Architecture-and Workload-Aware Heterogeneous Algorithms for Sparse Matrix Vector Multiplication