Executing Dynamic Data Rate Actor Networks on OpenCL Platforms

摘要

Heterogeneous computing platforms consisting of general purpose processors(GPPs) and graphics processing units (GPUs) have become commonplace in personalmobile devices and embedded systems. For years, programming of these platformswas very tedious and simultaneous use of all available GPP and GPU resourcesrequired low-level programming to ensure efficient synchronization and datatransfer between processors. However, in the last few years several high-levelprogramming frameworks have emerged, which enable programmers to describeapplications by means of abstractions such as dataflow or Kahn process networksand leave parallel execution, data transfer and synchronization to be handledby the framework. Unfortunately, even the most advanced high-level programming frameworks havehad shortcomings that limit their applicability to certain classes ofapplications. This paper presents a new, dataflow-flavored programmingframework targeting heterogeneous platforms, and differs from previousapproaches by allowing GPU-mapped actors to have data dependent consumption ofinputs / production of outputs. Such flexibility is essential for configurableand adaptive applications that are becoming increasingly common in signalprocessing. In our experiments it is shown that this feature allows up to 5xincrease in application throughput. The proposed framework is validated by application examples from the videoprocessing and wireless communications domains. In the experiments theframework is compared to a well-known reference framework and it is shown thatthe proposed framework enables both a higher degree of flexibility and betterthroughput.