Towards Production Code Effective Portability among Vector Machines and Microprocessor-Based Architectures

Effective portability of production codes is a goal that has been pursued over decades with modest success. This paper presents a technique that enhances efficiency of a single code fragment (advection) of a production weather forecast code on both a vector machine and an IA-32-based machine. The technique combines vector instructions with cache reuse and small memory footprint, using vector length as the single performance parameter. Performance results compare favorably with current performance levels for production weather forecast codes