And Tamas Papp writes: > My rule of thumb is that if I need to worry about the IEEE rounding mode > making a difference in my results, then my problem is really > ill-conditioned and rounding the least of my worries. So I usually just > ignore it.
You know of ill-conditioning, so you are not the concern. Round-to-even is meant to allay bias problems for the vast majority of programmers, like those responsible for the Vancouver stock exchange index in 1982. They used truncation which, over time, roughly halved the index. Rounding upward in the same scenario would have doubled it... The choices in 754 were made with a nearly revolting amount of deliberation. Always makes me sad to see them shrugged away.