PGI User Forum

mkcolg · Posted: Mon Jan 28, 2013 12:54 pm Post subject:

Hi Sotiris,

To achieve 2-D parallelism here, you either need to make "A" 2-D or add a reduction.

paokara · Joined: 06 Feb 2011 Posts: 19

Hello Mat,

With have a problem with our code and we can't figure out the reason. We work on N-body problem and we've changed some of our FORTRAN functions using some OpenACC directives. We have only one "heavy" part, the part we calculate the accelerations( a 2D loop) with the way with have discussed in this topic. Also we are using a Tesla C1060.
Our implementation works fine with a small number of bodies but when we use a large number for our system (for example 10.000*10.000 for the 2-D in accelaration part) we get NaN in our results. We believe that this is a memory issue. Is that possible?

Another important issue is that we are using double precision variables for our code. Tesla C1060 has only 1/10 of the cores for double precision calculations. Is it possible that the device can't execute all the double precision calculations and give us NaN?

Thanks,
Sotiris

mkcolg · Posted: Mon Mar 18, 2013 12:54 pm Post subject:

paokara · Joined: 06 Feb 2011 Posts: 19

Hi Mat and thanks for the quick reply,

Today we finally figure out what the problem was.
To avoid useless calculations we use 2 "if" statements inside our openACC regions. We have one DATA region and every N steps we copy our data back to the host. The first step of those N steps we must do the calculations inside those two IF statements i mentioned above. But in the remaining N-1 steps we don't want this calculations. The first if statement is inside a PARALLEL region and the second IF statement is inside a KERNEL region. Here is the structrure of our code

mkcolg · Posted: Wed Mar 20, 2013 9:39 am Post subject:

Hi Sotiris,

By using the "parallel" construct, you are defining that everything in this region will be moved over the device. Also, it's work shared, so unless you define using the "loop" directive to tell the compiler how you want the work divided, all thread in the region will execute the same code. With the "kerenl" construct, the compiler figures out the best way to divide up the work. (FYI, this article may be useful http://www.pgroup.com/lit/articles/insider/v4n2a1.htm in understanding the differences between the two).

Without seeing the code, I can't be sure if this is the source of the NaNs, but it is possible. Do you get wrong answers if you switch to using just "kernels"?

- Mat