Why this code is so slow if run in thread?

Because you are returning a ref to an array of arrays, which in the unthreaded code means simply returning a scalar; but in the threaded code, means duplicating the entire contents of that AoAs into the calling threads memory and then returning a reference to the copy.

This can be demonstrated without PDL or Encode like this:

#! perl -slw
use strict;
use threads;
use Time::HiRes qw[ time ];

sub test {
    my @b = map[[0,0],[0,0]], 0 .. 255;
    return \@b;
}

my $start = time;
my $nothread = @{ test() };
printf "Unthreaded took: %.6f\n", time() - $start;

$start = time;
my $thread = @{ threads->create( \&test )->join };
printf "Threaded took: %.6f\n", time() - $start;

__END__
c:\test>junk
Unthreaded took: 0.000851
Threaded took: 0.012510
[download]

The simple solution (for your specific posted example, which isn't using the contents of the AoA, but only the count of elements on the outer array ) ):

say 'Count: ', scalar @{ threads-> create( \&test, 'test.png' )-> join
+ };
[download]

would be to derive the count internal to the function and thus return a scalar rather than a compound data structure:

    ...
    shift @b;
    return scalar @b;
}
[download]

If your real code actually needs the AoAs returned to the main thread, then show an example of what you are doing with it there and I can perhaps suggest something else.