[Numpy-discussion] My GSoC Proposal to Implement a Subset of NumPy for PyPy

Pauli Virtanen pav@iki...
Sun Apr 18 05:46:36 CDT 2010


su, 2010-04-18 kello 19:05 +0900, David Cournapeau kirjoitti:
[clip]
> First, an aside: with the recent announcement from pypy concerning the
> new way of interfacing with C, wouldn't it make more sense to go the
> other way around  - i.e. starting from full numpy, and replace some
> parts in rpython ? Of course, this assumes that numpy can work with
> pypy using the new C API "emulation", but this may not be a lot of
> work. This would gives the advantage of having something useful from
> the start.
> 
> I think this project is extremely interesting in nature, and the parts
> which are the most interesting to implement are (IMHO of course):
>   - indexing, fancy indexing
>   - broadcasting
>   - basic ufunc support, although I am not sure how to plug this with
> the C math library for math functions (cos, log, etc...)

I agree here: if it's possible to supplement the existing Numpy code
base by interpreter-level (JIT-able) support for indexing, fancy
indexing, and broadcasting, this would reduce the roadblock in writing
effective algorithms without using vectorization.

Also, if you can incrementally replace parts of Numpy, the code could be
immediately usable in real world, being less of a proof-of-concept
subset.

-- 
Pauli Virtanen




More information about the NumPy-Discussion mailing list