.png){kind=icon}
.png){kind=icon}
Return
PESA: error sensitivity analysis tool for floating-point computational programs
.png){kind=icon}
PRE
AI
DOI：10.1007/s11227-025-06962-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Floating-point programs are currently extensively applied across various domains of scientific computing. Within such computational programs, there is a propensity to utilize high-precision floating-point variables to ensure the accuracy of the results. However, this practice concurrently leads to a decline in program performance. To address this issue, researchers have aimed to reduce the precision requirements of variables in programs without significant impact on results. Current mixed-precision tools utilize differential derivation to compute variable sensitivity, but this method is cumbersome and can cause memory crashes in intensive computational programs. This paper presents program error sensitivity analysis (PESA), an error sensitivity analysis tool based on variance decomposition, optimized for floating-point programs using LLVM. PESA uses the Sobol method to perform error sensitivity analysis and provides the user with the magnitude of error sensitivity for each variable in the program. Experimental results show that PESA produces error analysis results consistent with other tools and outperforms other sensitivity analysis tools overall by testing six benchmark files and small applications.
Keywords:
Global sensitivity analysis
Sobol
Error analysis
LLVM