Interface ITensorAllocator

Interfaces for tensor allocator

Namespace: Unity.Barracuda

Syntax

public interface ITensorAllocator : IDisposable

Methods

Alloc(TensorShape, AllocScope, DataType)

Allocate

Declaration

Tensor Alloc(TensorShape shape, AllocScope scope = AllocScope.LayerOutput, DataType dataType = DataType.Float)

Parameters

Type	Name	Description
TensorShape	shape	shape
AllocScope	scope	tensor lifetime scope
DataType	dataType	tensor data type

Returns

Type	Description
Tensor	allocated Tensor

Alloc(TensorShape, ITensorData, AllocScope, DataType)

Allocate with existing ITensorData buffer

Declaration

Tensor Alloc(TensorShape shape, ITensorData buffer, AllocScope scope = AllocScope.LayerOutput, DataType dataType = DataType.Float)

Parameters

Type	Name	Description
TensorShape	shape	shape
ITensorData	buffer	buffer
AllocScope	scope	tensor lifetime scope
DataType	dataType

Returns

Type	Description
Tensor	allocated Tensor

MoveToDevice(Tensor, ITensorData, ITensorData, Boolean)

Move Tensor to device

Declaration

void MoveToDevice(Tensor x, ITensorData newBuffer, ITensorData oldBuffer, bool disposeDetachedBufferHint)

Parameters

Type	Name	Description
Tensor	x	Tensor
ITensorData	newBuffer	new buffer
ITensorData	oldBuffer	old buffer
Boolean	disposeDetachedBufferHint	dispose detached buffer hint

PostLayerCleanup()

Allows ITensorAllocator to run cleanup operations such as clearing temporary buffers only used in the scope of the last layer executed.

Declaration

void PostLayerCleanup()

Release(Tensor, Boolean)

Release Tensor

Declaration

void Release(Tensor x, bool calledFromTensorDispose)

Parameters

Type	Name	Description
Tensor	x	Tensor
Boolean	calledFromTensorDispose	called from tensor dispose flag

Reset(Boolean)

Reset allocator

Declaration

void Reset(bool keepCachedMemory)

Parameters

Type	Name	Description
Boolean	keepCachedMemory	keep cached memory flag

WaiveOwnership(Tensor)

Waive ownership

Declaration

void WaiveOwnership(Tensor x)

Parameters

Type	Name	Description
Tensor	x	Tensor