patch in Python?

[Fernando Perez]

On Wed, 2 Oct 2002, Gerhard Häring wrote:

> > - How does this compare to gmpy?
> 
> No idea. It doesn't have rationals, which GPM has (not sure about gmpy).
> Btw. I think mxNumber from MAL is the GMP wrapper that has the most
> momentum nowadays. His aim is a similar one: NUMERIC support for his
> mxODBC library.

Great. I'll have to look at mxNumber when I get around to working on this
problem. It would be great if the Numpy support was in already, but if
needed that's something I could work on.

> > - And http://www.nersc.gov/~dhbailey/mpdist/mpdist.html? This one is supposed 
> > to be first rate.
> 
> These are written in Fortran and C++, it seems. See above. In my tries
> to wrap MAPM I have found, however, that creating a new Python number
> type is relatively easy. So if these are worth wrapping, I think it
> should be a matter of days, rather than weeks.

This one is a strong candidate b/c the algorithms for extended (_not_
arbitrary) precision have been very carefully tuned.

> > I'd love to have the ability not only to define extended/arbitrary precision 
> > scalars, but especially to have them carry over to Numeric. One could then 
> > develop a difficult algorithm in arbitrary precision and then move it over to 
> > only extended (or even normal) precision once everything is working reliably.
> 
> Yeah, this sounds like a neat idea.
> 
> One "problem" I currently have is this:
> 
> Under MAPM, certain operations (division, exponentiation, ...) require
> an additional parameter for the precision. My current API is this:
> 
> module mapm:
> def set_default_div_precision(n)
> def set_default_pow_precision(n)
> 
> class MAPM:
>     ...
>     def __div__(self, other, precision=None)
>     def __pow__(self, other, precision=None)
> 
> But under normal operation, such as in:
> 
>     a = mapm.MAPM("27.34")
>     b = mapm.MAPM("3.17")
>     c = a / b
> 
> you can't give the precision operator to __div__, so you'd have to set
> the precision before with something like:
> 
>     mapm.set_default_div_precision(10)
> 
> alternatively, you can do without the "/" operator:
> 
>     c = a.__div__(b, 10)
> 
> but this feels clumsy.
> 
> To put it short: do you have any suggestion for a better API wrt.
> precision?

Well, you could extend your MAPM class so that _each_ number in it carries
a precision flag with it. Operations would then be done with the precision
of the lower of the two operands. Pseudocode:

class MAPM:
	def __init__(self,data,precision=16):
	...

Then, 

     a = mapm.MAPM("27.34",32)
     b = mapm.MAPM("3.17",48)

     c = a/b  # c would be a mapm number with 32 digits.

You could have an overall flag for warnings, so that anytime that there is
a loss of precision (like in the above case) a warning is printed.

Cheers,

f.