A just-in-time modulo scheduling for virtual coarse-grained reconfigurable architectures

In the past decade, most solutions concerning the mapping of the compute-intensive loop kernels to accelerators have used heuristics and compiler-based strategies. These facts require that most of the decisions be taken at design time, thus precluding efficient solutions that can take run-time information into account. Any success in accelerating such applications greatly depends on two steps, extracting the loops and mapping them into the architecture. This last step is a challenge in itself since it is a NP-complete problem. In this paper, we propose a runtime solution that can provide speed ups of 3 to 6 orders of magnitude for the mapping step when compared to the state-of-the-art at minimal performance degradation, by the combined usage of 3 distinct mechanisms: 1) a simple and efficient modulo scheduling heuristic, 2) a crossbar network, which simplifies the placement and routing, 3) a virtual coarse-grained reconfigurable architecture (CGRA). Additionally, since the CGRA is a virtual layer on top of an FPGA, it is possible to use any off-the-shelf FPGA without the need of special tools or IP solutions. Although the mapping is NP-complete even for crossbar-based CGRAs, experimental results demonstrate a huge reduction in compilation time, as opposed to previous solutions that require seconds to map the applications, our solution requires only microseconds to find near optimal schedules. Besides the speed up, the proposed solution enables the use of just-in-time compilation, hence it is intrinsically adaptive to a changing scenario.