I still think you would need to log it before clustering it. The purpose of taking the log of data is to reduce the effect of outliers. That will apply here as well. I know in our software (OncoLand) we do this when looking at heatmaps of RPKM data and automatically clustering.

The choice may depend on why you are clustering it. I find that mean centering the data and representing it as z-scores (also called scaling the data) as you mentioned, is a generally useful way to group genes with common behaviors. There's a function in R called: scale() that does this. Though beware as it scales columns by default so you have to wrap your call in t() to transpose it.