FFTW OpenMP Parallelization?

user_10216005 0
Viewed 143

I've been having some trouble creating parallelized OMP DFTs with FFTW. I compiled with --enable-openmp and have the following code sample:

#include <fftw3.h>

int NUMBER_OF_THREADS = 40;
fftw_init_threads();
fftw_plan_with_nthreads(NUMBER_OF_THREADS);

std::vector<std::complex<double>> buf(num_trials);
fftw_plan p = fftw_plan_dft_1d(num_trials, reinterpret_cast<fftw_complex*>(buf.data()), reinterpret_cast<fftw_complex*>(buf.data()), -1, FFTW_MEASURE);

// plan is called here sequentially some number of times
for (dummy in dummy) {
   fftw_execute(p);
}

fftw_destroy_plan(p);

Unfortunately, the code doesn't seem to be parallelizing. I've used the following flags in the Makefile:

CXXFLAGS = -fopenmp -O3 -std=c++17 -funroll-loops -march=native -Wno-unused -Wall -I. -g
LDFLAGS = -I /usr/local/include -L /usr/local/lib -lfftw3_omp -lfftw3 -lm

Are there available code samples detailing how to use OpenMP FFTW? What am I doing wrong here? Reference FFTW doc.

fftw c++
1 Answers
  1. -fopenmp is (also?) a linker flag. Move it (or copy, to be safe) from CXXFLAGS to LDFLAGS. I hope this should help.
  2. Less important: -I /usr/include has no effect on the linker; move it to CXXFLAGS (or delete if unnecessary)
user_124512430
Related