I don't know of anyone who's done this and I'm not an expert personally in GPU-ifying that algorithm. It looks like GPU Gems 2 looked at doing FFT via a multipass shader. You could probably do the multipass aspect via a graph that's simply a linear chain via the toolkit. Any interpolated calculations happening in the vertex stage would need to be shoved into the fragment stage or out to the evalDependents/evalParameters stage.
Here's the article I was looking at to learn more: http://http.developer.nvidia.com/GPUGems2/gpugems2_chapter48.html
Sorry to dissapoint, but the toolkit doesn't support 32 bit image buffers in the GPU path and you'd want that for this application. We support them under the hood when doing CPU path computations, but for what you want to do, there wouldn't be any point since GPU speed is what you're hoping for.