[SciPy-dev] NumPy f2py GSOC project

Dag Sverre Seljebotn dagss@student.matnat.uio...
Fri Mar 27 04:02:06 CDT 2009


Dag Sverre Seljebotn wrote:
> David Cournapeau wrote:
>> If you are interested in working on scipy-related projects, something
>> which could help the transition is to replace C code by cython. This
>> would be particularly helpful in scipy - there are many modules which
>> are crufty and would benefit from code cleaning, but if you don't know
>> cython, maybe that's not something that interesting for you. Cython is
>> not difficult to use, though. scipy.signal comes to mind, to take
>> something where I could definitely help you - but there are other
>> modules as well, where other people would be more qualified
>> (scipy.ndimage, etc...). One module means that it would be easier for
>> you to work on, easier for use to help/coordinate, and at least in the
>> case of scipy.signal, there is definitely some new features which are
>> needed as well (new feature is generally more fun than code cleanup/bug
>> hunting). Of course, you should choose  modules which interest you.
> 
> If something like this is done, one idea I have for SciPy is quick 
> Cython callbacks for solvers/integrators. Perhaps Lisandro would be 
> interested in being Cython-side mentor? Though the main mentor would 
> need to be from SciPy.
> 
> People often write on the Cython list about code like this:
> 
> def f(double x): return x*x
> 
> some_scipy_function(f, 0, 10)
> 
> The problem here is that a Python overhead is encurred on all function 
> evaluations. What one could do is add a "Cython protocol" like this:
> 
> cdef class DoubleFunction:
>      cpdef double eval(double x): return 0
> 
> Then inside SciPy core (psuedo-code):
> 
> PyObject* some_scipy_function(...) {
>      PyObject* callback = parse argument...
>      ...
>      if (callback is instance of DoubleFunction) {
>          call using quick Cython dispatch
>      } else {
>          call using Python API
>      }
> }
> 
> Then in user code:
> 
> cdef class MyFunc(DoubleFunction):
>      cpdef double eval(double x): return x*x
> 
> some_scipy_function(MyFunc(), 0, 10)
> 
> Voila!, 50x speedup of solvers from Cython.
> 

That is, for algorithms where one cannot use vectorized functions. 
Integration would not benefit, but Newton's method would.

-- 
Dag Sverre


More information about the Scipy-dev mailing list