my @bits_in = map $bits_in{chr($_ & 0xFF)} + $bits_in{chr($_ >> 8)}, (0..65535) ;

# Use vec() and @bits_in -- faster than ord(substr(..))

sub loop_count_va {       # "lpva"
    my ( $bits ) = @_;

    my $count;
    $count += $bits_in[vec($bits, $_, 8)] for 0 .. length( $bits ) - 1;

    return $count;
}

# As "lpva" but 16 bits at a time

sub loop_count_VA {       # "lpVA"
    my ( $bits ) = @_;

    my $count;
    $count += $bits_in[vec($bits, $_, 16)] for 0 .. (length( $bits ) - 1) >> 1 ;

    return $count;
}

# The other other scheme

sub unpack_count {        # "unpk"
    my ( $bits ) = @_ ;

    my $count = unpack("%32b*", $bits);

    if ($count != $COUNT) { die "unpack_count $count != $COUNT" ; } ;

    return $count;
}