[Numpy-discussion] neighborhood iterator speed

[David Cournapeau]

On Mon, Oct 24, 2011 at 6:57 AM, Nadav Horesh <nadavh at visionsense.com> wrote:
> I am trying to replace an old code (biliteral filter) that rely on ndimage.generic_filter with the neighborhood iterator. In the old code, the generic_filter generates a contiguous copy of the neighborhood, thus the (cython) code could use C loop to iterate over the neighbourhood copy. In the new code version the  PyArrayNeighborhoodIter_Next must be called to retrieve every neighbourhood item. The results of rough benchmarking to compare bilateral filtering on a 1000x1000 array:
> Old code (ndimage.generic_filter):  16.5 sec
> New code (neighborhood iteration):  60.5 sec
> New code with PyArrayNeighborhoodIter_Next  omitted: 1.5 sec
>
> * The last benchmark is not "real" since the omitted call is a must. It just demonstrates the iterator overhead.
> * I assune the main overhead in the old code is the python function callback process. There are instructions in the manual how to wrap a C code for a faster callback, but I rather use the neighbourhood iterator as I consider it as more generic.
>

I am afraid the cost is unavoidable: you are really trading cpu for
memory. When using PyArrayNeighborhood_Next, there is a loop with a
condiational within, and I don't think those can easily be avoided
without losing genericity. Which mode are you using when creating the
neighborhood iterator ?

There used to be a PyArrayNeightborhoodIter_Next2d, I don't know why I
commented out. You could try to see if you can get faster.

> If the PyArrayNeighborhoodIter_Reset could (optionally) copy the relevant data (as the generic_filter does) it would provide a major speed up in many cases.

Optionally copying may be an option, but it would make more sense to
do it at creation time than during reset, no ? Something like a binary
and with the current mode flag,

cheers,

David