Mila::Dnn::TensorShape Struct Reference

Fixed-capacity inline shape descriptor for N-dimensional tensors. More...

Public Member Functions
	TensorShape ()=default
	TensorShape (const dim_t *first, const dim_t *last)
	Constructs from a contiguous pointer range of dimension sizes.
	TensorShape (std::initializer_list< dim_t > il)
	Constructs from a brace-enclosed list of dimension sizes.
dim_t &	back ()
dim_t	back () const
dim_t *	begin ()
const dim_t *	begin () const
dim_t *	data ()
const dim_t *	data () const
bool	empty () const
dim_t *	end ()
const dim_t *	end () const
bool	operator== (const TensorShape &other) const noexcept
	Equality compares only the active ndim dimensions.
dim_t &	operator[] (size_t i)
dim_t	operator[] (size_t i) const
size_t	size () const

Public Attributes
std::array< dim_t, MaxRank >	dims {}
uint8_t	ndim { 0 }

Static Public Attributes
static constexpr uint8_t	MaxRank = 6

Detailed Description

Fixed-capacity inline shape descriptor for N-dimensional tensors.

Stores up to MaxRank dimension sizes inline with no heap allocation. The active rank is tracked by ndim; unused slots are zero-initialized and ignored by all operations. An empty shape (ndim == 0) represents a scalar tensor.

MaxRank = 6 covers all current Mila architectures (2-4D) and common research extensions including video and spatiotemporal models (5D) with one dimension of safety margin. shape_t, stride_t, and index_t are all aliases of TensorShape since all three are structurally identical fixed-capacity int64 sequences.

Example

shape_t s{ 2, 1024, 768 };  // [B, T, C]
s[0] == 2; s.size() == 3;
for (auto d : s) { ... }

Constructor & Destructor Documentation

TensorShape() [1/3]

Mila::Dnn::TensorShape::TensorShape ( )

default

Here is the caller graph for this function:

TensorShape() [2/3]

Mila::Dnn::TensorShape::TensorShape ( std::initializer_list< dim_t > il )

inline

Constructs from a brace-enclosed list of dimension sizes.

Parameters

il	Dimension sizes in row-major order.

Exceptions

std::invalid_argument

if il.size() > MaxRank.

TensorShape() [3/3]

Mila::Dnn::TensorShape::TensorShape ( const dim_t * first, const dim_t * last )

inline

Constructs from a contiguous pointer range of dimension sizes.

Covers serialization boundaries where shapes arrive as contiguous storage (e.g. std::vector<int64_t>::data()). The range [first, last) must not exceed MaxRank elements.

Parameters

first	Pointer to the first dimension size.
last	One-past-end pointer.

Exceptions

std::invalid_argument

if the range length exceeds MaxRank.

Member Function Documentation

back() [1/2]

dim_t & Mila::Dnn::TensorShape::back ( )

inline

back() [2/2]

dim_t Mila::Dnn::TensorShape::back ( ) const

inline

Here is the caller graph for this function:

begin() [1/2]

dim_t * Mila::Dnn::TensorShape::begin ( )

inline

begin() [2/2]

const dim_t * Mila::Dnn::TensorShape::begin ( ) const

inline

Here is the caller graph for this function:

data() [1/2]

dim_t * Mila::Dnn::TensorShape::data ( )

inline

data() [2/2]

const dim_t * Mila::Dnn::TensorShape::data ( ) const

inline

Here is the caller graph for this function:

empty()

bool Mila::Dnn::TensorShape::empty ( ) const

inline

Examples

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

Here is the caller graph for this function:

end() [1/2]

dim_t * Mila::Dnn::TensorShape::end ( )

inline

end() [2/2]

const dim_t * Mila::Dnn::TensorShape::end ( ) const

inline

Here is the caller graph for this function:

operator==()

bool Mila::Dnn::TensorShape::operator== ( const TensorShape & other ) const

inlinenoexcept

Equality compares only the active ndim dimensions.

Unused slots (indices >= ndim) are ignored regardless of their values.

Here is the call graph for this function:

operator[]() [1/2]

dim_t & Mila::Dnn::TensorShape::operator[] ( size_t i )

inline

operator[]() [2/2]

dim_t Mila::Dnn::TensorShape::operator[] ( size_t i ) const

inline

size()

size_t Mila::Dnn::TensorShape::size ( ) const

inline

Examples

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

Here is the caller graph for this function:

Member Data Documentation

dims

std::array<dim_t, MaxRank> Mila::Dnn::TensorShape::dims {}

MaxRank

uint8_t Mila::Dnn::TensorShape::MaxRank = 6

staticconstexpr

ndim

uint8_t Mila::Dnn::TensorShape::ndim { 0 }

---

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

• /__w/Mila/Mila/Mila/Src/Dnn/Tensors/Tensor.Types.ixx