SignalPU: A Programming Model for DSP Applications on Parallel and Heterogeneous Clusters

Abstract

The biomedical imagery, the numeric communications, the acoustic signal processing and many others digital signal processing (DSP) applications are present more and more in the numeric world. They process growing data volume which is represented with more and more accuracy, and use complex algorithms with time constraints to satisfying. Consequently, a high requirement of computing power characterize them.To satisfy this need, it's inevitable today to use parallel and heterogeneous architectures in order to speedup the processing, where the best examples are today's supercomputers like "Tianhe-2" and "Titan" of Top500 ranking. These architectures with their multi-core nodes supported by many-core accelerators offer a good response to this problem. However, they are still hard to program to make performance because of many reasons: Parallelism expression, task synchronization, memory management, hardware specifications handling, load balancing . . . In the present work, we are characterizing DSP applications and propose a programming model based on their distinctiveness in order to implement them easily and efficiently on heterogeneous clusters.

A domain‐specific high‐level programming model

“…We also experiment with our approach on a real‐world application based on the primate retina, with the visual saliency model used to locate the regions of interest, that is, the capability of human vision to focus on particular places in a visual scene. For implementation, we use Algorithm 2 that was described in the preliminary work of our team . To implement the application with our programming models, the first step is to model Algorithm 2 with a DFG‐XML description using the SignalPU interface.…”

“…We also experiment with our approach on a real-world application based on the primate retina, with the visual saliency model used to locate the regions of interest, that is, the capability of human vision to focus on particular places in a visual scene. For implementation, we use Algorithm 2 that was described in the preliminary work of our team [39,40]. Node CPU GPU Archi1 Intel-i7 core (4 cores) 2 NVIDIA Quadro4000 + NVIDIA Quadro2000 Archi2 Intel-i7 core (4 cores) NVIDIA GTX TITAN + NVIDIA GTX780 Archi3 Intel-i7 core (4 cores) NVIDIA GTX680 + NVIDIA GTX680 Archi4 Intel-i7 core (4 cores) NVIDIA GTX480 Figure 8.…”