Trade-off Between Accuracy and Time for Automatically Generated Summation Algorithms
Résumé
We focus on numerical algorithms for which performances and accuracy do not cohabit well in practice. A widely studied example is the floating-point summation. In order to increase parallelism, expressions are reparsed implicitly using arithmetic properties like associativity or distributivity. In IEEE-754 floating-point arithmetic, these laws do not hold any longer for any arithmetic expression. The numerical accuracy of the algorithm may be strongly sensitive to reparsing. So increasing the parallelism of some algorithm may decrease its numerical accuracy, and, conversely, improving the accuracy of some computation may reduce its parallelism. We present an experimental analysis of this problem for the oating-point summation.