Indexed Binary Operations in the GraphBLAS

GraphBLAS is a sparse matrix library for graph algorithms. We could improve the GraphBLAS and support a wider range of applications if we had an indexed binary operator; i.e., a binary operator that depends on elements from two matrices and their positions (indices). We propose a design for these operators and illustrate their impact on performance and expressiveness for several algorithms including breadth-firstsearch, argmin/argmax, vector search, and finite-element assembly.