00:00imirkin: kisak: fwiw, vdpau + offloading isn't a well-tested path
00:00imirkin: (not like vaapi is any different)
00:00imirkin: also some videos may be more problematic than others. the decoder works fairly well for msot videos in my collection
00:00imirkin: but esp trailers i downloaded tended to have odd issues
00:00imirkin: like some motion buffer is insufficiently large
00:00imirkin: or reference frame mismanagement
00:01imirkin: i could never track it down
00:01imirkin: and eventually gave up
02:31imirkin: skeggsb: not sure if you saw my questions earlier this week ... any idea about the NOUVEAU_DMA_SKIPS thing?
02:32imirkin: i.e. should the submitted buffer also have 32 dwords' worth of 0's at the end
02:32imirkin: (or is it 64)
02:33skeggsb: no, i don't think so. i don't fully know for sure the real reason they're needed, but as i understand it, it's a workaround for a HW race condition with vram push buffers, when jumping back to the start of the push buffer
02:42imirkin: oh, so if a pushbuf is in GART then it's ok?
02:42skeggsb: as i understand it - yes.
02:43imirkin: basically i'm seeing some DMA_PUSHER errors
02:43imirkin: with a NV5
08:45mastermart: what you do is crap, this is not programming, inside the queues begins the real chess strategy, only that carries complexity. The shit you do is like wasting employers money and drinking tequila while doing it and primarly let all the users suffer.
08:51mastermart: in round robin you can pin the queues with anyway you like, by running all instructions in full length of the pipeline with zero operands, this is pinning the instruction operands and opcodes to the queues, next up you need to author logic that always run and then block dependencies, instructions that always run is first diagonal is filled with offset adjusting, instruction selection along with branch target determining&load
08:51mastermart: handler and writeback and, next diagonal belongs to with the very same dependencies readback and while the very same deps are also used to handle all the branches and instructions
08:56mastermart: so the second diagonal as i mentioned offset handling in the first diagonal than the second diagonal is all lsus with proper block selction dependencies
09:11mastermart: allthough let's say the first diagonal includes redirection so it technically read&writeback, and second diagonal then could be rereadback
09:13mastermart: cause the first one just takes adjusts the offsets, and grabs the in cache writeback of real instruction
09:13mastermart: then exports it to the second diagonal
09:17mastermart: even though i consider that type of formation one of the most effiecient ways to do it, there are gazillion of other ways in round-robin. where greedy-then-oldest and emulated greedy then oldest, are respecively totally easy and little less easy but still easy
10:11mastermart: so very special attention is made to writeback hazards, so the memory being written is in queues blocked in contrast to full lenth of the pipeline mode, i intend to run SSA based code, yeah wikipedia says too ssa can be interpreted just fine
10:26mastermart: so the memory location that the real instruction commits it's data to can not be overwritten by another one at the same time, this would lead to situation where some writebacks are blocked right
10:26mastermart: i mean lost sorry
10:31ouinonpeutetre: Hi, my GP107 doesn't work at all with Nouveau, only with nVidia and nomodeset
10:31ouinonpeutetre: It makes a black screen
10:32ouinonpeutetre: I don't know if theres people here
11:01mastermart: All would be screwed if export did not take 4 elements, but it is inherently needed cause of how graphics work
12:32uis: I traced program that causes nvc0_program_translate:640 - shader translation failed: -4
13:01mastermart: i took a little nap , so anyways queue mode always blocks on register, unless you free one element and test, did something land, note the index with another export from the target block, rerun the previously skipped one after that
13:06mastermart: so the result should be out-of-order execution, out-of-order commit, though can be madde with in-order commit as well, out-of-order commit is faster
13:13mastermart: what i personally think is, that libFLAME would generate all, you just need to know how your program would run, so you need to give gemm/chloskey/trigonmetric solvers the correct measures, but i have not looked yet
13:38karolherbst: uis: qapit?
13:39uis: apitrace trace
13:39karolherbst: weird extension though
13:39karolherbst: I mean, the g is kind of confusing
13:40karolherbst: also, that cdn is super slow
13:47karolherbst: or my network is messed up.. dunno.. it seems it will take an hour
13:47karolherbst: uis: what kind of GPU is that on?
13:48uis: 760 Ti
13:49karolherbst: mhhhh... okay.. I think I know what issue that is
13:49karolherbst: probably spilling is failing
13:56karolherbst: imirkin: ohh, nvidia already released a 440 beta: http://us.download.nvidia.com/XFree86/Linux-x86_64/440.26/README/randr14.html
13:56karolherbst: so there is no display offloading with nvidia
14:21uis: Are you downloaded replay?
14:24karolherbst: still needs 30 minutes
14:25uis: What speed?
14:29karolherbst: heh.. now it aborted
14:29karolherbst: it's around 250 kb/s
14:51uis: Any news?
15:27karolherbst: uis: when is that error supposed to happen?
15:27karolherbst: I don't have any kepler hardware in reach right now
15:28karolherbst: ehh.. gotta reboot :/
15:29ijwyn: hello everyone... my video card died on me last week and had a new one installed (GeForce GT 710)... but now the screen keeps freezing at random intervals (1h first time, then 8h and 2h30) with no apparent pattern
15:31ijwyn: it seems nouveau automatically recognized the card, from what I can tell, so I'm not sure what the problem is... I did try adding nouveau.noaccel=1 to kernel parameters, a proposed fix I found online, but it did not change anything
15:31ijwyn: added that to /etc/sysctl.d/99-sysctl.conf BTW
15:33ijwyn: I'm starting to get concerned about possible data loss because the only way to get the PC running again as a pressing the reset button on the front of the computer :(
15:33ijwyn: *is by
15:34uis: Are you shure, that GPU died?
15:35uis: Freezing at random time like a driver problem
16:08imirkin: uis: great. will try to have a look this afternoon.
16:09imirkin: ijwyn: it hangs with noaccel=1 ??
16:09imirkin: that's quite odd
16:10imirkin: is there anything in dmesg when the screen freezes?
16:34ijwyn: no sure imirkin, never was able to make sense of dmesg contents :o
16:46ijwyn: this, maybe? was about 13 seconds before everything froze:
16:46ijwyn: localhost kernel: NMI watchdog: BUG: soft lockup - CPU#0 stuck for 23s! [sshd:15168]
16:55ijwyn: also wondering if I placed that noaccel=1 thing in the right place... that 99-sysctl.conf file was empty, just had some commented text in it
17:13karolherbst: uis: okay.. so on maxwell+ the error doesn't reproduce, so it's indeed most likely the spilling issue we have..
17:13karolherbst: don't know how easy that would be to fix to be honest
17:32imirkin: ijwyn: noaccel=1 goes into the module options, or kernel cmdline as nouveau.noaccel=1
17:34imirkin: karolherbst: oh hm. if it doesn't repro on maxwell, i'll have to plug a kepler back in
17:34imirkin: and i don't have any kepler1's actually
17:34imirkin: (for the register pressure bit)
17:35karolherbst: imirkin: shader-db :p
17:38karolherbst: it doesn't trigger with shader-db either
17:39karolherbst: maybe it's fixed accidentally
17:39imirkin: shader-db isn't perfect
17:39karolherbst: well I do have a few shaders which spill when running against e7
17:40doobz: Thanks, Imirkin. Have filed a report as best I can here. https://bugs.freedesktop.org/show_bug.cgi?id=112070
17:40karolherbst: imirkin: by any chance, do you know against which version of mesa it was ran?
17:41karolherbst: mhh, probably not
17:41karolherbst: nothing in the log either
18:02ijwyn: so any thoughts anyone on how I can figure out what is making my screen freeze at random?
18:15imirkin: doobz: what's a "360 degree video"?
18:32doobz: A video filmed with a 360 degree camera. You can alter the direction of view by clicking on them while they play. Go to YouTube and search "360 video"
18:33doobz: Don't know a better way to describe them.
18:35ijwyn: screen just froze again :( lasted 3 hours this time...
18:36ijwyn: (I mean the PC held 3 hours before it froze again)
18:36ijwyn: I tried starting up with an older kernel, to see if it helps (doubt it)
18:37ijwyn: one thing I noticed though in startup messages was that it said "failed to start load kernel modules" so maybe it's not taking into account my "nouveau.noaccel=1" line???
18:40ijwyn: and what does this line mean? it's the last thing in /var/log/messages before the freeze...
18:40ijwyn: Oct 19 20:33:20 localhost org.a11y.atspi.Registry: Xlib: extension "XEVIE" missing on display ":0.0".
19:35uis: Any news?
19:58imirkin: doobz: neat. and your problems are specifically with youtube on firefox, yes?
19:58imirkin: (iceweasel == firefox)
20:00imirkin: doobz: try flipping layers.acceleration.disabled = true in about:config
20:01doobz: Not entirely b/c there is still the issue of the proprietary driver not installing correctly. If you google the error message I got you will find a bunch of folk on nv4b trying to install steam on various distros.
20:02imirkin: yeah, i have no comment anything to do with proprietary driver
20:02imirkin: i (semi) recently fixed some issues with nouveau starting up the steam client
20:03imirkin: should have been in mesa 19.0
20:03imirkin: what mesa are you using?
20:03doobz: But it should work when installed from the repositories on a fresh system, no?
20:03imirkin: fresh, yes. debian? who knows
20:37doobz: Looks like deb buster has mesa (18.3.6-2)
20:40imirkin: ah yeah, you definitely want something that's not quite as ancient
20:51imirkin: anyways, yeah, i fixed some long-standing bugs that affected the steam client, but could really affect anything
20:53imirkin: the steam client even starts on GeForce FX now, albeit nothing renders since the steam client requires NPOT texture support
20:54doobz: I've not investigated anything upstream yet. I will have to do so.
20:55doobz: Might have to try a different distro too maybe.
20:58imirkin: doobz: yeah, dunno how to operate debian, that's up to you. i'd strongly recommend either disabling GL accel or trying newer mesa
23:14imirkin: alkisg: so, success in the field? or haven't had a chance to deploy yet?
23:16imirkin: if it works well, i'll do a release
23:16imirkin: otherwise let me know how to repro any issues