Problems with FORTRAN Accelerator and subroutines

[nicolaprandi]

@mkcolg

You can mix the two and the PGI Accelerator model can access CUDA Fortran device variables. However, support for CUDA Fortran accessing PGI Accelerator model device variables is still in process. If you have an example of what you are try to do, I can put in a feature request.

I tried to have a kernel push data back to the host by using the mirror-reflected directives (by using the “update” clause inside the kernel), but now I understand it shouldn’t be possible.


Thanks again,

Nicola.

[nicolaprandi]

Hi again! After I got some good results with the CUDA FORTRAN version of the SWE1D code (and some holidays), I got back to the PGI Acclerator’s one in order to compare the performances between the two of them.

I got the PGI Accelerator version working, but only for small size arrays: if I try to use sizes of 10000* or more elements (I should be using sizes of 100000-200000 for my applications), I get “cuMemAlloc error 2” (aka “Out of Memory”).

*Actually, I tested a little bit more the code and discovered that the “memory out limit” is quite random: sometimes, it’s given by 1000 elemts and other times by 20000 elements.

Any suggestion? As the usual, I’ve uploaded the latest version of the code on MediaFire:

http://www.mediafire.com/?4dos3otya1dhvy2

In particular, the parameter NC inside the “Input.log” file determines the size of the arrays used by the code.


Thanks in advance, Nicola.