topic How to parallelize FFT computation in MKL? in Intel® oneAPI Math Kernel Library

How to parallelize FFT computation in MKL?

hadesmajesty — Sun, 07 Nov 2010 13:42:05 GMT

I try to parallelize the fft computation in each row of the matrix fu(m,2), but the result is not correct. The following is my code:

type(DFTI_DESCRIPTOR), SAVE,POINTER :: My_FFT_Handle

nthreads=omp_get_max_threads ();

Status = DftiCreateDescriptor(My_FFT_Handle, DFTI_DOUBLE, DFTI_REAL, 1, m)

Status = DftiSetValue (My_FFT_Handle, DFTI_NUMBER_OF_USER_THREADS, nThreads);

Status = DftiCommitDescriptor(My_FFT_Handle)

!$OMP PARALLEL DO

do ic=1,2

Status = DftiComputeForward(My_FFT_Handle,fu(:,ic))

end do

!$OMP END PARALLEL DO

stop

The results of fu(:,2) is sometimes not correct. Whats wrong in the code?

Thank you.

How to parallelize FFT computation in MKL?

Gennady_F_Intel — Sun, 07 Nov 2010 17:04:44 GMT

Please look at the example #4 from the KB "Different parallelization techniques and Intel MKL FFT"

How to parallelize FFT computation in MKL?

Dmitry_B_Intel — Mon, 08 Nov 2010 05:21:48 GMT

Hi, hidesmajesty,

The real in-place FFT specified by DftiCreateDescriptor(..., DFTI_REAL, 1, M) accepts input of M real values and needs space of M+2 real elements for output. That means the matrix should be declared/allocated thus:
double precision fu(M+2,2)

Thanks
Dima

How to parallelize FFT computation in MKL?

hadesmajesty — Mon, 08 Nov 2010 06:38:27 GMT

Thank you Dima.

It seems there is no difference by allocating fu(M+2,2) in sequential code, but by doing this in parallel code, I can obtain the correct result.

By the way, the additional data in fu(M+1:M+2,:) are meaningless and they are there to just make the routine work, are they?