#include <NetworkQuantizationScheme.hpp>

Inheritance diagram for QSymmS8QuantizationScheme:

Public Member Functions
OffsetScalePair	ComputeScheme (double min, double max) const override

int	NumBits () const override

DataType	GetDataType () const override

Public Member Functions inherited from IQuantizationScheme
virtual	~IQuantizationScheme ()

Detailed Description

Definition at line 100 of file NetworkQuantizationScheme.hpp.

Member Function Documentation

◆ ComputeScheme()

OffsetScalePair ComputeScheme	(	double	min,
		double	max
	)		const

inlineoverridevirtual

Implements IQuantizationScheme.

Definition at line 102 of file NetworkQuantizationScheme.hpp.

References IQuantizationScheme::NumBits().

     {
         if (min > max)
         {
             throw InvalidArgumentException("min > max will result in invalid quantization.");
         }
 
         // To avoid dividing by zero when quantizing a zero filled tensor
         if (min == 0.0 && max == 0.0)
         {
             max = 1.0;
         }
 
         double highest = (1 << (NumBits()-1)) - 1; // (numbits-1) accounts for the sign bit
 
         double extent = std::max(std::abs(min), std::abs(max));
         double scale = extent / highest;
 
         return std::make_pair(static_cast<float>(scale), 0);
     }