MLIR

Multi-Level IR Compiler Framework

The `Broadcastable` Trait

Description 

The Broadcastable trait enforces the following properties on an operation:

  • The operation has at least one input operand.

  • The operation has exactly one result.

  • All input operands and result are of type tensor or vector.

  • A shape inference mechanism is able to compute the result shape solely based on input operand shapes.

  • Input operands have broadcast-compatible shapes, according to the verification rules presented below.

  • The operation’s result shape is compatible with —though not necessarily identical to— the shape inferred from its input operands, according to the verification rules presented below.

Dimension inference 

Given an operation with two input operands, the size of dimension i of its result can be inferred from dimension i of the operands according to the table below. Here, dim0 and dim1 represent dimension i of the input operands in an interchangeable order, while inferredDim represents the inferred size for dimension i of the operation result. Dimensions are classified in three categories: dynamic ("?"), static equal to 1 (“1”), and static greater than 1 (">1").

dim0dim1inferredDimNotes
???If RuntimeSize(dim0) is 1, dimension dim0 is broadcast to RuntimeSize(dim1). If RuntimeSize(dim1) is 1, dimension dim1 is broadcast to RuntimeSize(dim0). The operation produces undefined behavior if both runtime sizes are greater than 1 and not equal.
?1?Dimension dim1 is broadcast to RuntimeSize(dim0).
?>1dim1If RuntimeSize(dim0) is 1, dim0 is broadcast to dim1. The operation produces undefined behavior if RuntimeSize(dim0) is greater than 1 and not equal to dim1.
111
1>1dim1Dimension dim0 is broadcast to dim1.
>1>1dim0The operation verifier produces a compile-time error if dim0 != dim1.

The following pseudo-function is a formal representation of the dimension inference process:

InferDim(dim0, dim1):
	switch (dim0, dim1):
		case (?, ?):
		case (?, 1):
		case (1, 1):
		case (>1, ?):
		case (>1, 1):
			return dim0
		case (?, >1):
		case (1, ?):
		case (1, >1):
			return dim1
		case (>1, >1):
			ERROR_IF(dim0 != dim1)
			return dim0

Shape inference 

The shape inference process begins by correcting rank differences in input operands. A shape is expanded by adding additional dimensions of size 1 on its left until the desired rank is reached, as shown here:

ExpandRank(shape, rank):
	while len(shape) < rank:
		shape.prepend(1)

Given the shapes of two ranked input operands, the result’s shape is inferred by equalizing input ranks and inferring individual dimensions, as shown here:

InferShape(shape0, shape1):

  # Equalize ranks
  rank = max(GetRank(shape0), GetRank(shape1))
  ExpandRank(shape0, rank)
  ExpandRank(shape1, rank)
	
  # Infer shape
  inferredShape = []
  for (dim0, dim1) in zip(shape0, shape1):
    inferredDim = InferDim(dim0, dim1)
    inferredShape.append(inferredDim)
  return inferredShape

The result shape for an operation with an arbitrary number of input operands is then inferred by discarding unranked operands, applying shape inference on the first ranked operand pair, and updating the inferred shape with each additional ranked operand. If the operation has no ranked operands, the result shape cannot be inferred. If the operation has exactly one ranked operand, its shape is directly provided as the inferred result shape. Formally:

InferResultShape(op):

	# Filter ranked operands
	rankedOperands = filter(op.operands, IsRanked)
	if len(rankedOperands) == 0:
		return None
	
	# Infer result shape
	inferredShape = GetShape(rankedOperands[0])
	for operand in rankedOperands[1:]:
		inferredShape = InferShape(inferredShape, GetShape(operand))
	return inferredShape

Verification 

The legality of an operation with the Broadcastable trait is verified by first running the shape inference process. If a failure occurs during shape inference, it is concluded that input operands are not broadcast-compatible, and verification fails. If shape inference succeeds, verification continues.

If either the result is unranked or all input operands are unranked, no further verification steps are needed, and the process ends here successfully. If, on the contrary, both the result and at least one input operand are ranked, verification continues by checking for a matching rank between the previously inferred shape and the result.

Once a rank match is guaranteed, each dimension of the inferred shape is compared with the corresponding dimension of the actual result shape according to the following table table:

inferredDimactualDimVerification outcome
??OK
?staticOK
A failure to guarantee that the runtime dimension size of the result is equal to actualDim causes undefined behavior. While unusual, this implicit dynamic-to-static cast is convenient in certain scenarios, such as an intermediate state of a shape inference pass. Ultimately, a static dimension in the result implies that all input dimension sizes are also known at compile time and may therefore become static as well, preferably.
static?OK
The actual result dimension may be dynamic even when a static size can be inferred at compile time. The programmer may choose to relax the specificity of the result dimension for forward compatibility of the result type.
staticstaticOK if equal
When both the inferred and actual dimensions are static, they must be set to the same size.

The full verification process can be formally specified as follows:

Verify(op):

	# Run shape inference
	inferredShape = InferResultShape(op.operands)

	# Done if result is unranked or all operands are unranked
	if not IsRanked(op.result) or inferredShape is None:
		return
	
	# Rank must match
	actualShape = GetShape(op.result):
	ERROR_IF(len(inferredShape) != len(actualShape))
	
	# Verify
	for (inferredDim, actualDim) in zip(inferredShape, actualShape):
		ERROR_IF(IsStatic(actualDim) and inferredDim != actualDim)

Examples 

The following are correct uses of broadcastable ops:

// Exact match of static sizes.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<1x2xi32>, tensor<1x2xi32) -> tensor<1x2xi32>

// Dynamic sizes match. The programmer must guarantee that the runtime sizes of
// %arg0 and %arg1 are equal at runtime.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<?xi32>, tensor<?xi32) -> tensor<?xi32>

// The shape of %arg0 is broadcast from tensor<1xi32> to tensor<4xi32>.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<1xi32>, tensor<4xi32) -> tensor<4xi32>

// The shape of %result is inferred as tensor<4xi32>, while the actual result
// type is tensor<?xi32>. The inferred shape is compatible with the actual shape.
%result = "test.broadcastable"(%arg0) : (tensor<4xi32) -> tensor<?xi32>

// The shape of %arg0 is first expanded to tensor<1x1x4xi32> and then broadcast
// to tensor<2x3x4xi32>.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<4xi32>, tensor<2x3x4xi32) -> tensor<2x3x4xi32>

// Input and results tensors have different element types (i1, i32, i64). The
// 'Broadcastable' trait has no restrictions on element types.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<2xi1>, tensor<2xi32) -> tensor<2xi64>

// No result shape verification is needed when the result is unranked.
%result = "test.broadcastable"(%arg0) : (tensor<2xi32>) -> tensor<*xi32>

// No result shape verification needed when all inputs are unranked.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<*xi32>, tensor<*xi32>) -> tensor<2xi32>

The following are incorrect uses of broadcastable ops:

// Dimension 0 of input operands is static but not equal.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<3xi32>, tensor<2xi32) -> tensor<?xi32>

// The inferred result shape is tensor<3xi32>, but the actual result shape is
// tensor<1x3xi32>. Inferred and actual shapes differ in rank.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<3xi32>, tensor<3xi32) -> tensor<1x3xi32>

// The inferred result shape is tensor<?xi32>, but the actual shape is
// tensor<4xi32>. The inferred shape is not compatible with the actual shape.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<?xi32>, tensor<?xi32) -> tensor<4xi32>

// The inferred result shape is tensor<2xi32>, but the actual result shape is
// tensor<4xi32>, which is not compatible.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<2xi32>, tensor<2xi32) -> tensor<4xi32>

// The inferred result shape is tensor<1xi32>, but the actual result shape is
// tensor<4xi32>. Broadcast semantics are not applicable for results.
%result = "test.broadcastable"(%arg0, %arg1) : (tensor<1xi32>, tensor<1xi32) -> tensor<4xi32>