Sorry for my earlier misdirection. By way of recompense I have what I believe (though it is essentially untested for lack of a suitable image), that addresses both the slowness of substr on utf strings within threads (which is just weird) and the problem I thought was the cause, that of cloning the returned array.

It avoids the former by doing away with the encoding, searching instead for runs of pairs of non-null characters in the unencoded pdl; and the latter by accumulating the counts in a packed binary array stored in a scalar.

sub test {
    my $fn  = shift;
    my $img = PDL::IO::Image-> new_from_file( $fn ) or die "Failed to 
+load image";
    my $pdl = $img->pixels_to_pdl->short;
    my $s   = cc8compt( $pdl != 0 );
    my $str = ${ $s-> get_dataref };

    my ( $w, $h ) = $s-> dims;

    my $bounds = pack 'n4', $w, 0, $h, 0;
    $bounds x= $s->max;

    for my $y ( 0 .. $h - 1 ) {
        my $s = substr( $str, 2 * $y * $w, 2 * $w );
        while( $s =~ m[(?:[^\0][^\0])+]g ) {
            my( $l, $r ) = ( $-[0]/2, (($+[0])-1)/2 );
            my $c = ord( $& );
            vec( $bounds, 4*$c+0, 16 ) = $l if $l < vec( $bounds, 4*$c
++0, 16 );
            vec( $bounds, 4*$c+1, 16 ) = $r if $r > vec( $bounds, 4*$c
++1, 16 );
            vec( $bounds, 4*$c+2, 16 ) = $y if $y < vec( $bounds, 4*$c
++2, 16 );
            vec( $bounds, 4*$c+3, 16 ) = $y if $y > vec( $bounds, 4*$c
++3, 16 );
        }
    }
    return $bounds;
}
[download]