Chris@19: Chris@19: Chris@19: One-dimensional distributions - FFTW 3.3.4 Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19: Chris@19:
Chris@19: Chris@19: Chris@19:

Chris@19: Previous: Transposed distributions, Chris@19: Up: MPI Data Distribution Chris@19:


Chris@19:
Chris@19: Chris@19:

6.4.4 One-dimensional distributions

Chris@19: Chris@19:

For one-dimensional distributed DFTs using FFTW, matters are slightly Chris@19: more complicated because the data distribution is more closely tied to Chris@19: how the algorithm works. In particular, you can no longer pass an Chris@19: arbitrary block size and must accept FFTW's default; also, the block Chris@19: sizes may be different for input and output. Also, the data Chris@19: distribution depends on the flags and transform direction, in order Chris@19: for forward and backward transforms to work correctly. Chris@19: Chris@19:

     ptrdiff_t fftw_mpi_local_size_1d(ptrdiff_t n0, MPI_Comm comm,
Chris@19:                      int sign, unsigned flags,
Chris@19:                      ptrdiff_t *local_ni, ptrdiff_t *local_i_start,
Chris@19:                      ptrdiff_t *local_no, ptrdiff_t *local_o_start);
Chris@19: 
Chris@19:

Chris@19: This function computes the data distribution for a 1d transform of Chris@19: size n0 with the given transform sign and flags. Chris@19: Both input and output data use block distributions. The input on the Chris@19: current process will consist of local_ni numbers starting at Chris@19: index local_i_start; e.g. if only a single process is used, Chris@19: then local_ni will be n0 and local_i_start will Chris@19: be 0. Similarly for the output, with local_no numbers Chris@19: starting at index local_o_start. The return value of Chris@19: fftw_mpi_local_size_1d will be the total number of elements to Chris@19: allocate on the current process (which might be slightly larger than Chris@19: the local size due to intermediate steps in the algorithm). Chris@19: Chris@19:

As mentioned above (see Load balancing), the data will be divided Chris@19: equally among the processes if n0 is divisible by the Chris@19: square of the number of processes. In this case, Chris@19: local_ni will equal local_no. Otherwise, they may be Chris@19: different. Chris@19: Chris@19:

For some applications, such as convolutions, the order of the output Chris@19: data is irrelevant. In this case, performance can be improved by Chris@19: specifying that the output data be stored in an FFTW-defined Chris@19: “scrambled” format. (In particular, this is the analogue of Chris@19: transposed output in the multidimensional case: scrambled output saves Chris@19: a communications step.) If you pass FFTW_MPI_SCRAMBLED_OUT in Chris@19: the flags, then the output is stored in this (undocumented) scrambled Chris@19: order. Conversely, to perform the inverse transform of data in Chris@19: scrambled order, pass the FFTW_MPI_SCRAMBLED_IN flag. Chris@19: Chris@19: Chris@19:

In MPI FFTW, only composite sizes n0 can be parallelized; we Chris@19: have not yet implemented a parallel algorithm for large prime sizes. Chris@19: Chris@19: Chris@19: Chris@19: