You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thanks for filing! This limitation originally came from Metal (Apple1 and Apple2 families) and Vulkan in #1863. However according to #1069 we do not support Apple1/Apple2 anymore anyway. So I checked using the gpuinfo-vulkan-query tool to see if we really need 16352 on Vulkan, and it appears we do not.
Unfortunately it doesn't look like we can bump to 32k right now - this would drop Metal family Apple3 (A9/A10) and Adreno 610/612 (though interestingly it doesn't drop any Adreno 500 devices). But it's a definite possibility for the future.
Activity
kainino0x commentedon May 6, 2022
Thanks for filing! This limitation originally came from Metal (Apple1 and Apple2 families)
and Vulkanin #1863. However according to #1069 we do not support Apple1/Apple2 anymore anyway. So I checked using the gpuinfo-vulkan-query tool to see if we really need 16352 on Vulkan, and it appears we do not.Unfortunately it doesn't look like we can bump to 32k right now - this would drop Metal family Apple3 (A9/A10) and Adreno 610/612 (though interestingly it doesn't drop any Adreno 500 devices). But it's a definite possibility for the future.
D3D SM5.0 guarantees 32kB (#1069 says we require SM5.1): https://microsoft.github.io/DirectX-Specs/d3d/archive/D3D11_3_FunctionalSpec.htm#18.6.6%20Inter-Thread%20Data%20Sharing
kainino0x commentedon May 6, 2022
Oh, I was going to look closer at Vulkan. Actually, it doesn't look like any Vulkan devices go below 16384:
http://vulkan.gpuinfo.org/displaydevicelimit.php?name=maxComputeSharedMemorySize&platform=all
So the restriction was from Metal only.
kainino0x commentedon May 6, 2022
BTW, this is only the required minimum value. Implementations can already expose 16384 or higher values.
kainino0x commentedon May 6, 2022
(Actually I'm going to open a PR for 16384 in V1 and recategorize this issue as post-V1 for 32768.)