If order need not be preserved I think a hash slice is quicker.

$ perl -MBenchmark=cmpthese -Mstrict -wE '
> my @arr; 
> push @arr, int rand 20 for 1 .. 1000;
> cmpthese(
>    -5,
>    {
>        grep  => sub
>           {
>               my %seen;
>               my @uniq = grep ! $seen{ $_ } ++, @arr;
>           },
>        slice => sub
>           {
>               my %seen;
>               @seen{ @arr } = ( 1 ) x @arr;
>               my @uniq = keys %seen;
>           },
>    }
>    );'
        Rate  grep slice
grep  3041/s    --  -29%
slice 4260/s   40%    --
$
[download]

However, take my benchmarks with a pinch of salt, I often make a complete dog's breakfast of them :-(

I hope this is of interest.