‘ADF’ Dialect

Types and operations for the ADF dialect

This is an MLIR dialect designed to capture the Xilinx Adaptive Data Flow (ADF) Graph semantics. This dialect is intended to be converible to/from the C++ graph syntax, but to represent the graph abstractly, rather than at the C++ syntactic level.

[TOC]

Operations

ADF.graph (::xilinx::ADF::GraphOp)

An ADF graph operation

Syntax:

operation ::= `ADF.graph` `(` $name `)` regions attr-dict

An operation that represents an ADF graph. This operation determines the name of the graph and contains a single region containing ADF dialect operations. For example,

ADF.graph("name") {
  your_graph_operations
}

Traits: AlwaysSpeculatableImplTrait, NoTerminator, SingleBlock

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
name	::mlir::StringAttr	string attribute

ADF.inout_port (::xilinx::ADF::GraphInOutOp)

Graph input/output port operation

Syntax:

operation ::= `ADF.inout_port` `(` $name `)` $inp attr-dict `:` `(`type($inp) `)` `->` type(results)

An operation representing a dual input/output of an ADF graph. This is usually used to represent inout parameter val=ues

  %2 = ADF.inout_port("name") %3 : (!ADF.interface<!ADF.int32>) -> !ADF.interface<!ADF.int32>                                           

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
name	::mlir::StringAttr	string attribute

Operands:

Operand	Description
inp

Results:

Result	Description
«unnamed»

ADF.input_port (::xilinx::ADF::GraphInputOp)

Graph input port operation

Syntax:

operation ::= `ADF.input_port` `(` $name `)` $value attr-dict `->` type($output)

An operation representing a primary input of an ADF graph. It takes an array attribute as the input argument. The constant values of the array attribute defines the interface of whole ADF graph. The first index is a bool integer and 1:i1 means the ADF graph is based on stream data access mechanism; 0:i1 means it’s an adf window data access. If it’s the stream mechanism, then the second index should set to -1, otherwise, the second index should set to the window size. e.g. 128 byte.

The input_port will return an output with “interface” type to propagate the interface values defined by the argument array attributes.

  %0 = ADF.input_port("name")  [1:i1, -1:i32] -> !ADF.interface<!ADF.int32>  // for ADF stream access mechanism

or

  %1 = ADF.input_port("name")  [0:i1, 128:i32] -> !ADF.interface<!ADF.int32> // for ADF window access mechanism 

Traits: AlwaysSpeculatableImplTrait, ConstantLike

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
value	::mlir::ArrayAttr	array attribute
name	::mlir::StringAttr	string attribute

Results:

Result	Description
output

ADF.kernel (::xilinx::ADF::KernelOp)

Call and instantiate a kernel node with proper connections

Syntax:

operation ::= `ADF.kernel` $callee `(` $kernel_inputs `)` attr-dict `:` functional-type($kernel_inputs, results)

An operation representing a primitive node in the graph. This operation refers to a function implementing the kernel. It takes “window” types for both its variadic inputs and the op_return to propagate the window size through entire ADF graph.

  %3 = ADF.kernel @kfunc2(%1, %2) :
          (!ADF.interface<!ADF.int32>, !ADF.interface<!ADF.int32>) -> !ADF.interface<!ADF.int32>

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
callee	::mlir::FlatSymbolRefAttr	flat symbol reference attribute

Operands:

Operand	Description
kernel_inputs	variadic of

Results:

Result	Description
«unnamed»

ADF.output_port (::xilinx::ADF::GraphOutputOp)

Graph output port operation

Syntax:

operation ::= `ADF.output_port` `(` $name `)` $inp attr-dict `:` `(`type($inp) `)` `->` type(results)

An operation representing a primary input of an ADF graph. The operation takes an argument with “window” type and also returns an “window” type output for final type checking.

  %2 = ADF.output_port("name") %3 : (!ADF.interface<!ADF.int32>) -> !ADF.interface<!ADF.int32>                                           

Traits: AlwaysSpeculatableImplTrait

Interfaces: ConditionallySpeculatable, NoMemoryEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

Attributes:

Attribute	MLIR Type	Description
name	::mlir::StringAttr	string attribute

Operands:

Operand	Description
inp

Results:

Result	Description
«unnamed»

Types

InterfaceType

Syntax:

!ADF.interface<
  mlir::Type   # type
>

Parameters:

Parameter	C++ type	Description
type	mlir::Type

ParameterType

Syntax:

!ADF.parameter<
  mlir::Type   # type
>

Parameters:

Parameter	C++ type	Description
type	mlir::Type

StreamType

Syntax:

!ADF.stream<
  mlir::Type   # type
>

Parameters:

Parameter	C++ type	Description
type	mlir::Type

WindowType

Syntax:

!ADF.window<
  mlir::Type,   # type
  int,   # size
  int   # overlap
>

Parameters:

Parameter	C++ type	Description
type	mlir::Type
size	int
overlap	int

acc48Type

Syntax: !ADF.acc48

acc80Type

Syntax: !ADF.acc80

accfloatType

Syntax: !ADF.accfloat

cacc48Type

Syntax: !ADF.cacc48

cacc80Type

Syntax: !ADF.cacc80

caccfloatType

Syntax: !ADF.caccfloat

cfloatType

Syntax: !ADF.cfloat

cint16Type

Syntax: !ADF.cint16

cint32Type

Syntax: !ADF.cint32

floatType

Syntax: !ADF.float

int8Type

Syntax: !ADF.int8

int16Type

Syntax: !ADF.int16

int32Type

Syntax: !ADF.int32

int64Type

Syntax: !ADF.int64

uint8Type

Syntax: !ADF.uint8

uint16Type

Syntax: !ADF.uint16

uint32Type

Syntax: !ADF.uint32

uint64Type

Syntax: !ADF.uint64