Public Types
using	MR = TMemoryResource

using	scalar_t = std::variant< int64_t, int32_t, half, float >

Public Member Functions
	Tensor ()

	Tensor (const std::vector< size_t > &shape, std::shared_ptr< TElementType > data_ptr)
	Constructs a tensor with the given shape and a shared pointer to allocated memory.

	Tensor (const std::vector< size_t > &shape, TElementType value=TElementType{})
	Constructs a tensor with the given shape and initializes it with the specified value.

	Tensor (const Tensor &other)
	Copy constructor (creates a shallow copy).

	Tensor (Tensor &&other) noexcept
	Move constructor.

	~Tensor ()=default

TElementType	at (const std::vector< size_t > &indices) const

Tensor< TElementType, TMemoryResource >	clone () const
	Creates a deep copy of this tensor.

template<typename SrcMemoryResource >
void	copyFrom (const Tensor< TElementType, SrcMemoryResource > &src)
	Copies data from another tensor into this tensor.

auto	data ()
	Gets a pointer to the tensor data with memory type safety.

auto	data () const
	Gets a const pointer to the tensor data with memory type safety.

const bool	empty () const

void	fill (const TElementType &value)

Tensor< TElementType, TMemoryResource > &	flatten ()
	Flattens the tensor to a 2D tensor by combining all dimensions except the last one.

Tensor< TElementType, TMemoryResource >	flattened () const
	Creates a flattened copy of this tensor.

std::string	get_uid () const

std::string	getBufferString (size_t start_index=0, size_t depth=0) const

std::string	getName () const

Tensor &	operator= (const Tensor &other)
	Copy assignment operator.

Tensor &	operator= (Tensor &&other) noexcept
	Move assignment operator.

template<typename... Args>
TElementType &	operator[] (Args... args)

template<typename... Args>
const TElementType &	operator[] (Args... args) const

size_t	rank () const

TElementType *	raw_data ()
	Gets a raw pointer to the tensor data for use in CUDA kernels.

const TElementType *	raw_data () const
	Gets a const raw pointer to the tensor data for use in CUDA kernels.

void	reshape (const std::vector< size_t > &new_shape)

void	set (const std::vector< size_t > &indices, TElementType value)

void	setName (const std::string &value)

const std::vector< size_t > &	shape () const

size_t	size () const

const std::vector< size_t > &	strides () const

template<typename TNewMR >
Tensor< TElementType, TNewMR >	to () const
	Converts this tensor to use a different memory resource.

template<typename T = TElementType, typename TMR = TMemoryResource>
std::enable_if_t< std::is_same_v< T, half >, Tensor< float, TMR > >	toFloat () const
	Converts a half precision tensor to float.

template<typename T = TElementType, typename TMR = TMemoryResource>
std::enable_if_t< std::is_same_v< T, float >, Tensor< half, TMR > >	toHalf () const
	Converts a float tensor to half precision.

template<typename HostAccessibleMR = Compute::HostMemoryResource>
Tensor< TElementType, HostAccessibleMR >	toHostAccessible () const

std::string	toString (bool showBuffer=false) const

Static Public Member Functions
template<typename MR = TMemoryResource>
static constexpr bool	is_device_accessible ()

template<typename MR = TMemoryResource>
static constexpr bool	is_host_accessible ()

Private Member Functions
void	allocateBuffer (TElementType value)

size_t	computeIndex (const std::vector< size_t > &indices) const

std::string	outputBuffer (size_t index, size_t depth) const

std::string	set_uid ()

void	validateIndices (const std::vector< size_t > &indices, const std::string &method_name) const

Static Private Member Functions
static size_t	computeSize (const std::vector< size_t > &shape)

static std::vector< size_t >	computeStrides (const std::vector< size_t > &shape)

Private Attributes
std::shared_ptr< TensorBuffer< TElementType, TMemoryResource > >	buffer_ { nullptr }

std::shared_ptr< TElementType >	external_memory_ptr_ { nullptr }

std::string	name_

std::vector< size_t >	shape_ {}

size_t	size_ { 0 }

std::vector< size_t >	strides_ {}

std::string	uid_

Friends
std::ostream &	operator<< (std::ostream &os, const Tensor &tensor)

Member Typedef Documentation

◆ MR

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

using Mila::Dnn::Tensor< TElementType, TMemoryResource >::MR = TMemoryResource

◆ scalar_t

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

using Mila::Dnn::Tensor< TElementType, TMemoryResource >::scalar_t = std::variant<int64_t, int32_t, half, float>

Constructor & Destructor Documentation

◆ Tensor() [1/5]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Mila::Dnn::Tensor< TElementType, TMemoryResource >::Tensor	(	const std::vector< size_t > &	shape,
		TElementType	value = `TElementType{}`
	)

inlineexplicit

Constructs a tensor with the given shape and initializes it with the specified value.

This constructor initializes the tensor with the provided shape and fills it with the given value. If no value is provided, the tensor is initialized with the default value of the type TElementType.

Parameters

shape	The shape of the tensor.
value	The value to initialize the tensor with. Defaults to the default value of type TElementType.

◆ Tensor() [2/5]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Mila::Dnn::Tensor< TElementType, TMemoryResource >::Tensor	(	const std::vector< size_t > &	shape,
		std::shared_ptr< TElementType >	data_ptr
	)

inline

Constructs a tensor with the given shape and a shared pointer to allocated memory.

This constructor initializes the tensor with the provided shape and uses the given shared pointer to allocated memory. The tensor does not take ownership of the memory.

Parameters

shape	The shape of the tensor.
data_ptr	Shared pointer to the allocated memory.

◆ Tensor() [3/5]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Mila::Dnn::Tensor< TElementType, TMemoryResource >::Tensor ( )

inline

Here is the call graph for this function:

◆ Tensor() [4/5]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

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

inline

Copy constructor (creates a shallow copy).

This constructor creates a new tensor that shares the underlying data buffer with the original tensor. Modifications to one tensor's data will affect the other. For a deep, independent copy, use the clone() method instead.

Parameters

other The tensor to copy from.

◆ Tensor() [5/5]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

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

inlinenoexcept

Move constructor.

This constructor moves the contents of the given tensor to this tensor.

Parameters

other The tensor to move from.

◆ ~Tensor()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Mila::Dnn::Tensor< TElementType, TMemoryResource >::~Tensor ( )

default

Member Function Documentation

◆ allocateBuffer()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::allocateBuffer ( TElementType value )

inlineprivate

Here is the caller graph for this function:

◆ at()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

TElementType Mila::Dnn::Tensor< TElementType, TMemoryResource >::at ( const std::vector< size_t > & indices ) const

inline

Here is the call graph for this function:

◆ clone()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Tensor< TElementType, TMemoryResource > Mila::Dnn::Tensor< TElementType, TMemoryResource >::clone ( ) const

inline

Creates a deep copy of this tensor.

Unlike the copy constructor which shares the underlying buffer, this method creates a completely independent copy with its own data buffer.

Returns: Tensor<TElementType, TMemoryResource> A deep copy of this tensor

Here is the call graph for this function:

◆ computeIndex()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

size_t Mila::Dnn::Tensor< TElementType, TMemoryResource >::computeIndex ( const std::vector< size_t > & indices ) const

inlineprivate

Here is the caller graph for this function:

◆ computeSize()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

static size_t Mila::Dnn::Tensor< TElementType, TMemoryResource >::computeSize ( const std::vector< size_t > & shape )

inlinestaticprivate

Here is the call graph for this function:

Here is the caller graph for this function:

◆ computeStrides()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

static std::vector< size_t > Mila::Dnn::Tensor< TElementType, TMemoryResource >::computeStrides ( const std::vector< size_t > & shape )

inlinestaticprivate

Here is the call graph for this function:

Here is the caller graph for this function:

◆ copyFrom()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename SrcMemoryResource >

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::copyFrom ( const Tensor< TElementType, SrcMemoryResource > & src )

inline

Copies data from another tensor into this tensor.

This method copies the contents of the source tensor to this tensor. Both tensors must have the same shape. The data is copied using the appropriate memory transfer mechanism based on the source and destination memory resource types.

Template Parameters

SrcMemoryResource The memory resource type of the source tensor.

Parameters

src	The source tensor to copy data from.

Exceptions

std::runtime_error If the shapes don't match or if a CUDA memory transfer operation fails.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ data() [1/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

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

inline

Gets a pointer to the tensor data with memory type safety.

Returns a memory-type-aware pointer that prevents unsafe host access to device memory at compile time.

Returns: A TensorPtr wrapper that enforces memory access safety

Here is the caller graph for this function:

◆ data() [2/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

auto Mila::Dnn::Tensor< TElementType, TMemoryResource >::data ( ) const

inline

Gets a const pointer to the tensor data with memory type safety.

Returns a memory-type-aware pointer that prevents unsafe host access to device memory at compile time.

Returns: A const TensorPtr wrapper that enforces memory access safety

◆ empty()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

const bool Mila::Dnn::Tensor< TElementType, TMemoryResource >::empty ( ) const

inline

Here is the caller graph for this function:

◆ fill()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::fill ( const TElementType & value )

inline

◆ flatten()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Tensor< TElementType, TMemoryResource > & Mila::Dnn::Tensor< TElementType, TMemoryResource >::flatten ( )

inline

Flattens the tensor to a 2D tensor by combining all dimensions except the last one.

This method reshapes the tensor from [D1, D2, ..., Dn-1, Dn] to [D1*D2*...*Dn-1, Dn]. This is particularly useful for operations that treat all but the last dimension as batch dimensions, such as Fully Connected operations.

Returns: A reference to this tensor after flattening.

Here is the call graph for this function:

◆ flattened()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Tensor< TElementType, TMemoryResource > Mila::Dnn::Tensor< TElementType, TMemoryResource >::flattened ( ) const

inline

Creates a flattened copy of this tensor.

Similar to flatten(), but returns a new tensor instead of modifying this one.

Returns: A new tensor with flattened shape.

Here is the call graph for this function:

◆ get_uid()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::get_uid ( ) const

inline

◆ getBufferString()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::getBufferString	(	size_t	start_index = `0`,
		size_t	depth = `0`
	)		const

inline

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getName()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::getName ( ) const

inline

Here is the caller graph for this function:

◆ is_device_accessible()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename MR = TMemoryResource>

static constexpr bool Mila::Dnn::Tensor< TElementType, TMemoryResource >::is_device_accessible ( )

inlinestaticconstexpr

◆ is_host_accessible()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename MR = TMemoryResource>

static constexpr bool Mila::Dnn::Tensor< TElementType, TMemoryResource >::is_host_accessible ( )

inlinestaticconstexpr

◆ operator=() [1/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Tensor & Mila::Dnn::Tensor< TElementType, TMemoryResource >::operator= ( const Tensor< TElementType, TMemoryResource > & other )

inline

Copy assignment operator.

This operator copies the contents of the given tensor to this tensor.

Parameters

other The tensor to copy from.

Returns: Tensor& A reference to this tensor.

◆ operator=() [2/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

Tensor & Mila::Dnn::Tensor< TElementType, TMemoryResource >::operator= ( Tensor< TElementType, TMemoryResource > && other )

inlinenoexcept

Move assignment operator.

This operator moves the contents of the given tensor to this tensor.

Parameters

other The tensor to move from.

Returns: Tensor& A reference to this tensor.

◆ operator[]() [1/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename... Args>

TElementType & Mila::Dnn::Tensor< TElementType, TMemoryResource >::operator[] ( Args... args )

inline

Here is the call graph for this function:

◆ operator[]() [2/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename... Args>

const TElementType & Mila::Dnn::Tensor< TElementType, TMemoryResource >::operator[] ( Args... args ) const

inline

Here is the call graph for this function:

◆ outputBuffer()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::outputBuffer	(	size_t	index,
		size_t	depth
	)		const

inlineprivate

Here is the call graph for this function:

Here is the caller graph for this function:

◆ rank()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

size_t Mila::Dnn::Tensor< TElementType, TMemoryResource >::rank ( ) const

inline

Here is the caller graph for this function:

◆ raw_data() [1/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

TElementType * Mila::Dnn::Tensor< TElementType, TMemoryResource >::raw_data ( )

inline

Gets a raw pointer to the tensor data for use in CUDA kernels.

This method is intended for internal use by CUDA operation implementations. Use with caution as it bypasses memory type safety.

Returns: Raw pointer to the tensor data

Here is the caller graph for this function:

◆ raw_data() [2/2]

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

const TElementType * Mila::Dnn::Tensor< TElementType, TMemoryResource >::raw_data ( ) const

inline

Gets a const raw pointer to the tensor data for use in CUDA kernels.

This method is intended for internal use by CUDA operation implementations. Use with caution as it bypasses memory type safety.

Returns: Const raw pointer to the tensor data

◆ reshape()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::reshape ( const std::vector< size_t > & new_shape )

inline

Here is the call graph for this function:

Here is the caller graph for this function:

◆ set()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::set	(	const std::vector< size_t > &	indices,
		TElementType	value
	)

inline

Here is the call graph for this function:

◆ set_uid()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::set_uid ( )

inlineprivate

Here is the call graph for this function:

◆ setName()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::setName ( const std::string & value )

inline

Here is the caller graph for this function:

◆ shape()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

const std::vector< size_t > & Mila::Dnn::Tensor< TElementType, TMemoryResource >::shape ( ) const

inline

Here is the caller graph for this function:

◆ size()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

size_t Mila::Dnn::Tensor< TElementType, TMemoryResource >::size ( ) const

inline

Here is the caller graph for this function:

◆ strides()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

const std::vector< size_t > & Mila::Dnn::Tensor< TElementType, TMemoryResource >::strides ( ) const

inline

Here is the caller graph for this function:

◆ to()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename TNewMR >

Tensor< TElementType, TNewMR > Mila::Dnn::Tensor< TElementType, TMemoryResource >::to ( ) const

inline

Converts this tensor to use a different memory resource.

This method creates a new tensor with the same shape and data but using a different memory resource. The data is copied from the current tensor to the new tensor using the appropriate memory transfer mechanism based on the source and destination memory resource types.

Template Parameters

TNewMR The target memory resource type.

Returns: Tensor<TElementType, TNewMR> A new tensor with the specified memory resource type.

Exceptions

std::runtime_error If a CUDA memory transfer operation fails.

Here is the call graph for this function:

◆ toFloat()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename T = TElementType, typename TMR = TMemoryResource>

std::enable_if_t< std::is_same_v< T, half >, Tensor< float, TMR > > Mila::Dnn::Tensor< TElementType, TMemoryResource >::toFloat ( ) const

inline

Converts a half precision tensor to float.

This method creates a new tensor with the same shape but with float precision elements. The conversion is performed in a type-safe manner, handling both host and device memory appropriately.

Template Parameters

T	The original element type (must be half)
TMR	The memory resource type

Returns: Tensor<float, TMR> A new tensor with float precision values

Exceptions

std::runtime_error If the source tensor is not of half type or if a CUDA operation fails

Here is the call graph for this function:

◆ toHalf()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename T = TElementType, typename TMR = TMemoryResource>

std::enable_if_t< std::is_same_v< T, float >, Tensor< half, TMR > > Mila::Dnn::Tensor< TElementType, TMemoryResource >::toHalf ( ) const

inline

Converts a float tensor to half precision.

This method creates a new tensor with the same shape but with half precision elements. The conversion is performed in a type-safe manner, handling both host and device memory appropriately.

Template Parameters

T	The original element type (must be float)
TMR	The memory resource type

Returns: Tensor<half, TMR> A new tensor with half precision values

Exceptions

std::runtime_error If the source tensor is not of float type or if a CUDA operation fails

Here is the call graph for this function:

◆ toHostAccessible()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

template<typename HostAccessibleMR = Compute::HostMemoryResource>

Tensor< TElementType, HostAccessibleMR > Mila::Dnn::Tensor< TElementType, TMemoryResource >::toHostAccessible ( ) const

inline

◆ toString()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::toString ( bool showBuffer = false ) const

inline

Here is the call graph for this function:

◆ validateIndices()

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

void Mila::Dnn::Tensor< TElementType, TMemoryResource >::validateIndices	(	const std::vector< size_t > &	indices,
		const std::string &	method_name
	)		const

inlineprivate

Here is the caller graph for this function:

Friends And Related Symbol Documentation

◆ operator<<

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::ostream & operator<<	(	std::ostream &	os,
		const Tensor< TElementType, TMemoryResource > &	tensor
	)

friend

Member Data Documentation

◆ buffer_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::shared_ptr<TensorBuffer<TElementType, TMemoryResource> > Mila::Dnn::Tensor< TElementType, TMemoryResource >::buffer_ { nullptr }

private

◆ external_memory_ptr_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::shared_ptr<TElementType> Mila::Dnn::Tensor< TElementType, TMemoryResource >::external_memory_ptr_ { nullptr }

private

◆ name_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::name_

private

◆ shape_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::vector<size_t> Mila::Dnn::Tensor< TElementType, TMemoryResource >::shape_ {}

private

◆ size_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

size_t Mila::Dnn::Tensor< TElementType, TMemoryResource >::size_ { 0 }

private

◆ strides_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::vector<size_t> Mila::Dnn::Tensor< TElementType, TMemoryResource >::strides_ {}

private

◆ uid_

template<typename TElementType , typename TMemoryResource = Compute::CudaMemoryResource>

std::string Mila::Dnn::Tensor< TElementType, TMemoryResource >::uid_

private

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

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Tensor/Tensor.ixx

Public Types

Public Member Functions

Static Public Member Functions

Private Member Functions

Static Private Member Functions

Private Attributes

Friends

Member Typedef Documentation

◆ MR

◆ scalar_t

Constructor & Destructor Documentation

◆ Tensor() [1/5]

◆ Tensor() [2/5]

◆ Tensor() [3/5]

◆ Tensor() [4/5]

◆ Tensor() [5/5]

◆ ~Tensor()

Member Function Documentation

◆ allocateBuffer()

◆ at()

◆ clone()

◆ computeIndex()

◆ computeSize()

◆ computeStrides()

◆ copyFrom()

◆ data() [1/2]

◆ data() [2/2]

◆ empty()

◆ fill()

◆ flatten()

◆ flattened()

◆ get_uid()

◆ getBufferString()

◆ getName()

◆ is_device_accessible()

◆ is_host_accessible()

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ operator[]() [1/2]

◆ operator[]() [2/2]

◆ outputBuffer()

◆ rank()

◆ raw_data() [1/2]

◆ raw_data() [2/2]

◆ reshape()

◆ set()

◆ set_uid()

◆ setName()

◆ shape()

◆ size()

◆ strides()

◆ to()

◆ toFloat()

◆ toHalf()

◆ toHostAccessible()

◆ toString()

◆ validateIndices()

Friends And Related Symbol Documentation

◆ operator<<

Member Data Documentation

◆ buffer_

◆ external_memory_ptr_

◆ name_

◆ shape_

◆ size_

◆ strides_

◆ uid_