Vlad Firoiu

@cpdurham: As I said, my real program is doing a complicated graph traversal involving IORefs at each node and thus is very difficult to express a pure function. Also, won't...

Well once I have a function in the form taken by run1 (Acc -> Acc) I can iterate it for a fixed number of times on the GPU with awhile....

To clarify, by Acc I mean a type that is stored on the GPU. `run :: Acc a -> a` removes the Acc wrapper and copies to the CPU. Maybe...

I realize now that the `compile :: (Acc a -> Acc b) -> Acc a -> Acc b` I gave is not possible, since Acc is backend-agnostic. Is there any...

After taking a look at some of the Accelerate code, I see that there are actually two different versions of Acc: a "HOAS" version defined in Data.Array.Accelerate.Smart, and a De...

Right now I have two semi-independent goals: 1. Generate ASTs where the matrix ops are explicitly let-shared. I'm not sure if this is even possible - maybe only non-functions are...

I've convinced myself that goal #1 is impossible, since the DeBruijn Alet constructor only takes acc values, and the HOAS functions (Data.Array.Accelerate.Trafo.Sharing) are constrained to be first-order. Goal #2 seems...

Bump? Does compiling kernels into Foreign functions sound like a workable solution?

Simplest solution for this might be a gecko code.

Was about to open an issue for this. The text currently references an "Option B", which doesn't seem to exist.