Multiplication of sparse Laurent polynomials and Poisson series on modern hardware architectures

TL;DR

This paper introduces two hardware-optimized algorithms for multiplying sparse Laurent polynomials and Poisson series, leveraging Kronecker substitution and cache-efficient data structures, with demonstrated performance improvements on modern multicore architectures.

Contribution

The paper presents novel algorithms tailored for modern hardware architectures that efficiently multiply sparse Laurent polynomials and Poisson series, with parallel implementations and benchmark comparisons.

Findings

Algorithms outperform existing computer algebra systems in benchmarks.

Parallelization enhances multiplication speed on multicore architectures.

Cache-optimized techniques improve memory efficiency during multiplication.

Abstract

In this paper we present two algorithms for the multiplication of sparse Laurent polynomials and Poisson series (the latter being algebraic structures commonly arising in Celestial Mechanics from the application of perturbation theories). Both algorithms first employ the Kronecker substitution technique to reduce multivariate multiplication to univariate multiplication, and then use the schoolbook method to perform the univariate multiplication. The first algorithm, suitable for moderately-sparse multiplication, uses the exponents of the monomials resulting from the univariate multiplication as trivial hash values in a one dimensional lookup array of coefficients. The second algorithm, suitable for highly-sparse multiplication, uses a cache-optimised hash table which stores the coefficient-exponent pairs resulting from the multiplication using the exponents as keys. Both algorithms have been implemented with attention to modern computer hardware architectures. Particular care has been devoted to the efficient exploitation of contemporary memory hierarchies through cache-blocking techniques and cache-friendly term ordering. The first algorithm has been parallelised for shared-memory multicore architectures, whereas the second algorithm is in the process of being parallelised. We present benchmarks comparing our algorithms to the routines of other computer algebra systems, both in sequential and parallel mode.