A Task Parallel Accelerator with Dynamic Pipeline Balancing

Coarse-grained reconfigurable array(CGRA) based ac- celerator is a promising architecture to accelerate data processing workloads. CGRA-based accelerator features more flexibility in adopting various workloads while remaining powerful in comparison to the traditional application specified accelerator. However, the strength of CGRAs is limited by irregular data access and dependence patterns. The previously proposed task- based execution model enabled work-aware dynamic scheduling, but the effectiveness is still limited by the regular execution resources. To address these issues, we proposed a heterogeneous architecture to improve parallelism for pipeline-enabled task streaming. We enabled dynamic pipeline balancing on this architecture with minor modifications to the task stream annotation. We compare the execution result on various heterogeneous structures with the regular configuration. Overall, we find that our architecture can improve performance with the same overall resources.