Profilování paralelních aplikací

Abstract

This thesis deals with visualising memory accesses of CUDA programs. It presents a profiling tool that instruments CUDA programs and generates their memory access traces. These can then be visualised to exploit memory access optimization opportunities. First part of the thesis introduces application profiling methods and technologies used to create the profiling tool. The second part describes existing tools that allow profiling graphics processing units. The last part discusses design and implementation of the profiling tool and introduces a web application that visualises the stored memory traces.