An Effective Genetic Algorithm for Mixed Precision

The precision of floating-point numbers is a critical task in high-performance computing. Many scientific applications rely on floating-point arithmetic, but excessive precision can lead to unnecessary computational overhead. Reducing precision may introduce unacceptable errors. Addressing this trad...

Full description

Saved in:
Bibliographic Details
Main Authors: Wanyu Zhang, Yu Shang, Min Tsao, Yiwei Li, Xiaoyu Song
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10948503/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850150940654436352
author Wanyu Zhang
Yu Shang
Min Tsao
Yiwei Li
Xiaoyu Song
author_facet Wanyu Zhang
Yu Shang
Min Tsao
Yiwei Li
Xiaoyu Song
author_sort Wanyu Zhang
collection DOAJ
description The precision of floating-point numbers is a critical task in high-performance computing. Many scientific applications rely on floating-point arithmetic, but excessive precision can lead to unnecessary computational overhead. Reducing precision may introduce unacceptable errors. Addressing this trade-off is essential for optimizing performance while ensuring numerical accuracy. In this paper, we present a genetic algorithm-based approach for tuning the precision of floating-point computations. Our method leverages algorithmic differentiation and first-order Taylor series approximation to assess the impact of precision variations efficiently. We employ stochastic partitioning algorithms with multiple precision combinations that meet the error requirements. Moreover, we present a genetic heuristic algorithm to determine the maximum number of variables that can sustain precision alterations without compromising the desired error threshold. The proposed approach is evaluated across various benchmark programs, analyzing the effects of precision tuning under increasing error thresholds. Our findings reveal that, for a majority of these programs, reducing precision through partitioning leads to significant performance enhancements, with improvements of up to 15%.
format Article
id doaj-art-e3d82a5c44384f59bf6dac3b07ddd48a
institution OA Journals
issn 2169-3536
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-e3d82a5c44384f59bf6dac3b07ddd48a2025-08-20T02:26:24ZengIEEEIEEE Access2169-35362025-01-0113627396275210.1109/ACCESS.2025.355750510948503An Effective Genetic Algorithm for Mixed PrecisionWanyu Zhang0https://orcid.org/0009-0008-9926-9961Yu Shang1Min Tsao2https://orcid.org/0009-0004-2670-3895Yiwei Li3Xiaoyu Song4https://orcid.org/0000-0002-6583-9400Department of Electrical and Computer Engineering, Portland State University, Portland, OR, USADepartment of Electrical and Computer Engineering, Portland State University, Portland, OR, USASiemens EDA, Fremont, CA, USASiemens EDA, Fremont, CA, USADepartment of Electrical and Computer Engineering, Portland State University, Portland, OR, USAThe precision of floating-point numbers is a critical task in high-performance computing. Many scientific applications rely on floating-point arithmetic, but excessive precision can lead to unnecessary computational overhead. Reducing precision may introduce unacceptable errors. Addressing this trade-off is essential for optimizing performance while ensuring numerical accuracy. In this paper, we present a genetic algorithm-based approach for tuning the precision of floating-point computations. Our method leverages algorithmic differentiation and first-order Taylor series approximation to assess the impact of precision variations efficiently. We employ stochastic partitioning algorithms with multiple precision combinations that meet the error requirements. Moreover, we present a genetic heuristic algorithm to determine the maximum number of variables that can sustain precision alterations without compromising the desired error threshold. The proposed approach is evaluated across various benchmark programs, analyzing the effects of precision tuning under increasing error thresholds. Our findings reveal that, for a majority of these programs, reducing precision through partitioning leads to significant performance enhancements, with improvements of up to 15%.https://ieeexplore.ieee.org/document/10948503/Floating-point arithmeticalgorithmic differentiationTaylor seriesdynamic program analysismixed precisiongenetic algorithm
spellingShingle Wanyu Zhang
Yu Shang
Min Tsao
Yiwei Li
Xiaoyu Song
An Effective Genetic Algorithm for Mixed Precision
IEEE Access
Floating-point arithmetic
algorithmic differentiation
Taylor series
dynamic program analysis
mixed precision
genetic algorithm
title An Effective Genetic Algorithm for Mixed Precision
title_full An Effective Genetic Algorithm for Mixed Precision
title_fullStr An Effective Genetic Algorithm for Mixed Precision
title_full_unstemmed An Effective Genetic Algorithm for Mixed Precision
title_short An Effective Genetic Algorithm for Mixed Precision
title_sort effective genetic algorithm for mixed precision
topic Floating-point arithmetic
algorithmic differentiation
Taylor series
dynamic program analysis
mixed precision
genetic algorithm
url https://ieeexplore.ieee.org/document/10948503/
work_keys_str_mv AT wanyuzhang aneffectivegeneticalgorithmformixedprecision
AT yushang aneffectivegeneticalgorithmformixedprecision
AT mintsao aneffectivegeneticalgorithmformixedprecision
AT yiweili aneffectivegeneticalgorithmformixedprecision
AT xiaoyusong aneffectivegeneticalgorithmformixedprecision
AT wanyuzhang effectivegeneticalgorithmformixedprecision
AT yushang effectivegeneticalgorithmformixedprecision
AT mintsao effectivegeneticalgorithmformixedprecision
AT yiweili effectivegeneticalgorithmformixedprecision
AT xiaoyusong effectivegeneticalgorithmformixedprecision