A model-driven partitioning and auto-tuning integrated framework for sparse matrix-vector multiplication on GPUs