DOI: 10.17587/prin.7.387-394

Development of High-Precision Arithmetic Package for Supercomputers with Graphics Processing Units

K. S. Isupov, ks_isupov@vyatsu.ru, V. S. Knyazkov, knyazkov@vyatsu.ru, A. S. Kuvaev, kyvaevy@gmail.com, M. V. Popov, mihailpopov1993@mail.ru, Vyatka State University, Kirov, 610000, Russian Federation

Corresponding author: Isupov Konstantin S., Leading Researcher, Vyatka State University, 610000, Kirov, Russian Federation, e-mail: ks_isupov@vyatsu.ru

A multiple-precision software library for parallel computing systems with hybrid CPU/GPU architecture is considered in this paper. Three numeric data types are supported: fixed-precision extended-range floats, multiple-precision integers, and multiple-precision floats. The feature of the library is the use of residue number system (RNS) to represent multiple-precision significands. This makes it possible to effective parallel processing of significand digits due to the lack of carry propagation. The interval estimation of relative (fractional) values in RNS is used to perform complex operations, such as magnitude comparison and overflow detection.

Keywords: multiple-precision computations, extended range, residue number system, CUDA, software

pp. 387–394

For citation:
Isupov K. S., Knyazkov V. S., Kuvaev A. S., Popov M. V. Development of High-Precision Arithmetic Package for Supercomputers with Graphics Processing Units, Programmnaya Ingeneria, 2016, vol. 7, no. 9, pp. 387—394.

This work was supported by the Russian Foundation for Basic Research, project no. 16-37-60003 mol_a_dk