Branch and Data Herding: Reducing Control and Memory Divergence for Error-tolerant GPU Applications