Understanding the efficiency of parallel incomplete Cholesky preconditioners on the performance of ICCG solvers for multi-core and GPU systems