Hacker News Reader

Comment by undershirt | Hacker News Reader

From a deleted comment I liked here from @stabbles:

> some things can be represented in finite digits in base x but require infinite digits in base y.

Very good summary. Binary to decimal is very straightforward until fractions require infinite digits. I don’t think dec64[1] is even a tradeoff—it’s just better. The significand stays a normal binary number— but it encodes the decimal point in gasp decimal. No infinities required for the numeric language that we all think in.

[1] https://en.wikipedia.org/wiki/Decimal64_floating-point_forma...

View on HN

> it's just better

This assertion does not withstand scrutiny. You may like being able to get True as the result of 0.1 + 0.2 == 0.3, but the landscape will still be littered with rounding errors as soon as you try to do anything nontrivial. (Or even plenty of trivial things like expecting 1/6 + 1/6 to add to 1/3). So all you gain is a false sense of security in exchange for less precision and slower computation.

(Of course, there are plenty of tasks for which floats are just wrong for the job, and you should transform the problem so that you can use integers or rationals instead. For example, when you are incrementing a number by (integer multiples of a) fixed delta, just change units so you can count numbers of increments as an integer, and change units back at the end.)

> ... dec64 is ...

Not to be confused with Douglas Crockford's DEC64 [1], which I believe is worse than binary floating points.

[1] https://www.crockford.com/dec64.html

Oh, thank you! I actually think I’m going with Crockford on this one, and that’s what I meant to post.

In which case I disagree ;-). Most strikingly DEC64 doesn't do normalization, so comparison will be a nightmare (as you have to normalize in order to compare!). He tried to special-case integer-only arguments, which hides the fact that non-integer cases are much, much slower thanks to added branches and complexity. If DEC64 were going to be "the only number type" in future languages, it had to be much better than this.

Good points! I think decimal64 doesn’t normalize the significand also. But I can’t assess what I haven’t used. Mine is a snap judgment in favor of understanding more of dec64 vs the wiki article. My general feeling is that it’s time for the scale of computing to tip away from total correctness and efficiency, and more toward non-awkward interfaces. But at bottom, I would try both and then talk about it.

Reply on news.ycombinator.com