Most floating point computations, in particular the additive operations, can have an annoying side-effect: The results can be inaccurate (the relative error can be any large) which may lead to qualitatively wrong answers to a computational problem. For the most prevalent applications there are advanced methods that avoid unreliable computational results in most of the cases. However, these methods usually decrease the performance of computations. In this paper, the most frequently occurring vector-vector addition and dot-product operations are investigated. We propose a faster stable add method which has been made possible by the appropriate utilization of the advanced SIMD architecture. In this work we focus only on the widely used Intel systems that are available for most of users. Based on it, stable vector addition and dot-product implementations are also introduced. Finally, we show the performance of these techniques.
分享
引用
