PGI User Forum

[KarlWilkinson85254]

Hi,

I am attempting to incorporate a layer of OpenMP code within an MPI program.
Unfortunately I seem to be falling foul of thread safety when attempting to perform FFTs within the OpenMP threads.

I am fairly convinced that thread safety is the issue as the code executes without issue if the call to dfftw_execute_dft is performed within an omp critical region but fails in parallel with a double free or corruption error.

As far as I can tell the code is structured in the manner advised by the fftw documentation: The plan is generated before the OpenMP region is created and dfftw_execute_dft is supposed to be a thread safe routine for executing the FFTs.

I have attempted to use the FFTW_UNALIGNED flag as specified in a post that describes a very similar problem on: http://fortran.itgroups.info/69/9/thread-1901648.html with no joy.

Before I resort to changing the manner in which the workload is distributed I wondered of there may be anything I missed?

Thanks in advance,

Karl

[mkcolg]

Hi Karl,

I'm not sure. Using the sample code in the link you reference, if I move the create and destroy plan calls outside the OpenMP region, the code runs correctly. I did make the mistake of forgetting to move the "plan_r2c" out of the private clause which caused a segv, but was find once made shared. Could this be your issue as well?

- Mat