Anonymous user
Talk:Safe addition: Difference between revisions
→Error propagation?: new section
(Summarize what my investigations into IEEE fp rounding leads me to understand) |
(→Error propagation?: new section) |
||
Line 4:
I misunderstood the original task, so my comments before are correct but distinctly off-topic. :-)<br>According to the documentation I can scare up with Google on IEEE rounding, the normal mode is round-to-nearest as that minimizes the error (to 0.5 ULP) given that its a single-valued result. The documentation also says that it is a really bad idea to call the standard math functions with the rounding mode set to anything else. Given that, the only sane method of implementing this task (assuming that the language has no portable way to set the rounding mode for the duration of the computation, e.g., [[C]] doesn't) is to compute and then widen with <code>nextafter</code>. —[[User:Dkf|Donal Fellows]] 08:56, 19 August 2009 (UTC)
== Error propagation? ==
Wouldn't it be more sensible to include error propagation in the task? Just two exact number adding to a sum +/- machine epsilon is of pretty limited use. --[[User:Ledrug|Ledrug]] 04:03, 10 June 2011 (UTC)
| |||