Re: [Caml-list] Comparing floats

[Goswin von Brederlow <goswin-v-b@web.de>]

On Thu, Jul 23, 2015 at 11:54:38AM +0200, Mr. Herr wrote:
> 
> 
> On 23.07.2015 11:01, Xavier Leroy wrote:
> >On 23/07/2015 10:35, Sébastien Hinderer wrote:
> >>What's the most efficient way to compare floats, please?
> >>Is it the polymorphic compare function, or is there a more specialized
> >>version of it?
> >You'll get good performance by type-specializing Pervasives.compare:
> >
> >let compare_float (x: float) (y: float) = compare x y
> >
> >If you're absolutely sure your floats are not NaN, you can shave a few
> >CPU cycles:
> >
> >let compare_float (x: float) (y: float) =
> >   if x < y then -1 else if x > y then 1 else 0
> >
> The assembler code says compare_float is directly compiled to a function
> that compares the 2 values
> in xmm0 and xmm1 registers, while Pervasives.compare is a library function
> written in C doing the same thing.
> 
> The assembler code looks very very good.
> 
> But I doubt that you could measure a difference. The type system will always
> yield the float_compare function, doesn't it? So far my quickcheck...

There are many cases where you have a polymorphic function (as in the
code does not allow to infere that the type is float) doing lots of
compares or you pass compare as argument to a polymorphic function. By
specifically forcing the type to be float you can gain a lot of
performance.

If the type system already inferes the float type itself then you gain
nothing. But that doesn't always happen (at the right time).

MfG
	Goswin