00:41 rhyskidd: mwk: thanks for the envytools review
01:06 Llmiseyhaa: So looks like the my GT730 likes to sit at ~46C idle.
01:08 rhyskidd: Llmiseyhaa: sounds reasonable
01:13 Llmiseyhaa: Yeah, that's what I thought.
09:03 paulk-gagarine: hey there, I see that you have updated the tasks lists for nouveau -- I'd like to suggest working on the GK20A firmwares so that the whole graphics pipeline can run with free software
09:04 paulk-gagarine: (which I care about a lot, as I'm very interested in running the Tegra K1, either on the Jetson or Chromebooks, with free software)
09:15 karolherbst: paulk-gagarine: yeah, you can do that. I am not entirely sure what is wrong with our kepler firmware on the tk1
09:37 paulk-gagarine: karolherbst, yeah I'd really like to figure it out, but never had the time to
09:37 paulk-gagarine: karolherbst, should I submit the idea anywhere in particular?
10:37 karolherbst: paulk-gagarine: well it could be a more general task, because we also use close firmware for videl acceleration on most boards
10:37 karolherbst: I keep that in mind
12:12 NanoSector: so, last revision is building now
12:12 NanoSector: it's given me way more good builds than bad ones though so i'm not convinced this is reliable
12:19 NanoSector: karolherbst, 0be75179df5e20306528800fc7c6a504b12b97db is the first bad commit
12:19 NanoSector: https://github.com/torvalds/linux/commit/0be75179df5e20306528800fc7c6a504b12b97db
12:28 NanoSector:compiles that and tests
12:32 NanoSector: yeah but that commit is fine as well :\
12:34 NanoSector:tries a bisect on drivers/gpu/drm/nouveau between 4.11 stable and 4.12 stable
12:57 karolherbst: NanoSector: :/
12:58 karolherbst: it might be always the case you did a mistake somewhere
12:58 karolherbst: most likely your last bad is good as well
12:59 NanoSector: i don't think i made a mistake
13:00 NanoSector: but i do builds in 10 minutes now (make localmodconfig helps a lot)
13:00 NanoSector: although
13:00 imirkin: paulk-gagarine: it's quite likely that the nouveau firmware Just Works (tm) on the TK1. just on one's ever tried.
13:00 NanoSector: i did select latest master as bad, which includes 4.13 and everything
13:03 paulk-gagarine: imirkin, IIRC tagr and gnurou were saying that it doesn't really work well
13:03 imirkin: i don't think they ever properly tried it
13:03 imirkin: i think there's a wrinkle in that it needs some of the new GK208 logic but in the v3 isa. or vice-versa.
13:14 paulk-gagarine: ohh
13:14 paulk-gagarine: if it's just that, it looks rather doable
13:23 imirkin: for all i know it won't work
13:23 imirkin: but iirc the initial hurdle was a moderately simple one
14:07 NanoSector: karolherbst, this next bisect points to 2579b8b0ece53248b815042f8662a4531acf120d
14:08 NanoSector: https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/commit/?id=2579b8b0ece53248b815042f8662a4531acf120d
14:08 NanoSector: which sounds illogical
14:09 NanoSector: or does it
14:09 NanoSector: hmm
14:10 NanoSector: well i'll be damned, now it works on kernel 4.13
14:10 NanoSector:flips a table
14:17 NanoSector: so running dolphin-emu does crash it on 4.13, but glxgears suddenly starts to work
14:17 NanoSector: managed to capture a dmesg: https://hastebin.com/efiyafobim.vbs
14:50 NanoSector: oh the dmesg got cut off, welp
14:51 NanoSector: here's the full thing: https://ptpb.pw/CzBH
14:52 NanoSector: so gk104_fifo_runlist_commit is in the stack traces of that dmesg
14:54 NanoSector:undoes the patch and tests
14:59 NanoSector: yeah it still hangs then. so unless anyone has other ideas i'm not sure i can be of more use here, sorry
15:23 karolherbst: :/
17:02 karolherbst: NanoSector: yeah, sometimes some effects aren't really connected to the issue so the wrong thing is actually bisected :/
17:04 NanoSector: :\
17:04 NanoSector: well, we tried
17:05 karolherbst: is GT200 third gen Tesla as well?
17:05 imirkin_: how do you count gens?
17:06 karolherbst: no clue
17:06 imirkin_: it's kinda in between G96, G98, and MCP77
17:06 imirkin_: but also has fp64 :)
17:06 karolherbst: I want to write down which GPUs have PMUs where we can upload our code to
17:07 karolherbst: in a way a random interested GSoC student would understand it
17:07 imirkin_: ooof....
17:07 karolherbst: ;)
17:07 imirkin_: i honestly don't remember
17:07 imirkin_: does G98 have such a PMU?
17:07 karolherbst: well GT215 is the first
17:07 imirkin_: if not, then there's no chance
17:07 karolherbst: imirkin_: NVIDIA GT21x, MCP77-MCP7A, MCP89, GFxxx, GKxxx or GM1xx chip ;)
17:08 imirkin_: wait, MCP77/79 have it?
17:08 karolherbst: I am sure my nvac has it
17:08 imirkin_: huh
17:08 karolherbst: or not? let me check
17:09 karolherbst: maybe only mcp89
17:09 imirkin_: mcp89 definitely does
17:09 karolherbst: okay, nouveau doesn't enable the PMU for MCP77-7a
17:10 karolherbst: well, then I would write "NVIDIA GT21x, MCP89, GFxxx, GKxxx or GM1xx chip" in the project idea
17:10 karolherbst: should be good enough
17:10 karolherbst: mhh
17:10 imirkin_: https://github.com/skeggsb/nouveau/tree/master/drm/nouveau/nvkm/subdev/pmu/fuc
17:10 karolherbst: tegra doesn't have i... argh
17:11 imirkin_: of course not ;)
17:11 karolherbst: oh wel
17:11 karolherbst: l
17:24 karolherbst: imirkin_: do you think it would be a feasable task for a student to implement access memory from system memory in draw calls for GF-GM GPUs in a GSoC project?
17:24 imirkin_: ?
17:25 imirkin_: that should already work
17:25 karolherbst: really?
17:25 imirkin_: perhaps i misunderstand the question
17:25 karolherbst: I thought for draw calls everything needs to be moved to vram?
17:25 imirkin_: i mean ... the way the code is written yeah, but just flip the vram_domain to gart, and you'll have it in gart.
17:26 skeggsb: it's... complicated... ;)
17:26 imirkin_: [this is done on K1]
17:26 karolherbst: skeggsb: well, yeah, but I though on fermi and newer this should be quite possble
17:26 karolherbst: imirkin_: okay well yeah, but the task of the student would be to implement a way to do that on demand
17:26 skeggsb: in theory, you can pass both VRAM/GART, and you'll get wherever you manage to get.. but, the way the ttm code works means if you do that, you'll lose large pages
17:27 karolherbst: and maybe even implement that "preferably in vram, but sysmem would be fine as well, but slower" thing
17:27 skeggsb: which means no compression, etc
17:27 imirkin_: skeggsb: i thought with fermi you could get large pages? or you mean hw supports it but not ttm?
17:27 skeggsb: the way it's implemented, mapping stuff into the VM isn't allowed to fail
17:27 imirkin_: [i mean, large pages in sysmem]
17:27 skeggsb: which means you can't migrate between small<->large pages for a buffer
17:28 skeggsb: (because you have to allocate PTs, which can fail)
17:28 skeggsb:is working on it
17:28 karolherbst: well yeah true, but imirkin_ was under the impression it should be easier on fermi?
17:29 imirkin_: skeggsb: is there anything you're not working on?
17:29 imirkin_: world peace?
17:29 skeggsb: karolherbst: it's *possible* on fermi, just not the way our ttm implementation works
17:29 karolherbst: skeggsb: anyhow, the question is rather if it would be a feasable task to do for a student or which part could be done by one
17:29 karolherbst: I see
17:29 imirkin_: skeggsb: wait, but does fermi not support large pages in sysmem?
17:30 skeggsb: you can't map 4K pages into 64/128K GPU pages :P
17:30 imirkin_: that's the CPU side of it
17:30 imirkin_: but from the GPU side it works OK, right?
17:30 skeggsb: if the CPU has a 64K page size, in theory, possibly
17:31 skeggsb: dunno if the GPU supports it or not
17:31 imirkin_: all you need is 64K or contiguous hw memory
17:31 skeggsb: i heard a hint that it doesn't somewhere
17:31 imirkin_: ah. so then that's fail, since you'd have to "resolve" the image to remove compression/etc
17:31 skeggsb: yes well, getting non-PAGE_SIZE blocks of memory gets really hard the longer the system is up... not really a feasible solution even if it does work
17:32 imirkin_: =/
17:32 imirkin_: so then wait... what does fermi do that G84 doesn't as far as this page migration stuff is concerned?
17:32 skeggsb: dual page tables
17:32 skeggsb: (4K and 64K page tables that cover the same virtual address range)
17:32 skeggsb: err.. 128K
17:32 skeggsb: (or 64K..)
17:32 skeggsb: it can be both, 128K by default
17:32 imirkin_: right
17:33 imirkin_: since configurable page sizes is definitely what all hw needs
17:33 skeggsb: pascal has 2MiB pages too!
17:33 skeggsb: ;)
17:33 imirkin_: and if possible, those page sizes shouldn't match what other hw vendors have
17:33 imirkin_: because then interop would be too simple
17:34 skeggsb: aaanyway, i've been having *loads* of fun in the horror of those codepaths lately, so yes, working on it actually
17:34 imirkin_: (x86_64 large pages are 1MB right?)
17:34 skeggsb: not a clue
17:34 imirkin_: or 4MB. but almsot definitely not 2MB
17:34 imirkin_: [and there are like 4GB huge pages]
17:34 skeggsb: i don't believe current non-SOC HW can use system memory large pages anyway, but i'm not sure where i heard that, or if it's even true
17:35 karolherbst: imirkin_: 4k, 2M, 1G
17:35 imirkin_: well it definitely doesn't sound like a good idea anyways... like you said, fragmentation becomes too big an issue.
17:35 imirkin_: oh. so then 2MB is a match. neat.
17:35 imirkin_: perhaps they *did* finally compare notes.
17:36 karolherbst: well yeah, what I am actually interested is, if you think we could make a project for a student out of this or if we should just deal with it ourselves
17:40 skeggsb: i don't think the DRM side of it is a big enough project, it basically amounts to lazily-mapping buffers into GPU address-space directly at submission time
17:41 skeggsb: as opposed to in the ttm move_notify() hook, which can't fail
17:41 skeggsb: making it sane on the userspace side, i can't comment on well enough atm...
17:42 karolherbst: skeggsb: okay, so it's most likely not much of a big deal, but would require a lot of knowledge about proper memory placement, so that we can implement good well enough scoring to not waste too much performance
17:42 skeggsb: in theory, userspace should send the kernel a list of "here are all the buffers this submission needs, and the places that are OK for it to be in for it" and the kernel deals with it
17:42 karolherbst: yeah, right
17:43 skeggsb: but, then you end up with hilarious situations like mesa doing system memory -> system memory blits with the cpu...
17:43 karolherbst: I already discuessed with imirkin_ that we could have for every buffer a "places where I can be", and "places where I _want_ to be"
17:44 skeggsb: *or* you don't have the kernel randomly place stuff, and somehow tell userspace to migrate shit itself, so it knows where its buffers are at any given moment
17:44 karolherbst: mhhh
17:44 skeggsb: (the kernel will still evict, but will put it back wherever userspace thought it was last)
17:44 NanoSector: karolherbst, i don't even think the bug i have is reliably reproducible, a while back I could open gnome-control-center no problem and now it doesn't work, on the very same kernel
17:44 karolherbst: skeggsb: maybe userspace could give a score for how bad it would be to be put into sys mem
17:45 karolherbst: and 1 means I don't care at all, and 100 means: super totally super bad
17:45 skeggsb: it can't really know that up front, i mean, how well do all our heuristics like that work out? ;)
17:45 karolherbst: but I doubt that's feasable at all
17:45 karolherbst: ;)
17:45 karolherbst: does it really matter if you run out of vram?
17:46 karolherbst: a bad heuristic is still better than messing up due to having no vram
17:46 skeggsb: that *should* already work, the kernel will kick buffers out
17:47 karolherbst: within one draw call?
17:47 skeggsb: yes, it'll go through the list of buffers, and hand out vram-only first, then gart-only, then whatever's left over goes to things marked as either
17:48 skeggsb: if you send a submission that has too many vram-only objects to fit at once, that's a userspace bug
17:48 NanoSector: speaking of gart, is it normal to have a gart of 1TB?
17:48 NanoSector: [ 2.562291] nouveau 0000:01:00.0: DRM: GART: 1048576 MiB
17:48 skeggsb: it's the size of the GPU address space
17:49 skeggsb: the message is a bit misleading, it comes from being AGP-only at one point
17:49 NanoSector: ah
17:50 karolherbst: skeggsb: I see
17:51 karolherbst: skeggsb: so it would make sense to adjust userspace to somehow say: here it matters or not, to not hurt perf too bad and don't mark a buffer as vram only, because of perf concerns
17:57 karolherbst: skeggsb: what do you think about reporting vram usage to userspace?
17:57 karolherbst: never looked into that at all and have no clue how much work that would be for nouveau
18:18 karolherbst: I've drafted some project ideas I came up with for GSoC/EVoC: https://gist.githubusercontent.com/karolherbst/f80890aad3983cd37a502888229d0978/raw/d66a66a99c993146be19fb4f4a1b505d01d2295f/gistfile1.txt
18:18 karolherbst: mupuf: ^
18:18 mupuf: cool!
18:20 karolherbst: I also added the three old ones for completness
18:21 karolherbst: Please tell me if some of those are good enough to be added to the Xorg wiki, so that I can simply add those there
18:21 imirkin_: skeggsb: my guess is we should look at what amdgpu does wrt vram/sysram placement, and copy that
18:23 karolherbst: and by the way: KHR-GL45.arrays_of_arrays_gl.SubroutineFunctionCalls1 shows a super annoying compiler bug regarding for loops
18:30 karolherbst: imirkin_: by any chance, do you know what is wrong with shaders@glsl-fs-lots-of-tex?
18:30 imirkin_: yeah. remove the test.
18:31 imirkin_: i thought it was gone finally
18:31 karolherbst: well, it passes on intel
18:31 karolherbst: that's why I am asking
18:31 imirkin_: both jekstrand and i sent patches to kill it
18:32 karolherbst: well, but it looks like we actually have a bug regarding that inside codegen or so
18:32 imirkin_: er no. i guess jekstrand sent a patch to adjust it. perhaps it was merged.
18:32 imirkin_: no.
18:32 imirkin_: the issue is a rounding one.
18:32 karolherbst: mhhh
18:32 karolherbst: 0.400000 vs 0.400000?
18:32 karolherbst: .-...
18:33 karolherbst: 0.400000 vs 0.266667
18:33 imirkin_: inside the texture evaluator
18:33 imirkin_: yep.
18:33 karolherbst: ohh, I see
18:33 imirkin_: the error vs what's expected gets magnified due to how the test is written.
18:33 karolherbst: mhh, I see
18:34 imirkin_: https://patchwork.freedesktop.org/patch/17718/
18:35 karolherbst: imirkin_: precise for the win maybe?
18:36 imirkin_: nope.
18:36 imirkin_: read the commit message.
18:36 karolherbst: ohh wait, there are simply adds
18:36 karolherbst: ohhhh
18:36 karolherbst: meh
18:36 karolherbst: k
18:37 imirkin_: jason changed it around a bit
18:37 imirkin_: but i don't think it addressed the source of the failure on nvidia
18:38 karolherbst: any knowledge about "spec@!opengl 1.0@gl-1.0-drawpixels-color-index" ?
18:38 imirkin_: yep
18:38 imirkin_: let that one go.
18:38 karolherbst: k
18:38 imirkin_: it's a st/mesa deficiency
18:39 imirkin_: apparently old-school GL supported indexed-color textures
18:39 karolherbst: we fail those amd_performance_monitor because of "Not enough free MP counter slots !"
18:39 imirkin_: this does not work well in combination with glDrawPixels()
18:39 karolherbst: ahh, I see
18:40 imirkin_: brian was nice enough to write the test, but not enough to fix the fail :)
18:40 karolherbst: the perf counter one?
18:40 imirkin_: the indexed color + drawpixels
18:40 karolherbst: k
18:42 karolherbst: "spec@arb_draw_indirect@arb_draw_indirect-draw-arrays-prim-restart" also kind of sounds unimportant
18:42 imirkin_: GL spec changed
18:43 karolherbst: k
18:43 imirkin_: 1-line fix in nouveau
18:43 imirkin_: https://cgit.freedesktop.org/mesa/mesa/tree/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c#n1013
18:43 imirkin_: change that to a "0" :)
18:43 karolherbst: fun
18:44 imirkin_: feel free to send that in if it works...
18:44 imirkin_: i've been neglecting that bit of it
18:44 imirkin_: coz i felt it was lame to change GL specs
18:44 karolherbst: :D
18:44 karolherbst: okay, so 0 would be the correct thing in the end?
18:44 imirkin_: for GL 4.5, iirc, yes
18:44 imirkin_: GL 4.4 is the old way
18:44 karolherbst: I see :/
18:45 imirkin_: like i said, "kinda lame"
18:45 karolherbst: "spec@arb_vertex_type_2_10_10_10_rev@draw-vertices-2101010" got my interest now
18:46 imirkin_: everyone fails that one.
18:47 karolherbst: nice...
18:47 imirkin_: pretty sure blob does too
18:48 skeggsb: so, er, what happens to clients that were written against 4.4 in that case?!
18:48 karolherbst: tough luck I guess
18:48 imirkin_: skeggsb: i thought i mentioned "kinda lame"
18:48 imirkin_: skeggsb: in practice, nobody cares about primitive restart with draw-arrays
18:48 karolherbst: but shouldn't we do it the old way on 4.4 contexts?
18:48 imirkin_: so ... it's just for conformance, and some hw didn't support it, so they just said "fuck it"
18:49 skeggsb: heh
18:49 imirkin_: although why nvidia made it configurable is a little beyond me
18:50 karolherbst: oh wow, like 70% of our piglit fails is due to multisampling
18:50 imirkin_: but it works out well now.
18:50 imirkin_: karolherbst: those ext_multisample_framebuffer tests?
18:50 karolherbst: not only those
18:50 imirkin_: good times...
18:50 karolherbst: spec.!opengl1.1 as well
18:50 imirkin_: i picked away at them over time, used to be worse =]
18:50 imirkin_: yeah, but those are the resolve issue
18:51 karolherbst: arb_texture_rg as well
18:51 imirkin_: ?
18:51 karolherbst: multisample-formats X gl_arb_texture_rg-int
18:51 imirkin_: hm
18:51 imirkin_: i don't remember off the top of my head
18:51 karolherbst: ext_framebuffer_multisample_blit_scaled in addition to ext_framebuffer_multisample
18:51 karolherbst: ext_texture_integer
18:52 karolherbst: for fun reason the latter has tests like "multisample-formats X gl_ext_texture_integer"
18:52 karolherbst: so ext_texture_integer == arb_texture_rg?
18:52 karolherbst: maybe
18:52 imirkin_: certainly rg-int ;)
18:53 karolherbst: as far as I can tell, every of those fails sound like resolv issues on glReadPixels
18:54 karolherbst: maybe not those multisample-format fails
18:54 karolherbst: https://mini.karolherbst.de/nouveau/piglit/nve6-cts/nve6-cts-0393cd92c5e121d5f93ac227dcc7638070f77443/spec@ext_texture_integer@multisample-formats%202%20gl_ext_texture_integer.html
18:55 mupuf: karolherbst: I updated the frameretracer and the dynamic reclocking one: https://pastebin.com/HNf4vVFJ
18:55 mupuf: you will never learn that experimenting in the PMU is a complete waste of time :p
18:56 karolherbst: mupuf: thanks
18:56 karolherbst: nope
18:56 karolherbst: I won't :p
18:57 mupuf: C is already a pain for signal processing, so asking people to do it in a foreign ASM is just suicidal
18:57 karolherbst: ohhh right, we can't do it on fermi either
18:57 karolherbst: mupuf: hey.... I was only thinking about adding the counter readout in asm, which I basically already wrote
18:57 karolherbst: ;)
18:57 mupuf: did I write the code for it already?
18:57 karolherbst: and those can be read out in userspace
18:57 mupuf: oh yeah, you made another one
18:58 mupuf: yes!
18:58 karolherbst: yeah
18:58 mupuf: this is something we need to do!
18:58 karolherbst: yeah
18:58 karolherbst: I already sent patches
18:58 mupuf: I don't care where the code comes from
18:58 karolherbst: it is even configureable from the host
18:58 mupuf: but we need to expose that to the userspace to enable the reclock experiments
18:58 karolherbst: mupuf: https://github.com/karolherbst/nouveau/commits/pmu_counters_v2
18:59 mupuf: shower time first
18:59 mupuf: then I will have a look at the nv4x fan issue that has been reported to me
18:59 mupuf: and then back again to the one on nvc1
18:59 mupuf: after a year without touching this, I have forgotten some aspects of it...
18:59 mupuf: and I will have to re-do experiments
19:00 mupuf: at least, I have a good model, so a couple of sample points will be enough
19:01 mupuf: I had forgotten so much that I did not even recognize my code in nvbios...
19:04 karolherbst: :O
19:08 karolherbst: imirkin_: we also fail some atomic tests
19:11 karolherbst: and that subtest argument is totaly broken
19:16 karolherbst: nice! one of those multisample fails shows a difference in apitrace, nice nice
19:34 karolherbst: imirkin_: do you know a fast hack to clear allocated buffers with 0s?
19:34 imirkin_: ->clear_buffer ?
19:35 karolherbst: well, I was more talking about adjusting just one place
19:35 karolherbst: like in the kernel or something
19:35 karolherbst: or did you mean inside libdrm?
19:36 imirkin_: i guess i'm not sure what the question is
19:37 karolherbst: well, I have the issue, that in qapitrace I see random content in surfaces and it would be cool if everything would be 0 by default
19:38 karolherbst: but I already kind of know what is goind on
19:38 karolherbst: anyway
19:39 imirkin_: ah.
19:39 imirkin_: should be easy to do in mesa
19:39 karolherbst: intel: https://i.imgur.com/UvJ6LCN.png nouveau: https://i.imgur.com/2cjE8FJ.png
19:40 imirkin_: yeah
19:40 imirkin_: i remember those ;)
19:40 imirkin_: good luck!
19:40 karolherbst: :D
19:40 imirkin_: iirc i spent some time randomly banging on the keyboard
19:40 imirkin_: but if i fixed some, others then broke
19:41 imirkin_: i think i may have a better understanding of the pipeline now though, and perhaps could give it another shot
19:41 karolherbst: I think we just have multisample disabled on the blit
19:41 karolherbst: and that's why it'smessed up
19:41 imirkin_: we fake it though
19:41 imirkin_: but the texture descriptor's MS thing isn't handled properly?
19:41 eliteqnvgb: p
19:42 karolherbst: I already had the code for that and threw it away...
19:49 karolherbst: imirkin_: is there a way to read the GL_DRAW_FRAMEBUFFER in qapitrace?
19:50 imirkin_: if it's not in the framebuffers view, then no
19:50 karolherbst: :/
19:50 karolherbst: okay
19:50 karolherbst: there are several glDrawArrays(mode = GL_TRIANGLES, first = 0, count = 3) calls drawing into that framebuffer
19:50 karolherbst: and then it's blit and the result is those pngs
19:52 karolherbst: would be cool to know if the blit goes wrong or if something is odd with the draws already
19:52 karolherbst: but
19:52 karolherbst: the blit is for 512 to 256
19:52 karolherbst: and those buffers are just 256x256
19:53 imirkin_: we do the blit as if the surfaces were not multi-sampled
19:53 karolherbst: okay
19:53 karolherbst: can the blit stretch the image?
19:54 imirkin_: there's a blit_scaled for msaa
19:54 imirkin_: [extension]
19:54 karolherbst: ahh
19:54 karolherbst: but I could just check how the src and dest is setup for the blit
19:56 karolherbst: uhhhhhhhh, buh buh buh
19:59 karolherbst: :/
20:00 karolherbst: imirkin_: you will totally love that
20:01 karolherbst: okay, now it is pixel perfect in apitrace
20:01 karolherbst: but we still fail the test, because glReadPixels still messes up
20:02 karolherbst: imirkin_: https://gist.github.com/karolherbst/546d62a8316aa9c3cdcf3d933685109b
20:02 imirkin_: lol
20:02 karolherbst: ;)
20:02 imirkin_: yeah, that one passes. now try the other ones.
20:04 karolherbst: imirkin_: the msaa 8 test of this also looks fine
20:05 karolherbst: but... the painful part is, the piglit test fails, becuase glReadPixels :( super annoying
20:07 karolherbst: it used the 3d blitter for this test
20:12 karolherbst: imirkin_: ohhh wait, I think I know what is going on
20:14 karolherbst: or not
20:15 imirkin_: yeah, so part of the trick is that the 2d blitter is sometimes used, other times the 3d blitter is used
20:15 imirkin_: loads of fun.
20:16 imirkin_: 3d blitter does texture lookups
20:16 imirkin_: on MSAA textures, but attempts to do it in a non-MSAA way
20:16 imirkin_: it's a bit of a hack ;)
20:16 imirkin_: should probably be adapted to use TXF instead of TEX
20:16 imirkin_: (or does it use TXF already?)
20:17 imirkin_: although of course that sucks for scaling
20:17 imirkin_: since you don't exactly want to implement filtering with TXF...
20:17 karolherbst: it seems to simply use tex 2D
20:19 imirkin_: you could break down and just use the u_blitter junk
20:19 imirkin_: which handles all these various things
20:19 imirkin_: i've resisted
20:19 karolherbst: mhh
20:19 karolherbst: well
20:20 karolherbst: we still need to fix glReadPixels first anyhow, which seems to be actually more annoying then to fix those blits
20:20 karolherbst: *than
20:20 karolherbst: sadly
20:24 karolherbst: imirkin_: also, what is so bad about the u_blitter stuff?
20:25 imirkin_: it's wasteful
20:26 karolherbst: okay well sure, but if it would work better than what we have now?
20:27 imirkin_: the path to hell is paved with good intentions
20:27 karolherbst: ahh
20:27 imirkin_: i'm a fan of the "leave it broken until it's properly fixed" approach rather than the "do a stupid thing to just fix it now because something better is coming Real Soon" approach
20:28 karolherbst: imirkin_: k, I see
20:28 imirkin_: since putting in crutches removes any impetus to properly resolve the issue
20:28 karolherbst: understandable
20:31 airlied: imirkin_: wasteful how?
20:31 imirkin_: airlied: doesn't use the 2D blitter
20:31 imirkin_: i suppose we could use it on the "3d blit" path only
20:32 karolherbst: mhh, then we end up always using the 3d blit, because it works and the 2d one doesn't
20:32 imirkin_: at which point ... well ... we cheat, whereas it can't cheat (since it's generic)
20:32 imirkin_: we make use of some things which are a bit nvidia-specific
20:32 imirkin_: karolherbst: actually 2d blit is the one that tends to work :)
20:33 karolherbst: ahh I see
20:33 imirkin_: anyways, u_blitter has a proper resolve... which is probably good
20:33 imirkin_: whereas we just do a scaled blit and hope for the best, which tends to work out nicely
20:33 imirkin_: except in some cases, apparently ;)
20:34 airlied: do you need shader based resolve?
20:34 karolherbst: airlied: glReadPixels doesn't return the right values
20:34 imirkin_: maybe?
20:34 imirkin_: karolherbst: that's entirely unrelated.
20:34 karolherbst: right, true
20:34 airlied: imirkin_: so 2d blitter can do resolve?
20:34 imirkin_: airlied: it can do scaled blits. and the samples are laid out in a grid.
20:35 imirkin_: so ... not perfect :)
20:35 imirkin_: for a high-quality resolve, we definitely need a shader.
20:35 airlied: does it so integer properly?
20:35 imirkin_: almost certainly not
20:36 airlied: since integer resoollve is pick first sample
20:36 imirkin_: eyah
20:36 imirkin_: i don't remember if 2d blit does integer at all tbh
20:37 airlied: radv uses a compute shader if the hw resolve doesnt work
21:57 terminestic: what should had been a fascinating experience and easy victory, you have turned into muddy yet continuing conspiracy, and all against me again and guess what of course i am struggling there one way or another, cause i have to argue to breath
22:01 terminestic: though i have won most or all the argues , it is still round about wasted time
22:01 NanoSector: ehwat
22:04 gnarface: some sort of fill bot i think
22:04 gnarface: i saw earlier parts of the rant a day or two ago
22:05 gnarface: dunno what the point is, but it always starts and ends mid-stream... just some contextless rant in broken english
22:05 gnarface: i was trying to figure out if it's auto-generated or copied from something else
22:06 gnarface: maybe some sort of AI test, or maybe just someone trying to compare how hard it is to get banned in different channels
22:07 gnarface: interesting. maybe the point is to advertise for this? ^
22:08 gnarface: anyway, i hope it's not what it's intended to look like at face value; a desperate anonymous plea for emotional support, because it's hopelessly inept as such.
22:13 imirkin_: pretty sure it's just a guy in estonia with schizophrenia.
22:15 oughttobe: this how it has been , every time they have come against bigger number in the business, allthough not faced neither comprehenvicely regognized, the hallucinating crap has been beaten hard, with usually absolutely no points scored against me, but unfortunently yet still acting as big conquerors
22:34 gnarface: interesting, he's also msging me from another connection now. he seems genuinely upset about this. still no specifics though.
22:34 gnarface: (still skeptical he's real)
22:35 gnarface: i sure know what it's like to be persecuted unfairly but i think perhaps the word "nouveau" must have some other context in his native language...
22:35 airlied: he's real, just ignore him