High Performance Rearrangement and Multiplication Routines for Sparse Tensor Arithmetic

Researchers from diverse disciplines are increasingly incorporating numeric highorder data, i.e., numeric tensors, within their practice. Just like the matrix-vector (MV) paradigm, the development of multi-purpose, but high-performance, sparse data structures and algorithms for arithmetic calculations, e.g., those found in Einstein-like notation, is crucial for the continued adoption of tensors. We use the motivating example of high-order differential operators to illustrate this need. As sparse tensor arithmetic represents an emerging research topic, with challenges distinct from the MV paradigm, many aspects require further articulation and development. This work focuses on three core facets. First, aligning with prominent voices in the field, we emphasise the importance of data structures able to accommodate the operational complexity of tensor arithmetic. However, we describe a linearised coordinate (LCO) data structure that provides faster and more memory-efficient sorting performance in support of this operational complexity. Second, flexible data structures, like the LCO, rely heavily on sorts and permutations. We introduce an innovative permutation algorithm, based on radix sort, that is tailored to rearrange already-sorted sparse data, producing significant performance gains. Third, we introduce a novel poly-algorithm for sparse tensor products, where hyper-sparsity is a possibility. Different manifestations of hyper-sparsity demand their own customised approach, which our multiplication poly-algorithm is the first to provide. These developments are incorporated within our LibNT and NTToolbox software libraries. Benchmarks, frequently drawn from the high-order differential operators example, demonstrate the practical impact of our routines, with speed-ups of 40% or higher compared to alternative high-performance implementations. Comparisons against the MATLAB Tensor Toolbox show over 10 times speed improvements. Thus, these advancements produce significant practical improvements for sparse tensor arithmetic.