Revisiting Online Autotuning for Sparse-Matrix Vector Multiplication Kernels on High-Performance Accelerators