Offloading Transprecision Calculation Using FPGA.

For reducing calculation time and energy consumption, transprecision, which calculates within proper precision in each phase in an application, comes into the spotlight in High-Performance Computing. However, we have to emulate it in CPU or GPU, and it is not easy to evaluate the whole application. In this paper, we implement transprecision calculation using HDL and High-Level Synthesis by C++ on FPGA. Furthermore, we evaluate it in matrix multiply and stencil computation by FPGA offloading execution. As a result, we realized an arithmetic unit with the expected arbitrary precision.We also confirmed the tendency to increase the throughput by increasing the parallelism using fewer bit lengths.