The 23.20.16.5018 driver appeared in Windows Update a few days ago.
I see that it is the first Windows driver with a device version string of "OpenCL 2.1 NEO".
The driver also supports Vulkan 1.1 and enables all subgroup operations across all stages. Impressive!
One question, will Vulkan shaders always be compiled with a subgroup width of 32 or will the compiler have latitude to select a subgroup size (8, 16, or 32) based on register occupancy like the OpenCL kernel compiler?
Presumably this OpenCL runtime is based on the recently open-sourced NEO codebase?
This is very exciting!