Vulkan 1.3 requirements

[billhollings]

Appendix D of the Vulkan spec and the VK_VERSION_1_3 man page summarize the extension list below, and adds additional Vulkan 1.3 enhancement requirements.

Extension	MoltenVK Support	Status Notes
VK_KHR_copy_commands2	✅
VK_KHR_dynamic_rendering	✅
VK_KHR_format_feature_flags2	✅	Added (#2108)
VK_KHR_maintenance4		In Dev (#2116)
VK_KHR_shader_integer_dot_product	✅	Added (#2119)
VK_KHR_shader_non_semantic_info	✅
VK_KHR_shader_terminate_invocation	✅	Added ( #2424)
VK_KHR_synchronization2	✅	Added (#2021)
VK_KHR_zero_initialize_workgroup_memory	✅	Added (#2450)
VK_EXT_4444_formats	✅
VK_EXT_extended_dynamic_state	✅	Added (#2036)
VK_EXT_extended_dynamic_state2	✅	Added (#2036)
VK_EXT_image_robustness	✅
VK_EXT_inline_uniform_block	✅
VK_EXT_pipeline_creation_cache_control	✅
VK_EXT_pipeline_creation_feedback	✅
VK_EXT_private_data	✅
VK_EXT_shader_demote_to_helper_invocation	✅
VK_EXT_subgroup_size_control	✅
VK_EXT_texel_buffer_alignment	✅
VK_EXT_texture_compression_astc_hdr	✅
VK_EXT_tooling_info	✅	Added (#2435)
VK_EXT_ycbcr_2plane_444_formats		Not required for 1.3 (see spec re promotion of this extension)

---

In addition to the promoted extensions described above, Vulkan 1.3 added required support for:

• SPIR-V version 1.6
  • SPIR-V 1.6 deprecates (but does not remove) the WorkgroupSize decoration.
• The bufferDeviceAddress feature which indicates support for accessing memory in shaders as storage buffers via vkGetBufferDeviceAddress.
• The vulkanMemoryModel and vulkanMemoryModelDeviceScope features, which indicate support for the corresponding Vulkan Memory Model capabilities.
• The maxInlineUniformTotalSize limit is added to provide the total size of all inline uniform block bindings in a pipeline layout.

Additional details of functionality and information that may need to be included in a 1.3 implementation can be found here.

[CarterLi]

Ref: haasn/libplacebo#174

[rcaridade145]

In regards to VK_KHR_shader_integer_dot_product

I came across

gpuweb/gpuweb#2677

Which led to me.

-https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/compiler/nir/nir_opcodes.py#L1367

Hopefully this can be useful.

[cdavis5e]

In regards to VK_KHR_shader_integer_dot_product

I came across

gpuweb/gpuweb#2677

Which led to me.

-https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/compiler/nir/nir_opcodes.py#L1367

Hopefully this can be useful.

That bit of code you highlighted is for a packed saturating vertical add. That's only part of three of the new instructions added by that extension. Metal already has that (as addsat()). For a dot product, we need a vertical multiply followed by a horizontal add (i.e. add all elements of the product vector together). AFAIK, Metal doesn't have either an overload of dot() that accepts integral input or any sort of horizontal add. We'd have to expand the final addition manually, which would be painful.

[billhollings]

We'd have to expand the final addition manually, which would be painful.

By that, you mean adding the required synthetic functions to SPIRV-Cross?

[cdavis5e]

We'd have to expand the final addition manually, which would be painful.

By that, you mean adding the required synthetic functions to SPIRV-Cross?

Yes.

[rcaridade145]

gfx-rs/naga#1689 People of webgpu seem to have dealt with dot product with integers on Metal.

[rcaridade145]

Been studying spirv-cross code and other alternatives like naga and tint (really nice code) .

Tint - https://dawn.googlesource.com/tint/+/refs/heads/main/src/tint/writer/msl/generator_impl.cc#681

It should be implemented something like that (to support some integer dot) here no?

https://github.com/KhronosGroup/SPIRV-Cross/blob/main/spirv_msl.cpp#L5414

[devshgraphicsprogramming]

btw vulkanMemoryModel starts to be quite important when SPV_KHR_physical_storage_buffer gets used because you start needing to use it on OpLoad and OpStore if you want Buffer Device Addresses to volatile/coherent variables