Abstract interface providing essential tensor information and data access. More...

Inheritance diagram for Mila::Dnn::ITensor:

[legend]

Public Member Functions
virtual	~ITensor ()=default
virtual size_t	elementSize () const =0
	Get the size in bytes of a single tensor element.
virtual TensorDataType	getDataType () const =0
	Get tensor element data type identifier.
virtual std::string	getDataTypeName () const =0
	Get human-readable name of the tensor's data type.
virtual Compute::DeviceId	getDeviceId () const =0
virtual Compute::DeviceType	getDeviceType () const =0
	Get the device type from the memory resource.
virtual Compute::MemoryResource *	getMemoryResource () const =0
	Get the memory resource managing this tensor's storage.
virtual std::string	getName () const =0
	Returns the tensor's optional user-assigned name.
virtual size_t	getStorageSize () const =0
	Get the total storage size in bytes backing the tensor's buffer.
virtual std::string	getUId () const =0
	Returns the tensor's unique identifier.
virtual bool	isScalar () const noexcept=0
	Check if the tensor is a scalar (rank 0).
template<typename T>
bool	isType () const
	Type-safe check for tensor element type.
virtual bool	isView () const =0
	Check if this tensor is a view of another tensor's data.
virtual const void *	rawData () const =0
	Get raw pointer to tensor data (const version).
virtual void *	rawData ()=0
	Get raw pointer to tensor data.
virtual const shape_t &	shape () const =0
	Get the tensor dimensional structure.
virtual size_t	size () const =0
	Get total number of elements in the tensor.

Detailed Description

Abstract interface providing essential tensor information and data access.

ITensor serves as a type-erased interface for accessing tensor data regardless of the specific element type or memory resource implementation. This interface enables polymorphic access to tensor properties and data, making it possible to write generic code that works with different tensor implementations.

Memory access levels:

Public Tensor::data() : Type-safe, host-accessible tensors only.
Protected rawData() : Type-erased void pointer for TensorOps implementations.
Device kernels : Raw pointers passed as kernel parameters (not through this interface).

See also: Tensor

Examples: /__w/Mila/Mila/Mila/Src/Dnn/Components/Normalization/Softmax.ixx.

Constructor & Destructor Documentation

◆ ~ITensor()

virtual Mila::Dnn::ITensor::~ITensor ( )

virtualdefault

Member Function Documentation

◆ elementSize()

virtual size_t Mila::Dnn::ITensor::elementSize ( ) const

pure virtual

Get the size in bytes of a single tensor element.

Returns: Byte size for the tensor's data type (e.g. 4 for FP32, 2 for FP16).

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getDataType()

virtual TensorDataType Mila::Dnn::ITensor::getDataType ( ) const

pure virtual

Get tensor element data type identifier.

Returns: TensorDataType enumeration value for the element type.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getDataTypeName()

virtual std::string Mila::Dnn::ITensor::getDataTypeName ( ) const

pure virtual

Get human-readable name of the tensor's data type.

Returns: String representation of the element type (e.g. "FP32", "BF16").

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getDeviceId()

virtual Compute::DeviceId Mila::Dnn::ITensor::getDeviceId ( ) const

pure virtual

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getDeviceType()

virtual Compute::DeviceType Mila::Dnn::ITensor::getDeviceType ( ) const

pure virtual

Get the device type from the memory resource.

Returns: DeviceType enum value (Cpu, Cuda, etc.).

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getMemoryResource()

virtual Compute::MemoryResource * Mila::Dnn::ITensor::getMemoryResource ( ) const

pure virtual

Get the memory resource managing this tensor's storage.

Returns: Pointer to the memory resource (never null for valid tensors).

◆ getName()

virtual std::string Mila::Dnn::ITensor::getName ( ) const

pure virtual

Returns the tensor's optional user-assigned name.

Returns: Name string (may be empty if not assigned).

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getStorageSize()

virtual size_t Mila::Dnn::ITensor::getStorageSize ( ) const

pure virtual

Get the total storage size in bytes backing the tensor's buffer.

Authoritative byte count for raw memory operations (memset, memcpy, etc.).

Returns: Number of bytes occupied by the tensor's logical elements, or 0 if no buffer.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getUId()

virtual std::string Mila::Dnn::ITensor::getUId ( ) const

pure virtual

Returns the tensor's unique identifier.

Stable identifier assigned at construction, useful for diagnostics.

Returns: Unique identifier string (implementation-defined format).

Here is the call graph for this function:

Here is the caller graph for this function:

◆ isScalar()

virtual bool Mila::Dnn::ITensor::isScalar ( ) const

pure virtualnoexcept

Check if the tensor is a scalar (rank 0).

◆ isType()

template<typename T>

bool Mila::Dnn::ITensor::isType ( ) const

inline

Type-safe check for tensor element type.

Template Parameters

T	The element type to check against.

Returns: true if the tensor's elements are of type T.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ isView()

virtual bool Mila::Dnn::ITensor::isView ( ) const

pure virtual

Check if this tensor is a view of another tensor's data.

Returns: true if this is a view, false if it owns its buffer.

◆ rawData() [1/2]

virtual const void * Mila::Dnn::ITensor::rawData ( ) const

pure virtual

Get raw pointer to tensor data (const version).

Returns: Raw const void pointer to tensor data, or nullptr if empty.

◆ rawData() [2/2]

virtual void * Mila::Dnn::ITensor::rawData ( )

pure virtual

Get raw pointer to tensor data.

Type-erased access for TensorOps implementations. For device-only memory, this pointer must not be dereferenced on the host — it is only valid for device operations (kernel launches, cudaMemcpy, etc.).

Returns: Raw void pointer to tensor data, or nullptr if empty.

Here is the caller graph for this function:

◆ shape()

virtual const shape_t & Mila::Dnn::ITensor::shape ( ) const

pure virtual

Get the tensor dimensional structure.

Returns: Immutable reference to the TensorShape describing each dimension size. Length (ndim) equals the tensor's rank. Scalars return an empty shape.

Examples: /__w/Mila/Mila/Mila/Src/Dnn/Components/Normalization/Softmax.ixx.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ size()

virtual size_t Mila::Dnn::ITensor::size ( ) const

pure virtual

Get total number of elements in the tensor.

Returns: Product of all dimension sizes. Scalars return 1; empty tensors return 0.

Here is the call graph for this function:

Here is the caller graph for this function:

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

/__w/Mila/Mila/Mila/Src/Dnn/Tensors/ITensor.ixx

Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ ~ITensor()

Member Function Documentation

◆ elementSize()

◆ getDataType()

◆ getDataTypeName()

◆ getDeviceId()

◆ getDeviceType()

◆ getMemoryResource()

◆ getName()

◆ getStorageSize()

◆ getUId()

◆ isScalar()

◆ isType()

◆ isView()

◆ rawData() [1/2]

◆ rawData() [2/2]

◆ shape()

◆ size()