On Sat, Feb 8, 2014 at 3:39 PM, Nicolas Trangez wrote:

Note there's some existing support for this in `base`:

On Sat, 2014-02-08 at 15:06 -0500, Omari Norman wrote:

...

but for a short explanation, go into ghci and type:

print $ 0.1 + 0.1 + 0.1

λ print $ (0.1 + 0.1 + 0.1 :: Double) 0.30000000000000004

λ :m + Data.Fixed

Key word there is "Fixed"; deka supports floating point with up to 34 digits of precision but, yes, Data.Fixed does solve some of this problem.

On Sat, Feb 8, 2014 at 4:12 PM, Roman Cheplyaka wrote:

So it is fixed, too? I'm not sure "Fixed" qualifies as a key word, then :)

Note that Data.Fixed doesn't put any restrictions on the (fixed) precision. You could easily define E34.

No, it is not fixed.

...

...

digits . unDeka . fromJust . strToDeka $ "3.00" 3

...

...

digits . unDeka . fromJust . strToDeka $ "3.0000" 5

The precision is determined at runtime. In contrast, with Data.Fixed, the programmer sets the precision at compile time. There is no difference between

...

...

3 :: Fixed E6

and

...

...

3.00 :: Fixed E6

These issues are discussed at http://speleotrove.com/decimal/decifaq1.html#tzeros but if your reaction is "so what, it's just some extra zeroes and at compile time I can just use E34", then by all means, keep using Fixed :) Also, note that you get 34 significant digits, but that doesn't mean the exponent is limited to 34.

...

...

let x = fromJust . strToDeka $ "3E-200" let y = fromJust . strToDeka $ "4E-200" x + y Deka {unDeka = 7E-200}

...which you could do in Fixed, true. If you defined E200 first. But, oops, now I need E250... Really this is a binding to decNumber, which implements the General Decimal Arithmetic Specification, so see if decNumber and that specification scratch an itch. If they don't, use Fixed, or Decimal (also on Hackage) or something else...

On Sat, Feb 8, 2014 at 5:00 PM, Roman Cheplyaka wrote:

* Omari Norman [2014-02-08 16:45:02-0500]

...

Also, note that you get 34 significant digits, but that doesn't mean the exponent is limited to 34.

...

...

...

let x = fromJust . strToDeka $ "3E-200" let y = fromJust . strToDeka $ "4E-200" x + y Deka {unDeka = 7E-200}

...which you could do in Fixed, true. If you defined E200 first. But, oops, now I need E250...

I see now. So this is a decimal floating-point numbers library, which also attempts to track the precision of numbers. Does this sound right?

That's roughly right, yes. I would say"attempts" is too soft a word. decNumber and, therefore, deka, tracks every significant digit and at least maintains that significance. If it trashes significant digits without notifying you, it's a bug; please report it. The differences between Data.Fixed and this library are so huge (Fixed manipulates Integers alone; decNumber stores numbers as binary-coded decimal coefficients and integer exponents) that I didn't even think to compare them, and I won't be the one to draw performance comparisons between the two because deka is for situations where solutions like Fixed are simply wrong. Fixed does me no good, no matter how fast it is, so there's no reason for me to know how it performs.