Control statements such as loops and branches poseserious challenges for their efficient utilization onGraphic Processing Units (GPUs) as those controlstatements will lead to a serialization of threads andconsequently ruin the occupancy and parallelism on GPUs. Unlike traditional central processingunits (CPUs), the GPU cannot leave the controlstatements to the CPU because fine-grain statementscheduling between GPU and CPU cannot begranted, as the GPU acts as a co-processing device.This paper analyzes the impact for using two leveltransformation techniques, namely loop/branchsplitting, which improves the register utilizationto manage the control statements on GPUs, inorder to implement the Lattice Boltzmann Method(LBM) benchmark application. Results executed inthe NVIDIA G80 architecture illustrate that thesetechniques are very efficient in term of parallelismand can lead to an increase in occupancy and adrastic improvement in performance, compared tonon-split version of the programs.