vector.atomic.reduce

← vector dialect

Atomic no-result scatter reduction/update into per-lane signed element offsets. Each lane atomically combines its value into origin + offsets[lane]; duplicate active addresses are valid and are serialized by the required ordering and scope attributes.

Operation contract

Property	Value
Semantic phase	—
Interfaces	CachePolicy, MemoryAccess
Memory effects	readwrite view

Signature

Kind	Name	Type	Cardinality	Description
Operand	value	vector	required	Fixed-width byte-addressable contribution for each lane.
Operand	view	view	required	Typed destination view.
Operand	offsets	vector	required	Per-lane signed element offsets from the logical origin.
Operand	indices	index	variadic	Dynamic logical origin indices.
Attribute	kind	enum AtomicKind	required	—
Attribute	memory_flags	flags AtomicMemoryFlags	optional	—
Attribute	ordering	enum AtomicOrdering	required	Required atomic memory ordering.
Attribute	scope	enum AtomicScope	required	Required atomic synchronization scope.
Attribute	cache_scope	enum CacheScope	optional	Optional cache/coherency scope required by target lowering.
Attribute	cache_temporal	enum CacheTemporal	optional	Optional temporal cache policy required by target lowering.
Attribute	static_indices	i64_array	required	Static logical origin indices with INT64_MIN sentinels for dynamics.

Verification constraints

• HasBytePatternElement(value)
• HasIndexOrNonI1IntegerElement(offsets)
• SameElementType(value, view)
• SameShape(offsets, value)

Examples

vector.atomic.reduce<addi> %v, %view[%row, %col][%offsets] {ordering = relaxed, scope = workgroup} : vector<4xi32>, view<[%m]x[%n]xi32, %layout>, vector<4xindex>