04:52 phomes: I have access to a few desktop chips I can test e.g. re-clocking on: Three different NV92's and a FX 3800. The last one is not listed on our wiki for code names. A google tells me that it is a G200GL. Are testing for those interesting?
04:57 karolherbst: phomes: try stuff out and check on which card it doesn't work and then you now the interesting ones ;)
05:01 phomes: karolherbst: Sure, I can do that. I don't have direct access but can borrow them as needed. It is a hassle though so I wanted to see if they known-good or something :)
05:02 pmoreau: phomes: NV92 are not advertised as supporting reclocking, but maybe it works on them. You should do the same as http://cgit.freedesktop.org/nouveau/linux-2.6/commit/?h=linux-4.4&id=0d42743dfa908a2ca4e349f883361906ebb4db95, but replace 0x94 by 0x92 of course. :-)
05:13 phomes: pmoreau: thanks! I am still very new to this and would not have known to do that :)
05:59 karolherbst: imirkin: can you tell me how you run the shader-db stuff on nouveau?
05:59 imirkin: karolherbst: http://cgit.freedesktop.org/mesa/shader-db
05:59 imirkin: grab a copy of that
05:59 imirkin: apply this patch:
05:59 imirkin: http://patchwork.freedesktop.org/patch/64194/
06:00 imirkin: and enjoy
06:00 imirkin: that mad thing, btw -- 0 changes in shader-db. what game was that with?
06:00 imirkin: karolherbst: also i made a more complete version of the patch...
06:01 imirkin: karolherbst: http://hastebin.com/imigasibow.pl
06:01 karolherbst: imirkin: gputest pixmark_piano
06:04 karolherbst: imirkin: mhh
06:04 karolherbst: the new patch doesn't work anymore :/
06:08 karolherbst: imirkin: mhh when I execute run.py it errors with 0 shaders found :/
06:09 karolherbst: "Exception: Only shader 0 found. shaders/orbital_explorer.shader_test"
06:24 imirkin: karolherbst: run, not run.py
06:24 imirkin: look at the intel instructions
06:24 karolherbst: yeah figured that out now, but I have to rebuild my mesa anyway, because I built i915 instead of i965 :/
06:25 imirkin: huh?
06:25 imirkin: why does that matter?
06:25 imirkin: oh, you also need to edit run.c and point it at the right rendernode
06:26 karolherbst: ohhhhhhhhh
06:26 karolherbst: that's why :D
06:26 karolherbst: but even under intel it won't run
06:26 karolherbst: ERROR: eglGetDisplay() failed
06:26 imirkin: you need --with-egl-platforms=drm,x11
06:27 karolherbst: k, thanks
06:27 karolherbst: mhhh
06:27 karolherbst: that didn't work for some reasons
06:27 karolherbst: clean build then
06:28 karolherbst: imirkin: but yeah, your new patch doesn't work, any idea why?
06:29 imirkin: no
06:29 imirkin: could explain why no changes in shaderdb :)
06:29 karolherbst: ohhh why OP_FMA?
06:30 imirkin: you're not supposed to split fma, while you can split up mad
06:30 karolherbst: mhh
06:30 karolherbst: do I have to use both patches?
06:33 imirkin: huh?
06:33 karolherbst: well first line of your patch: " case OP_FMA: {"
06:33 imirkin: right...
06:33 imirkin: and the line above that is OP_MAD
06:34 imirkin: not sure i get your point
06:34 karolherbst: ahhh
06:34 karolherbst: okay
07:23 Tom^: sir imirkin, busy day or did you find some time to test kepler on elemental demo?
07:23 imirkin: it's 10am, i have not yet found time, but i will before the day's out
07:24 Tom^: oh right you live in the past.
07:24 imirkin: yes, in a simpler time :)
07:26 fugdnscerd: having issues starting X windows on an imac G5 ppc using nouveau drivers, was hoping you could help. X begins starting and seems just fine until the display turns black and the computer seems to freeze for quite a while. X will then crash. In the Xorg log file I get NVLeaveVT is called and after some wait a backtrace is started
07:28 fugdnscerd: Thinking I need to enable more verbose logging but I cant find an Xorg.conf file
07:46 fugdnscerd: OK fixed that by removing the boot option video=ofonly
07:54 RSpliet: fugdnscerd: ah, good to know. sorry, didn't have a solution for you myself
07:55 imirkin_: fugdnscerd: i don't have scrollback here and gotta run, but please provide dmesg + xorg log + glxinfo
07:56 imirkin_: i'll have a look in an hour or two
08:38 fugdnscerd: i did not get a chance to save anything due to a seperate issue that was text bombing my console via dmesg. received the computer without a hard drive, i have since replaced but do not have the temperature sensor.
08:38 fugdnscerd: when i get things settled down, I will recreate the situation and post logs for before and after
09:32 imirkin_: fugdnscerd: ok, well i have a G5 with a NV34 which at least sorta works... which gpu do you have?
09:32 imirkin_: i fixed a bunch of stuff in mesa which should have made it to mesa 11.0.x too
09:39 imirkin_: Tom^: hm, i see the same effect on my GK208
09:39 Tom^: im not so special after all then
09:40 Tom^: first game and demo i notice it in tho
09:42 Tom^: imirkin_: do you like work at redhat or something, being able to test nouveau at work. i mean i want your job. :P
09:42 imirkin_: Tom^: i do not work at redhat, nor something :p
09:43 Yoshimo: not YET
09:49 karolherbst: lol
09:49 imirkin_: Tom^: i do think this might be a newer version of the demo though (directory layout is a bit different), so perhaps they changed something
09:49 imirkin_: i'll re-check at home with a fresh download
09:49 Tom^: ok
09:53 mlankhorst: noon
09:56 imirkin_: hm, i get this with a debug build. probably highly relevant: glretrace: codegen/nv50_ir_emit_gk110.cpp:581: void nv50_ir::CodeEmitterGK110::emitIMUL(const nv50_ir::Instruction*): Assertion `i->subOp != 1' failed.
10:00 Tom^: meaning my shaders are broken? :p
10:01 imirkin_: will investigate in a bit
10:06 imirkin_: hmmm looks like i never found the imul high bit for the limm case... let's search around
10:24 imirkin_: glretrace: main/teximage.c:813: init_teximage_fields_ms: Assertion `img->_BaseFormat != -1' failed.
10:24 imirkin_: well that's just *super*
10:25 imirkin_: nooope, that doesn't fix it
10:35 imirkin_: Tom^: minor fix pushed out for gk110, but that doesn't fix the trace.
10:36 Tom^: mk
10:36 imirkin_: probably a silent fail
10:43 fugdnscerd: imirkin_: I have the Nvidia GeForce FX 5200
10:44 fugdnscerd: think I botched some things up with the window manager trying to switch from lightdm to gdm so I decided to do a a fresh install.
10:45 imirkin_: fugdnscerd: ok, that's the same one i have...
10:45 imirkin_: fugdnscerd: i'd stay away from gnome-shell, that's unlikely to work well on top of nouveau/nv34
10:47 fugdnscerd: ahh, well its about 3/4 through install so we will see if it works and if not i might have to reinstall yet another time
10:47 fugdnscerd: have you ever changed the hd in it?
10:48 imirkin_: fugdnscerd: i've only booted mine over nfsroot. the main hdd has a fresh osx install that i've booted once or twice
10:48 imirkin_: i got it for the express purpose of making sure nouveau worked on big-endian.. there were a handful of annoying-to-diagnose issues
10:49 imirkin_: and there remain a bunch more, sadly
10:49 fugdnscerd: i changed my HD and now if i do a cold boot it doesnt see the hd, but if I run the install cd and let install get to the point where it detects hardware the hd appears in the /dev tree and I can do a warm reboot and all is good... wierd
10:49 imirkin_: you mean OF doesn't see the hd? that's odd.
10:50 fugdnscerd: nope, I snagged the HD out of an old stb dvr I had laying around. I heard of them locking them but I wouldnt expect it to unlock without a lot more work if that was the case
10:50 fugdnscerd: but ive never tried before either
10:51 imirkin_: i changed mine to always drop me into OF on boot
10:51 imirkin_: and then just type "boot enet:0" :)
10:52 fugdnscerd: hmm I may have to look into doing something like nfs boot, I'm setting up a small DAW with ardour and jack
10:52 imirkin_: DAW?
10:52 imirkin_: oh, digital audio workstation?
10:52 fugdnscerd: Ive tried once or twice before but the guides I found had some pretty big holes appearantly
10:52 fugdnscerd: digital audio workstation
10:53 imirkin_: it took me a while to figure out how to get OF stuff going properly
10:57 fugdnscerd: yea all this OF stuff is pretty new to me, I'm generally a PC kinda guy
10:57 fugdnscerd: the only other experience I have is with a G3 imac that I put debian on years ago. Wasnt good for much more than web browsing.
11:03 imirkin_: yeah, apparently they like to switch up a lot of the details
11:03 imirkin_: so you find all tehse guides that talk about OF, but it's a diff version fo OF
11:03 imirkin_: naturally i don't actually remember how i got mine working
11:03 imirkin_: just remember that it was a pain :)
11:08 karolherbst: imirkin_: this pixmark_piano shader is two lines total :/
11:08 karolherbst: https://gist.github.com/karolherbst/196067b7df380f0135e9
11:13 imirkin_: karolherbst: ;)
11:15 karolherbst: imirkin_: rerunning all passes: 3.4 seconds => 3.5 seconds
11:16 karolherbst: imirkin_: should I just do a foor loop after the first dead code and check if special things changed? like if the total instruction count has changed, just rerun everything until it doesn't
11:16 karolherbst: and in total 2 or 3 runs or something
11:17 karolherbst: mhhh
11:17 karolherbst: no I don't like that :D
11:32 karolherbst: imirkin_: mhh, this entire benchmark has like 63 calls per frame and all the other shaders are pretty small, still there is a rather big performance gap compared to the blob, could it be soley that the generated code is _that_ bad?
11:43 karolherbst: jkucia: you there?
11:45 glennk: karolherbst, you might want to look at the perf counters when running on the blob, and then compare with nouveau
12:02 karolherbst: glennk: I am sure this is the kind of optimisation where nvidia does unsafe stuff
12:03 karolherbst: this shader contains a lot of mathematical operations and stuff
12:03 karolherbst: and just does some cheap tricks :D
12:03 glennk: "sure" as in measured, or "sure" as in "i am totally guessing" ?
12:03 karolherbst: I am totally guessing
12:03 karolherbst: allthough I have the generated binary by nvidia on my disc
12:04 karolherbst: but this is a beast :/
12:26 karolherbst: how did a binary started in mmt again?
12:36 prg: just tried my gk106 again, this time on 4.4-rc4, started a game (Desperados), intro movie was playing fine, got to the menu, a few seconds later screen froze and https://paste.debian.net/342060/
12:36 prg: seems like it's trying to recover and failing
12:36 prg: would it be useful to report a bug about that? i'm getting lots of different exciting lockups though it seems
12:39 karolherbst: prg: I bet the voltage was just too low
12:39 karolherbst: or
12:40 karolherbst: who knows, maybe the command stream got curropted
12:40 karolherbst: you could try again and figure out if this was just bad luck or if this always happens after a few seconds
12:43 prg: will do
12:49 prg: karolherbst, same thing
12:49 prg: survived a few seconds more, got to click a few buttons this time
12:49 prg: but... same result, same message
12:53 Tom^: karolherbst: is the voltage on 6xx cards the same as on 7xx ?
12:53 Tom^: if so i would say its to low volt if he is clocking the core up :p
12:53 karolherbst: no, doesn't has to be related
12:54 karolherbst: the voltage can also be to low for stock clocks, but this is highly unlikely
12:54 karolherbst: prg: did you reclock?
12:54 prg: no
13:14 prg: karolherbst, so about this voltage... any way to find out if it was indeed too low? (and anything that can be done about it?)
13:16 karolherbst: it could be something else too though :/ if you don't reclock, it shouldnt be
13:16 prg: no, didn't reclock
13:19 karolherbst: prg: is the error always the same?
13:19 karolherbst: like this line
13:19 karolherbst: fifo: read fault at 0000013000 engine 07 [PBDMA0] client 07 [HOST_CPU] reason 02 [PTE] on channel 5 [023f927000 game.exe[23085]]
13:19 karolherbst: except the pid
13:21 prg: karolherbst, second one is fifo: read fault at 0000013000 engine 07 [PBDMA0] client 07 [HOST_CPU] reason 02 [PTE] on channel 5 [023f929000 game.exe[2808]]
13:22 prg: 023f927000 vs 023f929000 if that's important
13:22 karolherbst: imirkin_: mmt and then check what is happening?
13:23 karolherbst: prg: you could do a mmt and see what exactly goes wrong
13:23 karolherbst: prg: I guess the desktop itself doesn't hang or something, just the application?
13:24 imirkin_: karolherbst: fyi i pushed a change which could have a perf impact on kepler+
13:24 prg: karolherbst, X is gone, ssh still works fine
13:25 prg: mouse cursor is frozen
13:25 karolherbst: uhh
13:25 prg: will try the mmt trace
13:26 karolherbst: imirkin_: what exactly will this change?
13:27 imirkin_: karolherbst: no functional change. it'll just stop using indirect sampler fetches for texelFetch() when sampler > 0
13:27 karolherbst: this means less cpu overhead or something?
13:27 imirkin_: gpu
13:27 karolherbst: ohh okay
13:27 imirkin_: more like... more caching opportunity for the gpu
13:39 prg: anything special you need to do when tracing wine? ~/valgrind-mmt/bin/valgrind --tool=mmt --mmt-trace-nouveau-ioctls --log-file=desperados.mmt wine game.exe results in an empty file
13:39 prg: same command works fine with glxgears
14:19 prg: capturing a trace with apitrace works fine, just valgrind seems to be uncooperative
14:20 prg: tried adding --trace-children=yes, but this results in VALGRIND INTERNAL ERROR: Valgrind received a signal 11 (SIGSEGV) - valgrind: the 'impossible' happened: Killed by fatal signal
14:22 imirkin_: karolherbst: can you try the elemental demo on your kepler and see if you also see funny artifacts?
14:22 imirkin_: karolherbst: would be good to know if this is a kepler thing or a gk110 thing
14:31 karolherbst: unigine?
14:31 karolherbst: or unreal?
14:31 karolherbst: ahh unreal it is
14:33 karolherbst: imirkin_: where can I "safely" download it? :/
14:33 imirkin_: https://wiki.unrealengine.com/Linux_Demos
14:35 imirkin_: i think i found another fail
14:35 imirkin_:hopes this is it
14:36 imirkin_: er hm. not likely :(
14:59 prg: karolherbst, any ideas how to get an mmt trace of wine? seems to require --trace-children, but this makes valgrind just segfault
15:00 karolherbst: ohh mhhh
15:01 karolherbst: I bet a newer valgrind is needed for that
15:01 prg: this is git from like two weeks ago, not new enough?
15:02 prg: actually same segfault happens if i try to trace wine-preloader directly, without --trace-children
15:03 prg: https://paste.debian.net/342092/
15:04 karolherbst: wine and valgrind is just painful
15:04 karolherbst: the same way like mono and valgrind are painful
15:04 prg: nothing new to fetch from https://github.com/envytools/valgrind.git
15:04 karolherbst: but valgrind 3.11 should be run on top of mono and maybe even on wine
15:04 karolherbst: who knows
15:04 karolherbst: prg: yeah, but it is based on 3.10
15:05 prg: oh, is there 3.11 with this mmt stuff available somewhere?
15:06 karolherbst: no
15:06 prg: so i'm SOL?
15:19 karolherbst: imirkin_: what rendering issue?
15:19 imirkin_: karolherbst: if it's not obviously there, you don't have it
15:19 karolherbst: k
15:20 karolherbst: all looks fine
15:20 imirkin_: i guess you don't see anything obviously wrong?
15:20 imirkin_: and this is with GK106 + mesa git?
15:20 karolherbst: well, git from yesterday
15:21 karolherbst: current git also looks fine
15:22 imirkin_: wow, only took you one minute to build + test? :)
15:22 karolherbst: system one: built yesterday
15:22 karolherbst: and my git clone is from 1 hour or something
15:24 karolherbst: git-3c6c4d4 and git-65b1a71 are fine :p
15:24 imirkin_: ok awesome, thanks for testing
15:24 imirkin_: so it's some stupid gk110 issue that i just can't spot
15:24 karolherbst: ohh 65b1a71 is something locally here
15:25 karolherbst: 63b850403c90f33c295d3ad6be4ad749d4ea6274 is the newest upstream commit
15:25 imirkin_: well that's just GREAT... i just added an opt that hurts way more shaders than it helps =/
15:25 karolherbst: :D
15:25 karolherbst: why didn't you test?
15:26 imirkin_: i did just test.
15:26 imirkin_: that was the test result.
15:26 karolherbst: :D
15:26 karolherbst: ahh I see
15:26 imirkin_: oh. that's coz i have a bad bug.
15:27 imirkin_: "oops"
15:28 imirkin_: much better.
15:29 imirkin_: total instructions in shared programs : 6180361 -> 6178023 (-0.04%)
15:40 karolherbst: yay
15:56 imirkin_: i think what i need to do for that elemental thing is to get shader dumps on kepler1 and kepler2, and diff the nvdisasm output
15:56 imirkin_: coz it all seems fine :)
16:05 ravior: Hi. I know it's like beating a dead horse, but I've seen something new for bug: https://bugs.freedesktop.org/show_bug.cgi?id=71659
16:06 ravior: I'm getting a message from perf sometimes that the system is under higher load than usual: perf interrupt took too long (2551 > 2495), lowering kernel.perf_event_max_sample_rate to 50100
16:07 imirkin_: expected.
16:15 karolherbst: what is presin by the way?
16:16 imirkin_: sin is a 2-op operation
16:16 imirkin_: presin, then sin
16:16 imirkin_: you can share a presin with a cos though
16:17 karolherbst: okay
16:17 karolherbst: mhhh
16:17 karolherbst: the blob binary uses more regs than nouveau
16:17 imirkin_: but i guess it uses them better? :)
16:18 karolherbst: but 3902 vs 2844 instructions
16:19 karolherbst: 50 vs 56 regs
16:19 karolherbst: yeah, those less instructions might make a difference
16:19 imirkin_: perhaps.
16:20 imirkin_: could be that we unrolled a loop that it didn't
16:20 karolherbst: ohh right :/
16:22 karolherbst: but
16:23 karolherbst: there is a big gap between interp and its use
16:23 karolherbst: over 20 instructions
17:08 mwk: karolherbst: presin is more-or-less convertion to fixed-point, divide by 2*pi (and take the modulus while you're at it)
18:12 imirkin_: well this makes no sense... looks like with *just* dead code elimination and copyprop, the elemental demo looks messed up. but dropping copyprop makes it "work"
18:12 imirkin_: i suspect there's something funky afoot
18:12 imirkin_: like... messed up control flow
18:12 imirkin_: or... who knows. argh
18:13 imirkin_: the corruption is different every time, so perhaps messed up texturing
18:16 imirkin_: mwk: do you have a good explanation for tex p vs tex t?
18:55 imirkin_: GAH! there's a WAW hazard!
18:56 imirkin_: fml
19:11 imirkin_: oh man... this such a corner case of a corner case
19:27 imirkin_: for anyone curious: https://trello.com/c/kjLa338G/132-nve0-waw-hazards-with-texture-loads-and-exit-result-values
19:27 imirkin_: looks like kepler1 just happens to get lucky on elemental, this is an issue that could affect it too
19:41 imirkin_: Tom^: there's a patch in a comment on that card which should fix that demo for you, but it's not exactly a "correct" fix :)
21:14 Tom^: well since i dont have anny issues besides elemental demo i dont think i need that patch until something proper lands. :p
21:18 imirkin: EMIT: sat mul f32 $r64 $r63 0.500000 (8)
21:18 imirkin: 000003a0: c0007f01 03f00003 mul sat f32 $r0 $r63 0x3f000000
21:18 imirkin: is it just me or are those two ops different?
21:24 Tom^: could try the patch later and see if it regresses anything.
21:24 imirkin: shouldn't regress anything except performance :)
21:25 imirkin: grrr... so what do i do in these scenarios? moving the const prop to post-ra seems crazy
21:25 imirkin: but RA can assign unassignable registers
21:26 imirkin: i guess i could try to teach RA not to pick sad registers (and even spill if needed)