03:03kugel: hi. when I use the HWcursor=no option for X/nouveau then the cursor flickers badly
03:04kugel: i need the sw cursor to get my PRIME setup working
03:06karolherbst: kugel: why is that?
03:08kugel: I'm using reverse prime to render to the integrated display of my laptop, this requires sw cursor afaik
03:10kugel: I've reported this a few weeks ago, but only know I noticd that sw cursor even flickers without prime at all (with just the xorg.conf change)
03:12kugel: oh and to be clear, the fickering is only in a full screen game (dota 2) and not on the desktop or when the game is windowed
03:15karolherbst: I think that you don't need to enable sw cursors at all in anyway
03:15karolherbst: and that hw cursors are supported
03:16karolherbst: airlied: you know when your patches landed?
03:16kugel: someone said that in reverse prime sw cursor is automatically enabled
03:16karolherbst: kugel: http://lists.x.org/archives/xorg-devel/2015-June/046636.html
03:16karolherbst: mh better this one_: http://lists.x.org/archives/xorg-devel/2015-June/046634.html
03:16kugel: and in fact, I get the same flickering in my reverse prime setup (without changing xorg.conf) as now without reverse prime but with sw cursor
03:17karolherbst: I don't know if they were included and when
03:17karolherbst: kugel: which xserver version have you installed?
03:17kugel: X.Org X Server 1.18.0
03:18kugel: and I'm still on your stable reclock branch for nouveau.ko
03:19karolherbst: the nouveau version used shouldn't mach matter though :/ sadly I am currently not able to setup reverse prime locally, so I can't check that
03:20kugel: shouldn't sw cursor work regardless of whether it's required by reverse prime or not?
03:20karolherbst: I would say yes
03:26kugel: karolherbst: apparently 1.18.0 lacks that patch set. I'll try to compile myself
03:28karolherbst: kugel: it seems like it isn't even in master
03:28karolherbst: maybe there were some problems with the patches?
03:29kugel: we'll see :-)
03:30kugel: is there an easy way to download that series? :/
03:35karolherbst: no clue
03:37kugel: what does "State: Accepted" mean in patchwork , i.e. here https://patchwork.freedesktop.org/patch/52863/
03:41karolherbst: I don't see thoise changes in the master branch
03:42kugel: i looked at the 1.18 tarball and it's not there either
03:44kugel:tries to fetch the patches from airlied's git
03:56kugel: doesnt apply cleanly, of course :\
03:59karolherbst: yeah, wait for airlied then, maybe he reads our stuff and knows something about it
04:06kugel: i rebased and it turns out only the one patch you linked to is not in master yet
04:15karolherbst: ohh k
04:26kugel: meh, no improvement in reverse prime :(
04:27kugel: so yea, I need a working sw cursor i guess
04:34kugel: I created bug #93732
07:02karolherbst: kugel: do you have the talos principle or wasteland 2 by the way?
07:25kugel: karolherbst: no
07:27kugel: karolherbst: i have lots of other games to test on, but I play dota 2 most of the time
07:27karolherbst: I just somebody to verify that the pcie stuff works for those games also on desktop systems
07:27kugel: (most of these games dont run natively on linux though)
07:28karolherbst: I like a lot though
07:28kugel: i've got hl2, l4d2 and tropico 5 that run natively if it helps
07:28karolherbst: no idea
07:29karolherbst: didn't test it there
07:29kugel: also metro and metro 2033
07:29karolherbst: the non redux versions?
07:30kugel: ah. no
07:30orbea: kugel: you try metro 2030 on nouveau?
07:30karolherbst: so the redux ones?
07:30orbea: does it work
07:30karolherbst: they don't start for me
07:30karolherbst: they oom the kernel
07:30kugel: orbea: i can test later, not installed atm
07:30orbea: Last time I tried it turned my screen all red an black and then sends the system down
07:31orbea: but that was with the 4.3.3 kernel, haven't tried with 4.4.0 yet
07:32orbea: I should try to apitraceit again, but I'm coming to the conclusion that my system is a lot less stable any time I run steam...
07:32karolherbst: orbea: I have one
07:32karolherbst: you can have it, but you need like 5GB of ram to run it
07:32orbea: have what?
07:33karolherbst: the trace
07:33orbea: I have 8 gb ram
07:33karolherbst: the trace is just 1.8GB big :/
07:36karlmag: orbea: heh.. I was about to ask "on you machine og graphics card?" Weird to think that's an actual question these days :-P
07:36karlmag: (mind, I just came by and saw the ram amount comment, not the whole conversation)
07:38karolherbst: karlmag: oh that RAM comment was just, because metro redux allocates a lot of ram in the kernel
07:38karolherbst: once I had it allocating like 20GB+ ram
07:38karolherbst: all inside the kernel
07:38karlmag: wow.. that is quite a bit..
07:39karolherbst: I think it has something todo with buffer mapping and that only like 0.1% of the big buffer is actually used or something like that
07:39karlmag: first PC I had.. 4MB ram..
07:40karlmag:looks at his nearly-ready i7 w/64GB Stuff has happend.. :-P
07:41karlmag: one can wonder if all of it is for the better.. But alas..
07:42orbea: 20+ gb ram, damn
07:45karlmag: Well, that considered.. I *have* seen machines with 512GB (yeah gigabytes) of RAM go OOM..
07:45orbea: Idk what I'd even do with half that ram :D
07:46karlmag: file server on ram disk? ;-)
07:46orbea: hahah :D
07:48karolherbst: orbea: I have only 16GB physically ram though
07:48karlmag: if you're not serving that many files, it's possible; run a 2 or 3 way disk mirror with 1 or 2 being ssd/spinning disks and the last being ram..
07:48karlmag: and tune it to do the reads from ramdisk
07:49karlmag: karolherbst: swap can be rather unfun over time..
07:49karlmag: but sometimes needed
07:49karolherbst: well I have zram swap, 16GB
07:49karolherbst: I was thinking about also adding a disc swap with less priority, but mhh
07:49karlmag: My current primary workstation is running into swap quite a bit, but I "only" have 8GB ram on it.
07:50karolherbst: karlmag: check out zram then
07:50karolherbst: and set the zram size to something like 16GB or if you are risky, then 24GB
07:51karlmag: nah... it'll be demoted to test machine in not that long anyway, so I can't be bothered.
07:51karolherbst: just a thought, because then you won't need your disc swap anymore :D
07:52karlmag: well, it will create some overhead too
07:52karlmag: maybe not enough to be bothersome though
07:53karolherbst: the overhead will only be created when you run out of RAM ;)
07:53karolherbst: and I don't think it is worse than swapping to disc
07:53karolherbst: but yes, it draw a bit out from your normal ram and in the end more is swapped
07:54karolherbst: but I heard that some used it on like 2GB Ram machines to boost their effective RAM to over 5GB
07:56karlmag: that is quite impressive
08:31kugel: orbea: yea, i can confirm the metro behavior
08:32kugel: was loading something, and after a while the ram usage jumped up immediately and the game was oom-killed
08:34orbea: it works all the way into trying to start a new game for me, then just red and black, even after closing the game if I manage to kill it before a hard freeze
08:35orbea: like for example, by green on black weechat becomes red on black...
08:41kugel: orbea: this is what I jus had on metro last night
08:49orbea: my logs weren't very useful unfortunately, an apitrace would be better
08:50orbea: i'd try again, but Im enjoying my system not crashing right now
09:03karolherbst: here is the trace by the way: http://en.file-upload.net/download-11216915/metro.trace.xz.html
09:03karolherbst: you can replay it and have a look at your memory consumption
10:16n7025: can someone help me with mmio-tracing?
10:17n7025: the option inside the kernel is enabled. i cant find /sys/kernel/debug/tracing directory
10:17n7025: as i understood i have to enable it with "echo 64000 > /sys/kernel/debug/tracing/buffer_size_kb"
10:21pmoreau: That echo is not going to enable it, "echo mmiotrace > /sys/kernel/debug/tracing/current_tracer" will
10:21pmoreau: But if you can't see the directory, that's not going to work anyway.
10:22pmoreau: Are you trying as root?
10:22pmoreau: n7025: -^
10:22n7025: pmoreau: yes, as root. i didnt have a /sys/kernel/debug/tracing directory. mkdir /sys/kernel/debug/tracing also reports a error (run as root)
10:23pmoreau: Creating the foldre manually won't help :-)
10:23pmoreau: Is debugfs mounted?
10:23n7025: pmoreau: how can i mount debugfs and where should i mount it?
10:24pmoreau: I have "debugfs on /sys/kernel/debug type debugfs (rw,relatime)" on my laptop, so I guess you should run `mount -t debugfs /sys/kernel/debug` or maybe just `mount -t debugfs`
10:27n7025: yes, its mounted on /sys/kernel/debug now. still no tracing directory
10:29karolherbst: n7025: is there anything inside /sys/kernel/debug ?
10:29n7025: karolherbst: yes
10:29n7025: about 20 things
10:29karolherbst: weird, there should be a tracing directoy.
10:30karolherbst: but 20 isn't much though
10:30karolherbst: did you by any chance build your kernel yourself?
10:30n7025: 22 to be exact
10:32karolherbst: n7025: what does "gzip -d -c /proc/config.gz | grep MMIOTRACE_SUPPORT" give you?
10:34n7025: karolherbst: config_have_mmiotrace_suppport = y
10:34n7025: thats why i told first that the option inside the kernel is already enabled
10:35karolherbst: yeah, was just checking for sure
10:36karolherbst: pmoreau, n7025: "tracefs on /sys/kernel/debug/tracing type tracefs (rw,relatime)"
10:45n7025: karolherbst: what did you mean? what should i do now?
10:47n7025: i tried "mount -t tracefs". still no tracing directory
10:51imirkin: n7025: mount -t tracefs -o rw,nosuid,nodev,noexec,relatime none /sys/kernel/debug/tracing
10:56n7025: imirkin: mount point /sys/kernel/debug/tracing does not exist
10:56imirkin: ah yeah... there should be an empty dir there
10:56imirkin: if there isn't you probably forgot to enable CONFIG_TRACING o rosmething
10:56n7025: imirkin: mkdir ... not allowed
10:56imirkin: yeah you can't mkdir
10:57imirkin: n7025: grep tracefs /proc/filesystems
10:57n7025: config_tracing_support = y
10:58n7025: grep tracefs /proc/filesystems -> nothing
10:58imirkin: ok, well that's a problem.
10:59imirkin: zgrep TRACE /proc/config.gz ?
11:00n7025: imirkin: what are you searching for? i have to type out everything by hand
11:00imirkin: hm actually i don't see what option enables it
11:00imirkin: hold on
11:02n7025: imirkin: there is no "CONFIG_TRACING". There is <n7025> karolherbst: config_have_mmiotrace_suppport = y
11:02imirkin: you need CONFIG_TRACING=y
11:02n7025: and <n7025> config_tracing_support = y
11:02imirkin: those are good too
11:02imirkin: you need 'em all
11:02n7025: both are enabled as you see
11:03imirkin: that's great. you also need CONFIG_TRACING=y.
11:03joi: CONFIG_MMIOTRACE != CONFIG_HAVE_MMIOTRACE_SUPPORT
11:03imirkin: yeah, the HAVE_* are arch-level things and don't enable functionality
11:03imirkin: you also will want CONFIG_MMIOTRACE=y
11:03n7025: imirkin: shouldnt i get inside the grep a "config_tracing is not set" ?
11:04imirkin: n7025: maybe? i dunno what's included in the /proc version of the config file
11:08joi: n7025: CONFIG_TRACING_SUPPORT is not enough, you need CONFIG_TRACING (and CONFIG_MMIOTRACE)
11:09n7025: joi: why cant i see them in a grep listed as "config_tracing is not set" ?
11:11imirkin: does it matter?
11:11imirkin: if they're not showing up, they're not set :)
11:12n7025: how did i know that i can enable them?
11:12imirkin: we just told you.
11:13n7025: :D ok
11:13imirkin: if e.g. you don't have a whole menu, i think none of the suboptions end up showing up as "not set"
11:13imirkin: otherwise your thing would be 99% "not set" entries
11:18pmoreau: joi: How do you find out in a demmt trace which channels are the EVO ones? IIRC, you told me how to some time ago, but I forgot…
11:22karolherbst: meh, the hardest thing about dynamic reclocking is, to choose an algorithm so that the driver won't reclock every 2 seconds :/
11:23joi: pmoreau: IIRC demmt does not support EVO channels...
11:25pmoreau: Don't they appear in the trace as `w: some_address data`?
11:27karolherbst: is there a way to stop a timer on the pmu?
11:27karolherbst: or reset it in any way?
11:29joi: pmoreau: http://people.freedesktop.org/~cbrill/dri-log/?channel=nouveau&highlight_names=pmoreau%3Bjoi&date=2015-02-10
11:35pmoreau: joi: Ah awesome! Exactly the conversation I was looking for, but couldn't find when I searched it…
11:47n7025: is there any distribution with config_tracing and config_mmiotrace enabled by default available?
11:58pmoreau: Arch has MMIOtracing enabled by default
12:02imirkin: mwk: are you aware of any ops on fermi that would query image sizes? (of bound images) something like txq but ... imgq
12:24mwk: imirkin: nope
12:25imirkin: mwk: ok, so just have to be passed via constbuf params?
12:25mwk: I guess so
12:25imirkin: oh, and do you know what all the various su* ops do?
12:26imirkin: and if so, could you provide like 1-line summaries? coz i'm totally lost.
12:26mwk: they're a horrible mess
12:26mwk: that's about as much as I remember
12:27imirkin: i was able to make that out all by myself :)
12:27mwk: I recall something about loads being addressed by byte coordinates and stores by pixel coordinates
12:27mwk: or the other way around
12:27imirkin: and created with no regard to pronounciation... sueau? come on.
12:27imirkin: well there's sustP and sustB
12:28imirkin: i think p = pixel, b = byte
12:28imirkin: kepler's all different and might not have the *p variants
12:28mwk: right, so... unholy mess
12:29imirkin: i won't even mention the mess that opencl creates by not having the format known at compile-time
12:29mwk: seriously? ouch
12:30imirkin: you know the dims, but not the format
12:30mwk: I wonder how hard it'd be to make a gf100 gr init routine for hwtest
12:30imirkin: it's just a image2d_t or whatever
12:30mwk: I think I had a piece of code that did just that, except it worked only on GF100s with exactly 15 SMs
12:31imirkin: i think ben likes to use his user-space nouveau stuff
12:32mwk: that's an option
12:32mwk: but then I do have a grand plan for these tests
12:33mwk: and it involves looking directly into the graph context
12:33mwk: a proper driver messing with it is the last thing I'd want
13:35n7025: kubuntu 14.04 lts have tracing enabled by default. i installed it, i setup the card with nvidia binary and now i try to follow the howto but its fully outdated
13:36n7025: i cant start any x-server. startx didnt work. /etc/init.d/lightdm start also didnt work.
13:37pmoreau: Which error do you get?
13:37n7025: pmoreau: at what try?
13:37pmoreau: Trying to start X
13:38n7025: startx didnt seem to work because the filesystem is readonly and it fails with an error that it cant create such file. i had already to pipe the mmiotrace to a usb drive i mounted by hand
13:48n7025: someone should really make a live-linux that create a mmiotrace.log file on a usb-drive automaticly after boot...
14:08n7025: thanks. i get that working i think. there should be really some more simple way to make a mmiotrace for nouveau development so that normal user can just run something and they are done
14:51pmoreau: Hum… mkLoadv do not set the type of the destination
15:14karolherbst: mupuf: funny, a real 50% pmu counter memory utilization is interpreted as 71% by the userspace tools
15:17karolherbst: and 1 means 27%
15:19karolherbst: and then also it depends on the current clock :/
15:19karolherbst: how messy
15:23karolherbst: and nvidia doesn't care about the pcie load at all
23:45imirkin: ok, that was easier than i thought... doubles are kinda-sorta-semi done for ssbo, indirect blocks are done.