generalized fold (and other newbie question)

[Manuel M T Chakravarty]

Johannes Waldmann:
> Thanks for the answers.
> 
>> The problem with that algorithm is that, 
>> without delayed arrays (as in Repa), it's space requirements are n^3 for nxn matrices.
> 
> yes, I was assuming that the 3-dim arrays are "virtual"
> (you don't store the array elements but just the index computation).

This is what happens in Repa, but not in Accelerate (which is implemented in a entirely different manner).  

Manuel