Extending Architecture Modeling for Signal Processing towards GPUs

Efficient usage of heterogeneous computing architectures requires distribution of the workload to available processing elements. Traditionally, this mapping is done based on information acquired from application profiling. To reduce the high amount of manual work related to mapping, statistical application and architecture modeling can be applied for automating mapping exploration. Application modeling has been studied extensively, whereas architecture modeling has received less attention. Originally developed for signal processing systems, Linear System Level Architecture (LSLA) is the first architecture modeling approach that clearly distinguishes the underlying computation hardware from software. Up to now, LSLA has covered the modeling of multicore CPUs. This work proposes extending the LSLA model with GPU support, by including the notion of parallelism. The proposed GPU modeling extension is evaluated by performance estimation of three signal processing applications with various workload distributions on a desktop GPU, and a mobile GPU. The measured average fidelity of the proposed model is 93%.