Support for `VK_KHR_shader_subgroup_uniform_control_flow` or even `VK_KHR_shader_maximal_reconvergence`?

[tareksander]

I don't know anything about the uniformity model of Metal. Would it be possible to implement these extensions? Given that subgroup operations are supported, it wouldn't seem far-fetched that MSL supports reconvergence, too.

[oscarbg]

Curiously these two were removed from android 16 vulkan profile because “older” Mali GPUs don’t support it.. hope Apple Mx GPUs and Metal API support it by now..:
KhronosGroup/Vulkan-Profiles@a8846f0

[oscarbg]

Adding also VK_KHR_shader_quad_control here because is also the last one of “shader only” extensions added in roadmap 2024 missing with rotated extension already implemented..
https://github.com/KhronosGroup/Vulkan-Docs/blob/main/appendices/roadmap/Roadmap-2024.adoc
Implementing them should be mainly adding spirvcross support..

[oscarbg]

And this is surely unsupported right in MSL now, but another shader only interesting ext:
VK_KHR_shader_clock

Edit:Might not be the best way to ask/request to Apple devs but posted here:
https://developer.apple.com/forums/thread/782390

[oscarbg]

well, at least there is hope the two requested extensions in this issue being exposed in Honeykrisp driver (Mesa Vulkan driver for Apple GPUs):
https://vulkan.gpuinfo.org/compare.php?reports=37244,38576#extensions
VK_KHR_shader_quad_control,VK_KHR_shader_clock still aren't exposed in HoneyKrisp..

[squidbus]

The main two asks here pass tests on Apple GPUs and have been added here: #2522

For VK_KHR_shader_quad_control I've been taking a look and have some of the SPIRV-Cross instruction wiring done but currently failing dEQP-VK.subgroups.shader_quad_control.divergent_condition.

[squidbus]

For VK_KHR_shader_quad_control I've been taking a look and have some of the SPIRV-Cross instruction wiring done but currently failing dEQP-VK.subgroups.shader_quad_control.divergent_condition.

This seems likely to be a Metal bug as opening a trace in a debugger shows the shader outputting the correct values but the actual render target containing incorrect values.

[squidbus]

It seems actually that with some fixes for OpSMod regressions all tests are now passing, so it may have been my fault. Not sure what was up with the debugger view I was seeing though.

[squidbus]

Quad control is added here: #2533

For VK_KHR_shader_clock it does not seem to be possible to implement currently in MSL. But that's also outside the scope of the original request here.