I found out that all versions are extremely fast, and thus not easy to benchmark. So probably my benchmark is flawed, take it with a grain of salt.

use strict;
use warnings;
use Benchmark qw(cmpthese);

my $x = "0123456789abcdefg" x 6000 . "XXX";

sub subst   { my $s = $x; $s =~ s/(.{16})/$1 /sg; $s };
sub up      { join ' ', unpack '(A16)*', $x }
sub m_substr {
    my $s = $x;
    for (my $i=16*int(length($s)/16);$i;$i-=16) { 
       substr($s,$i,0)=" "; 
    }
    $s;
}

sub readline {
    local $/ = \16;
    my $s;
    open my $handle, '<', \$x;
    $s = <$handle>;
    while (<$handle>) {
        $s .= ' ' . $_;
    }
    $s;
}

my $ref = subst();

sub wrap {
    my ($impl, $sub) = @_;
    die "Wrong result for $impl!" unless $sub->() eq $ref;
}

cmpthese(-1, {
    subst   => wrap('subst', \&subst),
    unpack  => wrap('unpack',\&up),
    substr  => wrap('substr',\&m_substr),
    readl   => wrap('readline', \&readline),
})
[download]

Here are the results (with perl5.10.0):

              Rate  readl  subst unpack substr
readl   78783885/s     --   -17%   -18%   -24%
subst   95302528/s    21%     --    -1%    -8%
unpack  96105165/s    22%     1%     --    -8%
substr 103929656/s    32%     9%     8%     --
[download]

              Rate  subst  readl substr unpack
subst   81695687/s     --    -1%   -36%   -65%
readl   82899185/s     1%     --   -35%   -65%
substr 128382089/s    57%    55%     --   -46%
unpack 235800076/s   189%   184%    84%     --
[download]

I increased the length of the source string by a factor of 100, and started the benchmark again. It still runs (after about 15 minutes).

My conclusions are that the differences are so small (or my benchmark so flawed...) that you can basically neglect them.