FPGA-based Implementation of Memory-Intensive Application using OpenCL
Iman Firmansyah, Changdao Du, Norihisa Fujita, Yoshiki Yamaguchi, Taisuke Boku · 2019
Systems with heterogeneous architectures, such as field programmable gate arrays (FPGAs) and graphics processing units (GPUs), are expected to boost the throughput in high-performance computing applications. However, the user may struggle to write sophisticated programs that are performed effectively on complicated heterogeneous systems. High-level synthesis (HLS) is a solution that can reduce the development time. This study focuses on the implementation of OpenCL programming as a type of HLS design on FPGA boards. The Himeno benchmark, which is a suitable benchmark for the measurement of memory-intensive applications, is chosen as a verification program. We found that the OpenCL-based implementation achieves reasonable performance on FPGAs by demonstrating the implementation of temporal blocking combined with shift register implementation simultaneously. For Stratix V DE5-Net FPGA, our current implementation achieves 10.62 GFLOPS, or 75% of the theoretical performance. Meanwhile, for Arria 10 A10PL4 FPGA, the peak performance reaches 13.95 GFLOPS, or 76% of the theoretical performance.