Hybrid Single/Double Precision Floating-Point Computation on GPU Accelerators for 2-D FDTD