Vulkan profiles support?

[SRSaunders]

When experimenting with the Vulkan-Samples profiles example, I have found that MoltenVK is very close to supporting the VP_LUNARG_desktop_baseline_2022 profile, which depends on Vulkan 1.1.

The only divergence appears to be in the device properties.limits area, specifically as follows:

1. lineWidthGranularity = 0, but the profile requires a multiple of 1
2. minMemoryMapAlignment = 256 on AMD, but the profile requires <= 64 (which does work for Apple Silicon)
3. residencyAlignedMipSize = VK_FALSE, but the profile requires VK_TRUE

Is it possible to address these differences (perhaps even ignore item 2, since arm64 is the path forward for Apple)?

Or is profiles support not on the priority list, even for this relatively close one?

I have looked at some other profiles, but the divergence is larger with VP_LUNARG_desktop_baseline_2023 and VP_LUNARG_desktop_baseline_2024.

Note the profiles VP_LUNARG_minimum_requirements_1_0/1/2/3 are also close, but require:

1. lineWidthGranularity to be a multiple of 1
2. maxDescriptorSetSamplers >= 96 (which exceeds 80 at the moment)
3. maxSamplerAllocationCount >= 4000 (which exceeds 2048 on AMD, and 1024 on M1 Apple Silicon)
4. minMemoryMapAlignment <= 64 (which fails for AMD but does work for Apple Silicon)

[squidbus]

For lineWidthGranularity line width isn't supported at all without private APIs enabled, with private APIs it is set to 0.125. It could be valid to have it set to 1 when not supported since the lineWidthRange will ultimately define the supported widths as from 1 to 1, but I'm not sure.

For residencyAlignedMipSize, I think this comes down to sparse not being supported at all currently, so all of the VkPhysicalDeviceSparseProperties flags are false.

[squidbus]

Also for maxDescriptorSetSamplers I'm not sure how exactly that requirement was defined, as in the same profiles maxPerStageDescriptorSamplers is 16, and there are only 5 shader stages in a traditional graphics pipeline (5 * 16 = 80).

[SRSaunders]

Thank you @squidbus for your replies.

For lineWidthGranularity ... It could be valid to have it set to 1 when not supported since the lineWidthRange will ultimately define the supported widths as from 1 to 1, but I'm not sure.

Yes, I was wondering the same thing. That may be a valid interpretation to close off this difference.

For residencyAlignedMipSize, I think this comes down to sparse not being supported at all currently, so all of the VkPhysicalDeviceSparseProperties flags are false.

Yes, I believe you are confirming this is the main blocker for the VP_LUNARG_desktop_baseline_2022 profile, at least on Apple Silicon. Is sparse support on the roadmap for MoltenVK?

And for the VP_LUNARG_minimum_requirements_1_0/1/2/3 profiles:

Also for maxDescriptorSetSamplers I'm not sure how exactly that requirement was defined, as in the same profiles maxPerStageDescriptorSamplers is 16, and there are only 5 shader stages in a traditional graphics pipeline (5 * 16 = 80).

Could the difference be the lack of geometry shader support in MoltenVK? If so this may be a relatively permanent difference, as I understand that one is not a priority for future MoltenVK development.

[squidbus]

Yes, I believe you are confirming this is the main blocker for the VP_LUNARG_desktop_baseline_2022 profile, at least on Apple Silicon. Is sparse support on the roadmap for MoltenVK?

It seems it would be possible with Metal 4, may be a lot of work involved to get there. See #2560

Could the difference be the lack of geometry shader support in MoltenVK? If so this may be a relatively permanent difference, as I understand that one is not a priority for future MoltenVK development.

The 5 stages I'm counting includes geometry (vertex, tessellation control, tessellation evaluation, geometry, fragment), so unless I'm missing something I don't think so.

[SRSaunders]

It seems it would be possible with Metal 4, may be a lot of work involved to get there. See #2560

Thanks for the link. Indeed it looks like Metal4 may bring a lot of changes. Vulkan profiles support is of minor importance compared to that. I will leave this issue open for a bit in case others want to comment, but will likely close after that.