HPX-FFT: Multidimensional Distributed FFTs leveraging Asynchronous Tasks and Communication Abstractions

HPX-FFT is an open-source library for FFT computation. Leveraging the asynchronous many-task runtime HPX, we aim to combine the performance of asynchronous parallelism in C++ and a powerfull communication abstraction based on an Active Global Adress Space with support for industry standard FFT backends such as FFTW3.

Dependencies

HPX-FFT depends on HPX for asynchronous task-based parallelization. Further one-dimensional FFTs are computed via an FFTW backend.

Install dependencies

All dependencies can be installed using Spack. For HPX we recommend: spack install hpx@1.10.0 networking={tcp/mpi/lci} max_cpu_count=256 For FFTW we recommend spack install fftw@3.3.10 +mpi +openmp

Note that the LCI variant of HPX currently requires to add the Octo-Tiger repo to spack. Alternatively, both librarires can be built from source.

How to Compile HPX-FFT

We provide a simple compilation script for HPX-FFT. The script assumes an gcc compiler version 14.2.0. However, it can be easily tailored towards specific needs.

How to Compile Examples

HPX-FFT comes with several examples for HPX-FFT itself as well as for a FFTW reference. The examples can be compiled with the respective compilation scripts.

How to Run Examples

HPX-FFT

1=Base size of the problem
2=FFTW Planning flag (estimate/measure)
3=Print info into runtime file (true/false)
4=Number of HPX threads
./hpxfft_shared_loop --nx=$1 --ny=$1 --plan=$2 --header=$3 --hpx:threads=$4

1=Base size of the problem
2=FFTW Planning flag (estimate/measure)
3=Print info into runtime file (true/false)
4=Collective operation (scatter/all_to_all)
5=Number of HPX threads
./hpxfft_distributed_loop --nx=$1 --ny=$1 --plan=$2 --header=$3 --run=$4 --hpx:threads=$5

FFTW

//        threads N_X N_Y  plan     header
./fftw_omp   1      8  14 estimate    0

//   nodes ranks    prog    threads N_X N_Y    plan  header
srun -N 2  -n 4 fftw_mpi_omp   4     8   14  estimate  0
    mpirun -n 4 fftw_mpi_omp   4     8   14  estimate  0

How to Run Benchmarks (wip)

Shared

All:

• ./shared_benchmark.sh estimate/measure partition_name

Executables only:

• fft_hpx_loop_shared/fft_hpx_task_sync_shared/fft_hpx_task_shared/fft_hpx_task_naive_shared /fft_hpx_task_agas_shared/fft_hpx_loop/fft_hpx_task_agas: sbatch -p $PARTITION -N 1 -n 1 -c $THREADS run_hpx_shared.sh executable_name estimate/measure $THREAD_POW

• fftw_hpx/fftw_mpi_threads/fftw_mpi_omp: sbatch -p $PARTITION -N 1 -n 1 -c $THREADS run_fftw_shared.sh executable_name estimate/measure $THREAD_POW

Distributed

All:

• ./distributed_benchmark.sh estimate/measure partition_name

Executables only:

• fftw_hpx_loop/fftw_hpx_tasg_agas: sbatch -p $PARTITION -N $NODES -n $NODES -c $THREADS run_hpx_dist.sh executable_name estimate/measure $NODES

• fftw_mpi_threads/fftw_mpi_omp: sbatch -p $PARTITION -N $NODES -n $NODES -c $THREADS run_fftw_dist.sh executable_name estimate/measure $NODES $THREADS

The Team

The HPX-FFT library is developed by the Scientific Computing department at IPVS at the University of Stuttgart. The project is a joined effort of multiple graduate, and PhD students under the supervision of Prof. Dr. Dirk Pflüger. We specifically thank the follow contributors:

• Alexander Strack: Maintainer and initial framework.

How To Cite

@InProceedings{GPRat2025,
  author={Strack, Alexander and Taylor, Christopher and Diehl, Patrick and Pfl{\"u}ger, Dirk},
  title={{Experiences Porting Distributed Applications to Asynchronous Tasks: A Multidimensional FFT Case-study}},
  booktitle={Asynchronous Many-Task Systems and Applications},
  year={2024},
  publisher={Springer Nature}
}