Execution time analysis of audio algorithms

Abstract

Execution time analysis forms an important part of design space and hardware architectural exploration for hard real-time systems. Some approaches for execution time analysis require prerequisite knowledge of synchronous data flows or timed automata employed in model checkers. This is unsuitable for industry-level use as easy approaches without any prerequisite knowledge requirements are preferred by them. Most of the existing proposals addressing execution time analysis target 'fast' approaches rather than timing accuracy. Also, these existing proposals are validated with simple processors as majority of processors used for hard real-time embedded software are not so complex. However, for audio applications employing audio algorithms, modern processors with high performance are required, to be scalable with increasing audio algorithm complexity. Execution time analysis of audio algorithms on these modern processors is gaining importance as sampling frequencies are getting higher and deadlines getting shorter. In this thesis, we propose an industry acceptable model/simulation framework to perform execution time analysis of audio algorithms on modern mono core and multi-core processors (operating in asymmetric multiprocessing mode). Our framework combines open-source tools such as Gcov, POOSL modelling language (Parallel object-oriented specification language) and Gem5 which is a computer architecture simulator to determine WCET of an audio algorithm using dynamic means as static WCET techniques lead to huge overestimations. This solution framework was proposed after conducting experiments targeting execution time analysis at different abstraction levels and evaluating results based on accuracy, flexibility, hardware compatibility and scalability. Our proposed technique is flexible as Gem5 models several processor architectures and expressing audio algorithms at the basic block abstraction level allows parameters such as loop bounds to be changed easily. It can be easily extended to more cores using the expressive POOSL modelling language. Our technique requires no prerequisite knowledge of any concepts and also solves issues such as infuence of execution contexts of basic blocks on execution time in an implicit way without requiring additional analysis. We were able to achieve almost 99% timing accuracy for integer and floating point programs without long latency instructions using Gem5. For programs with long latency instructions, minor manipulations are presented which yielded almost 90% accuracy using Gem5.