Kahan algorithm based accumulator. More...

#include <SparseInverse.h>

Public Member Functions
Scalar	value ()

void	operator+= (Scalar increment)

Private Attributes
Scalar	_sum {}

Scalar	_correction {}

Detailed Description

template<typename Scalar>
class Eigen::KahanSum< Scalar >

Kahan algorithm based accumulator.

The Kahan sum algorithm guarantees to bound the error from floating point accumulation to a fixed value, regardless of the number of accumulations performed. Naive accumulation accumulates errors O(N), and pairwise O(logN). However pairwise also requires O(logN) memory while Kahan summation requires O(1) memory, but 4x the operations / latency.

NB! Do not enable associative math optimizations, they may cause the Kahan summation to be optimized out leaving you with naive summation again.

Member Function Documentation

◆ operator+=()

template<typename Scalar >

void Eigen::KahanSum< Scalar >::operator+= ( Scalar increment )

inline

                                     {
     const Scalar correctedIncrement = increment + _correction;
     const Scalar previousSum = _sum;
     _sum += correctedIncrement;
     _correction = correctedIncrement - (_sum - previousSum);
   }