Why it matters

Parallel FFT enables HPC + scientific computing. Understanding shapes parallel computing.

Advertisement

The architecture

Distribute along one axis.

Perform 1D FFTs locally.

All-to-all transpose.

Repeat.

Parallel FFT flowDistribute rowsone axisLocal 1D FFTsper rowAll-to-all transposeswap axesFFTW3 + cuFFT + heFFTe scale; all-to-all dominates cost at large scale
Parallel FFT.
Advertisement

How it works end to end

Slab decomposition + pencil decomposition.

All-to-all pattern.

cuFFT + FFTW MPI.