Device-aware N-dimensional tensor. More...

Inheritance diagram for Mila::Dnn::Tensor< TDataType, TMemoryResource >:

Public Types
using	DataType = TensorDataType
	Abstract data type enumeration.
using	DataTypeTraits = TensorDataTypeTraits<TDataType>
	Compile-time data type characteristics.
using	host_value_t = std::conditional_t<TensorDataTypeTraits<TDataType>::is_integer_type, int32_t, float>
	Host value type for scalars.
using	MemoryResource = TMemoryResource
	Memory resource type for this tensor.

Public Member Functions
	Tensor (Compute::DeviceId device_id, const shape_t &shape, std::string name={})
	Creates a tensor bound to the specified device id and shape.
	Tensor (const Tensor &other)=delete
	Copy constructor - explicitly deleted for performance safety.
	Tensor (Tensor &&other) noexcept
	Efficiently transfers ownership from another tensor.
	~Tensor ()=default
	Destructor with automatic resource cleanup via RAII.
constexpr const auto *	data () const noexcept
	Returns type-safe immutable pointer to tensor data with concrete host type.
constexpr auto *	data () noexcept
	Returns type-safe pointer to tensor data with concrete host type.
size_t	elementSize () const override
	Returns the size in bytes of a single tensor element.
bool	empty () const
	Checks if the tensor contains no elements.
Tensor< TDataType, TMemoryResource > &	flatten ()
	Flattens tensor to 2D shape in-place.
TensorDataType	getDataType () const override
	Returns the tensor's abstract data type identifier.
std::string	getDataTypeName () const override
	Returns human-readable name of the tensor's data type.
Compute::DeviceId	getDeviceId () const override
Compute::DeviceType	getDeviceType () const override
	Returns the device type of this tensor's memory resource.
Compute::MemoryResource *	getMemoryResource () const override
	Returns pointer to the memory resource managing this tensor's storage.
std::string	getName () const override
	Returns the tensor's optional user-assigned name.
size_t	getStorageSize () const override
	Returns the total storage size in bytes backing the tensor's buffer.
std::string	getUId () const override
	Returns the tensor's unique identifier.
bool	isScalar () const noexcept override
	Checks if tensor is a scalar (0-dimensional).
bool	isValid () const noexcept
	Check if tensor is in a valid state (not moved-from).
bool	isView () const override
	Check if this tensor is a view of another tensor's data.
auto &	item ()
	Gets the scalar value for 0-dimensional tensors.
const auto &	item () const
	Gets the scalar value for 0-dimensional tensors (const version).
Tensor &	operator= (const Tensor &other)=delete
	Copy assignment operator - explicitly deleted.
Tensor &	operator= (Tensor &&other) noexcept
	Efficiently moves resources from another tensor.
auto &	operator[] (const index_t &indices)
	Accesses tensor element using multi-dimensional indices.
const auto &	operator[] (const index_t &indices) const
	Accesses tensor element using multi-dimensional indices (const version).
template<typename... Indices> requires TMemoryResource::is_host_accessible
auto &	operator[] (Indices... indices) &&(std
	Variadic index access (non-const).
template<typename... Indices> requires TMemoryResource::is_host_accessible
const auto &	operator[] (Indices... indices) const &&(std
	Variadic index access (const).
size_t	rank () const
	Returns the number of dimensions in the tensor.
const void *	rawData () const override
	Returns raw pointer to tensor data (const version).
void *	rawData () override
	Returns raw pointer to tensor data (implements ITensor protected API).
void	reshape (const shape_t &new_shape)
	Modifies tensor shape while preserving total element count.
void	setName (const std::string &value)
	Assigns a descriptive name to the tensor.
const shape_t &	shape () const override
	Returns the tensor's dimensional shape vector.
dim_t	size () const override
	Returns the total number of logical elements in the tensor.
const stride_t &	strides () const
	Returns the tensor's memory stride information.
std::string	toString (bool showBuffer=false) const
	Generates string representation of the tensor.
Tensor	view (const shape_t &new_shape, dim_t offset=0) const
Public Member Functions inherited from Mila::Dnn::ITensor
template<typename T>
bool	isType () const
	Type-safe check for tensor element type.

Static Public Member Functions
static constexpr bool	is_device_accessible ()
	Checks if tensor data is accessible from device (GPU) code.
static constexpr bool	is_host_accessible ()
	Checks if tensor data is accessible from host (CPU) code.

Detailed Description

template<TensorDataType TDataType, typename TMemoryResource>
requires isValidTensor<TDataType, TMemoryResource>
class Mila::Dnn::Tensor< TDataType, TMemoryResource >

Device-aware N-dimensional tensor.

Move-only tensor parameterized by an abstract TensorDataType and a MemoryResource. Scalars are represented by an empty shape ({}), which yields size() == 1.

Template Parameters

TDataType	Abstract tensor data type from TensorDataType enumeration
TMemoryResource	Memory resource type defining storage location and access patterns

Constructor & Destructor Documentation

◆ Tensor() [1/3]

template<TensorDataType TDataType, typename TMemoryResource>

Mila::Dnn::Tensor< TDataType, TMemoryResource >::Tensor	(	Compute::DeviceId	device_id,
		const shape_t &	shape,
		std::string	name = {} )

inlineexplicit

Creates a tensor bound to the specified device id and shape.

Constructs a tensor using a DeviceId. Empty shape {} creates a scalar (0D tensor) with size 1. Shape containing a zero produces an empty tensor.

Parameters

device_id	Device identifier (type + index)
shape	Vector defining the size of each dimension in row-major order
name	Optional tensor name for diagnostics (default: empty)

Exceptions

std::invalid_argument	If device id is invalid (checked by validateDeviceId)
std::runtime_error	If device type doesn't match memory resource
std::bad_alloc	If memory allocation fails

◆ Tensor() [2/3]

template<TensorDataType TDataType, typename TMemoryResource>

Mila::Dnn::Tensor< TDataType, TMemoryResource >::Tensor ( const Tensor< TDataType, TMemoryResource > & other )

delete

Copy constructor - explicitly deleted for performance safety.

Use clone() for deep copies or std::move() for ownership transfer.

◆ Tensor() [3/3]

template<TensorDataType TDataType, typename TMemoryResource>

Mila::Dnn::Tensor< TDataType, TMemoryResource >::Tensor ( Tensor< TDataType, TMemoryResource > && other )

inlinenoexcept

Efficiently transfers ownership from another tensor.

Moves resources from the source tensor, leaving it in a moved-from state.

Member Function Documentation

◆ data()

template<TensorDataType TDataType, typename TMemoryResource>

auto * Mila::Dnn::Tensor< TDataType, TMemoryResource >::data ( )

inlinenodiscardconstexprnoexcept

Returns type-safe pointer to tensor data with concrete host type.

Only available for host-accessible memory resources.

◆ elementSize()

template<TensorDataType TDataType, typename TMemoryResource>

size_t Mila::Dnn::Tensor< TDataType, TMemoryResource >::elementSize ( ) const

inlineoverridevirtual

Returns the size in bytes of a single tensor element.

Public Types

Public Member Functions

Static Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ Tensor() [1/3]

◆ Tensor() [2/3]

◆ Tensor() [3/3]

Member Function Documentation

◆ data()

◆ elementSize()

◆ empty()

◆ flatten()

◆ getDataType()

◆ getDataTypeName()

◆ getDeviceId()

◆ getDeviceType()

◆ getMemoryResource()

◆ getName()

◆ getStorageSize()

◆ getUId()

◆ isScalar()

◆ isView()

◆ item() [1/2]

◆ item() [2/2]

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ operator[]()

◆ rawData() [1/2]

◆ rawData() [2/2]

◆ reshape()

◆ setName()

◆ shape()

◆ size()

◆ toString()