[Omp] Barrier constructs and regions

[Haab, Grant]

Here is the definition of region from the 2.5 specification, Section
1.2.2:

region:
All code encountered during a specific instance of the execution of a
given
construct or OpenMP library routine. A region includes any code in
called
routines as well as any implicit code introduced by the OpenMP
implementation.
COMMENTS:
A region may also be thought of as the dynamic or runtime extent of a
construct or OpenMP library routine.
During the execution of an OpenMP program, a construct may give
rise to many regions.


I think it is clear here that tbe language given in the first sentence:
"a given construct or OpenMP library routine"
makes it clear that a region can only be associated with a single 
construct, not more than one construct.

Thus, two distinct barrier constructs (in the code below) may not 
be associated with a single barrier region.
The two barrier constructs result in to two different barrier 
regions at run time, and thus, the program is non-conforming based 
on the restriction that all threads (or none at all) must execute each 
barrier region.

The non-conforming code below may work just fine on some compilers 
(in some situations), but is not guaranteed to work according to the 
specfication.

The example in the appendix is illustrating a completely different 
restriction than the restrictions Sanjiv mentions and is irrelevant to
the discussion.

- Grant
 

-----Original Message-----
From: Omp-bounces at openmp.org [mailto:Omp-bounces at openmp.org] On Behalf
Of Marina Kraeva
Sent: Tuesday, June 21, 2005 9:27 AM
To: Omp at openmp.org
Cc: Michael Suess
Subject: Re: [Omp] Barrier constructs and regions

Sanjiv,

Why do you say that the program below is non-conforming (in the quoted 
message you say it is "con-forming", but later you said it's 
non-conforming)?

Both restrictions that you quoted are satisfied - the barrier is called 
by all threads, and there is no work-sharing regions, so the sequence of

work-sharing regions and barrier regions is the same.

Look at the non-conforming example A.35.5 in the Specs 2.5. The barrier 
cannot be nested inside the critical region because only one thread AT A

TIME can enter the critical region. But in the Michael's example all 
threads can call barrier at the same time.

Marina.

Shah, Sanjiv wrote:

> This is a con-forming program that probably works with most
> implementations, although it's possible some flag an error (like maybe
> ours, I don't remember).  
> 
> The resons are as you quoted:
> 
> "The following restrictions apply to the barrier construct:
> - Each barrier region must be encountered by all threads in a team or
by
> none 
> at all
> - The sequence of work-sharing regions and barrier regions encountered
> must be
>   the same for every thread in a team."
> 
> Sanjiv 
> 
> 
> --
> Sanjiv, 217-403-4244
> 
> -----Original Message-----
> From: Omp-bounces at openmp.org [mailto:Omp-bounces at openmp.org] On Behalf
> Of Michael Suess
> Sent: Wednesday, June 15, 2005 9:20 AM
> To: Omp at openmp.org
> Subject: [Omp] Barrier constructs and regions
> 
> Hi,
> 
> I have a small request for clarification. Please consider the
following
> small 
> program:
> 
> #include <stdio.h>
> #include <omp.h>
> 
> #define THREADS 4
> 
> int main()
> {
> 
>         omp_set_num_threads(THREADS);
> 
> #       pragma omp parallel
>         {
>                 if (omp_get_thread_num() == 0) {
>                         #pragma omp barrier
>                 } else {
>                         #pragma omp barrier
>                 }
>         }
> 
>         return 0;
> }
> 
> The question is: Is this a standard compliant program (the problem are
> the 
> barriers here). The Spec has to say: 
> 
> "All of the threads of the team executing the binding parallel region
> must 
> execute the barrier region before any are allowed to continue
execution 
> beyond the barrier."
> 
> And:
> 
> "The following restrictions apply to the barrier construct:
> - Each barrier region must be encountered by all threads in a team or
by
> none 
> at all
> - The sequence of work-sharing regions and barrier regions encountered
> must be
>   the same for every thread in a team."
> 
> We have two barrier constructs. But do they create the same barrier
> region? If 
> yes, the program is compliant and should run just fine (it wont do
very
> much, 
> but thats not the point :-) ). If not, the program is not compliant
and 
> should hang (actually it's behaviour is undefined then and I would
> expect it 
> to hang). All compilers I have access to run the program just fine.
> 
> In the section about language terminology, it says in the comments
about
> 
> region: 
> "During the execution of an OpenMP program, a construct may give
> rise to many regions."
> 
> But can many constructs also give rise to the same region?
> 
> Thanks for your answers,
> Michael Suess
> 
> _______________________________________________
> Omp mailing list
> Omp at openmp.org
> http://openmp.org/mailman/listinfo/omp_openmp.org
> 
> _______________________________________________
> Omp mailing list
> Omp at openmp.org
> http://openmp.org/mailman/listinfo/omp_openmp.org
> 
> 

_______________________________________________
Omp mailing list
Omp at openmp.org
http://openmp.org/mailman/listinfo/omp_openmp.org