SHAFFT

SHAFFT is a scalable library for high-dimensional complex-to-complex Fast Fourier Transforms (FFTs) in distributed-memory environments. It implements the slab decomposition method introduced by Dalcin, Mortensen, and Keyes (arXiv:1804.09536), using MPI for communication across a Cartesian process topology.

Features

• N-dimensional distributed FFTs with slab decomposition
• 1D distributed FFTs via dedicated FFT1D class
• Flexible process grid topology (1 to N-1 distributed axes)
• Single and double precision complex transforms
• Backend-agnostic design with portable buffer API
• C++, C, and Fortran 2003 interfaces

Supported Backends

Backend	Target	Description
hipFFT	GPU	AMD and NVIDIA GPUs via ROCm/HIP
FFTW	CPU	Multi-threaded CPU execution

Quick Start

Requirements

• CMake >= 3.21
• MPI implementation
• C++17-compatible compiler (GCC >= 10)
• Backend: FFTW3 (CPU) or ROCm/HIP (GPU)

Build

# FFTW backend (CPU)
cmake -B build -S . \
    -DSHAFFT_ENABLE_FFTW=ON \
    -DCMAKE_INSTALL_PREFIX=/opt/shafft
 
cmake --build build --target install

# hipFFT backend (GPU)
cmake -B build -S . \
    -DSHAFFT_ENABLE_HIPFFT=ON \
    -DSHAFFT_GPU_AWARE_MPI=OFF \
    -DCMAKE_PREFIX_PATH=/opt/rocm \
    -DCMAKE_INSTALL_PREFIX=/opt/shafft
 
cmake --build build --target install

Example

#include <shafft/shafft.hpp>
#include <mpi.h>
 
int main(int argc, char** argv) {
    MPI_Init(&argc, &argv);
    
    std::vector<int> commDims = {0, 0, 0};  // auto-select
    std::vector<size_t> dims = {64, 64, 32};
    
    shafft::FFTND plan;
    plan.init(commDims, dims, shafft::FFTType::C2C, MPI_COMM_WORLD);
    plan.plan();
    
    size_t n = plan.allocSize();
    shafft::complexf *data, *work;
    shafft::allocBuffer(n, &data);
    shafft::allocBuffer(n, &work);
    
    plan.setBuffers(data, work);
    plan.execute(shafft::FFTDirection::FORWARD);
    plan.normalize();
    
    plan.execute(shafft::FFTDirection::BACKWARD);
    plan.normalize();
    
    plan.release();
    shafft::freeBuffer(data);
    shafft::freeBuffer(work);
    
    MPI_Finalize();
    return 0;
}

Validation (Optional)

cd build
ctest --output-on-failure

Documentation

Primary documentation:

• Getting Started - installation and build options
• User Guide - API usage for C++, C, and Fortran
• Linking Guide - compile and link instructions
• Backend Reference - backend-specific configuration
• Limitations - current constraints

Build local HTML documentation (optional):

doxygen docs/Doxyfile

Then open docs/html/index.html in your browser.

License

MIT License. See LICENSE for details.

Citation

If you use SHAFFT in your research, please cite the underlying method:

@article{dalcin2019fast,
  title={Fast parallel multidimensional FFT using advanced MPI},
  author={Dalcin, Lisandro and Mortensen, Mikael and Keyes, David E},
  journal={Journal of Parallel and Distributed Computing},
  volume={128},
  pages={137--150},
  year={2019},
  doi={10.1016/j.jpdc.2019.02.006}
}