Towards a Reliable Performance Evaluation of Accurate Summation Algorithms

Philippe Langlois (1) , Bernard Goossens (1) , David Parello (1)

1 DALI - Digits, Architectures et Logiciels Informatiques

Recent floating point summation algorithms compute the best accurate sum even for arbitrary ill-conditioned problems. So the actual run-time efficiency of these algorithms becomes the criteria to decide which is the best. Neither the classic flop count nor experimental timings are relevant measures of the actual performance of such core numerical algorithms. We justify these claims and present a reliable performance evaluation of some accurate summation algorithms thanks to an automatic instruction level parallelism analysis.

Domaines: Logiciel mathématique [cs.MS]; Arithmétique des ordinateurs

Philippe Langlois, Bernard Goossens, David Parello. Towards a Reliable Performance Evaluation of Accurate Summation Algorithms. SCE: Computational Science and Engineering, SIAM, Feb 2013, Boston, Ma, United States. ⟨hal-01233879⟩