kernel.async.tensor.load.to.lds

← kernel dialect

Initiate an AMDGPU gfx1250+ tensor-memory load from a global-like source view into a workgroup/LDS destination view using an explicit kernel.tensor.lds.descriptor. The descriptor supplies the exact hardware dgroups, while the source and destination views keep the logical rank, element type, layout, and memory-space facts visible. The endpoints must have the same rank in [1, 5], the same 1/2/4/8 byte element type, and memory spaces global/constant/descriptor to workgroup. The returned token must be committed to exactly one kernel.async.group.

Operation contract

Property	Value
Semantic phase	—
Target contracts	kernel.async, kernel.tensor_memory
Memory effects	read source, write dest

Signature

Kind	Name	Type	Cardinality	Description
Operand	source	view	required	Global-like tensor-memory source view.
Operand	dest	view	required	Workgroup/LDS tensor-memory destination view.
Operand	descriptor	any	required	Tensor LDS descriptor supplying AMDGPU dgroups.
Result	token	any	required	Opaque async-copy token for the tensor load.
Attribute	cache_scope	enum CacheScope	required	Required cache/coherency scope for the transfer.
Attribute	cache_temporal	enum CacheTemporal	required	Required temporal cache hint for the transfer.

Examples

%copy = kernel.async.tensor.load.to.lds %global_tile to %lds_tile using %desc {cache_scope = cu, cache_temporal = regular} : view<64x64xf32> to view<64x64xf32>, kernel.tensor.lds.descriptor -> kernel.async.token