Novel architecture for loop acceleration: a case study

In this paper, we show a novel approach to accelerate loops by tightly coupling a coprocessor to an ASIP. Latency hiding is used to exploit the parallelism available in this architecture. To illustrate the advantages of this approach, we investigate a JPEG encoding algorithm and accelerate one of its loop by implementing it in a coprocessor. We contrast the acceleration by implementing the critical segment as two different coprocessors and a set of customized instructions. The two different coprocessor approaches are: a high-level synthesis (HLS) approach; and a custom coprocessor approach. The HLS approach provides a faster method of generating coprocessors. We show that a loop performance improvement of 2.57x is achieved using the custom coprocessor approach, compared to 1.58x for the HLS approach and 1.33x for the customized instruction approach compared with just the main processor. Respective energy savings within the loop are 57%, 28% and 19%.