Accelerating GPU-Based Parallel FDTD With Advanced Operator Fusion

Abstract

The finite-difference time-domain (FDTD) method is one of the most widely used methods for solving Maxwell’s equations, but its efficiency is limited by the Courant–Friedrichs–Lewy stability condition. Recent research has extensively explored graphics processing unit (GPU)-based parallel implementations of the FDTD method to enhance computational performance. However, the inherent time-stepping and time-marching nature of the FDTD algorithm leads to frequent kernel launches and low memory efficiency on GPUs, still significantly impacting execution efficiency. This letter proposes a GPU-based FDTD framework enhanced with operator fusion to address this challenge. Within this framework, the FDTD algorithm is represented as a computation graph composed of operators. We classify these operators into different types based on their input and output relationships. Using this type of information, a rule-based strategy is developed to merge the operators into larger computational kernels, effectively enhancing GPU execution efficiency. Simulation results demonstrate that the proposed operator fusion framework does not introduce additional errors while achieving a 4× speedup compared to conventional GPU-based implementations.