Profiling Kernels Behavior to Improve CPU / GPU Interactions

Most modern computer and mobile devices have a graphical processing unit (GPU) available for any general purpose computation. GPU supports a programming model that is radically different from traditional sequential programming. As such, programming GPU is known to be hard and error prone, despite the large number of available APIs and dedicated programming languages. In this paper we describe a profiling technique that reports on the interaction between a CPU and GPUs. The resulting execution profile may then reveal anomalies and suboptimal situations, in particular due to an improper memory configuration. Our profiler has been effective at identifying suboptimal memory allocation usage in one image processing application.