[Omp] A question about threadprivate

[Dieter an Mey]

Sorry for generating the misunderstanding.

The "data race" concerned me and Larry answering your first mail at the 
same time.
There is no data race in your code. :-)=

Actually I think that the behaviour is clearly defined by the 2.5 
specification.

In case 2 the inner teams' master thread set their value of i to
	-omp_get_thread_num()
which is 0 of course.

So I don't see any problem here.
Open 3.0 will hopefully be more specific about threadprivate variables 
in inner parallel regions.

best regards
Dieter


Meadows, Lawrence F schrieb:
> So once again I state that the behavior in the first example is
> guaranteed by the (2.5) standard, and the behavior in the
> second is not specified.
> 
> I agree that it would be nice to have both specified. I'm not sure
> where the lang committee is on that topic. 
> 
> -----Original Message-----
> From: Marina Kraeva [mailto:kraeva at iastate.edu] 
> Sent: Wednesday, April 11, 2007 10:12 AM
> To: anmey at rz.rwth-aachen.de; Meadows, Lawrence F
> Cc: omp at openmp.org
> Subject: Re: [Omp] A question about threadprivate
> 
> Dieter and Larry,
> 
> Thanks for your input!
> 
> I did not understand whether Dieter agreed that array is guaranteed to 
> have values (0,1,2,3) in the first example, or not. And it looks like 
> Dieter thinks that in the second example array is guaranteed to have 
> values (-0,-0,-0,-0).
> 
> I think that the behavior in both cases should be specified in API 3.0, 
> or at least it should be said that that the behavior is implementation 
> defined. Otherwise it won't be safe to use threadprivate and nested 
> parallel in the same program.
> 
> Marina.
> 
> Dieter an Mey wrote:
> 
>> Hi,
>>
>> In general in the nested case each thread of the inner teams will get 
>> their own copies.
>> They will be initialized "at an unspecified point in the program
> porior 
>> to the first reference to that copy".
>> To my understanding of the specs ...
>> In case 1 the threadprivate variable is not referenced, so it is not 
>> specified whether the inner teams' slave thread's copy will be 
>> initialized at all.
>> References by the inner team's master threads "will be to the copy of 
>> the object in the thread which encountered the (remark: inner)
> parallel 
>> region".
>> Therefore in case 2 I would expect that the inner teams' master
> threads' 
>> copies will be modified in the inner parallel region.
>> Therefore array will be set to -0 for all indeces.
>>
>> best regards
>> Dieter
>>
>> Marina Kraeva schrieb:
>>
>>> Hello,
>>>
>>> Could you please share your opinion on what will happen in the 
>>> following two small examples.
>>>
>>> In the OpenMP API 2.5 (and in the earlier versions) it is said:
>>> "The values of data in the threadprivate objects of threads other
> than 
>>> the initial thread are guaranteed to persist between two consecutive 
>>> active parallel regions only if all the following conditions hold:
>>> * Neither parallel region is nested inside another parallel
> region..."
>>>
>>> According to API, is it guaranteed that in the following example the 
>>> values of 'i' will persist, so array will have values (0,1,2,3)?
>>>
>>> int i,array(0:3)
>>> threadprivate(i)
>>> !$omp parallel num_threads(4)
>>>     i=omp_get_thread_num()
>>>     !$omp parallel num_threads(2)
>>>         ... ('i' is not referenced)
>>>     !$omp end parallel
>>> !$omp end parallel
>>>
>>> !$omp parallel num_threads(4) shared(array)
>>>     array(omp_get_thread_num())=i
>>> !$omp end parallel
>>>
>>>
>>>
>>> What if inside the nested parallel region 'i' was modified:
>>>
>>> int i,array(0:3)
>>> threadprivate(i)
>>> !$omp parallel num_threads(4)
>>>     i=omp_get_thread_num()
>>>     !$omp parallel num_threads(2)
>>>                 i=-omp_get_thread_num()
>>>     !$omp end parallel
>>> !$omp end parallel
>>>
>>> !$omp parallel num_threads(4) shared(array)
>>>     array(omp_get_thread_num())=i
>>> !$omp end parallel
>>>
>>> Is it guaranteed, that array still will have values (0,1,2,3)?
>>>
>>> Thank you!
>>> Marina Kraeva.
>>> _______________________________________________
>>> Omp mailing list
>>> Omp at openmp.org
>>> http://openmp.org/mailman/listinfo/omp
>>>
> 

-- 
--------------------------------------------------------------------
Dieter an Mey
High Performance Computing               Hochleistungsrechnen
RWTH Aachen University                   Rechen- und Kommunikations-
Center for Computing and Communication   zentrum der RWTH Aachen
phone: ++49-(0)241-80-24377              Seffenter Weg 23
fax:   ++49-(0)241-80-22134              52074 Aachen, Germany
email: anmey at rz.rwth-aachen.de
--------------------------------------------------------------------