Design space exploration in heterogeneous platforms using OpenMP

Abstract

In the fields of high performance computing (HPC) and embedded systems, the current trend is to employ heterogeneous platforms which integrate general purpose CPUs with specialized accelerators such as GPUs and FPGAs. Programming these architectures to approach their theoretical performance limits is a complex issue. In this article, we present a design methodology targeting heterogeneous platforms which combines a novel dynamic offloading mechanism for OpenMP and a scheduling strategy for assigning tasks to accelerator devices. The current OpenMP offloading model depends on the compiler supporting each target device, with many architectures still unsupported by the most popular compilers, such as GCC and Clang. In our approach, the software and/or hardware design flows for programming the accelerators are dissociated from the host OpenMP compiler and the device-specific implementations are dynamically loaded at runtime. Moreover, the assignment of tasks to computing resources is dynamically evaluated at runtime, with the aim of maximizing performance when using the available resources. The proposed methodology has been applied to a video processing system as a test case. The results demonstrate the flexibility of the proposal by exploiting different heterogeneous platforms and design particularities of devices, leading to a significant performance improvement.