Abstract

GPU and multicore architectures are often used to accelerate scientific applications that rely on floating point computation. Current techniques for verifying the correctness of these implementations are quite primitive. Scientists expect their floating point code to produce different results on different architectures, and as a result do not test their code thoroughly.

The talk will consider two case studies: a parallel implementation of Pi and a real world medical application used in cancer diagnosis. Through these case studies, the sources of differences in CPU and GPU codes will be identified. Issues of correctness, including control issues such as thread synchronization as well as issues of floating point correctness will be discussed. I will give a brief survey of tools available for addressing these issues and discuss current research directions in identifying bugs in GPU and multicore implementations.