Improving accuracy of source level timing simulation for GPUs using a probabilistic resource model

After their success in the high performance and desktop market, Graphic Processing Units (GPUs), that can be used for general purpose computing are introduced for embedded systems on a chip (SOCs). Due to some advanced architectural features, like massive simultaneous multithreading, static performance analysis and high-level timing simulation are difficult to apply to code running on these systems. This paper extends a method for performance simulation of GPUs. The method uses automated performance annotations in the application's OpenCL C source code, and an extended performance model for derivation of a kernels runtime from metrics produced by the execution of annotated kernels. The final results are then generated using a probabilistic resource conflict model. The model reaches an accuracy of 90% on most test cases and delivers a higher average accuracy than previous methods.