02:03 Tom^: karolherbst: yea i went to bed early :p
02:03 karolherbst: :D
02:10 karolherbst: mupuf: maybe that's a bad idea, but maybe the blob decides through the resistance wich shunt to use?
02:11 karolherbst: ohh wait, that doesn't work
02:19 mupuf: :D
02:27 Tom^: hm nouveau isnt in such a bad state afterall, http://i.imgur.com/obEXb0r.png , http://i.imgur.com/be2Ur98.png im ~7 fps from beating all that amd can muster on both catalyst and opensource. :P
02:33 loonycyborg: will nouveau have frequency adjustment enabled by default soon? :P
02:39 Tom^: i think there are a few obstacles left before that, power sensors so it doesnt overshoot power budget. and figuring out when to raise clock and when not to etc
02:44 karolherbst: Tom^: :D
02:45 karolherbst: loonycyborg: well no, but experimental dynamic reclocking yes :p
02:45 karolherbst: and with soon I mean next year
02:45 Tom^: what.. i expected it this year
02:46 karolherbst: well I have some patches for that
02:46 Tom^: you still have 12 hours and 14 minutes on you.
02:46 karolherbst: but the pmu messes up sometimes on my gpu :/
02:46 karolherbst: Tom^: yeah done :p
02:47 karolherbst: loonycyborg: mupuf wants to figure out how to read out the power consumption on every kepler card first, so the other stuff has to wait for now :p
02:48 karolherbst: mupuf: by the way, is there anything I can help with that nvaforcebios thing?
02:48 mupuf: karolherbst: well, give me a few more minutes to settle back in my flat and we can have a look again
02:49 karolherbst: mhhh okay
02:49 mupuf: I need to modify nvagetbios and nvafakebios to re-upload the same data
02:49 mupuf: but when I had a look at it remotely, there was something funky
02:51 karolherbst: Tom^: post your result on the comments :D
02:51 karolherbst: *in
02:51 Tom^: heh nah not until some of it is sort of reproduceable on master :P
02:51 karolherbst: doesn't matter
02:52 Tom^: i dont even have a phoronix account either
02:52 karolherbst: I have to check what settings he uses anyway
02:53 karolherbst: Tom^: which article was it?
02:53 Tom^: i tried doing that, but understanding his phoronix-test-suite aint easy, so i just assumed he ran 1920x1080 at default startup settings.
02:54 karolherbst: mupuf: by the way, there wasn't a thing funky with his 780 ti, he just didn't manage to install nouveau the right way
02:54 karolherbst: well the only funky thing is, that 0a, 0d and 0f have the same core clock
02:55 Tom^: karolherbst: not sure which article, can ask the guy who gave me the url if he knows when he gets back online.
02:56 karolherbst: yeah doesn't matter then
02:56 mupuf: karolherbst: not unheard of
02:56 mupuf: my nvc4 had this IIRC
02:56 karolherbst: yeah, I also saw that sometimes
02:56 karolherbst: this just explained why 0a also messed up
02:58 Tom^: karolherbst: might actually be straight from openbenchmarking.org
03:02 karolherbst: mupuf: I think I am gonna head out now, maybe I find some time to come here later that day, otherwise we could discuss all that stuff tomorrow or something. If not, then maybe I take my time and figure out Fermi stuff or something else :/
03:02 karolherbst: mmiotracing is again broken for me
03:02 karolherbst: no idea what is going on there
05:29 jarnos: imirkin, dmesg for GF4 IPG http://pastebin.com/bVNV0aMM and xorg log: http://pastebin.com/qPWbHm2M
05:33 jarnos: imirkin, odd that video card 0000:02:08.0 was not detected, when the GF4 MX 420 was installed. Here is dmesg for that case: http://pastebin.com/GJUMrTeE
07:12 karolherbst: mupuf: by the way, did you already found out the algorithm how to choose the right voltage or do you only have a guess and want to verify that on your maxwell after faking the vbios works?
09:06 mlankhorst: happy ++year
09:07 pmoreau: happy new year(); // to you!
09:07 imirkin: yes... delete year; year = new year_t()
09:07 mlankhorst: too much class for an integer..
09:08 imirkin: in that case, maybe __sync_fetch_and_add(&year) ? :)
09:10 pmoreau: Is that the OpenCL syntax for `atomicAdd(&year, 6666);`?
09:11 Tom^: while(1){ year = addone(); sleep(yearinseconds); }
09:14 imirkin: aka the julian calendar
09:15 imirkin: actually i guess you could throw 365.25 years worth of seconds in there. but... still miss out on the leap seconds :)
09:15 Tom^: =D
09:15 karolherbst: ...
09:15 karolherbst: okay again
09:15 karolherbst: does anybody has issues tracing the 358.16 nvidia driver?
09:16 karolherbst: https://gist.github.com/karolherbst/f69e2a7b9c372e049525
09:32 joi: pq: ^
09:34 karolherbst: Tom^: ... do you know what?
09:34 Tom^: if only i did.
09:34 karolherbst: Tom^: comparing the R9 290 and 780 Ti. With the semi-official GFLOPS taken into account, nouveau does better ...
09:35 Tom^: i know =D
09:35 karolherbst: imirkin: maybe the nouveau compiler stuff isn't as bad as we think
09:35 Tom^: but still im not 100% sure i used the correct quality settings.
09:35 karolherbst: and the radeon binary compiler is just shit too or the nvidia one just exceptionally good
09:36 Tom^: since parsing phoronix test suite settings wasnt the easiest.
09:36 karolherbst: Tom^: mhh could be
09:36 karolherbst: maybe you should have benchmarked ultra?
09:37 Tom^: well the default settings valley chooses is high
09:37 Tom^: and i cant find anywhere that he changes it besides setting fullscreen and 1920x1080
09:37 Tom^: so idk
09:37 Tom^: i tried getting the test suite running but simply didnt with various errors.
09:38 karolherbst: well I got it work for me
09:38 karolherbst: so I could help
09:38 Tom^: benchmark it in the test suite, run it manually. see which settings is correct xD
09:39 Tom^: nvidia does a fair bit of application profiling too btw if you guys didnt know
09:40 Tom^: earlier when cs:go was freshly released you could gain like ~50fps by simply renaming it to the same as the hl2 binary.
09:40 Tom^: :P
09:40 Tom^: and that probably extends to various benchmarks to give nicer numbers.
09:41 imirkin: karolherbst: the nouveau compiler is actually pretty good
09:41 imirkin: the nvidia compiler is just better.
09:41 karolherbst: :D
09:42 karolherbst: yeah but as it seems it may be even better than the catalyst one
09:42 karolherbst: or Tom^ just benchmarked different settings
09:42 karolherbst: if nouveau is faster than radeon as it is, then I am already happy enough :D
09:43 karolherbst: Tom^: seems like pts uses default settings except setting to fullscreen
09:43 karolherbst: same for heaven
09:43 Tom^: well yea then i did use the right settings
09:43 karolherbst: yeah
09:44 karolherbst: I will still run the benchmark once
09:44 karolherbst: but the tests are all declared inside xml files
09:50 karolherbst: Tom^: well the benchmark doesn't show the quality settings :/
09:51 karolherbst: and I get only 20 fps :O
09:51 karolherbst: ohh wait, now more
09:51 imirkin: karolherbst: i've observed a very odd thing with the unigine benchmarks and prime
09:52 imirkin: if i move the mouse cursor away from the window, the fps counter in the window shoots way up
09:52 imirkin: but the actual demo has the feeling of slowing down
09:52 imirkin: giving the window focus again makes the fps counter slow down
09:52 imirkin: but the actual motion inside of the demo speed up
09:52 karolherbst: mhh
09:52 karolherbst: nope, not for me
09:52 imirkin: pretty sure this is with prime only, not with "regular" :)
09:52 Tom^: karolherbst: run it again yourself see if you get same settings. :D
09:52 Tom^: get same score i mean
09:53 imirkin: (move mouse cursor away == give focus to something else for me)
09:53 karolherbst: imirkin: no, I don't see that kind of thing
09:55 karolherbst: Tom^: yeah sure
09:56 karolherbst: I would be happy to get just half the score you get :D
09:56 Tom^: OC it
09:56 karolherbst: I could
09:56 karolherbst: that*s how I plan to get 50% :D
09:57 karolherbst: my gpu "only" has 1530 GFLOPS :/
09:57 karolherbst: and yours like 5046
09:57 Tom^: =D
09:57 Tom^: it did cost a fortune tho
09:57 karolherbst: well my gpu can run about 280 MHz above the base clock though
09:58 karolherbst: so I should reach 2k GFLOPs without issues
09:58 karolherbst: Tom^: yeah well, my laptop too
09:59 karolherbst: Tom^: you know what, nouveau with your gpu is kind of weak compared to mine though :D
10:00 Tom^: how so
10:00 karolherbst: I think i will get near 25fps
10:00 karolherbst: and if I get 25 you should get like 80
10:00 karolherbst: :D
10:00 Tom^: haha yea well i blame imirkin for not having multithreaded mesa yet
10:00 karolherbst: :D
10:00 Tom^: my cpu isnt sending things fast enough. !
10:00 karolherbst: ohh I am sure it does
10:01 karolherbst: valey cpu usage: <30%
10:01 karolherbst: ohhh I got only 21fps :O
10:01 karolherbst: so 69fps for you seems fine
10:01 karolherbst: well
10:02 karolherbst: I really thought I would get more
10:02 karolherbst: :D
10:05 karolherbst: imirkin: so no idea what you mean. Maybe your window manager sets the window as inactive and unigine drops performance when not in benchmark mode?
10:05 karolherbst: to safe power you know
10:05 karolherbst: *save
10:05 karolherbst: there are games that do that too
10:06 karolherbst: *does
10:06 imirkin: karolherbst: conceivable.
10:43 karolherbst: imirkin: did you check if benchmarking makes a difference?
10:43 imirkin: never
11:39 srg: I have a GTX 960. I'm on an Arch/Win dual-boot (Win for gaming). I don't do much in terms of graphics in Arch. I'm using the proprietary drivers now and they have some crashing issues.
11:39 srg: Does Nouveau work for a 960? I didn't see it in the FeatureMatrix
11:40 imirkin: srg: you should get modesetting, which should work fine. you will get no acceleration at all.
11:41 srg: Firefox is probably the most graphics-intensive thing I use on Arch, so it should be fine, heh
11:41 srg: I do use dual-monitors though
11:43 imirkin: well
11:43 imirkin: you might use gnome
11:43 imirkin: or kde
11:43 imirkin: which are plenty graphics-intensive.
11:43 srg: Xfce
11:43 imirkin: anyways, no real harm trying it...
11:43 srg: Yep, I'll try it
11:43 srg: Thanks!
11:44 imirkin: assuming you don't have a p60 in that thing, your cpu should handle the software 3d accel fine
11:44 srg: ok, cool
11:44 imirkin: which should be enough for the simple usage many applications need
12:30 karolherbst: funny card: https://gist.github.com/karolherbst/483953ac6196eba5666b
12:30 karolherbst: mupuf: there is not even a FAN table for that one :O
12:30 karolherbst: there is like nothing
12:31 imirkin: i.e. no boost :)
12:32 karolherbst: mupuf: the vbios has one thing though: the voltage table has the entry values set
12:32 karolherbst: and it is different than the header values
12:32 karolherbst: 825000 µV/837500 µV
12:32 karolherbst: and so on
12:35 Tom^: karolherbst: which card is that?
12:41 karolherbst: 650 ti
14:55 mwk: well well well
14:55 mwk: happy nouveau year!
14:57 pmoreau: :-)
14:58 pmoreau: Meh, we have to take new resolutions…
15:01 Wonka: Happy Date Rollover!
15:09 imirkin: mwk: good one :)
16:00 pmoreau: imirkin: Found two additional wrongly sized getScratch(): in `loadImm(double)` and `DataArray::acquire()` (not completely sure about that one, but I think it should set to eltSize and file).
16:01 imirkin: pmoreau: oops... send patches :)
16:01 pmoreau: Does it make sense to run piglit to check for potential breakage in that case? Or some other tests I should run before submitting the patch?
16:02 imirkin: pmoreau: mmmm.... my personal guess is that there's no actual problem
16:02 imirkin: just a very hypothetical one
16:02 pmoreau: K
16:02 imirkin: but hey, i could be wrong
16:02 imirkin: doesn't mean we shouldn't fix it
16:02 pmoreau: Well, I did hit one for mkLoadv ;-)
16:02 imirkin: but if you don't believe me, i challenge you to writing glsl that triggers the issue
16:03 imirkin: however if you have other code which in fact triggers the issue, by all means, make the fix
16:03 pmoreau: But not for those
16:03 imirkin: i'm not in favor of having broken code... it's just that hypothetical breakage doesn't always translate into actual breakage
16:04 imirkin: but 64-bit immediate support is pretty new... added by hansg
16:04 imirkin: now, i reviewed it and it seemed to make sense
16:04 imirkin: but i could easily have missed something
16:04 imirkin: and he's not exactly an expert in the codebase either
16:04 imirkin: doubles don't get a ton of usage, double immediates even less
16:05 pmoreau: I'm more looking forward to half-float :-)
16:05 pmoreau: But I'll send a patch
16:06 imirkin: pmoreau: yeah, so i need to eventually resuscitate my patches for making pack/unpack of half floats not horrible
16:06 pmoreau: I'll look back at the weird %r0 with no associated instruction as well
16:06 imirkin: i got side tracked like 20 times on diff things, and nothing ever uses it in practice
16:07 imirkin: and we're pretty good about optimizing everything else
16:07 pmoreau: I was thinking of Pascal which will handle op on packed 16-bit floats
16:07 imirkin: we'll see how pascal works
16:08 imirkin: there's practically no SIMD support in nv50 ir
16:08 imirkin: so... that will be a process :)
16:08 pmoreau: But that will be for my research work, where we use the blob
16:09 imirkin: there are already video instructions that we could use for various simd things on bytes/words
16:09 imirkin: but... in the end it'd be pretty rare for those to get used
16:09 pmoreau: I'll try to pick from time to time a task on trello, but spir-v will remain my main focus
16:09 imirkin: sure
16:10 imirkin: and feel free to send patches which fix hypothetical issues in codegen
16:10 imirkin: even if they're not actively affecting anything
16:10 pmoreau: Sure
16:10 pmoreau: I still have the MAD/MUL split op 64
16:11 imirkin: i'd be fine with upstreaming that before an active user
16:11 pmoreau: But since I haven't tested it apart from looking at the output of nouveau_compiler, I never submitted it
16:11 imirkin: the int64 stuff will need it anyways
16:11 imirkin: which mesa will get to at *some* point
16:11 imirkin: although probably not before 6 months from now, i'd guess
16:14 pmoreau: Which OGL version requires int64?
16:14 imirkin: none
16:14 imirkin: but there's a GL_ARB_gpu_shader_int64
16:14 imirkin: or something like that
16:14 pmoreau: Is it used by any games/applications?
16:14 imirkin: of course not
16:14 pmoreau: :-D
16:16 pmoreau: I haven't looked at the atomics yet, but I was wondering how much I could reuse from your work on it.
16:17 airlied: I wonder when gl4.6 will turn up
16:17 pmoreau: gl4.6 or will they jump to 5.0?
16:17 imirkin: pmoreau: almost everything
16:18 pmoreau: Nice! :-)
16:18 airlied: I suppose some of the features won't happen on dx11 hw, probably GL5.0 might happen
16:18 imirkin: pmoreau: assuming that SPIR-V looks something like the GL stuff
16:18 imirkin: airlied: well, all the bindless stuff is no-go on fermi
16:19 imirkin: so if they bring that into core, it'll be at least kepler+
16:20 pmoreau: imirkin: I have no idea how the GL stuff looks like, but it seems similar to CUDA https://www.khronos.org/registry/spir-v/specs/1.0/SPIRV.html#_a_id_atomic_a_atomic_instructions
16:20 pmoreau: Wasn't there some bindless texture in 4.5 or 4.4 already?
16:20 imirkin: pmoreau: yeah this is basically OP_ATOM with the various subops
16:20 imirkin: pmoreau: nope
16:20 imirkin: pmoreau: just an ext... ARB_bindless_texture
16:21 pmoreau: :-/
16:22 pmoreau: imirkin: BTW, do you need a piglit run on MCP79 or G96? I just updated Mesa and piglit.