#asahi-gpu · 2026-08-25 15:18
alyssa <struct name="Fragment occlusion query" size="4">#asahi-gpu · 2026-08-25 15:18
alyssawaltmck: #asahi-gpu · 2026-08-25 11:32
chaos_princessbut yea, looks more atv-related#asahi-gpu · 2026-08-25 11:32
chaos_princessthey did briefly do a mac mini as htpc thing, with an ir remote & stuff#asahi-gpu · 2026-08-25 11:32
chadmedmy only thinking is that its a remnant of some sort of apple tv specific format that predates both AGX and Interchange#asahi-gpu · 2026-08-25 11:31
chadmedand then i remembered i own 4 apple tvs :P #asahi-gpu · 2026-08-25 11:31
chadmedthat was my thinking, and i thought "no that sounds stupid when have apple ever done anything like an htpc" #asahi-gpu · 2026-08-25 11:30
jannauhome theater pc surface? would kind of fit to the scaler limitation of full output width#asahi-gpu · 2026-08-25 11:28
chadmedand i still have no idea what an "HTPC Surface" is#asahi-gpu · 2026-08-25 11:28
chadmeddoes seem like it would Just Work(tm) with the correct surface metadata though #asahi-gpu · 2026-08-25 11:28
chadmedah well interchange will suffice for now i suppose, not like macos ever does this anyway#asahi-gpu · 2026-08-25 11:27
chadmedi dont think im setting up the gpu tiling layout properly, no matter what i do iomfb complains about the plane size being too small for whatever sanity check it's calculated itself#asahi-gpu · 2026-08-25 05:41
JamesCalligeros[m]Theres also a list of tests we expect to fail somewhere in .gitlab-ci#asahi-gpu · 2026-08-25 04:35
JamesCalligeros[m]nice#asahi-gpu · 2026-08-25 04:35
waltmckI'm running it now, it is in nixpkgs so not too much trouble#asahi-gpu · 2026-08-25 04:14
JamesCalligeros[m]building and running it is a little uhhhh cumbersome#asahi-gpu · 2026-08-25 04:13
JamesCalligeros[m]yeah that one#asahi-gpu · 2026-08-25 04:05
waltmckyou mean https://github.com/KhronosGroup/VK-GL-CTS ?#asahi-gpu · 2026-08-25 04:05
waltmckI did not run the cts, I can do that#asahi-gpu · 2026-08-25 03:54
JamesCalligeros[m]We cannot accept gpu patches that regress the vulkan or opengl cts#asahi-gpu · 2026-08-25 03:54
JamesCalligeros[m]Did you run the CTS?#asahi-gpu · 2026-08-25 03:37
waltmckwould you accept a patch that raises the limit?#asahi-gpu · 2026-08-25 03:37
waltmckwith the current limit that takes 256KiB, with my 1M limit it is 8MiB#asahi-gpu · 2026-08-25 03:36
waltmckso you need to allocate the entire table up front regardless of how much is actually used#asahi-gpu · 2026-08-25 03:36
waltmcksee https://gitlab.freedesktop.org/mesa/mesa/-/blob/mesa-26.1.5/src/asahi/vulkan/hk_descriptor_table.c#L35#asahi-gpu · 2026-08-25 03:36
waltmckI don't really understand this stuff very well, but I think that the issue is that on honeykrisp the buffer object table can't grow#asahi-gpu · 2026-08-25 03:34
waltmckchaos_princess: increasing the occlusion query limit to 1M increases the bitmap from 4KiB to 128KiB. So not a very significant change#asahi-gpu · 2026-08-24 20:34
noopwafelright now I am mostly just going back and cleaning up the "# FIXME: alyssa: non-zero data in the padding" (where usually it is not, in fact, non-zero data in the padding..)#asahi-gpu · 2026-08-24 20:34
noopwafelso a test suite would be <3#asahi-gpu · 2026-08-24 20:33
noopwafel(with some way to stub the adt/gpu register reads)#asahi-gpu · 2026-08-24 20:33
noopwafelI was hoping that I could just wrap rust_fill_gpu_initdata somehow for regression testin#asahi-gpu · 2026-08-24 20:33
jannaumaybe we should capture adt and initdata and create a test suite#asahi-gpu · 2026-08-24 20:31
jannaustarlight-debug is merged in the downstream kernel. not sure how much help it's is in porting to to 14.x but at least it protects against m1n1 rust code regressing fro 12.x and 13.5#asahi-gpu · 2026-08-24 20:26
chaos_princessyes, i was going to suggest to use that to verify that the "churn" does not break thing#asahi-gpu · 2026-08-24 20:26
noopwafelchaos_princess: no, but is that a kernel branch comparing the kernel-side w/m1n1-side?#asahi-gpu · 2026-08-24 20:22
chaos_princessnoopwafel: do you know about the "starlight-debug" branch#asahi-gpu · 2026-08-24 20:21
noopwafelalthough it is actually what I am staring at .. again#asahi-gpu · 2026-08-24 20:21
noopwafeloops, sorry, that was me fighting ssh#asahi-gpu · 2026-08-24 20:21
noopwafelPorting all the initdata stuff back to rust (in m1n1) is the sensible next step on the driver side.. a lot of churn but mostly testable on M1/M2.#asahi-gpu · 2026-08-24 18:51
chaos_princessprobably because the driver uses a bitmap to track queries, while a set could be used instead#asahi-gpu · 2026-08-24 18:28
waltmckDo you own No Man's Sky on steam or gog? I can give instructions on how to reproducibly run it from a nix derivation#asahi-gpu · 2026-08-24 18:27
waltmckdoes apple silicon require different data structurs that have worse scaling than these other GPUs? I guess that might make sense#asahi-gpu · 2026-08-24 18:27
waltmckI guess this is in the weeds, but in that case how come there is no limit in any of the other drivers I could find?#asahi-gpu · 2026-08-24 18:16
chaos_princessthere are some data structures that scale with it#asahi-gpu · 2026-08-24 18:13
waltmckI'm honestly not sure why there is such a small limit on occlusion queries in honeykrisp considering it seems to run fine with it increased#asahi-gpu · 2026-08-24 18:13
waltmckLet me verify everything and I'll publish the game (packaged with Nix). You should be able to run it and test it yourself then if you're curious#asahi-gpu · 2026-08-24 18:12
waltmckto -> go#asahi-gpu · 2026-08-24 18:12
waltmckIf I patch the driver to increase AGX_MAX_OCCLUSION_QUERIES to 1M, the artifacts to away#asahi-gpu · 2026-08-24 18:11
waltmckOK, the reason for the crash was actually something else. The game runs without increasing AGX_MAX_OCCLUSION_QUERIES, but there are flickering artifacts#asahi-gpu · 2026-08-24 16:35
waltmckIt looks like this is pretty unusual, the other drivers I looked at (radv, anv, nvk, turnip) don't have any limit on the number of occlusion queries