src/gromacs/ewald/pme-spread.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "pme-spread.h"
  40
  41 #include "config.h"
  42
  43 #include <assert.h>
  44
  45 #include <algorithm>
  46
  47 #include "gromacs/ewald/pme.h"
  48 #include "gromacs/fft/parallel_3dfft.h"
  49 #include "gromacs/simd/simd.h"
  50 #include "gromacs/utility/basedefinitions.h"
  51 #include "gromacs/utility/exceptions.h"
  52 #include "gromacs/utility/fatalerror.h"
  53 #include "gromacs/utility/smalloc.h"
  54
  55 #include "pme-grid.h"
  56 #include "pme-internal.h"
  57 #include "pme-simd.h"
  58 #include "pme-spline-work.h"
  59
  60 /* TODO consider split of pme-spline from this file */
  61
  62 static void calc_interpolation_idx(struct gmx_pme_t *pme, pme_atomcomm_t *atc,
  63                                    int start, int grid_index, int end, int thread)
  64 {
  65     int             i;
  66     int            *idxptr, tix, tiy, tiz;
  67     real           *xptr, *fptr, tx, ty, tz;
  68     real            rxx, ryx, ryy, rzx, rzy, rzz;
  69     int             nx, ny, nz;
  70     int            *g2tx, *g2ty, *g2tz;
  71     gmx_bool        bThreads;
  72     int            *thread_idx = NULL;
  73     thread_plist_t *tpl        = NULL;
  74     int            *tpl_n      = NULL;
  75     int             thread_i;
  76
  77     nx  = pme->nkx;
  78     ny  = pme->nky;
  79     nz  = pme->nkz;
  80
  81     rxx = pme->recipbox[XX][XX];
  82     ryx = pme->recipbox[YY][XX];
  83     ryy = pme->recipbox[YY][YY];
  84     rzx = pme->recipbox[ZZ][XX];
  85     rzy = pme->recipbox[ZZ][YY];
  86     rzz = pme->recipbox[ZZ][ZZ];
  87
  88     g2tx = pme->pmegrid[grid_index].g2t[XX];
  89     g2ty = pme->pmegrid[grid_index].g2t[YY];
  90     g2tz = pme->pmegrid[grid_index].g2t[ZZ];
  91
  92     bThreads = (atc->nthread > 1);
  93     if (bThreads)
  94     {
  95         thread_idx = atc->thread_idx;
  96
  97         tpl   = &atc->thread_plist[thread];
  98         tpl_n = tpl->n;
  99         for (i = 0; i < atc->nthread; i++)
 100         {
 101             tpl_n[i] = 0;
 102         }
 103     }
 104
 105     const real shift = c_pmeMaxUnitcellShift;
 106
 107     for (i = start; i < end; i++)
 108     {
 109         xptr   = atc->x[i];
 110         idxptr = atc->idx[i];
 111         fptr   = atc->fractx[i];
 112
 113         /* Fractional coordinates along box vectors, add a positive shift to ensure tx/ty/tz are positive for triclinic boxes */
 114         tx = nx * ( xptr[XX] * rxx + xptr[YY] * ryx + xptr[ZZ] * rzx + shift );
 115         ty = ny * (                  xptr[YY] * ryy + xptr[ZZ] * rzy + shift );
 116         tz = nz * (                                   xptr[ZZ] * rzz + shift );
 117
 118         tix = static_cast<int>(tx);
 119         tiy = static_cast<int>(ty);
 120         tiz = static_cast<int>(tz);
 121
 122         /* Because decomposition only occurs in x and y,
 123          * we never have a fraction correction in z.
 124          */
 125         fptr[XX] = tx - tix + pme->fshx[tix];
 126         fptr[YY] = ty - tiy + pme->fshy[tiy];
 127         fptr[ZZ] = tz - tiz;
 128
 129         idxptr[XX] = pme->nnx[tix];
 130         idxptr[YY] = pme->nny[tiy];
 131         idxptr[ZZ] = pme->nnz[tiz];
 132
 133 #ifdef DEBUG
 134         range_check(idxptr[XX], 0, pme->pmegrid_nx);
 135         range_check(idxptr[YY], 0, pme->pmegrid_ny);
 136         range_check(idxptr[ZZ], 0, pme->pmegrid_nz);
 137 #endif
 138
 139         if (bThreads)
 140         {
 141             thread_i      = g2tx[idxptr[XX]] + g2ty[idxptr[YY]] + g2tz[idxptr[ZZ]];
 142             thread_idx[i] = thread_i;
 143             tpl_n[thread_i]++;
 144         }
 145     }
 146
 147     if (bThreads)
 148     {
 149         /* Make a list of particle indices sorted on thread */
 150
 151         /* Get the cumulative count */
 152         for (i = 1; i < atc->nthread; i++)
 153         {
 154             tpl_n[i] += tpl_n[i-1];
 155         }
 156         /* The current implementation distributes particles equally
 157          * over the threads, so we could actually allocate for that
 158          * in pme_realloc_atomcomm_things.
 159          */
 160         if (tpl_n[atc->nthread-1] > tpl->nalloc)
 161         {
 162             tpl->nalloc = over_alloc_large(tpl_n[atc->nthread-1]);
 163             srenew(tpl->i, tpl->nalloc);
 164         }
 165         /* Set tpl_n to the cumulative start */
 166         for (i = atc->nthread-1; i >= 1; i--)
 167         {
 168             tpl_n[i] = tpl_n[i-1];
 169         }
 170         tpl_n[0] = 0;
 171
 172         /* Fill our thread local array with indices sorted on thread */
 173         for (i = start; i < end; i++)
 174         {
 175             tpl->i[tpl_n[atc->thread_idx[i]]++] = i;
 176         }
 177         /* Now tpl_n contains the cummulative count again */
 178     }
 179 }
 180
 181 static void make_thread_local_ind(pme_atomcomm_t *atc,
 182                                   int thread, splinedata_t *spline)
 183 {
 184     int             n, t, i, start, end;
 185     thread_plist_t *tpl;
 186
 187     /* Combine the indices made by each thread into one index */
 188
 189     n     = 0;
 190     start = 0;
 191     for (t = 0; t < atc->nthread; t++)
 192     {
 193         tpl = &atc->thread_plist[t];
 194         /* Copy our part (start - end) from the list of thread t */
 195         if (thread > 0)
 196         {
 197             start = tpl->n[thread-1];
 198         }
 199         end = tpl->n[thread];
 200         for (i = start; i < end; i++)
 201         {
 202             spline->ind[n++] = tpl->i[i];
 203         }
 204     }
 205
 206     spline->n = n;
 207 }
 208
 209 /* Macro to force loop unrolling by fixing order.
 210  * This gives a significant performance gain.
 211  */
 212 #define CALC_SPLINE(order)                     \
 213     {                                              \
 214         for (int j = 0; (j < DIM); j++)            \
 215         {                                          \
 216             real dr, div;                          \
 217             real data[PME_ORDER_MAX];              \
 218                                                    \
 219             dr  = xptr[j];                         \
 220                                                \
 221             /* dr is relative offset from lower cell limit */ \
 222             data[order-1] = 0;                     \
 223             data[1]       = dr;                          \
 224             data[0]       = 1 - dr;                      \
 225                                                \
 226             for (int k = 3; (k < order); k++)      \
 227             {                                      \
 228                 div       = 1.0/(k - 1.0);               \
 229                 data[k-1] = div*dr*data[k-2];      \
 230                 for (int l = 1; (l < (k-1)); l++)  \
 231                 {                                  \
 232                     data[k-l-1] = div*((dr+l)*data[k-l-2]+(k-l-dr)* \
 233                                        data[k-l-1]);                \
 234                 }                                  \
 235                 data[0] = div*(1-dr)*data[0];      \
 236             }                                      \
 237             /* differentiate */                    \
 238             dtheta[j][i*order+0] = -data[0];       \
 239             for (int k = 1; (k < order); k++)      \
 240             {                                      \
 241                 dtheta[j][i*order+k] = data[k-1] - data[k]; \
 242             }                                      \
 243                                                \
 244             div           = 1.0/(order - 1);                 \
 245             data[order-1] = div*dr*data[order-2];  \
 246             for (int l = 1; (l < (order-1)); l++)  \
 247             {                                      \
 248                 data[order-l-1] = div*((dr+l)*data[order-l-2]+    \
 249                                        (order-l-dr)*data[order-l-1]); \
 250             }                                      \
 251             data[0] = div*(1 - dr)*data[0];        \
 252                                                \
 253             for (int k = 0; k < order; k++)        \
 254             {                                      \
 255                 theta[j][i*order+k]  = data[k];    \
 256             }                                      \
 257         }                                          \
 258     }
 259
 260 static void make_bsplines(splinevec theta, splinevec dtheta, int order,
 261                           rvec fractx[], int nr, int ind[], real coefficient[],
 262                           gmx_bool bDoSplines)
 263 {
 264     /* construct splines for local atoms */
 265     int   i, ii;
 266     real *xptr;
 267
 268     for (i = 0; i < nr; i++)
 269     {
 270         /* With free energy we do not use the coefficient check.
 271          * In most cases this will be more efficient than calling make_bsplines
 272          * twice, since usually more than half the particles have non-zero coefficients.
 273          */
 274         ii = ind[i];
 275         if (bDoSplines || coefficient[ii] != 0.0)
 276         {
 277             xptr = fractx[ii];
 278             assert(order >= 4 && order <= PME_ORDER_MAX);
 279             switch (order)
 280             {
 281                 case 4:  CALC_SPLINE(4);     break;
 282                 case 5:  CALC_SPLINE(5);     break;
 283                 default: CALC_SPLINE(order); break;
 284             }
 285         }
 286     }
 287 }
 288
 289 /* This has to be a macro to enable full compiler optimization with xlC (and probably others too) */
 290 #define DO_BSPLINE(order)                            \
 291     for (ithx = 0; (ithx < order); ithx++)                    \
 292     {                                                    \
 293         index_x = (i0+ithx)*pny*pnz;                     \
 294         valx    = coefficient*thx[ithx];                          \
 295                                                      \
 296         for (ithy = 0; (ithy < order); ithy++)                \
 297         {                                                \
 298             valxy    = valx*thy[ithy];                   \
 299             index_xy = index_x+(j0+ithy)*pnz;            \
 300                                                      \
 301             for (ithz = 0; (ithz < order); ithz++)            \
 302             {                                            \
 303                 index_xyz        = index_xy+(k0+ithz);   \
 304                 grid[index_xyz] += valxy*thz[ithz];      \
 305             }                                            \
 306         }                                                \
 307     }
 308
 309
 310 static void spread_coefficients_bsplines_thread(pmegrid_t                         *pmegrid,
 311                                                 pme_atomcomm_t                    *atc,
 312                                                 splinedata_t                      *spline,
 313                                                 struct pme_spline_work gmx_unused *work)
 314 {
 315
 316     /* spread coefficients from home atoms to local grid */
 317     real          *grid;
 318     int            i, nn, n, ithx, ithy, ithz, i0, j0, k0;
 319     int       *    idxptr;
 320     int            order, norder, index_x, index_xy, index_xyz;
 321     real           valx, valxy, coefficient;
 322     real          *thx, *thy, *thz;
 323     int            pnx, pny, pnz, ndatatot;
 324     int            offx, offy, offz;
 325
 326 #if defined PME_SIMD4_SPREAD_GATHER && !defined PME_SIMD4_UNALIGNED
 327     GMX_ALIGNED(real, GMX_SIMD4_WIDTH)  thz_aligned[GMX_SIMD4_WIDTH*2];
 328 #endif
 329
 330     pnx = pmegrid->s[XX];
 331     pny = pmegrid->s[YY];
 332     pnz = pmegrid->s[ZZ];
 333
 334     offx = pmegrid->offset[XX];
 335     offy = pmegrid->offset[YY];
 336     offz = pmegrid->offset[ZZ];
 337
 338     ndatatot = pnx*pny*pnz;
 339     grid     = pmegrid->grid;
 340     for (i = 0; i < ndatatot; i++)
 341     {
 342         grid[i] = 0;
 343     }
 344
 345     order = pmegrid->order;
 346
 347     for (nn = 0; nn < spline->n; nn++)
 348     {
 349         n           = spline->ind[nn];
 350         coefficient = atc->coefficient[n];
 351
 352         if (coefficient != 0)
 353         {
 354             idxptr = atc->idx[n];
 355             norder = nn*order;
 356
 357             i0   = idxptr[XX] - offx;
 358             j0   = idxptr[YY] - offy;
 359             k0   = idxptr[ZZ] - offz;
 360
 361             thx = spline->theta[XX] + norder;
 362             thy = spline->theta[YY] + norder;
 363             thz = spline->theta[ZZ] + norder;
 364
 365             switch (order)
 366             {
 367                 case 4:
 368 #ifdef PME_SIMD4_SPREAD_GATHER
 369 #ifdef PME_SIMD4_UNALIGNED
 370 #define PME_SPREAD_SIMD4_ORDER4
 371 #else
 372 #define PME_SPREAD_SIMD4_ALIGNED
 373 #define PME_ORDER 4
 374 #endif
 375 #include "pme-simd4.h"
 376 #else
 377                     DO_BSPLINE(4);
 378 #endif
 379                     break;
 380                 case 5:
 381 #ifdef PME_SIMD4_SPREAD_GATHER
 382 #define PME_SPREAD_SIMD4_ALIGNED
 383 #define PME_ORDER 5
 384 #include "pme-simd4.h"
 385 #else
 386                     DO_BSPLINE(5);
 387 #endif
 388                     break;
 389                 default:
 390                     DO_BSPLINE(order);
 391                     break;
 392             }
 393         }
 394     }
 395 }
 396
 397 static void copy_local_grid(struct gmx_pme_t *pme, pmegrids_t *pmegrids,
 398                             int grid_index, int thread, real *fftgrid)
 399 {
 400     ivec local_fft_ndata, local_fft_offset, local_fft_size;
 401     int  fft_my, fft_mz;
 402     int  nsy, nsz;
 403     ivec nf;
 404     int  offx, offy, offz, x, y, z, i0, i0t;
 405     int  d;
 406     pmegrid_t *pmegrid;
 407     real *grid_th;
 408
 409     gmx_parallel_3dfft_real_limits(pme->pfft_setup[grid_index],
 410                                    local_fft_ndata,
 411                                    local_fft_offset,
 412                                    local_fft_size);
 413     fft_my = local_fft_size[YY];
 414     fft_mz = local_fft_size[ZZ];
 415
 416     pmegrid = &pmegrids->grid_th[thread];
 417
 418     nsy = pmegrid->s[YY];
 419     nsz = pmegrid->s[ZZ];
 420
 421     for (d = 0; d < DIM; d++)
 422     {
 423         nf[d] = std::min(pmegrid->n[d] - (pmegrid->order - 1),
 424                          local_fft_ndata[d] - pmegrid->offset[d]);
 425     }
 426
 427     offx = pmegrid->offset[XX];
 428     offy = pmegrid->offset[YY];
 429     offz = pmegrid->offset[ZZ];
 430
 431     /* Directly copy the non-overlapping parts of the local grids.
 432      * This also initializes the full grid.
 433      */
 434     grid_th = pmegrid->grid;
 435     for (x = 0; x < nf[XX]; x++)
 436     {
 437         for (y = 0; y < nf[YY]; y++)
 438         {
 439             i0  = ((offx + x)*fft_my + (offy + y))*fft_mz + offz;
 440             i0t = (x*nsy + y)*nsz;
 441             for (z = 0; z < nf[ZZ]; z++)
 442             {
 443                 fftgrid[i0+z] = grid_th[i0t+z];
 444             }
 445         }
 446     }
 447 }
 448
 449 static void
 450 reduce_threadgrid_overlap(struct gmx_pme_t *pme,
 451                           const pmegrids_t *pmegrids, int thread,
 452                           real *fftgrid, real *commbuf_x, real *commbuf_y,
 453                           int grid_index)
 454 {
 455     ivec local_fft_ndata, local_fft_offset, local_fft_size;
 456     int  fft_nx, fft_ny, fft_nz;
 457     int  fft_my, fft_mz;
 458     int  buf_my = -1;
 459     int  nsy, nsz;
 460     ivec localcopy_end, commcopy_end;
 461     int  offx, offy, offz, x, y, z, i0, i0t;
 462     int  sx, sy, sz, fx, fy, fz, tx1, ty1, tz1, ox, oy, oz;
 463     gmx_bool bClearBufX, bClearBufY, bClearBufXY, bClearBuf;
 464     gmx_bool bCommX, bCommY;
 465     int  d;
 466     int  thread_f;
 467     const pmegrid_t *pmegrid, *pmegrid_g, *pmegrid_f;
 468     const real *grid_th;
 469     real *commbuf = NULL;
 470
 471     gmx_parallel_3dfft_real_limits(pme->pfft_setup[grid_index],
 472                                    local_fft_ndata,
 473                                    local_fft_offset,
 474                                    local_fft_size);
 475     fft_nx = local_fft_ndata[XX];
 476     fft_ny = local_fft_ndata[YY];
 477     fft_nz = local_fft_ndata[ZZ];
 478
 479     fft_my = local_fft_size[YY];
 480     fft_mz = local_fft_size[ZZ];
 481
 482     /* This routine is called when all thread have finished spreading.
 483      * Here each thread sums grid contributions calculated by other threads
 484      * to the thread local grid volume.
 485      * To minimize the number of grid copying operations,
 486      * this routines sums immediately from the pmegrid to the fftgrid.
 487      */
 488
 489     /* Determine which part of the full node grid we should operate on,
 490      * this is our thread local part of the full grid.
 491      */
 492     pmegrid = &pmegrids->grid_th[thread];
 493
 494     for (d = 0; d < DIM; d++)
 495     {
 496         /* Determine up to where our thread needs to copy from the
 497          * thread-local charge spreading grid to the rank-local FFT grid.
 498          * This is up to our spreading grid end minus order-1 and
 499          * not beyond the local FFT grid.
 500          */
 501         localcopy_end[d] =
 502             std::min(pmegrid->offset[d] + pmegrid->n[d] - (pmegrid->order - 1),
 503                      local_fft_ndata[d]);
 504
 505         /* Determine up to where our thread needs to copy from the
 506          * thread-local charge spreading grid to the communication buffer.
 507          * Note: only relevant with communication, ignored otherwise.
 508          */
 509         commcopy_end[d]  = localcopy_end[d];
 510         if (pmegrid->ci[d] == pmegrids->nc[d] - 1)
 511         {
 512             /* The last thread should copy up to the last pme grid line.
 513              * When the rank-local FFT grid is narrower than pme-order,
 514              * we need the max below to ensure copying of all data.
 515              */
 516             commcopy_end[d] = std::max(commcopy_end[d], pme->pme_order);
 517         }
 518     }
 519
 520     offx = pmegrid->offset[XX];
 521     offy = pmegrid->offset[YY];
 522     offz = pmegrid->offset[ZZ];
 523
 524
 525     bClearBufX  = TRUE;
 526     bClearBufY  = TRUE;
 527     bClearBufXY = TRUE;
 528
 529     /* Now loop over all the thread data blocks that contribute
 530      * to the grid region we (our thread) are operating on.
 531      */
 532     /* Note that fft_nx/y is equal to the number of grid points
 533      * between the first point of our node grid and the one of the next node.
 534      */
 535     for (sx = 0; sx >= -pmegrids->nthread_comm[XX]; sx--)
 536     {
 537         fx     = pmegrid->ci[XX] + sx;
 538         ox     = 0;
 539         bCommX = FALSE;
 540         if (fx < 0)
 541         {
 542             fx    += pmegrids->nc[XX];
 543             ox    -= fft_nx;
 544             bCommX = (pme->nnodes_major > 1);
 545         }
 546         pmegrid_g = &pmegrids->grid_th[fx*pmegrids->nc[YY]*pmegrids->nc[ZZ]];
 547         ox       += pmegrid_g->offset[XX];
 548         /* Determine the end of our part of the source grid.
 549          * Use our thread local source grid and target grid part
 550          */
 551         tx1 = std::min(ox + pmegrid_g->n[XX],
 552                        !bCommX ? localcopy_end[XX] : commcopy_end[XX]);
 553
 554         for (sy = 0; sy >= -pmegrids->nthread_comm[YY]; sy--)
 555         {
 556             fy     = pmegrid->ci[YY] + sy;
 557             oy     = 0;
 558             bCommY = FALSE;
 559             if (fy < 0)
 560             {
 561                 fy    += pmegrids->nc[YY];
 562                 oy    -= fft_ny;
 563                 bCommY = (pme->nnodes_minor > 1);
 564             }
 565             pmegrid_g = &pmegrids->grid_th[fy*pmegrids->nc[ZZ]];
 566             oy       += pmegrid_g->offset[YY];
 567             /* Determine the end of our part of the source grid.
 568              * Use our thread local source grid and target grid part
 569              */
 570             ty1 = std::min(oy + pmegrid_g->n[YY],
 571                            !bCommY ? localcopy_end[YY] : commcopy_end[YY]);
 572
 573             for (sz = 0; sz >= -pmegrids->nthread_comm[ZZ]; sz--)
 574             {
 575                 fz = pmegrid->ci[ZZ] + sz;
 576                 oz = 0;
 577                 if (fz < 0)
 578                 {
 579                     fz += pmegrids->nc[ZZ];
 580                     oz -= fft_nz;
 581                 }
 582                 pmegrid_g = &pmegrids->grid_th[fz];
 583                 oz       += pmegrid_g->offset[ZZ];
 584                 tz1       = std::min(oz + pmegrid_g->n[ZZ], localcopy_end[ZZ]);
 585
 586                 if (sx == 0 && sy == 0 && sz == 0)
 587                 {
 588                     /* We have already added our local contribution
 589                      * before calling this routine, so skip it here.
 590                      */
 591                     continue;
 592                 }
 593
 594                 thread_f = (fx*pmegrids->nc[YY] + fy)*pmegrids->nc[ZZ] + fz;
 595
 596                 pmegrid_f = &pmegrids->grid_th[thread_f];
 597
 598                 grid_th = pmegrid_f->grid;
 599
 600                 nsy = pmegrid_f->s[YY];
 601                 nsz = pmegrid_f->s[ZZ];
 602
 603 #ifdef DEBUG_PME_REDUCE
 604                 printf("n%d t%d add %d  %2d %2d %2d  %2d %2d %2d  %2d-%2d %2d-%2d, %2d-%2d %2d-%2d, %2d-%2d %2d-%2d\n",
 605                        pme->nodeid, thread, thread_f,
 606                        pme->pmegrid_start_ix,
 607                        pme->pmegrid_start_iy,
 608                        pme->pmegrid_start_iz,
 609                        sx, sy, sz,
 610                        offx-ox, tx1-ox, offx, tx1,
 611                        offy-oy, ty1-oy, offy, ty1,
 612                        offz-oz, tz1-oz, offz, tz1);
 613 #endif
 614
 615                 if (!(bCommX || bCommY))
 616                 {
 617                     /* Copy from the thread local grid to the node grid */
 618                     for (x = offx; x < tx1; x++)
 619                     {
 620                         for (y = offy; y < ty1; y++)
 621                         {
 622                             i0  = (x*fft_my + y)*fft_mz;
 623                             i0t = ((x - ox)*nsy + (y - oy))*nsz - oz;
 624                             for (z = offz; z < tz1; z++)
 625                             {
 626                                 fftgrid[i0+z] += grid_th[i0t+z];
 627                             }
 628                         }
 629                     }
 630                 }
 631                 else
 632                 {
 633                     /* The order of this conditional decides
 634                      * where the corner volume gets stored with x+y decomp.
 635                      */
 636                     if (bCommY)
 637                     {
 638                         commbuf = commbuf_y;
 639                         /* The y-size of the communication buffer is set by
 640                          * the overlap of the grid part of our local slab
 641                          * with the part starting at the next slab.
 642                          */
 643                         buf_my  =
 644                             pme->overlap[1].s2g1[pme->nodeid_minor] -
 645                             pme->overlap[1].s2g0[pme->nodeid_minor+1];
 646                         if (bCommX)
 647                         {
 648                             /* We index commbuf modulo the local grid size */
 649                             commbuf += buf_my*fft_nx*fft_nz;
 650
 651                             bClearBuf   = bClearBufXY;
 652                             bClearBufXY = FALSE;
 653                         }
 654                         else
 655                         {
 656                             bClearBuf  = bClearBufY;
 657                             bClearBufY = FALSE;
 658                         }
 659                     }
 660                     else
 661                     {
 662                         commbuf    = commbuf_x;
 663                         buf_my     = fft_ny;
 664                         bClearBuf  = bClearBufX;
 665                         bClearBufX = FALSE;
 666                     }
 667
 668                     /* Copy to the communication buffer */
 669                     for (x = offx; x < tx1; x++)
 670                     {
 671                         for (y = offy; y < ty1; y++)
 672                         {
 673                             i0  = (x*buf_my + y)*fft_nz;
 674                             i0t = ((x - ox)*nsy + (y - oy))*nsz - oz;
 675
 676                             if (bClearBuf)
 677                             {
 678                                 /* First access of commbuf, initialize it */
 679                                 for (z = offz; z < tz1; z++)
 680                                 {
 681                                     commbuf[i0+z]  = grid_th[i0t+z];
 682                                 }
 683                             }
 684                             else
 685                             {
 686                                 for (z = offz; z < tz1; z++)
 687                                 {
 688                                     commbuf[i0+z] += grid_th[i0t+z];
 689                                 }
 690                             }
 691                         }
 692                     }
 693                 }
 694             }
 695         }
 696     }
 697 }
 698
 699
 700 static void sum_fftgrid_dd(struct gmx_pme_t *pme, real *fftgrid, int grid_index)
 701 {
 702     ivec local_fft_ndata, local_fft_offset, local_fft_size;
 703     pme_overlap_t *overlap;
 704     int  send_index0, send_nindex;
 705     int  recv_nindex;
 706 #if GMX_MPI
 707     MPI_Status stat;
 708 #endif
 709     int  recv_size_y;
 710     int  ipulse, size_yx;
 711     real *sendptr, *recvptr;
 712     int  x, y, z, indg, indb;
 713
 714     /* Note that this routine is only used for forward communication.
 715      * Since the force gathering, unlike the coefficient spreading,
 716      * can be trivially parallelized over the particles,
 717      * the backwards process is much simpler and can use the "old"
 718      * communication setup.
 719      */
 720
 721     gmx_parallel_3dfft_real_limits(pme->pfft_setup[grid_index],
 722                                    local_fft_ndata,
 723                                    local_fft_offset,
 724                                    local_fft_size);
 725
 726     if (pme->nnodes_minor > 1)
 727     {
 728         /* Major dimension */
 729         overlap = &pme->overlap[1];
 730
 731         if (pme->nnodes_major > 1)
 732         {
 733             size_yx = pme->overlap[0].comm_data[0].send_nindex;
 734         }
 735         else
 736         {
 737             size_yx = 0;
 738         }
 739 #if GMX_MPI
 740         int datasize = (local_fft_ndata[XX] + size_yx)*local_fft_ndata[ZZ];
 741
 742         int send_size_y = overlap->send_size;
 743 #endif
 744
 745         for (ipulse = 0; ipulse < overlap->noverlap_nodes; ipulse++)
 746         {
 747             send_index0   =
 748                 overlap->comm_data[ipulse].send_index0 -
 749                 overlap->comm_data[0].send_index0;
 750             send_nindex   = overlap->comm_data[ipulse].send_nindex;
 751             /* We don't use recv_index0, as we always receive starting at 0 */
 752             recv_nindex   = overlap->comm_data[ipulse].recv_nindex;
 753             recv_size_y   = overlap->comm_data[ipulse].recv_size;
 754
 755             sendptr = overlap->sendbuf + send_index0*local_fft_ndata[ZZ];
 756             recvptr = overlap->recvbuf;
 757
 758             if (debug != NULL)
 759             {
 760                 fprintf(debug, "PME fftgrid comm y %2d x %2d x %2d\n",
 761                         local_fft_ndata[XX], send_nindex, local_fft_ndata[ZZ]);
 762             }
 763
 764 #if GMX_MPI
 765             int send_id = overlap->send_id[ipulse];
 766             int recv_id = overlap->recv_id[ipulse];
 767             MPI_Sendrecv(sendptr, send_size_y*datasize, GMX_MPI_REAL,
 768                          send_id, ipulse,
 769                          recvptr, recv_size_y*datasize, GMX_MPI_REAL,
 770                          recv_id, ipulse,
 771                          overlap->mpi_comm, &stat);
 772 #endif
 773
 774             for (x = 0; x < local_fft_ndata[XX]; x++)
 775             {
 776                 for (y = 0; y < recv_nindex; y++)
 777                 {
 778                     indg = (x*local_fft_size[YY] + y)*local_fft_size[ZZ];
 779                     indb = (x*recv_size_y        + y)*local_fft_ndata[ZZ];
 780                     for (z = 0; z < local_fft_ndata[ZZ]; z++)
 781                     {
 782                         fftgrid[indg+z] += recvptr[indb+z];
 783                     }
 784                 }
 785             }
 786
 787             if (pme->nnodes_major > 1)
 788             {
 789                 /* Copy from the received buffer to the send buffer for dim 0 */
 790                 sendptr = pme->overlap[0].sendbuf;
 791                 for (x = 0; x < size_yx; x++)
 792                 {
 793                     for (y = 0; y < recv_nindex; y++)
 794                     {
 795                         indg = (x*local_fft_ndata[YY] + y)*local_fft_ndata[ZZ];
 796                         indb = ((local_fft_ndata[XX] + x)*recv_size_y + y)*local_fft_ndata[ZZ];
 797                         for (z = 0; z < local_fft_ndata[ZZ]; z++)
 798                         {
 799                             sendptr[indg+z] += recvptr[indb+z];
 800                         }
 801                     }
 802                 }
 803             }
 804         }
 805     }
 806
 807     /* We only support a single pulse here.
 808      * This is not a severe limitation, as this code is only used
 809      * with OpenMP and with OpenMP the (PME) domains can be larger.
 810      */
 811     if (pme->nnodes_major > 1)
 812     {
 813         /* Major dimension */
 814         overlap = &pme->overlap[0];
 815
 816         ipulse = 0;
 817
 818         send_nindex   = overlap->comm_data[ipulse].send_nindex;
 819         /* We don't use recv_index0, as we always receive starting at 0 */
 820         recv_nindex   = overlap->comm_data[ipulse].recv_nindex;
 821
 822         recvptr = overlap->recvbuf;
 823
 824         if (debug != NULL)
 825         {
 826             fprintf(debug, "PME fftgrid comm x %2d x %2d x %2d\n",
 827                     send_nindex, local_fft_ndata[YY], local_fft_ndata[ZZ]);
 828         }
 829
 830 #if GMX_MPI
 831         int datasize = local_fft_ndata[YY]*local_fft_ndata[ZZ];
 832         int send_id  = overlap->send_id[ipulse];
 833         int recv_id  = overlap->recv_id[ipulse];
 834         sendptr      = overlap->sendbuf;
 835         MPI_Sendrecv(sendptr, send_nindex*datasize, GMX_MPI_REAL,
 836                      send_id, ipulse,
 837                      recvptr, recv_nindex*datasize, GMX_MPI_REAL,
 838                      recv_id, ipulse,
 839                      overlap->mpi_comm, &stat);
 840 #endif
 841
 842         for (x = 0; x < recv_nindex; x++)
 843         {
 844             for (y = 0; y < local_fft_ndata[YY]; y++)
 845             {
 846                 indg = (x*local_fft_size[YY]  + y)*local_fft_size[ZZ];
 847                 indb = (x*local_fft_ndata[YY] + y)*local_fft_ndata[ZZ];
 848                 for (z = 0; z < local_fft_ndata[ZZ]; z++)
 849                 {
 850                     fftgrid[indg+z] += recvptr[indb+z];
 851                 }
 852             }
 853         }
 854     }
 855 }
 856
 857 void spread_on_grid(struct gmx_pme_t *pme,
 858                     pme_atomcomm_t *atc, pmegrids_t *grids,
 859                     gmx_bool bCalcSplines, gmx_bool bSpread,
 860                     real *fftgrid, gmx_bool bDoSplines, int grid_index)
 861 {
 862     int nthread, thread;
 863 #ifdef PME_TIME_THREADS
 864     gmx_cycles_t c1, c2, c3, ct1a, ct1b, ct1c;
 865     static double cs1     = 0, cs2 = 0, cs3 = 0;
 866     static double cs1a[6] = {0, 0, 0, 0, 0, 0};
 867     static int cnt        = 0;
 868 #endif
 869
 870     nthread = pme->nthread;
 871     assert(nthread > 0);
 872
 873 #ifdef PME_TIME_THREADS
 874     c1 = omp_cyc_start();
 875 #endif
 876     if (bCalcSplines)
 877     {
 878 #pragma omp parallel for num_threads(nthread) schedule(static)
 879         for (thread = 0; thread < nthread; thread++)
 880         {
 881             try
 882             {
 883                 int start, end;
 884
 885                 start = atc->n* thread   /nthread;
 886                 end   = atc->n*(thread+1)/nthread;
 887
 888                 /* Compute fftgrid index for all atoms,
 889                  * with help of some extra variables.
 890                  */
 891                 calc_interpolation_idx(pme, atc, start, grid_index, end, thread);
 892             }
 893             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 894         }
 895     }
 896 #ifdef PME_TIME_THREADS
 897     c1   = omp_cyc_end(c1);
 898     cs1 += (double)c1;
 899 #endif
 900
 901 #ifdef PME_TIME_THREADS
 902     c2 = omp_cyc_start();
 903 #endif
 904 #pragma omp parallel for num_threads(nthread) schedule(static)
 905     for (thread = 0; thread < nthread; thread++)
 906     {
 907         try
 908         {
 909             splinedata_t *spline;
 910             pmegrid_t *grid = NULL;
 911
 912             /* make local bsplines  */
 913             if (grids == NULL || !pme->bUseThreads)
 914             {
 915                 spline = &atc->spline[0];
 916
 917                 spline->n = atc->n;
 918
 919                 if (bSpread)
 920                 {
 921                     grid = &grids->grid;
 922                 }
 923             }
 924             else
 925             {
 926                 spline = &atc->spline[thread];
 927
 928                 if (grids->nthread == 1)
 929                 {
 930                     /* One thread, we operate on all coefficients */
 931                     spline->n = atc->n;
 932                 }
 933                 else
 934                 {
 935                     /* Get the indices our thread should operate on */
 936                     make_thread_local_ind(atc, thread, spline);
 937                 }
 938
 939                 grid = &grids->grid_th[thread];
 940             }
 941
 942             if (bCalcSplines)
 943             {
 944                 make_bsplines(spline->theta, spline->dtheta, pme->pme_order,
 945                               atc->fractx, spline->n, spline->ind, atc->coefficient, bDoSplines);
 946             }
 947
 948             if (bSpread)
 949             {
 950                 /* put local atoms on grid. */
 951 #ifdef PME_TIME_SPREAD
 952                 ct1a = omp_cyc_start();
 953 #endif
 954                 spread_coefficients_bsplines_thread(grid, atc, spline, pme->spline_work);
 955
 956                 if (pme->bUseThreads)
 957                 {
 958                     copy_local_grid(pme, grids, grid_index, thread, fftgrid);
 959                 }
 960 #ifdef PME_TIME_SPREAD
 961                 ct1a          = omp_cyc_end(ct1a);
 962                 cs1a[thread] += (double)ct1a;
 963 #endif
 964             }
 965         }
 966         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 967     }
 968 #ifdef PME_TIME_THREADS
 969     c2   = omp_cyc_end(c2);
 970     cs2 += (double)c2;
 971 #endif
 972
 973     if (bSpread && pme->bUseThreads)
 974     {
 975 #ifdef PME_TIME_THREADS
 976         c3 = omp_cyc_start();
 977 #endif
 978 #pragma omp parallel for num_threads(grids->nthread) schedule(static)
 979         for (thread = 0; thread < grids->nthread; thread++)
 980         {
 981             try
 982             {
 983                 reduce_threadgrid_overlap(pme, grids, thread,
 984                                           fftgrid,
 985                                           pme->overlap[0].sendbuf,
 986                                           pme->overlap[1].sendbuf,
 987                                           grid_index);
 988             }
 989             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 990         }
 991 #ifdef PME_TIME_THREADS
 992         c3   = omp_cyc_end(c3);
 993         cs3 += (double)c3;
 994 #endif
 995
 996         if (pme->nnodes > 1)
 997         {
 998             /* Communicate the overlapping part of the fftgrid.
 999              * For this communication call we need to check pme->bUseThreads
1000              * to have all ranks communicate here, regardless of pme->nthread.
1001              */
1002             sum_fftgrid_dd(pme, fftgrid, grid_index);
1003         }
1004     }
1005
1006 #ifdef PME_TIME_THREADS
1007     cnt++;
1008     if (cnt % 20 == 0)
1009     {
1010         printf("idx %.2f spread %.2f red %.2f",
1011                cs1*1e-9, cs2*1e-9, cs3*1e-9);
1012 #ifdef PME_TIME_SPREAD
1013         for (thread = 0; thread < nthread; thread++)
1014         {
1015             printf(" %.2f", cs1a[thread]*1e-9);
1016         }
1017 #endif
1018         printf("\n");
1019     }
1020 #endif
1021 }