Support subgroups

[VarLad]

See https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#cl_khr_subgroups

[VarLad]

This probably needs to be done in SPIRVIntrinsics
https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_group_and_subgroup_instructions

Also need to figure out what do with optional extensions:
https://github.khronos.org/SPIRV-Registry/extensions/INTEL/SPV_INTEL_subgroups

both are supported in both Rusticl+Zink and pocl

[VarLad]

More useful info from Slack from Tim:

basically https://llvm.org/docs/SPIRVUsage.html#builtin-functions; we depend on the LLVM SPIR-V back-end now, and thus calling SPIR-V intrinsics directoy, and not on the translator upgrading OpenCL intrinsics.
see e.g.

OpenCL.jl/lib/intrinsics/src/synchronization.jl

Lines 39 to 49 in 5355310

	@device_function memory_barrier(scope, semantics) =
	Base.llvmcall(("""
	declare void @_Z21__spirv_MemoryBarrierjj(i32, i32) #0
	define void @entry(i32 %scope, i32 %semantics) #1 {
	call void @_Z21__spirv_MemoryBarrierjj(i32 %scope, i32 %semantics)
	ret void
	}
	attributes #0 = { convergent }
	attributes #1 = { alwaysinline }
	""", "entry"),
	Cvoid, Tuple{UInt32, UInt32}, convert(UInt32, scope), convert(UInt32, semantics))

note the manual IA64 mangling

[simeonschaub]

Note that we do already have some subgroup intrinsics wrapped in https://github.com/JuliaGPU/OpenCL.jl/blob/master/lib/intrinsics/src/work_item.jl. I have also been playing around with subgroup shuffles in #356