Automatically Tuning Sparse Matrix-Vector Multiplication for GPU Architectures