Hi,
relative_irf only normalizes the variables by the shock size so that the shock sizes are comparable. It does not rescale the other variables by their steady state values to make them percentage deviations!
The respective Dynare code for this option used in stoch_simul.m is
if options_.relative_irf
y = 100*y/cs(i,i);
end
where y is the standard IRF without relative_irf and cs is the cholesky factorization of the covariance matrix of the shocks, i.e. the IRFs are scaled by the standard deviation of the respective shock. If you want the model to be log-linearized, you have to put everything in exp() (search the forum).