kernel.async.gather

← kernel dialect

Initiate a subgroup-collective asynchronous gather from each invocation's source view into a lane-contiguous workgroup destination view. The destination view has one leading subgroup-lane axis and a trailing lane slot with enough static bytes to hold one source payload. If the lane slot is larger than the source footprint, the extra destination bytes are padding bytes with unspecified contents. The destination denotes the subgroup-uniform base tile; the current subgroup lane is applied by the op semantics and must not be pre-applied by forming a lane subview. This directly represents AMDGPU global_load_lds-style staging, including padded narrow loads, without requiring a later pass to rediscover that a set of per-lane copies was really one subgroup LDS DMA operation.

Operation contract

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

Signature

Kind	Name	Type	Cardinality	Description
Operand	source	view	required	Per-invocation global-like source fragment.
Operand	dest	view	required	Subgroup-uniform workgroup destination tile with a leading subgroup-lane axis.
Result	token	any	required	Opaque async-copy token for the subgroup gather.
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.gather %src_lane to %lds_tile {cache_scope = cu, cache_temporal = regular} : view<4xi8> to view<[%wave]x4xi8> -> kernel.async.token

%copy = kernel.async.gather %src_lane to %lds_tile {cache_scope = cu, cache_temporal = regular} : view<12xi8> to view<64x16xi8> -> kernel.async.token