Deep::Tensor Class Reference

A device-aware, RAII-managed, move-only buffer – see the file-level note above for the full ownership/device contract. More...

#include <Tensor.h>

Public Member Functions
	Tensor (IComputeBackend *backend, DeviceType device, size_t numFloats)
	Allocates numFloats on the given backend/device, zero-initialized.
	Tensor (IComputeBackend *backend, DeviceType device, const float *hostData, size_t numFloats)
	Allocates numFloats on the given backend/device, then copies hostData in via backend->CopyFromHost(). This is the "load weights saved on CPU, run on GPU" pattern – works identically for device=DEVICE_CPU too, where CopyFromHost() is just a memcpy.
	Tensor (IComputeBackend *backend, DeviceType device, const std::vector< float > &hostData)
	Same as above, taking a std::vector directly – the expected common case (read a saved model's weights into a std::vector, then hand it straight to this constructor).
	~Tensor ()
	Frees the underlying device allocation, if any (not called on a moved-from Tensor – see the move constructor).
	Tensor (const Tensor &)=delete
Tensor &	operator= (const Tensor &)=delete
	Tensor (Tensor &&other) noexcept
	Move constructor: takes over other's allocation, leaving it empty (Free()-safe, zero size).
Tensor &	operator= (Tensor &&other) noexcept
	Move assignment: frees this Tensor's current allocation (if any), then takes over other's, leaving it empty.
float *	Data () noexcept
	Returns the raw device buffer. Only safe to dereference directly from code that already knows it's running on the same device this Tensor lives on (e.g. CPUBackend's own methods for a DEVICE_CPU Tensor) – never dereference a DEVICE_GPU Tensor's Data() from host code directly.
const float *	Data () const noexcept
	const overload of Data() above.
size_t	Size () const noexcept
	Number of floats this Tensor holds.
DeviceType	Device () const noexcept
	Which device this Tensor lives on.
IComputeBackend *	Backend () const noexcept
	The backend this Tensor was allocated through.
void	CopyToHost (float *hostDst) const noexcept
	Copies this Tensor's data out to a host buffer, e.g. for Save()/inspection. A no-op memcpy if this Tensor is already DEVICE_CPU.
void	CopyToHost (std::vector< float > &hostDst) const
	Same as above, resizing hostDst to fit if needed.

Detailed Description

A device-aware, RAII-managed, move-only buffer – see the file-level note above for the full ownership/device contract.

Constructor & Destructor Documentation

Tensor() [1/5]

Deep::Tensor::Tensor ( IComputeBackend * backend, DeviceType device, size_t numFloats )

inline

Allocates numFloats on the given backend/device, zero-initialized.

Parameters

backend	Non-owning pointer to the backend that will perform the allocation; must outlive this Tensor.
device	Which device this Tensor lives on – must match what backend itself actually allocates on (a CPUBackend paired with DEVICE_GPU, or vice versa, is a caller error).
numFloats	Number of floats to allocate.

Tensor() [2/5]

Deep::Tensor::Tensor ( IComputeBackend * backend, DeviceType device, const float * hostData, size_t numFloats )

inline

Allocates numFloats on the given backend/device, then copies hostData in via backend->CopyFromHost(). This is the "load weights saved on CPU, run on GPU" pattern – works identically for device=DEVICE_CPU too, where CopyFromHost() is just a memcpy.

Tensor() [3/5]

Deep::Tensor::Tensor ( IComputeBackend * backend, DeviceType device, const std::vector< float > & hostData )

inline

Same as above, taking a std::vector directly – the expected common case (read a saved model's weights into a std::vector, then hand it straight to this constructor).

~Tensor()

Deep::Tensor::~Tensor ( )

inline

Frees the underlying device allocation, if any (not called on a moved-from Tensor – see the move constructor).

Tensor() [4/5]

Deep::Tensor::Tensor ( const Tensor & )

delete

Tensor() [5/5]

Deep::Tensor::Tensor ( Tensor && other )

inlinenoexcept

Move constructor: takes over other's allocation, leaving it empty (Free()-safe, zero size).

Member Function Documentation

Backend()

IComputeBackend * Deep::Tensor::Backend ( ) const

inlinenoexcept

The backend this Tensor was allocated through.

CopyToHost() [1/2]

void Deep::Tensor::CopyToHost ( float * hostDst ) const

inlinenoexcept

Copies this Tensor's data out to a host buffer, e.g. for Save()/inspection. A no-op memcpy if this Tensor is already DEVICE_CPU.

Parameters

hostDst

Must point to at least Size() floats.

CopyToHost() [2/2]

void Deep::Tensor::CopyToHost ( std::vector< float > & hostDst ) const

inline

Same as above, resizing hostDst to fit if needed.

Data() [1/2]

const float * Deep::Tensor::Data ( ) const

inlinenoexcept

const overload of Data() above.

Data() [2/2]

float * Deep::Tensor::Data ( )

inlinenoexcept

Returns the raw device buffer. Only safe to dereference directly from code that already knows it's running on the same device this Tensor lives on (e.g. CPUBackend's own methods for a DEVICE_CPU Tensor) – never dereference a DEVICE_GPU Tensor's Data() from host code directly.

Device()

DeviceType Deep::Tensor::Device ( ) const

inlinenoexcept

Which device this Tensor lives on.

operator=() [1/2]

Tensor & Deep::Tensor::operator= ( const Tensor & )

delete

operator=() [2/2]

Tensor & Deep::Tensor::operator= ( Tensor && other )

inlinenoexcept

Move assignment: frees this Tensor's current allocation (if any), then takes over other's, leaving it empty.

Size()

size_t Deep::Tensor::Size ( ) const

inlinenoexcept

Number of floats this Tensor holds.

---

The documentation for this class was generated from the following file:

• include/deepity/backend/Tensor.h