Simplified Graphical Processor Acceleration of the Standard and High-Order Finite-Difference Time-Domain Algorithms

Abstract

A computationally efficient and streamlined approach for coding a high-order finite-difference time-domain algorithm on both central and graphical processing units is presented. This objective was achieved through extending the update equations of the convolutional perfectly-matched layer absorbing boundary conditions throughout the numerical domain with appropriate parameter selections. It is demonstrated that the resulting appreciable increase in the floating-point operations count would result in only a negligible loss of overall computational efficiency using either central processing units or graphical processing units. This achievement translates into sizable reductions in code complexity and development costs. Comparative analyses were also presented for the standard finite-difference time-domain method, resulting in an efficient accelerated algorithm that does not degrade with model size reduction.