Previous: Transposed distributions, Up: MPI Data Distribution


6.4.4 One-dimensional distributions

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

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

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

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

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

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