#asahi-gpu · 2026-09-27 16:32
janochaos_princess: yes, that is super helpful, thanks! I saw just the docs.html output before, didn't realize there was whole tooling around that#asahi-gpu · 2026-09-27 16:21
janojannau: disassembly of my own trivial shaders#asahi-gpu · 2026-09-27 16:21
janojannau: I am aware of that discussion, but did not look into any code shared there#asahi-gpu · 2026-09-27 16:16
jannaujano: what do you mean by "decompiling trivial shader probes"? looking at the disassembly of your own trivial shaders is ok#asahi-gpu · 2026-09-27 16:14
chaos_princesshttps://github.com/dougallj/applegpu - this also should be helpful, and if you do the same technique for shader analysis that it is doing, it is most likely kosher re-wise#asahi-gpu · 2026-09-27 16:12
jannauare you aware of https://github.com/ggml-org/llama.cpp/issues/10982 ? I would caution against looking at implementations there since you can't know they are clean room #asahi-gpu · 2026-09-27 16:02
janoI would like to get the support added; am I allowed to try to understand the bit representation of the instruction by decompiling trivial shader probes compiled with metal? am I allowed to then use this knowledge to build support for mesa/asahi, or does it need another person to avoid tainting?#asahi-gpu · 2026-09-27 15:59
janohey guys, I noticed that Mesa doesn't report support for cooperative matrix on Asahi and from metal probes the HW seems to support something that resembles a matrix fma instruction (testing on M1 ultra)#asahi-gpu · 2026-09-08 20:16
dyldgood luck#asahi-gpu · 2026-09-02 14:51
karolherbstwhat does the "failed to validate RA source 0 offs 0: actual.var == idx.value" error mean from a high level pov?#asahi-gpu · 2026-08-29 10:33
JamesCalligeros[m]Yeah i was just doing the wrong thing and setting 32 bytes per compression subtile#asahi-gpu · 2026-08-29 08:21
jannauchadmed: so that luma and chroma have the same number of tiles and a minimal tile size of 16x16#asahi-gpu · 2026-08-29 04:18
chadmed(for interchange)#asahi-gpu · 2026-08-29 04:18
chadmedohh the luma plane has 32 bytes of metadata per tile because there's 4 16x16 compression subtiles lol#asahi-gpu · 2026-08-26 01:47
waltmckoh wow huh it is vulkan#asahi-gpu · 2026-08-26 01:16
alyssaaccording to https://www.pcgamingwiki.com/wiki/No_Man's_Sky#API it's a native Vulkan title on windows (not DXVK)#asahi-gpu · 2026-08-26 01:15
alyssasorry, sloppy language on my part#asahi-gpu · 2026-08-26 01:15
alyssawaltmck: By Proton I meant DXVK / vkd3d-proton#asahi-gpu · 2026-08-26 00:27
waltmckIt's using DXVK#asahi-gpu · 2026-08-26 00:27
waltmckalyssa: this is actually a WINE game, I'm honestly not sure what proton does besides wrap WINE#asahi-gpu · 2026-08-25 23:15
alyssabut it's not clear to me what exactly the least incorrect patch would be#asahi-gpu · 2026-08-25 23:15
alyssawaltmck: also worst case, we could upstream a driconf for the game#asahi-gpu · 2026-08-25 22:45
alyssawaltmck: Proton games are all broken in the same way, native Vulkan games are each broken in their own unique and special way :p#asahi-gpu · 2026-08-25 21:47
waltmckopticron: not that I could find, beyond patching the binary (which I am considering)#asahi-gpu · 2026-08-25 21:46
waltmckJamesCalligeros[m]: yeah I had to do something else so I stopped it. will run it later when I have time#asahi-gpu · 2026-08-25 21:05
JamesCalligeros[m]Also the vk cts should not still be running lol, it should only take half an hour or so#asahi-gpu · 2026-08-25 20:57
opticronwaltmck, is it possible to tell NMS to decrease the pool size?#asahi-gpu · 2026-08-25 20:02
waltmckprobably I'll just keep running my local patch with the limit raised since this shouldn't be upstreamed#asahi-gpu · 2026-08-25 20:01
waltmckI don't think I've ever seen it use more than 32K queries at once, it just starts a query pool that is massive#asahi-gpu · 2026-08-25 20:00
waltmckbut yeah it sounds like the driver is doing the "correct" behavior and NMS is just not handling it well#asahi-gpu · 2026-08-25 19:59
waltmckvisually, no mans sky looks completely fine with the limit raised and looks completely broken with it left as is#asahi-gpu · 2026-08-25 19:59
waltmckhuh okay. I'll do some more testing and see if it breaks anything#asahi-gpu · 2026-08-25 19:58
alyssawaltmck: presumbaly the first 32k queries will work and the rest will not#asahi-gpu · 2026-08-25 19:58
waltmckhaven't finished running*#asahi-gpu · 2026-08-25 19:58
waltmckalyssa: very interesting, thank you. What is expected to happen if I just raise the limit anyway? Everything seems to work fine, but I haven't running the vulkan conformance tests#asahi-gpu · 2026-08-25 15:35
alyssaHope this helps#asahi-gpu · 2026-08-25 15:33
alyssaThe relevant limit is "Maximum visibility query offset", which on all M1-M5 chips is 256 KB = 32k queries at 64-bits each#asahi-gpu · 2026-08-25 15:32
alyssaSee https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf#asahi-gpu · 2026-08-25 15:24
alyssawhich comes out to 16k/32k queries depending if 32-bit/64-bit#asahi-gpu · 2026-08-25 15:24
alyssa#define PVR_SUBALLOCATOR_VIS_TEST_SIZE (128 * 1024)#asahi-gpu · 2026-08-25 15:24
alyssaYou'll see the powervr vk driver has a similar limit, in the form of:#asahi-gpu · 2026-08-25 15:22
alyssa(the hardware pool)#asahi-gpu · 2026-08-25 15:22
alyssa& yes, this is unusual, a bad design inherited from PowerVR#asahi-gpu · 2026-08-25 15:20
alyssa& I never saw a workload needing more than that anyway#asahi-gpu · 2026-08-25 15:20
alyssa(not impossible but complicated + perf impacts)#asahi-gpu · 2026-08-25 15:20
alyssaso hk limits to 32k occlusion queries VkDevice-wide, because doing otherwise would require major shenanigans#asahi-gpu · 2026-08-25 15:18
alyssaso architecturally M1 can only do 32k queries in a hardware pool#asahi-gpu · 2026-08-25 15:18
alyssa </struct>#asahi-gpu · 2026-08-25 15:18
alyssa <field name="Index" size="15" start="17" type="uint"/>#asahi-gpu · 2026-08-25 15:18
alyssa <!-- 64-bit index into occlusion result buffer -->