vector.mma

← vector dialect

Compute a matrix multiply-accumulate over target-shaped vector fragments using native matrix arithmetic. Floating intermediate precision, association, rounding, subnormal handling, and special-value behavior follow the selected target's arithmetic contract; the result type does not promise the bits of an ordered scalar FMA sequence. A target-independent reference decomposition converts each logical input to the accumulator type and applies scalar.fmaf in increasing K order from the initial accumulator. That reference evaluation remains permitted even when native execution differs. Use explicit scalar.fmaf operations or vector.dotf on widened inputs when ordered fused accumulation is required. The op consumes only the physical lhs, rhs, and init vectors; logical M/N/K shape, fragment role, packed storage schema, scales, codebooks, sparse metadata, and other interpretation data are carried by vector.fragment facts on those operands. Lowering queries those facts without baking target-specific witnesses into the MMA syntax. The optional saturate flag clamps an integer accumulation result to its destination range.

Operation contract

Property	Value
Semantic phase	executable
Traits	Pure, SafeToSpeculate
Target contracts	vector.contraction

Signature

Kind	Name	Type	Cardinality	Description
Operand	lhs	vector	required	Physical lhs fragment vector.
Operand	rhs	vector	required	Physical rhs fragment vector.
Operand	init	vector	required	Physical accumulator/addend fragment vector.
Result	result	vector	required	Updated accumulator/result fragment vector.
Attribute	flags	flags MmaFlags	optional	—

Verification constraints

• SameType(init, result)

Examples

%r = vector.mma %lhs, %rhs, %init : vector<8xf16>, vector<8xf16>, vector<8xf32>

%r = vector.mma<saturate> %lhs, %rhs, %init : vector<32xi8>, vector<32xi8>, vector<8xi32>

%r = vector.mma %lhs, %rhs, %init : vector<6xi32>, vector<6xi32>, vector<8xf32>