Detailed Description

TensorShape describes the shape of a Tensor.

A Tensor's shape is described by three parameters: its dimensions, alignment padding, and a DataLayout.

Dimensions: an N dimensional vector. The outermost dimension is at index 0, and innermost dimemnsion is at index N -1.
Alignment: used to zeropad the innermost dimension to be a multiple of a hardware backend's data alignment requirements (for example, if an accelerator expects 8-wide vectors, the innermost dimension must be a multiple of 8). This can be zero if no alignment is required.
DataLayout: the meaning of each dimension, e.g. "NCHW" means dimensions[0] = N, dimensions[3] = W, etc.

Definition at line 35 of file tensor.h.

Public Member Functions
	TensorShape (std::vector< int > _dims, DataLayout _layout, int _alignment=0)

	TensorShape (std::initializer_list< int > _dims, DataLayout _layout, int _alignment=0)

	TensorShape (const TensorShape &shape)

	TensorShape (const TensorShapeProto &shapeProto)

const std::vector< int > &	dims () const

const std::vector< int > &	padding () const
	Returns a vector of padding along each dimension.

int	operator[] (int index) const

int &	operator[] (int index)

int	getStorageDim (int index) const
	Returns the alignment-padded size of the specified dimension.

bool	operator== (const TensorShape &other) const

DataLayout	getLayout () const

int	ndims () const

int	size () const

int	storageSize () const

int	getAlignment () const

int	getPadding (int index) const

TensorShapeProto *	asTensorShapeProto ()
	Return a TensorShapeProto that serializes this TensorShape.

Protected Attributes
std::vector< int >	dims_

std::vector< int >	padding_
	Padding along each dimension. More...

DataLayout	layout

int	alignment

Member Data Documentation

std::vector<int> smaug::TensorShape::padding_

protected

Padding along each dimension.

Only the last element be nonzero.

Definition at line 99 of file tensor.h.

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