Easing development of precision-sensitive applications with a beyond-quad-precision library
Résumé
The IEEE754 Standard offers essentially two binary floating-point formats, binary32 and binary64, natively supported by current hardware. Whenever these precisions do not suffice, developers are restricted to arbitrary precision libraries, such as MPFR. These libraries however leave a gap in the mid-precision range (64 to 512 bits). Their very nature –as object code to be linked in– prevents modern compilers from inlining the code or optimizing it e.g. with loop unrolling. We propose the libwidefloat software meant to fill this gap. It offers precisions from 64 through 512 bits. It supports all basic operations (+, −, ×, /, √ , FMA, comparisons etc.) It is fully implemented in header files for automatic optimization.
Domaines
Arithmétique des ordinateursOrigine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...