use Benchmark qw(:hireswallclock timethese);
timethese( 1000000,{
  one => sub { require Benchmark },
  two => sub { require Benchmark; require Benchmark },
});
__END__
$ perl 1
Benchmark: timing 1000000 iterations of one, two...
 one: 2.43888 wallclock secs ( 1.70 usr +  0.00 sys =  1.70 CPU) @ 588
+235.29/s (n=1000000)
 two: 5.11982 wallclock secs ( 2.77 usr +  0.00 sys =  2.77 CPU) @ 361
+010.83/s (n=1000000)
[download]

Also, for some reason your list flattening with map() does not work. Not sure whether this is a bug in Perl, or a conceptual problem with using map() and $_. Observe:

use Encode qw(_utf8_on);
$a = [['']];
foreach (@$a) { _utf8_on($_) foreach @$_ } # my way
print utf8::is_utf8( $a->[0][0] ),$/;
$a = [['']];
_utf8_on( $_ ) for map @$_, @{$a}; # Aristotle's way
print utf8::is_utf8( $a->[0][0] ),$/;
__END__
1

[download]

But also from a performance point of view, the extra list flattening with map() is not very efficient:

use Encode qw(_utf8_on);
use Benchmark qw(:hireswallclock timethese);
push @$a,[(0) x 10] foreach 1..10;
timethese( 10000,{
  liz => sub {
    foreach (@{$a}) {
        _utf8_on( $_ ) foreach @{$_};
    }
},
  Aristotle => sub {
    _utf8_on( $_ ) for map @$_, @{$a};
},
});
__END__
Benchmark: timing 10000 iterations of Aristotle, liz...
 Aristotle: 4.37344 wallclock secs ( 3.73 usr + 0.00 sys = 3.73 CPU) @
+ 2680.97/s (n=10000)
 liz: 4.06957 wallclock secs ( 2.73 usr + 0.00 sys = 2.73 CPU) @ 3663.
+00/s (n=10000)
[download]

Liz

Update:
The way Aristotle proposed doesn't work because map() creates a copy of the elements, on which the UTF-8 flag is set and then discarded. See $_ and list flattening with map() for more info.