Package io.github.kirstenali.deepj.tensor

Class TensorStorageAccess

java.lang.Object

io.github.kirstenali.deepj.tensor.TensorStorageAccess

Direct Known Subclasses:

TensorManipulations

public abstract class TensorStorageAccess
extends Object

Shared physical storage plumbing; logical Tensor operations remain on the concrete facade.

Method Summary

Modifier and Type	Method	Description
final boolean	coversStorage()	Whether immutable logical geometry covers the whole allocation contiguously from offset zero.
final DType	dtype()
final float	get(int... indices)
final float	get(int row, int column)
final Object	getGpuTag()
final GradientState	gradientState()
final boolean	hasHostStorage()	False for lazy storage that has never been touched on the host; such storage reads as zeros.
final boolean	hasSameShape(TensorStorageAccess other)
final boolean	hasShape(int... dimensions)	Tests exact logical dimensions without copying shape metadata or reading tensor values.
final boolean	isContiguous()
abstract void	materialize()
final int	matrixCols()
final int	matrixRows()	Checked rank-two dimensions for matrix kernels; other ranks are never implicitly flattened.
final int	numel()
final int	rank()
final float[]	rawData()	Unsafe backing access bypasses mutation tracking; use managed Tensor get/set for application code.
final void	releaseDeviceBufferRetention()
final boolean	retainsDeviceBuffer()
final void	set(int row, int column, float value)
final void	setAt(float value, int... indices)
final void	setGpuTag(Object tag)
final int[]	shape()
final int	size(int axis)
final long	storageBytes()	Physical bytes in the shared allocation; views may address only part of it.
final int	storageOffset()
final long[]	strides()
final float[]	toFloatArray()	Returns a logical row-major copy, converting compact storage to FLOAT32 values when necessary.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

coversStorage

public final boolean coversStorage()

Whether immutable logical geometry covers the whole allocation contiguously from offset zero.

rank

public final int rank()

shape

public final int[] shape()

hasShape

public final boolean hasShape(int... dimensions)

Tests exact logical dimensions without copying shape metadata or reading tensor values.

size

public final int size(int axis)

numel

public final int numel()

hasSameShape

public final boolean hasSameShape(TensorStorageAccess other)

strides

public final long[] strides()

storageOffset

public final int storageOffset()

isContiguous

public final boolean isContiguous()

storageBytes

public final long storageBytes()

Physical bytes in the shared allocation; views may address only part of it.

matrixRows

public final int matrixRows()

Checked rank-two dimensions for matrix kernels; other ranks are never implicitly flattened.

matrixCols

public final int matrixCols()

materialize

public abstract void materialize()

get

public final float get(int row,
 int column)

get

public final float get(int... indices)

set

public final void set(int row,
 int column,
 float value)

setAt

public final void setAt(float value,
 int... indices)

toFloatArray

public final float[] toFloatArray()

Returns a logical row-major copy, converting compact storage to FLOAT32 values when necessary.

gradientState

public final GradientState gradientState()

getGpuTag

public final Object getGpuTag()

setGpuTag

public final void setGpuTag(Object tag)

dtype

public final DType dtype()

hasHostStorage

public final boolean hasHostStorage()

False for lazy storage that has never been touched on the host; such storage reads as zeros.

rawData

public final float[] rawData()

Unsafe backing access bypasses mutation tracking; use managed Tensor get/set for application code.

retainsDeviceBuffer

public final boolean retainsDeviceBuffer()

releaseDeviceBufferRetention

public final void releaseDeviceBufferRetention()