Hi Jim,
what is the purpose of the "large data structure"? Is it a sort of
scratchpad, which is not directly used in the final result? In which
case I'd expect something like:
public class TestVal {
var result : Double;
public def someOperation() {
var reg: Region{rank==1} = [1..8];
var dist: Dist{rank==1} = Dist.make(reg);
val sharedData = 2.71828;
finish ateach((i) in dist) {
val c = new LargeObject(sharedData, ...);
// complicated work on c
result += c.result();
}
}
}
Or is the large data structure used to hold a partial result for each
iteration, and the partial results are accumulated to form the final
result? In which case:
public class TestVal {
public def someOperation() {
var reg: Region{rank==1} = [1..8];
var dist: Dist{rank==1} = Dist.make(reg);
val sharedData = 2.71828;
val partialResults = Array.make[LargeObject](dist, (Point)=> new
LargeObject(sharedData, ...));
finish ateach((i) in partialResults) {
val c = partialResults(i);
// complicated work on c
}
result = partialResults.reduce((a:LargeObject,b.LargeObject) => a.add(b),
LargeObject.NULL);
}
}
The second case could be seen as a complicated analogue of an OpenMP
reduction variable. There has been some discussion on x10-core about
introducing a "collecting finish" which would simplify this greatly.
See
https://sourceforge.net/mailarchive/forum.php?thread_name=z2w7a30b9461004211429u1e6349b3u517c1c576e97d8%40mail.gmail.com&forum_name=x10-core
Cheers,
Josh
Jim LaGrone wrote:
>>> I'm still looking for an answer to parallelizing a loop with common
>>> data for each iteration.
>>>
>> Sorry, things have been a little hectic and this must have slipped through
>> the cracks.
>>
>>
>>> It is my understanding that the value vC
>>> would be private to each iteration of the ateach loop. This doesn't
>>> seem to be the case. Output is below. I'm using version 1.7.7.
>>>
>> vC is private, but the object it is pointing to is not.
>>
>> I believe in 1.7.7 Complex is an object,
>>
>
> I wrote my own class for this.
>
>
>> so although each iteration has its
>> own private pointer to C, they are still pointing to the same object.
>> Therefore the updates from one iteration are visible in others because
>> there is a single object that is being accessed by all the iterations.
>>
>> Does that make sense?
>>
>
> That makes sense, but obviously this is not what I want. Is there a simple
> way in X10 to get a such an object to be privatized among the iterations?
> This code snippet is an example of the code I'm working with. The actual code
> involves a large data structure which needs to be unique to each iteration
> with no loop carried dependences. Sequentially this is no problem.
> Concurrently, each iteration needs a separate copy.
>
> Any help is appreciated.
>
> Jim
>
>
>>> public class TestVal{
>>> public static def main(args:Rail[String]){
>>> var c: Complex = new Complex();
>>> var reg: Region{rank==1} = [1..8];
>>> var dist: Dist{rank==1} = Dist.make(reg);
>>>
>>> val vC = c;
>>> finish ateach((i) in dist)
>>> {
>>> vC.imag = i as Double;
>>> for ( var n: Int = 0; n < 100; n++ );
>>> Console.OUT.println(vC);
>>> }
>>> }
>>> }
>>>
>>>
>>> Output: Notice the absence of (0,8) and the replication of (0,2). It
>>> seems there is a data race on the vC, but my understanding is that
>>> vC would be local to each thread.
>>>
>>> $ x10 TestVal
>>> (0.0, 1.0)
>>> (0.0, 7.0)
>>> (0.0, 6.0)
>>> (0.0, 5.0)
>>> (0.0, 4.0)
>>> (0.0, 3.0)
>>> (0.0, 2.0)
>>> (0.0, 2.0)
>>>
>>>
>>> If I am misunderstanding this, can anyone offer a correction? In
>>> this example, I want vC to be private (local) to each iteration.
>>> I.e., each iteration gets its own copy.
>>>
>>> Thanks,
>>>
>>> Jim
>>>
>>
>>
>
> ------------------------------------------------------------------------------
>
> _______________________________________________
> X10-users mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/x10-users
>
------------------------------------------------------------------------------
_______________________________________________
X10-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/x10-users
