Task-DAG support in single-source PHAST library: Enabling flexible assignment of tasks to CPUs and GPUs in heterogeneous architectures

Nowadays, the majority of desktop, mobile, and embedded devices in the consumer and industrial markets are heterogeneous, as they contain at least multi-core CPU and GPU resources in the same system. However, exploiting the performance and energy-efficiency of these diverse processing elements does not come for free from a software point of view: programmers need to a) code each activity through the specific approaches, libraries, and frameworks suitable for their target architecture (e.g., CPUs and GPUs) along with the orchestration of such heterogeneous execution, and b) decide the distribution of sequential and parallel activities towards the different parallel hardware resources available. Current frameworks typically provide either low-abstraction-level target-specific and/or generic but not high-performance interfaces, which complicate the exploration of different task assignments, with DAG1 precedence relationship, to the available heterogeneous resources. To enable this, tasks would typically need to be coded one time for each target architecture due to the profound differences in their programming. In this work, we include the support of tasks and DAGs of data-parallel tasks within the single-source PHAST library, which currently supports both multi-core CPUs and NVIDIA GPUs, so that tasks are coded in a target-agnostic fashion and their targeting to multi-core or GPU architectures is automatic and efficient. The integration of this coding approach with tasks can help to postpone the choice of the execution platform for each task up to the testing, or even to the runtime, phase. Finally, we demonstrate the effects of this approach in the case of a sample image pipeline benchmark from the computer vision domain. We compare our implementation to a SYCL implementation from a productivity point of view. Also, we show that various task assignments can be seamlessly explored by implementing both the PEFT2 mapping technique along with an exhaustive search in the mapping space