Vector

vector

DYNAMIC_INDEX: int = -2 ** 63 module-attribute

Vector = Dialect('vector', [BitcastOp, BroadcastOp, CreateMaskOp, ExtractOp, FMAOp, InsertOp, LoadOp, MaskedLoadOp, MaskedStoreOp, PrintOp, ReductionOp, ShuffleOp, StoreOp, TransferReadOp, TransferWriteOp], [CombiningKindAttr]) module-attribute

LoadOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class LoadOp(IRDLOperation):
    name = "vector.load"
    base = operand_def(MemRefType)
    indices = var_operand_def(IndexType)
    result = result_def(VectorType)
    nontemporal = opt_prop_def(BoolAttr, default_value=BoolAttr.from_bool(False))

    irdl_options = (ParsePropInAttrDict(),)
    assembly_format = (
        "$base `[` $indices `]` attr-dict `:` type($base) `,` type($result)"
    )

    def __init__(
        self,
        ref: SSAValue | Operation,
        indices: Sequence[SSAValue | Operation],
        result_type: VectorType,
    ):
        super().__init__(
            operands=(ref, indices),
            result_types=(result_type,),
        )

    def verify_(self):
        assert isa(self.base.type, MemRefType)
        assert isa(self.result.type, VectorType[Attribute])

        if self.base.type.element_type != self.result.type.element_type:
            raise VerifyException(
                "MemRef element type should match the Vector element type."
            )

        if self.base.type.get_num_dims() != len(self.indices):
            raise VerifyException("Expected an index for each dimension.")

name = 'vector.load' class-attribute instance-attribute

base = operand_def(MemRefType) class-attribute instance-attribute

indices = var_operand_def(IndexType) class-attribute instance-attribute

result = result_def(VectorType) class-attribute instance-attribute

nontemporal = opt_prop_def(BoolAttr, default_value=(BoolAttr.from_bool(False))) class-attribute instance-attribute

irdl_options = (ParsePropInAttrDict(),) class-attribute instance-attribute

assembly_format = '$base `[` $indices `]` attr-dict `:` type($base) `,` type($result)' class-attribute instance-attribute

__init__(ref: SSAValue | Operation, indices: Sequence[SSAValue | Operation], result_type: VectorType)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    ref: SSAValue | Operation,
    indices: Sequence[SSAValue | Operation],
    result_type: VectorType,
):
    super().__init__(
        operands=(ref, indices),
        result_types=(result_type,),
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    assert isa(self.base.type, MemRefType)
    assert isa(self.result.type, VectorType[Attribute])

    if self.base.type.element_type != self.result.type.element_type:
        raise VerifyException(
            "MemRef element type should match the Vector element type."
        )

    if self.base.type.get_num_dims() != len(self.indices):
        raise VerifyException("Expected an index for each dimension.")

StoreOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class StoreOp(IRDLOperation):
    name = "vector.store"
    vector = operand_def(VectorType)
    base = operand_def(MemRefType)
    indices = var_operand_def(IndexType)
    nontemporal = opt_prop_def(BoolAttr, default_value=BoolAttr.from_bool(False))

    irdl_options = (ParsePropInAttrDict(),)
    assembly_format = (
        "$vector `,` $base `[` $indices `]` attr-dict `:` type($base) `,` type($vector)"
    )

    def __init__(
        self,
        vector: SSAValue | Operation,
        base: SSAValue | Operation,
        indices: Sequence[SSAValue | Operation],
        nontemporal: BoolAttr | None = None,
    ):
        super().__init__(
            operands=[vector, base, indices],
            properties={"nontemporal": nontemporal},
        )

    def verify_(self):
        assert isa(self.base.type, MemRefType)
        assert isa(self.vector.type, VectorType[Attribute])

        if self.base.type.element_type != self.vector.type.element_type:
            raise VerifyException(
                "MemRef element type should match the Vector element type."
            )

        if self.base.type.get_num_dims() != len(self.indices):
            raise VerifyException("Expected an index for each dimension.")

name = 'vector.store' class-attribute instance-attribute

vector = operand_def(VectorType) class-attribute instance-attribute

base = operand_def(MemRefType) class-attribute instance-attribute

indices = var_operand_def(IndexType) class-attribute instance-attribute

nontemporal = opt_prop_def(BoolAttr, default_value=(BoolAttr.from_bool(False))) class-attribute instance-attribute

irdl_options = (ParsePropInAttrDict(),) class-attribute instance-attribute

assembly_format = '$vector `,` $base `[` $indices `]` attr-dict `:` type($base) `,` type($vector)' class-attribute instance-attribute

__init__(vector: SSAValue | Operation, base: SSAValue | Operation, indices: Sequence[SSAValue | Operation], nontemporal: BoolAttr | None = None)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    vector: SSAValue | Operation,
    base: SSAValue | Operation,
    indices: Sequence[SSAValue | Operation],
    nontemporal: BoolAttr | None = None,
):
    super().__init__(
        operands=[vector, base, indices],
        properties={"nontemporal": nontemporal},
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    assert isa(self.base.type, MemRefType)
    assert isa(self.vector.type, VectorType[Attribute])

    if self.base.type.element_type != self.vector.type.element_type:
        raise VerifyException(
            "MemRef element type should match the Vector element type."
        )

    if self.base.type.get_num_dims() != len(self.indices):
        raise VerifyException("Expected an index for each dimension.")

ShuffleResultConstraint dataclass

Bases: AttrConstraint[VectorType]

Source code in xdsl/dialects/vector.py

@dataclass(frozen=True)
class ShuffleResultConstraint(AttrConstraint[VectorType]):
    element_constr: AttrConstraint
    v1_shape_constr: VarConstraint
    v2_shape_constr: VarConstraint
    mask_constraint: VarConstraint

    def verify(self, attr: Attribute, constraint_context: ConstraintContext) -> None:
        # We can only verify the element type here, and not the relations to other shapes
        VectorType.constr(self.element_constr).verify(attr, constraint_context)
        attr = cast(VectorType, attr)
        if not attr.shape.data:
            raise VerifyException("Result vector type must not be 0-D.")

    def can_infer(self, var_constraint_names: AbstractSet[str]) -> bool:
        res = self.element_constr.can_infer(var_constraint_names) and (
            self.v1_shape_constr.name in var_constraint_names
            and self.v2_shape_constr.name in var_constraint_names
            and self.mask_constraint.name in var_constraint_names
        )
        assert res
        return res

    def infer(self, context: ConstraintContext) -> VectorType:
        v1_shape = context.get_variable(self.v1_shape_constr.name)
        v2_shape = context.get_variable(self.v2_shape_constr.name)
        mask = context.get_variable(self.mask_constraint.name)
        assert v1_shape is not None
        assert v2_shape is not None
        assert mask is not None
        assert _IntArrayConstr.verifies(v1_shape)
        assert _IntArrayConstr.verifies(v2_shape)
        assert _MaskConstr.verifies(mask)

        result_trailing: tuple[IntAttr, ...]
        if not v1_shape:
            assert not v2_shape
            result_trailing = ()
        else:
            result_trailing = v1_shape.data[1:]

        element_type = self.element_constr.infer(context)
        result_leading = len(mask)
        shape = (
            (IntAttr(result_leading), *result_trailing)
            if result_leading
            else result_trailing
        )
        return VectorType(element_type, ArrayAttr(shape))

    def mapping_type_vars(
        self, type_var_mapping: Mapping[TypeVar, AttrConstraint | IntConstraint]
    ) -> AttrConstraint[VectorType]:
        return ShuffleResultConstraint(
            self.element_constr.mapping_type_vars(type_var_mapping),
            self.v1_shape_constr.mapping_type_vars(type_var_mapping),
            self.v2_shape_constr.mapping_type_vars(type_var_mapping),
            self.mask_constraint.mapping_type_vars(type_var_mapping),
        )

element_constr: AttrConstraint instance-attribute

v1_shape_constr: VarConstraint instance-attribute

v2_shape_constr: VarConstraint instance-attribute

mask_constraint: VarConstraint instance-attribute

__init__(element_constr: AttrConstraint, v1_shape_constr: VarConstraint, v2_shape_constr: VarConstraint, mask_constraint: VarConstraint) -> None

verify(attr: Attribute, constraint_context: ConstraintContext) -> None

Source code in xdsl/dialects/vector.py

def verify(self, attr: Attribute, constraint_context: ConstraintContext) -> None:
    # We can only verify the element type here, and not the relations to other shapes
    VectorType.constr(self.element_constr).verify(attr, constraint_context)
    attr = cast(VectorType, attr)
    if not attr.shape.data:
        raise VerifyException("Result vector type must not be 0-D.")

can_infer(var_constraint_names: AbstractSet[str]) -> bool

Source code in xdsl/dialects/vector.py

def can_infer(self, var_constraint_names: AbstractSet[str]) -> bool:
    res = self.element_constr.can_infer(var_constraint_names) and (
        self.v1_shape_constr.name in var_constraint_names
        and self.v2_shape_constr.name in var_constraint_names
        and self.mask_constraint.name in var_constraint_names
    )
    assert res
    return res

infer(context: ConstraintContext) -> VectorType

Source code in xdsl/dialects/vector.py

def infer(self, context: ConstraintContext) -> VectorType:
    v1_shape = context.get_variable(self.v1_shape_constr.name)
    v2_shape = context.get_variable(self.v2_shape_constr.name)
    mask = context.get_variable(self.mask_constraint.name)
    assert v1_shape is not None
    assert v2_shape is not None
    assert mask is not None
    assert _IntArrayConstr.verifies(v1_shape)
    assert _IntArrayConstr.verifies(v2_shape)
    assert _MaskConstr.verifies(mask)

    result_trailing: tuple[IntAttr, ...]
    if not v1_shape:
        assert not v2_shape
        result_trailing = ()
    else:
        result_trailing = v1_shape.data[1:]

    element_type = self.element_constr.infer(context)
    result_leading = len(mask)
    shape = (
        (IntAttr(result_leading), *result_trailing)
        if result_leading
        else result_trailing
    )
    return VectorType(element_type, ArrayAttr(shape))

mapping_type_vars(type_var_mapping: Mapping[TypeVar, AttrConstraint | IntConstraint]) -> AttrConstraint[VectorType]

Source code in xdsl/dialects/vector.py

def mapping_type_vars(
    self, type_var_mapping: Mapping[TypeVar, AttrConstraint | IntConstraint]
) -> AttrConstraint[VectorType]:
    return ShuffleResultConstraint(
        self.element_constr.mapping_type_vars(type_var_mapping),
        self.v1_shape_constr.mapping_type_vars(type_var_mapping),
        self.v2_shape_constr.mapping_type_vars(type_var_mapping),
        self.mask_constraint.mapping_type_vars(type_var_mapping),
    )

ShuffleOp

Bases: IRDLOperation

The shuffle operation constructs a permutation (or duplication) of elements from two input vectors, returning a vector with the same element type as the input and a length that is the same as the shuffle mask. The two input vectors must have the same element type, same rank , and trailing dimension sizes and shuffles their values in the leading dimension (which may differ in size) according to the given mask. The legality rules are: * the two operands must have the same element type as the result - Either, the two operands and the result must have the same rank and trailing dimension sizes, viz. given two k-D operands v1 : and v2 : we have s_i = t_i for all 1 < i <= k - Or, the two operands must be 0-D vectors and the result is a 1-D vector. * the mask length equals the leading dimension size of the result * numbering the input vector indices left to right across the operands, all mask values must be within range, viz. given two k-D operands v1 and v2 above, all mask values are in the range [0,s_1+t_1)

Note, scalable vectors are not supported.

Example:

%0 = vector.shuffle %a, %a [0, 3]
            : vector<2xf32>, vector<2xf32>       ; yields vector<2xf32>
%1 = vector.shuffle %c, %b [0, 1, 2]
            : vector<2x16xf32>, vector<1x16xf32> ; yields vector<3x16xf32>
%2 = vector.shuffle %a, %a [3, 2, 1, 0]
             : vector<2xf32>, vector<2xf32>      ; yields vector<4xf32>
%3 = vector.shuffle %d, %d [0, 1]
            : vector<f32>, vector<f32>           ; yields vector<2xf32>

See external documentation.

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class ShuffleOp(IRDLOperation):
    """
    The shuffle operation constructs a permutation (or duplication) of elements
    from two input vectors, returning a vector with the same element type as
    the input and a length that is the same as the shuffle mask. The two input
    vectors must have the same element type, same rank , and trailing dimension
    sizes and shuffles their values in the
    leading dimension (which may differ in size) according to the given mask.
    The legality rules are:
    * the two operands must have the same element type as the result
      - Either, the two operands and the result must have the same
        rank and trailing dimension sizes, viz. given two k-D operands
                v1 : <s_1 x s_2 x .. x s_k x type> and
                v2 : <t_1 x t_2 x .. x t_k x type>
        we have s_i = t_i for all 1 < i <= k
      - Or, the two operands must be 0-D vectors and the result is a 1-D vector.
    * the mask length equals the leading dimension size of the result
    * numbering the input vector indices left to right across the operands, all
      mask values must be within range, viz. given two k-D operands v1 and v2
      above, all mask values are in the range [0,s_1+t_1)

    Note, scalable vectors are not supported.

    Example:

    ```mlir
    %0 = vector.shuffle %a, %a [0, 3]
                : vector<2xf32>, vector<2xf32>       ; yields vector<2xf32>
    %1 = vector.shuffle %c, %b [0, 1, 2]
                : vector<2x16xf32>, vector<1x16xf32> ; yields vector<3x16xf32>
    %2 = vector.shuffle %a, %a [3, 2, 1, 0]
                 : vector<2xf32>, vector<2xf32>      ; yields vector<4xf32>
    %3 = vector.shuffle %d, %d [0, 1]
                : vector<f32>, vector<f32>           ; yields vector<2xf32>
    ```

    See external [documentation](https://mlir.llvm.org/docs/Dialects/Vector/#vectorshuffle-vectorshuffleop).
    """

    name = "vector.shuffle"

    T: ClassVar = VarConstraint("T", AnyAttr())
    V1_SHAPE: ClassVar = VarConstraint("V1_SHAPE", _IntArrayConstr)
    V2_SHAPE: ClassVar = VarConstraint("V2_SHAPE", _IntArrayConstr)
    MASK: ClassVar = VarConstraint("MASK", _MaskConstr)
    RES: ClassVar = ShuffleResultConstraint(T, V1_SHAPE, V2_SHAPE, MASK)

    v1 = operand_def(VectorType.constr(T, shape=V1_SHAPE))
    v2 = operand_def(VectorType.constr(T, shape=V2_SHAPE))
    mask = prop_def(MASK)
    result = result_def(RES)

    irdl_options = (ParsePropInAttrDict(),)
    traits = traits_def(NoMemoryEffect())

    assembly_format = "operands $mask attr-dict `:` type(operands)"

    def __init__(
        self,
        v1: SSAValue,
        v2: SSAValue,
        mask: DenseArrayBase[I64],
        *,
        result_type: VectorType,
    ):
        super().__init__(
            operands=(v1, v2),
            result_types=(result_type,),
            properties={"mask": mask},
        )

    def verify_(self):
        assert isa(self.v1.type, VectorType)
        assert isa(self.v2.type, VectorType)
        assert isa(self.result.type, VectorType)

        v1_shape = self.v1.type.get_shape()
        v2_shape = self.v2.type.get_shape()
        result_shape = self.result.type.get_shape()
        mask = self.mask.get_values()

        result_leading_dim = result_shape[0]

        if len(mask) != result_leading_dim:
            # the mask length equals the leading dimension size of the result
            raise VerifyException(
                f"Length of mask {self.mask} must equal leading dim of result {self.result.type}."
            )

        if not v1_shape or not v2_shape:
            if v1_shape or v2_shape:
                raise VerifyException(
                    "Inputs must either both be non-0-D or both be 0-D"
                )

            if len(result_shape) != 1:
                raise VerifyException("If inputs are 0-D output must be 1-D")

            v1_leading_dim = 1
            v2_leading_dim = 1
        else:
            v1_leading_dim, *v1_trailing = v1_shape
            v2_leading_dim, *v2_trailing = v2_shape

            if v1_trailing != v2_trailing:
                raise VerifyException("Input trailing dimensions must match")

        dim_bound = v1_leading_dim + v2_leading_dim
        for dim in mask:
            if not (-1 <= dim < dim_bound):
                raise VerifyException(
                    f"Mask value {dim} out of range [-1, {dim_bound})"
                )

name = 'vector.shuffle' class-attribute instance-attribute

T: ClassVar = VarConstraint('T', AnyAttr()) class-attribute instance-attribute

V1_SHAPE: ClassVar = VarConstraint('V1_SHAPE', _IntArrayConstr) class-attribute instance-attribute

V2_SHAPE: ClassVar = VarConstraint('V2_SHAPE', _IntArrayConstr) class-attribute instance-attribute

MASK: ClassVar = VarConstraint('MASK', _MaskConstr) class-attribute instance-attribute

RES: ClassVar = ShuffleResultConstraint(T, V1_SHAPE, V2_SHAPE, MASK) class-attribute instance-attribute

v1 = operand_def(VectorType.constr(T, shape=V1_SHAPE)) class-attribute instance-attribute

v2 = operand_def(VectorType.constr(T, shape=V2_SHAPE)) class-attribute instance-attribute

mask = prop_def(MASK) class-attribute instance-attribute

result = result_def(RES) class-attribute instance-attribute

irdl_options = (ParsePropInAttrDict(),) class-attribute instance-attribute

traits = traits_def(NoMemoryEffect()) class-attribute instance-attribute

assembly_format = 'operands $mask attr-dict `:` type(operands)' class-attribute instance-attribute

__init__(v1: SSAValue, v2: SSAValue, mask: DenseArrayBase[I64], *, result_type: VectorType)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    v1: SSAValue,
    v2: SSAValue,
    mask: DenseArrayBase[I64],
    *,
    result_type: VectorType,
):
    super().__init__(
        operands=(v1, v2),
        result_types=(result_type,),
        properties={"mask": mask},
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    assert isa(self.v1.type, VectorType)
    assert isa(self.v2.type, VectorType)
    assert isa(self.result.type, VectorType)

    v1_shape = self.v1.type.get_shape()
    v2_shape = self.v2.type.get_shape()
    result_shape = self.result.type.get_shape()
    mask = self.mask.get_values()

    result_leading_dim = result_shape[0]

    if len(mask) != result_leading_dim:
        # the mask length equals the leading dimension size of the result
        raise VerifyException(
            f"Length of mask {self.mask} must equal leading dim of result {self.result.type}."
        )

    if not v1_shape or not v2_shape:
        if v1_shape or v2_shape:
            raise VerifyException(
                "Inputs must either both be non-0-D or both be 0-D"
            )

        if len(result_shape) != 1:
            raise VerifyException("If inputs are 0-D output must be 1-D")

        v1_leading_dim = 1
        v2_leading_dim = 1
    else:
        v1_leading_dim, *v1_trailing = v1_shape
        v2_leading_dim, *v2_trailing = v2_shape

        if v1_trailing != v2_trailing:
            raise VerifyException("Input trailing dimensions must match")

    dim_bound = v1_leading_dim + v2_leading_dim
    for dim in mask:
        if not (-1 <= dim < dim_bound):
            raise VerifyException(
                f"Mask value {dim} out of range [-1, {dim_bound})"
            )

BroadcastOp

Bases: IRDLOperation

See external documentation.

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class BroadcastOp(IRDLOperation):
    """
    See external [documentation](https://mlir.llvm.org/docs/Dialects/Vector/#vectorbroadcast-vectorbroadcastop).
    """

    name = "vector.broadcast"
    source = operand_def()
    vector = result_def(VectorType)
    traits = traits_def(Pure())

    assembly_format = "$source attr-dict `:` type($source) `to` type($vector)"

    def __init__(self, source: Operation | SSAValue, result_type: VectorType):
        super().__init__(operands=(source,), result_types=(result_type,))

    def verify_(self):
        if isa(self.source.type, VectorType):
            element_type = self.source.type.element_type
        else:
            element_type = self.source.type

        if element_type != self.vector.type.element_type:
            raise VerifyException(
                "Source operand and result vector must have the same element type."
            )

name = 'vector.broadcast' class-attribute instance-attribute

source = operand_def() class-attribute instance-attribute

vector = result_def(VectorType) class-attribute instance-attribute

traits = traits_def(Pure()) class-attribute instance-attribute

assembly_format = '$source attr-dict `:` type($source) `to` type($vector)' class-attribute instance-attribute

__init__(source: Operation | SSAValue, result_type: VectorType)

Source code in xdsl/dialects/vector.py

def __init__(self, source: Operation | SSAValue, result_type: VectorType):
    super().__init__(operands=(source,), result_types=(result_type,))

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    if isa(self.source.type, VectorType):
        element_type = self.source.type.element_type
    else:
        element_type = self.source.type

    if element_type != self.vector.type.element_type:
        raise VerifyException(
            "Source operand and result vector must have the same element type."
        )

FMAOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class FMAOp(IRDLOperation):
    name = "vector.fma"

    T: ClassVar = VarConstraint("T", VectorType.constr(AnyFloatConstr))

    lhs = operand_def(T)
    rhs = operand_def(T)
    acc = operand_def(T)
    res = result_def(T)
    traits = traits_def(Pure())

    assembly_format = "$lhs `,` $rhs `,` $acc attr-dict `:` type($lhs)"

    def __init__(
        self,
        lhs: Operation | SSAValue,
        rhs: Operation | SSAValue,
        acc: Operation | SSAValue,
    ):
        acc = SSAValue.get(acc)
        super().__init__(operands=(lhs, rhs, acc), result_types=(acc.type,))

name = 'vector.fma' class-attribute instance-attribute

T: ClassVar = VarConstraint('T', VectorType.constr(AnyFloatConstr)) class-attribute instance-attribute

lhs = operand_def(T) class-attribute instance-attribute

rhs = operand_def(T) class-attribute instance-attribute

acc = operand_def(T) class-attribute instance-attribute

res = result_def(T) class-attribute instance-attribute

traits = traits_def(Pure()) class-attribute instance-attribute

assembly_format = '$lhs `,` $rhs `,` $acc attr-dict `:` type($lhs)' class-attribute instance-attribute

__init__(lhs: Operation | SSAValue, rhs: Operation | SSAValue, acc: Operation | SSAValue)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    lhs: Operation | SSAValue,
    rhs: Operation | SSAValue,
    acc: Operation | SSAValue,
):
    acc = SSAValue.get(acc)
    super().__init__(operands=(lhs, rhs, acc), result_types=(acc.type,))

MaskedLoadOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class MaskedLoadOp(IRDLOperation):
    name = "vector.maskedload"
    base = operand_def(MemRefType)
    indices = var_operand_def(IndexType)
    mask = operand_def(VectorBaseTypeAndRankConstraint(i1, 1))
    pass_thru = operand_def(VectorType)
    result = result_def(VectorRankConstraint(1))

    assembly_format = "$base `[` $indices `]` `,` $mask `,` $pass_thru attr-dict `:` type($base) `,` type($mask) `,` type($pass_thru) `into` type($result)"  # noqa: E501

    def __init__(
        self,
        base: SSAValue | Operation,
        indices: Sequence[SSAValue | Operation],
        mask: SSAValue | Operation,
        pass_thru: SSAValue | Operation,
        result_type: VectorType | None = None,
    ):
        pass_thru = SSAValue.get(pass_thru, type=VectorType)
        if result_type is None:
            result_type = pass_thru.type
        super().__init__(
            operands=[base, indices, mask, pass_thru],
            result_types=[result_type],
        )

    def verify_(self):
        memref_type = self.base.type
        assert isa(memref_type, MemRefType)
        memref_element_type = memref_type.element_type

        res_type = self.result.type
        assert isa(res_type, VectorType[Attribute])
        res_element_type = res_type.element_type

        passthrough_type = self.pass_thru.type
        assert isa(passthrough_type, VectorType[Attribute])
        passthrough_element_type = passthrough_type.element_type

        if memref_element_type != res_element_type:
            raise VerifyException(
                "MemRef element type should match the result vector and passthrough vector "
                "element type. Found different element types for memref and result."
            )
        elif memref_element_type != passthrough_element_type:
            raise VerifyException(
                "MemRef element type should match the result vector and passthrough vector "
                "element type. Found different element types for memref and passthrough."
            )

        if memref_type.get_num_dims() != len(self.indices):
            raise VerifyException("Expected an index for each memref dimension.")

name = 'vector.maskedload' class-attribute instance-attribute

base = operand_def(MemRefType) class-attribute instance-attribute

indices = var_operand_def(IndexType) class-attribute instance-attribute

mask = operand_def(VectorBaseTypeAndRankConstraint(i1, 1)) class-attribute instance-attribute

pass_thru = operand_def(VectorType) class-attribute instance-attribute

result = result_def(VectorRankConstraint(1)) class-attribute instance-attribute

assembly_format = '$base `[` $indices `]` `,` $mask `,` $pass_thru attr-dict `:` type($base) `,` type($mask) `,` type($pass_thru) `into` type($result)' class-attribute instance-attribute

__init__(base: SSAValue | Operation, indices: Sequence[SSAValue | Operation], mask: SSAValue | Operation, pass_thru: SSAValue | Operation, result_type: VectorType | None = None)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    base: SSAValue | Operation,
    indices: Sequence[SSAValue | Operation],
    mask: SSAValue | Operation,
    pass_thru: SSAValue | Operation,
    result_type: VectorType | None = None,
):
    pass_thru = SSAValue.get(pass_thru, type=VectorType)
    if result_type is None:
        result_type = pass_thru.type
    super().__init__(
        operands=[base, indices, mask, pass_thru],
        result_types=[result_type],
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    memref_type = self.base.type
    assert isa(memref_type, MemRefType)
    memref_element_type = memref_type.element_type

    res_type = self.result.type
    assert isa(res_type, VectorType[Attribute])
    res_element_type = res_type.element_type

    passthrough_type = self.pass_thru.type
    assert isa(passthrough_type, VectorType[Attribute])
    passthrough_element_type = passthrough_type.element_type

    if memref_element_type != res_element_type:
        raise VerifyException(
            "MemRef element type should match the result vector and passthrough vector "
            "element type. Found different element types for memref and result."
        )
    elif memref_element_type != passthrough_element_type:
        raise VerifyException(
            "MemRef element type should match the result vector and passthrough vector "
            "element type. Found different element types for memref and passthrough."
        )

    if memref_type.get_num_dims() != len(self.indices):
        raise VerifyException("Expected an index for each memref dimension.")

MaskedStoreOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class MaskedStoreOp(IRDLOperation):
    name = "vector.maskedstore"
    base = operand_def(MemRefType)
    indices = var_operand_def(IndexType)
    mask = operand_def(VectorBaseTypeAndRankConstraint(i1, 1))
    value_to_store = operand_def(VectorRankConstraint(1))

    assembly_format = "$base `[` $indices `]` `,` $mask `,` $value_to_store attr-dict `:` type($base) `,` type($mask) `,` type($value_to_store)"  # noqa: E501

    def verify_(self):
        memref_type = self.base.type
        assert isa(memref_type, MemRefType)
        memref_element_type = memref_type.element_type

        value_to_store_type = self.value_to_store.type
        assert isa(value_to_store_type, VectorType[Attribute])

        mask_type = self.mask.type
        assert isa(mask_type, VectorType[Attribute])

        if memref_element_type != value_to_store_type.element_type:
            raise VerifyException(
                "MemRef element type should match the stored vector type. "
                "Obtained types were "
                + str(memref_element_type)
                + " and "
                + str(value_to_store_type.element_type)
                + "."
            )

        if memref_type.get_num_dims() != len(self.indices):
            raise VerifyException("Expected an index for each memref dimension.")

    def __init__(
        self,
        memref: SSAValue | Operation,
        indices: Sequence[SSAValue | Operation],
        mask: SSAValue | Operation,
        value_to_store: SSAValue | Operation,
    ):
        super().__init__(operands=[memref, indices, mask, value_to_store])

name = 'vector.maskedstore' class-attribute instance-attribute

base = operand_def(MemRefType) class-attribute instance-attribute

indices = var_operand_def(IndexType) class-attribute instance-attribute

mask = operand_def(VectorBaseTypeAndRankConstraint(i1, 1)) class-attribute instance-attribute

value_to_store = operand_def(VectorRankConstraint(1)) class-attribute instance-attribute

assembly_format = '$base `[` $indices `]` `,` $mask `,` $value_to_store attr-dict `:` type($base) `,` type($mask) `,` type($value_to_store)' class-attribute instance-attribute

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    memref_type = self.base.type
    assert isa(memref_type, MemRefType)
    memref_element_type = memref_type.element_type

    value_to_store_type = self.value_to_store.type
    assert isa(value_to_store_type, VectorType[Attribute])

    mask_type = self.mask.type
    assert isa(mask_type, VectorType[Attribute])

    if memref_element_type != value_to_store_type.element_type:
        raise VerifyException(
            "MemRef element type should match the stored vector type. "
            "Obtained types were "
            + str(memref_element_type)
            + " and "
            + str(value_to_store_type.element_type)
            + "."
        )

    if memref_type.get_num_dims() != len(self.indices):
        raise VerifyException("Expected an index for each memref dimension.")

__init__(memref: SSAValue | Operation, indices: Sequence[SSAValue | Operation], mask: SSAValue | Operation, value_to_store: SSAValue | Operation)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    memref: SSAValue | Operation,
    indices: Sequence[SSAValue | Operation],
    mask: SSAValue | Operation,
    value_to_store: SSAValue | Operation,
):
    super().__init__(operands=[memref, indices, mask, value_to_store])

PrintOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class PrintOp(IRDLOperation):
    name = "vector.print"
    source = operand_def()

    def __init__(self, source: SSAValue | Operation):
        super().__init__(operands=[SSAValue.get(source)])

name = 'vector.print' class-attribute instance-attribute

source = operand_def() class-attribute instance-attribute

__init__(source: SSAValue | Operation)

Source code in xdsl/dialects/vector.py

def __init__(self, source: SSAValue | Operation):
    super().__init__(operands=[SSAValue.get(source)])

CreateMaskOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class CreateMaskOp(IRDLOperation):
    name = "vector.create_mask"
    mask_dim_sizes = var_operand_def(IndexType)
    mask_vector = result_def(VectorBaseTypeConstraint(i1))

    assembly_format = "$mask_dim_sizes attr-dict `:` type(results)"

    def __init__(
        self, mask_operands: list[Operation | SSAValue], result_type: VectorType
    ):
        super().__init__(operands=(mask_operands,), result_types=(result_type,))

    def verify_(self):
        assert isa(self.mask_vector.type, VectorType[Attribute])
        if self.mask_vector.type.get_num_dims() != len(self.mask_dim_sizes):
            raise VerifyException(
                "Expected an operand value for each dimension of resultant mask."
            )

name = 'vector.create_mask' class-attribute instance-attribute

mask_dim_sizes = var_operand_def(IndexType) class-attribute instance-attribute

mask_vector = result_def(VectorBaseTypeConstraint(i1)) class-attribute instance-attribute

assembly_format = '$mask_dim_sizes attr-dict `:` type(results)' class-attribute instance-attribute

__init__(mask_operands: list[Operation | SSAValue], result_type: VectorType)

Source code in xdsl/dialects/vector.py

def __init__(
    self, mask_operands: list[Operation | SSAValue], result_type: VectorType
):
    super().__init__(operands=(mask_operands,), result_types=(result_type,))

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    assert isa(self.mask_vector.type, VectorType[Attribute])
    if self.mask_vector.type.get_num_dims() != len(self.mask_dim_sizes):
        raise VerifyException(
            "Expected an operand value for each dimension of resultant mask."
        )

ExtractOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class ExtractOp(IRDLOperation):
    name = "vector.extract"

    _T: ClassVar = VarConstraint(
        "T", base(IntegerType) | base(IndexType) | AnyFloatConstr
    )
    _V: ClassVar = VarConstraint("V", VectorType.constr(_T))

    static_position = prop_def(DenseArrayBase.constr(i64))

    vector = operand_def(_V)
    dynamic_position = var_operand_def(IndexTypeConstr)

    result = result_def(
        VectorType.constr(
            _T,
            shape=MessageConstraint(
                ArrayAttr.constr(RangeOf(base(IntAttr)).of_length(AtLeast(1))),
                "Cannot extract 0d vector.",
            ),
        )
        | _T
    )

    traits = traits_def(Pure())

    DYNAMIC_INDEX: ClassVar = DYNAMIC_INDEX
    """This value is used to indicate that a position is a dynamic index."""

    assembly_format = (
        "$vector `` custom<DynamicIndexList>($dynamic_position, $static_position)"
        " attr-dict `:` type($result) `from` type($vector)"
    )

    custom_directives = (DynamicIndexList,)

    def get_mixed_position(self) -> list[SSAValue | int]:
        """
        Returns the list of positions, represented as either an SSAValue or an int
        """
        static_positions = self.static_position.get_values()
        return get_dynamic_index_list(
            static_positions,
            self.dynamic_position,
            ExtractOp.DYNAMIC_INDEX,
        )

    def verify_(self):
        # Check that static position attribute and dynamic position operands
        # are compatible.
        static_values = self.static_position.get_values()
        verify_dynamic_index_list(
            static_values,
            self.dynamic_position,
            self.DYNAMIC_INDEX,
        )

        num_indices = len(self.static_position)
        vector_type = self.vector.type
        assert isa(vector_type, VectorType[Attribute])
        # Check that the number of dimensions match
        if isa(self.result.type, VectorType):
            if (
                num_indices + self.result.type.get_num_dims()
                != vector_type.get_num_dims()
            ):
                raise VerifyException(
                    f"Expected position attribute rank ({num_indices}) + result rank "
                    f"({self.result.type.get_num_dims()}) to "
                    f"match source vector rank ({vector_type.get_num_dims()})."
                )
        else:
            if num_indices != vector_type.get_num_dims():
                raise VerifyException(
                    f"Expected position attribute rank ({num_indices}) to match "
                    f"source vector rank ({vector_type.get_num_dims()})."
                )

    def __init__(
        self,
        vector: SSAValue,
        positions: Sequence[SSAValue | int],
        result_type: Attribute,
    ):
        static_positions, dynamic_positions = split_dynamic_index_list(
            positions, ExtractOp.DYNAMIC_INDEX
        )

        super().__init__(
            operands=[vector, dynamic_positions],
            result_types=[result_type],
            properties={
                "static_position": DenseArrayBase.from_list(i64, static_positions)
            },
        )

name = 'vector.extract' class-attribute instance-attribute

static_position = prop_def(DenseArrayBase.constr(i64)) class-attribute instance-attribute

vector = operand_def(_V) class-attribute instance-attribute

dynamic_position = var_operand_def(IndexTypeConstr) class-attribute instance-attribute

result = result_def(VectorType.constr(_T, shape=(MessageConstraint(ArrayAttr.constr(RangeOf(base(IntAttr)).of_length(AtLeast(1))), 'Cannot extract 0d vector.'))) | _T) class-attribute instance-attribute

traits = traits_def(Pure()) class-attribute instance-attribute

DYNAMIC_INDEX: ClassVar = DYNAMIC_INDEX class-attribute instance-attribute

This value is used to indicate that a position is a dynamic index.

assembly_format = '$vector `` custom<DynamicIndexList>($dynamic_position, $static_position) attr-dict `:` type($result) `from` type($vector)' class-attribute instance-attribute

custom_directives = (DynamicIndexList,) class-attribute instance-attribute

get_mixed_position() -> list[SSAValue | int]

Returns the list of positions, represented as either an SSAValue or an int

Source code in xdsl/dialects/vector.py

def get_mixed_position(self) -> list[SSAValue | int]:
    """
    Returns the list of positions, represented as either an SSAValue or an int
    """
    static_positions = self.static_position.get_values()
    return get_dynamic_index_list(
        static_positions,
        self.dynamic_position,
        ExtractOp.DYNAMIC_INDEX,
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    # Check that static position attribute and dynamic position operands
    # are compatible.
    static_values = self.static_position.get_values()
    verify_dynamic_index_list(
        static_values,
        self.dynamic_position,
        self.DYNAMIC_INDEX,
    )

    num_indices = len(self.static_position)
    vector_type = self.vector.type
    assert isa(vector_type, VectorType[Attribute])
    # Check that the number of dimensions match
    if isa(self.result.type, VectorType):
        if (
            num_indices + self.result.type.get_num_dims()
            != vector_type.get_num_dims()
        ):
            raise VerifyException(
                f"Expected position attribute rank ({num_indices}) + result rank "
                f"({self.result.type.get_num_dims()}) to "
                f"match source vector rank ({vector_type.get_num_dims()})."
            )
    else:
        if num_indices != vector_type.get_num_dims():
            raise VerifyException(
                f"Expected position attribute rank ({num_indices}) to match "
                f"source vector rank ({vector_type.get_num_dims()})."
            )

__init__(vector: SSAValue, positions: Sequence[SSAValue | int], result_type: Attribute)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    vector: SSAValue,
    positions: Sequence[SSAValue | int],
    result_type: Attribute,
):
    static_positions, dynamic_positions = split_dynamic_index_list(
        positions, ExtractOp.DYNAMIC_INDEX
    )

    super().__init__(
        operands=[vector, dynamic_positions],
        result_types=[result_type],
        properties={
            "static_position": DenseArrayBase.from_list(i64, static_positions)
        },
    )

InsertOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class InsertOp(IRDLOperation):
    name = "vector.insert"

    _T: ClassVar = VarConstraint(
        "T", base(IntegerType) | base(IndexType) | AnyFloatConstr
    )
    _V: ClassVar = VarConstraint("V", VectorType.constr(_T))

    static_position = prop_def(DenseArrayBase.constr(i64))

    source = operand_def(
        VectorType.constr(
            _T,
            shape=MessageConstraint(
                ArrayAttr.constr(RangeOf(base(IntAttr)).of_length(AtLeast(1))),
                "Cannot insert 0d vector.",
            ),
        )
        | _T
    )
    dest = operand_def(_V)
    dynamic_position = var_operand_def(IndexTypeConstr)

    result = result_def(_V)

    traits = traits_def(Pure())

    DYNAMIC_INDEX: ClassVar = -(2**63)
    """This value is used to indicate that a position is a dynamic index."""

    assembly_format = (
        "$source `,` $dest custom<DynamicIndexList>($dynamic_position, $static_position)"
        "attr-dict `:` type($source) `into` type($dest)"
    )

    custom_directives = (DynamicIndexList,)

    def get_mixed_position(self) -> list[SSAValue | int]:
        """
        Returns the list of positions, represented as either an SSAValue or an int.
        """
        static_positions = self.static_position.get_values()
        return get_dynamic_index_list(
            static_positions,
            self.dynamic_position,
            InsertOp.DYNAMIC_INDEX,
        )

    def verify_(self):
        # Check that static position attribute and dynamic position operands
        # are compatible.
        static_values = self.static_position.get_values()
        verify_dynamic_index_list(
            static_values,
            self.dynamic_position,
            self.DYNAMIC_INDEX,
        )

        num_indices = len(self.static_position)
        # Check that the number of dimensions match
        if isa(self.source.type, VectorType):
            if (
                num_indices + self.source.type.get_num_dims()
                != self.result.type.get_num_dims()
            ):
                raise VerifyException(
                    f"Expected position attribute rank ({num_indices}) + source rank "
                    f"({self.source.type.get_num_dims()}) to "
                    f"match dest vector rank ({self.result.type.get_num_dims()})."
                )
        else:
            if num_indices != self.result.type.get_num_dims():
                raise VerifyException(
                    f"Expected position attribute rank ({num_indices}) to match "
                    f"dest vector rank ({self.result.type.get_num_dims()})."
                )

    def __init__(
        self,
        source: SSAValue,
        dest: SSAValue,
        positions: Sequence[SSAValue | int],
        result_type: Attribute | None = None,
    ):
        static_positions, dynamic_positions = split_dynamic_index_list(
            positions, InsertOp.DYNAMIC_INDEX
        )

        if result_type is None:
            result_type = dest.type

        super().__init__(
            operands=[source, dest, dynamic_positions],
            result_types=[result_type],
            properties={
                "static_position": DenseArrayBase.from_list(i64, static_positions)
            },
        )

name = 'vector.insert' class-attribute instance-attribute

static_position = prop_def(DenseArrayBase.constr(i64)) class-attribute instance-attribute

source = operand_def(VectorType.constr(_T, shape=(MessageConstraint(ArrayAttr.constr(RangeOf(base(IntAttr)).of_length(AtLeast(1))), 'Cannot insert 0d vector.'))) | _T) class-attribute instance-attribute

dest = operand_def(_V) class-attribute instance-attribute

dynamic_position = var_operand_def(IndexTypeConstr) class-attribute instance-attribute

result = result_def(_V) class-attribute instance-attribute

traits = traits_def(Pure()) class-attribute instance-attribute

DYNAMIC_INDEX: ClassVar = -2 ** 63 class-attribute instance-attribute

This value is used to indicate that a position is a dynamic index.

assembly_format = '$source `,` $dest custom<DynamicIndexList>($dynamic_position, $static_position)attr-dict `:` type($source) `into` type($dest)' class-attribute instance-attribute

custom_directives = (DynamicIndexList,) class-attribute instance-attribute

get_mixed_position() -> list[SSAValue | int]

Returns the list of positions, represented as either an SSAValue or an int.

Source code in xdsl/dialects/vector.py

def get_mixed_position(self) -> list[SSAValue | int]:
    """
    Returns the list of positions, represented as either an SSAValue or an int.
    """
    static_positions = self.static_position.get_values()
    return get_dynamic_index_list(
        static_positions,
        self.dynamic_position,
        InsertOp.DYNAMIC_INDEX,
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    # Check that static position attribute and dynamic position operands
    # are compatible.
    static_values = self.static_position.get_values()
    verify_dynamic_index_list(
        static_values,
        self.dynamic_position,
        self.DYNAMIC_INDEX,
    )

    num_indices = len(self.static_position)
    # Check that the number of dimensions match
    if isa(self.source.type, VectorType):
        if (
            num_indices + self.source.type.get_num_dims()
            != self.result.type.get_num_dims()
        ):
            raise VerifyException(
                f"Expected position attribute rank ({num_indices}) + source rank "
                f"({self.source.type.get_num_dims()}) to "
                f"match dest vector rank ({self.result.type.get_num_dims()})."
            )
    else:
        if num_indices != self.result.type.get_num_dims():
            raise VerifyException(
                f"Expected position attribute rank ({num_indices}) to match "
                f"dest vector rank ({self.result.type.get_num_dims()})."
            )

__init__(source: SSAValue, dest: SSAValue, positions: Sequence[SSAValue | int], result_type: Attribute | None = None)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    source: SSAValue,
    dest: SSAValue,
    positions: Sequence[SSAValue | int],
    result_type: Attribute | None = None,
):
    static_positions, dynamic_positions = split_dynamic_index_list(
        positions, InsertOp.DYNAMIC_INDEX
    )

    if result_type is None:
        result_type = dest.type

    super().__init__(
        operands=[source, dest, dynamic_positions],
        result_types=[result_type],
        properties={
            "static_position": DenseArrayBase.from_list(i64, static_positions)
        },
    )

VectorTransferOperation dataclass

Bases: IRDLOperation, ABC

Encodes properties of a vector.transfer_read or vector.transfer_write operation. Vector transfer ops have:

• A shaped value that the op reads from/writes to: a memref or a tensor.
• A vector, either as a result or as an operand.
• Indices that describe where the transfer from/to the shaped value starts.
• An optional mask.
• An optional in_bounds array to indicate transfer dimensions that are guaranteed to be in-bounds.
• A permutation map to indicate transposes and broadcasts.

The "vector rank" is the rank of the vector type. E.g.:

// Transfer with shaped value rank 2 and vector (transfer) rank 1.
%0 = vector.transfer_read %arg0[%c3, %c3], %f0
    {permutation_map = affine_map<(d0, d1) -> (d0)>}
    : memref<?x?xf32>, vector<128xf32>

The "vector transfer rank" is the number of dimensions that participate in the transfer and broadcasts, and matches the number of results in the permutation map. In most cases, the vector rank matches the vector transfer rank; the only exception is when a vector is flattened as part of the transfer (see permutation_map).

Mirrors VectorTransferOpInterface from MLIR

Source code in xdsl/dialects/vector.py

class VectorTransferOperation(IRDLOperation, ABC):
    """
    Encodes properties of a `vector.transfer_read` or `vector.transfer_write`
    operation. Vector transfer ops have:

    - A shaped value that the op reads from/writes to: a memref or a tensor.
    - A vector, either as a result or as an operand.
    - Indices that describe where the transfer from/to the shaped value starts.
    - An optional mask.
    - An optional in_bounds array to indicate transfer dimensions that are
      guaranteed to be in-bounds.
    - A permutation map to indicate transposes and broadcasts.

    The "vector rank" is the rank of the vector type. E.g.:
    ```mlir
    // Transfer with shaped value rank 2 and vector (transfer) rank 1.
    %0 = vector.transfer_read %arg0[%c3, %c3], %f0
        {permutation_map = affine_map<(d0, d1) -> (d0)>}
        : memref<?x?xf32>, vector<128xf32>
    ```

    The "vector transfer rank" is the number of dimensions that participate in
    the transfer and broadcasts, and matches the number of results in the
    permutation map. In most cases, the vector rank matches the vector transfer
    rank; the only exception is when a vector is flattened as part of the
    transfer (see `permutation_map`).

    Mirrors VectorTransferOpInterface from [MLIR](https://github.com/llvm/llvm-project/blob/main/mlir/include/mlir/Interfaces/VectorInterfaces.td)
    """

    permutation_map = prop_def(AffineMapAttr)
    """
    The permutation map that describes the mapping of vector
    dimensions to source dimensions, as well as broadcast dimensions.

    The permutation result has one result per vector transfer dimension.
    Each result is either a dim expression, indicating the corresponding
    dimension in the source operand, or a constant "0" expression,
    indicating a broadcast dimension.

    Note: Nested vector dimensions that are flattened by this op are not
    accounted for in the permutation map. E.g.:
    ```mlir
    // Vector type has rank 4, but permutation map has only 2 results. That
    // is because there are only 2 transfer dimensions.
    %0 = vector.transfer_read %arg1[%c3, %c3], %vf0
        {permutation_map = affine_map<(d0, d1) -> (d0, d1)>}
        : memref<?x?xvector<4x3xf32>>, vector<1x1x4x3xf32>
    ```
    """

    in_bounds = prop_def(ArrayAttr[BoolAttr])
    """
    For every vector dimension, the boolean array attribute `in_bounds` specifies if the
    transfer is guaranteed to be within the source bounds. If set to `“false”`, accesses
    (including the starting point) may run out-of-bounds along the respective vector
    dimension as the index increases. Non-vector dimensions must always be in-bounds.
    The `in_bounds` array length has to be equal to the vector rank. This attribute has
    a default value: `false` (i.e. “out-of-bounds”). When skipped in the textual IR, the
    default value is assumed. Similarly, the OP printer will omit this attribute when
    all dimensions are out-of-bounds (i.e. the default value is used).
    """

    @staticmethod
    def infer_transfer_op_mask_type(
        vec_type: VectorType, perm_map: AffineMap
    ) -> VectorType[I1]:
        """
        Given a resulting vector type and a permutation map from the dimensions of the
        shaped type to the vector type dimensions, return the vector type of the mask.
        """
        unused_dims_bit_vector = tuple(
            not dim for dim in perm_map.used_dims_bit_vector()
        )
        inv_perm_map = perm_map.drop_dims(unused_dims_bit_vector).inverse_permutation()
        assert inv_perm_map is not None, "Inversed permutation map couldn't be computed"
        mask_shape = inv_perm_map.eval(vec_type.get_shape(), ())
        scalable_dims = ArrayAttr(
            BoolAttr.from_bool(bool(b))
            for b in inv_perm_map.eval(vec_type.get_scalable_dims(), ())
        )
        res = VectorType(i1, mask_shape, scalable_dims)
        return res

    @staticmethod
    def get_transfer_minor_identity_map(
        shaped_type: TensorType | MemRefType, vector_type: VectorType
    ) -> AffineMap:
        """
        Get the minor identity map for a transfer operation.

        This is a helper function to compute the default permutation map for
        transfer operations when none is specified.
        """
        element_vector_rank = 0
        element_type = shaped_type.element_type
        if isa(element_type, VectorType):
            element_vector_rank += element_type.get_num_dims()

        # 0-d transfers are to/from tensor<t>/memref<t> and vector<1xt>.
        # TODO: replace once we have 0-d vectors.
        if shaped_type.get_num_dims() == 0 and vector_type.get_shape() == (1,):
            return AffineMap.constant_map(0)

        return AffineMap.minor_identity(
            shaped_type.get_num_dims(),
            vector_type.get_num_dims() - element_vector_rank,
        )

    def _print_attrs(self, printer: Printer):
        reserved_attr_names = {"operandSegmentSizes"}
        if self.permutation_map.data.is_minor_identity():
            reserved_attr_names.add("permutation_map")
        if not any(self.in_bounds):
            reserved_attr_names.add("in_bounds")
        printer.print_op_attributes(
            self.attributes | self.properties, reserved_attr_names=reserved_attr_names
        )

    @staticmethod
    def resolve_attrs(
        parser: Parser,
        attributes_dict: dict[str, Attribute],
        shaped_type: TensorType | MemRefType,
        vector_type: VectorType,
        mask_start_pos: Position | None,
        mask_end_pos: Position | None,
        mask: UnresolvedOperand | None,
        types_pos: Position,
    ):
        # Create default permutation_map if not provided in attributes
        permutation_map = None
        if attributes_dict and "permutation_map" in attributes_dict:
            permutation_map = attributes_dict["permutation_map"]
            assert isinstance(permutation_map, AffineMapAttr)
        else:
            # Create identity permutation map for the shaped type's rank
            permutation_map = AffineMapAttr(
                VectorTransferOperation.get_transfer_minor_identity_map(
                    shaped_type, vector_type
                )
            )

        # Create in_bounds attribute if not provided
        in_bounds = None
        if attributes_dict and "in_bounds" in attributes_dict:
            in_bounds = cast(ArrayAttr[BoolAttr], attributes_dict["in_bounds"])
        else:
            # Default: all dimensions are out-of-bounds
            in_bounds = ArrayAttr(
                (BoolAttr.from_bool(False),) * len(permutation_map.data.results)
            )

        if mask is not None:
            if isa(shaped_type.element_type, VectorType):
                assert mask_start_pos is not None
                assert mask_end_pos is not None
                parser.raise_error(
                    "does not support masks with vector element type",
                    at_position=mask_start_pos,
                    end_position=mask_end_pos,
                )
            if vector_type.get_num_dims() != len(permutation_map.data.results):
                parser.raise_error(
                    "expected the same rank for the vector and the "
                    "results of the permutation map",
                    types_pos,
                )
            # Instead of adding the mask type as an op type, compute it based on the
            # vector type and the permutation map (to keep the type signature small).
            mask_type = VectorTransferOperation.infer_transfer_op_mask_type(
                vector_type, permutation_map.data
            )
            resolved_mask = parser.resolve_operand(mask, mask_type)
        else:
            resolved_mask = None

        return resolved_mask, permutation_map, in_bounds

    def has_broadcast_dim(self):
        """
        Return "true" if at least one of the vector dimensions is a broadcasted dimension.
        """
        return any(
            isinstance(expr, AffineConstantExpr) and expr.value == 0
            for expr in self.permutation_map.data.results
        )

    @staticmethod
    def verify_op(
        op: TransferReadOp | TransferWriteOp,
        shaped_type: MemRefType | TensorType,
        vector_type: VectorType,
        mask_type: VectorType[I1] | None,
        inferred_mask_type: VectorType[I1] | None,
        permutation_map: AffineMap,
        in_bounds: ArrayAttr[BoolAttr],
    ):
        """
        This mirrors VectorOps.cpp -> verifyTransferOp from MLIR
        """

        element_type = shaped_type.element_type
        vector_element_type = vector_type.element_type

        if isa(element_type, VectorType):
            # Memref or tensor has vector element type
            # TODO verify vector element type
            pass
        else:
            # Memref of tensor has scalar element type
            if isa(vector_element_type, IndexType):
                if not isa(element_type, IndexType):
                    raise VerifyException(
                        "Element type of source is index, expected element type of vector also to be index"
                    )
            else:
                assert isa(vector_element_type, IntegerType | AnyFloat)
                assert isa(element_type, IntegerType | AnyFloat)

                minor_size = (
                    1
                    if vector_type.get_num_dims() == 0
                    else vector_type.get_shape()[-1]
                )
                result_vec_size = vector_element_type.bitwidth * minor_size
                if result_vec_size % element_type.bitwidth != 0:
                    raise VerifyException(
                        f'"{op.name}" requires the bitwidth of the minor 1-D vector to be '
                        "an integral multiple of the bitwidth of the source element type"
                    )

            # Check that permutation map results match rank of vector type.
            if len(permutation_map.results) != vector_type.get_num_dims():
                raise VerifyException(
                    f'"{op.name}" requires a permutation_map with result dims of the same rank as the vector type'
                )

        if permutation_map.num_symbols != 0:
            raise VerifyException(
                f'"{op.name}" requires permutation_map without symbols'
            )

        if permutation_map.num_dims != shaped_type.get_num_dims():
            raise VerifyException(
                f'"{op.name}" requires a permutation_map with input dims of the same rank as the source type'
            )

        if mask_type:
            if mask_type != inferred_mask_type:
                raise VerifyException(
                    f'"{op.name}" inferred mask type ({inferred_mask_type}) and mask operand type ({mask_type}) don\'t match'
                )

        if len(in_bounds) != len(permutation_map.results):
            raise VerifyException(
                f'"{op.name}" expects the in_bounds attr of same rank as permutation_map results: '
                f"{permutation_map!s} vs in_bounds of of size {len(in_bounds)}"
            )

    @staticmethod
    def verify_permutation_map(
        op: TransferReadOp | TransferWriteOp,
        permutation_map: AffineMap,
    ):
        """
        This mirrors VectorOps.cpp -> verifyPermutationMap
        """

        seen: list[bool] = [False for _ in range(permutation_map.num_dims)]

        for expr in permutation_map.results:
            if isa(expr, AffineConstantExpr):
                if expr.value != 0:
                    raise VerifyException(
                        f'"{op.name}" requires a projected permutation_map '
                        "(at most one dim or the zero constant can appear in each result)"
                    )
                continue
            if not isa(expr, AffineDimExpr):
                raise VerifyException(
                    f'"{op.name}" requires a projected permutation_map '
                    "(at most one dim or the zero constant can appear in each result)"
                )
            if seen[expr.position]:
                raise VerifyException(
                    f'"{op.name}" requires a permutation_map that is a permutation '
                    "(found one dim used more than once)"
                )
            seen[expr.position] = True

permutation_map = prop_def(AffineMapAttr) class-attribute instance-attribute

The permutation map that describes the mapping of vector dimensions to source dimensions, as well as broadcast dimensions.

The permutation result has one result per vector transfer dimension. Each result is either a dim expression, indicating the corresponding dimension in the source operand, or a constant "0" expression, indicating a broadcast dimension.

Note: Nested vector dimensions that are flattened by this op are not accounted for in the permutation map. E.g.:

// Vector type has rank 4, but permutation map has only 2 results. That
// is because there are only 2 transfer dimensions.
%0 = vector.transfer_read %arg1[%c3, %c3], %vf0
    {permutation_map = affine_map<(d0, d1) -> (d0, d1)>}
    : memref<?x?xvector<4x3xf32>>, vector<1x1x4x3xf32>

in_bounds = prop_def(ArrayAttr[BoolAttr]) class-attribute instance-attribute

For every vector dimension, the boolean array attribute in_bounds specifies if the transfer is guaranteed to be within the source bounds. If set to “false”, accesses (including the starting point) may run out-of-bounds along the respective vector dimension as the index increases. Non-vector dimensions must always be in-bounds. The in_bounds array length has to be equal to the vector rank. This attribute has a default value: false (i.e. “out-of-bounds”). When skipped in the textual IR, the default value is assumed. Similarly, the OP printer will omit this attribute when all dimensions are out-of-bounds (i.e. the default value is used).

infer_transfer_op_mask_type(vec_type: VectorType, perm_map: AffineMap) -> VectorType[I1] staticmethod

Given a resulting vector type and a permutation map from the dimensions of the shaped type to the vector type dimensions, return the vector type of the mask.

Source code in xdsl/dialects/vector.py

@staticmethod
def infer_transfer_op_mask_type(
    vec_type: VectorType, perm_map: AffineMap
) -> VectorType[I1]:
    """
    Given a resulting vector type and a permutation map from the dimensions of the
    shaped type to the vector type dimensions, return the vector type of the mask.
    """
    unused_dims_bit_vector = tuple(
        not dim for dim in perm_map.used_dims_bit_vector()
    )
    inv_perm_map = perm_map.drop_dims(unused_dims_bit_vector).inverse_permutation()
    assert inv_perm_map is not None, "Inversed permutation map couldn't be computed"
    mask_shape = inv_perm_map.eval(vec_type.get_shape(), ())
    scalable_dims = ArrayAttr(
        BoolAttr.from_bool(bool(b))
        for b in inv_perm_map.eval(vec_type.get_scalable_dims(), ())
    )
    res = VectorType(i1, mask_shape, scalable_dims)
    return res

get_transfer_minor_identity_map(shaped_type: TensorType | MemRefType, vector_type: VectorType) -> AffineMap staticmethod

Get the minor identity map for a transfer operation.

This is a helper function to compute the default permutation map for transfer operations when none is specified.

Source code in xdsl/dialects/vector.py

@staticmethod
def get_transfer_minor_identity_map(
    shaped_type: TensorType | MemRefType, vector_type: VectorType
) -> AffineMap:
    """
    Get the minor identity map for a transfer operation.

    This is a helper function to compute the default permutation map for
    transfer operations when none is specified.
    """
    element_vector_rank = 0
    element_type = shaped_type.element_type
    if isa(element_type, VectorType):
        element_vector_rank += element_type.get_num_dims()

    # 0-d transfers are to/from tensor<t>/memref<t> and vector<1xt>.
    # TODO: replace once we have 0-d vectors.
    if shaped_type.get_num_dims() == 0 and vector_type.get_shape() == (1,):
        return AffineMap.constant_map(0)

    return AffineMap.minor_identity(
        shaped_type.get_num_dims(),
        vector_type.get_num_dims() - element_vector_rank,
    )

resolve_attrs(parser: Parser, attributes_dict: dict[str, Attribute], shaped_type: TensorType | MemRefType, vector_type: VectorType, mask_start_pos: Position | None, mask_end_pos: Position | None, mask: UnresolvedOperand | None, types_pos: Position) staticmethod

Source code in xdsl/dialects/vector.py

@staticmethod
def resolve_attrs(
    parser: Parser,
    attributes_dict: dict[str, Attribute],
    shaped_type: TensorType | MemRefType,
    vector_type: VectorType,
    mask_start_pos: Position | None,
    mask_end_pos: Position | None,
    mask: UnresolvedOperand | None,
    types_pos: Position,
):
    # Create default permutation_map if not provided in attributes
    permutation_map = None
    if attributes_dict and "permutation_map" in attributes_dict:
        permutation_map = attributes_dict["permutation_map"]
        assert isinstance(permutation_map, AffineMapAttr)
    else:
        # Create identity permutation map for the shaped type's rank
        permutation_map = AffineMapAttr(
            VectorTransferOperation.get_transfer_minor_identity_map(
                shaped_type, vector_type
            )
        )

    # Create in_bounds attribute if not provided
    in_bounds = None
    if attributes_dict and "in_bounds" in attributes_dict:
        in_bounds = cast(ArrayAttr[BoolAttr], attributes_dict["in_bounds"])
    else:
        # Default: all dimensions are out-of-bounds
        in_bounds = ArrayAttr(
            (BoolAttr.from_bool(False),) * len(permutation_map.data.results)
        )

    if mask is not None:
        if isa(shaped_type.element_type, VectorType):
            assert mask_start_pos is not None
            assert mask_end_pos is not None
            parser.raise_error(
                "does not support masks with vector element type",
                at_position=mask_start_pos,
                end_position=mask_end_pos,
            )
        if vector_type.get_num_dims() != len(permutation_map.data.results):
            parser.raise_error(
                "expected the same rank for the vector and the "
                "results of the permutation map",
                types_pos,
            )
        # Instead of adding the mask type as an op type, compute it based on the
        # vector type and the permutation map (to keep the type signature small).
        mask_type = VectorTransferOperation.infer_transfer_op_mask_type(
            vector_type, permutation_map.data
        )
        resolved_mask = parser.resolve_operand(mask, mask_type)
    else:
        resolved_mask = None

    return resolved_mask, permutation_map, in_bounds

has_broadcast_dim()

Return "true" if at least one of the vector dimensions is a broadcasted dimension.

Source code in xdsl/dialects/vector.py

def has_broadcast_dim(self):
    """
    Return "true" if at least one of the vector dimensions is a broadcasted dimension.
    """
    return any(
        isinstance(expr, AffineConstantExpr) and expr.value == 0
        for expr in self.permutation_map.data.results
    )

verify_op(op: TransferReadOp | TransferWriteOp, shaped_type: MemRefType | TensorType, vector_type: VectorType, mask_type: VectorType[I1] | None, inferred_mask_type: VectorType[I1] | None, permutation_map: AffineMap, in_bounds: ArrayAttr[BoolAttr]) staticmethod

This mirrors VectorOps.cpp -> verifyTransferOp from MLIR

Source code in xdsl/dialects/vector.py

@staticmethod
def verify_op(
    op: TransferReadOp | TransferWriteOp,
    shaped_type: MemRefType | TensorType,
    vector_type: VectorType,
    mask_type: VectorType[I1] | None,
    inferred_mask_type: VectorType[I1] | None,
    permutation_map: AffineMap,
    in_bounds: ArrayAttr[BoolAttr],
):
    """
    This mirrors VectorOps.cpp -> verifyTransferOp from MLIR
    """

    element_type = shaped_type.element_type
    vector_element_type = vector_type.element_type

    if isa(element_type, VectorType):
        # Memref or tensor has vector element type
        # TODO verify vector element type
        pass
    else:
        # Memref of tensor has scalar element type
        if isa(vector_element_type, IndexType):
            if not isa(element_type, IndexType):
                raise VerifyException(
                    "Element type of source is index, expected element type of vector also to be index"
                )
        else:
            assert isa(vector_element_type, IntegerType | AnyFloat)
            assert isa(element_type, IntegerType | AnyFloat)

            minor_size = (
                1
                if vector_type.get_num_dims() == 0
                else vector_type.get_shape()[-1]
            )
            result_vec_size = vector_element_type.bitwidth * minor_size
            if result_vec_size % element_type.bitwidth != 0:
                raise VerifyException(
                    f'"{op.name}" requires the bitwidth of the minor 1-D vector to be '
                    "an integral multiple of the bitwidth of the source element type"
                )

        # Check that permutation map results match rank of vector type.
        if len(permutation_map.results) != vector_type.get_num_dims():
            raise VerifyException(
                f'"{op.name}" requires a permutation_map with result dims of the same rank as the vector type'
            )

    if permutation_map.num_symbols != 0:
        raise VerifyException(
            f'"{op.name}" requires permutation_map without symbols'
        )

    if permutation_map.num_dims != shaped_type.get_num_dims():
        raise VerifyException(
            f'"{op.name}" requires a permutation_map with input dims of the same rank as the source type'
        )

    if mask_type:
        if mask_type != inferred_mask_type:
            raise VerifyException(
                f'"{op.name}" inferred mask type ({inferred_mask_type}) and mask operand type ({mask_type}) don\'t match'
            )

    if len(in_bounds) != len(permutation_map.results):
        raise VerifyException(
            f'"{op.name}" expects the in_bounds attr of same rank as permutation_map results: '
            f"{permutation_map!s} vs in_bounds of of size {len(in_bounds)}"
        )

verify_permutation_map(op: TransferReadOp | TransferWriteOp, permutation_map: AffineMap) staticmethod

This mirrors VectorOps.cpp -> verifyPermutationMap

Source code in xdsl/dialects/vector.py

@staticmethod
def verify_permutation_map(
    op: TransferReadOp | TransferWriteOp,
    permutation_map: AffineMap,
):
    """
    This mirrors VectorOps.cpp -> verifyPermutationMap
    """

    seen: list[bool] = [False for _ in range(permutation_map.num_dims)]

    for expr in permutation_map.results:
        if isa(expr, AffineConstantExpr):
            if expr.value != 0:
                raise VerifyException(
                    f'"{op.name}" requires a projected permutation_map '
                    "(at most one dim or the zero constant can appear in each result)"
                )
            continue
        if not isa(expr, AffineDimExpr):
            raise VerifyException(
                f'"{op.name}" requires a projected permutation_map '
                "(at most one dim or the zero constant can appear in each result)"
            )
        if seen[expr.position]:
            raise VerifyException(
                f'"{op.name}" requires a permutation_map that is a permutation '
                "(found one dim used more than once)"
            )
        seen[expr.position] = True

TransferReadOp

Bases: VectorTransferOperation

Reads a supervector from memory into an SSA vector value.

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class TransferReadOp(VectorTransferOperation):
    "Reads a supervector from memory into an SSA vector value."

    name = "vector.transfer_read"

    source = operand_def(TensorType | MemRefType)
    indices = var_operand_def(IndexType)
    padding = operand_def()
    mask = opt_operand_def(VectorType[I1])

    permutation_map = prop_def(AffineMapAttr)

    result = result_def(VectorType)

    irdl_options = (
        AttrSizedOperandSegments(as_property=True),
        ParsePropInAttrDict(),
    )

    def __init__(
        self,
        source: SSAValue | Operation,
        indices: Sequence[SSAValue | Operation],
        padding: SSAValue | Operation,
        result_type: Attribute,
        in_bounds: ArrayAttr[BoolAttr],
        permutation_map: AffineMapAttr,
        mask: SSAValue | Operation | None = None,
    ):
        super().__init__(
            operands=[source, indices, padding, mask],
            result_types=[result_type],
            properties={"in_bounds": in_bounds, "permutation_map": permutation_map},
        )

    def print(self, printer: Printer):
        printer.print_string(" ", indent=0)
        printer.print_ssa_value(self.source)
        printer.print_string("[", indent=0)
        printer.print_list(self.indices, printer.print_ssa_value)
        printer.print_string("], ", indent=0)
        printer.print_ssa_value(self.padding)
        if self.mask is not None:
            printer.print_string(", ", indent=0)
            printer.print_ssa_value(self.mask)
        self._print_attrs(printer)
        printer.print_string(" : ", indent=0)
        printer.print_attribute(self.source.type)
        printer.print_string(", ", indent=0)
        printer.print_attribute(self.result.type)

    @classmethod
    def parse(cls, parser: Parser) -> TransferReadOp:
        source = parser.parse_unresolved_operand()
        indices = parser.parse_comma_separated_list(
            Parser.Delimiter.SQUARE, parser.parse_operand
        )
        parser.parse_punctuation(",")
        padding = parser.parse_operand()
        if parser.parse_optional_punctuation(","):
            mask_start_pos = parser.pos
            mask = parser.parse_unresolved_operand()
            mask_end_pos = parser.pos
        else:
            mask_start_pos = None
            mask = None
            mask_end_pos = None
        attributes_dict = parser.parse_optional_attr_dict()

        types_pos = parser.pos
        parser.parse_punctuation(":")
        shaped_type = parser.parse_type()
        parser.parse_punctuation(",")
        vector_type = parser.parse_type()

        source = parser.resolve_operand(source, shaped_type)

        if not isa(shaped_type, MemRefType | TensorType):
            parser.raise_error(
                "requires memref or ranked tensor type", at_position=types_pos
            )

        if not isa(vector_type, VectorType):
            parser.raise_error("requires vector type", at_position=types_pos)

        mask, permutation_map, in_bounds = VectorTransferOperation.resolve_attrs(
            parser,
            attributes_dict,
            shaped_type,
            vector_type,
            mask_start_pos,
            mask_end_pos,
            mask,
            types_pos,
        )

        # Create and return the TransferReadOp
        return TransferReadOp(
            source=source,
            indices=indices,
            padding=padding,
            mask=mask,
            permutation_map=permutation_map,
            in_bounds=in_bounds,
            result_type=vector_type,
        )

    def verify_(self):
        assert isa(self.source.type, MemRefType | TensorType)
        assert isa(self.result.type, VectorType)
        if self.mask:
            assert isa(self.mask.type, VectorType[I1])
            mask_type = self.mask.type
        else:
            mask_type = None

        if len(self.indices) != self.source.type.get_num_dims():
            raise VerifyException("Expected an index for each memref/tensor dimension.")

        if mask_type:
            inferred_mask_type = VectorTransferOperation.infer_transfer_op_mask_type(
                self.result.type,
                self.permutation_map.data,
            )
        else:
            inferred_mask_type = VectorType(i1, [])

        VectorTransferOperation.verify_op(
            self,
            self.source.type,
            self.result.type,
            mask_type,
            inferred_mask_type,
            self.permutation_map.data,
            self.in_bounds,
        )

        if isa(self.source.type.element_type, VectorType):
            # TODO verify vector element type
            pass
        else:
            # source memref/tensor has scalar element type
            # TODO verify that padding type is a valid element_type for a vector
            if self.source.type.element_type != self.padding.type:
                raise VerifyException(
                    f'"{self.name}" requires formal padding and source of the same elemental type'
                )

        VectorTransferOperation.verify_permutation_map(
            self,
            self.permutation_map.data,
        )

name = 'vector.transfer_read' class-attribute instance-attribute

source = operand_def(TensorType | MemRefType) class-attribute instance-attribute

indices = var_operand_def(IndexType) class-attribute instance-attribute

padding = operand_def() class-attribute instance-attribute

mask = opt_operand_def(VectorType[I1]) class-attribute instance-attribute

permutation_map = prop_def(AffineMapAttr) class-attribute instance-attribute

result = result_def(VectorType) class-attribute instance-attribute

irdl_options = (AttrSizedOperandSegments(as_property=True), ParsePropInAttrDict()) class-attribute instance-attribute

__init__(source: SSAValue | Operation, indices: Sequence[SSAValue | Operation], padding: SSAValue | Operation, result_type: Attribute, in_bounds: ArrayAttr[BoolAttr], permutation_map: AffineMapAttr, mask: SSAValue | Operation | None = None)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    source: SSAValue | Operation,
    indices: Sequence[SSAValue | Operation],
    padding: SSAValue | Operation,
    result_type: Attribute,
    in_bounds: ArrayAttr[BoolAttr],
    permutation_map: AffineMapAttr,
    mask: SSAValue | Operation | None = None,
):
    super().__init__(
        operands=[source, indices, padding, mask],
        result_types=[result_type],
        properties={"in_bounds": in_bounds, "permutation_map": permutation_map},
    )

print(printer: Printer)

Source code in xdsl/dialects/vector.py

def print(self, printer: Printer):
    printer.print_string(" ", indent=0)
    printer.print_ssa_value(self.source)
    printer.print_string("[", indent=0)
    printer.print_list(self.indices, printer.print_ssa_value)
    printer.print_string("], ", indent=0)
    printer.print_ssa_value(self.padding)
    if self.mask is not None:
        printer.print_string(", ", indent=0)
        printer.print_ssa_value(self.mask)
    self._print_attrs(printer)
    printer.print_string(" : ", indent=0)
    printer.print_attribute(self.source.type)
    printer.print_string(", ", indent=0)
    printer.print_attribute(self.result.type)

parse(parser: Parser) -> TransferReadOp classmethod

Source code in xdsl/dialects/vector.py

@classmethod
def parse(cls, parser: Parser) -> TransferReadOp:
    source = parser.parse_unresolved_operand()
    indices = parser.parse_comma_separated_list(
        Parser.Delimiter.SQUARE, parser.parse_operand
    )
    parser.parse_punctuation(",")
    padding = parser.parse_operand()
    if parser.parse_optional_punctuation(","):
        mask_start_pos = parser.pos
        mask = parser.parse_unresolved_operand()
        mask_end_pos = parser.pos
    else:
        mask_start_pos = None
        mask = None
        mask_end_pos = None
    attributes_dict = parser.parse_optional_attr_dict()

    types_pos = parser.pos
    parser.parse_punctuation(":")
    shaped_type = parser.parse_type()
    parser.parse_punctuation(",")
    vector_type = parser.parse_type()

    source = parser.resolve_operand(source, shaped_type)

    if not isa(shaped_type, MemRefType | TensorType):
        parser.raise_error(
            "requires memref or ranked tensor type", at_position=types_pos
        )

    if not isa(vector_type, VectorType):
        parser.raise_error("requires vector type", at_position=types_pos)

    mask, permutation_map, in_bounds = VectorTransferOperation.resolve_attrs(
        parser,
        attributes_dict,
        shaped_type,
        vector_type,
        mask_start_pos,
        mask_end_pos,
        mask,
        types_pos,
    )

    # Create and return the TransferReadOp
    return TransferReadOp(
        source=source,
        indices=indices,
        padding=padding,
        mask=mask,
        permutation_map=permutation_map,
        in_bounds=in_bounds,
        result_type=vector_type,
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    assert isa(self.source.type, MemRefType | TensorType)
    assert isa(self.result.type, VectorType)
    if self.mask:
        assert isa(self.mask.type, VectorType[I1])
        mask_type = self.mask.type
    else:
        mask_type = None

    if len(self.indices) != self.source.type.get_num_dims():
        raise VerifyException("Expected an index for each memref/tensor dimension.")

    if mask_type:
        inferred_mask_type = VectorTransferOperation.infer_transfer_op_mask_type(
            self.result.type,
            self.permutation_map.data,
        )
    else:
        inferred_mask_type = VectorType(i1, [])

    VectorTransferOperation.verify_op(
        self,
        self.source.type,
        self.result.type,
        mask_type,
        inferred_mask_type,
        self.permutation_map.data,
        self.in_bounds,
    )

    if isa(self.source.type.element_type, VectorType):
        # TODO verify vector element type
        pass
    else:
        # source memref/tensor has scalar element type
        # TODO verify that padding type is a valid element_type for a vector
        if self.source.type.element_type != self.padding.type:
            raise VerifyException(
                f'"{self.name}" requires formal padding and source of the same elemental type'
            )

    VectorTransferOperation.verify_permutation_map(
        self,
        self.permutation_map.data,
    )

TransferWriteOp

Bases: VectorTransferOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class TransferWriteOp(VectorTransferOperation):
    name = "vector.transfer_write"

    vector = operand_def(VectorType)
    source = operand_def(TensorType | MemRefType)
    indices = var_operand_def(IndexType)
    mask = opt_operand_def(VectorType[I1])

    permutation_map = prop_def(AffineMapAttr)

    result = opt_result_def(TensorType)

    irdl_options = (
        AttrSizedOperandSegments(as_property=True),
        ParsePropInAttrDict(),
    )

    def __init__(
        self,
        vector: SSAValue | Operation,
        source: SSAValue | Operation,
        indices: Sequence[SSAValue | Operation],
        in_bounds: ArrayAttr[BoolAttr],
        mask: SSAValue | Operation | None = None,
        permutation_map: AffineMapAttr | None = None,
        result_type: TensorType | None = None,
    ):
        super().__init__(
            operands=[vector, source, indices, mask],
            properties={"in_bounds": in_bounds, "permutation_map": permutation_map},
            result_types=[result_type],
        )

    def print(self, printer: Printer):
        printer.print_string(" ", indent=0)
        printer.print_operand(self.vector)
        printer.print_string(", ", indent=0)
        printer.print_operand(self.source)
        printer.print_string("[", indent=0)
        printer.print_list(self.indices, printer.print_operand)
        printer.print_string("]", indent=0)
        if self.mask is not None:
            printer.print_string(", ", indent=0)
            printer.print_ssa_value(self.mask)
        self._print_attrs(printer)
        printer.print_string(" : ", indent=0)
        printer.print_attribute(self.vector.type)
        printer.print_string(", ", indent=0)
        printer.print_attribute(self.source.type)

    @classmethod
    def parse(cls, parser: Parser) -> TransferWriteOp:
        vector = parser.parse_unresolved_operand()
        parser.parse_punctuation(",")
        source = parser.parse_unresolved_operand()
        indices = parser.parse_comma_separated_list(
            Parser.Delimiter.SQUARE, parser.parse_operand
        )
        if parser.parse_optional_punctuation(","):
            mask_start_pos = parser.pos
            mask = parser.parse_unresolved_operand()
            mask_end_pos = parser.pos
        else:
            mask_start_pos = None
            mask = None
            mask_end_pos = None
        attributes_dict = parser.parse_optional_attr_dict()

        types_pos = parser.pos
        parser.parse_punctuation(":")
        vector_type = parser.parse_type()
        parser.parse_punctuation(",")
        shaped_type = parser.parse_type()

        vector = parser.resolve_operand(vector, vector_type)
        source = parser.resolve_operand(source, shaped_type)

        if not isa(shaped_type, MemRefType | TensorType):
            parser.raise_error(
                "requires memref or ranked tensor type", at_position=types_pos
            )

        if not isa(vector_type, VectorType):
            parser.raise_error("requires vector type", at_position=types_pos)

        mask, permutation_map, in_bounds = VectorTransferOperation.resolve_attrs(
            parser,
            attributes_dict,
            shaped_type,
            vector_type,
            mask_start_pos,
            mask_end_pos,
            mask,
            types_pos,
        )

        # Create and return the TransferReadOp
        return TransferWriteOp(
            vector=vector,
            source=source,
            indices=indices,
            mask=mask,
            permutation_map=permutation_map,
            in_bounds=in_bounds,
            result_type=shaped_type if isinstance(shaped_type, TensorType) else None,
        )

    def verify_(self):
        assert isa(self.source.type, MemRefType | TensorType)
        assert isa(self.vector.type, VectorType)
        if self.mask:
            assert isa(self.mask.type, VectorType[I1])
            mask_type = self.mask.type
        else:
            mask_type = None

        if len(self.indices) != self.source.type.get_num_dims():
            raise VerifyException("Expected an index for each memref/tensor dimension.")

        if self.has_broadcast_dim():
            raise VerifyException(
                f'"{self.name}" should not have broadcast dimensions.'
            )

        if mask_type:
            inferred_mask_type = VectorTransferOperation.infer_transfer_op_mask_type(
                self.vector.type,
                self.permutation_map.data,
            )
        else:
            inferred_mask_type = VectorType(i1, [])

        VectorTransferOperation.verify_op(
            self,
            self.source.type,
            self.vector.type,
            mask_type,
            inferred_mask_type,
            self.permutation_map.data,
            self.in_bounds,
        )

        VectorTransferOperation.verify_permutation_map(
            self,
            self.permutation_map.data,
        )

name = 'vector.transfer_write' class-attribute instance-attribute

vector = operand_def(VectorType) class-attribute instance-attribute

source = operand_def(TensorType | MemRefType) class-attribute instance-attribute

indices = var_operand_def(IndexType) class-attribute instance-attribute

mask = opt_operand_def(VectorType[I1]) class-attribute instance-attribute

permutation_map = prop_def(AffineMapAttr) class-attribute instance-attribute

result = opt_result_def(TensorType) class-attribute instance-attribute

irdl_options = (AttrSizedOperandSegments(as_property=True), ParsePropInAttrDict()) class-attribute instance-attribute

__init__(vector: SSAValue | Operation, source: SSAValue | Operation, indices: Sequence[SSAValue | Operation], in_bounds: ArrayAttr[BoolAttr], mask: SSAValue | Operation | None = None, permutation_map: AffineMapAttr | None = None, result_type: TensorType | None = None)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    vector: SSAValue | Operation,
    source: SSAValue | Operation,
    indices: Sequence[SSAValue | Operation],
    in_bounds: ArrayAttr[BoolAttr],
    mask: SSAValue | Operation | None = None,
    permutation_map: AffineMapAttr | None = None,
    result_type: TensorType | None = None,
):
    super().__init__(
        operands=[vector, source, indices, mask],
        properties={"in_bounds": in_bounds, "permutation_map": permutation_map},
        result_types=[result_type],
    )

print(printer: Printer)

Source code in xdsl/dialects/vector.py

def print(self, printer: Printer):
    printer.print_string(" ", indent=0)
    printer.print_operand(self.vector)
    printer.print_string(", ", indent=0)
    printer.print_operand(self.source)
    printer.print_string("[", indent=0)
    printer.print_list(self.indices, printer.print_operand)
    printer.print_string("]", indent=0)
    if self.mask is not None:
        printer.print_string(", ", indent=0)
        printer.print_ssa_value(self.mask)
    self._print_attrs(printer)
    printer.print_string(" : ", indent=0)
    printer.print_attribute(self.vector.type)
    printer.print_string(", ", indent=0)
    printer.print_attribute(self.source.type)

parse(parser: Parser) -> TransferWriteOp classmethod

Source code in xdsl/dialects/vector.py

@classmethod
def parse(cls, parser: Parser) -> TransferWriteOp:
    vector = parser.parse_unresolved_operand()
    parser.parse_punctuation(",")
    source = parser.parse_unresolved_operand()
    indices = parser.parse_comma_separated_list(
        Parser.Delimiter.SQUARE, parser.parse_operand
    )
    if parser.parse_optional_punctuation(","):
        mask_start_pos = parser.pos
        mask = parser.parse_unresolved_operand()
        mask_end_pos = parser.pos
    else:
        mask_start_pos = None
        mask = None
        mask_end_pos = None
    attributes_dict = parser.parse_optional_attr_dict()

    types_pos = parser.pos
    parser.parse_punctuation(":")
    vector_type = parser.parse_type()
    parser.parse_punctuation(",")
    shaped_type = parser.parse_type()

    vector = parser.resolve_operand(vector, vector_type)
    source = parser.resolve_operand(source, shaped_type)

    if not isa(shaped_type, MemRefType | TensorType):
        parser.raise_error(
            "requires memref or ranked tensor type", at_position=types_pos
        )

    if not isa(vector_type, VectorType):
        parser.raise_error("requires vector type", at_position=types_pos)

    mask, permutation_map, in_bounds = VectorTransferOperation.resolve_attrs(
        parser,
        attributes_dict,
        shaped_type,
        vector_type,
        mask_start_pos,
        mask_end_pos,
        mask,
        types_pos,
    )

    # Create and return the TransferReadOp
    return TransferWriteOp(
        vector=vector,
        source=source,
        indices=indices,
        mask=mask,
        permutation_map=permutation_map,
        in_bounds=in_bounds,
        result_type=shaped_type if isinstance(shaped_type, TensorType) else None,
    )

verify_()

Source code in xdsl/dialects/vector.py

def verify_(self):
    assert isa(self.source.type, MemRefType | TensorType)
    assert isa(self.vector.type, VectorType)
    if self.mask:
        assert isa(self.mask.type, VectorType[I1])
        mask_type = self.mask.type
    else:
        mask_type = None

    if len(self.indices) != self.source.type.get_num_dims():
        raise VerifyException("Expected an index for each memref/tensor dimension.")

    if self.has_broadcast_dim():
        raise VerifyException(
            f'"{self.name}" should not have broadcast dimensions.'
        )

    if mask_type:
        inferred_mask_type = VectorTransferOperation.infer_transfer_op_mask_type(
            self.vector.type,
            self.permutation_map.data,
        )
    else:
        inferred_mask_type = VectorType(i1, [])

    VectorTransferOperation.verify_op(
        self,
        self.source.type,
        self.vector.type,
        mask_type,
        inferred_mask_type,
        self.permutation_map.data,
        self.in_bounds,
    )

    VectorTransferOperation.verify_permutation_map(
        self,
        self.permutation_map.data,
    )

CombiningKindFlag

Bases: StrEnum

Values specifying the kind of combining operation.

Source code in xdsl/dialects/vector.py

class CombiningKindFlag(StrEnum):
    """
    Values specifying the kind of combining operation.
    """

    ADD = "add"
    MUL = "mul"
    MINUI = "minui"
    MINSI = "minsi"
    MINNUMF = "minnumf"
    MAXUI = "maxui"
    MAXSI = "maxsi"
    MAXNUMF = "maxnumf"
    AND = "and"
    OR = "or"
    XOR = "xor"
    MAXIMUMF = "maximumf"
    MINIMUMF = "minimumf"

ADD = 'add' class-attribute instance-attribute

MUL = 'mul' class-attribute instance-attribute

MINUI = 'minui' class-attribute instance-attribute

MINSI = 'minsi' class-attribute instance-attribute

MINNUMF = 'minnumf' class-attribute instance-attribute

MAXUI = 'maxui' class-attribute instance-attribute

MAXSI = 'maxsi' class-attribute instance-attribute

MAXNUMF = 'maxnumf' class-attribute instance-attribute

AND = 'and' class-attribute instance-attribute

OR = 'or' class-attribute instance-attribute

XOR = 'xor' class-attribute instance-attribute

MAXIMUMF = 'maximumf' class-attribute instance-attribute

MINIMUMF = 'minimumf' class-attribute instance-attribute

CombiningKindAttr dataclass

Bases: EnumAttribute[CombiningKindFlag]

A mirror of LLVM's vector.kind attribute.

Source code in xdsl/dialects/vector.py

@irdl_attr_definition
class CombiningKindAttr(EnumAttribute[CombiningKindFlag]):
    """
    A mirror of LLVM's vector.kind attribute.
    """

    name = "vector.kind"

    def print_parameter(self, printer: Printer) -> None:
        with printer.in_angle_brackets():
            printer.print_string(self.data)

    @classmethod
    def parse_parameter(cls, parser: AttrParser) -> CombiningKindFlag:
        with parser.in_angle_brackets():
            return CombiningKindFlag(parser.parse_identifier())

name = 'vector.kind' class-attribute instance-attribute

print_parameter(printer: Printer) -> None

Source code in xdsl/dialects/vector.py

def print_parameter(self, printer: Printer) -> None:
    with printer.in_angle_brackets():
        printer.print_string(self.data)

parse_parameter(parser: AttrParser) -> CombiningKindFlag classmethod

Source code in xdsl/dialects/vector.py

@classmethod
def parse_parameter(cls, parser: AttrParser) -> CombiningKindFlag:
    with parser.in_angle_brackets():
        return CombiningKindFlag(parser.parse_identifier())

ReductionOp

Bases: IRDLOperation

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class ReductionOp(IRDLOperation):
    name = "vector.reduction"

    _T: ClassVar = VarConstraint("T", AnyAttr())

    vector = operand_def(VectorType.constr(_T))
    acc = opt_operand_def(_T)
    dest = result_def(_T)
    kind = prop_def(CombiningKindAttr)
    fastmath = prop_def(FastMathFlagsAttr, default_value=FastMathFlagsAttr("none"))

    assembly_format = "$kind `,` $vector (`,` $acc^)? (`fastmath` `` $fastmath^)? attr-dict `:` type($vector) `into` type($dest)"

    def __init__(
        self,
        vector: SSAValue | Operation,
        kind: CombiningKindAttr,
        acc: SSAValue | Operation | None = None,
        fastmath: FastMathFlagsAttr | None = None,
    ):
        vector = SSAValue.get(vector)
        super().__init__(
            operands=[vector, acc],
            result_types=[vector.type],
            properties={
                "kind": kind,
                "fastmath": fastmath,
            },
        )

name = 'vector.reduction' class-attribute instance-attribute

vector = operand_def(VectorType.constr(_T)) class-attribute instance-attribute

acc = opt_operand_def(_T) class-attribute instance-attribute

dest = result_def(_T) class-attribute instance-attribute

kind = prop_def(CombiningKindAttr) class-attribute instance-attribute

fastmath = prop_def(FastMathFlagsAttr, default_value=(FastMathFlagsAttr('none'))) class-attribute instance-attribute

assembly_format = '$kind `,` $vector (`,` $acc^)? (`fastmath` `` $fastmath^)? attr-dict `:` type($vector) `into` type($dest)' class-attribute instance-attribute

__init__(vector: SSAValue | Operation, kind: CombiningKindAttr, acc: SSAValue | Operation | None = None, fastmath: FastMathFlagsAttr | None = None)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    vector: SSAValue | Operation,
    kind: CombiningKindAttr,
    acc: SSAValue | Operation | None = None,
    fastmath: FastMathFlagsAttr | None = None,
):
    vector = SSAValue.get(vector)
    super().__init__(
        operands=[vector, acc],
        result_types=[vector.type],
        properties={
            "kind": kind,
            "fastmath": fastmath,
        },
    )

BitcastOp

Bases: IRDLOperation

Bitcast between vectors.

See external documentation.

Source code in xdsl/dialects/vector.py

@irdl_op_definition
class BitcastOp(IRDLOperation):
    """
    Bitcast between vectors.

    See [external documentation](https://mlir.llvm.org/docs/Dialects/Vector/#vectorbitcast-vectorbitcastop).
    """

    name = "vector.bitcast"

    source = operand_def(
        VectorType.constr(base(IntegerType) | base(IndexType) | AnyFloatConstr)
    )
    result = result_def(
        VectorType.constr(base(IntegerType) | base(IndexType) | AnyFloatConstr)
    )

    assembly_format = "$source attr-dict `:` type($source) `to` type($result)"

    def __init__(
        self,
        source: SSAValue | Operation,
        result_type: Attribute,
    ):
        super().__init__(
            operands=[source],
            result_types=[result_type],
        )

    def verify_(self) -> None:
        s_t = self.source.type
        r_t = self.result.type

        assert isa(s_t, VectorType)
        assert isa(r_t, VectorType)

        s_elem_t = s_t.get_element_type()
        r_elem_t = r_t.get_element_type()
        s_shape = s_t.get_shape()
        r_shape = r_t.get_shape()

        # technically only support index -> index conversions if sizes unknown,
        # and they must have the same shape
        s_elem_t_sized = isinstance(s_elem_t, FixedBitwidthType)
        r_elem_t_sized = isinstance(r_elem_t, FixedBitwidthType)

        if not s_elem_t_sized or not r_elem_t_sized:
            # if they are both unsized and have the same shape
            if not (s_elem_t_sized ^ r_elem_t_sized) and s_shape == r_shape:
                return

            raise VerifyException(
                "For element types of undefined bitwidth, expect "
                + "both types to have undefined bitwidth and shape to be equal"
            )

        source_size = prod(s_shape) * s_elem_t.bitwidth
        result_size = prod(r_shape) * r_elem_t.bitwidth

        # if sizes are known, they must match perfectly
        if not source_size == result_size:
            raise VerifyException(
                "The source and result types do not have an equal bitwidth"
            )

name = 'vector.bitcast' class-attribute instance-attribute

source = operand_def(VectorType.constr(base(IntegerType) | base(IndexType) | AnyFloatConstr)) class-attribute instance-attribute

result = result_def(VectorType.constr(base(IntegerType) | base(IndexType) | AnyFloatConstr)) class-attribute instance-attribute

assembly_format = '$source attr-dict `:` type($source) `to` type($result)' class-attribute instance-attribute

__init__(source: SSAValue | Operation, result_type: Attribute)

Source code in xdsl/dialects/vector.py

def __init__(
    self,
    source: SSAValue | Operation,
    result_type: Attribute,
):
    super().__init__(
        operands=[source],
        result_types=[result_type],
    )

verify_() -> None

Source code in xdsl/dialects/vector.py

def verify_(self) -> None:
    s_t = self.source.type
    r_t = self.result.type

    assert isa(s_t, VectorType)
    assert isa(r_t, VectorType)

    s_elem_t = s_t.get_element_type()
    r_elem_t = r_t.get_element_type()
    s_shape = s_t.get_shape()
    r_shape = r_t.get_shape()

    # technically only support index -> index conversions if sizes unknown,
    # and they must have the same shape
    s_elem_t_sized = isinstance(s_elem_t, FixedBitwidthType)
    r_elem_t_sized = isinstance(r_elem_t, FixedBitwidthType)

    if not s_elem_t_sized or not r_elem_t_sized:
        # if they are both unsized and have the same shape
        if not (s_elem_t_sized ^ r_elem_t_sized) and s_shape == r_shape:
            return

        raise VerifyException(
            "For element types of undefined bitwidth, expect "
            + "both types to have undefined bitwidth and shape to be equal"
        )

    source_size = prod(s_shape) * s_elem_t.bitwidth
    result_size = prod(r_shape) * r_elem_t.bitwidth

    # if sizes are known, they must match perfectly
    if not source_size == result_size:
        raise VerifyException(
            "The source and result types do not have an equal bitwidth"
        )