Accuracy, Memory, and Speed Strategies in GPU-Based Finite-Element Matrix-Generation