tenzr::ReduceWindowNode¶

#include <node.h>

Inherits from tenzr::Node, std::enable_shared_from_this< Node >

Public Functions¶

	Name
	ReduceWindowNode(std::shared_ptr< Node > input, std::shared_ptr< Node > init_value, std::vector< int64_t > window_dimensions, std::vector< int64_t > window_strides, std::vector< int64_t > padding_low, std::vector< int64_t > padding_high, std::vector< int64_t > window_dilations, std::string Reduction)
virtual std::string	OpName() const override StableHLO-style name of this node's operation.
const std::shared_ptr< Node > &	Input() const
const std::shared_ptr< Node > &	InitValue() const Initial value of the reduction.
const std::vector< int64_t > &	WindowDimensions() const Window extent per dimension.
const std::vector< int64_t > &	WindowStrides() const Window stride per dimension.
const std::vector< int64_t > &	PaddingLow() const Low edge padding per dimension.
const std::vector< int64_t > &	PaddingHigh() const High edge padding per dimension.
const std::vector< int64_t > &	WindowDilations() const Window dilation per dimension.
const std::string &	Reduction() const Name of the reduction, such as "add" or "max".

Additional inherited members¶

Public Functions inherited from tenzr::Node

	Name
	Node(NodeKind kind, Shape shape, DType dtype, ShardingSpec sharding =ShardingSpec::Replicated())
virtual	~Node() =default
NodeKind	Kind() const The node's operation kind.
const Shape &	shape() const The node's output shape.
DType	Dtype() const The node's output dtype.
const ShardingSpec &	Sharding() const The node's sharding spec.
void	SetSharding(ShardingSpec spec) Replace the sharding spec.
const StridedLayout &	Layout() const The node's strided layout.
void	SetLayout(StridedLayout layout) Replace the layout.
const std::vector< std::shared_ptr< Node > > &	Inputs() const The node's input operands.
uint64_t	Id() const The node's unique id.

Protected Functions inherited from tenzr::Node

	Name
uint64_t	NextId()

Protected Attributes inherited from tenzr::Node

	Name
NodeKind	kind_
Shape	shape_
DType	dtype_
ShardingSpec	sharding_
StridedLayout	layout_
std::vector< std::shared_ptr< Node > >	inputs_
uint64_t	id_

Detailed Description¶

class tenzr::ReduceWindowNode;

Sliding-window reduction over an operand, used to lower pooling operations.

Public Functions Documentation¶

function ReduceWindowNode¶

ReduceWindowNode(
    std::shared_ptr< Node > input,
    std::shared_ptr< Node > init_value,
    std::vector< int64_t > window_dimensions,
    std::vector< int64_t > window_strides,
    std::vector< int64_t > padding_low,
    std::vector< int64_t > padding_high,
    std::vector< int64_t > window_dilations,
    std::string Reduction
)

Parameters:

input Input node.
init_value Reduction identity.
window_dimensions Window extent per dimension.
window_strides Window stride per dimension.
padding_low Low padding per dimension.
padding_high High padding per dimension.
window_dilations Window dilation per dimension.
Reduction Reduction operator name.

Create a windowed reduction node, normalizing default window parameters.

function OpName¶

inline virtual std::string OpName() const override

StableHLO-style name of this node's operation.

Reimplements: tenzr::Node::OpName

function Input¶

inline const std::shared_ptr< Node > & Input() const

function InitValue¶

inline const std::shared_ptr< Node > & InitValue() const

Initial value of the reduction.

function WindowDimensions¶

inline const std::vector< int64_t > & WindowDimensions() const

Window extent per dimension.

function WindowStrides¶

inline const std::vector< int64_t > & WindowStrides() const

Window stride per dimension.

function PaddingLow¶

inline const std::vector< int64_t > & PaddingLow() const

Low edge padding per dimension.

function PaddingHigh¶

inline const std::vector< int64_t > & PaddingHigh() const

High edge padding per dimension.

function WindowDilations¶

inline const std::vector< int64_t > & WindowDilations() const

Window dilation per dimension.

function Reduction¶

inline const std::string & Reduction() const

Name of the reduction, such as "add" or "max".