(AS)2: Accelerator synthesis using algorithmic skeletons for rapid design space exploration

Hardware accelerators in heterogeneous multiprocessor system-on-chips are becoming popular as a means of meeting performance and energy efficiency requirements of modern embedded systems. Current design methods for accelerator synthesis, such as High-Level Synthesis, are not fully automated. Therefore, time consuming manual iterations are required to explore efficient accelerator alternatives: the programmer is still required to think in terms of the underlying architecture. In this paper, we present (AS)2: a design flow for Accelerator Synthesis using Algorithmic Skeletons. Skeletonization separates the structure of a parallel computation from an algorithms' functionality, enabling efficient implementations without requiring the programmer to have hardware knowledge. We define three such skeletons (for three image processing kernels) enabling FPGA specific parallelization techniques and optimizations. As a case study, we present a design space exploration of these skeletons and show how multiple design points with area-performance trade-offs for the accelerators can be efficiently and rapidly synthesized. We show that (AS)2 is a promising direction for accelerator synthesis as it generates a pareto front of 8 design points in under half an hour for each of the three image processing kernels.