topic problem with data share about calling dll in openmp in Intel® Moderncode for Parallel Architectures

problem with data share about calling dll in openmp

ollehyas — Mon, 18 Jul 2011 06:49:29 GMT

Hi, guys, long time no see.

Recently, I am busy with some numerical computation. In the work, I have to call a Fortran function from a dynamic link library in my c++ code. The Fortran dynamic link library was made with old syntax-F77, having many COMMON statements. The Fortran function is called many times in a loop for some purpose with different input parameters. To speed up, I parallel the loop like this:

#pragma omp paralle for schedule(static)
for(int i=0;i<1000;i++)
{
....
Fortran_function(parameter_1,parameter_2);
.....
}
when debuging the code with Intel Parallel Debugger Extension, I found some data race. But, after fixed the bugs, the results from parallel is rather different from sequential results. In the loopall variabls are defined as local, so I doubt whether the reason is from the Fortran DLL. Is the COMMON statement in Fortran DLL causing share conflict? How to fix it?

problem with data share about calling dll in openmp

jimdempseyatthecove — Tue, 19 Jul 2011 22:30:49 GMT

If your DLL function is using "local" arrays .AND. if the function is not declared with RECURSIVE (or with -openpm) then the local arrays are attributed withSAVE. IOW

real:: vec(3)

is a SAVE'e (shared)array.

Your DLL function has to be compiled either a) with RECURSIVE or b) with -openmp option (eventhough it is not using !$OMP... code.

Jim Dempsey

problem with data share about calling dll in openmp

TimP — Wed, 20 Jul 2011 18:28:45 GMT

Jim's comment is correct for Intel Fortran. A subroutine which is to be active in a parallel region should have the RECURSIVE attribute, or be compiled with a non-default thread safety option e.g. -auto -openmp .....
A C compiler with default auto definition of an array within a parallel region will make the array private automatically. You would have trouble if a shared array (as when defined outside the parallel region, including a static) is updated by more than one thread.