"Printing Floating-Point Numbers Quickly and Accurately"

Discussion in 'Python' started by Michael Hudson, Jul 23, 2004.

1. Michael HudsonGuest

Does anyone (Tim?) have (ideally Python, but I can cope) code
implementing the fixed-format algorithm from the Subject: named paper
by Burger & Dybvig?

Cheers,
mwh

--
<Erwin> #python FAQ: How do I build X? A: Wait for twisted.X.
-- from Twisted.Quotes

Michael Hudson, Jul 23, 2004

2. Tim PetersGuest

[Michael Hudson]
> Does anyone (Tim?) have (ideally Python, but I can cope) code
> implementing the fixed-format algorithm from the Subject: named paper
> by Burger & Dybvig?

I don't, and you're too young if you think anyone else might <wink>.

The paper gives Scheme code, you know! And there's a Haskell variant here:

As the paper says at the end,

[David] Gay ... showed that floating-point arithmetic is sufficiently
accurate in most cases when the requested number of digits is small.
The fixed-format printing algorithm described in this paper is useful when
these heuristics fail.

IOW, the point of Burger & Dybvig was to run faster than the
algorithms in the earlier Steele & White paper, but David Gay's code
*usually* beats everything on speed (if you don't care about speed,
code for this task is quite simple; if you do care about speed, it's
mind-numbingly complicated), so there's little incentive to implement
this algorithm. Gay's code is written in C, & available from Netlib:

http://www.netlib.org/fp/

Tim Peters, Jul 24, 2004

3. Michael HudsonGuest

Tim Peters <> writes:

> [Michael Hudson]
> > Does anyone (Tim?) have (ideally Python, but I can cope) code
> > implementing the fixed-format algorithm from the Subject: named paper
> > by Burger & Dybvig?

>
> I don't, and you're too young if you think anyone else might <wink>.
>
> The paper gives Scheme code, you know!

Not for the fixed format algorithm it doesn't. The paper says:

The rational arithmetic used in fixed-format printing can wbe
converted into high-precision integer arithmetic by introducing a
common denominator as before. Because there are several more cases
to consider, however, the resulting code is lengthy and has
therefore been omitted from this paper.

I was hoping someone else had considered all the cases for me

I've already translated the free format code from the paper into
Python, unfortunately it's not what I actually need...

> And there's a Haskell variant here:
>

That's the free format algorithm again, unless I've gone blind.

> As the paper says at the end,
>
> [David] Gay ... showed that floating-point arithmetic is sufficiently
> accurate in most cases when the requested number of digits is small.
> The fixed-format printing algorithm described in this paper is useful when
> these heuristics fail.
>
> IOW, the point of Burger & Dybvig was to run faster than the
> algorithms in the earlier Steele & White paper, but David Gay's code
> *usually* beats everything on speed (if you don't care about speed,
> code for this task is quite simple; if you do care about speed, it's
> mind-numbingly complicated), so there's little incentive to implement
> this algorithm.

I'm probably suffering from an attack of perfectionism. Burger &
Dubvig's algorithm is so neat!

> Gay's code is written in C, & available from Netlib:
>
> http://www.netlib.org/fp/

Now *that* code is over-the-top, even for me

Cheers,
mwh

--
. <- the point your article -> .
|------------------------- a long way ------------------------|
-- Christophe Rhodes, ucam.chat

Michael Hudson, Jul 24, 2004