I'd be grateful if anyone could point me to the actual code for this. I'd like to give it a try for myself, but all I seem to be able to dig up is research papers.
The Haskell community is fond of mistaking research papers (in PDF!) for documentation, it's quite likely you've found all there is. Just as they are also fond of mistaking a list of function names and type signatures for documentation. (On good days, they may deign to add a sentence about what the function does.)
(Haskell guys, I love ya, but you have some of the oddest blind spots sometimes.)
But, seriously, they've been talking about this GPU work for upwards of a 1.5 years now. Usually we get something to play with, even if it's pretty badly broken (I'm thinking NDP here), so the lack of code here is pretty atypical. The accelerate link above is news to me, though, so that's helpful. Hope to see more soon!
I think it's because of overexposure to category theory. It gets to the point where they see a name, the see a type signature, they assume the developer isn't pathological, and figure out what they need.
(Haskell guys, I love ya, but you have some of the oddest blind spots sometimes.)