Improving CADNA Performance on GPUs
Abstract
The quantification of rounding errors is crucial for numerical simulations on massively parallel architectures such as GPUs. The CADNA library enables one to estimate rounding errors in simulation programs. A version of CADNA for GPUs had been proposed to show the feasibility of numerical validation on such architectures. In this paper we show how the performance of CADNA on GPUs has been improved. Thanks to various optimizations that have been validated on several benchmarks, the performance gain is up to 61% with respect to the original prototype. Furthermore the GPU version of CADNA has been completed with features such as the accuracy estimation for double precision computation.