Performance Improvement of Iterative Methods using a Bit-Representation Technique for Coefficient Matrices

Practical simulators require high-performance iterative methods and efficient implementations of various boundary conditions, which reflect real-world effects. A nobel Bit-representation technique is proposed to enhance the performance of such a code. This technique is applied to the implementation of iterative kernels that treat various boundary conditions. The first advantage of this approach is that it reduces memory traffic and effectively utilizes Single Instruction Multiple Data stream (SIMD) units with cache. Secondly, the proposed implementation can replace if-branch statements with mask operations using a bit expression. This promotes the optimization of code during compilation and run-time. Consequently, the proposed approach achieves 3.5 times faster than a näıve implementation on both Intel and Fujitsu Sparc architectures.