kernel.workgroup.reduce

← kernel dialect

Reduce one or more independent scalar or rank-1 vector payloads at one workgroup convergence point. Each result has the type of the operand in the same position. All payloads share the combining kind and math requirement; no combination occurs between payloads. Each active contribution participates once per component; association is unspecified. Floating sums and products default to native target arithmetic, which may change signed zeros, flush subnormals, or relax nonfinite handling. The optional preserve requirement instead evaluates each combine with the source format's arithmetic semantics. Neither mode promises serial-reference bits, reproducibility, or a correctly rounded real-number sum. Integer arithmetic and the selected floating-extremum kind keep their exact semantics.

Operation contract

Property	Value
Semantic phase	executable
Traits	Convergent, HasAnyAncestor(kernel.def, pipeline.strand)
Target contracts	kernel.synchronization

Signature

Kind	Name	Type	Cardinality	Description
Operand	values	any	variadic	Nonempty independent per-invocation payloads to reduce.
Result	results	any	variadic	Workgroup-uniform results corresponding positionally to the payloads.
Attribute	kind	enum CombiningKind	required	Combining operation.
Attribute	math	flags CollectiveMathFlags	optional	Optional floating-point preservation requirement; omission selects native arithmetic.

Verification constraints

• VariadicValuesMatch(values, results)

Examples

%sum = kernel.workgroup.reduce<addf> %v : f32

%sum, %moment = kernel.workgroup.reduce<addf> %value, %squared : f32, vector<4xf32>