skainet-lang-core/sk.ainet.lang.tensor/TensorView

TensorView

interface TensorView<T : DType, V> : Tensor<T, V> (source)

A tensor view interface that represents a zero-copy slice or subset of a parent tensor.

TensorView extends the base Tensor interface while adding view-specific functionality for efficient tensor slicing operations. Views provide memory-efficient access to portions of larger tensors without copying the underlying data.

Key Characteristics

Zero-Copy: Views reference the parent tensor's data without duplication
Shape Transformation: Views can have different shapes than their parent
Index Mapping: Coordinate transformation between view space and parent space
Memory Efficiency: Multiple views can reference the same underlying data

Usage Patterns

Views are particularly useful for:

Batch processing (extracting mini-batches from larger datasets)
Channel extraction (accessing specific feature maps in NCHW tensors)
Spatial regions (selecting rectangular areas in image tensors)
Sequence windows (sliding window operations on temporal data)

Type Parameters

T

the data type constraint extending DType, defining the numerical precision

V

the actual value type that will be stored and accessed

Inheritors

SlicedTensorView

Properties

abstract val indexMapping: IndexMapper

The index mapper responsible for coordinate transformation between view and parent.

abstract val parentTensor: Tensor<T, V>

The parent tensor that this view references.

open override val shape: Shape

Override the shape property to return the view-specific shape.

abstract val viewShape: Shape

The shape of this view, which may differ from the parent tensor's shape.

Functions

analyzeContiguity

fun <T : DType, V> TensorView<T, V>.analyzeContiguity(): ContiguityAnalysis

Provides a more detailed contiguity analysis with diagnostic information.

fun <T : DType, V> TensorView<T, V>.canMaterialize(): Boolean

Checks if this tensor view can be materialized using the default strategy.

fun <T : DType, V> TensorView<T, V>.canMaterialize(strategy: MaterializationStrategy<T, V>): Boolean

Checks if this tensor view can be materialized using a specific strategy.

copyMaterialize

fun <T : DType, V> TensorView<T, V>.copyMaterialize(): Tensor<T, V>

Explicitly copies this view into a standalone contiguous tensor.

estimateMaterializationCost

fun <T : DType, V> TensorView<T, V>.estimateMaterializationCost(): Long

Estimates the memory overhead of materializing this tensor view.

fun <T : DType, V> TensorView<T, V>.estimateMaterializationCost(strategy: MaterializationStrategy<T, V>): Long

Estimates the memory overhead of materializing this tensor view with a specific strategy.

fun <T : DType, V> TensorView<T, V>.isContiguous(): Boolean

Determines if this tensor view represents a contiguous memory layout.

fun <T : DType, V> TensorView<T, V>.materialize(): Tensor<T, V>

Materializes this tensor view into a standalone tensor using the default strategy.

fun <T : DType, V> TensorView<T, V>.materialize(strategy: MaterializationStrategy<T, V>): Tensor<T, V>

Materializes this tensor view using a specific materialization strategy.

fun <T : DType, V> TensorView<T, V>.realizeAlias(): Tensor<T, V>

Realizes this view as an alias — returns a tensor that shares the parent's backing data when the view is a simple contiguous slice.

fun <T : DType, V> TensorView<T, V>.sliceView(builder: TensorSliceBuilder<T, V>.() -> Unit): TensorView<T, V>

Creates a view of an existing tensor view, enabling view chaining.