'dataflow' Dialect

An dataflow out-of-tree MLIR dialect This dialect contains dedicated operations, types, and attributes designed for representing dataflow-specific structures and components in MLIR.

• Operations
  • dataflow.buffer (streamtensor::dataflow::BufferOp)
  • dataflow.itensor_buffer (streamtensor::dataflow::ITensorBufferOp)
  • dataflow.itensor_cast (streamtensor::dataflow::ITensorCastOp)
  • dataflow.itensor_chunk (streamtensor::dataflow::ITensorChunkOp)
  • dataflow.itensor_concat (streamtensor::dataflow::ITensorConcatOp)
  • dataflow.itensor_empty (streamtensor::dataflow::ITensorEmptyOp)
  • dataflow.itensor_fork (streamtensor::dataflow::ITensorForkOp)
  • dataflow.itensor_instance (streamtensor::dataflow::ITensorInstanceOp)
  • dataflow.itensor_join (streamtensor::dataflow::ITensorJoinOp)
  • dataflow.itensor_read (streamtensor::dataflow::ITensorReadOp)
  • dataflow.itensor_reassociate (streamtensor::dataflow::ITensorReassociateOp)
  • dataflow.itensor_to_stream (streamtensor::dataflow::ITensorToStreamOp)
  • dataflow.itensor_write (streamtensor::dataflow::ITensorWriteOp)
  • dataflow.kernel (streamtensor::dataflow::KernelOp)
  • dataflow.stream (streamtensor::dataflow::StreamOp)
  • dataflow.stream_cast (streamtensor::dataflow::StreamCastOp)
  • dataflow.stream_read (streamtensor::dataflow::StreamReadOp)
  • dataflow.stream_to_itensor (streamtensor::dataflow::StreamToITensorOp)
  • dataflow.stream_write (streamtensor::dataflow::StreamWriteOp)
  • dataflow.task (streamtensor::dataflow::TaskOp)
  • dataflow.yield (streamtensor::dataflow::YieldOp)
• Attributes
  • PartitionLayoutAttr
• Types
  • ITensorType
  • StreamType

Operations

dataflow.buffer (streamtensor::dataflow::BufferOp)

Represent a dataflow buffer

Syntax:

operation ::= `dataflow.buffer` (`layout` $layout^)? attr-dict `:` type(results)

Interfaces: MemoryEffectOpInterface

Attributes:

Attribute	MLIR Type	Description
layout	::mlir::streamtensor::dataflow::PartitionLayoutAttr	Layout of array partition

Results:

Result	Description
memref	statically shaped memref of any type values

dataflow.itensor_buffer (streamtensor::dataflow::ITensorBufferOp)

Buffer an iterative tensor at a specific position

Syntax:

operation ::= `dataflow.itensor_buffer` $source `buffer` $bufferType `before` `loop` $loopIndex `dim` $dimIndex
              attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
bufferType	::mlir::TypeAttr	type attribute of ranked tensor of any type values
loopIndex	::mlir::IntegerAttr	64-bit signed integer attribute
dimIndex	::mlir::IntegerAttr	64-bit signed integer attribute

Operands:

Operand	Description
source	itensor of any type values

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_cast (streamtensor::dataflow::ITensorCastOp)

Cast an iterative tensor to a different type

Syntax:

operation ::= `dataflow.itensor_cast` $source attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, ITensorViewLikeOpInterface, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
source	itensor of any type values

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_chunk (streamtensor::dataflow::ITensorChunkOp)

Chunk a tensor to multiple tensors

Syntax:

operation ::= `dataflow.itensor_chunk` `iter_dim` `(` $iterDim `)` `dim` `(` $dim `)` `size` `(` $size `)` $source
              attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
iterDim	::mlir::IntegerAttr	64-bit signed integer attribute
dim	::mlir::IntegerAttr	64-bit signed integer attribute
size	::mlir::IntegerAttr	64-bit signed integer attribute

Operands:

Operand	Description
source	itensor of any type values

Results:

Result	Description
results	variadic of itensor of any type values

dataflow.itensor_concat (streamtensor::dataflow::ITensorConcatOp)

Concat a tensor to multiple tensors

Syntax:

operation ::= `dataflow.itensor_concat` `iter_dim` `(` $iterDim `)` `dim` `(` $dim `)` $sources attr-dict `:`
              functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
iterDim	::mlir::IntegerAttr	64-bit signed integer attribute
dim	::mlir::IntegerAttr	64-bit signed integer attribute

Operands:

Operand	Description
sources	variadic of itensor of any type values

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_empty (streamtensor::dataflow::ITensorEmptyOp)

A empty iterative tensor

Syntax:

operation ::= `dataflow.itensor_empty` attr-dict `:` type(results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_fork (streamtensor::dataflow::ITensorForkOp)

Fork an iterative tensor to multiple ones

Syntax:

operation ::= `dataflow.itensor_fork` $source attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
source	itensor of any type values

Results:

Result	Description
results	variadic of itensor of any type values

dataflow.itensor_instance (streamtensor::dataflow::ITensorInstanceOp)

An iterative tensor instance that is not pure

Syntax:

operation ::= `dataflow.itensor_instance` `depth` $depth attr-dict `:` type(results)

Attributes:

Attribute	MLIR Type	Description
depth	::mlir::IntegerAttr	64-bit signed integer attribute

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_join (streamtensor::dataflow::ITensorJoinOp)

Join multiple iterative tensors to one

Syntax:

operation ::= `dataflow.itensor_join` $sources attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
sources	variadic of itensor of any type values

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_read (streamtensor::dataflow::ITensorReadOp)

Read an element from an iterative tensor

Syntax:

operation ::= `dataflow.itensor_read` $source (`init` $init^)? attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, DestinationStyleOpInterface, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
source	itensor of any type values
init	ranked tensor of any type values

Results:

Result	Description
value	any type

dataflow.itensor_reassociate (streamtensor::dataflow::ITensorReassociateOp)

Reassociate the shape and/or iteration of an iterative tensor

Syntax:

operation ::= `dataflow.itensor_reassociate` $source (`expand` $expandShape^):(`collapse`)? `shape` $shapeReassociation
              (`expand` $expandIteration^):(`collapse`)? `iteration`
              $iterationReassociation attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, ITensorViewLikeOpInterface, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
expandShape	::mlir::UnitAttr	unit attribute
shapeReassociation	::mlir::ArrayAttr	Array of 64-bit integer array attributes
expandIteration	::mlir::UnitAttr	unit attribute
iterationReassociation	::mlir::ArrayAttr	Array of 64-bit integer array attributes

Operands:

Operand	Description
source	itensor of any type values

Results:

Result	Description
result	itensor of any type values

dataflow.itensor_to_stream (streamtensor::dataflow::ITensorToStreamOp)

Convert an itensor to a stream channel

Syntax:

operation ::= `dataflow.itensor_to_stream` $iTensor attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
iTensor	An iterative tensor type

Results:

Result	Description
stream	stream of any type values

dataflow.itensor_write (streamtensor::dataflow::ITensorWriteOp)

Write an element to an iterative tensor

Syntax:

operation ::= `dataflow.itensor_write` $value `into` $dest attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
value	any type
dest	itensor of any type values

Results:

Result	Description
result	itensor of any type values

dataflow.kernel (streamtensor::dataflow::KernelOp)

Represent a dataflow kernel containing multiple tasks

Syntax:

operation ::= `dataflow.kernel` $sym_name (`inputs` $inputs^)? `inputs_packed` $inputsPacked
              `results_packed` $resultsPacked $body attr-dict `:`
              functional-type(operands, results)

Traits: SingleBlockImplicitTerminator<YieldOp>, SingleBlock

Attributes:

Attribute	MLIR Type	Description
sym_name	::mlir::StringAttr	string attribute
inputsPacked	::mlir::DenseBoolArrayAttr	i1 dense array attribute
resultsPacked	::mlir::DenseBoolArrayAttr	i1 dense array attribute

Operands:

Operand	Description
inputs	variadic of ranked tensor of any type values

Results:

Result	Description
results	variadic of ranked tensor of any type values

dataflow.stream (streamtensor::dataflow::StreamOp)

Instantiate a stream channel

Syntax:

operation ::= `dataflow.stream` `depth` $depth attr-dict `:` type(results)

Interfaces: MemoryEffectOpInterface

Attributes:

Attribute	MLIR Type	Description
depth	::mlir::IntegerAttr	64-bit signed integer attribute

Results:

Result	Description
stream	stream of any type values

dataflow.stream_cast (streamtensor::dataflow::StreamCastOp)

Cast a stream to a different type

Syntax:

operation ::= `dataflow.stream_cast` $source attr-dict `:` functional-type(operands, results)

Interfaces: StreamViewLikeOpInterface

Operands:

Operand	Description
source	stream of any type values

Results:

Result	Description
result	stream of any type values

dataflow.stream_read (streamtensor::dataflow::StreamReadOp)

Read an element from stream channel

Syntax:

operation ::= `dataflow.stream_read` $source attr-dict `:` functional-type(operands, results)

Read/pop a value from a stream channel.

Interfaces: MemoryEffectOpInterface

Operands:

Operand	Description
source	stream of any type values

Results:

Result	Description
value	any type

dataflow.stream_to_itensor (streamtensor::dataflow::StreamToITensorOp)

Convert a stream channel to a itensor

Syntax:

operation ::= `dataflow.stream_to_itensor` $stream attr-dict `:` functional-type(operands, results)

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
stream	stream of any type values

Results:

Result	Description
iTensor	itensor of any type values

dataflow.stream_write (streamtensor::dataflow::StreamWriteOp)

Write an element to stream channel

Syntax:

operation ::= `dataflow.stream_write` $value `into` $dest attr-dict `:` type(operands)

Write/push a value to a stream channel.

Interfaces: MemoryEffectOpInterface

Operands:

Operand	Description
value	any type
dest	stream of any type values

dataflow.task (streamtensor::dataflow::TaskOp)

Represent a dataflow task

Syntax:

operation ::= `dataflow.task` (`kernel` $kernel^)? $sym_name (`inits` $inits^)? $body attr-dict
              (`:` functional-type($inits, $results)^)?

Traits: SingleBlockImplicitTerminator<YieldOp>, SingleBlock

Attributes:

Attribute	MLIR Type	Description
sym_name	::mlir::StringAttr	string attribute
kernel	::mlir::UnitAttr	unit attribute

Operands:

Operand	Description
inits	variadic of any type

Results:

Result	Description
results	variadic of any type

dataflow.yield (streamtensor::dataflow::YieldOp)

Terminate and yield results of a schedule or task op

Syntax:

operation ::= `dataflow.yield` $results attr-dict `:` type($results)

Traits: AlwaysSpeculatableImplTrait, HasParent<KernelOp, TaskOp>, ReturnLike, Terminator

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface), RegionBranchTerminatorOpInterface

Effects: MemoryEffects::Effect{}

Operands:

Operand	Description
results	variadic of any type

Attributes

PartitionLayoutAttr

Layout of array partition

Syntax:

#dataflow.partition<
  ::llvm::ArrayRef<dataflow::PartitionKind>,   # kinds
  ::llvm::ArrayRef<int64_t>   # factors
>

Parameters:

Parameter	C++ type	Description
kinds	::llvm::ArrayRef<dataflow::PartitionKind>
factors	::llvm::ArrayRef<int64_t>

Types

ITensorType

An iterative tensor type

Syntax:

!dataflow.itensor<
  mlir::Type,   # elementType
  ::llvm::ArrayRef<int64_t>,   # iterTripCounts
  ::llvm::ArrayRef<int64_t>,   # iterSteps
  mlir::AffineMap   # iterMap
>

Iterative tensor representing a tile or window of a full tensor that iteratively traverses the full tensor. For example, given a tensor with type tensor<64x128xf32>, if we tile this tensor into 16x16 tiles and traverse them with two nested loops in a natural order, the corresponding itensor type would be:

itensor<f32, iter_space: [4, 8] * [16, 16], iter_map: (d0, d1) -> (d0, d1)>

Specifically, the iterTripCounts parameter specifies the trip counts of each loop, while the iterSteps parameter specifies the step sizes of each loop. The iterMap parameter specifies the mapping from the iteration space to the full tensor space, i.e., the loop induction variables to the tensor indices. The following is an example of how to read the itensor above:

%0 = any_defining_op : itensor<f32, iter_space: [4, 8] * [16, 16],
  iter_map: (d0, d1) -> (d0, d1)>
scf.for %i = %c0 to %c4 step %c1 {
  scf.for %j = %c0 to %c8 step %c1 {
    %1 = tensor.empty : tensor<16x16xf32>
    %2 = dataflow.itensor_read %0 init %1 : itensor<f32, iter_space:[4, 8] *
      [16, 16], iter_map: (d0, d1) -> (d0, d1)> -> tensor<16x16xf32>
    ...
  }
}

An itensor shares similar properties with the tensor, which means it is not mutable and cannot be updated in place. The itensor ops can be converted to stream ops for code generation. This process is similar to the bufferize process from tensor ops to buffer (memref) ops.

Parameters:

Parameter	C++ type	Description
elementType	mlir::Type
iterTripCounts	::llvm::ArrayRef<int64_t>
iterSteps	::llvm::ArrayRef<int64_t>
iterMap	mlir::AffineMap

StreamType

A stream type

Syntax:

!dataflow.stream<
  mlir::Type,   # elementType
  mlir::Attribute   # memorySpace
>

A stream of any type that represents a FIFO between the producer and the consumer module.

Parameters:

Parameter	C++ type	Description
elementType	mlir::Type
memorySpace	mlir::Attribute