Re: [Cython] access to numpy functions?

On May 1, 2008, at 7:39 AM, Dag Sverre Seljebotn wrote:

>
>> API?  If I have two objects in Cython/Pyrex, like:
>>
>>
>> c_numpy.ndarray x,y
>>
>> and I want to do a dot product, or add the two arrays, is there
>> access to the numpy C-version of this?  I know I could write my own
>> for the dot product, but I'd rather not reinvent the wheel when
>> possible.
>>
>
> First off, you can use the Python functions directly, like this  
> (Cython
> syntax):
>
> ---
> import numpy
> cimport c_numpy
>
> cdef c_numpy.ndarray a = ..., b = ...
> cdef c_numpy.ndarray c = numpy.dot(a, b)
> ---
>
> As for calling the NumPy C functions directly: The a place to start is
> looking seems to be core/src/multiarraymodule.c. (You may want to  
> ask in
> the numpy mailing list as well though. I'm not very familiar with  
> NumPy
> internals.)
>
> My first impression is that the implementations (say,
> PyArray_InnerProduct for instance) seem to be a) very generic (any
> number of dimensions/strides etc. and b) coded using the Python/C API
> directly, so there is no "inner layer" to communicate with. I.e.,
> arguments to functions are Python objects.
>
> So you're probably just as well off by calling the functions  
> through the
> Python API like stated above.
>
> If that is not fast enough, there doesn't seem to be much to be gained
> by "direct calls" to the C functions; and a new implementation in  
> Cython
> seems in order.

And fortunately this is another GSoC project.

> The main reason is that to make this faster you will
> want to make use of additional information -- if you know the  
> number of
> dimensions (or even their size) and datatype at compile-time, you can
> make some optimizations that functions in the NumPy library will never
> be able to do. (For most if not all cases I'm guessing that we're
> talking very small gains if any gains here; though if you tend to
> multiply really small arrays a lot (like doing coordinate
> transformation...) then I suppose you must write your own code.)
>
> -- 
> Dag Sverre
>
> _______________________________________________
> Cython-dev mailing list
> Cython-dev-F/[email protected]
> http://codespeak.net/mailman/listinfo/cython-dev