Sparse tensor transpositions in the tensor algebra compiler

Abstract

The current state of the art for transposing sparse tensors involves converting the sparse tensor into a list of coordinates, sorting the list of coordinates and finally packing the list of coordinates into the desired sparse tensor format. This thesis explores the potential for faster methodologies. Its main contributions are an algorithm that exploits partial sortedness to minimize sorting passes and an implementation that demonstrates that this transposition algorithm is competitive with state of the art. In particular the algorithm takes advantage of the ordering that already exists to apply selective sorting passes and thereby reduce the amount of work that needs to be done to reorder the tensor.