If you are familiar with Theano, you can check out the LSTM unit implementation here: https://github.com/fchollet/keras/blob/master/keras/layers/r...

Each input is propagated through the unit (all timesteps) before the gradient is computed. No RTRL. It works well in practice.