src/mdlib/gmx_parallel_3dfft.c

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2 *
   3 *
   4 * Gromacs                               Copyright (c) 1991-2005
   5 * David van der Spoel, Erik Lindahl, University of Groningen.
   6 *
   7 * This program is free software; you can redistribute it and/or
   8 * modify it under the terms of the GNU General Public License
   9 * as published by the Free Software Foundation; either version 2
  10 * of the License, or (at your option) any later version.
  11 *
  12 * To help us fund GROMACS development, we humbly ask that you cite
  13 * the research papers on the package. Check out http://www.gromacs.org
  14 *
  15 * And Hey:
  16 * Gnomes, ROck Monsters And Chili Sauce
  17 */
  18 #ifdef HAVE_CONFIG_H
  19 #include <config.h>
  20 #endif
  21
  22 #include <stdlib.h>
  23 #include <string.h>
  24 #include <errno.h>
  25
  26 #ifdef GMX_LIB_MPI
  27 #include <mpi.h>
  28 #endif
  29 #ifdef GMX_THREADS
  30 #include "tmpi.h"
  31 #endif
  32
  33 #include "smalloc.h"
  34 #include "gmx_parallel_3dfft.h"
  35 #include "gmx_fft.h"
  36 #include "gmxcomplex.h"
  37 #include "gmx_fatal.h"
  38
  39 #include "fft5d.h"
  40
  41 struct gmx_parallel_3dfft  {
  42     fft5d_plan p1,p2;
  43 };
  44
  45
  46 static int *copy_int_array(int n,int *src)
  47 {
  48     int *dest,i;
  49
  50     dest = (int*)malloc(n*sizeof(int));
  51     for(i=0; i<n; i++)
  52         dest[i] = src[i];
  53
  54     return dest;
  55 }
  56
  57 static int *make_slab2grid(int nnodes,int ngrid)
  58 {
  59     int *s2g,i;
  60
  61     s2g = (int*)malloc((nnodes+1)*sizeof(int));
  62     for(i=0; i<nnodes+1; i++) {
  63         /* We always round up */
  64         s2g[i] = (i*ngrid + nnodes - 1)/nnodes;
  65     }
  66
  67     return s2g;
  68 }
  69
  70 int
  71 gmx_parallel_3dfft_init   (gmx_parallel_3dfft_t *    pfft_setup,
  72                            ivec                      ndata,
  73                                                    real **                   real_data,
  74                                                    t_complex **              complex_data,
  75                            MPI_Comm                  comm[2],
  76                            int *                     slab2index_major,
  77                            int *                     slab2index_minor,
  78                            gmx_bool                      bReproducible,
  79                            int nthreads)
  80 {
  81     int rN=ndata[2],M=ndata[1],K=ndata[0];
  82     int flags = FFT5D_REALCOMPLEX | FFT5D_ORDER_YZ; /* FFT5D_DEBUG */
  83     MPI_Comm rcomm[]={comm[1],comm[0]};
  84     int Nb,Mb,Kb; /* dimension for backtransform (in starting order) */
  85     t_complex *buf1, *buf2; /*intermediate buffers - used internally.*/
  86
  87     snew(*pfft_setup,1);
  88     if (bReproducible) flags |= FFT5D_NOMEASURE;
  89
  90     if (!(flags&FFT5D_ORDER_YZ)) {
  91         Nb=M;Mb=K;Kb=rN;
  92     } else {
  93         Nb=K;Mb=rN;Kb=M;  /* currently always true because ORDER_YZ always set */
  94     }
  95
  96     (*pfft_setup)->p1 = fft5d_plan_3d(rN,M,K,rcomm, flags, (t_complex**)real_data, complex_data, &buf1, &buf2, nthreads);
  97
  98     (*pfft_setup)->p2 = fft5d_plan_3d(Nb,Mb,Kb,rcomm,
  99                                       (flags|FFT5D_BACKWARD|FFT5D_NOMALLOC)^FFT5D_ORDER_YZ, complex_data, (t_complex**)real_data, &buf1, &buf2, nthreads);
 100
 101     return (*pfft_setup)->p1 != 0 && (*pfft_setup)->p2 !=0;
 102 }
 103
 104
 105 static int
 106 fft5d_limits(fft5d_plan p,
 107              ivec                      local_ndata,
 108              ivec                      local_offset,
 109              ivec                      local_size)
 110 {
 111     int N1,M0,K0,K1,*coor;
 112     fft5d_local_size(p,&N1,&M0,&K0,&K1,&coor);  /* M0=MG/P[0], K1=KG/P[1], NG,MG,KG global sizes */
 113
 114     local_offset[2]=0;
 115     local_offset[1]=p->oM[0];  /*=p->coor[0]*p->MG/p->P[0]; */
 116     local_offset[0]=p->oK[0];  /*=p->coor[1]*p->KG/p->P[1]; */
 117
 118     local_ndata[2]=p->rC[0];
 119     local_ndata[1]=p->pM[0];
 120     local_ndata[0]=p->pK[0];
 121
 122     if ((!(p->flags&FFT5D_BACKWARD)) && (p->flags&FFT5D_REALCOMPLEX)) {
 123         local_size[2]=p->C[0]*2;
 124     } else {
 125         local_size[2]=p->C[0];
 126     }
 127     local_size[1]=p->pM[0];
 128     local_size[0]=p->pK[0];
 129     return 0;
 130 }
 131
 132 int
 133 gmx_parallel_3dfft_real_limits(gmx_parallel_3dfft_t      pfft_setup,
 134                                ivec                      local_ndata,
 135                                ivec                      local_offset,
 136                                ivec                      local_size) {
 137     return fft5d_limits(pfft_setup->p1,local_ndata,local_offset,local_size);
 138 }
 139
 140 static void reorder_ivec_yzx(ivec v)
 141 {
 142     real tmp;
 143
 144     tmp   = v[0];
 145     v[XX] = v[2];
 146     v[ZZ] = v[1];
 147     v[YY] = tmp;
 148 }
 149
 150 int
 151 gmx_parallel_3dfft_complex_limits(gmx_parallel_3dfft_t      pfft_setup,
 152                                   ivec                      complex_order,
 153                                   ivec                      local_ndata,
 154                                   ivec                      local_offset,
 155                                   ivec                      local_size)
 156 {
 157     int ret;
 158
 159     /* For now everything is in-order, but prepare to save communication by avoiding transposes */
 160     complex_order[0] = 0;
 161     complex_order[1] = 1;
 162     complex_order[2] = 2;
 163
 164     ret = fft5d_limits(pfft_setup->p2,local_ndata,local_offset,local_size);
 165
 166     reorder_ivec_yzx(local_ndata);
 167     reorder_ivec_yzx(local_offset);
 168     reorder_ivec_yzx(local_size);
 169
 170     return ret;
 171 }
 172
 173
 174 int
 175 gmx_parallel_3dfft_execute(gmx_parallel_3dfft_t    pfft_setup,
 176                            enum gmx_fft_direction  dir,
 177                            void *                  in_data,
 178                            void *                  out_data,
 179                            int                     thread,
 180                            gmx_wallcycle_t         wcycle) {
 181     if ((!(pfft_setup->p1->flags&FFT5D_REALCOMPLEX)) ^ (dir==GMX_FFT_FORWARD ||dir==GMX_FFT_BACKWARD))
 182     {
 183         gmx_fatal(FARGS,"Invalid transform. Plan and execution don't match regarding reel/complex");
 184     }
 185     if (dir==GMX_FFT_FORWARD || dir==GMX_FFT_REAL_TO_COMPLEX) {
 186         fft5d_execute(pfft_setup->p1,thread,wcycle);
 187     } else {
 188         fft5d_execute(pfft_setup->p2,thread,wcycle);
 189     }
 190     return 0;
 191 }
 192
 193 int
 194 gmx_parallel_3dfft_destroy(gmx_parallel_3dfft_t    pfft_setup) {
 195     fft5d_destroy(pfft_setup->p2);
 196     fft5d_destroy(pfft_setup->p1);
 197     sfree(pfft_setup);
 198     return 0;
 199 }
 200
 201
 202
 203
 204