Parallel Pipeline on Heterogeneous Multi-processing Architectures

We address the problem of providing support for executing single streaming applications implemented as a pipeline of stages that run on heterogeneous chips comprised of several cores and one on-chip GPU. In this paper, we present an API that allows the user to specify the type of parallelism exploited by each pipeline stage running on the CPU multicore, the mapping of the pipeline stages to the devices (GPU or CPU), and the number of active threads. Using as case of study a real streaming application, we evaluate how these parameters affect the performance and energy efficiency of a heterogeneous on-chip processor (Exynos 5 Octa) that has three different computational cores: a GPU, an A15 quad-core and an A7 quad-core. We also explore some memory optimizations and find that while their performance impact depends on the granularity type, they usually reduce energy consumption.