It also argues that most of these entries are not used for typical data with limited range. In any case, it claims at most a factor of two benefit from the “fast” method vs. the “small” method; I suspect that most cases requiring an exactly rounded sum are not performance-critical enough for this to matter. It would be nice to have an implementation of at least the “small” method in a Julia package for people who need exactly rounded floating-point sums.
1 Like