[Numpy-discussion] Need help for implementing a fast clip in numpy (was slow clip)

[David Cournapeau]

Travis Oliphant wrote:
>
> This is one thing I've exposed (and made use of in more than one place) 
> with NumPy.  In Numeric, the magic was in a few lines of the ufuncobject 
> file).  Now, it is exposed in the concept of an array iterator.  Anybody 
> can take advantage of it as it there is a C-API call to get an array 
> iterator from the array (it's actually the object returned by the .flat 
> method).   You can even get an iterator that iterates over one-less 
> dimension than the array has (with the dimension using the smallest 
> strides left "un-iterated" so that you can call an inner loop with it).
The thing which confuses me is whether this is useful when you only one 
item of one array at a time. When I was implementing some functions for 
LPC, I took a look at your examples for array iterators and explanations 
in the numpy ebook, and it looked really helpful, indeed. For this kind 
of code, I needed to operate on several contiguous elements at a time.

But here, for cliping with scalar min and max, I only need to access to 
one item at a time from the input array, and that's it; in particular, I 
don't care about the order of iteration. So the question really boils 
down to:

"for a numpy array a of eg float32, am I guaranteed that 
a->data[sizeof(float32) * i] for 0 <= i < a.size gives me all the items 
of a, even for non contiguous arrays ?"

cheers,

David