05:45pmoreau: imirkin: I'll try
06:07karolherbst: pmoreau: are you on 4.5 already?
06:08karolherbst: ahh okay. Then you can simply use my stable_reclocking_kepler_v2 branch
06:08pmoreau: Cool! So it includes the serie you sent yesterday?
06:08karolherbst: and master
06:09karolherbst: my master_4.5 branch is skeggsb/master just compatible with 4.5
06:09karolherbst: and I base all my stuff on it (or older master_$ver branches)
06:15pmoreau: I need to clone Linux, and I'll build your branch and try it.
06:15mithro: karolherbst: any info I can provide from my setup which would be useful to you?
06:17karolherbst: mithro: mhh you was the one with the three displays right?
06:17mithro: karolherbst: the 3x4k - yeah
06:17karolherbst: well I never looked into the display stuff myself
06:17karolherbst: maybe mupuf? He got 3x4k working
06:17karolherbst: mupuf: but only if he boots with all dispalys connected
06:18karolherbst: mithro: hotplugging still doesn't work, right?
06:18mithro: I'm not sure, my displays seem to disappear when I turn them off and then reconnect when they turn back on
06:18karolherbst: ohh okay
06:18mithro: and that seems to work okay on the x side of things (not so much on the window manager side)
06:19karolherbst: mithro: are you using plasma5 by any chance?
06:19karolherbst: or something qt5 based?
06:19mithro: nope - awesomewm with gnome
06:19karolherbst: there is a qt5 bug
06:19karolherbst: mithro: but booting with both DP displays and then plugging in the hdmi one later still doesn't work, right?
06:20mithro: karolherbst: I haven't tested that recently, I can if you want
06:20karolherbst: yeah, would be nice
06:20karolherbst: if this is a "stable" bug, then we should look into that
06:40mupuf: karolherbst: sorry, never looked at the display. I know imirkin and pmoreau did
06:40karolherbst: mupuf: ahh okay
06:40karolherbst: I never looked there too, so
06:41pmoreau: Not to much success though, but I did look at it true
06:43pmoreau: Now that I have a external screen at home, I'll be able to look back at some problems I had with external screens on my laptop.
06:44pmoreau: Hans also looked a bit at display, but I don't know how far he went.
06:45karolherbst: well the issue mithro had was that he had 2 4k dispalys connected through DP and when he connected a third one through HDMI nouveau couldn't read the EDID
06:45karolherbst: but booting with all connected seems to work fine
07:02pmoreau: imirkin: Should I make a fix of the fix, since the commit has already been pushed to master?
08:06imirkin: pmoreau: go ahead
08:34pmoreau: karolherbst: Back on Nouveau with your patches, let's see if everything goes well
08:35arizo: what do those patches do?
08:35karolherbst: stabilize kepler reclocking
08:36karolherbst: in a way, that not working reclocking is a bug :p
08:36arizo: don't have that kepler
08:37pmoreau: karolherbst: Well, I can go up to 0f without any crash, nor "failed to change voltage" messages.
08:37karolherbst: yeah, this should always work with these patches
08:37karolherbst: heavy load is more interessting
08:37pmoreau: But I should probably try to put some more load
08:38pmoreau: What would you recommend for heavy loading?
08:38karolherbst: full load, run heaven or something
08:38pmoreau: Ok, let's see how to get it
08:38karolherbst: or furmark
08:38karolherbst: or well, anything should do as long as vsync is disabled
08:39karolherbst: I still have to implement some stuff (reclock/revolt on temperature change)
08:40arizo: why not work over fermi
08:40karolherbst: because fermi doesn't have memory reclocking
08:40karolherbst: and it does affect fermi
08:40arizo: all the rest is already better
08:41karolherbst: arizo: yeah well, memory reclocking is really hard to get right, you have to spend months for this
08:41arizo: can i help?
08:41arizo: i can try to find that bit
08:41karolherbst: it is more like a script
08:41karolherbst: you have to generate a working script on the fly
08:41karolherbst: it isn't trivial
08:42karolherbst: it includes a lot of tracing of the nvidia driver and finding the generic algorithm which is behind what nvidia does and so on
08:43karolherbst: well you could try to do that, but it is not a beginner task and yeah well, it isn't easy
08:44arizo: can we write to nvidia and ask them to help?
08:44arizo: then i can't help i thought u need just simple trying attempts to get it activated
08:45imirkin: pmoreau: do you happen to have dEQP set up on a machine with i965?
08:46karolherbst: arizo: well you could find other things which doesn't work and try to solve them first, over time you might able to actually help with the reclocking bits
08:47CustosLimen: so I'm having some weird situation here: https://bpaste.net/show/a0db917ded2a
08:47CustosLimen: I can see "1920x1200" in output of xrandr
08:47CustosLimen: but when I try and use It I get xrandr: cannot find mode 1920x1200
08:47arizo: karolherbst: wich things?
08:47karolherbst: arizo: no idea, whatever doesn't work on your gpu
08:48imirkin: CustosLimen: fyi, you're not using nouveau
08:48karolherbst: CustosLimen: and try out arandr (or the settings of your desktop environment), maybe you just get the parameters wrong
08:48CustosLimen: karolherbst, arandr gives same error
08:49CustosLimen: karolherbst, I used it and switched to xrandr cos it did not work
08:49CustosLimen: imirkin, how do you know ?
08:49imirkin: CustosLimen: i'm guessing the issue is around the fact that the fb is too small
08:49imirkin: CustosLimen: the names of the connectors... either radeon or an old xf86-video-modesetting
08:49imirkin: CustosLimen: try xrandr --output DisplayPort-0 --right-of DVI-0
08:50CustosLimen: imirkin: xrandr: cannot find output 0x46
08:50CustosLimen: let me check xorg log
08:51imirkin: are both displays up and running and mirroring each other?
08:51pmoreau: imirkin: I should have i965 on this laptop, but… I haven't played much with it. Only once, to set it as the default "boot" card, and I ended up with a complete black screen, as soon as it booted.
08:52CustosLimen: imirkin, yes
08:52imirkin: pmoreau: ah ok. don't worry about it then.
08:54CustosLimen: imirkin, here is my xorg log: http://termbin.com/b13f
08:54karolherbst: pmoreau: this is a macbook with the on the fly swithching support?
08:54CustosLimen: glxinfo output: http://termbin.com/gzez - and lspci output: https://bpaste.net/show/6cffe8071acc
08:55CustosLimen: brb - let me reboot - I just compiled with glamor support
08:55karolherbst: pmoreau: I thought this was fixed in the gmux code?
08:56imirkin: [ 9.554] Current Operating System: Linux hadean 4.1.15-gentoo-r1 #1 SMP Mon Jan 25 03:01:31 SAST 2016 x86_64
08:56karolherbst: uhh gentoo stable
08:56imirkin: CustosLimen: i would recommend something newer for better GM20x support. fyi, there's no acceleration for now.
08:57imirkin: karolherbst: yeah, some crazy people out there don't update to linus's master branch every day :p
08:58karolherbst: I am always annoyed by the name of "git-sources" because it is totally a package for the stable kernel
08:58karolherbst: there is nothing gitish about it
08:58pmoreau: karolherbst: Could be. I think it's one of the rMBPs which caused troubles to l1k.
08:58karolherbst: pmoreau: I had one of these once, but sandybridge based + radeon
08:58karolherbst: I got intel to work
08:59karolherbst: there is a magic reg you have to set (if it is still the same) to pipe the output through intel
08:59karolherbst: I mean, well
08:59karolherbst: you know what I mean
08:59CustosLimen: imirkin, latest stable gentoo kernel is 4.1.15-r1
08:59CustosLimen: imirkin, what version would you recommend ?
08:59CustosLimen: imirkin, I can install others - just mentioning it
09:00karolherbst: pmoreau: but I am sure I lost my script to do that...
09:00CustosLimen: I have some nvrm stuff in my dmesg
09:00CustosLimen: just came when I restart xorg
09:01pmoreau: karolherbst: Unigine did put some pressure (running it 1440p at 8x MSAA and extreme tesselation), and made Nouveau unhappy (mostly some "nv50cal_space: -16").
09:01CustosLimen: what is this ?
09:01pmoreau: karolherbst: So, seem to be stable
09:01karolherbst: pmoreau: MacBookPro8,2
09:01karolherbst: this was what I had
09:01karolherbst: pmoreau: ohh 8xmsaa is broken anyway
09:01karolherbst: try with 4x
09:01karolherbst: this should work
09:02imirkin: CustosLimen: i dunno, 4.5 was just released, you could try that :) although if you want to live on the edge, you could grab airlied's drm-next branch which should include acceleration support for your GPU
09:02imirkin: CustosLimen: (you'll need to download some firmware too for that)
09:02CustosLimen: see - my real problem is this - I have this card http://www.gigabyte.com/products/product-page.aspx?pid=5212#ov
09:03karolherbst: pmoreau: https://wiki.archlinux.org/index.php/MacBookPro8,1/8,2/8,3_(2011)#EFI_Boot the outb stuff
09:03CustosLimen: and if I run it with nvidia drivers - its noisy as hell
09:03CustosLimen: but anyway, maybe I should just use my intel onboard gfx
09:03CustosLimen: I dont game allot anyway
09:03CustosLimen: maybe starcraft 2 once every few months
09:03CustosLimen: and that works ok with intel
09:03karolherbst: but there was something else
09:03CustosLimen: should really just not buy gigabyte
09:05CustosLimen: will try in a few months when newer version of linux is stable in gentoo
09:05karolherbst: pmoreau: but good to know that it also works for you without issues :)
09:09pmoreau: karolherbst: it's more fluid with MSAA x4 and tesselation set to normal. Almost 3fps all the time! :-D
09:09pmoreau: MBP11.3 for me
09:13arizo: karolherbst: i will try to write to nvidia corp
09:13arizo: what should i ask from them? tell me the recloking bit?
09:22karolherbst: pmoreau: wow :D
09:23karolherbst: pmoreau: well if you boot with config=NvBoost=2 you get slightly higher clocks maybe
09:51karolherbst: pmoreau: can you still upload your vbios, so that I know which parts are tested by you
09:53pmoreau: I'll try again, see if I can manage to retrieve it
09:56pmoreau: karolherbst: Ok, it worked. Need to clone the repo and put it in there
09:57tsermyat: karolherbst: what should i ask from them? tell me the recloking bit?
09:57imirkin: or else!
10:04imirkin: hakzsam: btw, happy to explain stuff in my textureGrad cube fix
10:04karolherbst: tsermyat: there is no single bit and if you don't know what to ask for better don't ask at all
10:05imirkin: hakzsam: i suspect there are about 10 things that prevent you from understand wtf is going on :) so i'm not surprised you're confused by it
10:05karolherbst: pmoreau: thanks a lo
10:06tsermyat: karolherbst: tell me what to ask and i will write to them
10:08pmoreau: karolherbst: Could you please upload it for me? :-D My new SSH key isn't recognised by the git user, and my USB key doesn't want to work. --"
10:08karolherbst: tsermyat: if you really want to help, you should start with something much simplier, some issue you encounter on your card. the memory stuff is most likely 10+ pages full of documentation and I really doubt they just give it to somebody asking a question
10:08karolherbst: pmoreau: k :) send it to me then
10:09tsermyat: they must provide support
10:09tsermyat: i bought their card
10:09tsermyat: and can't use it
10:10imirkin: a fool and his money are soon parted?
10:10pmoreau: karolherbst: Thanks: https://phabricator.pmoreau.org/F38124, the strap_peek is in the comment
10:11karolherbst: some are really optimistic with that kind of stuff
10:11imirkin: karolherbst: more like 100+ pages of docs :)
10:11karolherbst: well 100 is a bit much
10:11imirkin: memory is complicated.
10:11karolherbst: that would mean that they do internal documentation real good
10:11imirkin: you're just trying to make it work the expected way
10:12imirkin: but there's a TON of stuff you're not trying to do
10:12imirkin: coz the blob doesn't do it
10:12karolherbst: I was only talking about the stuff we would need
10:13karolherbst: k, now lets have a look at the vbios
10:13karolherbst: what the
10:14pmoreau: Did I messed up something?
10:14karolherbst: 8 extedecs?
10:14karolherbst: no, it is fine, I am just buffled by the amount of extdevs
10:14karolherbst: 2xCHIL_I2C, 5xADS1112 and one unknown
10:15karolherbst: LCD4 Lamp Status. LCD4 corresponds to the LCD4 defined in LCD ID field in Connector Table.
10:16karolherbst: ohh wrong table
10:16karolherbst: k, that unknown isn't documented
10:18karolherbst: pstate 0
10:18karolherbst: pmoreau: your vbios is something, something totally different
10:18karolherbst: still fine in the end, but still different
10:19imirkin: hakzsam: anyways, let me know if you want me to explain... otherwise i'll just push.
10:19karolherbst: pmoreau: nouveau clocked to 926MHz right?
10:20pmoreau: I didn't checked which clock it picked
10:20pmoreau: Let me try again
10:20pmoreau: I still have the results: it picked 925MHz
10:21hakzsam: imirkin, yes, it would be good to have some explanations but I don't really have time right now, a bit later in one hour or so if you are still here :)
10:21imirkin: ok, we'll see
10:21karolherbst: pmoreau: k, and also, your board/cpu only supports pcie 2.0?
10:21pmoreau: karolherbst: And 405 for perflvl 07
10:23karolherbst: ohh it is only a U cpu
10:23karolherbst: they don't seem to have pcie 3.0 support
10:23karolherbst: and only 12 lanes in total
10:23pmoreau: "Capabilities:  Express (v2)" for the GK107M, so v2 I guess
10:24karolherbst: it is also v2 for pcie 3.0
10:24karolherbst: the lnkCap should be only at 5.0
10:24pmoreau: LnkCap: Port #0, Speed 8GT/s, Width x16
10:25pmoreau: For the GPU at least
10:25karolherbst: pmoreau: and on 0f lnkSta is at 5.0 width x16?
10:26karolherbst: and huh, the max entries can be set to 0xff
10:26karolherbst: I have to check that in my patches
10:27karolherbst: there is just no proper GPUboost on that gpu
10:27karolherbst: no revolt on temp changes
10:27karolherbst: every clock has a static voltage
10:27pmoreau: LnkSta: Speed 5GT/s, Width x8
10:28karolherbst: k that makes sense from the CPU documentation
10:28karolherbst: 4258U? or other cpu
10:29pmoreau: i7-4850HQ according to `/proc/cpuinfo`
10:29karolherbst: ahh optional cpu
10:29pmoreau: optional cpu, what?
10:29pmoreau: How can the CPU be optional
10:30karolherbst: configure option of the model
10:30karolherbst: pmoreau: boot with nouveau.config=NvBoost=2
10:31karolherbst: nouveau should then clock to 1028MHz
10:31karolherbst: 1029 actually, but doesn't matter
10:32karolherbst: and also run some benchmarks with that, just to see how stable that is
10:32karolherbst: and then I would like to know what nvidia does with that gpu
10:32karolherbst: especially the pcie part
10:33pmoreau: Argh, too many things to do! :-D
10:33karolherbst: well you can skip the NvBoost thing
10:33karolherbst: it would be more interessting to know what nvidia does
10:34karolherbst: pmoreau: you actually only need 4 of those patches...
10:37karolherbst: pmoreau: so I try to get at least those patches merged for the next release because these are rather trivial
10:37pmoreau: Nice :-)
10:39karolherbst: pmoreau: I will create a branch with all the stuff you need, then you can check if that branch also works for you and then if you want you can review those patches ;)
10:40karolherbst: pmoreau: https://github.com/karolherbst/nouveau/commits/pmoreau
10:40pmoreau: Will there be more stuff than on the stable_reclocking_kepler_v2 one? Or is it a subset?
10:40karolherbst: the pmu stuff is only there for stability
10:40karolherbst: pmoreau: a small subset
10:40karolherbst: 5 out of 19 patches actually
10:40pmoreau: Sounds good :-)
10:41karolherbst: but if I can get it merged in smaller pieces I am also happy with that
10:43karolherbst: pmoreau: in the end your gpu will be the only one where we will use the same voltage as nvidia does, because the tables are rather trivial
11:13pmoreau: imirkin: Where is the constant propagation stuff happening?
11:21l1k: pmoreau: what you got an MBP 11,3 now?
11:21pmoreau: l1k: It's my work laptop
11:21l1k: karolherbst: this outb stuff is no longer necessary with 4.6 to switch to intel on an MBP 8,2 :)
11:22l1k: pmoreau: nice!
11:22pmoreau: I received it ~two weeks ago
11:22l1k: pmoreau: that's a retina. no GPU switching on that machine just yet.
11:23pmoreau: Right, but I wanted to have the Intel as main, and use PRIME for Nouveau
11:23pmoreau: That should work, shouldn't it?
11:24l1k: pmoreau: yes if you use gpu-switch tool to switch to intel on the next boot.
11:25pmoreau: That's what I did, but I ended up with a complete black screen.
11:25l1k: pmoreau: the intel card is normally hidden on this machine. you need either a kernel patch or a grub patch to make it visible.
11:25l1k: if you boot with the nvidia card, does lspci show the intel card at all?
11:25pmoreau: I don't think so
11:26l1k: that's the problem right there :)
11:26pmoreau: In which kernel can that patch be found?
11:26pmoreau: (I'm not using grub, and not planning to go back to it ;-))
11:28l1k: quote: "Otherwise you will end up with a powered-down integrated graphics card and a black screen." :)
11:28pmoreau: I can confirm! :-D
11:29pmoreau: The oatch was never merged?
11:29l1k: no, not yet
11:29l1k: I'm not using grub either. gummiboot. :)
11:30pmoreau: I was using gummiboot, but now bootctl as it merged in gummiboot
11:32pmoreau: I'll try that patch out, thanks for the link!
11:35karolherbst: l1k: ohh then it doesn't work cause of retina... right
11:37pmoreau: Oh, but I should not get the Intel working… I will be tempted to play with Vulkan… :s
11:42l1k: pmoreau: I've implemented runtime pm for Thunderbolt, saves about 1.5 W... so far I've only tested it with the Light Ridge controller built into my older MBP 9,1. if you have some time to spare maybe you could give those patches a spin on your MBP 11,3. it has a Falcon Ridge controller, very different. would be interesting to know if it works there as well. https://github.com/l1k/linux/commits/thunderbolt_runpm_v1
11:43pmoreau: l1k: Nice!
11:43pmoreau: I'll try to give it a try tomorrow
11:49imirkin: pmoreau: the ConstantFolding pass :)
11:49pmoreau: The name was too obvious, I thought it was a trap! :-D
11:51pmoreau: I'll have a look there to see if I can simplify `mad r0d r1d 0` for example
11:51imirkin: pmoreau: i thought i took care of that... maybe i only did it for F32
11:52pmoreau: In my case, I have u64.
11:54pmoreau: I'll clean up a bit the split64BitOp patch for MUL and MAD, and CVT for U64 and S64, before sending them and looking at cst folding
11:55imirkin: pmoreau: ah yeah, if it's done at split time, that's *after* all the opt passes
11:57pmoreau: So, it's missing some constant folding for u64 (at least)
11:58imirkin: my point is that if the ops are created by split64BitOp
11:58imirkin: then no amount of logic in ConstantFolding will fix it
11:58imirkin: since all the opts happen *before* the splitting, not after
12:00pmoreau: But I'm just splitting an existing 64 bit mad/mul to two 32 bit mad/mul. If there was a 64bit mul involving two immediates, or 0, that should be taken care before reaching the splitting
12:00imirkin: ah yeah, good point
12:00pmoreau: My patch is only adding the split from 64 mad/mul to 32 mad/mul, nothing else.
12:00imirkin: i think a lot of that stuff is keyed to only work with 32-bit quantites
12:01pmoreau: Who uses 64bit values anyway
12:01imirkin: there's no real way for U64 stuff to exist atm, so it's not well-handled
12:01imirkin: and the F64 support is minimal too
12:01pmoreau: It makes sense
12:06karolherbst: what I hate most is, when applications are messing with mouse/display settings...
12:07imirkin: karolherbst: do you have a mesa 11.2-rcN install locally perchance?
12:07pmoreau: I never experienced applications messing with mouse settings… but… certainly looking forward to that day!
12:07karolherbst: imirkin: only mesa master
12:07imirkin: do you use risen?
12:07karolherbst: I know the issue
12:07imirkin: does it happen for you?
12:07karolherbst: the vegetation thingy, right?
12:08imirkin: coz i tested on my system mesa and i saw weirdness
12:08karolherbst: yeah well, it is d3d8
12:08imirkin: and then i tested with my current tree and it works fine
12:08karolherbst: ohh mhh
12:08karolherbst: you need a setting set
12:08imirkin: i do have a few local patches
12:08karolherbst: enhanced vegetation or something
12:08imirkin: well, it was a trace
12:09imirkin: so whatever settings were used to record the trace repro'd the issue initially
12:14imirkin: checking to see if it repros with origin...
12:14imirkin: coz i'm 99.999999% sure it's not my textureGrad fixes :)
12:15imirkin: (who uses explicit derivatives and cubemaps... that's just crazytalk)
12:15karolherbst: ohh so the issue goes away with the trace?
12:15karolherbst: ahh okay
12:18imirkin: yeah, looks fine with origin
12:18karolherbst: pmoreau: it would be nice if you would leave comments on the patches then if you can test them. Maybe we can convince ben to merge some of them together :D
12:18imirkin: karolherbst: mind testing it out real quick and see if it's all fixed for you too?
12:18karolherbst: imirkin: have to download the game, but yeah, will test it
12:18imirkin: well just grab the trace from the bug
12:18imirkin: that might be faster
12:18karolherbst: yeah, but I want to test it for real
12:18karolherbst: and I am already 70% done downloading
12:19imirkin: hehe ok
12:19karolherbst: and I have to update mesa anyway, my build is like more than a month old
12:21imirkin: the only additional difference i can think of is debug vs non-debug
12:21imirkin: anyways, the trace replays fine @origin for me
12:21imirkin: so... yeah.
12:21karolherbst: somebody fixed something :D
12:21imirkin: at some point.
12:22imirkin: that's why i was asking about 11.2-rc
12:22imirkin: it'd be nice if the fix made it into the release
12:22karolherbst: when was 11.2 branched?
12:22imirkin: like a month ago
12:23imirkin: but it's had fixes cherry-picked
12:23karolherbst: I could try it out though
12:28karolherbst: mhh now the game crashes for me... maybe my wine is messed up
12:28imirkin: could you just try the trace in that bug?
12:28imirkin: i bet it runs 100x faster for you
12:28imirkin: coz you don't have a tricycle as your gpu
12:29karolherbst: k, download trace
12:29karolherbst: ohh maybe I won't even see the issue
12:30imirkin: well, if you try it with mesa 11.1.x, i bet you will
12:32karolherbst: I ment because of the replay speed
12:32imirkin: oh, well it was pretty visible
12:34imirkin: pmoreau: do you have push privileges on mesa?
12:34karolherbst: imirkin: by the way, how do I replay that race, because it is just black for me
12:34imirkin: karolherbst: just wait a bit
12:34imirkin: it pops up like 75 full screen windows
12:34imirkin: and stays black for a minute
12:36karolherbst: regarding the speed: Rendered 6461 frames in 21.9763 secs, average of 293.998 fps :D
12:36karolherbst: and yeah, seems also be fixed for me
12:37karolherbst: will wait until mesa-11.2 compiles
12:37karolherbst: and then I test again
12:39imirkin: karolherbst: is that with mesa master, or mesa master from 1 month ago?
12:39karolherbst: master from now
12:39karolherbst: compiling the 11.2 branch now
12:40karolherbst: imirkin: does it gets to the ingame thing faster if you disable vsync?
12:40karolherbst: because with vsync it was slow as hell for me too
12:40pmoreau: imirkin: I don't, nor have I a freedesktop account.
12:40pmoreau: Planning to ask for one one day
12:41imirkin: pmoreau: ah ok
12:41imirkin: karolherbst: didn't try
12:42karolherbst: imirkin: 11.2.0-rc3 also broken
12:43karolherbst: do you want to bisect or should I?
12:43imirkin: i don't want to bisect
12:43imirkin: i'm also not going to make you do it unless you want to :)
12:43karolherbst: well it should be done in a few minutes
12:44imirkin: done bisecting??
12:44imirkin: my bet is that the pbo upload changes "fixed" it
12:44karolherbst: not that fast :D
12:45karolherbst: ohh meh, I have to bisect reverse....
12:45imirkin: yeah, flip your brain around for what's good and bad
12:45imirkin: i wish bisect had a "reverse" mode
12:45imirkin: that did it for you
12:45karolherbst: yeah or it would notice by itself
12:47karolherbst: imirkin: which commit would it be?
12:48imirkin: from nha...
12:49imirkin: commit 757071ca7 first started adding them, but PBO uploads were broken on nouveau until 2b938a390c
12:52karolherbst: then I will try 2b938a390c and if not, I just let git decide what to test next
13:26karolherbst: imirkin: it seems something rather new
13:32karolherbst: maybe even this month
13:33karolherbst: still broken with this: https://cgit.freedesktop.org/mesa/mesa/commit/?id=60c08aa90bce4c8766a747c8517f7ff6987937f0
13:33imirkin: perhaps the SQRT stuff?
13:33imirkin: karolherbst: c1e4a6bfbf0
13:33karolherbst: that might make sense actually
13:34karolherbst: I could imagine it also fixes talos principle
13:35imirkin: you have quite a vivid imagination
13:36imirkin: could also be the blit fail fix: a651bc027d
13:37imirkin: although it's unlikely it'd be using such large fb sizes
13:45karolherbst: imirkin: on c1e4a6bfbf015801c6a8b0ae694482421a22c2d9 it is fixed
13:45karolherbst: checking what it really fixed though
13:50karolherbst: imirkin: one of these commits: https://github.com/karolherbst/mesa/compare/8be9efcce70f671afcaf3cfc76c624b3467b7834...c1e4a6bfbf015801c6a8b0ae694482421a22c2d9
13:50karolherbst: I would assume it is one of yours though :D
13:51imirkin: seems likely
13:53karolherbst: c1e4a6bfbf015801c6a8b0ae694482421a22c2d9 it is
13:55karolherbst: yep, I can verify, that this commit realy fixes it :)
13:56imirkin: ok cool
13:56imirkin: can you make a note in the bug description?
13:56karolherbst: already done
13:56imirkin: cool :)
13:56karolherbst: I also encountered the issue, but was too laty to debug it :D
13:58karolherbst: but the mesa build system needs to be fixed, because I am a little bit annoyed by the fact, that I have to make clean...
14:00imirkin: you shouldn't need to... i never do
14:00imirkin: you must be doing something odd
14:01karolherbst: out of tree build
14:01imirkin: you should let xexaxo know - he tends to deal with such issues
14:01imirkin: although i think he's on vacation or something
14:02karolherbst: I can't do in tree builds, because I need 32bit builds sometimes, but with in tree build it works
14:02karolherbst: I guess some path is just wrong or something
14:02imirkin: yeah dunno
14:02karolherbst: anyway, when I modify a header, it doesn't rebuild anything
14:03imirkin: that's messed up
14:03karolherbst: leads to all kind of crappy issues
14:33pmoreau: karolherbst: Works for me with an out-of-tree build
14:37karolherbst: pmoreau: the branch I made for you?
14:38pmoreau: Nah, the Mesa stuff :-)
14:38karolherbst: mhh odd
14:38pmoreau: I haven't tested your new branch yet, but it's all build and ready to try as soon as I reboot
14:39karolherbst: sadly those patches help nearly nobody though
14:39karolherbst: but better to have some merged than nothing
19:54imirkin: hakzsam: so the bits you were probably missing on that patch...
19:57imirkin: (a) lod selection - this happens based on implicit derivatives - frag shader invocations execute in 2x2 quads, and you look at the coordinates in the other quads to figure out the derivatives
19:57imirkin: there's some way that a particular derivative maps onto a particular lod, the details of which are unimportant
19:57imirkin: (b) cubemaps - the idea is that you have a cube, i.e. 6 faces, but you provide spherical coordinates (i.e. 3 axes), and whereever that vector points onto the cube, that's the texel you select
19:57imirkin: however there are some ambiguities with which face to pick, and there's the concept of a "major axis", which determines this
19:57imirkin: (and there's something known as seamless cubemap filtering, which is when the derivatives are such that the other pixels in the quad map onto a different face)
19:57imirkin: so under normal circumstances, each shader execution gets its coordiantes, and the derivatives are calculated from that, and all is well
19:57imirkin: however under special circumstances, like when using textureGrad() you want to be able to provide explicit derivatives rather than the implicit ones as above
19:57imirkin: nvc0 has *some* support for that, but only for 1d/2d textures, not 3d/cube. and nv50 has support for none of that.
19:57imirkin: so the idea is that you load up the coordinates into each lane *as if* it were the normal case (using quadop cleverness)
19:57imirkin: and then just run the texture op, but save off only the results from the lane in question
19:57imirkin: so instead of doing just 1 texture op for all 4 lanes, you end up doing 4 separate ones
19:58imirkin: so the thing here was that we were normalizing the cube coordinates at the wrong time
19:58imirkin: instead of "load coordinates; normalize", coordinates would get modified *after* the normalization
19:59imirkin: i've (hopefully) fixed that
20:55imirkin: if someone gets a chance, could use mmt traces of dEQP-GLES3.functional.fbo.depth.depth_test_clamp.depth_component24 and dEQP-GLES3.functional.fbo.depth.depth_write_clamp.depth_component24 against the blob...
20:55imirkin: (including whether they pass, but even if they fail, could use the traces anyways)