Poster De Conférence Année : 2025

PROMISE: Floating-point Autotuning with Customized Precisions

Résumé

Achieving optimal performance in numerical computations often hinges on aggressively quantizing arithmetic and data to low-precision formats or rigorous rounding error analysis to retain numerical accuracy. The PROMISE (PRecision OptiMISE) software provides a task-specific validation using CADNA (https://cadna.lip6.fr) for automated precision tuning, enabling a balance between computational efficiency and numerical fidelity. Here, we advance PROMISE with customized precision formats to examine the low-precision potential in a suite of numerical algorithms for performance gains, offering actionable insights into selecting optimal precision levels for diverse algorithms.

Fichier principal
Vignette du fichier
Poster_PROMISE.pdf (3.99 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05291790 , version 1 (01-10-2025)

Licence

Identifiants

  • HAL Id : hal-05291790 , version 1

Citer

Xinye Chen, Thibault Hilaire, Fabienne Jézéquel. PROMISE: Floating-point Autotuning with Customized Precisions. Workshop on Approximate Computing in Numerical Linear Algebra, Oct 2025, Paris, France. ⟨hal-05291790⟩
1335 Consultations
190 Téléchargements

Partager

  • More