include/fftw_mpi.h

   1 /*
   2  * $Id$
   3  *
   4  *       This source code is part of
   5  *
   6  *        G   R   O   M   A   C   S
   7  *
   8  * GROningen MAchine for Chemical Simulations
   9  *
  10  *               VERSION 2.0
  11  *
  12  * Copyright (c) 1991-1999
  13  * BIOSON Research Institute, Dept. of Biophysical Chemistry
  14  * University of Groningen, The Netherlands
  15  *
  16  * Please refer to:
  17  * GROMACS: A message-passing parallel molecular dynamics implementation
  18  * H.J.C. Berendsen, D. van der Spoel and R. van Drunen
  19  * Comp. Phys. Comm. 91, 43-56 (1995)
  20  *
  21  * Also check out our WWW page:
  22  * http://md.chem.rug.nl/~gmx
  23  * or e-mail to:
  24  * gromacs@chem.rug.nl
  25  *
  26  * And Hey:
  27  * Good ROcking Metal Altar for Chronical Sinners
  28  */
  29 #ifndef FFTW_MPI_H
  30 #define FFTW_MPI_H
  31
  32 static char *SRCID_fftw_mpi_h = "$Id$";
  33
  34 #include "fftw.h"
  35 #include <mpi.h> /* need access to the MPI type definitions */
  36
  37 #ifdef __cplusplus
  38 extern "C" {
  39 #endif /* __cplusplus */
  40
  41 /***********************************************************************/
  42
  43 typedef fftw_real TRANSPOSE_EL_TYPE;
  44
  45 typedef struct {
  46      int block_num, dest_pe, send_size, recv_size;
  47 } transpose_mpi_exchange;
  48
  49 typedef struct {
  50      MPI_Comm comm;
  51      int n_pes, my_pe;
  52
  53      int nx,ny,local_nx,local_ny;
  54
  55      transpose_mpi_exchange *exchange;
  56      int num_steps, send_block_size, recv_block_size;
  57
  58      MPI_Datatype el_type;
  59
  60      MPI_Request request[2];
  61
  62      int *perm_block_dest;
  63      int num_perm_blocks, perm_block_size;
  64
  65      int all_blocks_equal;
  66      int *send_block_sizes, *send_block_offsets;
  67      int *recv_block_sizes, *recv_block_offsets;
  68
  69      char *move;
  70      int move_size;
  71 } transpose_mpi_plan_struct;
  72
  73 typedef transpose_mpi_plan_struct *transpose_mpi_plan;
  74
  75 extern void transpose_mpi_get_local_size(int n, int my_pe, int n_pes,
  76                                          int *local_n, int *local_start);
  77 extern int transpose_mpi_get_local_storage_size(int nx, int ny,
  78                                                 int my_pe, int n_pes);
  79
  80 extern transpose_mpi_plan transpose_mpi_create_plan(int nx, int ny,
  81                                                     MPI_Comm comm);
  82 extern void transpose_mpi_destroy_plan(transpose_mpi_plan p);
  83
  84 extern void transpose_mpi(transpose_mpi_plan p, int el_size,
  85                           TRANSPOSE_EL_TYPE *local_data,
  86                           TRANSPOSE_EL_TYPE *work);
  87
  88 typedef enum { BEFORE_TRANSPOSE, AFTER_TRANSPOSE } transpose_in_place_which;
  89
  90 typedef enum { TRANSPOSE_SYNC, TRANSPOSE_ASYNC } transpose_sync_type;
  91
  92 extern void transpose_in_place_local(transpose_mpi_plan p,
  93                               int el_size, TRANSPOSE_EL_TYPE *local_data,
  94                               transpose_in_place_which which);
  95
  96 extern TRANSPOSE_EL_TYPE *transpose_allocate_send_buf(transpose_mpi_plan p,
  97                                                       int el_size);
  98 extern void transpose_get_send_block(transpose_mpi_plan p, int step,
  99                                      int *block_y_start, int *block_ny);
 100 extern void transpose_start_exchange_step(transpose_mpi_plan p,
 101                                           int el_size,
 102                                           TRANSPOSE_EL_TYPE *local_data,
 103                                           TRANSPOSE_EL_TYPE *send_buf,
 104                                           int step,
 105                                           transpose_sync_type sync_type);
 106 extern void transpose_finish_exchange_step(transpose_mpi_plan p, int step);
 107
 108 /***********************************************************************/
 109
 110 typedef struct {
 111      fftw_plan p_fft_x;  /* plan for first dimension */
 112      fftwnd_plan p_fft;  /* plan for subsequent dimensions */
 113      transpose_mpi_plan p_transpose, p_transpose_inv;
 114      fftw_complex *work; /* extra workspace, if needed */
 115 } fftwnd_mpi_plan_data;
 116
 117 typedef fftwnd_mpi_plan_data *fftwnd_mpi_plan;
 118
 119 typedef enum {
 120     FFTW_NORMAL_ORDER,
 121     FFTW_TRANSPOSED_ORDER
 122 } fftwnd_mpi_output_order;
 123
 124 extern fftwnd_mpi_plan fftwnd_mpi_create_plan(MPI_Comm comm,
 125                                               int rank, const int *n,
 126                                               fftw_direction dir,
 127                                               int flags);
 128 extern fftwnd_mpi_plan fftw2d_mpi_create_plan(MPI_Comm comm,
 129                                               int nx, int ny,
 130                                           fftw_direction dir, int flags);
 131 extern fftwnd_mpi_plan fftw3d_mpi_create_plan(MPI_Comm comm,
 132                                               int nx, int ny, int nz,
 133                                           fftw_direction dir, int flags);
 134
 135 extern void fftwnd_mpi_destroy_plan(fftwnd_mpi_plan p);
 136
 137 extern void fftwnd_mpi_local_sizes(fftwnd_mpi_plan p,
 138                                    int *local_nx,
 139                                    int *local_x_start,
 140                                    int *local_ny_after_transpose,
 141                                    int *local_y_start_after_transpose,
 142                                    int *total_local_size);
 143
 144 extern void fftwnd_mpi(fftwnd_mpi_plan p,
 145                        int n_fields,
 146                        fftw_complex *local_data, fftw_complex *work,
 147                        fftwnd_mpi_output_order output_order);
 148
 149 extern void fftw_mpi_die(const char *error_string);
 150
 151 /***********************************************************************/
 152
 153 typedef struct fftw_mpi_twiddle_struct {
 154      int rows, rowstart, cols, n;
 155      fftw_complex *W;
 156      int refcount;
 157      struct fftw_mpi_twiddle_struct *next;
 158 } fftw_mpi_twiddle;
 159
 160 typedef struct fftw_mpi_plan_struct {
 161      int n, m, r, local_m, local_m_start, local_r, local_r_start;
 162      fftw_complex *fft_work;
 163      fftw_mpi_twiddle *tw;
 164      transpose_mpi_plan p_transpose, p_transpose_inv;
 165      fftw_plan pm, pr;
 166      int flags;
 167      fftw_direction dir;
 168 } *fftw_mpi_plan;
 169
 170 /* new flags for the MPI planner: */
 171 #define FFTW_SCRAMBLED_INPUT (8192)
 172 #define FFTW_SCRAMBLED_OUTPUT (16384)
 173
 174 extern void fftw_mpi_local_sizes(fftw_mpi_plan p,
 175                                  int *local_n,
 176                                  int *local_start,
 177                                  int *local_n_after_transform,
 178                                  int *local_start_after_transform,
 179                                  int *total_local_size);
 180
 181 extern fftw_mpi_plan fftw_mpi_create_plan(MPI_Comm comm,
 182                                           int n,
 183                                           fftw_direction dir, int flags);
 184
 185 extern void fftw_mpi_destroy_plan(fftw_mpi_plan p);
 186
 187 extern void fftw_mpi(fftw_mpi_plan p, int n_fields,
 188                      fftw_complex *local_data, fftw_complex *work);
 189
 190 extern void fftw_mpi_print_plan(fftw_mpi_plan p);
 191
 192 /***********************************************************************/
 193
 194 #ifdef __cplusplus
 195 } /* extern "C" */
 196 #endif /* __cplusplus */
 197
 198 #endif /* FFTW_MPI_H */