PDLapi warnings in parallel calculations with PDL::Opt::NonLinear

[YuryPakhomov]

Hello all!

I use PDL::Opt::NonLinear module in my code. It works fine but slowly due to complex minimization function.
The minimization function can be paralleled with Parallel::ForkManager. It works fine in separate call.
But when I try to use the function in PDL::Opt::NonLinear module, many warnings appear:

	(in cleanup) INVALID PDL MAGICNO, got hex=0xf02100 (37593680)
 at /usr/local/share/perl5/Parallel/ForkManager/Child.pm line 26.
	eval {...} called at /usr/local/share/perl5/Parallel/ForkManager/Child.pm line 26
Warning: special data without datasv is not freed currently!! at test.pl line 0.
	eval {...} called at test.pl line 0
Warning: special data without datasv is not freed currently!! at test.pl line 0.
	eval {...} called at test.pl line 0
Warning: special data without datasv is not freed currently!! at test.pl line 0.
	eval {...} called at test.pl line 0
Warning: special data without datasv is not freed currently!! at test.pl line 0.
	eval {...} called at test.pl line 0
Warning: special data without datasv is not freed currently!! at test.pl line 0.
	eval {...} called at test.pl line 0

The simple code to reproduce this output is:

#! /usr/bin/perl
use PDL;
use PDL::NiceSlice;
use PDL::Opt::NonLinear;
use Parallel::ForkManager;

my $pm = Parallel::ForkManager->new(4);

my $x=pdl(1);
my $fx=pdl(0);
my $gx=zeroes(nelem($x));

# This single call works fine
#fg_func($fx,$gx,$x); exit;

# Optimization algorithm
                my $bounds =  zeroes(nelem($x),2);
                $bounds(0,0).=  pdl(0.0);
                $bounds(0,1).= pdl(10.0);
                
                my $tbounds = zeroes(nelem($x));
                $tbounds.=2;

                my $gtol = pdl(0.9);
                my $pgtol = pdl(0.1);
                my $factr = pdl(1e7);
                my $m = pdl(10);
                my $print = pdl(-1);
                my $maxit = pdl(long,15);
                my $info = pdl(long,1);
                my $iv = zeroes(long,44);
                my $v = zeroes(29);

                lbfgsb($fx, $gx, $x, $m, $bounds, $tbounds, $maxit, $factr, $pgtol, $gtol,
                $print, $info,$iv, $v,\&fg_func);

sub fg_func{
 my ($f, $g, $x) = @_;
 $f .= 0;
 $g .= pdl(0);

# Parallel calculations block. This is demo without any calculation. Create external thread and then terminate it.
# If to comment this block then program works fine
 DATA_LOOP:
 for my $i (0 .. nelem($x)-1){
  my $ppid = $pm->start and next DATA_LOOP;
  $pm->finish($i); # Terminates the child process
 }
 $pm->wait_all_children;

 return 0;
}

The Warning points to "Child.pm" line 26, which contains CORE::exit($x || 0) of subroutine finish:

sub finish {
  my ($s, $x, $r)=@_;

  $s->store($r);

    CORE::exit($x || 0);
}

The messages "INVALID PDL MAGICNO" and "Warning: special data without datasv is not freed currently!" are generated by Basic/Core/pdlapi.c

[mohawk2]

The good news is that I can reproduce that locally! More to follow.

[mohawk2]

I suspect this is some horrible interaction with garbage collection and processes exiting. If we change your script from using $pm->finish to POSIX::_exit, the error doesn't occur. I don't understand why yet, since the PDL objects should be destroyed as normal in sub-processes.

There's a clue in that turning on PDL::Core::set_debugging(1) shows that the ndarray that's getting the cleanup message has state PDL_DONTTOUCHDATA, which triggers the message. That's being set in PDL::Opt::NonLinear's C wrapper for Perl functions (like your fg_func), which is why you don't see errors when you call your function standalone, it's specifically a P:O:NL thing.

[mohawk2]

I think that's somewhat the answer; P:O:NL isn't designed to be used in the way you're using it. It sets up some very thin ndarrays (with a data pointer, a dims array, and a bit of state to say "don't touch"), then calls the user-supplied Perl function with them. On return, it then cleans those up, but your call to exit in the middle stops the last part, so they're visible to Perl which tries to call DESTROY.

A workaround is to use POSIX::_exit as identified above. Also, PDL is working as designed, so I'm transferring this issue to P:O:NL.

[mohawk2]

There is nothing else that can be done on this issue.