Hacker News Reader

In my case I'm more interested in the positions where floating point accuracy is lost: I'm currently working on a huge code base where we moved from double to single precision. If a unittest does not validate in single precision I want to have a tool telling me where and why.

View on HN

It seems to be a very wrong decision moving a "huge" base from double to single.

The sane approach is to move only that which must be moved, and there must be a valid reason for the change. Moving "huge" seems wrong by default, if the subject is numeric. It's not comparable with renaming your variables or methods or whatever you can do on the huge base and still keep the result the same.

That's not a trivial question to answer. I suggest reading about unums in John Gustafsson's "the end of error". It includes nice failures of floats.

Disclaimer: I'm working on an implementation

Reply on news.ycombinator.com