This package is a set of routines to perform 2d and 3d complex-to-complex Fast Fourier Transforms (FFTs) efficiently on parallel computers.
It was developed at Sandia National Laboratories, a US Department of Energy facility, with funding from the DOE. It is an open-source code, distributed freely under the terms of the GNU Public License (GPL) - see the LICENSE file in the distribution.
The author of this package is Steve Plimpton, who can be contacted at firstname.lastname@example.org This download site at www.cs.sandia.gov/~sjplimp/download.html is where the latest version of this package can be downloaded.
These routines are designed for distributed-memory parallel machines and use MPI as their message-passing protocol. Actually the routines only perform the data movement tasks necessary to compute multi-dimensional FFTs in parallel; the transforms themselves are computed by on-processor 1d FFT routines provided by the machine vendor or the freely available FFTW package. The data remapping routines can also be called directly by the user (independent of the FFTs) to change the layout of an application's 2d or 3d arrays across processors.
The parallel FFT and remapping routines are written in C, are callable from Fortran, are portable to any parallel machine that supports MPI, can be run on any number of processors (including a single processor), work with any size arrays (so long as the native 1d FFT routines support the array dimensions), and allow considerable flexibility in the choice of initial and final data layout across processors.
There are other distributed-memory parallel FFTs available including some provided by vendors and in the FFTW package. I wrote my own because I wanted flexibility in the grid sizes, # of procesors, and in the data layout used as input/output by the FFTs. Also I needed data remapping routines for other parts of my applications; once you have written parallel remapping routines, turning them into parallel FFTs is a straightforward extension.
I use these FFTs in materials modeling applications, specifically for long-range Coulombic solvers in molecular dynamics simulations and for plane-wave electronic structure computations. See my home page for more information on these applications, such as the LAMMPS molecular dynamics code.
These FFT routines perform their data remapping similar to the way an MPI_Alltoall function works: each processor in a group sends a message to every other processor in the group. The group can be all the processors or a subset, depending on the data layout. An alternative method that I've used previously is to do recursive swapping in a hypercube style of communication -- see J. S. Nelson, S. J. Plimpton, M. P. Sears, Phys Rev B, 47, 1765-1774 (1993).
The latter method requires less messages, but a larger volume of data to be shipped. It's also hard to implement on a non-power-of-two # of processors. I believe the all-to-all style works better on current parallel machines with low communication latencies (e.g. the Intel and Cray T3E) and when processor groups can be made small (3d FFTs); I'm not sure which approach is better on high-latency machines such as clusters or for 2d FFTs.
The appropriate routines from this package should be compiled and linked with your application. For a 2d FFT, use fft_2d.c, remap_2d.c, and pack_2d.c and their associated *.h files. The analogous 3d files are used for 3d FFTs. To use only the remapping routines, the fft_* files are not needed.
You must compile the fft routines with a -D compiler flag appropriate to the 1d FFT machine libraries you intend to use. Available switches are FFT_INTEL (i.e. -DFFT_INTEL) for the Intel Paragon and Intel Tflops, FFT_SGI for SGI machines, FFT_DEC for Dec machines, and FFT_T3E for the Cray T3E. If the portable FFTW library is installed on your machine, the FFT_FFTW switch may be used instead. To use only the remapping routines an FFT switch is not needed. On the T3E you also must use the T3E_KLUDGE switch (i.e. -DT3E_KLUDGE) for both the fft and remapping routines. You must also enable the needed compiler and link options appropriate for the 1d FFT libraries you are using and any associated include files they require.
Before giving the details of how to call the 3 functions, I'll discuss data layout across processors and within a processor's memory. Consider 2d and 3d arrays which are operated on by 2d and 3d FFTs and remaps. In your application you may wish, for parallel efficiency, to distribute an array across processors in a variety of ways. The FFT and remap routines allow you to specify how an array is mapped to processors on input to the routine and on output. The only requirement for a mapping is that each processor own a sub-section of the 2d or 3d array. In 2d this is a rectangular-shaped section of the 2d array; in 3d this is a brick-shaped section of the 3d array. Or in array notation, each processor must own a subsection (ilo:ihi,jlo:jhi) of the global 2d array and similarly for 3d arrays.
On each processor, the elements for its subsection must be stored contiguously in memory with a fast-varying index, a mid-varying index (for 3d), and a slow-varying index. This means the routines can be called with either C-style (last index varies fastest) or Fortran-style (first index varies fastest) arrays; the calling routine only specifies the range of the fast, mid, and slow indices. It also means that C-style arrays of pointers to pointers are not allowed; the data must be contiguous in memory. For a good discussion of why this is the best way to store data on current RISC processors, see the FFTW documentation. Basically it's cheaper to do integer arithmetic to compute an array offset than it is to dereference pointers since the latter requires additional memory fetches.
The FFT and remap routines allow the option of permuting the order of the fast, mid, and slow indices on output. In a 2-d FFT for example, a processor can own data in a row-wise ordering on input to the FFT and in a column-wise ordering on output.
Note that there is an implicit assumption that the subsections owned by all the processors do not overlap and that their union exactly tiles the global 2d or 3d array. It is also permissible for a particular processor owns no data on input and/or output, for example if the processor's subsection is input as (ilo:ihi,jlo:jhi) and ilo > ihi.
Here are some examples of the data layouts that the FFT and remap routines will thus allow:
* Each processor initially owns a few rows (or columns) of a 2d or 3d array and the transformed data is returned in the same layout.
* Each processor initally owns a few rows of the array; to save inter-processor communication inside the FFT, it is returned with each processor owning a few columns. Then a convolution can be performed by the application, followed by an inverse FFT that returns the data in the original row-wise layout.
* Each processor initially owns a 2d or 3d subsection of the grid and the transformed data is returned in the same layout. Or it could be returned in a column-wise layout as in the previous convolution example.
What is NOT allowed in a data layout is for a procsesor to own a scattered or random set of rows, columns, or subsections. Such a data distribution might be natural, for example, in a torus-wrap mapping of a matrix to processors. If this is the case in your application, you will need to write your own remapping routine that puts the data in an acceptable layout before calling the FFTs.
While the FFT routines allow for a wide variety of input and output data layouts, they work fastest with layouts directly usable by the parallel FFTs, without pre- or post-remappings being necessary. This is discussed in the optimization section.
To perform an FFT (or remap) there are 3 function calls to understand. Before performing an FFT for the 1st time, you create a "plan" that pre-computes and stores internally all the information needed to do an FFT in parallel (see Acknowledgements). The plan is unique to a particular sized array and an initial and final data layout across processors. You can create as many plans as you need in your application, though one is often sufficient. Then you perform one or more FFTs using a specific plan. Finally when you are finished using a plan for the last time, you destroy it to free up the memory it consumes.
For example, to perform 2d FFTs in C you do somthing like the following. Note that in C, the indices used as calling arguments and to index into the data arrays are C-style, meaning that they run from 0 to N-1. When calling the routines from Fortran, use Fortran-style indices that run from 1 to N. You can see Fortran examples for calling the routines in the test codes provided in the distribution.
#include "fft_2d.h" fft_2d_plan *plan; FFT_DATA *data; int in_ilo,in_ihi,in_jlo,in_jhi; int out_ilo,out_ihi,out_jlo,out_jhi; int bufsize; int me; MPI_Comm_rank(MPI_COMM_WORLD,&me); nx,ny = size of global FFT grid in_ilo,in_ihi,in_jlo,in_jhi = portion of grid to be owned by this proc on input out_ilo,out_ihi,out_jlo,out_jhi = portion of grid to be owned by this proc on output plan = fft_2d_create_plan(MPI_COMM_WORLD,nx,ny, in_ilo,in_ihi,in_jlo,in_jhi, out_ilo,out_ihi,out_jlo,out_jhi, 1,0,&bufsize); if (plan == NULL) printf("ERROR: FFT plan failed on proc %d\n",me); data = ... /* allocate data array and fill with values owned by this proc */ /* forward & reverse FFTs */ fft_2d(data,data,1,plan); /* do this as many */ ... /* times as */ fft_2d(data,data,-1,plan); /* needed */ fft_2d_destroy_plan(plan);
The create_plan routines return NULL if there was an error creating a plan, typically due to a lack of memory.
In MPI parlance, all the FFT and remap calls are collective in the sense they need to be called by all the processors in the communicator in order to complete successfully. Typically this would be MPI_COMM_WORLD (all the processors), but it can be a communicator for a subset of the processors. This allows multiple sets of processors to be doing their own FFTs simultaneously on their own data arrays, if desired.
When calling the FFT routines from C, you include the file "fft_2d.h" or "fft_3d.h" in your application's C file. This defines a complex data type FFT_DATA which is used to store the data you pass into and receive back from the FFT routines. When calling the FFT routines from Fortran (F90 or F77) no header file is needed; you just pass in the usual Fortran complex array. Similarly, when calling the remap routines from your C application, the "remap_2d.h" or "remap_3d.h" header file should be included. The data arrays passed to the remap routines are just C doubles.
struct fft_plan_2d *fft_2d_create_plan( MPI_Comm comm, int nfast, int nslow, int in_ilo, int in_ihi, int in_jlo, int in_jhi, int out_ilo, int out_ihi, int out_jlo, int out_jhi, int scaled, int permute, int *nbuf)
void fft_2d(FFT_DATA *in, FFT_DATA *out, int flag, fft_plan_2d *plan)
void fft_2d_destroy_plan(struct fft_plan_2d *plan)
struct fft_plan_3d *fft_3d_create_plan( MPI_Comm comm, int nfast, int nmid, int nslow, int in_ilo, int in_ihi, int in_jlo, int in_jhi, int in_klo, int in_khi, int out_ilo, int out_ihi, int out_jlo, int out_jhi, int out_klo, int out_khi, int scaled, int permute, int *nbuf)
void fft_3d(FFT_DATA *in, FFT_DATA *out, int flag, fft_plan_3d *plan)
void fft_3d_destroy_plan(struct fft_plan_3d *plan)
struct remap_plan_2d *remap_2d_create_plan( MPI_Comm comm, int in_ilo, int in_ihi, int in_jlo, int in_jhi, int out_ilo, int out_ihi, int out_jlo, int out_jhi, int nqty, int permute, int memory, int precision)
void remap_2d(double *in, double *out, double *buf, struct remap_plan_2d *plan)
void remap_2d_destroy_plan(struct remap_plan_2d *plan)
struct remap_plan_3d *remap_3d_create_plan( MPI_Comm comm, int in_ilo, int in_ihi, int in_jlo, int in_jhi, int in_klo, int in_khi, int out_ilo, int out_ihi, int out_jlo, int out_jhi, int out_klo, int out_khi, int nqty, int permute, int memory, int precision)
void remap_3d(double *in, double *out, double *buf, struct remap_plan_3d *plan)
void remap_3d_destroy_plan(struct remap_plan_3d *plan)
As noted above, you can minimize communication in the FFTs by choosing appropriate input and output data layouts. For 2d and 3d FFTs an optimal input layout is one where each processor owns the entire fast-varying dimension of the data array. For 2d and 3d FFTs an optimal output layout is one where each proc owns the entire slow-varying dimension and permutation is specified as 1 for 2d and as 2 for 3d. Note that these output layouts may not make sense for your application, but you can still reduce communication by choosing the optimal input data layout.
The FFT and remapping routines can be called from Fortran using the wrapper routines in fft_2d_f.c and remap_2d_f.c and similarly for 3d. There are 2 differences in the calling syntax from Fortran. All the indices run from 1 to N in standard Fortran style, instead of from 0 to N-1. Also, the create_plan routines require an extra final argument, a double precision (8 byte) Fortran variable to store the returned plan. Since this variable stores a pointer to the C data structure for the plan, it cannot be examined directly from the Fortran program to check it for a NULL value. Instead the wrappers print an error message if the create_plan routines failed.
The FFT routines can be used in either single or double precision mode be setting #define FFT_PRECISION at the top of fft_2d.h or fft_3d.h. The precision of the remap routines is a calling parameter.
The pack routines have 3 options which can be selected with compiler switches. Setting ARRAY, POINTER, or MEMCPY chooses between the 3 different sets of pack/unpack routines - e.g. -DPOINTER. The ARRAY option is the default. These routines are the kernels for parallel data movement since messages sent to other proecssors have to be packed and unpacked into and out of message buffers. On some machines one version of these tends to be faster than the others; on some machines it makes little difference. You can experiment with which works best by toggling the appropriate flag when compilng the pack_2d.c and pack_3d.c routines.
My thanks to Bruce Hendrickson and Sue Minkoff at Sandia for useful discussions about parallel FFT strategies and to the FFTW authors for the idea of using "plans" as an object-oriented tool for hiding FFT and remap details from the user.