Work

Multithread Approximation: An OpenMP Constructor

Year: 2026

Type: article

Abstract: ABSTRACT This study introduces an OpenMP construct designed to simplify and unify the integration of approximate computing techniques into shared‐memory parallel programs. Approximate Computing leverages the inherent error tolerance of many applications to trade computational accuracy for gains in performance and energy efficiency. Unlike prior OpenMP‐based approximation proposals, which typically focus on a single technique, our constructor provides a unified and portable interface that integrates multiple software approximation strategies. The proposed construct provides native support for software Approximate Computing strategies, including task dropping , loop perforation , floating‐point relaxation , and memoization , enabling developers to selectively apply these techniques with minimal code disruption. The experimental evaluation across a set of representative benchmark applications reveals performance improvements of up to 1272.90%, with a degradation of 83.3% accuracy and energy savings reaching up to 99.34% in specific configurations. These results underscore the viability of Approximate Computing alongside parallel programming models and open the door for further exploration of compiler and runtime support for controlled inaccuracy in high‐performance computing.

Source: Concurrency and Computation Practice and Experience

DOI: 10.1002/cpe.70584

Authors: João Briganti de Oliveira, Rogério Aparecido Gonçalves, João Fabrício Filho

Institution: Universidade Tecnológica Federal do Paraná

Language: English

Topic: Parallel Computing and Optimization Techniques

Subfield: Hardware and Architecture

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $5,140

Funders: Coordenação de Aperfeiçoamento de Pessoal de Nível Superior, Universidade Tecnológica Federal do Paraná