I think you want the same datastructures moving up and down all layers of stack, otherwise you pay a marshalling penalty at each layer which is terrible for per

I think you want the same datastructures moving up and down all layers of stack, otherwise you pay a marshalling penalty at each layer which is terrible for perf. If this is true, then we'd need to see fast immutable memory models inside GPUs that could natively represent persistent trees (moving away from pointers to mutable array buffers). Otherwise, the imperative way will always be at least C times faster, which seems a dealbreaker for the trailblazing researchers who forge the path the application layer follows.