dmesg andi have to reboot. it doesnt seem to happen if i rmmod nouveau
12:47mupuf: when I say "like the blob does", I do not imply that the blob is hardcoding it
12:48RSpliet: sarnex: oh those kind of stalls, sounds like nouveau is trying to access a GPU that is suspended?
12:48karolherbst: sarnex: trace pls
12:48RSpliet: karolherbst: hold your horses ;-)
12:48sarnex: karolherbst: its unpredictable, but how can i get a trace of this
12:48karolherbst: and with trace I mean stacktrace
12:48RSpliet: sarnex: what kernel?
12:48karolherbst: sarnex: find the stuck kworker
12:49karolherbst: and cat the trace out of /proc/$pid
12:49sarnex: 4.3.2 but i tried all between 4.1 and 4.3.2 and they all bad
12:49karolherbst: there should be a dead kworker
12:49karolherbst: or a stalled one or whatever
12:49sarnex: i cant reproduce i just have to use the kernel for a day or two and see if it happens
12:49karolherbst: ohh wait
12:49karolherbst: the stall in dmesg should also print a stacktrace, right?
12:49karolherbst: I fear the worst, but you could post that one :D
12:50karolherbst: I bet it is a wait call waiting on a kworker to finish
12:50sarnex: ill modprobe nouveua and see if it happens
12:50karolherbst: I bet you have a dmesg log somewhere?
12:50sarnex: oh maybe
12:50karolherbst: well with systemd you have one for sure :p
12:50sarnex: yeah can you tell me how to find it with systemd
12:50karolherbst: and without systemd you also should have one
12:50sarnex: is it journalctl ?
12:50sarnex: got it
12:51karolherbst: mupuf: got it :D
12:51mupuf: karolherbst: that was fast :p
12:51karolherbst: my brain works like a stack :p
12:51sarnex: karolherbst: http://pastebin.com/5kuFJbSa
12:52sarnex: ignore the date all the stalls print the same thing
12:52karolherbst: sarnex: yeah well that is somehow mhhh nouveau unrelated :D
12:52sarnex: are you sure? rmmoding nouveau prevents it?
12:52sarnex: no question mark at the end
12:52karolherbst: well if I say it is nouveau unrelated I mean there isn't nouveau involved in the stack at least
12:53karolherbst: isn't there more?
12:54karolherbst: okay, it seems like the cpu is in a state where it can't put into a sleep state or whatever
12:59sarnex: karolherbst: yeah heres one from yesterday http://pastebin.com/raw.php?i=vEpsd8Vz
12:59sarnex: its upside down because journalctl -r
12:59sarnex: i have no idea how to begin diagnosing this
13:00karolherbst: there should be a dead or stalled kworker
13:01karolherbst: htop can show you this easily if you let it display kernel threads
13:02sarnex: yeah i dont see kworker in htop
13:03karolherbst: you have to disable the option first
13:03karolherbst: "Hide kernel threads"
13:03sarnex: oh sorry i havent used htop before
13:03sarnex: got it
13:04sarnex: how can i find the stalling one with this
13:04karolherbst: then the S coloumn may display D for the broken one
13:05sarnex: operation not permitted to attach but im running as root
13:05karolherbst: why attack?
13:05karolherbst: it is a kernel thread
13:05karolherbst: of course you can't attach to it ;)
13:05sarnex: sorry lol
13:05karolherbst: you have to get the stack from /proc/$pid
13:06sarnex: ok thanks ill insmod and wait for it to die
13:19karolherbst: mupuf: found it :) 2010c
13:19mupuf: here you go, sir :p
13:19mupuf: document that :)
13:20mupuf: mwk: your server seems down: http://ng.0x04.net/~mwk/scans/
13:20imirkin: iirc ng is down for good
13:20imirkin: try www
13:21imirkin: mupuf: http://0x04.net/~mwk/scans/
13:23mupuf: imirkin: thx!
13:23karolherbst: mupuf: I think with PTHERM.I2C_SLAVE.THRESHOLD_8 there can be a interrupt or something be triggered from the gpu to the host, but this is currently in a guessing state
13:23karolherbst: I think with that we could remove that temp polling stuff
13:23mupuf: I am sure you can, that's the spirit
13:24mupuf: I did this for nv50
13:24karolherbst: ohh k
13:24mupuf: it uses the interupts
13:24karolherbst: how can I check if I get such?
13:24mupuf: but I stopped doing it because I was lacking interupts or it was not super safe
13:24karolherbst: or will the pmu send one?
13:24karolherbst: I don't know from where to expect this interrupt so
13:24mupuf: nope, ptherm sends it
13:24karolherbst: ohh okay
13:24mupuf: but please, don't go this way
13:25karolherbst: why not?
13:25mupuf: because that's an optimisation and we have more important stuff to do
13:25mupuf: in the end, the pmu will do the monitoring
13:26mupuf: so, why bother implementing chipset-dependent logic when you can just poll every second on it?
13:26karolherbst: okay, my road is something like that: 1. fsrm ground work 2. find out which thresholds to use 3. finalize fsrm 4. finalize my kepler reclock patches
13:26karolherbst: ohh right
13:28mupuf: 1-3 should be 1% of 4, but sure :p
13:28mupuf: don't try to over-do the fsrm work, it is mostly trivial
13:28mupuf: and even if we hardcode conservative thresholds, that would be enough
13:29karolherbst: yeah I know
13:29karolherbst: 95°C and 100°C? or is this too high already?
13:30karolherbst: ohh wait, we have three values
13:30karolherbst: blob uses 99/102/104 for me
13:31mupuf: on my kepler, it was only two thresholds
13:31mupuf: 97 and 100
13:31karolherbst: which ones? the high fsrm and emergency disconnect?
13:31mupuf: no idea
13:31karolherbst: the low_fsrm one sets a div between 0 and 7, the high one between 8 and f
13:32mupuf: what is this low_fsrm you are talking about?
13:32karolherbst: PTHERM.I2C_SLAVE.THRESHOLD_6 is the low one
13:32karolherbst: PTHERM.I2C_SLAVE.THRESHOLD_2 the high one
13:33karolherbst: the blob uses 99/102/104 °C with 2/e/off as the configured "effect"
13:33karolherbst: well off maybe just means stop all work and disconnect from bus or something
13:40karolherbst: I think my EC just shut down my laptop :O
13:43karolherbst: mupuf: input clock: 405 MHz, div 2: 202MHz, div 3: 166MHz, div e: 50 MHz
13:53Tom^: imirkin: im never upgrading then, ever.
13:54imirkin: Tom^: but... elemental? :)
13:56karolherbst: how does that looks like? http://plotshare.com/sessions/617433473/Plot1.png
13:58karolherbst: ohh I think I got it
13:58karolherbst: maybe not
14:02mupuf: karolherbst: off = cut the voltage regulator
14:02mupuf: this has altready been REed
14:02mupuf: and it is documented and explained in ... the paper :D
14:02karolherbst: okay, I got the formular to calculate the fsrmed freq
14:03karolherbst: real_freq = set_freq / ((fsrm_div / 2) + 1)
14:03karolherbst: means, if 862 MHz is set, and the divider is 15 (max) you get a real freq of 101MHz
14:13mupuf: why / 2?
14:13mupuf: that is weird
14:13karolherbst: no idea, but this is what the data say to me
14:14mupuf: what would be the formula in this form? real = set >> div;
14:14mupuf: find div :D
14:14mupuf: that's what the hw does
14:14karolherbst: well maybe nvidia-settings reports wrong clocks then?
14:16mupuf: no, trust it more than my jibberish
14:16karolherbst: if I divide real/fsrmed_clock I get stuf like: 1, 1.5, 2, 2.5, 3, 3.5, 4... and so on
14:16karolherbst: so this is the only formular which would really make sense
14:16karolherbst: but it is a bit strange though
14:17mupuf: yep :s Maybe they used a DLL instead of divider by two
14:17mupuf: but it definitely was one at the time I reversed it
14:17mupuf: I used the perf counters
14:17karolherbst: yeah maybe it is different with kepler now
14:18karolherbst: I also only get 50% power consumption with the highest divider
14:18karolherbst: well "only"
14:18karolherbst: so that much doesn't change
14:18karolherbst: allthough 50% less powe consumption is still a big win if the gpu gets too hot :D
14:19karolherbst: it makes no difference if the fsrm_div is 2 or e for the performance if you set the threshold to 60°C and keep load running
14:19karolherbst: without forcing a temp
14:19karolherbst: it will make a difference though when the fan still can't keep up with that
14:20karolherbst: but as long as the current temp stays with the thresholds, the performance loss is the same with all dividers
14:20mupuf: it's not supposed to
14:20karolherbst: why not?
14:21karolherbst: higher div: less heat => faster cooling => faster clock restore
14:21karolherbst: well for the eye there was a difference though, but no difference in avarage
14:22mupuf: the FSRM is ridiculous for power efficiency
14:22mupuf: because the voltage is super high
14:22mupuf: just downclock!
14:22mupuf: the FSRM is a safety feature
14:22karolherbst: yeah I know
14:23karolherbst: I was just playing around to check what the real effect is of the FSRM
14:24karolherbst: so I would say we would set two thresholds with two different dividers. Then we should dicide on how much power consumption we want to cut with the low one and set the high one to max
14:24mupuf: don;t try to be smart
14:24mupuf: use what nvidia uses :D
14:24karolherbst: nvidia uses half clock on the low one :D
14:24mupuf: I do not want us to tweak stuff per GPU
14:24karolherbst: so 2
14:25karolherbst: no, me neither
14:25mupuf: 2 and e are what I have seen in nature, but check mmiotraces
14:25mupuf: and then we need to check the thresholds
14:25karolherbst: yeah I also have 2 and e
14:25karolherbst: but we should use 2 and f I suppose
14:25mupuf: if they are constant across a lot of boards, then let's just hardcode them
14:25mupuf: if we can't find them in the bios, we might as well take the lowest values we find and put them there
14:26mupuf: we should never reach this temp anyway
14:26karolherbst: this is for cards with broken fans and stuff I guess
14:26karolherbst: or worse
14:26karolherbst: australian summer, broken fan
14:28karolherbst: mupuf: nvkm_clk_read should be adjusted accordingly?
14:28mupuf: Nah, ignore it
14:28mupuf: devs can just peek the reg
14:28karolherbst: okay, I will add a todo then :D
14:29mupuf: which todo?
14:29karolherbst: no idea
14:29karolherbst: maybe we should create a list of task especially suited for beginners or something, this could be one of them
14:29karlmag: .. there's always a todo... :-P
14:33mupuf: karolherbst: you have edit rights on the trello, right?
14:45karolherbst: right I could add it there then :D
14:46karolherbst: mupuf: I suppose I should also do "PM: Add PCIe link management support" at some point then :D
14:46karolherbst: meh documentation :D
14:47karolherbst: k, will go to bed now anyway
16:42imirkin: nice... got the buffer stuff into a much more working state
16:51imirkin: pretty soon it'll be cleanup time for that feature
16:51imirkin: maybe even tonight
16:51imirkin: [this is ARB_shader_storage_buffer_object + ARB_shader_atomic_counters, for those following at home]
16:52Tom^: im following!
16:52Tom^: but i dont understand :(
16:53imirkin: well, in addition to the functionality presented by those exts themselves, it will also make it possible to start working on ARB_compute_shader and ARB_framebuffer_no_attachments
16:55Tom^: cool i soon have gl 4.2 then :p
16:55imirkin: well, gl 4.2 has images too
16:56imirkin: which is another biggie
16:56imirkin: and also all the surface stuff needs to be worked out for GK110 (it kinda is for GK104 already)
16:56imirkin: on the bright side, i have a GK208 so i'll be able to do it
16:57imirkin: anyways, assuming other people pitch in on some work, it's conceivable that nouveau/radeonsi/i965 will all have GL 4.3 with the next mesa release
16:57Tom^: now we just need some software actually using GL 4.3 ! :P
16:57imirkin: i think unreal does
16:57imirkin: and a bunch of the DX11 games coming out...
17:07chillfan: directx :/
17:08Tom^: lets hope they create more vulkan games then dx11.
17:08chillfan: hm vulkan, I guess I'd need a new card to use it?
17:08imirkin: perhaps i should have said "ports of DX11 games"
17:09imirkin: vulkan's not a thing yet
17:09chillfan: oh ports, over to opengl?
17:09chillfan: that's something eh, what games are they?
17:10imirkin: i dunno... a bunch. GRID Autosport just came out
17:10imirkin: (and had some issues with mesa)
17:10imirkin: i fixed one of them, dunno if there are more
17:10chillfan: ah will have to keep my eyes open
17:11imirkin: there are a few groups doing porting
17:11imirkin: feral, eon, and virtual programming
17:12imirkin: they've all developed various wrapper tech, and are just wrapping games in it
17:13chillfan: glad someone is doing something other than steam then
17:14imirkin: huh? they're on steam...
17:14imirkin: steam is just an application distribution thing
17:14chillfan: no i mean presumably you can get the game without steam?
17:14imirkin: usually yeah
17:14imirkin: i dunno the specifics tbh
17:42Tom^: ugh compiling llvm is slow. i need xeons.
19:53Tom^: im ricing so bad, wasting hours on compiling stuff with optimization flags, even the kernel even tried 1000hz and i didnt gain or lose a single score or fps in unigine. xD
19:55Tom^: Dezponia: you have me beaten, for now.
19:55imirkin: i thought you were ahead
19:55imirkin: did the mesa update make it that much worse?
19:56Tom^: i thought he had me with like 3 points or so
19:56Tom^: but nah the mesa update didnt affect it, and elemental demo runs flawlessly now :p
19:58Tom^: yea he has me with 4 in total score but my max fps was 0.3 higher. ;)
19:59Tom^: so i hoped my ricing could get him when it was so close, but as ive always assumed. ricing is placebo. :P
20:15Tom^: imirkin: amazing work tho, all the UE demos seems to run fine now
21:05Tom^: imirkin: sadly its introduced somewhere in the -git and not in 11.0.7
21:06imirkin: did i miss something?
21:07Tom^: perhaps it gone lost in the DC i got. :P
21:07Tom^: the longer im ingame in cs:go the more the fps drops from each shot fired i would assume from the muzzle fire.
21:07Tom^: until im at the point of it being 120fps and drops to 5 everytime i shoot
21:07Tom^: almost like a cache is overfilling or similiar, and its not happening on 11.0.7
21:08imirkin: yeah.... i can't think of anything offhand that'd be different there -- all of my resource handling stuff went into 11.0.x
21:08imirkin: but there's probably something i'm forgetting
21:08imirkin: or it could be something else entirely
21:08imirkin: perhaps a bisect? :)
21:08Tom^: sure thing, however im going to work in uh 15 minutes.
21:09Tom^: gonna take a while to bisect since it isnt happening directly it takes a ~4 minutes of playing :P
21:09Tom^: can i check how much vram is used?
21:09imirkin: yeah, sad
21:10imirkin: i don't think we expose that anywhere
21:10Tom^: was just wondering, so i could see if it leaks somewhere :P
21:10imirkin: i wonder if 53cbb117 manages to cause the regression
21:11Tom^: how much has llvm in play in things like this?
21:11imirkin: Tom^: check if 11.1-rc3 is affected
21:11imirkin: absolutely none... nouveau doesn't use llvm
21:11Tom^: because i sort of have to build llvm-svn to get mesa-git runnning and running repo llvm 3.7 for repo mesa
21:11Tom^: ah ok
21:11imirkin: [except in extreme cases where st/mesa falls back to software rendering, like for GL_SELECT... basically the only time]
21:12Tom^: imirkin: oki il try rc3 when i get home otherwise il just start bisecting. :p
21:34gnurou: imirkin: good news! looks like I have permission to release the Maxwell texture headers documentation!
21:34gnurou: imirkin: looking at the rnndb format, any particular advice as to how I should write it? Is direct XML ok?
21:35gnurou: imirkin: also there are actually 5 versions of the header, with the fields changing depending on which version is used, how can I reflect that? Is the <stripe> tag the correct way to do it?
21:36imirkin: gnurou: have a look at https://github.com/envytools/envytools/blob/master/rnndb/graph/g80_texture.xml for the current definition
21:36imirkin: here is an example of how to do versions: <bitfield pos="9" name="CUBE_SEAMLESS" type="boolean" variants="GK104-"/>
21:37imirkin: presumably there aren't really *5* versions and we haven't noticed -- just a couple of fields changing here and there
21:37gnurou: imirkin: right, but these versions depend on the chip - for the Maxwell headers, the version used depends on the value of a particular field in the first word of the header
21:37imirkin: that said i know that G80 is missing a bunch of stuff that G84+ has
21:37imirkin: ohhh i see
21:38imirkin: define an enum (much like chipset is defined), and then do variants on that
21:38imirkin: and yeah, stripe is somehow related... look at other examples :)
21:39imirkin: i don't really know the right way, i just copy stuff and make sure it works when i use 'lookup'
21:39imirkin: gnurou: if you have trouble, do it in *some* way, and i can clean it up later
21:40imirkin: what's important is having the right info... getting the xml all pretty can come later ;)
21:41imirkin: anyways... i'm off to sleep
21:42gnurou: imirkin: thanks - good night!
23:15Dezponia: Tom^: A valiant effort, but non shall compare to my mighty 6GB of VRAM which is totally required at all times to gain better FPS! Non I say!
23:20Dezponia: Tom^: It is interesting how consistant the score is in Heaven though, even when the FPS seemingly changes a fair bit from run-to-run the score almost always comes out the same, even accross both our cards which seems to indicate it really is a very accurate benchmark