‘symbol’ Dialect

Symbol dialect to represent and expresses dynamic shape’s relationship.

Operations

symbol.bind_symbolic_shape (symbol::BindSymbolicShapeOp)

Binds shape expressions to tensors using an affine map indexed by shape symbols

The symbol.bind_symbolic_shape operation binds shape expressions useful to compute the dynamic dimensions of a tensor. It takes a variadic of SSA symbols that map 1:1 to the local symbols declared in the affine map. The affine map contains a list of affine shape expressions for each dim where the terminals are from the declared symbols.

Example:

symbol.bind_symbolic_shape %arg0, [%0, %1], affine_map<()[s0, s1] -> (s0, s1, 3)> : tensor<?x?x3xf32>
symbol.bind_symbolic_shape %out0, [%0, %1, %2], affine_map<()[s0, s1, s2] -> (s0, s1 * 2 + s2, 3)> : tensor<?x?x3xf32>

Attributes:

Attribute	MLIR Type	Description
shape_expressions	::mlir::AffineMapAttr	An Attribute containing an AffineMap object {{% markdown %}} Syntax: ``` affine-map-attribute ::= `affine_map` `<` affine-map `>` ``` Examples: ```mlir affine_map<(d0) -> (d0)> affine_map<(d0, d1, d2) -> (d0, d1)> ``` {{% /markdown %}}

Operands:

Operand	Description
operand	shaped of any type values
shape_symbols	variadic of index

symbol.symbolic_int (symbol::SymbolicIntOp)

Represents a symbolic integer with a constrained range

The “symbolic_int” operation represents a symbolic integer value by name (via a symbol attribute) and returns it as an index-typed value. It includes range constraints through min_val and max_val attributes that specify the inclusive bounds for the symbolic value.

This operation is typically used to represent dynamic dimensions or other symbolic integer values that have known constraints.

Example:

%0 = symbol.symbolic_int @s0 {min_val = 5, max_val = 10} : index
%1 = symbol.symbolic_int @s1 {min_val = 2, max_val = 20} : index
%2 = symbol.symbolic_int @s2 [%0, %1], affine_map<()[s1, s2] -> (s1 * s2)> {min_val = 2, max_val = 20} : index

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface), OpAsmOpInterface

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
symbol_name	::mlir::FlatSymbolRefAttr	flat symbol reference attribute
min_val	::mlir::IntegerAttr	64-bit signless integer attribute
max_val	::mlir::IntegerAttr	64-bit signless integer attribute
int_expressions	::mlir::AffineMapAttr	An Attribute containing an AffineMap object {{% markdown %}} Syntax: ``` affine-map-attribute ::= `affine_map` `<` affine-map `>` ``` Examples: ```mlir affine_map<(d0) -> (d0)> affine_map<(d0, d1, d2) -> (d0, d1)> ``` {{% /markdown %}}

Operands:

Operand	Description
int_symbols	variadic of index

Results:

Result	Description
result	index