You make a good point and the decision was by no means cut and dry. However I made a point of developing some test code using some the newer array data types and looked at maintaining the array in Haskell and then directly calling Blas etc. I even had a nice polymorphic matrix class going. However I found the array interface just a bit too 'clunky' to use a technical term. The withArray interface is not very appealing. The layers of lambda notation was giving me a headache.
The idea of separating the imperative code into an imperative language was appealing to me. Moreover using a stack based architecture for matrix operations makes the C end of things very easy to implement. The big bugaboo of memory management issues pretty much disappears, and thinking of the world state in the IO monad as a stack of matrices has a nice intuitive appeal for me. It seems to work well so far as I said earlier. I'm not sure all my issues would have gone away if I had tried to do more of the matrix op.s in Haskell. It is pretty much a given that I need to interface to external optimized libraries, that's where the big number crunching is occuring and those libraries have had teams of Ph.D.s working on them for years. I want to leverage that. My approach actually minimizes the amount of marshalling I have to do between C and Haskell. The stack manipulations are simply scripted in a do clause, with nary an argument being passed.
Has any one created binding for the numarray library ala Python Numarray? I've heard it's a very good and fast array interface. Perhaps it would be useful for this sort of task. Jason