LSSTApplications  18.0.0+106,18.0.0+50,19.0.0,19.0.0+1,19.0.0+10,19.0.0+11,19.0.0+13,19.0.0+17,19.0.0+2,19.0.0-1-g20d9b18+6,19.0.0-1-g425ff20,19.0.0-1-g5549ca4,19.0.0-1-g580fafe+6,19.0.0-1-g6fe20d0+1,19.0.0-1-g7011481+9,19.0.0-1-g8c57eb9+6,19.0.0-1-gb5175dc+11,19.0.0-1-gdc0e4a7+9,19.0.0-1-ge272bc4+6,19.0.0-1-ge3aa853,19.0.0-10-g448f008b,19.0.0-12-g6990b2c,19.0.0-2-g0d9f9cd+11,19.0.0-2-g3d9e4fb2+11,19.0.0-2-g5037de4,19.0.0-2-gb96a1c4+3,19.0.0-2-gd955cfd+15,19.0.0-3-g2d13df8,19.0.0-3-g6f3c7dc,19.0.0-4-g725f80e+11,19.0.0-4-ga671dab3b+1,19.0.0-4-gad373c5+3,19.0.0-5-ga2acb9c+2,19.0.0-5-gfe96e6c+2,w.2020.01
LSSTDataManagementBasePackage
Public Member Functions | List of all members
lsst::geom::polynomials::SafeSum< T > Class Template Reference

A numerically stable summation algorithm for floating-point numbers. More...

#include <SafeSum.h>

Public Member Functions

 SafeSum (T initial=static_cast< T >(0)) noexcept
 
SafeSumoperator= (T value) noexcept
 
SafeSumoperator+= (T value) noexcept
 
SafeSumoperator-= (T value) noexcept
 
 operator T () const noexcept
 

Detailed Description

template<typename T>
class lsst::geom::polynomials::SafeSum< T >

A numerically stable summation algorithm for floating-point numbers.

SafeSum implements += and -= operators that can be used to accumulate floating point numbers with very different magnitudes, with accuracy limited only by the usual floating-point inaccuracy in represented the final sum.

SafeSum is explicitly convertible to and from its underlying floating-point type and only supports in-place addition and subtraction, in order to avoid cases where a combination of implicit conversion and multiple overloaded operators could lead to accidental use of regular floating-point operations.

SafeSum uses the Kahan-Neumaier algorithm (though this should be considered an implementation detail by callers), which maintains a lower-order-bit correction that compensates for the loss of precision in the main sum. Particularly aggressive compiler optimizations that do not preserve IEEE floating point behavior (e.g. gcc's -fassociative-math) may optimize away the correction and reduce SafeSum's behavior to a standard unsafe sum.

Definition at line 62 of file SafeSum.h.

Constructor & Destructor Documentation

◆ SafeSum()

template<typename T>
lsst::geom::polynomials::SafeSum< T >::SafeSum ( initial = static_cast<T>(0))
inlineexplicitnoexcept

Definition at line 65 of file SafeSum.h.

65  :
66  _sum(initial),
67  _correction(static_cast<T>(0))
68  {}

Member Function Documentation

◆ operator T()

template<typename T>
lsst::geom::polynomials::SafeSum< T >::operator T ( ) const
inlineexplicitnoexcept

Definition at line 93 of file SafeSum.h.

93  {
94  return _sum + _correction;
95  }

◆ operator+=()

template<typename T>
SafeSum& lsst::geom::polynomials::SafeSum< T >::operator+= ( value)
inlinenoexcept

Definition at line 76 of file SafeSum.h.

76  {
77  T t = _sum + value;
78  // update _correction to account for lost low-order bits of the
79  // greater of _sum and value
80  if (std::abs(_sum) >= std::abs(value)) {
81  _correction += (_sum - t) + value;
82  } else {
83  _correction += (value - t) + _sum;
84  }
85  _sum = t;
86  return *this;
87  }
Angle abs(Angle const &a)
Definition: Angle.h:106

◆ operator-=()

template<typename T>
SafeSum& lsst::geom::polynomials::SafeSum< T >::operator-= ( value)
inlinenoexcept

Definition at line 89 of file SafeSum.h.

89  {
90  return operator+=(-value);
91  }
SafeSum & operator+=(T value) noexcept
Definition: SafeSum.h:76

◆ operator=()

template<typename T>
SafeSum& lsst::geom::polynomials::SafeSum< T >::operator= ( value)
inlinenoexcept

Definition at line 70 of file SafeSum.h.

70  {
71  _sum = value;
72  _correction = static_cast<T>(0);
73  return *this;
74  }

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