04:08karolherbst: TimurTabi: Wiring up compute shouldn't be too hard on the GA100. Just needs NVK to skip setting up any 3D stuff on a GA100 (is this the case for _all_ GA100?) and then zink needs to respect "PIPE_CONTEXT_COMPUTE_ONLY" to skip initializing any 3D bits...
10:22mohamexiety[d]: GA100 should still have 3D hardware :thonk:
10:24mohamexiety[d]: Even GH200/100 does, albeit restricted to only two TPCs iirc. It’s mainly GB200 which just completely removed all the 3D stuff
20:15sonicadvance1[d]: The joke when a massive rack cluster is on stage at GTC and Jensen exclaims, "Can it run Crysis." and it doesn't have any rasterization units :BlobSweat:
21:51mohamexiety[d]: software rendering, time to use all that compute :evil_gears:
22:10phomes_[d]: new hud version:
22:10phomes_[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1472354446439612426/Screenshot_From_2026-02-14_22-14-16.png?ex=6992446f&is=6990f2ef&hm=310397f1c76e4d121b9024d5e8c821e98a2b51c588e6fe602c38a4c03dd66658&
22:51asch: Hello. I'm using a RTX 2060 GPU on a musl-based Linux system and no vulkan apps seem to be working since at least December. They all segfault.
22:53asch: I tried multiple versions of mesa, cbindgen, glslang and vulkanloader, vulkan-headers, spirv-loaders, spirv-headers etc. but I'm not managing to solve this.
22:53mhenning[d]: Can you get a stack trace from gdb?
22:54asch: I could try getting from lldb, would that be good enough?
22:54mhenning[d]: sure
22:55asch: I'm not really a pro at backtracing, if you could help me, I would be glad. I will just load vulkaninfo, tell it to continue and paste the output.
22:56asch: https://pastebin.com/ujzthjyu
22:57asch: I'm also using a custom kernel with nouveau and firmware files builtin. This happens if I use nouveau as a separate module as well.
22:58mhenning[d]: If you run `bt` in lldb after the crash, what does it print?
22:59asch: (lldb) bt
22:59asch: * thread #1, name = 'vulkaninfo', stop reason = signal SIGSEGV: address not mapped to object (fault address=0x0) * frame #0: 0x00007ffff7c7dcf0 libvulkan_nouveau.so
23:00mhenning[d]: Oh, that's the entire stack?
23:00asch: Looks like so
23:00asch: any vulkan app segfaults immediately
23:00asch: even vulkaninfo
23:01mhenning[d]: Does your distro provide debug symbols?
23:01asch: I'm on Gentoo, ussing Clang, is it enough to just recompile mesa/vulkaninfo with debug flags?
23:02asch: btw, dmesg says the following: [ 2440.279729] vulkaninfo[6569]: segfault at 0 ip 000074878be7dcf0 sp 00007ffe3d6739d0 error 4 in libvulkan_nouveau.so[c7ccf0,74878b9ac000+504000] likely on CPU 0 (core 0, socket 0)
23:02mhenning[d]: yes
23:02asch: and [ 2440.279739] Code: 8b 05 c4 92 09 00 eb 22 66 90 4a 8b 34 30 ba 08 00 00 00 48 89 df 48 83 c6 08 e8 df 87 02 00 48 8b 05 a4 92 09 00 49 83 c6 10 <4a> 8b 4c 30 f8 48 83 f9 19 74 d5 48 85 c9 75 ec 48 8b 44 24 08 48
23:03mhenning[d]: specifically, it's the mesa debug symbols that would be useful here
23:03asch: What would be the best debug compilation options? Would COMMON_FLAGS="-O0 -march=native -gfull -gdwarf -fno-omit-frame-pointer -pipe" suffice?
23:04asch: mesa takes about 5 mins to rebuild, vulkan-tools not as much
23:05mhenning[d]: yeah, that should be good
23:05mhenning[d]: You can also turn on assertions in mesa, although I suspect that won't matter here
23:06mhenning[d]: assertions are enabled by passing `-D b_ndebug=false` to meson
23:10asch: vulkan-tools is already compiled with those flags. I'm on gentoo and now rebuilding mesa with those flags and the debug USE enabled, that should probably do just this.
23:11asch: I'm however downgrading to 25.2.7 from 25.2.8 becauses that's the version in the Alpine Linux tree
23:13mhenning[d]: Oh, 25.2 is a bit old we only support 25.3 or 26.0 upstream at this point
23:19asch: Yes, but everything segfaults immediately with them, not only vulkan apps
23:19asch: might be because of Zink
23:20asch: I mean, it happens with every higher version of mesa I tested
23:23asch: Now I get this: https://pastebin.com/5PSfKQDr
23:24asch: I'm on irssi so we could try a newer version of mesa on tty if you so like. Are newer versions of mesa also loading different firmware files?
23:37HdkR: When crashing in malloc. Spicy.
23:38asch: I noticed I'm missing debugedit here. I think I will recompile again with debugedit present.
23:41karolherbst: crashing in malloc usually means something wrote somewhere where it shouldn't
23:46asch: I read somewhere that musl uses different stack sizes and it might not be enough for mesa. Does this make sense?
23:46asch: They amended the issue with an extra LDFLAG, but that didn't seem to work here.