vector.fragment

← vector dialect

Attach a matrix-fragment interpretation to a physical vector value without changing the physical vector type. The role selects how the shape operands are interpreted: lhs is [m, k], rhs is [k, n], and init/result are [m, n]. A leading block extent forms independent batched fragments [b, m, k], [b, k, n], and [b, m, n]. Dense/default fragments need only the data value and shape SSA values. Encoded fragments carry schema and scale/table/sparse metadata values in the keyed using dictionary so bulk runtime data remains ordinary SSA while lowering can consume a compact resolved fragment fact.

Operation contract

Property	Value
Semantic phase	executable
Traits	Pure, RefinableResultTypeRefs, ValueAlias

Signature

Kind	Name	Type	Cardinality	Description
Operand	data	vector	required	Physical vector lanes or packed registers.
Operand	blocks	index	optional	Optional independent matrix block count.
Operand	rows	index	required	Logical matrix row count for this fragment role.
Operand	columns	index	required	Logical matrix column count for this fragment role.
Operand	params	any	variadic	Optional schema, scale, table, metadata, or online state operands.
Result	result	vector	required	The same physical vector value with fragment facts.
Attribute	role	enum VectorFragmentRole	required	—
Attribute	param_names	dict	optional	Sorted fragment parameter names mapped to parameter operand ordinals.
Attribute	predicates	predicate_list	optional	Optional local facts constraining fragment shape or parameter values.

Verification constraints

• SameType(data, result)

Examples

%fragment = vector.fragment<lhs> %payload shape [%m, %k] : vector<4xi32>

%fragment = vector.fragment<rhs> %payload shape [%k, %n] using {scale = %scale : vector<1xf16>, schema = %schema : encoding<schema>} : vector<4xi32>