Int8Quantization Class Reference

Public Member Functions
template<typename T>
void	constrain_scale (T &scale) const noexcept
	Keep learned scales in the numerically safe, strictly-positive range.
template<typename T>
void	constrain_zero_point (T &zero_point) const noexcept
template<typename T>
void	constrain_parameters (T *scales, T *zero_points) const noexcept
	Project all learned parameters into this backend quantization scheme.

Static Public Member Functions
static constexpr Int8Quantization	activation () noexcept
	Per-tensor asymmetric activation quantization.
static constexpr Int8Quantization	weights (std::size_t axis_size, std::size_t inner_size) noexcept

Description

LiteRT/CMSIS-NN/Ethos-U signed int8 quantization layout.

Member Function Documentation

activation()

static constexpr Int8Quantization activation ( )

inlinestaticconstexprnoexcept

Per-tensor asymmetric activation quantization.

Returns

Asymmetric int8 layout with one scale and zero-point per tensor.

constrain_parameters()

template<typename T>

void constrain_parameters ( T * scales, T * zero_points ) const

inlinenoexcept

Project all learned parameters into this backend quantization scheme.

Template Parameters

T	Quantization parameter scalar type.

Parameters

[in,out]	scales	Array of parameter_count() learned scales.
[in,out]	zero_points	Array of parameter_count() learned zero-points.

constrain_scale()

template<typename T>

void constrain_scale ( T & scale ) const

inlinenoexcept

Keep learned scales in the numerically safe, strictly-positive range.

Template Parameters

T	Scale scalar type.

Parameters

[in,out]

scale

Learned scale to constrain.

constrain_zero_point()

template<typename T>

void constrain_zero_point ( T & zero_point ) const

inlinenoexcept

Keep learned zero-points in this scheme's integer domain. Symmetric weight quantization always forces them to zero.

Template Parameters

T	Zero-point scalar type.

Parameters

[in,out]

zero_point

Learned zero-point to constrain.

weights()

static constexpr Int8Quantization weights ( std::size_t axis_size, std::size_t inner_size )

inlinestaticconstexprnoexcept

Symmetric per-axis weight quantization.

Parameters

axis_size	Number of scales along the quantized axis.
inner_size	Product of dimensions following the quantized axis.

Returns

Symmetric int8 layout with one scale per axis entry.

axis_size is the number of scales and inner_size is the product of the dimensions following the quantized axis in a row-major tensor. Thus a fully-connected [output, input] matrix uses (output, input), Conv2D OHWI uses (output, H*W*I), and depthwise Conv2D uses (channels, 1).