02:21 imirkin: cosurgi: from ben: https://github.com/skeggsb/nouveau/commit/dd09ebc623e3b3f2ee1ebd9df53bb0754b1dc79b
02:21 imirkin: for your SIGBUS woes
02:21 imirkin: just a guess.
11:02 RSpliet: imirkin: is that u64 export of primitive id / layer 64-bit aligned?
11:02 imirkin: RSpliet: 0x60 / 0x64. so yes.
11:02 imirkin: otherwise we wouldn't merge them
11:04 imirkin: when i merge all 4 into a single export, then they all fail
11:04 imirkin: but if i merge (primitive id, layer) and (viewport index, point size), then only the former pair fails
12:14 imirkin_: tagr: any chance you could have a look at any errata about the above --^ under the guise of helping tegra, of course :)
12:15 imirkin_: specifically the issue is with AST.64 in (at least) geometry shaders, for writing 0x60 + 0x64 in one go
12:16 imirkin_: [and probably only ever geom shaders, since gl_PtimitiveID is only an output in gs]
12:17 imirkin_: i _really_ don't think this is us emitting code wrong, since the exact same logic works for the (viewport index, point size) pair at 0x68/0x6c
12:18 imirkin_: btw AST.128 over that range (i.e. 0x60...0x6c) also fails
12:44 karolherbst: imirkin_: I guess an AST.32 works?
13:11 imirkin_: karolherbst: yes, of course
13:11 imirkin_: technically called "AST" :)
13:11 karolherbst: right
13:12 imirkin_: karolherbst: if you have any non-pascal GPUs plugged in, mind trying out that piglit?
13:12 karolherbst: I don't
13:12 imirkin_: hah
13:12 karolherbst: still not at home
13:12 imirkin_: oh ok
13:12 karolherbst: next week I have access to all my stuff
13:12 imirkin_: enjoy the holidays then!
13:12 karolherbst: well, technically I work this week though :p
13:13 imirkin_: pfft. technicalities.
13:21 karolherbst:about to destory his grub installation
13:22 imirkin_: just write debug.com into your bootloader
13:23 imirkin_: a real man should be able to boot from that.
13:33 karolherbst: grub is just broken *sigh*
18:31 CarolusFridericu: Eine Musik-Nachrichten-Sitzung wurde angefragt. Bitte klicken Sie das MM-Symbol zum Akzeptieren.
19:08 pounce: hi, on my computer XWayland keeps jumping up to 100% cpu utilization and im getting the error `nouveau 0000:01:00.0: Xwayland[706]: nv50cal_space: -16`
19:10 imirkin_: that means commands are being submitted too quickly, and the GPU can't keep up
19:10 imirkin_: or alternatively, the GPU has hung, and we evenutally run out of buffer space for the commands
19:10 pounce: makes sense
19:10 pounce: so probably a bug with xwayland?
19:10 imirkin_: unfortunately with XWayland, it uses a 3d draw-based backend for X11 accel
19:11 imirkin_: (known as "GLAMOR")
19:11 imirkin_: no - nouveau should be handling such situations more gracefully
19:11 imirkin_: there's nothing a 3d application should be doing differently
19:11 imirkin_: but the reality is that we don't
19:11 imirkin_: and furthermore, the 3d driver is very bad at handling any sort of submission faliure.
19:14 pounce: yeah i think it crashed xwayland
19:14 imirkin_: what GPU are you on?
19:16 pounce: GeForce GT 730
19:17 imirkin_: the GK208 variant?
19:17 imirkin_: or the GK107?
19:17 imirkin_: (or, *gasp*, GF108)
19:17 pounce: GK208B
19:17 imirkin_: ah ok. i have the same board.
19:17 imirkin_: it was weirdly unstable for me, uncharacteristically so
19:17 imirkin_: never quite tracked down why =/
19:18 pounce: oh nice! this is my work computer so it was what was installed when i got here. it does seem to have quite a few problems that im only really solving by using wayland programs
19:19 pounce: im always thinking about doing nouveau work but i don't have that many nvidia GPUs and most of them are buggy as hell
19:19 imirkin_: if you feel like helping me out
19:19 imirkin_: i'm looking for people to run a piglit test and tell me the result
19:20 imirkin_: on various GPUs
19:20 imirkin_: there's a weird errata, and i want to know where it applies
19:20 pounce: if you have this one the only other one i have is a 750TI at home
19:20 imirkin_: well
19:20 imirkin_: i have a lot of ones
19:20 imirkin_: it's not plugged in at the moment
19:20 imirkin_: the one that's plugged in is a GT 1030 (GP108)
19:20 imirkin_: so results from the GK208 would actually be nice
19:21 imirkin_: this is the patch: https://patchwork.freedesktop.org/patch/347445/
19:21 imirkin_: but i have at least one GPU per generation since Riva TNT2 ... one tends to build up a collection when one works on nouveau
19:22 imirkin_: but swapping GPUs, rebooting, etc - esp on my main desktop - not my favorite activity
19:22 pounce: sure lol
19:23 imirkin_: not sure if you've ever built piglit, but if not, it's pretty easy. just make sure you do *not* install it - that way lies madness.
19:23 pounce: can i just apply that patch to the repo i got from anongit.freedesktop.org/git/piglit
19:24 imirkin_: yes
19:24 imirkin_: there should be a "download patch" link
19:31 pounce: should i exit my wayland compositor before running tests?
19:37 imirkin_: no
19:37 imirkin_: wait
19:37 imirkin_: only run one test
19:37 imirkin_: bin/glsl-1.50-transform-feedback-builtins -fbo -auto
19:37 imirkin_: pounce: --^
19:43 imirkin_: hm, i really hope he's not running the full suite
19:43 imirkin_: that way lie crashes =/
19:48 pounce: just compiling everything right now..
19:49 imirkin_: ok cool. but only run that one test :)
19:49 pounce: ye
19:54 pounce: PIGLIT: {"result": "pass" }
19:55 imirkin_: huh
19:55 pounce: sorry, pretty boring result
19:55 imirkin_: interesting.
19:55 imirkin_: can you run it with NV50_PROG_DEBUG=1 in the environment?
19:55 imirkin_: want to double-check something
19:55 imirkin_: (might not cause anything extra to be printed, depends on the mesa build settings)
19:56 imirkin_: and you applied my patch, right? :)
19:57 pounce: mhm! patch is applied. nothing else is printed though
19:57 pounce: i can rebuild mesa
19:57 imirkin_: nah, don't bother
19:57 imirkin_: this is interesting though!
19:58 imirkin_: there's no great reason that it should fail in the first place, it just does on GP108...
19:58 pounce: i was thinking i should run some of these tests to see what's wrong with my gpu.. but you said some of them crash things?
19:58 imirkin_: yeah
19:58 imirkin_: more likely than not, you'll get a gpu hang at some point
19:58 imirkin_: note that you should be able to reclock
19:59 imirkin_: to get higher speeds, hopefully mitigating some of those "draw too fast" issues
19:59 imirkin_: cat /sys/kernel/debug/dri/0/pstate
20:00 pounce: don't have those debug files either; probably need to enable them in the kernel module?
20:00 imirkin_: probably just aren't root
20:00 pounce: oh ya see them now
20:01 imirkin_: you probably see 2 pstates ... 07 and 0f?
20:01 imirkin_: if you echo '0f' into that file, should increase clocks
20:02 imirkin_: (the 'AC' line shows curernt clock settings)
20:02 pounce: 3; 07: core 405 MHz memory 810 MHz 0f: core 653-901 MHz memory 1800 MHz AC: core 405 MHz memory 810 MHz
20:02 pounce: makes sense
20:02 pounce: oh
20:02 imirkin_: looks like you have the DDR3 variant of that board
20:02 imirkin_: not exactly a speed demon...
20:03 pounce: lol
20:03 imirkin_: (i do too... came with a dell of some sort)
20:03 imirkin_: (that's also where my GM107 and GP108 came from... dell's been good to me.)
20:03 imirkin_: coz i just remove those boards, since they do more harm than good
20:03 imirkin_: and call it profit :)
20:05 imirkin_: anyways, thanks a lot for testing!
20:05 imirkin_: appreciate any and all help
20:05 pounce: no problem. feel free to ping me if u want anything else
20:06 imirkin_: as for the stability issues ... there was a recent update to the firmware code that nouveau uses
20:06 imirkin_: (really just regenerating it)
20:06 imirkin_: so perhaps a more recent kernel will help
20:07 pounce: im on 5.4.8. more recent than that?
20:07 imirkin_: probably, hold on, might not even be merged upstream =/
20:08 imirkin_: skeggsb: bleh, what's going on? those commits have been in your tree for ages =/
20:08 imirkin_: pounce: https://github.com/skeggsb/nouveau/commit/79017dd04347a1fb7f38ed7016ecd0ed0d850423
20:08 imirkin_: minor massaging would have to be used to get it to apply to an actual kernel tree
20:08 imirkin_: but it shouldn't be too hard
20:09 imirkin_: (note that appending .patch to the github url will get a proper patch)
20:15 tanriol: imirkin_: the test seems to fail on GM107 if you're interested in this one
20:17 imirkin_: tanriol: yes, i am, thank you!
20:17 imirkin_: very interesting.
20:17 imirkin_: so it works on SM35 (and presumably earlier), but fails on SM50+
20:17 imirkin_: i wonder if there's something even more subtle going on
20:17 imirkin_: like not waiting enough cycles
20:17 imirkin_: or something dumb like that
20:18 imirkin_: although why AST.32 and AST.64 would be different somehow is not immediately apparent
20:18 imirkin_: tanriol: presumably it comlains about gl_Layer and gl_PrimitiveID, but doesn't mention gl_ViewportIndex?
20:19 imirkin_:will mess with delays later on
20:20 tanriol:sent a long message: < https://matrix.org/_matrix/media/r0/download/matrix.org/aTCHFGnarWlHWXQNLtfHHVJR >
20:24 imirkin_: tanriol: right. thanks.
20:24 imirkin_: the random alues are slightly unexpected, but meh. i tended to just get 0's there...
20:25 imirkin_: could be that my memory was mostly 0'd out after runnign so many tests :)
20:31 imirkin_: tagr: if you do end up looking, i guess SM50 would be where it starts
23:23 Lyude: imirkin_: how exactly did you guys go about testing ovlys, and do you have any idea when the last time you tested them was? been playing around with trying to figure out what's up with them on my gk104, it seems like we're not programming the surface location correctly (wherever it's scanning out from is overlapping with the crc notifier contexts we allocate in the GPU's vram as well, the garbage you see in
23:23 Lyude: the hw plane moves in a very consistent pattern when you turn on crc capture) https://lyude.net/~lyudess/tmp/nv_kms_setplane.png
23:23 imirkin_: Lyude: i used my bare hands to test them
23:23 imirkin_: i probably did it in the past 6 months
23:23 imirkin_: when i was doing the LUT stuff
23:24 imirkin_: the 'modetest' tool supports all this
23:24 imirkin_: you just have to feed it a sufficiently large quantity of arguments
23:24 Lyude: imirkin_: i mean I have a way of testing it already
23:24 Lyude: that's from the kms_plane test on igt
23:24 imirkin_: sounds like the format is off?
23:24 imirkin_: try modetest.
23:25 Lyude: yeah that's my other guess, i'll give modetest a shot
23:26 imirkin_: not all primary plane formats work on overlay btw
23:38 Lyude: imirkin_: yeah-but it looks like we're only exposing the right ones, although I might double check to make sure. modetest seems to show the same behavior
23:38 imirkin_: ok
23:38 imirkin_: well i've never personally tested on GK104
23:38 Lyude: I'd expect XR24 to work though :S
23:38 imirkin_: or GK10x
23:38 Lyude: ahh
23:38 imirkin_: i've tested on GK208, i'm pretty sure, though
23:38 imirkin_: which should be a similar display class
23:39 Lyude: imirkin_: pretty sure they use the same ovly methods for the most part, just different format support
23:39 imirkin_: erm, well, i have a GP108 plugged in, i can check on that
23:39 imirkin_: what command are you using, precisely?
23:40 Lyude: oh nope, same overlay channel
23:40 imirkin_: can you try YUYV?
23:40 imirkin_: [for overlay format]
23:40 Lyude: imirkin_: sudo modetest -P 43@47:64x64+100+100 (43 == ovly, 47 == head-0)
23:40 Lyude: and yeah sure
23:41 imirkin_: erm
23:41 imirkin_: does that work?
23:41 imirkin_: you need to like ... set a mode, resolution, primary format
23:41 Lyude: imirkin_: that command doesn't, no. the hardware plane comes up but I'm still seeing garbage in it
23:41 imirkin_: i've never used modetest like that
23:42 imirkin_: i always give it a LOT more arguments
23:43 Lyude: imirkin_: https://lyude.net/~lyudess/tmp/nv_modetest_01.png
23:43 Lyude: imirkin_: I will try more args
23:43 imirkin_: you should have at least like 5lbs of args...
23:44 imirkin_: like one of those big rice bags worth :)
23:48 Lyude: mh :s, doesn't help, honestly i'm almost certain we are setting something up incorrectly in hw. going to try playing around with the color formats
23:48 imirkin_: it's been known to happen.
23:49 imirkin_: but it definitely worked for some people at some point.