Year: 2025
Type: conference-paper
Abstract: Mixed-precision computing improves performance by reducing variable precision where high accuracy is unnecessary, but manually tuning large programs is infeasible. We present Perci-AD, an automated tool that applies algorithmic differentiation to perform static error analysis and identify precision-sensitive variables directly at the source code level. Perci-AD generates an initial mixed-precision configuration and transforms the source code while preserving its structure and readability. It also serves as a preprocessing step for dynamic tuning by pruning the search space, avoiding unnecessary high-precision evaluations. Evaluated on five floating-point programs, Perci-AD achieves up to \(40 \times\) speedup over long double precision baselines in Simpsons and accelerates dynamic tuning by over 1.2× in all benchmarks. Our results demonstrate that combining algorithmic differentiation with static analysis enables effective and extensible mixed-precision optimization.
Authors: Yifan Zhou, Lei Li, Hua Liang, Chao Zheng, Kuan Li, Hao Jiang
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4413987067