Bas van Dijk
v.dijk.bas at gmail.com
Thu Nov 10 00:55:02 CET 2011
On 9 November 2011 22:33, kaffeepause73 <kaffeepause73 at yahoo.de> wrote:
> Repa is indeed very Interesting, but I have changing vector length in the
> second dimension and later on only want to generate Data on demand. If I use
> Matrices, I will use loads of space for no reason.
Even if it is possible to create an unboxed vector of unboxed vectors,
if the inner unboxed vectors have variable lengths as you require,
indexing will become O(n) instead of O(1) because you need to traverse
the inner unboxed vectors and check their length to find the desired
index. I'm not sure that's what you want.
> Seems like sticking to Boxed Vector for now is best Choice for me.
Yes your second alternative: a boxed vector of unboxed vectors seems
to do what you want.
> isn't data.vector also providing multidimensional arrays?
I don't think so. All indexing functions get a single Int argument. Of
course it's easy to build a layer on top that adds more dimensions.
> So is Repa just another Version of Data.Vector or is it building another level on top.
The latter, repa provides a layer on top of vector.
Note that you can also convert Vectors to repa Arrays using:
fromVector :: Shape sh => sh -> Vector a -> Array sh a
I believe its O(1).
> And when to use best which of the two ?
I guess when your vectors are multi-dimensional and you want to
benefit from parallelism you should use repa instead of vector.
More information about the Haskell-Cafe