← Optimizing Parallel Reduction in CUDA