Task list for VK_NV_cooperative_matrix release

[jeffbolznv]

The Vulkan 1.1.101 spec release includes the VK_NV_cooperative_matrix extension. This extension adds support for using cooperative matrix types in SPIR-V. Cooperative matrix types are medium-sized matrices that are primarily supported in compute shaders, where the storage for the matrix is spread across all invocations in some scope (usually a subgroup) and those invocations cooperate to efficiently perform matrix multiplies. These operations are accelerated by NVIDIA Tensor Cores on Turing GPUs.

The task list for the VK_NV_cooperative_matrix extension release is:

• Vulkan specification (https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VK_NV_cooperative_matrix)
• SPIRV-Registry (Add SPV_NV_cooperative_matrix SPIRV-Registry#20)
• SPIRV-Headers (Add SPV_NV_cooperative_matrix SPIRV-Headers#94)
• SPIRV-Tools (Add validation for SPV_NV_cooperative_matrix SPIRV-Tools#2404)
• GLSL extension (Add GL_NV_cooperative_matrix GLSL#55)
• glslang implementation (Implement GL_NV_cooperative_matrix glslang#1701)
• validation layer (Validate VK_NV_cooperative_matrix Vulkan-ValidationLayers#742)
• Conformance tests (https://github.com/KhronosGroup/VK-GL-CTS/releases/tag/vulkan-cts-1.1.4.0)
• Sample application/benchmark (https://github.com/jeffbolznv/vk_cooperative_matrix_perf)

As each component is made public, the task will be checked off. When all tasks have been completed this issue will be closed and the extension will be fully released.

[pdaniell-nv]

Driver versions 419.09 (Windows) and 418.31.03 (Linux) implement this extension and can be found here: https://developer.nvidia.com/vulkan-driver.

[oscarbg]

@pdaniell-nv can't see this interesting extension (VK_NV_cooperative_matrix) exposed on Titan V (Volta) on 419.09..
https://vulkan.gpuinfo.org/compare.php?compare=compare&id%5B5264%5D=on&id%5B4960%5D=on
is that expected? it's because Volta Tensor cores aren't "rich enough" to support it?
thanks..

[jeffbolznv]

Right, this is only currently supported on Turing.

[minisparrow]

could the warpgroup mma(H100 WGMMA) using this extension?