Automatic mapping of control-intensive kernels onto coarse-grained reconfigurable array architecture with speculative execution

Coarse-grained reconfigurable array architectures have drawn increasing attention due to their good performance and flexibility. In general, they show high performance for compute-intensive kernel code, but cannot handle control-intensive parts efficiently, thereby degrading the overall performance. In this paper, we present automatic mapping of control-intensive kernels onto coarse-grained reconfigurable array architecture by using kernel-level speculative execution. Experimental results show that our automatic mapping tool successfully handles control-intensive kernels for coarse-grained reconfigurable array architecture. In particular, it improves the performance of the H.264 deblocking filters for luma and chroma over 26 and 16 times respectively compared to conventional software implementation. Compared to the approach using predicated execution, the proposed approach achieves 2.27 times performance enhancement.