New module (written in C) for using the high-precision QD library

Stefan Behnel stefan_ml at behnel.de
Fri Jul 31 10:40:14 CEST 2015


Chris Angelico schrieb am 31.07.2015 um 09:37:
> On Fri, Jul 31, 2015 at 5:26 PM, Stefan Behnel wrote:
>> Your C code seems to be only about 1500 lines, not too late to translate
>> it. That should save you a couple of hundred lines and at the same time
>> make it work with Python 3 (which it currently doesn't, from what I see).
> 
> To what extent does Cython make this easier? The biggest barrier I
> would expect to see is the bytes/text distinction

Yes, that tends to be a barrier. Cython is mostly just Python, so you can write

    if isinstance(s, unicode):
        s = (<unicode> s).encode('utf8')

and be happy with it ("<type>" is a cast in Cython). Such simple code looks
uglier when spelled out using the C-API and wouldn't be any CPU cycle faster.

But there's also the PyInt/PyLong unification, which can easily get in the
way for a number processing library. In Cython, you can write

    if isinstance(x, (int, long)):
        try:
            c_long = <long> x
        except OverflowError:
            ...  # do slow conversion of large integer here
        else:
            ...  # do fast conversion from c_long here

or something like that and it'll work in Py2.6 through Py3.5 because Cython
does the necessary adaptations internally for you. This code snippet
already has a substantially faster fast-path than what the OP's code does
and it will still be much easier to tune later, in case you notice that the
slow path is too slow after all.

And then there are various helpful little features in the language like,
say, C arrays assigning by value, or freelists for extension types using a
decorator. The OP's code would clearly benefit from those, if only for
readability.

Python is much easier to write and maintain than C. Cython inherits that
property and expands it across C data types. And it generates C code for
you that automatically adapts to the different Python versions in various
ways, both in terms of compatibility and performance.

Stefan




More information about the Python-list mailing list