[pypy-dev] Questions on the pypy+numpy project

[Peter Cock]

On Wed, Oct 19, 2011 at 12:57 PM, Antonio Cuni <anto.cuni at gmail.com> wrote:
> On 19/10/11 13:42, Antonio Cuni wrote:
>
>> I'm not sure to interpret your sentence correctly.
>> Are you saying that you would still want a pypy+numpy+scipy,
>> even if it ran things slower than CPython? May I ask why?
>
> ah sorry, I think I misunderstood your email.
>
> You would like pypy+numpy+scipy so that you could write fast
> python-only algorithms and still use the existing libraries.  I
> suppose this is a perfectly reasonable usecase, and indeed
> the current plan does not focus on this.

I want this too - well actually pypy+numpy+xxx where xxx uses
bits of the numpy C API. I don't care if the numpy bits are a *bit*
slower under PyPy than C Python - 100% compatibility is more
important to me.

> However, I'd like to underline that to write "fast python-only
> algorithms", you most probably still need a fast numpy in the
> way it is written right now (unless you want to write your
> algorithms without using numpy at all).  If we went to the
> slow-but-scipy-compatible approach, any pure python
> algorithm which interfaces with numpy arrays would be
> terribly slow.

I'd be happy with "close to numpy under C Python" speeds
for my code using numpy under PyPy, with fast python-only
bits. That covers quite a lot of use cases I would think, but
if we'd get "terribly slow" for the numpy using bits that is
less tempting. Depending on your value of terrible ;)

Right now the PyPy micronumpy is far too limited to be of
real use even where I'm using only the Python interface.
e.g. there is no numpy.linalg module:
https://bugs.pypy.org/issue915

Peter