Is there any clever way to embed a dynamic, large, relatively sparse, bipartite graph in a CUDA-friendly way?
The queries will generally be to return the list of edges incident to a given vertex.
The vertex sets are fixed in advance, the modifications will be to insert or delete all the edges incident to a vertex given in the first part.
The part of the problem that makes CUDA attractive is that given a vertex in the first part, computation of the edges incident to it (in order to add them) is well suited to CUDA, although by a method that does not “live” in the graph representation.
There are lots of abstract data structures which can handle this, I’m just wondering if anyone has something like this - keeping the whole algorithm on the GPU might be advantageous if the representation is not too bad.