[Numpy-discussion] Cython numerical syntax revisited

[Francesc Alted]

A Thursday 05 March 2009, Dag Sverre Seljebotn escrigué:
> Francesc Alted wrote:
> > A Thursday 05 March 2009, Dag Sverre Seljebotn escrigué:
> >> No, one could do the same thing that NumPy does (I think, never
> >> looked into it in detail), i.e:
> >>
> >> decide on dimension to do innermost dynamically from strides and
> >> sizes save the stride in that dimension for each array
> >> for loop using n-dimensional iterator with larger per-loop
> >> overhead: save offsets
> >>    for loop on the innermost dimension with lower per-loop
> >> overhead: component-wise operation using offsets and innermost
> >> strides
> >
> > I see.  Yes, it seems definitely doable.  However, I don't
> > understand very well when you say that you have to "decide on
> > dimension to do innermost dynamically".  For me, this dimension
> > should always be the trailing dimension, in order to maximize the
> > locality of data.  Or I'm missing something?
>
> For a transposed array (or Fortran-ordered one) it will be the
> leading.

Good point.  I was not aware of this subtlity.  In fact, numexpr does 
not get well with transposed views of NumPy arrays.  Filed the bug in:

http://code.google.com/p/numexpr/issues/detail?id=18

> Not sure whether it is possible with other kinds of views 
> (where e.g. a middle dimension varies fastest), but the NumPy model
> doesn't preclude it and I suppose it would be possible with
> stride_tricks.

Middle dimensions varying first?  Oh my! :)

-- 
Francesc Alted