GSTACO: A Generalized Sparse Tensor Algebra Compiler

Abstract

Many applications in engineering and computer science are characterized by sparse multi-dimensional data. Therefore, optimizations for sparse tensor algebra have received a lot of attention lately. Several hardware and software solutions have emerged in order to speed up the computation of certain tensor expressions, but none of them provides an interface that is general and comprehensive enough to meet the requirements of complex applications like graph analysis. In this work we attempt to identify where previous solutions fell short and build the Generalized Sparse Tensor Algebra Compiler (GSTACO), a new compiler aiming to fill in the engineering gaps of efficient sparse computation.