Class ComputeTensorData

Tensor data storage for GPU backends

Inheritance
object
ComputeTensorData
Implements
Inherited Members
Namespace: Unity.Barracuda
Assembly: Unity.Barracuda.dll
Syntax
public class ComputeTensorData : ITensorData, IDisposable

Constructors

ComputeTensorData(TensorShape, string, ChannelsOrder, bool)

Create ComputeTensorData

Declaration
public ComputeTensorData(TensorShape shape, string buffername, ComputeInfo.ChannelsOrder onDeviceChannelsOrder, bool clearOnInit = true)
Parameters
Type Name Description
TensorShape shape

shape
string buffername

buffer name
ComputeInfo.ChannelsOrder onDeviceChannelsOrder

channel order
bool clearOnInit

clear on init

ComputeTensorData(ComputeBuffer, TensorShape, int, string, ChannelsOrder)

Create ComputeTensorData with specified buffer

Declaration
public ComputeTensorData(ComputeBuffer buffer, TensorShape shape, int offset, string buffername, ComputeInfo.ChannelsOrder onDeviceChannelsOrder)
Parameters
Type Name Description
ComputeBuffer buffer

buffer
TensorShape shape

shape
int offset

offset
string buffername

buffer name
ComputeInfo.ChannelsOrder onDeviceChannelsOrder

channels order

Fields

name

Parent Tensor name

Declaration
public string name
Field Value
Type Description
string

Properties

buffer

Data storage as ComputeBuffer

Declaration
public ComputeBuffer buffer { get; }
Property Value
Type Description
ComputeBuffer

channelsOrder

Channel order channels-first vs channels-last

Declaration
public ComputeInfo.ChannelsOrder channelsOrder { get; }
Property Value
Type Description
ComputeInfo.ChannelsOrder

maxCapacity

Returns the maximum number of element this tensorData can contain.

Declaration
public virtual int maxCapacity { get; }
Property Value
Type Description
int

offset

Offset in the data storage buffer

Declaration
public int offset { get; }
Property Value
Type Description
int

Methods

Dispose()

Dispose internal storage

Declaration
public virtual void Dispose()

Download(TensorShape)

Returns an array filled with the values of a tensor. Depending on the implementation and underlying device this array might be a copy or direct reference to the tensor values. This is a blocking call, unless data from device was requested via ScheduleAsyncDownload beforehand and has already arrived.

Declaration
public virtual float[] Download(TensorShape shape)
Parameters
Type Name Description
TensorShape shape

the TensorShape (and thus length) of the data to copy
Returns
Type Description
float[]

Tensor data as float arrary

~ComputeTensorData()

Finalizer

Declaration
protected ~ComputeTensorData()

Reserve(int)

Reserve uninitialized memory.

Declaration
public virtual void Reserve(int count)
Parameters
Type Name Description
int count

element count to reserve

ScheduleAsyncDownload(int)

Schedule an asynchronous download from device memory. count is the number of element to readback.

Declaration
public virtual bool ScheduleAsyncDownload(int count)
Parameters
Type Name Description
int count

count of elements to download
Returns
Type Description
bool

false until data from device arrives to CPU and is ready for access

SharedAccess(out int)

Returns an array filled with the values of multiple tensors that share the same tensorData on device. Depending on the implementation and underlying device this array might be a copy or direct reference to tensor values, no conversion from on device memory layout will occur. This is a blocking call, unless data from device was requested via ScheduleAsyncDownload beforehand and has already arrived.

Declaration
public virtual float[] SharedAccess(out int offset)
Parameters
Type Name Description
int offset

This function outputs offset from the beginning of the array to location of values for specific tensor. offset parameters is specified in float elements
Returns
Type Description
float[]

array filled with the values of multiple tensors that share the same tensorData on device

ToString()

Summary

Declaration
public override string ToString()
Returns
Type Description
string

summary
Overrides

Upload(float[], TensorShape, int)

Initialize with data. shape is the TensorShape (and thus length) of the data to copy. managedBufferStartIndex is the offset where to start the copy in the data

Declaration
public virtual void Upload(float[] data, TensorShape shape, int managedBufferStartIndex = 0)
Parameters
Type Name Description
float[] data

data as float array
TensorShape shape

Tensor shape
int managedBufferStartIndex

managed buffer start index

Implements

Extension Methods