High Performance Sparse Tensor Contraction

Tensors are widely encountered in several application domains such as scientific computing, machine learning and data analytics. Tensor contraction is a key algebraic operation in many applications involving multi-dimensional data. It is a higher dimensional analog of matrix-matrix multiplication. Sparse tensor contraction suffers from poor data-locality and irregular accesses, which poses a significant performance challenge. We explore the use of hashing-based methods to improve the efficiency of sparse tensor contraction operation on shared-memory systems.