04:58 imirkin: karol - looks like you were right - no modifications necessary in the compiler for enhanced layouts. nha didn't implement it quite the way i had originally thought, and the spec legislates out certain scenarios that we don't handle.
08:00 karolherbst: imirkin: was that regarding that broken test or something else? No idea where enhanced layouts came into play in what I was talking about
08:05 karolherbst: imirkin: by the way your patch regarding preret: "helped inst shaders/orbital_explorer.shader_test - 2 5167 -> 5163" :)
10:37 hakzsam_: imirkin, yep, just checked myself, all arb_enhanced_layouts tests pass here :)
10:38 hakzsam_: OpenGL core profile version string: 4.5 (Core Profile) Mesa 12.1.0-devel (git-6d4b1ed)
10:38 hakzsam_: awesome :)
11:24 pmoreau: hakzsam_: For which chipsets? O.O
11:48 hakzsam_: pmoreau, fermi and kepler
11:49 pmoreau: Nice!
11:49 hakzsam_: thanks to Nicolai actually :)
11:49 pmoreau: In time for 12.1 or not?
11:49 pmoreau: Ok
11:49 hakzsam_: probably in time for mesa 13 yes
11:50 pmoreau: :-)
12:00 hakzsam_: GL 4.5 for Maxwell won't be enabled in mesa 13 though
12:01 hakzsam_: coz my scheduler is definitely not ready yet :)
13:06 Guest63229: hi
13:07 mupuf: Guest63229: hey
13:07 Guest63229: is it technically possible to backport nouveau driver for 4.8 to 4.4 ?
13:07 mupuf: well, sure, it just is a little of a pain
13:07 mupuf: the easiest is to backport the entire drm/ folder
13:08 Guest63229: the api changed probably ?
13:09 mupuf: yes, it does constantly, but nothing to dramatic
13:09 mupuf: why do you want to backport? Why mot use linux 4.8?
13:09 Guest63229: what's a bit anoying is that the kernel doens't favor binary drivers wich is ok, but at the same time, we can see that it doesn't work with linux distrubutions
13:10 Guest63229: mupuf, because i use ubuntu xenial, and it's on kernel 4.4
13:10 Guest63229: but the GTX 1070 is supported by nouveau only in kernel 4.8
13:10 mupuf: I see, well, there are PPAs for ubuntu to get a newer kernel
13:10 mupuf: "supported"
13:11 mupuf: be aware that only modesetting works, no 3D acceleration
13:11 mupuf: AKA, you will see something, but all the rendering will be done on your CPU
13:12 Guest63229: mupuf, i just needed to be able to have 1080p
13:12 mupuf: that should work
13:12 Guest63229: mupuf, because i only used the 1070 for a 2 days on that computer
13:12 mupuf: but why buy a 1070 if you only need a framebuffer
13:12 mupuf: hehe :)
13:12 Guest63229: also other solution is to just install nvidia proprietary drivers
13:13 Guest63229: but I wanted to avoid that, because it requiers I think to have a xorg.conf file, wich I would have to delete when putting back amd video card
13:15 Guest63229: mupuf, it's a computer mailny on linux, but with a dual boot to play on windows sometimes. It's at my parents home and I wanted to show them the htc vive, so i brought my 1070. but my mom wanted to print something, and she usually does that from linux
13:15 Guest63229: that's why it's better not to ask what makes sens
13:15 Guest63229: hahaha
13:15 mupuf: ah ah
13:15 mupuf: indeed :D
13:15 mupuf: ack, well, backporting is a bad idea
13:15 Guest63229: so I wanted to avoid using a kernel ppa, because I prefer stability for such a parent computer
13:16 mupuf: a PPA may be a good one
13:16 mupuf: well, maybe the PPA will install the kernel under a different name
13:16 Guest63229: ha ok, that could be a good idea, I would only use that kernel when the video card is in
13:17 Guest63229: but well, maybe installing nvidia proprietary drivers would be a better option. I don't forget to remove the xorg.conf
13:18 Guest63229: well anyway it's a long time now, but i still think there is a problem with linux and the way dependencies works
13:26 mupuf: imirkin_: got the NAS, 3TB + 250GB of SSD
13:26 mupuf: and root access on the bloody thing
13:26 mupuf: so, next step, installing the new PC
13:26 mupuf: hakzsam_: is it OK if I disconnect wtrpm for a bit?
13:28 hakzsam_: mupuf, I'm running piglit, but yeah ok
13:28 hakzsam_: just don't poweroff reator :)
13:28 mupuf: the new machine will have a fermi, a c1
13:28 mupuf: that's the only single profile GPU I have that fits in the box :s
13:29 hakzsam_: nice
13:29 hakzsam_: you are going to set up the CI system? :)
13:29 mupuf: hmm, the GTX 710 is another option
13:29 mupuf: well, one step at a time, yeah
13:29 hakzsam_: sure
13:30 mupuf: so far, I want two modes: testing patch series and tracking mainline mesa
13:30 hakzsam_: yup
13:30 mupuf: and, when ezbench is smart enough to recompile and bisect multiple components at the same time, we can do full stack testing
13:31 hakzsam_: awesome
13:31 mupuf: but, yeah, that's not happening in the coming months
13:31 mupuf: in the mean time, the machines will be accesible for devs
13:31 mupuf: and the NAS will be there to share data between machines
13:31 hakzsam_: that's already a nice news :)
13:32 mupuf: that will already be a nv86, a c1 and 2 more in reator
13:32 mupuf: I may get back my nvd9 laptop too
13:32 mupuf: but it fails to run benchmarks
13:32 mupuf: I may try with the blob's firmwares
13:33 hakzsam_: okay
14:11 mupuf: it is not at all a mess in this cupboard
14:12 mupuf: and I still need to add a few more cable for wtrpm for the new machine
14:12 mupuf: all is well in mupuf world
14:12 mupuf: needed a second switch also, glad I had a spare one
14:37 mupuf: hmm, I will need a new power supply for this PC, it is noisy as hell!
15:50 imirkin: hakzsam_: btw, i think that we're still not showing GL 4.3 on GK20A - should probably figure out what's necessary there. i'm guessing just a small handful of things.
15:51 hakzsam_: imirkin, yep
15:51 hakzsam_: imirkin, I'm looking at F1 again
15:51 imirkin: heh
15:52 hakzsam_: that hang is crazy
15:52 imirkin: are you sure it's called "F1" and not "Moby Dick"?
15:52 hakzsam_: ahah
16:08 mupuf: ok, one more machine available
16:08 mupuf: oh, forgot to make it available to the outside world
16:09 mupuf: done
16:10 mupuf: 2222 -> reator, 2223 -> jetson tk1, 2224 -> laptop-g86, 2225 -> barebone (nvc1)
16:14 mupuf: next task, replace the laptop's hdd by an ssd (which I got)
16:14 mupuf: 50e for 120GB, that should be sufficient
16:54 Yoshimo: nice work on GLSL 4.5 btw
18:05 mupuf: ok, added the NFS shares on reator
18:05 mupuf: there are two, one for the ssd and one for the hdd
18:05 mupuf: they are in /mnt
18:05 mupuf: respectively: fast and storage
18:38 orbea: you guys recently made dolphin-emu almost full speed with some games, thanks :)
18:39 imirkin: i'd guess it was a dolphin-emu change rather than a nouveau change
18:39 imirkin: although there have also been general mesa improvements lately too
18:40 orbea: my guess was GL_KHR_robustness, but im not really sure
18:40 imirkin: doubtful - that just flips on the extension string - no new functionality
18:41 orbea: ah
18:44 orbea: i think its mesa still, just tried an old dolphin build from when it was certainly slower and its still fast
18:50 imirkin: well, don't complain *too* loudly ;)
18:50 orbea: heh
19:09 mwk: what the fuck, nv3
19:10 mwk: the 4 supported surface formats are... Y16, Y8, D1R5G5B5, and... D1X1R2G2B2R8G8B8
19:14 imirkin: meaning what exactly?
19:14 imirkin: 2 R channels? or subsample format?
19:14 mwk: it's just R10G10B10
19:14 mwk: but swizzled
19:14 mwk: the R2G2B2 part is low 2 bits of all channels
19:15 mwk: so that you can just stuff R8G8B8 data into the surface and it works
19:19 imirkin: oh. fun!
19:23 mwk: the Y16 format is weird, too
19:23 mwk: it seems completely crippled
19:23 imirkin: well, those high-range formats are all the rage - 30-bit screens - were those around when nv3 came out? doubtful...
19:23 mwk: well, with VGA, it's all about DAC performance
19:23 mwk: so technically yes
19:24 imirkin: heh
19:24 mwk: but 30bpp was already supported by NV1, FWIW
19:25 mwk: NV1's surface formats are Y8, D1R5G5B5, D1X1R10G10B10
19:25 mwk: R8G8B8 is not given a fuck about.
19:26 mwk: anyhow
19:26 mwk: it seems the only thing you can do with Y16 is copy it around
19:27 mwk: enabling color key, plane mask, any operation other than SRCCOPY, or color format conversion all raise format error
19:27 mwk: I wonder why it even exists
19:29 karolherbst: hakzsam: 4.5 \o/
19:39 mwk: alright
19:39 mwk: R5G5B5 dithering algorithm is unchanged from NV1, thank gods
19:40 mwk: now, what happened to chroma key... it should've hit me in the face by now
19:43 imirkin: hakzsam_: those loops are pretty common. you run a set of optimizations over and over again until you hit a so-called "fixed" point, i.e. no further progress is made
19:43 imirkin: but if you're clever about how you implement the opts and order them, you can avoid it
19:44 hakzsam_: yeah, that's why I think it's not always the right solution :)
19:45 karolherbst: yeah
19:45 karolherbst: I just wrote that thing as an experiment to see where passes suck
19:45 karolherbst: and -0.5% is still a good improvmenet
19:47 imirkin: most compilers do them
19:47 imirkin: it's impossible to write all the later opts to take the earlier ones into account
19:47 imirkin: nouveau gets away with it because we don't do anything too crazy
19:48 karolherbst: true
19:49 karolherbst: there is still a lot of potential, but maybe we get away by analyzing what we miss or don't opt in an optimal way
20:34 mwk: ugh
20:34 mwk: I get format errors if I use Y16 with any operation other than SRCCOPY, but it *does* support other operations if you ignore the FORMAT_ERROR
20:52 mwk: alright, everything except blend and cliprects is now tested
21:38 karolherbst: any add(abs(add), x) combination can be turned into a sad, right? we would just need to add another neg in the worst case
21:43 imirkin: i don't remember what sad does...
21:43 imirkin: sum of absolute differences
21:43 karolherbst: abs(a - b) + c
21:43 imirkin: it's like |a - b| + |b - c| or something like that
21:43 imirkin: mmmm... i dunno if that's right
21:43 karolherbst: it is
21:44 imirkin: what makes you say that?
21:44 karolherbst: cuda docs: "d = c + ((a<b) ? b-a : a-b);"
21:44 imirkin: ah, fun
21:44 karolherbst: yeah
21:44 karolherbst: not _that_ use less
21:44 karolherbst: but you know what makes it uselesS?
21:44 karolherbst: it only works for ints
21:44 karolherbst: ....
21:45 imirkin: it's used for specific algorithsm i think
21:45 karolherbst: yeah sure
21:45 karolherbst: it actually happens in the sahder we have
21:45 karolherbst: but only for floats