02:36 Lynne: bnieuwenhuizen: you wrote the gdeflate decoder, right?
02:36 bnieuwenhuizen: Context?
02:37 Lynne: https://github.com/BNieuwenhuizen/gdeflate-benchmark/blob/main/third_party/gdeflate-shader/gdeflate.comp
02:37 Lynne: or at least the glsl version
02:37 Lynne: I'm wondering if it uses the resynchronization property of huffman streams or does it use restart markers?
02:39 bnieuwenhuizen: I have no idea, I don't think I ever tried to understand the algorithm
02:40 bnieuwenhuizen: Anyhow can check tomorrow but I'm not a compression expert by trade
02:51 Lynne: normally there would be a spec, but this is the best we get: https://github.com/microsoft/DirectStorage/blob/main/GDeflate/GDeflate/README.md
02:52 Lynne: apparently not, its just slicing across 32 separate deflate streams, lame
02:56 Lynne: the idea is that even if you start at an arbitrary offset, codes that comply with the kraft–mcmillan inequality (like huffman) will resync
02:57 Lynne: as long as the alphabet stays the same for about ~128 bytes
05:26 MoeIcenowy: karolherbst: therotically they should add a verison in the soname
08:41 jannau: karolherbst: there is https://bugzilla.redhat.com/show_bug.cgi?id=2424159 in tghe case you haven't seen it already
13:48 karolherbst: jannau: yeah, I've asked on the darktable bug tracker to file a fedora bug, I assume they are the same person
13:59 airlied: karolherbst: static link it :-), not sure how in fedora to force downstream rebuilds on updates
14:02 karolherbst: heh... that would also be an option
14:03 karolherbst: but then fedora has this "everything should be shared lniked" policy and then package maintainers have to ignore upstream doing static linking and argue why for why it should be indeed statically linked :P
14:09 kwizart: I used to have a c++ package (live555) with very bad ABI stability, I used to enforce a strict version lock by dropping a live555_version macro that using packages can have Requires = %{?live555_version}
14:46 MoeIcenowy: karolherbst: well, I think no ABI stability is a good reason for static linking
16:03 karolherbst: airlied: oh yeah.. I think you can do it via adding a so version to spirv-tools (at least the linker) and then depend on it
16:03 karolherbst: or I think the fedora tooling does it on its own
16:03 karolherbst: because the dependency is on the so name
16:05 kwizart: if you can do symbol versioning, the RPM tooling will auto-detect indeed
16:10 karolherbst: I think it requires patching the cmake file a bit to add it.. shouldn't be too difficult
16:15 mareko: konstantin: memory atomics on AMD GPUs are quite fast
16:21 HdkR: robclark: Apparently some user is spouting that Freedreno is buggy purely because some tessellation tests from "GPU Caps Viewer" flicker heavily, but don't with zink+turnip. Not sure how much you would care.
16:24 robclark: hmm, what is GPU Caps Viewer?
16:26 robclark: windows thing?
16:26 robclark: anyways, it isn't buggy if there isn't a bug report ;-)
16:27 HdkR: Yea, some Windows thing that has some benchmarks in it
16:27 HdkR: I told them that as well but they aren't the type of user to report it
16:28 HdkR: My X1E laptop is out on loan as well so reproducing myself will take time :D
16:32 robclark: only have the x13s with me atm.. but if I can reproduce it I can fix it
16:36 mlankhorst: Lyude: I was looking at drm-misc-fixes to see if I should send something out, noticed 560271e10b2c ("drm/nouveau/dispnv50: Don't call drm_atomic_get_crtc_state() in prepare_fb")
16:37 HdkR: robclark: Theoretically building arm64ec wine and fex should get it working but I would need to check on my devices as well.
16:37 mlankhorst: The IS_ERR() chunk can be and should be removed, as in prepare_fb() plane_state->crtc is always part of state