[PPL-devel] OpenMP & PPL

[Thomas LE MÉZO]

Thank you, I better understand the behavior of Thread_Init now.

The issue with OpenMP is that one thread of the parallel part will be kept (the master thread) to execute the remaining lines. For this thread, a PPL::Thread_Init variable must not be deleted. This is why the integration with OpenMP require some additional line of code:


#pragma omp parallel
{
        PPL::Thread_Init *thread_init = NULL;
        bool thread_init_valid = true;
        #pragma omp master
        {
            thread_init_valid = false;
        }
        if(thread_init_valid)
            thread_init = new PPL::Thread_Init();

        #pragma omp for
        for(int i=0; i<1000; i++){
            PPL::C_Polyhedron ph1(3);
            PPL::C_Polyhedron ph2(3);
            ph1.intersection_assign(ph2);
         }
         if(thread_init_valid)
            delete(thread_init);
}
PPL::C_Polyhedron ph3(3); // ==> Ok (executed by the master thread)


Regards,
Thomas.


________________________________
De : Enea Zaffanella <zaffanella.enea at gmail.com> de la part de Enea Zaffanella <zaffanella at cs.unipr.it>
Envoyé : jeudi 4 janvier 2018 14h53
À : Thomas LE MÉZO
Cc : The Parma Polyhedra Library developers' list
Objet : Re: [PPL-devel] OpenMP & PPL


On 04/01/2018 11:30, Thomas LE MÉZO wrote:

I have found an issue with the use of Thread_Init object, the following code does not work:


#pragma omp parallel
{
   PPL::Thread_Init thread_init;
#pragma omp for
    for(int i=0; i<1000; i++){
        PPL::C_Polyhedron ph1(3);
        PPL::C_Polyhedron ph2(3);
        ph1.intersection_assign(ph2);
    }
}
PPL::C_Polyhedron ph3(3); // => Segment fault

It seems that when all the thread_init variables are delete, it is then impossible to instantiate or proceed to any operation with ppl.

That is the expected behavior.
Each thread willing to call the PPL (even the master thread) should first call the thread initialization code.

A workaround is to instantiate a new Thread_Init variable but this is a bit heavy.

I can't see why this is "heavy".
Note: you can place the thread initialization code for the master thread in the main function.
The thread will be initialized once at program start and finalized when exiting the program.

Regards,
Enea.

#pragma omp parallel
{
   PPL::Thread_Init thread_init;
#pragma omp for
    for(int i=0; i<1000; i++){
        PPL::C_Polyhedron ph1(3);
        PPL::C_Polyhedron ph2(3);
        ph1.intersection_assign(ph2);
    }
}
PPL::Thread_Init thread_init;
PPL::C_Polyhedron ph3(3); // => OK

Any other idea?

Thomas

________________________________
De : Thomas LE MÉZO
Envoyé : mercredi 3 janvier 2018 12:26:09
À : enea.zaffanella at unipr.it<mailto:enea.zaffanella at unipr.it>
Cc : The Parma Polyhedra Library developers' list
Objet : RE: [PPL-devel] OpenMP & PPL


It seems to work! To avoid multiple instantiation of thread_init, a solution is to break down the "omp parallel for" in two lines:


#pragma omp parallel
{
   PPL::Thread_Init thread_init;
#pragma omp for
    for(int i=0; i<1000; i++){
        PPL::C_Polyhedron ph1(3);
        PPL::C_Polyhedron ph2(3);
        ph1.intersection_assign(ph2);
    }
}

Thank you very much,
Thomas


________________________________
De : Enea Zaffanella <zaffanella.enea at gmail.com><mailto:zaffanella.enea at gmail.com> de la part de Enea Zaffanella <zaffanella at cs.unipr.it><mailto:zaffanella at cs.unipr.it>
Envoyé : mercredi 3 janvier 2018 11h17
À : The Parma Polyhedra Library developers' list; Thomas LE MÉZO
Objet : Re: [PPL-devel] OpenMP & PPL



On 31/12/2017 17:33, Thomas LE MÉZO wrote:

Dear Enea, Dear all,


I tried recently to implement a multi-thread version of my thesis code using PPL library. I used OpenMP and the last version of ppl on devel git branch as I read you have been working on a thread-safe implementation (thread-safe option with configure). I didn't succeed in writing a correct program; see for instance this simple code that return a segmentation fault :


#include <omp.h>

namespace PPL = Parma_Polyhedra_Library;


int main(){

#pragma omp parallel for
    for(int i=0; i<1000; i++){
        PPL::C_Polyhedron ph1(3);  // => segmentation fault at line 83 of Linear_Expression.cc
        PPL::C_Polyhedron ph2(3);
        ph1.intersection_assign(ph2);
    }
    return 0;
}


Would you have any idea about what is wrong?

Each thread should be properly initialized by creating a PPL::Thread_Init object (and finalized by destroying it).
See for instance test01() in tests/Polyhedron/threadsafe1.cc
As an alternative, you could use the make_threadable wrapper (see test02() in the same test).


Is PPL compatible with using OpenMP?

I never tried using it with OpenMP ... however, as long as thread initializations (and finalizations) are performed as suggested above, things should be working (there is also a library initialization part, but that is automatically done when working in C++).

In you code, I would try adding

  PPL::Thread_Init thread_init;

as the very first line of the loop body.
This would maybe create more init objects than needed,
but (if I remember correctly) that should not be a problem.

Let us know how it goes.

Enea



Thanks a lot in advance,

Thomas




_______________________________________________
PPL-devel mailing list
PPL-devel at cs.unipr.it<mailto:PPL-devel at cs.unipr.it>
http://www.cs.unipr.it/mailman/listinfo/ppl-devel



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.cs.unipr.it/pipermail/ppl-devel/attachments/20180105/3903eea7/attachment.htm>