00:21 gfxstrand[d]: I'm still getting a bunch of misaligned address exceptions
00:21 gfxstrand[d]: No idea where they're coming from or how to avoid them
01:45 swee: update: mesa-vulkan-nouveau got added on alpine!! \o/
01:45 gfxstrand[d]: \o/
01:46 swee: https://pkgs.alpinelinux.org/package/edge/main/x86_64/mesa-vulkan-nouveau
01:46 gfxstrand[d]: And I just assigned Kepler images to Marge.
01:49 gfxstrand[d]: \o/
01:51 swee: https://p.swee.codes/swee/bf2072f8960a462cb651ae74dbef9650
01:51 swee: oh and turbowarp app now works!!
01:53 swee: using Nouveau GL works, but using Zink will cause segfault
02:48 swee: https://p.swee.codes/swee/d58b5bd2d3014b9293937bc0288c3419
03:51 HdkR: :thonk:
04:01 mangodev[d]: swee: hmmm
04:01 mangodev[d]: does running without kopper fix it?
04:25 swee: kopper?
08:26 farengrubnils: So i share the final results: 173-122=51 where this was the last index 8th with four digits gaps, so 24th element altogether. 173 comes as 132+8+26=166/2=83 256-83=173 we had compile time 118 to be lifted as bigger of the two, so compiler generates 118-104-52-76 equalling to buddy of 114/2=57, now 118+57-120+61(61 comes from 232-120-122-51=-61 51 comes as 173-122=51)=116 173-116=57 and
08:26 farengrubnils: rest of it as before shown by sensei. So first two elements per logics are done hence like so: 192-110-112-86=-116, samewise 200-114-112-79=-105 hence 198-161=37 and 193-152=41 where buddies were 74/2 and 82/2 which themselves came as 118-144-72+24=-74 118-136-68+4=-82 hence 82+118=200 where 74+118=192, so 86 came as 198-112 and 79 as 193-114=79 hence the calculations before sensei
08:26 farengrubnils: filtering are 118+37-110+116=161 198-161=37 similarly 118+41-112+105=152 193-152=41....All indexes pulled in the same way to get the value as return. Hence i said your mentally ill wanker stories are your delusion and self-defense mechanism to justify your failures in life, why you spam that is i do not know why, but you will be treated with force there soon.
09:00 snowycoder[d]: gfxstrand[d]: with great sadness I shall inform you that `iset` only sets the register to either 0 or -1 (or, with the .bf modifier, 1.0 f32).
09:00 snowycoder[d]: Even CUDA to set 0 or 1 applies a negation to iset
09:16 kingparrot: snowycoder[d]: very sad news is that you are entirely illborn folks so to speak. You are also as dead as possible. I saw many of such abortion leftovers in life. Trying to take over my and my fathers hard earned assets. You are in bad history column as finished soon. Your fucking gangsta mob and anuses we crush them soon in a single day entirely. All my people come together to do that with
09:16 kingparrot: honor and proud and justice will be served. What i think the doctors are landed missiles into their asses, all this conspiracy ends with my victory i am not about to listen your dick science at all.
09:18 HdkR: 🎉
12:33 gfxstrand[d]: snowycoder[d]: I can't say I'm surprised. That's the D3D 32-bit true.
12:47 karolherbst[d]: mhh if one wants some perf tuning, xcom2 runs slower on my ampere GPU (high-end) than on my intel igpu (xe though) 🙃
12:52 gfxstrand[d]: 🙃
12:59 snowycoder[d]: gfxstrand[d]: Should I wire it up anyways? It seems pretty useless unless we have some kind of very special compiler magic.
13:11 gfxstrand[d]: <a:shrug_anim:1096500513106841673>
13:12 gfxstrand[d]: Probably no point maybe stick it in a branch in case we ever want it so we don't have to retype the tests.
13:13 gfxstrand[d]: There are probably still cases where, even after all NIR's attempts to get rid of them, some D3D booleans exist. But IDK that we care that much.
13:14 gfxstrand[d]: Honestly, the only reason I wired up fset was for fsign.
13:14 snowycoder[d]: Makes sense
13:24 zmike[d]: gfxstrand[d]: do you plan to look at that wsi issue anytime soon?
13:44 gfxstrand[d]: Which one is that?
13:54 zmike[d]: https://gitlab.freedesktop.org/mesa/mesa/-/issues/11901 https://gitlab.freedesktop.org/mesa/mesa/-/issues/13173
13:55 karolherbst[d]: gfxstrand[d]: I suspect it's PCIe overhead
13:55 karolherbst[d]: maybe
14:01 kar1m0[d]: I am trying to uninstall nvidia drivers and it freezes my pc completely
14:02 kar1m0[d]: I did install nouveau though
14:02 kar1m0[d]: It can't switch to nouveau drivers for some reason
14:02 snowycoder[d]: After what command does the pc freeze? The switch should only happen after a reboot
14:03 kar1m0[d]: When I remove nvidia drivers
14:03 kar1m0[d]: Completely
14:04 kar1m0[d]: Sudo pacman -Rns nvidia-utils nvidia-open-dkms
14:06 snowycoder[d]: Ohh, dkms, ok yep that will cause trouble.
14:06 snowycoder[d]: If nouveau is installed then it should still work after a reboot
14:07 kar1m0[d]: yes but how do I switch between the drivers if I can't uninstall dkms
14:28 gfxstrand[d]: zmike[d]: Ugh... Right...
14:37 gfxstrand[d]: Anybody wanna ack https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30403 ?
14:37 gfxstrand[d]: Note we haven't enabled Kepler by default yet so we still have time to fix the last few bugs.
14:48 kar1m0[d]: snowycoder[d]: yeah I probably should have chosen the open source drivers version...
14:49 kar1m0[d]: seems like I am stuck with nvidia dkms until I reinstall the whole os
14:55 kar1m0[d]: although I am not sure if I installed everything for nouveau
15:36 snowycoder[d]: Might be a stupid question, but what does `r-b` mean? It's not mentioned in mesa docs.
15:40 pendingchaos: snowycoder[d]: short form of Reviewed-by: https://docs.mesa3d.org/submittingpatches.html#reviewing-patches
15:41 gfxstrand[d]: snowycoder[d]: "Reviewed-by"
16:04 snowycoder[d]: Thanks!
16:50 gfxstrand[d]: Okay, I've avoided it long enough. Time to figure out why message passing tests are failing
16:54 gfxstrand[d]: gfxstrand[d]: I feel like that deserves a blog post...
16:56 gfxstrand[d]: Nah. I should have merged it like 6 months ago. Users don't care. Assigned to Marge.
17:02 karolherbst[d]: heh
19:07 karolherbst[d]: Is the copy in NAK supposed to be a scalar copy or can it also copy vectors?
19:07 karolherbst[d]: because apparently nothing in the code prevents it from copy from a vector to a scalar
19:31 gfxstrand[d]: It's a scalar
19:31 gfxstrand[d]: What does copying from a vector to a scalar even mean?
19:34 gfxstrand[d]: I might have Kepler B passing the CTS
19:35 gfxstrand[d]: I still don't know where these misaligned addresses are coming from
19:48 marysaka[d]: gfxstrand[d]: you are excluding the waivers for Kepler btw? not sure if that would be related
19:50 marysaka[d]: <https://github.com/KhronosGroup/VK-GL-CTS/blob/main/external/vulkancts/mustpass/main/waivers.xml#L236>
19:50 gfxstrand[d]: I think those get automatically included
19:51 gfxstrand[d]: But the misaligned address fails are different
19:51 marysaka[d]: I see okay
19:56 karolherbst[d]: gfxstrand[d]: nothing, I just saw nak doing that, so I was wondering if it was supposed to copy a vector to a vector, or scalar to scalar
20:18 gfxstrand[d]: If NAK is doing that, it's a bug.
20:19 gfxstrand[d]: I feel like I should do different CTS submissions for Kepler A and B. They're different compiler back-ends after all.
20:21 gfxstrand[d]: gfxstrand[d]: The misaligned address warns are from a set of CTS tests I asked for 6 years ago. 😂 They test Image load/store with the wrong format in the shader to make sure it reads/writes junk but doesn't wedge. If the formats have different sizes, the result on Kepler is an address calculation to the right pixel in memory but a load that might be too big. The errors are harmless.
20:22 gfxstrand[d]: Unfortunately, without access to the ESP report mask register, I can't shut them off.
20:23 gfxstrand[d]: The OOR_ADDR errors are simiarly already ignored on Maxwell+
20:55 gfxstrand[d]: `Pass: 529326, Fail: 1, Skip: 832665, Flake: 8, Duration: 38:02, Remaining: 38:35`
20:56 gfxstrand[d]: I'm hoping that fail is a timeout
21:12 karolherbst[d]: I mean.. I'm probably doing weird things to NAK, so it ends up doing that 🙃
21:13 karolherbst[d]: gfxstrand[d]: might be the same bit in a different register
21:13 karolherbst[d]: we do have the reg names
21:18 gfxstrand[d]: The problem, I thought, was that we don't have a mechanism to bang on those from userspace. We don't have FALCON pre-Maxwell
21:20 gfxstrand[d]: We could revive the SW method stuff but ugh...
21:25 gfxstrand[d]: At the time we were mucking about with all that we were like, "Eh, we'll get to Kepler when we get to it." Oops. 😳
21:31 asdqueerfromeu[d]: marysaka[d]: I wonder if this is a hardware-level issue
21:32 gfxstrand[d]: It is
21:32 gfxstrand[d]: Well, there are details. mhenning[d] and I are arguing about some of them on GitLab.
21:41 karolherbst[d]: gfxstrand[d]: we can set it unconditionally for every context
21:42 karolherbst[d]: the kernel creates a template context (called golden context or something) that every new context is derived from, and the kernel can just set up a state that every new context will inherit
21:42 karolherbst[d]: I think I have a patch somewhere..
21:43 karolherbst[d]: gfxstrand[d]: https://gitlab.freedesktop.org/karolherbst/nouveau/-/commit/10a2421f1e44b2ff6050a2c94c9622410c47d3f8
21:46 karolherbst[d]: upstream patch: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=1cb9e2ef66d53b020842b18762e30d0eb4384de8
21:46 karolherbst[d]: just need to do something similar for other bits
22:03 asdqueerfromeu[d]: And I assume this was the NVK part?: <https://gitlab.freedesktop.org/nouveau/mesa/-/merge_requests/203>
22:14 karolherbst[d]: that's for GPUs where we can use the firmware to do it for us
23:04 snowycoder[d]: What's falcon?
23:15 mhenning[d]: falcon is an architecture used for different microcontrollers on nvidia gpus
23:15 mhenning[d]: falcon functions have been moving to riscv more recently
23:16 mhenning[d]: but on eg. kepler there are a few falcon cores that do a grab bag of different things
23:17 mhenning[d]: https://envytools.readthedocs.io/en/latest/hw/falcon/intro.html has some more info
23:19 gfxstrand[d]: mhenning[d]: Do we know CUDA does for shared on Kepler?
23:20 mhenning[d]: not off the top of my head
23:22 mhenning[d]: Sounds like it's partitioned out of L1 like usual https://docs.nvidia.com/cuda/archive/10.2/kepler-tuning-guide/index.html#shared-memory-capacity
23:23 mhenning[d]: I'd be surprised if we could actually select different cache settings for shared in any meaningful way
23:23 mhenning[d]: but I don't know what bits actually get set