00:17imirkin_: Lyude: i have a branch which adds 10bpc support
00:35imirkin_: Lyude: btw, a lot of people implied that it all worked fine with 4.13 but errors come up in 4.14 -- have you checked that claim?
01:38pmoreau: I need to test the Meson build + check I haven’t made too many mistakes regarding the coding style, and then it’s time to resend the clover series.
01:39pmoreau: I think with the linker gone away, and by implementing the OpenCL 1.2 extension rather than the OpenCL 2.1 core support, it should have an easier time getting merged.
01:40imirkin_: skeggsb: no compression in stolen vram?
01:43imirkin_: MCP89 too?
01:44nyef``: ... PCI pass-through on an MCP89 is a bad idea, isn't it?
01:47imirkin_: nyef``: depends on the end goal
01:47imirkin_: if it's to see how many sparks you can make fly out of your computer at once... it's a pretty good idea
01:48nyef``: If I wanted sparks to fly out of a computer, I'd use thermite.
01:49imirkin_: the easy way.
01:49imirkin_: c'mon, be creative
01:49nyef``: If I wanted sparcs to fly out of a computer, I'd overvolt one a SunFire v210.
01:50imirkin_: wasn't there a user-space instruction that'd set a sparc-1 on fire?
01:50nyef``: s/one /
01:50nyef``: That sounds fun.
01:50nyef``: The classic "HCF" instruction?
01:51imirkin_: i dunno
01:51imirkin_: this is harkening to a forgotten age
01:51skeggsb: imirkin_: actually, those do support it.. we have had some troubles reported with that in the past
01:52imirkin_: skeggsb: huh, ok.
01:52imirkin_: troubles relating to compression?
01:52imirkin_: or just troubles?
01:53skeggsb: yeah, like 3d apps getting "holes" in them when using it etc
01:53imirkin_: i'm aware of nvaf troubles in general which felt like context switches being broken
01:53skeggsb: it only exists in that werido macbook
01:53imirkin_: so ... maybe they don't quite support them?
01:53skeggsb: nvidia use it
01:53imirkin_: nah, there was also a mac mini
01:53imirkin_: perhaps nvidia knows something we don't?
01:53skeggsb: i think they know a lot we don't ;)
01:54nyef``: Mac Mini 2010 has MCP89.
01:56airlied: skeggsb: I think we still have that wierd macbook in the office
01:56skeggsb: yeah, i thought we did
02:00nyef``: ... Do you also have one of the HDMI displays that the MCP89 audio doesn't work with?
02:01nyef``: (Confirmed with the PlayStation 3D display, but there may be others.)
02:01imirkin_: airlied has every display
07:23karolherbst: imirkin: I will test your maxwell bindless texture patches over the weekend and will probably add support for that in my nir pass as well if that works out
10:24karolherbst: mupuf: I kind of need fast access to the apitraces :(
10:53karolherbst: imirkin: [29/29] skip: 1, pass: 14, fail: 8, crash: 6
10:54karolherbst: that was with nir enabled...
10:54karolherbst: TGSI: [29/29] skip: 1, pass: 28
10:55karolherbst: I should force 4.5 as well
10:55karolherbst: [29/29] pass: 29
10:57RSpliet: karolherbst: was your emphasis on fast? As in, the git repo is slow to operate and that's wasting your and your bosses time? If this sort of infra is important for your performance, it wouldn't harm checking whether RH can take over hosting this from mupufs hobby machine ;-)
10:58karolherbst: RSpliet: I don't want to wait 2h+ until I cloned a repository
10:59RSpliet: karolherbst: Ah in that case my point stands. I don't think it's fair to expect greater expenses from mupuf as a hobbyist, so if this requires changes you might want to talk to your boss as well and see if RH can play a role here that will benefit the community as a whole
11:00karolherbst: RSpliet: I already suggested we could just use FTP for traces
11:00karolherbst: and who said I need it for RH work related stuff :p
11:01RSpliet: Hah, I call your proposal and raise you SFTP ;-)
11:01karolherbst: works for me
11:04RSpliet: And I think you should be saying you'll use it for RH related work. If we need an upgraded infrastructure they are the most likely ones able to help out... politics :-P
11:05karolherbst: I doubt that
11:05karolherbst: first point on my list is some ezbench machine and this will take a while
11:06RSpliet: Is that one of those machines with 8 full-blown PCI-E ports?
11:06karolherbst: that wouldn't make any sense
11:11karolherbst: imirkin: will test dow3 as well
12:34karolherbst: pmoreau: do you have some mostly cleaned up clover patches somewhere?
12:35pmoreau: My clover_spirv_support branch is the very latest, and quite cleaned up. Only needs some checking, and that will be the version I’ll send to the list.
12:36pmoreau: I need to rebase to “squash” the patches I made yesterday with the other patches, but otherwise should be good.
12:48karolherbst: pmoreau: nice :)
12:48karolherbst: pmoreau: I plan to do some experiments on how to just get some super trivial stuff working
12:49karolherbst: opencl -> spirv -> nir -> nvir
12:49pmoreau: Sounds great!
12:49karolherbst: maybe even just pass a spir-v kernel instead of a cl one
12:49pmoreau: Yup, you can totally do that with my series
12:49karolherbst: I think there were some patches recently to adjust the current spir-v to nir pass to handle such stuff
12:50pmoreau: I’ll most likely still send the series over the weekend without any of your changes, but it shouldn’t be hard to integrate them afterwards, or send a new series.
12:51karolherbst: yeah, that should be fine
12:54karolherbst: imirkin: regression in piglit/bin/arb_shader_image_size-builtin --quick -auto -fbo
12:54karolherbst: suq 1D $r255 $s31 f32 $r0 $r0
12:54karolherbst: ../src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp:3457: virtual bool nv50_ir::CodeEmitterGM107::emitInstruction(nv50_ir::Instruction*): Assertion `!"invalid opcode"'
13:11karolherbst: pmoreau: but that clover_spirv_support branch contains everything, right?
13:11pmoreau: Oh yeah, I don’t have a branch with just the clover patches
13:12pmoreau: Cause I need to check that I haven’t broken anything in the process, so I still need a SPIR-V consumer :-)
13:12karolherbst: I see
13:14pmoreau: Commit messages should be clear enough (hopefully). You can pick from 08d0b3749cd4 “clover/api: Fix tab indentation to spaces”, to d74994ff18a1 “[OpenCL 2.1] clover: Use OpenCL 2.1 defines in place of cl_khr_il_program”, and then from 543ddf8fb3d8 “clover: Move device extensions definitions to core/device.cpp” to the top of the branch.
13:14pmoreau: If you only want the clover SPIR-V patches.
14:03imirkin: karolherbst: thanks
14:05imirkin: i'm out for a week, so won't be looking at it in that time -- feel free to take a stab at fixing it if you like
14:06imirkin: karolherbst: but the first results were just for the bindless tests presumably? if so that's nice.
14:07imirkin: that suq should have gotten converted to a texquery -- weird!
14:07imirkin: i must have messed something very simple up
14:07karolherbst: it is the only regression afaik
14:07imirkin: like forgetting to change the opcode
14:07imirkin: or something else dumb
14:10karolherbst: maybe because you don't call handleSUQ anymore
14:10karolherbst: or inside handleSUQGM107 the opcode isn't changed ;)
14:11karolherbst: should become a SUQ->TXQ right?
14:12karolherbst: uhh wait
14:20imirkin: yes, it should
14:21karolherbst: imirkin: by the way, we have a lowering_gm107 file, so handleSUQ could be simply overloaded
14:22karolherbst: uhmm, whatever term is correct here
14:23karolherbst: a "suq->op = OP_TXQ;" didn't fix the test. it doesn't crash, but all subtests are just failing now
14:27imirkin: ok. so i just messed up the whole image size thing. super.
14:27imirkin: o well. i'll figure it out ... later.
14:32karolherbst: yeah, no worries
14:33karolherbst: imirkin: I went through my nir patches and I think I am okay with those now. I tracked down a few other bugs and now I have around 30 fails more than the tgsi path
14:34karolherbst: would be nice if we could get it in after the branching, so we have around 3 months of testing by whoever is interested
14:52imirkin_: well, like i said, i'm out for a week
14:52imirkin_: starting tomorrow
14:53karolherbst: yeah, no worries about that
15:23karolherbst: nice, seems like tomb raider also runs just fine :) now checking if the precise thing works as well
15:27karolherbst: nice precise also works as epxected :)
15:57karolherbst: dEQP-GLES31.functional.draw_indirect.compute_interop.large.drawarrays_separate_grid_200x200_drawcount_800: nouveau: kernel rejected pushbuf: Device or resource busy
15:58karolherbst: well I have another 368 fails I can look into :)
15:59imirkin_: i think that's the "submitting too fast" thing
16:01karolherbst: it triggers quite reliable
16:08karolherbst: I think we have a bug in a peephole
16:09karolherbst: ohhh no, it is fine
16:09karolherbst: I thought we miss a shladd
16:16karolherbst: uhh 80 fails due to missing helper invocation system value handling
16:33karolherbst: *sigh* my favourite fails: TGSI == NIR shader :(
16:34imirkin_: helper invoc is easy
16:34imirkin_: just don't get it backwards
16:34imirkin_: (coz it's really easy to get it backwards :) )
16:35karolherbst: to fix it was to add 2 locs
16:35karolherbst: another issue I found was inputs with sample flag
16:35karolherbst: and memory barriers for buffers were also missing
16:36karolherbst: sad that piglit doesn't cover those
16:36karolherbst: now I have 6 weirdo compute related fails
16:36karolherbst: only min/max
16:36karolherbst: I have an idea
16:37karolherbst: "atom min s32 $r5 s[0x0] $r0" vs "atom min u32 $r5 s[0x0] $r0"
16:37karolherbst: I am sure the type difference is the issue here
16:38karolherbst: it is even shared_atomic_imin in nir
16:38karolherbst: why do I end up with a u32?
16:39karolherbst: ohhh right
16:39karolherbst: because intrinsics have no sign information
16:41imirkin_: umin vs imin are different
16:41imirkin_: nir better have separate intrinsics.
16:41imirkin_: otherwise it's buggy.
16:42imirkin_: and can't work
16:42karolherbst: it has
16:42imirkin_: iirc that was the case initially, actually, but it got fixed
16:42imirkin_: [iirc i complained about it, but might have been someone else]
16:42karolherbst: but for alu_instr you get the signess out of that info struct
16:42karolherbst: with intrinsics you can't
16:42imirkin_: ah yeah, that might be.
16:42karolherbst: so I always treat intrinsic dests as unsigned ints
16:42imirkin_: oops :)
16:43imirkin_: all the atomic variants have the min vs max thing
16:43imirkin_: (shared, ssbo, atomic)
16:43karolherbst: not all
16:43karolherbst: image atomic doesn't
16:43karolherbst: only shared and ssbo
16:43imirkin_: that means the info comes from the image type
16:43imirkin_: i.e. if the image is uimage2D vs iimage2D
16:44imirkin_: [i assume. i don't know nir *that* well.]
16:45karolherbst: anyway, up to that pushbuf crash I have now the same passrate for deqp :)
16:45karolherbst: I am sure I fixed some piglit tests as well, because I had some weird atomic fails as well
16:45imirkin_: make sure you fix the image min/max stuff
16:45imirkin_: otherwise you'll forget.
16:46karolherbst: ohh and there is a helper invocation tests in piglit, I just overlooked it
16:47imirkin_: once we pushed it, iirc every impl was backwards.
16:47imirkin_: so yeah. it's an easy one to get wrong. i think hw naturally returns gl_NotHelperInvocation
16:47karolherbst: I added a TODO at least for that image stuff
16:47imirkin_: and everyone just forgot to invert
16:48imirkin_: although iirc on nvidia it might actually be the straightforward thing
17:41Lyude: imirkin_: I can in just a sec
18:21Lyude: imirkin_: yep; doesn't appear to happen before that
18:21Lyude: although I kind of expected that since 4.14 was the version that skeggsb rewrote a good bit of the vmm code
18:22Lyude: at least I thought it was
18:22imirkin_: so ... bisect
18:22imirkin_: vmm is rewritten in 4.15
18:22imirkin_: i don't think 4.14 had substantial nouveau changes. i forget.
18:22Lyude: lemme see
18:23Lyude: yes you are right, I feel silly for not noticing that
18:23imirkin_: could be a change in ttm
18:23imirkin_: or something entirely unrelated
18:25Lyude: heh, well I learned about memory management on nouveau at least :)
18:26Lyude: anyway; bisecting now
18:37endrift: Does Nouveau have support for the TX1?
18:37endrift: ...I can probably find this if I look hard enough, but mostly the info is for GeForce cards
18:43imirkin_: look for GM20B
18:43imirkin_: i've never tried it
18:48imirkin_: the integration between the gpu and display is ... suboptimal
19:31endrift: hmm I'd try it out at some point but first I'd need to figure out how to install Linux on it :P
19:32endrift: I have four TX1 devices and I'm pretty sure the only one I can install Linux on with documented instructions is the one I want to install Linux on the least :(
22:33hh444: yo! any kmscon fans here? (terminal emu)
22:34imirkin_: (or if they are, they keep quiet about it)
22:35hh444: patience is a virtue
22:36imirkin_: don't i know it
22:37hh444: ever used dos back in the day?
22:37imirkin_: that's what i learned programming on...
22:37hh444: awesome that you had to use cli just to play
22:37imirkin_: tasm, turboc, etc.
22:37hh444: o snap!
22:38imirkin_: 16-bit real mode is kind of an ideal learning environment
22:38imirkin_: nowadays you have all this silly protection
22:38imirkin_: can't do anything without asking permission
22:38hh444: cool, i played around with dosbox today and its really learnful, get a game running
22:57Lyude: tracked down the PTE fault issue on kepler to # first bad commit: [b88fa004e8a349eba38a3df7a3ace9dd1730b6e0] dma-buf: fix reservation_object_wait_timeout_rcu to wait correctly v2
22:58Lyude: which partially eplaisn why things aren't getting fenced properly