Accurate Cross-Architecture Performance Modeling for Sparse Matrix-Vector Multiplication (SpMV) on GPUs