21 lines
975 B
Plaintext
21 lines
975 B
Plaintext
// KahanSum: the sum of a list of FLOAT values by compensated summation
|
|
// (Kahan, 1965). C holds the low-order part the last addition lost, and the
|
|
// next addition puts it back, so the error bound, about twice the unit
|
|
// roundoff times the sum of the absolute values (Higham, section 4.3), does
|
|
// not grow with the length of the list, to first order. It is relative to
|
|
// that sum of absolute values, not to the result: KahanSum([1.0e100, 1.0,
|
|
// -1.0e100]) is 0.0. In exact arithmetic C is always 0 and this is the
|
|
// plain sum: the compensation exists only because of rounding. A compiler
|
|
// allowed to reorder FLOAT arithmetic (-ffast-math) can simplify C away.
|
|
FUNCTION KahanSum(X : LIST OF FLOAT) RETURNS FLOAT
|
|
DECLARE Total : FLOAT <- 0.0
|
|
DECLARE C : FLOAT <- 0.0
|
|
FOR I <- 0 TO LENGTH(X)
|
|
DECLARE Y : FLOAT <- X[I] - C
|
|
DECLARE T : FLOAT <- Total + Y
|
|
C <- (T - Total) - Y
|
|
Total <- T
|
|
NEXT I
|
|
RETURN Total
|
|
ENDFUNCTION
|