07:54 mupuf: wow, I do not remember anyone ever hitting the limit of hwsw
07:54 mupuf: q
07:54 mupuf: it must be a change from RSpliet. The hwsq codespace started very limited and got doubled at some point
10:57 karolherbst: skeggsb: !!!!!
10:57 karolherbst: skeggsb: awesome work :D
10:57 karolherbst: he just ported the table offsets to 32bit
10:57 karolherbst: but I suppose there is even more
11:01 mupuf: what the heck
11:01 mupuf: how could we have screwed this up that much?
11:01 karolherbst: uhh
11:01 karolherbst: this is there since a long time
11:01 karolherbst: no idea
11:02 karolherbst: I even create a trello card I think
11:02 mupuf: oh, I missed this then
11:02 karolherbst: mupuf: https://trello.com/c/xk6T1MYa/164-port-16bit-table-offsets-to-32bit
11:02 mupuf:is banging his head on the rounding function
11:02 mupuf: for the fan
11:02 mupuf: this is actually the likeliest reason why I get weird constants
11:03 karolherbst: mupuf: always remember the sense table where the actualy row interpretation depends on an entry in another table :p
11:03 mupuf: and it should fix the frequency
11:03 mupuf: yeah.... I read your patch
11:03 mupuf: kuddos to mwk for forcing us to split printing and parxsing
11:04 karolherbst: will push my pascal patches today
11:05 karolherbst: mupuf: especially this one: https://github.com/karolherbst/envytools/commit/68085465e310ea403fded44a85f69d4c218e3256
11:12 mupuf: karolherbst: yeah, go for it :)
11:44 karolherbst: wanna more pascal vbios
11:47 karolherbst: that new memory table is for sure gddr5x related, maybe even general memory reclocking stuff....
11:47 karolherbst: or maybe hbm2
11:48 Yoshimo: not all pascal cards are using that, right?
11:48 karolherbst: nope
11:48 karolherbst: only the super high end
12:02 karolherbst: this table looks like crap, serously: https://gist.github.com/karolherbst/c327327fe46121bb37aadbb3a9db6904
12:17 karolherbst: mupuf: maybe the unk0 T table helps you?
12:21 karolherbst: okay, no pcie stuff
12:21 karolherbst: *now
13:08 bloblo: Hello
13:11 bloblo: I switch from gts 250 to gt 630 this day for test echo 0f to pstate, but not working.
13:11 karolherbst: bloblo: machine freezes?
13:11 bloblo: i watch the box of gt 630, i was wrong is not 386 core
13:12 karolherbst: bloblo: let me guess, gf108?
13:12 bloblo: not freeze work fine
13:12 bloblo: yes
13:12 bloblo: 108
13:12 bloblo: only problem is is showing just 1024M
13:12 karolherbst: true
13:12 karolherbst: fermi won't work
13:13 bloblo: i check box is 2048M 96 core, gainward geforce gt 630, gainward have 2 model same spect just 1 gb and 2 gb change
13:13 karolherbst: bloblo: did you get a significant perf increase on the gt 250?
13:13 bloblo: 128 bit*
13:13 bloblo: after kernel modification yeah in game is not droping much on "game smoke"
13:14 bloblo: only problem is memory clock not change
13:14 karolherbst: yeah
13:15 karolherbst: bloblo: ask mupuf, maybe he could help
13:15 karolherbst: *can
13:16 bloblo: for memory MB i where i need repport debian or nouveau website ?
13:16 karolherbst: mupuf: it was doublted for g94 actually from 64 to 128
13:16 karolherbst: bloblo: it depends on the gpu
13:16 karolherbst: some cards have more than others
13:16 karolherbst: even if they are the same model
13:18 karolherbst: bloblo: drivers/gpu/drm/nouveau/nvkm/subdev/bus/nv50.c
13:18 karolherbst: bloblo: .hwsq_size = 64,
13:18 karolherbst: make it 128
13:18 karolherbst: for the gt 250
13:19 bloblo: i need to change + "modification lastday" or without
13:19 karolherbst: bloblo: with
13:20 bloblo: ok i check log and modify
13:25 bloblo: compil started
13:44 mupuf: oh dear, 64 bits only
13:45 mupuf: bytes*
13:45 mupuf: that is tough
13:45 mupuf: unless it was 64 instructions
13:45 mupuf: funny thing about hwsq
13:45 mupuf: it was all documented in a patent...
13:45 mupuf: and we reversed it
13:45 mupuf: :D
13:45 mupuf: I only found out about it later on
13:45 bloblo: kernel compiled
13:46 bloblo: @karolherbst you want any thing about gt 630 with this new modification kernel ?
13:46 bloblo: if not i switch cards
13:47 karolherbst: mupuf: for reasons I don'T know, the g92 is more treated like a g86 than a g94
13:50 mupuf: karolherbst: because it came out before the g92
13:50 mupuf: what makes you think this would be OK?
13:50 karolherbst: mupuf: well, I am actually surprised that those things are name g94_
13:50 karolherbst: and I would expect that the g92 is actullay more close to g94
13:51 karolherbst: but maybe sanity was first introduced with kepler
13:51 karolherbst: so...
13:51 mupuf: you know that the release order was g86, g92, g94, right?
13:51 mupuf: yeah, tesla was a big clusterfuck
13:52 mupuf: anyway, what is this thing with vram on bloblo's?
13:52 karolherbst: mupuf: gt 630 is hell of a model
13:52 karolherbst: mupuf: 4 chipsets
13:52 karolherbst: gk107, two different gf108, gk204
13:53 karolherbst: 1GB. 2GB and 4GB vram models
13:53 karolherbst: even DDR3 and GDDR5
13:53 mupuf: yeah, the marketting people had a field day :D
13:54 karolherbst: the gt 640 is even worse though
13:55 mupuf: karolherbst: as for your patch for pcie, we'll see how it goes
13:55 mupuf: but be ready to revert it, because tesla is that fucked up
13:56 mupuf: gt2xx brought some uniformity between gpus
13:56 mupuf: before this, it was insane
13:56 karolherbst: mupuf: it is already enabled for g94
13:56 karolherbst: mupuf: g8x is broken
13:56 bloblo: hi
13:56 karolherbst: and I tested it with bloblo on his g92
13:56 karolherbst: even looked in traces
13:57 mupuf: one datapoint ;) But yeah, quite possible it works
13:57 karolherbst: mupuf: 8 traces
13:57 mupuf:had no success with pcie reclocking in the old days :p
13:57 karolherbst: :p
13:57 karolherbst: yeah, it i broken on g8x gpus
13:57 mupuf: karolherbst: you may have missed out some crucial elements needed for the g92 only ;)
13:57 karolherbst: controller gets upset
13:57 karolherbst: throws the gpu out
13:57 karolherbst: mupuf: nope, I didn't enable it for g92, because I lacked traces
13:57 mupuf: nvidia was having fun with it on my g86 iirc
13:57 mupuf: ack
13:57 karolherbst: sure they had
13:57 karolherbst: but it is broken
13:58 mupuf: why did you lack the traces?
13:58 karolherbst: because I didn't search for .xz files
13:58 karolherbst: just noticed yesterday there are actually g92 traces
13:58 karolherbst: .....
13:59 karolherbst: mupuf: anyway, if you look at the entire pcie code, you see what kind of big mess that entire thing is
13:59 karolherbst: and order is important as well
13:59 karolherbst: sometimes even the reg already indicate 5.0
13:59 karolherbst: but the card is still on 2.5
13:59 karolherbst: so you need to poke the same value again
14:01 mupuf: ahah
14:01 karolherbst: it is funny though, that you can even put the kepler cards to v1 mode and the driver even switches back to v2
14:01 karolherbst: protocol I think
14:02 karolherbst: it is all a big black box anyway
14:02 mupuf: yop
14:02 karolherbst: 16.0 will be fun
14:04 karolherbst: anyway, in theory the g92 implementation should work on g8x, there are the same regs with same meanings and everything
14:04 karolherbst: but yeah... I guess they messed up to implement the specs
14:05 karolherbst: but switching to v2 mode works....
14:05 bloblo: hi
14:05 karolherbst: ohhw ait
14:05 karolherbst: no it doesn't
14:05 karolherbst: bloblo: had fun reclocking the gt250?
14:06 bloblo: without set 0f, is working normal without crash, also with set 03, but with 0f screen immediatly blacked and keyboard not response
14:06 karolherbst: bloblo: would it be possible to grab a dmesg?
14:07 bloblo: http://pastebin.com/VRi0FGje
14:07 karolherbst: bloblo: I meant from the last boot
14:08 bloblo: is the last i take
14:08 bloblo: some line adding with set 03
14:08 bloblo: you want ?
14:09 bloblo: but i have no control system after set 0f
14:09 karolherbst: bloblo: maybe it is something inside your sys logger
14:09 karolherbst: but maybe it simply won't work at all
14:10 karolherbst: I guess somebody would need to spend time to check out how that works on a g92
14:41 karolherbst: hum...
14:48 karolherbst: I try to find a pattern within these numbers: https://gist.github.com/karolherbst/afea1460986eb0df89029b4925054d25
14:48 karolherbst: I think either of them matches one of two masks
14:49 karolherbst: 0x1 and 0x20 are already used for something different
14:50 karolherbst: all values | together is 0xde
14:51 karolherbst: mhh
14:51 karolherbst: maybe I should remove that silly 0 value
14:56 Lekensteyn: karolherbst: there are three groups starting at 00, 40, 80 and c0 followed by steps by two until some value (00..0x1e, 00..5e, ...))
14:57 karolherbst: Lekensteyn: ... well, true if the mask is 0xde
14:58 Lekensteyn: no idea if it helps, this is what I found: sort gistfile1.txt | awk '{n=strtonum("0x"$1); print $1, n, (n-l); l = n}'
14:58 Lekensteyn: first column is the number you presented, second the decimal, third the difference
14:58 Lekensteyn: (totally missing context btw :D)
14:58 karolherbst: Lekensteyn: crappy gnu function :p
14:58 Lekensteyn: s/awk/gawk/ then :)
14:59 karolherbst: mawk is thousend times faster than gawk!
15:02 karolherbst: I don't get nvidia here
15:02 karolherbst: why overwriting fast speeds with lower
15:02 karolherbst: 0x1 is for 2.5 GT/s and overwrites 0x20 being 5.0
15:02 karolherbst: I guess default is 8.0 then
15:02 karolherbst: ....
15:31 random-nick: hello, I use Arch Linux and nouveau with the graphics card NVIDIA GeForce 9600m GT 512 MB. I'm trying to launch the game Minecraft, and it crashes with this error: https://bpaste.net/show/32d5be753bff The game works when I run it with LIBGL_ALWAYS_SOFTWARE=1
15:33 random-nick: also, sometimes the error is just the last line and sometimes it's a java error that the process received SIGSEGV while in the nouveau libdrm shared object
15:56 random-nick: I think that error is caused by the game's texture uploading stage
15:58 karolherbst: might be
15:59 karolherbst: tesla is in a little mess right now
16:15 karolherbst: pmoreau: didn't you want me get a pascal vbios somehow?
16:16 pmoreau: I was supposed to get you one, yes, but somehow forgot about it… I'll try to think to boot with iomem=relaxed tomorrow morning
16:16 karolherbst: :) thanks
16:21 Yoshimo: i admire your effort for cards that we can't properly use yet
16:24 imirkin: karolherbst: highly doubtful tesla is in any different situation than other gpu's...
16:24 imirkin: random-nick: i believe minecraft uses concurrent gl calls, nouveau doesn't support that.
16:28 random-nick: ok
16:29 imirkin: random-nick: i have a branch which half-assedly addresses some of these issues if you want to build your own
16:30 random-nick: so it's basically a race condition?
16:30 mooch2: hey, imirkin, is 750ti gm107 or gm108?
16:32 imirkin: mooch2: GM107
16:32 imirkin: afaik GM108's are only mobile chips
16:32 mooch2: ah, okay thank you
16:32 mooch2: ah alright
16:33 mooch2: well, it looks like i have a gm107 in my system now
16:33 imirkin: you can always find out with 'lspci -nn -d 10de:'
16:33 mooch2: well, i'm on windows currently so
16:33 imirkin: the pci.ids db is pretty good about those codes
16:33 imirkin: ah
16:33 mooch2: currently trying to install blob drivers
16:46 karolherbst: imirkin: there is a desktop gm108 though: GT 945A
16:47 mooch2: dam, i didn't know there was a gt 945
16:47 imirkin: karolherbst: ah, fun. i have a GTX 745, which is a GM107 as well.
16:50 karolherbst: odd
16:50 karolherbst: ohh right, the 745 is a gm107
16:51 imirkin: 04:00.0 VGA compatible controller [0300]: NVIDIA Corporation GM107 [GeForce GTX 745] [10de:1382] (rev a2)
16:55 karolherbst: imirkin: I wrote 945 though
16:55 imirkin: i know
16:55 karolherbst: maybe they indeed stopped to sell low end gpus at all now
16:56 karolherbst: 945 is like the slowest from the 900 series
16:59 Yoshimo: there is always ebay
17:02 karolherbst: what kind of table header is this: 10 03 08
17:03 karolherbst: uhh
17:03 karolherbst: no table header size
18:57 karolherbst: mhh, found a game where my nvidia card isn't faster than intel, interesting
18:59 imirkin: with nouveau?
18:59 karolherbst: yes
18:59 imirkin: =/
19:00 karolherbst: checking pmu counters now
19:00 karolherbst: maybe memory limited
19:00 karolherbst: might explain it
19:00 imirkin: or maybe it ends up doing something we didn't anticipate with persistently-mapped buffers
19:00 karolherbst: maybe
19:00 imirkin: those all go to GART now, but there could be an argument for sticking the non-coherent ones into VRAM
19:01 karolherbst: tell me how to check that and I will figure things out
19:01 imirkin: check what?
19:01 karolherbst: uhhhh
19:01 imirkin: whether it's using persistently-mapped-but-not-coherent buffers? check for glBufferData() calls with one but not the other flag :)
19:01 karolherbst: that game doesn't use my nvidia card
19:01 karolherbst: ....
19:01 imirkin: hehehe
19:01 imirkin: that could explain it
19:01 karolherbst: ohh wait
19:02 karolherbst: I set DRI_PRIME=0 to check out intel perf
19:02 karolherbst: the heck...
19:03 karolherbst: now I set DRI_PRIME=1 and still intel
19:03 karolherbst: I hate those "uber smart" engines
19:04 mlankhorst: is the nvidia card linked? :p
19:04 imirkin: he's using dri3
19:05 karolherbst: the hell does this game do
19:05 karolherbst: now I even added that into the launcher script
19:06 karolherbst: https://gist.github.com/karolherbst/487a563fa1e71c5c3b5cfada195fa819
19:06 karolherbst: ...
19:07 karolherbst: ....
19:08 karolherbst: what is wrong with those guys
19:08 imirkin: check glxinfo just in case
19:08 karolherbst: glxinfo works
19:08 karolherbst: I do "LD_LIBRARY_PATH=lib/32/ DRI_PRIME=1 ./bin/32/kAGE" now
19:08 karolherbst: there is no wrapper whatsoever in between
19:08 imirkin: does LD_LIBRARY_PATH work with relative paths?
19:08 imirkin: i know LD_PRELOAD doesn't
19:08 karolherbst: yes
19:09 karolherbst: even LD_PRELOAD does
19:09 imirkin: but the game could chdir
19:09 karolherbst: it could
19:09 imirkin: i'd recommend full path ;)
19:09 karolherbst: but otherwise the game wouldn't even start
19:10 karolherbst: ohhhhh no.
19:11 karolherbst: ....
19:11 karolherbst: please no
19:11 karolherbst: no, god, no, please....
19:11 karolherbst: libGL: Using DRI2 for screen 0
19:11 imirkin: muahahaha
19:11 karolherbst: wait
19:11 karolherbst: it gets better
19:11 imirkin: mlankhorst was right ;)
19:11 karolherbst: nope
19:12 karolherbst: it is worse
19:12 karolherbst: glxinfo: OpenGL version string: 3.0 Mesa 13.1.0-devel (git-2e2562c)
19:12 karolherbst: game log: GL_VERSION = 3.0 Mesa 11.1.0-devel (git-bd6e516)
19:12 karolherbst: !
19:12 imirkin: oh super.
19:12 karolherbst: how...
19:12 imirkin: they're shipping a random git snapshot of mesa.
19:12 karolherbst: staticly linked
19:12 imirkin: what could possibly go wrong
19:12 imirkin: on the bright side, it's an upstream commit
19:12 imirkin: not a local one ;)
19:12 karolherbst: :D
19:12 karolherbst: yeah
19:12 karolherbst: no patches...
19:13 karolherbst: ey....
19:13 karolherbst: how stupid
19:13 orbea: what game is this?
19:14 karolherbst: book of the unwritten tales
19:14 karolherbst: seriously
19:14 orbea: heh
19:14 imirkin: just a random commit taken shortly after the 11.0-branchpoint
19:15 karolherbst: this is frigging sick
19:15 karolherbst: I bet they even hacked around some gl stuff, cause they do silly things or so
19:15 imirkin: yeah, i mean they could obviously make the version id thing say whatever they want
19:17 mlankhorst: static linked? most distros don't even ship that
19:17 mlankhorst: requires some violence..
19:17 karolherbst: mlankhorst: because it doesn't make frigging sense with mesa
19:17 imirkin: mlankhorst: sounds like statically linked *into the game binary*
19:17 imirkin: which would take some doing
19:17 karolherbst: yes
19:17 mlankhorst: that's what I said ;)
19:18 imirkin: well, even shipping a mostly-statically-linked libGL.so or whatever would be tricky
19:18 imirkin: which would be the sane thing to do
19:18 imirkin: i wonder how it works with anything non-mesa
19:18 karolherbst: I bet they only load mesa directly...
19:18 mlankhorst: is there really no libGL somewhere there?
19:19 imirkin: karolherbst: but if it's linked into the game binary...
19:19 karolherbst: imirkin: no frigging idea
19:19 karolherbst: let me try with nvidia
19:19 imirkin: i guess they could have enabled name mangling
19:19 imirkin: and they flip their loader to the mangled names
19:20 karolherbst: sure
19:20 karolherbst: but bumblebee works
19:20 karolherbst: fine, just fine
19:21 imirkin: karolherbst: try changing the GL_VENDOR :)
19:21 karolherbst: there is a ./lib/32/RenderSystem_GL.so file
19:27 karolherbst: it makes no sense
19:27 karolherbst: my system mesa is even loaded...
19:27 karolherbst: even /usr/lib32/libdrm_nouveau.so.2.0.0
19:29 imirkin: just run strings over their binaries until you find the one that has Mesa in it
19:30 karolherbst: they link against some mesa gl calls
19:31 karolherbst: https://gist.github.com/karolherbst/6a02b830ce66adc75fcb54d4c2b194a6
19:32 karolherbst: ohh I have an idea
19:32 karolherbst: I search for that version string
19:36 karolherbst: okay... what are they doing..
19:38 karolherbst: mhh
19:38 karolherbst: seems like they have a custom libGL loader or something
19:39 airlied: karolherbst: that's glew
19:39 airlied: those strings are from
19:39 karolherbst: I know, but well, my nouveau driver isn't picked up
19:40 karolherbst: no whater I try
19:40 karolherbst: imirkin: that old version was a local binary on my system though
19:40 karolherbst: still
19:40 karolherbst: they do something silly cause no other 32bit application ever picked up that old file
19:41 imirkin: ah. probably a hard-coded search path in there
19:41 imirkin: try setting LIBGL_DRIVERS_PATH
19:41 imirkin: (or is it LIBGL_DRIVER_PATH)
19:41 karolherbst: well, my new intel is picked up, where also nouveau is as well
19:42 karolherbst: https://gist.github.com/karolherbst/bacac63dbb288612945eaf716e92e7d9
19:42 karolherbst: the last part
19:42 karolherbst: let me try LIBGL_DRIVERS_PATH
19:43 karolherbst: no luck
19:43 karolherbst: anyway
19:43 karolherbst: it is weird that dri3 is disabled
19:45 karolherbst: imirkin: maybe they staticly linked against libGL
19:45 imirkin: karolherbst: strace -f -e open answers a lot of questions :)
19:45 imirkin: karolherbst: also ltrace
19:46 karolherbst: https://gist.githubusercontent.com/karolherbst/3c1d11ea4dcf3bacaaaef10b1c17eebe/raw/f0e74584f004452296cfe3c5a783a37bbb440b60/gistfile1.txt
19:59 karolherbst: *sigh* let me test dri2 offloading then
20:03 karolherbst: fun...
20:03 karolherbst: the nouveau ddx doesn't load at all for me
20:06 karolherbst: imirkin: "(EE) Screen 1 deleted because of no matching config section."
20:06 imirkin: do you have AutoAddGPU = 0?
20:07 karolherbst: I've disabled it
20:07 karolherbst: should I set it to 1?
20:11 karolherbst: fun
20:11 karolherbst: how disabling xorg.conf helped
20:13 karolherbst: uhh
20:13 karolherbst: k
20:13 karolherbst: compostior crap now
20:14 Dezponia: Hey karolherbst! Whats the best "performance per $" card right now on nouveau you think? :)
20:15 imirkin: Dezponia: the free one from the recycle bin
20:15 karolherbst: Dezponia: no clue
20:15 karolherbst: Dezponia: 780 ti? .D
20:15 Dezponia: Probably actually, no working reclocking on the GTX760 right now?
20:15 karolherbst: sure
20:15 karolherbst: why not?
20:16 imirkin: all keplers should reclock ok now, esp with drm-next
20:16 Dezponia: Nice imirkin
20:16 imirkin: props go to karol
20:16 karolherbst: imirkin: well, now with enabled dri2 it works....
20:16 imirkin: (and ben, for doing a lot of the initial work)
20:16 karolherbst: yeah, ben baiscally did the memory reclocking things I guess
20:16 karolherbst: especially for gddr5
20:16 Dezponia: Thanks a lot karolherbst and ben then :)
20:17 Dezponia: beep-eep: Seems all Keppler (aka GTX700 cards except 750 series) should work pretty ok then :)
20:17 karolherbst: imirkin: what do you think, maybe they just linked staticly against an X library with disabled dri3?
20:17 karolherbst: Dezponia: 750 works as well
20:18 Dezponia: karolherbst: yeah but not with a linux-libre kernel, right?
20:18 karolherbst: ohh right, no firmware no fun
20:18 karolherbst: ohh wait
20:18 karolherbst: nope
20:18 karolherbst: gm107 is fine
20:18 karolherbst: no firmware required
20:18 Dezponia: karolherbst: Even for 3D acceleration and reclocking?
20:18 karolherbst: but mesa supports maxwell not as good
20:18 karolherbst: Dezponia: yes
20:18 karolherbst: Dezponia: gm20x are the bad ones
20:19 Dezponia: Ah, so probably better to get a GTX760 anyway then for people that want to run say, Xonotic at 1920x1008 at decent speeds? :)
20:19 Dezponia: karolherbst: BTW I haven't checked in a while, whats the status for automatic re-clocking?
20:19 imirkin: Dezponia: better get people AMD chips
20:19 orbea: xonotic doesn't take much to be full speed
20:19 karolherbst: yes
20:19 karolherbst: orbea: well, there is also stability
20:20 Dezponia: imirkin: Sure but they're not good for Linux-libre users I'm afraid :)
20:20 karolherbst: ohhh true :D
20:20 karolherbst: imirkin: they can only use nouveau
20:20 karolherbst: :D
20:20 imirkin: lol
20:20 karolherbst: firmware files !=nope
20:20 imirkin: the whole linux-libre thing is pretty hilarious
20:20 imirkin: meh, nouveau requires firmware to boot as well
20:21 imirkin: it just happens to be stored on a chip on the gpu
20:21 karolherbst: imirkin: well, open source firmware is fine though
20:21 karolherbst: ohh right
20:21 karolherbst: the vbios stuff
20:21 imirkin: yup
20:21 karolherbst: well
20:21 karolherbst: you can't easily replace it, so users can't do much about it
20:21 orbea: other than to not buy it
20:21 imirkin: the distinctions are hilarious, like i said
20:21 karolherbst: but you can easily replace the blob files with open replacements
20:22 imirkin: karolherbst: apparently not - none are available for amd
20:22 Dezponia: Anyway so Xonotic would run nicely with a GTX760. I'll pass that on :) But what was the status of automatic re-clocking? Still requires manual re-clocking?
20:22 imirkin: Dezponia: still manual
20:22 karolherbst: imirkin: well, it makes a little sense, because the ones on the chip are generally less important
20:22 Dezponia: imirkin: Ok, thanks. Thats probably fine :)
20:22 imirkin: Dezponia: if you're using the GPU for display, there's some small-ish chance of a screwup on every reclock
20:22 karolherbst: imirkin: well, somebody could write some :D
20:23 imirkin: this chance doesn't add up quickly with manual reclocks, but with automatic it'd basically be 100% cumulative pretty quickly
20:23 karolherbst: imirkin: well, except if you keep mem to max
20:23 karolherbst: my current branch does this
20:23 karolherbst: allthough by accidence
20:23 imirkin: karolherbst: ah, so still manual switch between pstates, but auto-adjust cstates?
20:24 karolherbst: nono, I just screwed up
20:24 imirkin: that's a good intermediate step
20:24 karolherbst: I had it working once
20:24 imirkin: ok, but it's still something that's probably worth shipping
20:24 karolherbst: and then I did something silly
20:24 Dezponia: beep-eep: So there you have it, get whatever you want from the GTX760 series but expect that you'll have to manually reclock the card to get decent performance. GTX760 will probably run Xonotic just fine :)
20:24 karolherbst: imirkin: yeah, I want to rewrite that stuff again and make it more clean
20:24 karolherbst: imirkin: it depends on the other half o my reclocking patches anyway
20:24 imirkin: karolherbst: since you're not going to figure out the line buffer stuff anytime soon
20:25 karolherbst: well
20:25 karolherbst: I would need hardware for this, wouldn't I?
20:25 imirkin: karolherbst: but the low cstates of high pstates are probably still much lower power
20:25 imirkin: karolherbst: and *lots* of patience
20:25 karolherbst: imirkin: yeah somewhat. Memory draws a lot of power though
20:26 imirkin: right, so it'd be better if we could do both ;)
20:26 imirkin: i'm not arguing that point
20:26 karolherbst: imirkin: for me 07 -> 0f nearly doubles power consumption even with the same clock
20:26 karolherbst: ohh I see
20:26 karolherbst: yeah
20:26 karolherbst: clock gating helps though
20:26 beep-eep: Dezponia: Oh yeah, sorry I was not paying attention, I was setting up something in the meantime and my wm did not alert me of the mentions, but thanks for asking for me!
20:26 imirkin: but as an intermediate point, we could ship something that only does cstates automatically
20:26 karolherbst: mhhh, good idea actually
20:27 karolherbst: I will think of that while rewriting it
20:27 karolherbst: I want to write some usefull implementation already and that means some code needs to be written for the pmu
20:27 karolherbst: and pmu counter have to be checked
20:27 karolherbst: and the memory load counter is pretty much... crazy
21:07 vita_cell: karolherbst, with DRI_PRIME and Nouveau I got fine performance, but Bumblebee, optirun, primusrun, with Nvidia drivers I get very very bad performance, can I do something?
21:08 RSpliet: mupuf: I don't think I extended the hwsq codespace. I just changed the pointer size from 8 to 16 bits to avoid overflow on most chips
21:08 mupuf: ok
21:16 karolherbst: RSpliet: any idea if memory reclocking should work at all on g92?
21:26 RSpliet: karolherbst: I recall looking into it briefly with someone here, and fixed it for one card maybe?
21:27 RSpliet: but even then, GDDR3 only I think, and lkely there's bits missing
21:28 karolherbst: I see
21:29 karolherbst: somebody with a g92 gddr3 card was here and we managed to reclock the engines
21:29 karolherbst: but reclocking memory crashed the system
21:35 karolherbst: RSpliet: bloblo was the one with the card
21:35 bloblo: i am the guy :D
21:37 RSpliet: okay... hmm... I haven't really had any time to sink into this recently I'm afraid
21:38 RSpliet: but if you get me an mmiotrace from the official driver doing a memory reclock, and a trace of nouveau doing it with nouveau.debug=NvMemExec=0 I can see if it'll be a quick-fix or a bigger project
21:40 bloblo: is new for me, i dont know what i suppose do
21:42 RSpliet: bloblo: sorry, I currently don't have a lot of time to talk you through the process of everything
21:42 bloblo: not problem
21:42 RSpliet: I think there is an MMIOTrace wiki page on the nouveau website, otherwise ubuntu had a decent one I think
21:42 RSpliet: (and if you're lucky, karolherbst can help you gather the traces required ;-))
21:43 karolherbst: :D meh, plan backfired
21:44 karolherbst: imirkin: actually, it is much easier to have a dynamic reclocking implementation without memory
21:45 imirkin: karolherbst: yes, i'm sure.
21:47 karolherbst: done .D
21:47 karolherbst: nice
21:47 karolherbst: it works
21:47 karolherbst: I can switch the pstate as usual, but the engine clocks depend on the load
21:49 karolherbst: imirkin: dynamic_reclocking branch, should be in good state already
21:49 karolherbst: even hardened against pmu communication errors
21:50 imirkin: ship it
21:50 karolherbst: it's dirty
21:50 imirkin: :)
21:50 karolherbst: https://github.com/karolherbst/nouveau/commit/7b4b6fe2766f60dddea8c76ed6d9532726a1bad1 ;)
21:50 karolherbst: uhh, that hwmon stuff can be removed actually
22:52 karolherbst: imirkin: uhh by the way, if you find some time, could you take a look at the mad load propagation patches again?
22:53 imirkin: link
22:54 karolherbst: imirkin: github: https://github.com/karolherbst/mesa/commits/postraconstfolding ML: https://lists.freedesktop.org/archives/mesa-dev/2016-November/134419.html
22:55 imirkin: thanks
22:58 imirkin: i just figured out another swr oopsie, so might flip back to nouveau for a bit. we'll see.
22:59 karolherbst: is swr so much fun?
22:59 imirkin: it has simpler bugs :)
23:00 karolherbst: :D
23:00 karolherbst: boring
23:00 karolherbst: I like the difficult ones
23:00 imirkin: heh
23:00 karolherbst: if I want to fix easy bugs, I go to work
23:00 karolherbst: :p
23:01 karolherbst: allthough sometimes I also find hard ones there
23:02 imirkin: mostly it looked like those guys needed help in areas where i was able to provide it
23:02 imirkin: and it wasn't a bottomless pit of sadness like nouveau is these days
23:02 karolherbst: :O
23:03 karolherbst: how cruel
23:03 RSpliet: karolherbst: there's a void where imirkin's heart used to be... just like there's a void in modern NVIDIA GPUs where firmware used to be (uploaded by nouveau)
23:05 karolherbst: :(
23:05 karolherbst: poor guy
23:05 imirkin: perhaps the rot will attract more people interested in fixing it up
23:06 airlied: imirkin: multithreaded pits of sadness are much cooler
23:06 karolherbst: I guess my next system will be a desktop and I would just ask mupu to give me one card to fix all desktop issues as well on kepler :O
23:06 RSpliet: NVIDIA promised a cryptographically signed official heart... but release is stalled a bit :-(
23:06 karolherbst: airlied: ! I totally agree
23:06 imirkin: airlied: yeah, that's basically the thing that did me in...
23:06 karolherbst: I would try to fix it myself, sadly nouveau doesn't crash my GPU ..........
23:07 airlied: just run plasma desktop
23:07 karolherbst: I have a laptop
23:07 karolherbst: I even ran it prime offloaded
23:07 karolherbst: nothing
23:07 RSpliet: or some Java applications... or a 3D accelerated VM offloaded through VirtualBox if you'd like some bugs ;-)
23:08 imirkin: that and also the prevailing winds around wayland, glamor, etc.
23:08 karolherbst: what java application?
23:08 imirkin: karolherbst: minecraft
23:10 RSpliet: idk, to be fair I find multithreaded applications in Java the epitome of stupidigy^w^w^w^w^w^w^w^w^wbraveness
23:11 karolherbst: imirkin: mhhh, another one pls
23:11 imirkin: karolherbst: no idea. that's the one people tend to come in complaining about.
23:11 karolherbst: or wait
23:11 karolherbst: I think I still have it installed
23:11 imirkin: minetest. something like that? i dunno.
23:13 karolherbst: silly account migration, can't log in
23:13 karolherbst: ...
23:14 imirkin: airlied: anyways, i might be convinced to look at the threading stuff again when/if ben adds in the userspace fence wait thing.
23:15 karolherbst: let me try something
23:17 imirkin: airlied: unfortunately all this crashy stuff messes with my workflow bigtime since i test directly on my dev machine, so there's a pretty high cost to failure.
23:17 imirkin: and nouveau recovery is sucky/non-existent
23:18 skeggsb: imirkin: that's not really true, the kernel does a pretty decent job at recovering the gpu
23:18 skeggsb: userspace, however...
23:18 imirkin: skeggsb: dunno. not on my box.
23:18 imirkin: it's basically in an unrecoverable hard hang. if it killed X, i'd be fine with that (albeit slightly unhappy)
23:19 imirkin: if i kill the offending app fast enough sometimes it'll come back
23:19 imirkin: [from a second computer]
23:20 imirkin: skeggsb: perhaps your recent fix to serialize channel preemption will actually help
23:20 imirkin: i've only recently started testing with that
23:20 imirkin: of course since my raid5 array is currently degraded, crashing isn't an option.
23:21 imirkin: but i think you know that parallel piglit on my box has always caused hangs... when it's on a secondary gpu, i can get nouveau to unload/reload the gpu via sysfs, but if it's on primary, the console messes with that
23:22 imirkin: also if you want an insta-crash, try running F1 2015 :)
23:23 imirkin: that one has defeated hakzsam_'s spirit
23:23 imirkin: looks like we mess something up which in turn causes an infinite loop in a shader.
23:23 skeggsb: isn't there a watchdog for that stuff?
23:24 imirkin: theoretically
23:24 imirkin: but maybe it wasn't ready for *every* frag shader invocation to hang
23:24 imirkin: anyways - nouveau never recovers from that one either.
23:25 skeggsb: what does the kernel say during one of those hangs?
23:25 imirkin: SCHED_ERROR i believe. but don't quote me on that.
23:25 skeggsb: that's what i'd expect, CTXSW_TIMEOUT, to be specific\
23:25 karolherbst: okay
23:26 karolherbst: startkde does something differently to what sddm does
23:26 karolherbst: odd
23:26 imirkin: yeah. that's the one.
23:26 imirkin: [at least that sounds familiar]
23:26 imirkin: skeggsb: another reliable way to nuke the box is to run the deqp test that makes a single patch generate 4M primitives. that one hangs nvidia blob too :)
23:27 imirkin: (that's with tess factors maxed out, and each GS emitting the max number of vertices)
23:31 karolherbst: hum
23:32 karolherbst: kwin crashes
23:33 karolherbst: huh
23:33 karolherbst: kwin doens't detect GLX_EXT_texture_from_pixmap
23:34 imirkin: uhhh that's kinda an important one
23:34 karolherbst: I guess
23:35 karolherbst: it is inside glxinfo though
23:35 karolherbst: use count of nouveau is 16 ...
23:35 karolherbst: never had it that high
23:36 karolherbst: https://gist.githubusercontent.com/karolherbst/a8a274ef4a576b68c419cae5b4f67b49/raw/e3cd019a7b09f821a212bcf0349d2da130bec197/gistfile1.txt
23:47 karolherbst: maybe the egl path works
23:48 karolherbst: "kwin_core: Failed to initialize compositing, compositing disabled" meh
23:50 karolherbst: xrender works... meh