#include <NetworkQuantizationScheme.hpp>

Inheritance diagram for QAsymmS8QuantizationScheme:

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 64 of file NetworkQuantizationScheme.hpp.

Member Function Documentation

◆ ComputeScheme()

OffsetScalePair ComputeScheme	(	double	min,
		double	max
	)		const

inlineoverridevirtual

Implements IQuantizationScheme.

Definition at line 66 of file NetworkQuantizationScheme.hpp.

References IQuantizationScheme::NumBits().

     {
         if (min > max)
         {
             throw InvalidArgumentException("min > max will result in invalid quantization.");
         }
 
         double highest = (1 << NumBits()) - 1;
 
         min = std::min(0.0, min); // min <= 0.0
         max = std::max(0.0, max); // max >= 0.0
 
         // To avoid dividing by zero when quantizing a zero filled tensor
         if (min == 0.0 && max == 0.0)
         {
             max = 1.0;
         }
 
         // Assumes quantization range [0-255]
         double scale = (max-min) / highest ;
         double offset = - min / scale;
 
         //Clamp 0 to Highest
         offset = std::max(0.0, std::min(highest, offset));
 
         //-128 on offset to cast to signed range
         return std::make_pair(static_cast<float>(scale), static_cast<int>(std::round(offset)-128));
     }