A reminder: normality is assumed for the residuals of a modelling exercise, not the raw data. You must do the regression first and then check the residuals distribution.
Now, if your fish counts are real observed counts (and not some kind of scaled estimate) they are more likely to be distributed as Poisson variates, possibly with overdispersion. SAS offers many tools to model those: GENMOD, GLIMMIX, COUNTREG.
More details about the nature and origin of your data could get you better advice.
PG
PG