Optimizing Performance of Recurrent Neural Networks on GPUs