Public Member Functions
	QTensor (const std::vector< int > &dims, const unsigned char precision, const bool signbit=false)
	Creates a quantized tensor of the given dimension. More...

void	Resize (std::vector< int > dim_source)

void	SetBitAtIndex (const unsigned char bit, const size_t index, const bool value)

bool	GetBitAtIndex (const unsigned char bit, const size_t index) const

void	SetPrecision (const unsigned char precision)

void	SetSigned (const bool make_signed=true)

void	SetScale (const double scale)

void	SetBias (const double bias)

unsigned char *	mutable_data ()

const unsigned char *	data () const

size_t	size () const

unsigned char	alignment () const

unsigned char	precision () const

const vector< int > &	dims () const

bool	is_signed () const

int	ndim () const
	Returns the number of dimensions of the data.

size_t	aligned_size () const

size_t	nbytes () const

double	scale () const

double	bias () const

int	dim32 (const int i) const
	Returns the i-th dimension of the qtensor in int.

int	canonical_axis_index (int axis_index) const
	Returns the 'canonical' version of a (usually) user-specified axis, allowing for negative indexing (e.g., -1 for the last axis). More...

TIndex	size_from_dim (int k) const
	Return product of all dimensions starting from K.

TIndex	size_to_dim (int k) const
	Product of all dims up to.

Protected Attributes
std::vector< int >	dims_

size_t	size_ = 0

unsigned char	precision_ = CHAR_BIT

unsigned char	alignment_ = CHAR_BIT

std::shared_ptr< unsigned char >	data_

double	scale_

double	bias_

bool	signed_

size_t	capacity_ = 0

Detailed Description

template<class Context>
class caffe2::QTensor< Context >

Definition at line 17 of file qtensor.h.

Constructor & Destructor Documentation

◆ QTensor()

template<class Context>

caffe2::QTensor< Context >::QTensor	(	const std::vector< int > &	dims,
		const unsigned char	precision,
		const bool	signbit = `false`
	)

inlineexplicit

Creates a quantized tensor of the given dimension.

Note that the actual data allocation is not going to be carried out until the first time mutable_data() is called.

The underlying storage of the quantized tensor interleaves elements by bit depth.

Labeled memory for tensor of size 6, precision 3 [ E1[0] E2[0] E3[0] E4[0] E5[0] E6[0] ] // Least significant Bits [ E1[1] E2[1] E3[1] E4[1] E5[1] E6[1] ] [ E1[2] E2[2] E3[2] E4[2] E5[2] E6[2] ]

In the case of sign bits (see enable_sign argument), an extra bit per element is added:

Labeled memory for tensor of size 6, precision 3, sign bit enabled [ E1[0] E2[0] E3[0] E4[0] E5[0] E6[0] ] [ E1[1] E2[1] E3[1] E4[1] E5[1] E6[1] ] [ E1[2] E2[2] E3[2] E4[2] E5[2] E6[2] ] [ E1[s] E2[s] E3[s] E4[s] E5[s] E6[s] ] Where 's' is 1 if E is negative

The reason for this layout is the ability to efficiently multiply many low precision integers as a sum of popcnt(A & B) * 1 << bit. Explained here: https://arxiv.org/abs/1606.06160

Definition at line 49 of file qtensor.h.

Member Function Documentation

◆ canonical_axis_index()

template<class Context>

int caffe2::QTensor< Context >::canonical_axis_index ( int axis_index ) const

inline

Returns the 'canonical' version of a (usually) user-specified axis, allowing for negative indexing (e.g., -1 for the last axis).

Parameters

axis_index the axis index. If 0 <= index < ndim(), return index. If -ndim <= index <= -1, return (ndim() - (-index)), e.g., the last axis index (ndim() - 1) if index == -1, the second to last if index == -2, etc. Dies on out of range index.

Definition at line 201 of file qtensor.h.

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

caffe2/core/qtensor.h

Public Member Functions

Protected Attributes