Investigate offloading fast fourier transform to gpu.
Initial plan is to investigate:
- 2decomp&FFT:, which may be a full replacement for the fft component, interleaving MPI comms, cuda, and cuda aware mpi.
- a gpu enabled implementation of fftpack - this would likely replace the 1d fft calls, leaving current MPI and pencil code largely intact.
- fallback options are vendor fft libraries such as hipFFT/cuFFT.
Investigate offloading fast fourier transform to gpu.
Initial plan is to investigate: