Chris@19: Chris@19: Chris@19: Basic and advanced distribution interfaces - 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: Next: , Chris@19: Previous: MPI Data Distribution, Chris@19: Up: MPI Data Distribution Chris@19:


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

6.4.1 Basic and advanced distribution interfaces

Chris@19: Chris@19:

As with the planner interface, the ‘fftw_mpi_local_size’ Chris@19: distribution interface is broken into basic and advanced Chris@19: (‘_many’) interfaces, where the latter allows you to specify the Chris@19: block size manually and also to request block sizes when computing Chris@19: multiple transforms simultaneously. These functions are documented Chris@19: more exhaustively by the FFTW MPI Reference, but we summarize the Chris@19: basic ideas here using a couple of two-dimensional examples. Chris@19: Chris@19:

For the 100 × 200 complex-DFT example, above, we would find Chris@19: the distribution by calling the following function in the basic Chris@19: interface: Chris@19: Chris@19:

     ptrdiff_t fftw_mpi_local_size_2d(ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm,
Chris@19:                                       ptrdiff_t *local_n0, ptrdiff_t *local_0_start);
Chris@19: 
Chris@19:

Chris@19: Given the total size of the data to be transformed (here, n0 = Chris@19: 100 and n1 = 200) and an MPI communicator (comm), this Chris@19: function provides three numbers. Chris@19: Chris@19:

First, it describes the shape of the local data: the current process Chris@19: should store a local_n0 by n1 slice of the overall Chris@19: dataset, in row-major order (n1 dimension contiguous), starting Chris@19: at index local_0_start. That is, if the total dataset is Chris@19: viewed as a n0 by n1 matrix, the current process should Chris@19: store the rows local_0_start to Chris@19: local_0_start+local_n0-1. Obviously, if you are running with Chris@19: only a single MPI process, that process will store the entire array: Chris@19: local_0_start will be zero and local_n0 will be Chris@19: n0. See Row-major Format. Chris@19: Chris@19: Chris@19:

Second, the return value is the total number of data elements (e.g., Chris@19: complex numbers for a complex DFT) that should be allocated for the Chris@19: input and output arrays on the current process (ideally with Chris@19: fftw_malloc or an ‘fftw_alloc’ function, to ensure optimal Chris@19: alignment). It might seem that this should always be equal to Chris@19: local_n0 * n1, but this is not the case. FFTW's Chris@19: distributed FFT algorithms require data redistributions at Chris@19: intermediate stages of the transform, and in some circumstances this Chris@19: may require slightly larger local storage. This is discussed in more Chris@19: detail below, under Load balancing. Chris@19: Chris@19: Chris@19:

The advanced-interface ‘local_size’ function for multidimensional Chris@19: transforms returns the same three things (local_n0, Chris@19: local_0_start, and the total number of elements to allocate), Chris@19: but takes more inputs: Chris@19: Chris@19:

     ptrdiff_t fftw_mpi_local_size_many(int rnk, const ptrdiff_t *n,
Chris@19:                                         ptrdiff_t howmany,
Chris@19:                                         ptrdiff_t block0,
Chris@19:                                         MPI_Comm comm,
Chris@19:                                         ptrdiff_t *local_n0,
Chris@19:                                         ptrdiff_t *local_0_start);
Chris@19: 
Chris@19:

Chris@19: The two-dimensional case above corresponds to rnk = 2 and an Chris@19: array n of length 2 with n[0] = n0 and n[1] = n1. Chris@19: This routine is for any rnk > 1; one-dimensional transforms Chris@19: have their own interface because they work slightly differently, as Chris@19: discussed below. Chris@19: Chris@19:

First, the advanced interface allows you to perform multiple Chris@19: transforms at once, of interleaved data, as specified by the Chris@19: howmany parameter. (hoamany is 1 for a single Chris@19: transform.) Chris@19: Chris@19:

Second, here you can specify your desired block size in the n0 Chris@19: dimension, block0. To use FFTW's default block size, pass Chris@19: FFTW_MPI_DEFAULT_BLOCK (0) for block0. Otherwise, on Chris@19: P processes, FFTW will return local_n0 equal to Chris@19: block0 on the first P / block0 processes (rounded down), Chris@19: return local_n0 equal to n0 - block0 * (P / block0) on Chris@19: the next process, and local_n0 equal to zero on any remaining Chris@19: processes. In general, we recommend using the default block size Chris@19: (which corresponds to n0 / P, rounded up). Chris@19: Chris@19: Chris@19:

For example, suppose you have P = 4 processes and n0 = Chris@19: 21. The default will be a block size of 6, which will give Chris@19: local_n0 = 6 on the first three processes and local_n0 = Chris@19: 3 on the last process. Instead, however, you could specify Chris@19: block0 = 5 if you wanted, which would give local_n0 = 5 Chris@19: on processes 0 to 2, local_n0 = 6 on process 3. (This choice, Chris@19: while it may look superficially more “balanced,” has the same Chris@19: critical path as FFTW's default but requires more communications.) Chris@19: Chris@19: Chris@19: