A journey from single-GPU to optimized multi-GPU SPH with CUDA