openmp for near2far calculation #868
Conversation
| double freq = freq_min + i * dfreq; | ||
| #ifdef HAVE_OPENMP | ||
| # pragma omp parallel for | ||
| #endif |
There was a problem hiding this comment.
For some reason, I am unable to compile with OpenMP via --with-openmp on my local machine using this #HAVE_OPENMP macro (i.e., the build is successful but the examples/binary_grating_n2f.py test uses just a single thread/process for the get_farfields calculation regardless of the value for OMP_NUM_THREADS). To fix this problem, I had to revert back to using just #pragma omp parallel for (i.e., removing the #ifdef and #endif lines) and CXX="g++ -fopenmp". Perhaps this is related to the openmp -related changes in configure.ac in this PR?
There was a problem hiding this comment.
Whoop, I forgot to add OPENMP_CXXFLAGS to CXXFLAGS … try it again.
|
This is working and giving linear speed up as demonstrated previously in #861 and again as shown below for both serial and MPI (
As part of this PR, it would be good to also provide documentation in two places: (1) the The related tutorial example Near to Far Field Spectra/Diffraction Spectrum of a Finite Binary Grating will be updated in a separate PR with these results. |
* try openmp for near2far loop * add --with-openmp configure flag * use --with and not --enable * whoops need to add OPENMP_CXXFLAGS to CXXFLAGS * docs * typo
This adds an experimental
--with-openmpflag to theconfigurescript that turns on OpenMP compilation, and uses it for thenear2farcalculation for part 1 of #861 (only parallelizing if you are computing for multiple frequencies, however).To use it,
./configure --with-openmpand set theOMP_NUM_THREADSenvironment variable before running Meep (or Python).In the long run, it would be nice to use more comprehensive OpenMP parallelism (#228).
Although this parallelizes over the different frequencies because that was simplest, we could alternatively parallelize at over spatial points in the near or far field.