1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
11 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
12 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
13 * Copyright (c) 2001-2004, The GROMACS development team,
14 * check out http://www.gromacs.org for more information.
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version 2
19 * of the License, or (at your option) any later version.
21 * If you want to redistribute modifications, please consider that
22 * scientific software is very special. Version control is crucial -
23 * bugs must be traceable. We will be happy to consider code for
24 * inclusion in the official distribution, but derived work must not
25 * be called official GROMACS. Details are found in the README & COPYING
26 * files - if they are missing, get the official version at www.gromacs.org.
28 * To help us fund GROMACS development, we humbly ask that you cite
29 * the papers on the package - you can find them in the top README file.
31 * For more info, check our website at http://www.gromacs.org
34 * GROwing Monsters And Cloning Shrimps
36 /* IMPORTANT FOR DEVELOPERS:
38 * Triclinic pme stuff isn't entirely trivial, and we've experienced
39 * some bugs during development (many of them due to me). To avoid
40 * this in the future, please check the following things if you make
41 * changes in this file:
43 * 1. You should obtain identical (at least to the PME precision)
44 * energies, forces, and virial for
45 * a rectangular box and a triclinic one where the z (or y) axis is
46 * tilted a whole box side. For instance you could use these boxes:
48 * rectangular triclinic
53 * 2. You should check the energy conservation in a triclinic box.
55 * It might seem an overkill, but better safe than sorry.
63 #include "gromacs/fft/parallel_3dfft.h"
64 #include "gromacs/utility/gmxmpi.h"
73 #include "gmxcomplex.h"
77 #include "gmx_fatal.h"
82 #include "gmx_wallcycle.h"
84 #include "gmx_cyclecounter.h"
89 /* Include the SIMD macro file and then check for support */
90 #include "gmx_simd_macros.h"
91 #if defined GMX_HAVE_SIMD_MACROS && defined GMX_SIMD_HAVE_EXP
92 /* Turn on arbitrary width SIMD intrinsics for PME solve */
96 /* Include the 4-wide SIMD macro file */
97 #include "gmx_simd4_macros.h"
98 /* Check if we have 4-wide SIMD macro support */
99 #ifdef GMX_HAVE_SIMD4_MACROS
100 /* Do PME spread and gather with 4-wide SIMD.
101 * NOTE: SIMD is only used with PME order 4 and 5 (which are the most common).
103 #define PME_SIMD4_SPREAD_GATHER
105 #ifdef GMX_SIMD4_HAVE_UNALIGNED
106 /* With PME-order=4 on x86, unaligned load+store is slightly faster
107 * than doubling all SIMD operations when using aligned load+store.
109 #define PME_SIMD4_UNALIGNED
114 /* #define PRT_FORCE */
115 /* conditions for on the fly time-measurement */
116 /* #define TAKETIME (step > 1 && timesteps < 10) */
117 #define TAKETIME FALSE
119 /* #define PME_TIME_THREADS */
122 #define mpi_type MPI_DOUBLE
124 #define mpi_type MPI_FLOAT
127 #ifdef PME_SIMD4_SPREAD_GATHER
128 #define SIMD4_ALIGNMENT (GMX_SIMD4_WIDTH*sizeof(real))
130 /* We can use any alignment, apart from 0, so we use 4 reals */
131 #define SIMD4_ALIGNMENT (4*sizeof(real))
134 /* GMX_CACHE_SEP should be a multiple of the SIMD and SIMD4 register size
135 * to preserve alignment.
137 #define GMX_CACHE_SEP 64
139 /* We only define a maximum to be able to use local arrays without allocation.
140 * An order larger than 12 should never be needed, even for test cases.
141 * If needed it can be changed here.
143 #define PME_ORDER_MAX 12
145 /* Internal datastructures */
151 int recv_size; /* Receive buffer width, used with OpenMP */
162 int *send_id, *recv_id;
163 int send_size; /* Send buffer width, used with OpenMP */
164 pme_grid_comm_t *comm_data;
170 int *n; /* Cumulative counts of the number of particles per thread */
171 int nalloc; /* Allocation size of i */
172 int *i; /* Particle indices ordered on thread index (n) */
186 int dimind; /* The index of the dimension, 0=x, 1=y */
193 int *node_dest; /* The nodes to send x and q to with DD */
194 int *node_src; /* The nodes to receive x and q from with DD */
195 int *buf_index; /* Index for commnode into the buffers */
202 int *count; /* The number of atoms to send to each node */
204 int *rcount; /* The number of atoms to receive */
211 gmx_bool bSpread; /* These coordinates are used for spreading */
214 rvec *fractx; /* Fractional coordinate relative to the
215 * lower cell boundary
218 int *thread_idx; /* Which thread should spread which charge */
219 thread_plist_t *thread_plist;
220 splinedata_t *spline;
227 ivec ci; /* The spatial location of this grid */
228 ivec n; /* The used size of *grid, including order-1 */
229 ivec offset; /* The grid offset from the full node grid */
230 int order; /* PME spreading order */
231 ivec s; /* The allocated size of *grid, s >= n */
232 real *grid; /* The grid local thread, size n */
236 pmegrid_t grid; /* The full node grid (non thread-local) */
237 int nthread; /* The number of threads operating on this grid */
238 ivec nc; /* The local spatial decomposition over the threads */
239 pmegrid_t *grid_th; /* Array of grids for each thread */
240 real *grid_all; /* Allocated array for the grids in *grid_th */
241 int **g2t; /* The grid to thread index */
242 ivec nthread_comm; /* The number of threads to communicate with */
247 #ifdef PME_SIMD4_SPREAD_GATHER
248 /* Masks for 4-wide SIMD aligned spreading and gathering */
249 gmx_simd4_pb mask_S0[6], mask_S1[6];
251 int dummy; /* C89 requires that struct has at least one member */
256 /* work data for solve_pme */
272 typedef struct gmx_pme {
273 int ndecompdim; /* The number of decomposition dimensions */
274 int nodeid; /* Our nodeid in mpi->mpi_comm */
277 int nnodes; /* The number of nodes doing PME */
282 MPI_Comm mpi_comm_d[2]; /* Indexed on dimension, 0=x, 1=y */
284 MPI_Datatype rvec_mpi; /* the pme vector's MPI type */
287 gmx_bool bUseThreads; /* Does any of the PME ranks have nthread>1 ? */
288 int nthread; /* The number of threads doing PME on our rank */
290 gmx_bool bPPnode; /* Node also does particle-particle forces */
291 gmx_bool bFEP; /* Compute Free energy contribution */
292 int nkx, nky, nkz; /* Grid dimensions */
293 gmx_bool bP3M; /* Do P3M: optimize the influence function */
297 pmegrids_t pmegridA; /* Grids on which we do spreading/interpolation, includes overlap */
299 /* The PME charge spreading grid sizes/strides, includes pme_order-1 */
300 int pmegrid_nx, pmegrid_ny, pmegrid_nz;
301 /* pmegrid_nz might be larger than strictly necessary to ensure
302 * memory alignment, pmegrid_nz_base gives the real base size.
305 /* The local PME grid starting indices */
306 int pmegrid_start_ix, pmegrid_start_iy, pmegrid_start_iz;
308 /* Work data for spreading and gathering */
309 pme_spline_work_t *spline_work;
311 real *fftgridA; /* Grids for FFT. With 1D FFT decomposition this can be a pointer */
312 real *fftgridB; /* inside the interpolation grid, but separate for 2D PME decomp. */
313 int fftgrid_nx, fftgrid_ny, fftgrid_nz;
315 t_complex *cfftgridA; /* Grids for complex FFT data */
316 t_complex *cfftgridB;
317 int cfftgrid_nx, cfftgrid_ny, cfftgrid_nz;
319 gmx_parallel_3dfft_t pfft_setupA;
320 gmx_parallel_3dfft_t pfft_setupB;
322 int *nnx, *nny, *nnz;
323 real *fshx, *fshy, *fshz;
325 pme_atomcomm_t atc[2]; /* Indexed on decomposition index */
329 pme_overlap_t overlap[2]; /* Indexed on dimension, 0=x, 1=y */
331 pme_atomcomm_t atc_energy; /* Only for gmx_pme_calc_energy */
333 rvec *bufv; /* Communication buffer */
334 real *bufr; /* Communication buffer */
335 int buf_nalloc; /* The communication buffer size */
337 /* thread local work data for solve_pme */
340 /* Work data for PME_redist */
341 gmx_bool redist_init;
349 int redist_buf_nalloc;
351 /* Work data for sum_qgrid */
352 real * sum_qgrid_tmp;
353 real * sum_qgrid_dd_tmp;
357 static void calc_interpolation_idx(gmx_pme_t pme, pme_atomcomm_t *atc,
358 int start, int end, int thread)
361 int *idxptr, tix, tiy, tiz;
362 real *xptr, *fptr, tx, ty, tz;
363 real rxx, ryx, ryy, rzx, rzy, rzz;
365 int start_ix, start_iy, start_iz;
366 int *g2tx, *g2ty, *g2tz;
368 int *thread_idx = NULL;
369 thread_plist_t *tpl = NULL;
377 start_ix = pme->pmegrid_start_ix;
378 start_iy = pme->pmegrid_start_iy;
379 start_iz = pme->pmegrid_start_iz;
381 rxx = pme->recipbox[XX][XX];
382 ryx = pme->recipbox[YY][XX];
383 ryy = pme->recipbox[YY][YY];
384 rzx = pme->recipbox[ZZ][XX];
385 rzy = pme->recipbox[ZZ][YY];
386 rzz = pme->recipbox[ZZ][ZZ];
388 g2tx = pme->pmegridA.g2t[XX];
389 g2ty = pme->pmegridA.g2t[YY];
390 g2tz = pme->pmegridA.g2t[ZZ];
392 bThreads = (atc->nthread > 1);
395 thread_idx = atc->thread_idx;
397 tpl = &atc->thread_plist[thread];
399 for (i = 0; i < atc->nthread; i++)
405 for (i = start; i < end; i++)
408 idxptr = atc->idx[i];
409 fptr = atc->fractx[i];
411 /* Fractional coordinates along box vectors, add 2.0 to make 100% sure we are positive for triclinic boxes */
412 tx = nx * ( xptr[XX] * rxx + xptr[YY] * ryx + xptr[ZZ] * rzx + 2.0 );
413 ty = ny * ( xptr[YY] * ryy + xptr[ZZ] * rzy + 2.0 );
414 tz = nz * ( xptr[ZZ] * rzz + 2.0 );
420 /* Because decomposition only occurs in x and y,
421 * we never have a fraction correction in z.
423 fptr[XX] = tx - tix + pme->fshx[tix];
424 fptr[YY] = ty - tiy + pme->fshy[tiy];
427 idxptr[XX] = pme->nnx[tix];
428 idxptr[YY] = pme->nny[tiy];
429 idxptr[ZZ] = pme->nnz[tiz];
432 range_check(idxptr[XX], 0, pme->pmegrid_nx);
433 range_check(idxptr[YY], 0, pme->pmegrid_ny);
434 range_check(idxptr[ZZ], 0, pme->pmegrid_nz);
439 thread_i = g2tx[idxptr[XX]] + g2ty[idxptr[YY]] + g2tz[idxptr[ZZ]];
440 thread_idx[i] = thread_i;
447 /* Make a list of particle indices sorted on thread */
449 /* Get the cumulative count */
450 for (i = 1; i < atc->nthread; i++)
452 tpl_n[i] += tpl_n[i-1];
454 /* The current implementation distributes particles equally
455 * over the threads, so we could actually allocate for that
456 * in pme_realloc_atomcomm_things.
458 if (tpl_n[atc->nthread-1] > tpl->nalloc)
460 tpl->nalloc = over_alloc_large(tpl_n[atc->nthread-1]);
461 srenew(tpl->i, tpl->nalloc);
463 /* Set tpl_n to the cumulative start */
464 for (i = atc->nthread-1; i >= 1; i--)
466 tpl_n[i] = tpl_n[i-1];
470 /* Fill our thread local array with indices sorted on thread */
471 for (i = start; i < end; i++)
473 tpl->i[tpl_n[atc->thread_idx[i]]++] = i;
475 /* Now tpl_n contains the cummulative count again */
479 static void make_thread_local_ind(pme_atomcomm_t *atc,
480 int thread, splinedata_t *spline)
482 int n, t, i, start, end;
485 /* Combine the indices made by each thread into one index */
489 for (t = 0; t < atc->nthread; t++)
491 tpl = &atc->thread_plist[t];
492 /* Copy our part (start - end) from the list of thread t */
495 start = tpl->n[thread-1];
497 end = tpl->n[thread];
498 for (i = start; i < end; i++)
500 spline->ind[n++] = tpl->i[i];
508 static void pme_calc_pidx(int start, int end,
509 matrix recipbox, rvec x[],
510 pme_atomcomm_t *atc, int *count)
515 real rxx, ryx, rzx, ryy, rzy;
518 /* Calculate PME task index (pidx) for each grid index.
519 * Here we always assign equally sized slabs to each node
520 * for load balancing reasons (the PME grid spacing is not used).
526 /* Reset the count */
527 for (i = 0; i < nslab; i++)
532 if (atc->dimind == 0)
534 rxx = recipbox[XX][XX];
535 ryx = recipbox[YY][XX];
536 rzx = recipbox[ZZ][XX];
537 /* Calculate the node index in x-dimension */
538 for (i = start; i < end; i++)
541 /* Fractional coordinates along box vectors */
542 s = nslab*(xptr[XX]*rxx + xptr[YY]*ryx + xptr[ZZ]*rzx);
543 si = (int)(s + 2*nslab) % nslab;
550 ryy = recipbox[YY][YY];
551 rzy = recipbox[ZZ][YY];
552 /* Calculate the node index in y-dimension */
553 for (i = start; i < end; i++)
556 /* Fractional coordinates along box vectors */
557 s = nslab*(xptr[YY]*ryy + xptr[ZZ]*rzy);
558 si = (int)(s + 2*nslab) % nslab;
565 static void pme_calc_pidx_wrapper(int natoms, matrix recipbox, rvec x[],
568 int nthread, thread, slab;
570 nthread = atc->nthread;
572 #pragma omp parallel for num_threads(nthread) schedule(static)
573 for (thread = 0; thread < nthread; thread++)
575 pme_calc_pidx(natoms* thread /nthread,
576 natoms*(thread+1)/nthread,
577 recipbox, x, atc, atc->count_thread[thread]);
579 /* Non-parallel reduction, since nslab is small */
581 for (thread = 1; thread < nthread; thread++)
583 for (slab = 0; slab < atc->nslab; slab++)
585 atc->count_thread[0][slab] += atc->count_thread[thread][slab];
590 static void realloc_splinevec(splinevec th, real **ptr_z, int nalloc)
592 const int padding = 4;
595 srenew(th[XX], nalloc);
596 srenew(th[YY], nalloc);
597 /* In z we add padding, this is only required for the aligned SIMD code */
598 sfree_aligned(*ptr_z);
599 snew_aligned(*ptr_z, nalloc+2*padding, SIMD4_ALIGNMENT);
600 th[ZZ] = *ptr_z + padding;
602 for (i = 0; i < padding; i++)
605 (*ptr_z)[padding+nalloc+i] = 0;
609 static void pme_realloc_splinedata(splinedata_t *spline, pme_atomcomm_t *atc)
613 srenew(spline->ind, atc->nalloc);
614 /* Initialize the index to identity so it works without threads */
615 for (i = 0; i < atc->nalloc; i++)
620 realloc_splinevec(spline->theta, &spline->ptr_theta_z,
621 atc->pme_order*atc->nalloc);
622 realloc_splinevec(spline->dtheta, &spline->ptr_dtheta_z,
623 atc->pme_order*atc->nalloc);
626 static void pme_realloc_atomcomm_things(pme_atomcomm_t *atc)
628 int nalloc_old, i, j, nalloc_tpl;
630 /* We have to avoid a NULL pointer for atc->x to avoid
631 * possible fatal errors in MPI routines.
633 if (atc->n > atc->nalloc || atc->nalloc == 0)
635 nalloc_old = atc->nalloc;
636 atc->nalloc = over_alloc_dd(max(atc->n, 1));
640 srenew(atc->x, atc->nalloc);
641 srenew(atc->q, atc->nalloc);
642 srenew(atc->f, atc->nalloc);
643 for (i = nalloc_old; i < atc->nalloc; i++)
645 clear_rvec(atc->f[i]);
650 srenew(atc->fractx, atc->nalloc);
651 srenew(atc->idx, atc->nalloc);
653 if (atc->nthread > 1)
655 srenew(atc->thread_idx, atc->nalloc);
658 for (i = 0; i < atc->nthread; i++)
660 pme_realloc_splinedata(&atc->spline[i], atc);
666 static void pmeredist_pd(gmx_pme_t pme, gmx_bool forw,
667 int n, gmx_bool bXF, rvec *x_f, real *charge,
669 /* Redistribute particle data for PME calculation */
670 /* domain decomposition by x coordinate */
675 if (FALSE == pme->redist_init)
677 snew(pme->scounts, atc->nslab);
678 snew(pme->rcounts, atc->nslab);
679 snew(pme->sdispls, atc->nslab);
680 snew(pme->rdispls, atc->nslab);
681 snew(pme->sidx, atc->nslab);
682 pme->redist_init = TRUE;
684 if (n > pme->redist_buf_nalloc)
686 pme->redist_buf_nalloc = over_alloc_dd(n);
687 srenew(pme->redist_buf, pme->redist_buf_nalloc*DIM);
695 /* forward, redistribution from pp to pme */
697 /* Calculate send counts and exchange them with other nodes */
698 for (i = 0; (i < atc->nslab); i++)
702 for (i = 0; (i < n); i++)
704 pme->scounts[pme->idxa[i]]++;
706 MPI_Alltoall( pme->scounts, 1, MPI_INT, pme->rcounts, 1, MPI_INT, atc->mpi_comm);
708 /* Calculate send and receive displacements and index into send
713 for (i = 1; i < atc->nslab; i++)
715 pme->sdispls[i] = pme->sdispls[i-1]+pme->scounts[i-1];
716 pme->rdispls[i] = pme->rdispls[i-1]+pme->rcounts[i-1];
717 pme->sidx[i] = pme->sdispls[i];
719 /* Total # of particles to be received */
720 atc->n = pme->rdispls[atc->nslab-1] + pme->rcounts[atc->nslab-1];
722 pme_realloc_atomcomm_things(atc);
724 /* Copy particle coordinates into send buffer and exchange*/
725 for (i = 0; (i < n); i++)
727 ii = DIM*pme->sidx[pme->idxa[i]];
728 pme->sidx[pme->idxa[i]]++;
729 pme->redist_buf[ii+XX] = x_f[i][XX];
730 pme->redist_buf[ii+YY] = x_f[i][YY];
731 pme->redist_buf[ii+ZZ] = x_f[i][ZZ];
733 MPI_Alltoallv(pme->redist_buf, pme->scounts, pme->sdispls,
734 pme->rvec_mpi, atc->x, pme->rcounts, pme->rdispls,
735 pme->rvec_mpi, atc->mpi_comm);
739 /* Copy charge into send buffer and exchange*/
740 for (i = 0; i < atc->nslab; i++)
742 pme->sidx[i] = pme->sdispls[i];
744 for (i = 0; (i < n); i++)
746 ii = pme->sidx[pme->idxa[i]];
747 pme->sidx[pme->idxa[i]]++;
748 pme->redist_buf[ii] = charge[i];
750 MPI_Alltoallv(pme->redist_buf, pme->scounts, pme->sdispls, mpi_type,
751 atc->q, pme->rcounts, pme->rdispls, mpi_type,
754 else /* backward, redistribution from pme to pp */
756 MPI_Alltoallv(atc->f, pme->rcounts, pme->rdispls, pme->rvec_mpi,
757 pme->redist_buf, pme->scounts, pme->sdispls,
758 pme->rvec_mpi, atc->mpi_comm);
760 /* Copy data from receive buffer */
761 for (i = 0; i < atc->nslab; i++)
763 pme->sidx[i] = pme->sdispls[i];
765 for (i = 0; (i < n); i++)
767 ii = DIM*pme->sidx[pme->idxa[i]];
768 x_f[i][XX] += pme->redist_buf[ii+XX];
769 x_f[i][YY] += pme->redist_buf[ii+YY];
770 x_f[i][ZZ] += pme->redist_buf[ii+ZZ];
771 pme->sidx[pme->idxa[i]]++;
777 static void pme_dd_sendrecv(pme_atomcomm_t *atc,
778 gmx_bool bBackward, int shift,
779 void *buf_s, int nbyte_s,
780 void *buf_r, int nbyte_r)
786 if (bBackward == FALSE)
788 dest = atc->node_dest[shift];
789 src = atc->node_src[shift];
793 dest = atc->node_src[shift];
794 src = atc->node_dest[shift];
797 if (nbyte_s > 0 && nbyte_r > 0)
799 MPI_Sendrecv(buf_s, nbyte_s, MPI_BYTE,
801 buf_r, nbyte_r, MPI_BYTE,
803 atc->mpi_comm, &stat);
805 else if (nbyte_s > 0)
807 MPI_Send(buf_s, nbyte_s, MPI_BYTE,
811 else if (nbyte_r > 0)
813 MPI_Recv(buf_r, nbyte_r, MPI_BYTE,
815 atc->mpi_comm, &stat);
820 static void dd_pmeredist_x_q(gmx_pme_t pme,
821 int n, gmx_bool bX, rvec *x, real *charge,
824 int *commnode, *buf_index;
825 int nnodes_comm, i, nsend, local_pos, buf_pos, node, scount, rcount;
827 commnode = atc->node_dest;
828 buf_index = atc->buf_index;
830 nnodes_comm = min(2*atc->maxshift, atc->nslab-1);
833 for (i = 0; i < nnodes_comm; i++)
835 buf_index[commnode[i]] = nsend;
836 nsend += atc->count[commnode[i]];
840 if (atc->count[atc->nodeid] + nsend != n)
842 gmx_fatal(FARGS, "%d particles communicated to PME node %d are more than 2/3 times the cut-off out of the domain decomposition cell of their charge group in dimension %c.\n"
843 "This usually means that your system is not well equilibrated.",
844 n - (atc->count[atc->nodeid] + nsend),
845 pme->nodeid, 'x'+atc->dimind);
848 if (nsend > pme->buf_nalloc)
850 pme->buf_nalloc = over_alloc_dd(nsend);
851 srenew(pme->bufv, pme->buf_nalloc);
852 srenew(pme->bufr, pme->buf_nalloc);
855 atc->n = atc->count[atc->nodeid];
856 for (i = 0; i < nnodes_comm; i++)
858 scount = atc->count[commnode[i]];
859 /* Communicate the count */
862 fprintf(debug, "dimind %d PME node %d send to node %d: %d\n",
863 atc->dimind, atc->nodeid, commnode[i], scount);
865 pme_dd_sendrecv(atc, FALSE, i,
866 &scount, sizeof(int),
867 &atc->rcount[i], sizeof(int));
868 atc->n += atc->rcount[i];
871 pme_realloc_atomcomm_things(atc);
875 for (i = 0; i < n; i++)
878 if (node == atc->nodeid)
880 /* Copy direct to the receive buffer */
883 copy_rvec(x[i], atc->x[local_pos]);
885 atc->q[local_pos] = charge[i];
890 /* Copy to the send buffer */
893 copy_rvec(x[i], pme->bufv[buf_index[node]]);
895 pme->bufr[buf_index[node]] = charge[i];
901 for (i = 0; i < nnodes_comm; i++)
903 scount = atc->count[commnode[i]];
904 rcount = atc->rcount[i];
905 if (scount > 0 || rcount > 0)
909 /* Communicate the coordinates */
910 pme_dd_sendrecv(atc, FALSE, i,
911 pme->bufv[buf_pos], scount*sizeof(rvec),
912 atc->x[local_pos], rcount*sizeof(rvec));
914 /* Communicate the charges */
915 pme_dd_sendrecv(atc, FALSE, i,
916 pme->bufr+buf_pos, scount*sizeof(real),
917 atc->q+local_pos, rcount*sizeof(real));
919 local_pos += atc->rcount[i];
924 static void dd_pmeredist_f(gmx_pme_t pme, pme_atomcomm_t *atc,
928 int *commnode, *buf_index;
929 int nnodes_comm, local_pos, buf_pos, i, scount, rcount, node;
931 commnode = atc->node_dest;
932 buf_index = atc->buf_index;
934 nnodes_comm = min(2*atc->maxshift, atc->nslab-1);
936 local_pos = atc->count[atc->nodeid];
938 for (i = 0; i < nnodes_comm; i++)
940 scount = atc->rcount[i];
941 rcount = atc->count[commnode[i]];
942 if (scount > 0 || rcount > 0)
944 /* Communicate the forces */
945 pme_dd_sendrecv(atc, TRUE, i,
946 atc->f[local_pos], scount*sizeof(rvec),
947 pme->bufv[buf_pos], rcount*sizeof(rvec));
950 buf_index[commnode[i]] = buf_pos;
957 for (i = 0; i < n; i++)
960 if (node == atc->nodeid)
962 /* Add from the local force array */
963 rvec_inc(f[i], atc->f[local_pos]);
968 /* Add from the receive buffer */
969 rvec_inc(f[i], pme->bufv[buf_index[node]]);
976 for (i = 0; i < n; i++)
979 if (node == atc->nodeid)
981 /* Copy from the local force array */
982 copy_rvec(atc->f[local_pos], f[i]);
987 /* Copy from the receive buffer */
988 copy_rvec(pme->bufv[buf_index[node]], f[i]);
997 gmx_sum_qgrid_dd(gmx_pme_t pme, real *grid, int direction)
999 pme_overlap_t *overlap;
1000 int send_index0, send_nindex;
1001 int recv_index0, recv_nindex;
1003 int i, j, k, ix, iy, iz, icnt;
1004 int ipulse, send_id, recv_id, datasize;
1006 real *sendptr, *recvptr;
1008 /* Start with minor-rank communication. This is a bit of a pain since it is not contiguous */
1009 overlap = &pme->overlap[1];
1011 for (ipulse = 0; ipulse < overlap->noverlap_nodes; ipulse++)
1013 /* Since we have already (un)wrapped the overlap in the z-dimension,
1014 * we only have to communicate 0 to nkz (not pmegrid_nz).
1016 if (direction == GMX_SUM_QGRID_FORWARD)
1018 send_id = overlap->send_id[ipulse];
1019 recv_id = overlap->recv_id[ipulse];
1020 send_index0 = overlap->comm_data[ipulse].send_index0;
1021 send_nindex = overlap->comm_data[ipulse].send_nindex;
1022 recv_index0 = overlap->comm_data[ipulse].recv_index0;
1023 recv_nindex = overlap->comm_data[ipulse].recv_nindex;
1027 send_id = overlap->recv_id[ipulse];
1028 recv_id = overlap->send_id[ipulse];
1029 send_index0 = overlap->comm_data[ipulse].recv_index0;
1030 send_nindex = overlap->comm_data[ipulse].recv_nindex;
1031 recv_index0 = overlap->comm_data[ipulse].send_index0;
1032 recv_nindex = overlap->comm_data[ipulse].send_nindex;
1035 /* Copy data to contiguous send buffer */
1038 fprintf(debug, "PME send node %d %d -> %d grid start %d Communicating %d to %d\n",
1039 pme->nodeid, overlap->nodeid, send_id,
1040 pme->pmegrid_start_iy,
1041 send_index0-pme->pmegrid_start_iy,
1042 send_index0-pme->pmegrid_start_iy+send_nindex);
1045 for (i = 0; i < pme->pmegrid_nx; i++)
1048 for (j = 0; j < send_nindex; j++)
1050 iy = j + send_index0 - pme->pmegrid_start_iy;
1051 for (k = 0; k < pme->nkz; k++)
1054 overlap->sendbuf[icnt++] = grid[ix*(pme->pmegrid_ny*pme->pmegrid_nz)+iy*(pme->pmegrid_nz)+iz];
1059 datasize = pme->pmegrid_nx * pme->nkz;
1061 MPI_Sendrecv(overlap->sendbuf, send_nindex*datasize, GMX_MPI_REAL,
1063 overlap->recvbuf, recv_nindex*datasize, GMX_MPI_REAL,
1065 overlap->mpi_comm, &stat);
1067 /* Get data from contiguous recv buffer */
1070 fprintf(debug, "PME recv node %d %d <- %d grid start %d Communicating %d to %d\n",
1071 pme->nodeid, overlap->nodeid, recv_id,
1072 pme->pmegrid_start_iy,
1073 recv_index0-pme->pmegrid_start_iy,
1074 recv_index0-pme->pmegrid_start_iy+recv_nindex);
1077 for (i = 0; i < pme->pmegrid_nx; i++)
1080 for (j = 0; j < recv_nindex; j++)
1082 iy = j + recv_index0 - pme->pmegrid_start_iy;
1083 for (k = 0; k < pme->nkz; k++)
1086 if (direction == GMX_SUM_QGRID_FORWARD)
1088 grid[ix*(pme->pmegrid_ny*pme->pmegrid_nz)+iy*(pme->pmegrid_nz)+iz] += overlap->recvbuf[icnt++];
1092 grid[ix*(pme->pmegrid_ny*pme->pmegrid_nz)+iy*(pme->pmegrid_nz)+iz] = overlap->recvbuf[icnt++];
1099 /* Major dimension is easier, no copying required,
1100 * but we might have to sum to separate array.
1101 * Since we don't copy, we have to communicate up to pmegrid_nz,
1102 * not nkz as for the minor direction.
1104 overlap = &pme->overlap[0];
1106 for (ipulse = 0; ipulse < overlap->noverlap_nodes; ipulse++)
1108 if (direction == GMX_SUM_QGRID_FORWARD)
1110 send_id = overlap->send_id[ipulse];
1111 recv_id = overlap->recv_id[ipulse];
1112 send_index0 = overlap->comm_data[ipulse].send_index0;
1113 send_nindex = overlap->comm_data[ipulse].send_nindex;
1114 recv_index0 = overlap->comm_data[ipulse].recv_index0;
1115 recv_nindex = overlap->comm_data[ipulse].recv_nindex;
1116 recvptr = overlap->recvbuf;
1120 send_id = overlap->recv_id[ipulse];
1121 recv_id = overlap->send_id[ipulse];
1122 send_index0 = overlap->comm_data[ipulse].recv_index0;
1123 send_nindex = overlap->comm_data[ipulse].recv_nindex;
1124 recv_index0 = overlap->comm_data[ipulse].send_index0;
1125 recv_nindex = overlap->comm_data[ipulse].send_nindex;
1126 recvptr = grid + (recv_index0-pme->pmegrid_start_ix)*(pme->pmegrid_ny*pme->pmegrid_nz);
1129 sendptr = grid + (send_index0-pme->pmegrid_start_ix)*(pme->pmegrid_ny*pme->pmegrid_nz);
1130 datasize = pme->pmegrid_ny * pme->pmegrid_nz;
1134 fprintf(debug, "PME send node %d %d -> %d grid start %d Communicating %d to %d\n",
1135 pme->nodeid, overlap->nodeid, send_id,
1136 pme->pmegrid_start_ix,
1137 send_index0-pme->pmegrid_start_ix,
1138 send_index0-pme->pmegrid_start_ix+send_nindex);
1139 fprintf(debug, "PME recv node %d %d <- %d grid start %d Communicating %d to %d\n",
1140 pme->nodeid, overlap->nodeid, recv_id,
1141 pme->pmegrid_start_ix,
1142 recv_index0-pme->pmegrid_start_ix,
1143 recv_index0-pme->pmegrid_start_ix+recv_nindex);
1146 MPI_Sendrecv(sendptr, send_nindex*datasize, GMX_MPI_REAL,
1148 recvptr, recv_nindex*datasize, GMX_MPI_REAL,
1150 overlap->mpi_comm, &stat);
1152 /* ADD data from contiguous recv buffer */
1153 if (direction == GMX_SUM_QGRID_FORWARD)
1155 p = grid + (recv_index0-pme->pmegrid_start_ix)*(pme->pmegrid_ny*pme->pmegrid_nz);
1156 for (i = 0; i < recv_nindex*datasize; i++)
1158 p[i] += overlap->recvbuf[i];
1167 copy_pmegrid_to_fftgrid(gmx_pme_t pme, real *pmegrid, real *fftgrid)
1169 ivec local_fft_ndata, local_fft_offset, local_fft_size;
1170 ivec local_pme_size;
1174 /* Dimensions should be identical for A/B grid, so we just use A here */
1175 gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
1180 local_pme_size[0] = pme->pmegrid_nx;
1181 local_pme_size[1] = pme->pmegrid_ny;
1182 local_pme_size[2] = pme->pmegrid_nz;
1184 /* The fftgrid is always 'justified' to the lower-left corner of the PME grid,
1185 the offset is identical, and the PME grid always has more data (due to overlap)
1190 char fn[STRLEN], format[STRLEN];
1192 sprintf(fn, "pmegrid%d.pdb", pme->nodeid);
1193 fp = ffopen(fn, "w");
1194 sprintf(fn, "pmegrid%d.txt", pme->nodeid);
1195 fp2 = ffopen(fn, "w");
1196 sprintf(format, "%s%s\n", pdbformat, "%6.2f%6.2f");
1199 for (ix = 0; ix < local_fft_ndata[XX]; ix++)
1201 for (iy = 0; iy < local_fft_ndata[YY]; iy++)
1203 for (iz = 0; iz < local_fft_ndata[ZZ]; iz++)
1205 pmeidx = ix*(local_pme_size[YY]*local_pme_size[ZZ])+iy*(local_pme_size[ZZ])+iz;
1206 fftidx = ix*(local_fft_size[YY]*local_fft_size[ZZ])+iy*(local_fft_size[ZZ])+iz;
1207 fftgrid[fftidx] = pmegrid[pmeidx];
1209 val = 100*pmegrid[pmeidx];
1210 if (pmegrid[pmeidx] != 0)
1212 fprintf(fp, format, "ATOM", pmeidx, "CA", "GLY", ' ', pmeidx, ' ',
1213 5.0*ix, 5.0*iy, 5.0*iz, 1.0, val);
1215 if (pmegrid[pmeidx] != 0)
1217 fprintf(fp2, "%-12s %5d %5d %5d %12.5e\n",
1219 pme->pmegrid_start_ix + ix,
1220 pme->pmegrid_start_iy + iy,
1221 pme->pmegrid_start_iz + iz,
1237 static gmx_cycles_t omp_cyc_start()
1239 return gmx_cycles_read();
1242 static gmx_cycles_t omp_cyc_end(gmx_cycles_t c)
1244 return gmx_cycles_read() - c;
1249 copy_fftgrid_to_pmegrid(gmx_pme_t pme, const real *fftgrid, real *pmegrid,
1250 int nthread, int thread)
1252 ivec local_fft_ndata, local_fft_offset, local_fft_size;
1253 ivec local_pme_size;
1254 int ixy0, ixy1, ixy, ix, iy, iz;
1256 #ifdef PME_TIME_THREADS
1258 static double cs1 = 0;
1262 #ifdef PME_TIME_THREADS
1263 c1 = omp_cyc_start();
1265 /* Dimensions should be identical for A/B grid, so we just use A here */
1266 gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
1271 local_pme_size[0] = pme->pmegrid_nx;
1272 local_pme_size[1] = pme->pmegrid_ny;
1273 local_pme_size[2] = pme->pmegrid_nz;
1275 /* The fftgrid is always 'justified' to the lower-left corner of the PME grid,
1276 the offset is identical, and the PME grid always has more data (due to overlap)
1278 ixy0 = ((thread )*local_fft_ndata[XX]*local_fft_ndata[YY])/nthread;
1279 ixy1 = ((thread+1)*local_fft_ndata[XX]*local_fft_ndata[YY])/nthread;
1281 for (ixy = ixy0; ixy < ixy1; ixy++)
1283 ix = ixy/local_fft_ndata[YY];
1284 iy = ixy - ix*local_fft_ndata[YY];
1286 pmeidx = (ix*local_pme_size[YY] + iy)*local_pme_size[ZZ];
1287 fftidx = (ix*local_fft_size[YY] + iy)*local_fft_size[ZZ];
1288 for (iz = 0; iz < local_fft_ndata[ZZ]; iz++)
1290 pmegrid[pmeidx+iz] = fftgrid[fftidx+iz];
1294 #ifdef PME_TIME_THREADS
1295 c1 = omp_cyc_end(c1);
1300 printf("copy %.2f\n", cs1*1e-9);
1309 wrap_periodic_pmegrid(gmx_pme_t pme, real *pmegrid)
1311 int nx, ny, nz, pnx, pny, pnz, ny_x, overlap, ix, iy, iz;
1317 pnx = pme->pmegrid_nx;
1318 pny = pme->pmegrid_ny;
1319 pnz = pme->pmegrid_nz;
1321 overlap = pme->pme_order - 1;
1323 /* Add periodic overlap in z */
1324 for (ix = 0; ix < pme->pmegrid_nx; ix++)
1326 for (iy = 0; iy < pme->pmegrid_ny; iy++)
1328 for (iz = 0; iz < overlap; iz++)
1330 pmegrid[(ix*pny+iy)*pnz+iz] +=
1331 pmegrid[(ix*pny+iy)*pnz+nz+iz];
1336 if (pme->nnodes_minor == 1)
1338 for (ix = 0; ix < pme->pmegrid_nx; ix++)
1340 for (iy = 0; iy < overlap; iy++)
1342 for (iz = 0; iz < nz; iz++)
1344 pmegrid[(ix*pny+iy)*pnz+iz] +=
1345 pmegrid[(ix*pny+ny+iy)*pnz+iz];
1351 if (pme->nnodes_major == 1)
1353 ny_x = (pme->nnodes_minor == 1 ? ny : pme->pmegrid_ny);
1355 for (ix = 0; ix < overlap; ix++)
1357 for (iy = 0; iy < ny_x; iy++)
1359 for (iz = 0; iz < nz; iz++)
1361 pmegrid[(ix*pny+iy)*pnz+iz] +=
1362 pmegrid[((nx+ix)*pny+iy)*pnz+iz];
1371 unwrap_periodic_pmegrid(gmx_pme_t pme, real *pmegrid)
1373 int nx, ny, nz, pnx, pny, pnz, ny_x, overlap, ix;
1379 pnx = pme->pmegrid_nx;
1380 pny = pme->pmegrid_ny;
1381 pnz = pme->pmegrid_nz;
1383 overlap = pme->pme_order - 1;
1385 if (pme->nnodes_major == 1)
1387 ny_x = (pme->nnodes_minor == 1 ? ny : pme->pmegrid_ny);
1389 for (ix = 0; ix < overlap; ix++)
1393 for (iy = 0; iy < ny_x; iy++)
1395 for (iz = 0; iz < nz; iz++)
1397 pmegrid[((nx+ix)*pny+iy)*pnz+iz] =
1398 pmegrid[(ix*pny+iy)*pnz+iz];
1404 if (pme->nnodes_minor == 1)
1406 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1407 for (ix = 0; ix < pme->pmegrid_nx; ix++)
1411 for (iy = 0; iy < overlap; iy++)
1413 for (iz = 0; iz < nz; iz++)
1415 pmegrid[(ix*pny+ny+iy)*pnz+iz] =
1416 pmegrid[(ix*pny+iy)*pnz+iz];
1422 /* Copy periodic overlap in z */
1423 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1424 for (ix = 0; ix < pme->pmegrid_nx; ix++)
1428 for (iy = 0; iy < pme->pmegrid_ny; iy++)
1430 for (iz = 0; iz < overlap; iz++)
1432 pmegrid[(ix*pny+iy)*pnz+nz+iz] =
1433 pmegrid[(ix*pny+iy)*pnz+iz];
1440 /* This has to be a macro to enable full compiler optimization with xlC (and probably others too) */
1441 #define DO_BSPLINE(order) \
1442 for (ithx = 0; (ithx < order); ithx++) \
1444 index_x = (i0+ithx)*pny*pnz; \
1445 valx = qn*thx[ithx]; \
1447 for (ithy = 0; (ithy < order); ithy++) \
1449 valxy = valx*thy[ithy]; \
1450 index_xy = index_x+(j0+ithy)*pnz; \
1452 for (ithz = 0; (ithz < order); ithz++) \
1454 index_xyz = index_xy+(k0+ithz); \
1455 grid[index_xyz] += valxy*thz[ithz]; \
1461 static void spread_q_bsplines_thread(pmegrid_t *pmegrid,
1462 pme_atomcomm_t *atc, splinedata_t *spline,
1463 pme_spline_work_t *work)
1466 /* spread charges from home atoms to local grid */
1469 int b, i, nn, n, ithx, ithy, ithz, i0, j0, k0;
1471 int order, norder, index_x, index_xy, index_xyz;
1472 real valx, valxy, qn;
1473 real *thx, *thy, *thz;
1474 int localsize, bndsize;
1475 int pnx, pny, pnz, ndatatot;
1476 int offx, offy, offz;
1478 #if defined PME_SIMD4_SPREAD_GATHER && !defined PME_SIMD4_UNALIGNED
1479 real thz_buffer[12], *thz_aligned;
1481 thz_aligned = gmx_simd4_align_real(thz_buffer);
1484 pnx = pmegrid->s[XX];
1485 pny = pmegrid->s[YY];
1486 pnz = pmegrid->s[ZZ];
1488 offx = pmegrid->offset[XX];
1489 offy = pmegrid->offset[YY];
1490 offz = pmegrid->offset[ZZ];
1492 ndatatot = pnx*pny*pnz;
1493 grid = pmegrid->grid;
1494 for (i = 0; i < ndatatot; i++)
1499 order = pmegrid->order;
1501 for (nn = 0; nn < spline->n; nn++)
1503 n = spline->ind[nn];
1508 idxptr = atc->idx[n];
1511 i0 = idxptr[XX] - offx;
1512 j0 = idxptr[YY] - offy;
1513 k0 = idxptr[ZZ] - offz;
1515 thx = spline->theta[XX] + norder;
1516 thy = spline->theta[YY] + norder;
1517 thz = spline->theta[ZZ] + norder;
1522 #ifdef PME_SIMD4_SPREAD_GATHER
1523 #ifdef PME_SIMD4_UNALIGNED
1524 #define PME_SPREAD_SIMD4_ORDER4
1526 #define PME_SPREAD_SIMD4_ALIGNED
1529 #include "pme_simd4.h"
1535 #ifdef PME_SIMD4_SPREAD_GATHER
1536 #define PME_SPREAD_SIMD4_ALIGNED
1538 #include "pme_simd4.h"
1551 static void set_grid_alignment(int *pmegrid_nz, int pme_order)
1553 #ifdef PME_SIMD4_SPREAD_GATHER
1555 #ifndef PME_SIMD4_UNALIGNED
1560 /* Round nz up to a multiple of 4 to ensure alignment */
1561 *pmegrid_nz = ((*pmegrid_nz + 3) & ~3);
1566 static void set_gridsize_alignment(int gmx_unused *gridsize, int gmx_unused pme_order)
1568 #ifdef PME_SIMD4_SPREAD_GATHER
1569 #ifndef PME_SIMD4_UNALIGNED
1572 /* Add extra elements to ensured aligned operations do not go
1573 * beyond the allocated grid size.
1574 * Note that for pme_order=5, the pme grid z-size alignment
1575 * ensures that we will not go beyond the grid size.
1583 static void pmegrid_init(pmegrid_t *grid,
1584 int cx, int cy, int cz,
1585 int x0, int y0, int z0,
1586 int x1, int y1, int z1,
1587 gmx_bool set_alignment,
1596 grid->offset[XX] = x0;
1597 grid->offset[YY] = y0;
1598 grid->offset[ZZ] = z0;
1599 grid->n[XX] = x1 - x0 + pme_order - 1;
1600 grid->n[YY] = y1 - y0 + pme_order - 1;
1601 grid->n[ZZ] = z1 - z0 + pme_order - 1;
1602 copy_ivec(grid->n, grid->s);
1605 set_grid_alignment(&nz, pme_order);
1610 else if (nz != grid->s[ZZ])
1612 gmx_incons("pmegrid_init call with an unaligned z size");
1615 grid->order = pme_order;
1618 gridsize = grid->s[XX]*grid->s[YY]*grid->s[ZZ];
1619 set_gridsize_alignment(&gridsize, pme_order);
1620 snew_aligned(grid->grid, gridsize, SIMD4_ALIGNMENT);
1628 static int div_round_up(int enumerator, int denominator)
1630 return (enumerator + denominator - 1)/denominator;
1633 static void make_subgrid_division(const ivec n, int ovl, int nthread,
1636 int gsize_opt, gsize;
1641 for (nsx = 1; nsx <= nthread; nsx++)
1643 if (nthread % nsx == 0)
1645 for (nsy = 1; nsy <= nthread; nsy++)
1647 if (nsx*nsy <= nthread && nthread % (nsx*nsy) == 0)
1649 nsz = nthread/(nsx*nsy);
1651 /* Determine the number of grid points per thread */
1653 (div_round_up(n[XX], nsx) + ovl)*
1654 (div_round_up(n[YY], nsy) + ovl)*
1655 (div_round_up(n[ZZ], nsz) + ovl);
1657 /* Minimize the number of grids points per thread
1658 * and, secondarily, the number of cuts in minor dimensions.
1660 if (gsize_opt == -1 ||
1661 gsize < gsize_opt ||
1662 (gsize == gsize_opt &&
1663 (nsz < nsub[ZZ] || (nsz == nsub[ZZ] && nsy < nsub[YY]))))
1675 env = getenv("GMX_PME_THREAD_DIVISION");
1678 sscanf(env, "%d %d %d", &nsub[XX], &nsub[YY], &nsub[ZZ]);
1681 if (nsub[XX]*nsub[YY]*nsub[ZZ] != nthread)
1683 gmx_fatal(FARGS, "PME grid thread division (%d x %d x %d) does not match the total number of threads (%d)", nsub[XX], nsub[YY], nsub[ZZ], nthread);
1687 static void pmegrids_init(pmegrids_t *grids,
1688 int nx, int ny, int nz, int nz_base,
1690 gmx_bool bUseThreads,
1695 ivec n, n_base, g0, g1;
1696 int t, x, y, z, d, i, tfac;
1697 int max_comm_lines = -1;
1699 n[XX] = nx - (pme_order - 1);
1700 n[YY] = ny - (pme_order - 1);
1701 n[ZZ] = nz - (pme_order - 1);
1703 copy_ivec(n, n_base);
1704 n_base[ZZ] = nz_base;
1706 pmegrid_init(&grids->grid, 0, 0, 0, 0, 0, 0, n[XX], n[YY], n[ZZ], FALSE, pme_order,
1709 grids->nthread = nthread;
1711 make_subgrid_division(n_base, pme_order-1, grids->nthread, grids->nc);
1718 for (d = 0; d < DIM; d++)
1720 nst[d] = div_round_up(n[d], grids->nc[d]) + pme_order - 1;
1722 set_grid_alignment(&nst[ZZ], pme_order);
1726 fprintf(debug, "pmegrid thread local division: %d x %d x %d\n",
1727 grids->nc[XX], grids->nc[YY], grids->nc[ZZ]);
1728 fprintf(debug, "pmegrid %d %d %d max thread pmegrid %d %d %d\n",
1730 nst[XX], nst[YY], nst[ZZ]);
1733 snew(grids->grid_th, grids->nthread);
1735 gridsize = nst[XX]*nst[YY]*nst[ZZ];
1736 set_gridsize_alignment(&gridsize, pme_order);
1737 snew_aligned(grids->grid_all,
1738 grids->nthread*gridsize+(grids->nthread+1)*GMX_CACHE_SEP,
1741 for (x = 0; x < grids->nc[XX]; x++)
1743 for (y = 0; y < grids->nc[YY]; y++)
1745 for (z = 0; z < grids->nc[ZZ]; z++)
1747 pmegrid_init(&grids->grid_th[t],
1749 (n[XX]*(x ))/grids->nc[XX],
1750 (n[YY]*(y ))/grids->nc[YY],
1751 (n[ZZ]*(z ))/grids->nc[ZZ],
1752 (n[XX]*(x+1))/grids->nc[XX],
1753 (n[YY]*(y+1))/grids->nc[YY],
1754 (n[ZZ]*(z+1))/grids->nc[ZZ],
1757 grids->grid_all+GMX_CACHE_SEP+t*(gridsize+GMX_CACHE_SEP));
1765 grids->grid_th = NULL;
1768 snew(grids->g2t, DIM);
1770 for (d = DIM-1; d >= 0; d--)
1772 snew(grids->g2t[d], n[d]);
1774 for (i = 0; i < n[d]; i++)
1776 /* The second check should match the parameters
1777 * of the pmegrid_init call above.
1779 while (t + 1 < grids->nc[d] && i >= (n[d]*(t+1))/grids->nc[d])
1783 grids->g2t[d][i] = t*tfac;
1786 tfac *= grids->nc[d];
1790 case XX: max_comm_lines = overlap_x; break;
1791 case YY: max_comm_lines = overlap_y; break;
1792 case ZZ: max_comm_lines = pme_order - 1; break;
1794 grids->nthread_comm[d] = 0;
1795 while ((n[d]*grids->nthread_comm[d])/grids->nc[d] < max_comm_lines &&
1796 grids->nthread_comm[d] < grids->nc[d])
1798 grids->nthread_comm[d]++;
1802 fprintf(debug, "pmegrid thread grid communication range in %c: %d\n",
1803 'x'+d, grids->nthread_comm[d]);
1805 /* It should be possible to make grids->nthread_comm[d]==grids->nc[d]
1806 * work, but this is not a problematic restriction.
1808 if (grids->nc[d] > 1 && grids->nthread_comm[d] > grids->nc[d])
1810 gmx_fatal(FARGS, "Too many threads for PME (%d) compared to the number of grid lines, reduce the number of threads doing PME", grids->nthread);
1816 static void pmegrids_destroy(pmegrids_t *grids)
1820 if (grids->grid.grid != NULL)
1822 sfree(grids->grid.grid);
1824 if (grids->nthread > 0)
1826 for (t = 0; t < grids->nthread; t++)
1828 sfree(grids->grid_th[t].grid);
1830 sfree(grids->grid_th);
1836 static void realloc_work(pme_work_t *work, int nkx)
1840 if (nkx > work->nalloc)
1843 srenew(work->mhx, work->nalloc);
1844 srenew(work->mhy, work->nalloc);
1845 srenew(work->mhz, work->nalloc);
1846 srenew(work->m2, work->nalloc);
1847 /* Allocate an aligned pointer for SIMD operations, including extra
1848 * elements at the end for padding.
1851 simd_width = GMX_SIMD_WIDTH_HERE;
1853 /* We can use any alignment, apart from 0, so we use 4 */
1856 sfree_aligned(work->denom);
1857 sfree_aligned(work->tmp1);
1858 sfree_aligned(work->eterm);
1859 snew_aligned(work->denom, work->nalloc+simd_width, simd_width*sizeof(real));
1860 snew_aligned(work->tmp1, work->nalloc+simd_width, simd_width*sizeof(real));
1861 snew_aligned(work->eterm, work->nalloc+simd_width, simd_width*sizeof(real));
1862 srenew(work->m2inv, work->nalloc);
1867 static void free_work(pme_work_t *work)
1873 sfree_aligned(work->denom);
1874 sfree_aligned(work->tmp1);
1875 sfree_aligned(work->eterm);
1881 /* Calculate exponentials through SIMD */
1882 inline static void calc_exponentials(int start, int end, real f, real *d_aligned, real *r_aligned, real *e_aligned)
1885 const gmx_mm_pr two = gmx_set1_pr(2.0);
1888 gmx_mm_pr tmp_d1, d_inv, tmp_r, tmp_e;
1890 f_simd = gmx_set1_pr(f);
1891 for (kx = 0; kx < end; kx += GMX_SIMD_WIDTH_HERE)
1893 tmp_d1 = gmx_load_pr(d_aligned+kx);
1894 d_inv = gmx_inv_pr(tmp_d1);
1895 tmp_r = gmx_load_pr(r_aligned+kx);
1896 tmp_r = gmx_exp_pr(tmp_r);
1897 tmp_e = gmx_mul_pr(f_simd, d_inv);
1898 tmp_e = gmx_mul_pr(tmp_e, tmp_r);
1899 gmx_store_pr(e_aligned+kx, tmp_e);
1904 inline static void calc_exponentials(int start, int end, real f, real *d, real *r, real *e)
1907 for (kx = start; kx < end; kx++)
1911 for (kx = start; kx < end; kx++)
1915 for (kx = start; kx < end; kx++)
1917 e[kx] = f*r[kx]*d[kx];
1923 static int solve_pme_yzx(gmx_pme_t pme, t_complex *grid,
1924 real ewaldcoeff, real vol,
1926 int nthread, int thread)
1928 /* do recip sum over local cells in grid */
1929 /* y major, z middle, x minor or continuous */
1931 int kx, ky, kz, maxkx, maxky, maxkz;
1932 int nx, ny, nz, iyz0, iyz1, iyz, iy, iz, kxstart, kxend;
1934 real factor = M_PI*M_PI/(ewaldcoeff*ewaldcoeff);
1935 real ets2, struct2, vfactor, ets2vf;
1936 real d1, d2, energy = 0;
1938 real virxx = 0, virxy = 0, virxz = 0, viryy = 0, viryz = 0, virzz = 0;
1939 real rxx, ryx, ryy, rzx, rzy, rzz;
1941 real *mhx, *mhy, *mhz, *m2, *denom, *tmp1, *eterm, *m2inv;
1942 real mhxk, mhyk, mhzk, m2k;
1945 ivec local_ndata, local_offset, local_size;
1948 elfac = ONE_4PI_EPS0/pme->epsilon_r;
1954 /* Dimensions should be identical for A/B grid, so we just use A here */
1955 gmx_parallel_3dfft_complex_limits(pme->pfft_setupA,
1961 rxx = pme->recipbox[XX][XX];
1962 ryx = pme->recipbox[YY][XX];
1963 ryy = pme->recipbox[YY][YY];
1964 rzx = pme->recipbox[ZZ][XX];
1965 rzy = pme->recipbox[ZZ][YY];
1966 rzz = pme->recipbox[ZZ][ZZ];
1972 work = &pme->work[thread];
1977 denom = work->denom;
1979 eterm = work->eterm;
1980 m2inv = work->m2inv;
1982 iyz0 = local_ndata[YY]*local_ndata[ZZ]* thread /nthread;
1983 iyz1 = local_ndata[YY]*local_ndata[ZZ]*(thread+1)/nthread;
1985 for (iyz = iyz0; iyz < iyz1; iyz++)
1987 iy = iyz/local_ndata[ZZ];
1988 iz = iyz - iy*local_ndata[ZZ];
1990 ky = iy + local_offset[YY];
2001 by = M_PI*vol*pme->bsp_mod[YY][ky];
2003 kz = iz + local_offset[ZZ];
2007 bz = pme->bsp_mod[ZZ][kz];
2009 /* 0.5 correction for corner points */
2011 if (kz == 0 || kz == (nz+1)/2)
2016 p0 = grid + iy*local_size[ZZ]*local_size[XX] + iz*local_size[XX];
2018 /* We should skip the k-space point (0,0,0) */
2019 if (local_offset[XX] > 0 || ky > 0 || kz > 0)
2021 kxstart = local_offset[XX];
2025 kxstart = local_offset[XX] + 1;
2028 kxend = local_offset[XX] + local_ndata[XX];
2032 /* More expensive inner loop, especially because of the storage
2033 * of the mh elements in array's.
2034 * Because x is the minor grid index, all mh elements
2035 * depend on kx for triclinic unit cells.
2038 /* Two explicit loops to avoid a conditional inside the loop */
2039 for (kx = kxstart; kx < maxkx; kx++)
2044 mhyk = mx * ryx + my * ryy;
2045 mhzk = mx * rzx + my * rzy + mz * rzz;
2046 m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
2051 denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
2052 tmp1[kx] = -factor*m2k;
2055 for (kx = maxkx; kx < kxend; kx++)
2060 mhyk = mx * ryx + my * ryy;
2061 mhzk = mx * rzx + my * rzy + mz * rzz;
2062 m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
2067 denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
2068 tmp1[kx] = -factor*m2k;
2071 for (kx = kxstart; kx < kxend; kx++)
2073 m2inv[kx] = 1.0/m2[kx];
2076 calc_exponentials(kxstart, kxend, elfac, denom, tmp1, eterm);
2078 for (kx = kxstart; kx < kxend; kx++, p0++)
2083 p0->re = d1*eterm[kx];
2084 p0->im = d2*eterm[kx];
2086 struct2 = 2.0*(d1*d1+d2*d2);
2088 tmp1[kx] = eterm[kx]*struct2;
2091 for (kx = kxstart; kx < kxend; kx++)
2093 ets2 = corner_fac*tmp1[kx];
2094 vfactor = (factor*m2[kx] + 1.0)*2.0*m2inv[kx];
2097 ets2vf = ets2*vfactor;
2098 virxx += ets2vf*mhx[kx]*mhx[kx] - ets2;
2099 virxy += ets2vf*mhx[kx]*mhy[kx];
2100 virxz += ets2vf*mhx[kx]*mhz[kx];
2101 viryy += ets2vf*mhy[kx]*mhy[kx] - ets2;
2102 viryz += ets2vf*mhy[kx]*mhz[kx];
2103 virzz += ets2vf*mhz[kx]*mhz[kx] - ets2;
2108 /* We don't need to calculate the energy and the virial.
2109 * In this case the triclinic overhead is small.
2112 /* Two explicit loops to avoid a conditional inside the loop */
2114 for (kx = kxstart; kx < maxkx; kx++)
2119 mhyk = mx * ryx + my * ryy;
2120 mhzk = mx * rzx + my * rzy + mz * rzz;
2121 m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
2122 denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
2123 tmp1[kx] = -factor*m2k;
2126 for (kx = maxkx; kx < kxend; kx++)
2131 mhyk = mx * ryx + my * ryy;
2132 mhzk = mx * rzx + my * rzy + mz * rzz;
2133 m2k = mhxk*mhxk + mhyk*mhyk + mhzk*mhzk;
2134 denom[kx] = m2k*bz*by*pme->bsp_mod[XX][kx];
2135 tmp1[kx] = -factor*m2k;
2138 calc_exponentials(kxstart, kxend, elfac, denom, tmp1, eterm);
2140 for (kx = kxstart; kx < kxend; kx++, p0++)
2145 p0->re = d1*eterm[kx];
2146 p0->im = d2*eterm[kx];
2153 /* Update virial with local values.
2154 * The virial is symmetric by definition.
2155 * this virial seems ok for isotropic scaling, but I'm
2156 * experiencing problems on semiisotropic membranes.
2157 * IS THAT COMMENT STILL VALID??? (DvdS, 2001/02/07).
2159 work->vir[XX][XX] = 0.25*virxx;
2160 work->vir[YY][YY] = 0.25*viryy;
2161 work->vir[ZZ][ZZ] = 0.25*virzz;
2162 work->vir[XX][YY] = work->vir[YY][XX] = 0.25*virxy;
2163 work->vir[XX][ZZ] = work->vir[ZZ][XX] = 0.25*virxz;
2164 work->vir[YY][ZZ] = work->vir[ZZ][YY] = 0.25*viryz;
2166 /* This energy should be corrected for a charged system */
2167 work->energy = 0.5*energy;
2170 /* Return the loop count */
2171 return local_ndata[YY]*local_ndata[XX];
2174 static void get_pme_ener_vir(const gmx_pme_t pme, int nthread,
2175 real *mesh_energy, matrix vir)
2177 /* This function sums output over threads
2178 * and should therefore only be called after thread synchronization.
2182 *mesh_energy = pme->work[0].energy;
2183 copy_mat(pme->work[0].vir, vir);
2185 for (thread = 1; thread < nthread; thread++)
2187 *mesh_energy += pme->work[thread].energy;
2188 m_add(vir, pme->work[thread].vir, vir);
2192 #define DO_FSPLINE(order) \
2193 for (ithx = 0; (ithx < order); ithx++) \
2195 index_x = (i0+ithx)*pny*pnz; \
2199 for (ithy = 0; (ithy < order); ithy++) \
2201 index_xy = index_x+(j0+ithy)*pnz; \
2206 for (ithz = 0; (ithz < order); ithz++) \
2208 gval = grid[index_xy+(k0+ithz)]; \
2209 fxy1 += thz[ithz]*gval; \
2210 fz1 += dthz[ithz]*gval; \
2219 static void gather_f_bsplines(gmx_pme_t pme, real *grid,
2220 gmx_bool bClearF, pme_atomcomm_t *atc,
2221 splinedata_t *spline,
2224 /* sum forces for local particles */
2225 int nn, n, ithx, ithy, ithz, i0, j0, k0;
2226 int index_x, index_xy;
2227 int nx, ny, nz, pnx, pny, pnz;
2229 real tx, ty, dx, dy, qn;
2230 real fx, fy, fz, gval;
2232 real *thx, *thy, *thz, *dthx, *dthy, *dthz;
2234 real rxx, ryx, ryy, rzx, rzy, rzz;
2237 pme_spline_work_t *work;
2239 #if defined PME_SIMD4_SPREAD_GATHER && !defined PME_SIMD4_UNALIGNED
2240 real thz_buffer[12], *thz_aligned;
2241 real dthz_buffer[12], *dthz_aligned;
2243 thz_aligned = gmx_simd4_align_real(thz_buffer);
2244 dthz_aligned = gmx_simd4_align_real(dthz_buffer);
2247 work = pme->spline_work;
2249 order = pme->pme_order;
2250 thx = spline->theta[XX];
2251 thy = spline->theta[YY];
2252 thz = spline->theta[ZZ];
2253 dthx = spline->dtheta[XX];
2254 dthy = spline->dtheta[YY];
2255 dthz = spline->dtheta[ZZ];
2259 pnx = pme->pmegrid_nx;
2260 pny = pme->pmegrid_ny;
2261 pnz = pme->pmegrid_nz;
2263 rxx = pme->recipbox[XX][XX];
2264 ryx = pme->recipbox[YY][XX];
2265 ryy = pme->recipbox[YY][YY];
2266 rzx = pme->recipbox[ZZ][XX];
2267 rzy = pme->recipbox[ZZ][YY];
2268 rzz = pme->recipbox[ZZ][ZZ];
2270 for (nn = 0; nn < spline->n; nn++)
2272 n = spline->ind[nn];
2273 qn = scale*atc->q[n];
2286 idxptr = atc->idx[n];
2293 /* Pointer arithmetic alert, next six statements */
2294 thx = spline->theta[XX] + norder;
2295 thy = spline->theta[YY] + norder;
2296 thz = spline->theta[ZZ] + norder;
2297 dthx = spline->dtheta[XX] + norder;
2298 dthy = spline->dtheta[YY] + norder;
2299 dthz = spline->dtheta[ZZ] + norder;
2304 #ifdef PME_SIMD4_SPREAD_GATHER
2305 #ifdef PME_SIMD4_UNALIGNED
2306 #define PME_GATHER_F_SIMD4_ORDER4
2308 #define PME_GATHER_F_SIMD4_ALIGNED
2311 #include "pme_simd4.h"
2317 #ifdef PME_SIMD4_SPREAD_GATHER
2318 #define PME_GATHER_F_SIMD4_ALIGNED
2320 #include "pme_simd4.h"
2330 atc->f[n][XX] += -qn*( fx*nx*rxx );
2331 atc->f[n][YY] += -qn*( fx*nx*ryx + fy*ny*ryy );
2332 atc->f[n][ZZ] += -qn*( fx*nx*rzx + fy*ny*rzy + fz*nz*rzz );
2335 /* Since the energy and not forces are interpolated
2336 * the net force might not be exactly zero.
2337 * This can be solved by also interpolating F, but
2338 * that comes at a cost.
2339 * A better hack is to remove the net force every
2340 * step, but that must be done at a higher level
2341 * since this routine doesn't see all atoms if running
2342 * in parallel. Don't know how important it is? EL 990726
2347 static real gather_energy_bsplines(gmx_pme_t pme, real *grid,
2348 pme_atomcomm_t *atc)
2350 splinedata_t *spline;
2351 int n, ithx, ithy, ithz, i0, j0, k0;
2352 int index_x, index_xy;
2354 real energy, pot, tx, ty, qn, gval;
2355 real *thx, *thy, *thz;
2359 spline = &atc->spline[0];
2361 order = pme->pme_order;
2364 for (n = 0; (n < atc->n); n++)
2370 idxptr = atc->idx[n];
2377 /* Pointer arithmetic alert, next three statements */
2378 thx = spline->theta[XX] + norder;
2379 thy = spline->theta[YY] + norder;
2380 thz = spline->theta[ZZ] + norder;
2383 for (ithx = 0; (ithx < order); ithx++)
2385 index_x = (i0+ithx)*pme->pmegrid_ny*pme->pmegrid_nz;
2388 for (ithy = 0; (ithy < order); ithy++)
2390 index_xy = index_x+(j0+ithy)*pme->pmegrid_nz;
2393 for (ithz = 0; (ithz < order); ithz++)
2395 gval = grid[index_xy+(k0+ithz)];
2396 pot += tx*ty*thz[ithz]*gval;
2409 /* Macro to force loop unrolling by fixing order.
2410 * This gives a significant performance gain.
2412 #define CALC_SPLINE(order) \
2416 real data[PME_ORDER_MAX]; \
2417 real ddata[PME_ORDER_MAX]; \
2419 for (j = 0; (j < DIM); j++) \
2423 /* dr is relative offset from lower cell limit */ \
2424 data[order-1] = 0; \
2428 for (k = 3; (k < order); k++) \
2430 div = 1.0/(k - 1.0); \
2431 data[k-1] = div*dr*data[k-2]; \
2432 for (l = 1; (l < (k-1)); l++) \
2434 data[k-l-1] = div*((dr+l)*data[k-l-2]+(k-l-dr)* \
2437 data[0] = div*(1-dr)*data[0]; \
2439 /* differentiate */ \
2440 ddata[0] = -data[0]; \
2441 for (k = 1; (k < order); k++) \
2443 ddata[k] = data[k-1] - data[k]; \
2446 div = 1.0/(order - 1); \
2447 data[order-1] = div*dr*data[order-2]; \
2448 for (l = 1; (l < (order-1)); l++) \
2450 data[order-l-1] = div*((dr+l)*data[order-l-2]+ \
2451 (order-l-dr)*data[order-l-1]); \
2453 data[0] = div*(1 - dr)*data[0]; \
2455 for (k = 0; k < order; k++) \
2457 theta[j][i*order+k] = data[k]; \
2458 dtheta[j][i*order+k] = ddata[k]; \
2463 void make_bsplines(splinevec theta, splinevec dtheta, int order,
2464 rvec fractx[], int nr, int ind[], real charge[],
2465 gmx_bool bFreeEnergy)
2467 /* construct splines for local atoms */
2471 for (i = 0; i < nr; i++)
2473 /* With free energy we do not use the charge check.
2474 * In most cases this will be more efficient than calling make_bsplines
2475 * twice, since usually more than half the particles have charges.
2478 if (bFreeEnergy || charge[ii] != 0.0)
2483 case 4: CALC_SPLINE(4); break;
2484 case 5: CALC_SPLINE(5); break;
2485 default: CALC_SPLINE(order); break;
2492 void make_dft_mod(real *mod, real *data, int ndata)
2497 for (i = 0; i < ndata; i++)
2500 for (j = 0; j < ndata; j++)
2502 arg = (2.0*M_PI*i*j)/ndata;
2503 sc += data[j]*cos(arg);
2504 ss += data[j]*sin(arg);
2506 mod[i] = sc*sc+ss*ss;
2508 for (i = 0; i < ndata; i++)
2512 mod[i] = (mod[i-1]+mod[i+1])*0.5;
2518 static void make_bspline_moduli(splinevec bsp_mod,
2519 int nx, int ny, int nz, int order)
2521 int nmax = max(nx, max(ny, nz));
2522 real *data, *ddata, *bsp_data;
2528 snew(bsp_data, nmax);
2534 for (k = 3; k < order; k++)
2538 for (l = 1; l < (k-1); l++)
2540 data[k-l-1] = div*(l*data[k-l-2]+(k-l)*data[k-l-1]);
2542 data[0] = div*data[0];
2545 ddata[0] = -data[0];
2546 for (k = 1; k < order; k++)
2548 ddata[k] = data[k-1]-data[k];
2550 div = 1.0/(order-1);
2552 for (l = 1; l < (order-1); l++)
2554 data[order-l-1] = div*(l*data[order-l-2]+(order-l)*data[order-l-1]);
2556 data[0] = div*data[0];
2558 for (i = 0; i < nmax; i++)
2562 for (i = 1; i <= order; i++)
2564 bsp_data[i] = data[i-1];
2567 make_dft_mod(bsp_mod[XX], bsp_data, nx);
2568 make_dft_mod(bsp_mod[YY], bsp_data, ny);
2569 make_dft_mod(bsp_mod[ZZ], bsp_data, nz);
2577 /* Return the P3M optimal influence function */
2578 static double do_p3m_influence(double z, int order)
2585 /* The formula and most constants can be found in:
2586 * Ballenegger et al., JCTC 8, 936 (2012)
2591 return 1.0 - 2.0*z2/3.0;
2594 return 1.0 - z2 + 2.0*z4/15.0;
2597 return 1.0 - 4.0*z2/3.0 + 2.0*z4/5.0 + 4.0*z2*z4/315.0;
2600 return 1.0 - 5.0*z2/3.0 + 7.0*z4/9.0 - 17.0*z2*z4/189.0 + 2.0*z4*z4/2835.0;
2603 return 1.0 - 2.0*z2 + 19.0*z4/15.0 - 256.0*z2*z4/945.0 + 62.0*z4*z4/4725.0 + 4.0*z2*z4*z4/155925.0;
2606 return 1.0 - 7.0*z2/3.0 + 28.0*z4/15.0 - 16.0*z2*z4/27.0 + 26.0*z4*z4/405.0 - 2.0*z2*z4*z4/1485.0 + 4.0*z4*z4*z4/6081075.0;
2608 return 1.0 - 8.0*z2/3.0 + 116.0*z4/45.0 - 344.0*z2*z4/315.0 + 914.0*z4*z4/4725.0 - 248.0*z4*z4*z2/22275.0 + 21844.0*z4*z4*z4/212837625.0 - 8.0*z4*z4*z4*z2/638512875.0;
2615 /* Calculate the P3M B-spline moduli for one dimension */
2616 static void make_p3m_bspline_moduli_dim(real *bsp_mod, int n, int order)
2618 double zarg, zai, sinzai, infl;
2623 gmx_fatal(FARGS, "The current P3M code only supports orders up to 8");
2630 for (i = -maxk; i < 0; i++)
2634 infl = do_p3m_influence(sinzai, order);
2635 bsp_mod[n+i] = infl*infl*pow(sinzai/zai, -2.0*order);
2638 for (i = 1; i < maxk; i++)
2642 infl = do_p3m_influence(sinzai, order);
2643 bsp_mod[i] = infl*infl*pow(sinzai/zai, -2.0*order);
2647 /* Calculate the P3M B-spline moduli */
2648 static void make_p3m_bspline_moduli(splinevec bsp_mod,
2649 int nx, int ny, int nz, int order)
2651 make_p3m_bspline_moduli_dim(bsp_mod[XX], nx, order);
2652 make_p3m_bspline_moduli_dim(bsp_mod[YY], ny, order);
2653 make_p3m_bspline_moduli_dim(bsp_mod[ZZ], nz, order);
2657 static void setup_coordinate_communication(pme_atomcomm_t *atc)
2665 for (i = 1; i <= nslab/2; i++)
2667 fw = (atc->nodeid + i) % nslab;
2668 bw = (atc->nodeid - i + nslab) % nslab;
2671 atc->node_dest[n] = fw;
2672 atc->node_src[n] = bw;
2677 atc->node_dest[n] = bw;
2678 atc->node_src[n] = fw;
2684 int gmx_pme_destroy(FILE *log, gmx_pme_t *pmedata)
2690 fprintf(log, "Destroying PME data structures.\n");
2693 sfree((*pmedata)->nnx);
2694 sfree((*pmedata)->nny);
2695 sfree((*pmedata)->nnz);
2697 pmegrids_destroy(&(*pmedata)->pmegridA);
2699 sfree((*pmedata)->fftgridA);
2700 sfree((*pmedata)->cfftgridA);
2701 gmx_parallel_3dfft_destroy((*pmedata)->pfft_setupA);
2703 if ((*pmedata)->pmegridB.grid.grid != NULL)
2705 pmegrids_destroy(&(*pmedata)->pmegridB);
2706 sfree((*pmedata)->fftgridB);
2707 sfree((*pmedata)->cfftgridB);
2708 gmx_parallel_3dfft_destroy((*pmedata)->pfft_setupB);
2710 for (thread = 0; thread < (*pmedata)->nthread; thread++)
2712 free_work(&(*pmedata)->work[thread]);
2714 sfree((*pmedata)->work);
2722 static int mult_up(int n, int f)
2724 return ((n + f - 1)/f)*f;
2728 static double pme_load_imbalance(gmx_pme_t pme)
2733 nma = pme->nnodes_major;
2734 nmi = pme->nnodes_minor;
2736 n1 = mult_up(pme->nkx, nma)*mult_up(pme->nky, nmi)*pme->nkz;
2737 n2 = mult_up(pme->nkx, nma)*mult_up(pme->nkz, nmi)*pme->nky;
2738 n3 = mult_up(pme->nky, nma)*mult_up(pme->nkz, nmi)*pme->nkx;
2740 /* pme_solve is roughly double the cost of an fft */
2742 return (n1 + n2 + 3*n3)/(double)(6*pme->nkx*pme->nky*pme->nkz);
2745 static void init_atomcomm(gmx_pme_t pme, pme_atomcomm_t *atc,
2746 int dimind, gmx_bool bSpread)
2748 int nk, k, s, thread;
2750 atc->dimind = dimind;
2755 if (pme->nnodes > 1)
2757 atc->mpi_comm = pme->mpi_comm_d[dimind];
2758 MPI_Comm_size(atc->mpi_comm, &atc->nslab);
2759 MPI_Comm_rank(atc->mpi_comm, &atc->nodeid);
2763 fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", atc->dimind, atc->nslab, atc->nodeid);
2767 atc->bSpread = bSpread;
2768 atc->pme_order = pme->pme_order;
2772 /* These three allocations are not required for particle decomp. */
2773 snew(atc->node_dest, atc->nslab);
2774 snew(atc->node_src, atc->nslab);
2775 setup_coordinate_communication(atc);
2777 snew(atc->count_thread, pme->nthread);
2778 for (thread = 0; thread < pme->nthread; thread++)
2780 snew(atc->count_thread[thread], atc->nslab);
2782 atc->count = atc->count_thread[0];
2783 snew(atc->rcount, atc->nslab);
2784 snew(atc->buf_index, atc->nslab);
2787 atc->nthread = pme->nthread;
2788 if (atc->nthread > 1)
2790 snew(atc->thread_plist, atc->nthread);
2792 snew(atc->spline, atc->nthread);
2793 for (thread = 0; thread < atc->nthread; thread++)
2795 if (atc->nthread > 1)
2797 snew(atc->thread_plist[thread].n, atc->nthread+2*GMX_CACHE_SEP);
2798 atc->thread_plist[thread].n += GMX_CACHE_SEP;
2800 snew(atc->spline[thread].thread_one, pme->nthread);
2801 atc->spline[thread].thread_one[thread] = 1;
2806 init_overlap_comm(pme_overlap_t * ol,
2816 int lbnd, rbnd, maxlr, b, i;
2819 pme_grid_comm_t *pgc;
2821 int fft_start, fft_end, send_index1, recv_index1;
2825 ol->mpi_comm = comm;
2828 ol->nnodes = nnodes;
2829 ol->nodeid = nodeid;
2831 /* Linear translation of the PME grid won't affect reciprocal space
2832 * calculations, so to optimize we only interpolate "upwards",
2833 * which also means we only have to consider overlap in one direction.
2834 * I.e., particles on this node might also be spread to grid indices
2835 * that belong to higher nodes (modulo nnodes)
2838 snew(ol->s2g0, ol->nnodes+1);
2839 snew(ol->s2g1, ol->nnodes);
2842 fprintf(debug, "PME slab boundaries:");
2844 for (i = 0; i < nnodes; i++)
2846 /* s2g0 the local interpolation grid start.
2847 * s2g1 the local interpolation grid end.
2848 * Because grid overlap communication only goes forward,
2849 * the grid the slabs for fft's should be rounded down.
2851 ol->s2g0[i] = ( i *ndata + 0 )/nnodes;
2852 ol->s2g1[i] = ((i+1)*ndata + nnodes-1)/nnodes + norder - 1;
2856 fprintf(debug, " %3d %3d", ol->s2g0[i], ol->s2g1[i]);
2859 ol->s2g0[nnodes] = ndata;
2862 fprintf(debug, "\n");
2865 /* Determine with how many nodes we need to communicate the grid overlap */
2871 for (i = 0; i < nnodes; i++)
2873 if ((i+b < nnodes && ol->s2g1[i] > ol->s2g0[i+b]) ||
2874 (i+b >= nnodes && ol->s2g1[i] > ol->s2g0[i+b-nnodes] + ndata))
2880 while (bCont && b < nnodes);
2881 ol->noverlap_nodes = b - 1;
2883 snew(ol->send_id, ol->noverlap_nodes);
2884 snew(ol->recv_id, ol->noverlap_nodes);
2885 for (b = 0; b < ol->noverlap_nodes; b++)
2887 ol->send_id[b] = (ol->nodeid + (b + 1)) % ol->nnodes;
2888 ol->recv_id[b] = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
2890 snew(ol->comm_data, ol->noverlap_nodes);
2893 for (b = 0; b < ol->noverlap_nodes; b++)
2895 pgc = &ol->comm_data[b];
2897 fft_start = ol->s2g0[ol->send_id[b]];
2898 fft_end = ol->s2g0[ol->send_id[b]+1];
2899 if (ol->send_id[b] < nodeid)
2904 send_index1 = ol->s2g1[nodeid];
2905 send_index1 = min(send_index1, fft_end);
2906 pgc->send_index0 = fft_start;
2907 pgc->send_nindex = max(0, send_index1 - pgc->send_index0);
2908 ol->send_size += pgc->send_nindex;
2910 /* We always start receiving to the first index of our slab */
2911 fft_start = ol->s2g0[ol->nodeid];
2912 fft_end = ol->s2g0[ol->nodeid+1];
2913 recv_index1 = ol->s2g1[ol->recv_id[b]];
2914 if (ol->recv_id[b] > nodeid)
2916 recv_index1 -= ndata;
2918 recv_index1 = min(recv_index1, fft_end);
2919 pgc->recv_index0 = fft_start;
2920 pgc->recv_nindex = max(0, recv_index1 - pgc->recv_index0);
2924 /* Communicate the buffer sizes to receive */
2925 for (b = 0; b < ol->noverlap_nodes; b++)
2927 MPI_Sendrecv(&ol->send_size, 1, MPI_INT, ol->send_id[b], b,
2928 &ol->comm_data[b].recv_size, 1, MPI_INT, ol->recv_id[b], b,
2929 ol->mpi_comm, &stat);
2933 /* For non-divisible grid we need pme_order iso pme_order-1 */
2934 snew(ol->sendbuf, norder*commplainsize);
2935 snew(ol->recvbuf, norder*commplainsize);
2939 make_gridindex5_to_localindex(int n, int local_start, int local_range,
2940 int **global_to_local,
2941 real **fraction_shift)
2949 for (i = 0; (i < 5*n); i++)
2951 /* Determine the global to local grid index */
2952 gtl[i] = (i - local_start + n) % n;
2953 /* For coordinates that fall within the local grid the fraction
2954 * is correct, we don't need to shift it.
2957 if (local_range < n)
2959 /* Due to rounding issues i could be 1 beyond the lower or
2960 * upper boundary of the local grid. Correct the index for this.
2961 * If we shift the index, we need to shift the fraction by
2962 * the same amount in the other direction to not affect
2964 * Note that due to this shifting the weights at the end of
2965 * the spline might change, but that will only involve values
2966 * between zero and values close to the precision of a real,
2967 * which is anyhow the accuracy of the whole mesh calculation.
2969 /* With local_range=0 we should not change i=local_start */
2970 if (i % n != local_start)
2977 else if (gtl[i] == local_range)
2979 gtl[i] = local_range - 1;
2986 *global_to_local = gtl;
2987 *fraction_shift = fsh;
2990 static pme_spline_work_t *make_pme_spline_work(int order)
2992 pme_spline_work_t *work;
2994 #ifdef PME_SIMD4_SPREAD_GATHER
2995 real tmp[12], *tmp_aligned;
2996 gmx_simd4_pr zero_S;
2997 gmx_simd4_pr real_mask_S0, real_mask_S1;
3000 snew_aligned(work, 1, SIMD4_ALIGNMENT);
3002 tmp_aligned = gmx_simd4_align_real(tmp);
3004 zero_S = gmx_simd4_setzero_pr();
3006 /* Generate bit masks to mask out the unused grid entries,
3007 * as we only operate on order of the 8 grid entries that are
3008 * load into 2 SIMD registers.
3010 for (of = 0; of < 8-(order-1); of++)
3012 for (i = 0; i < 8; i++)
3014 tmp_aligned[i] = (i >= of && i < of+order ? -1.0 : 1.0);
3016 real_mask_S0 = gmx_simd4_load_pr(tmp_aligned);
3017 real_mask_S1 = gmx_simd4_load_pr(tmp_aligned+4);
3018 work->mask_S0[of] = gmx_simd4_cmplt_pr(real_mask_S0, zero_S);
3019 work->mask_S1[of] = gmx_simd4_cmplt_pr(real_mask_S1, zero_S);
3028 void gmx_pme_check_restrictions(int pme_order,
3029 int nkx, int nky, int nkz,
3032 gmx_bool bUseThreads,
3034 gmx_bool *bValidSettings)
3036 if (pme_order > PME_ORDER_MAX)
3040 *bValidSettings = FALSE;
3043 gmx_fatal(FARGS, "pme_order (%d) is larger than the maximum allowed value (%d). Modify and recompile the code if you really need such a high order.",
3044 pme_order, PME_ORDER_MAX);
3047 if (nkx <= pme_order*(nnodes_major > 1 ? 2 : 1) ||
3048 nky <= pme_order*(nnodes_minor > 1 ? 2 : 1) ||
3053 *bValidSettings = FALSE;
3056 gmx_fatal(FARGS, "The PME grid sizes need to be larger than pme_order (%d) and for dimensions with domain decomposition larger than 2*pme_order",
3060 /* Check for a limitation of the (current) sum_fftgrid_dd code.
3061 * We only allow multiple communication pulses in dim 1, not in dim 0.
3063 if (bUseThreads && (nkx < nnodes_major*pme_order &&
3064 nkx != nnodes_major*(pme_order - 1)))
3068 *bValidSettings = FALSE;
3071 gmx_fatal(FARGS, "The number of PME grid lines per node along x is %g. But when using OpenMP threads, the number of grid lines per node along x should be >= pme_order (%d) or = pmeorder-1. To resolve this issue, use less nodes along x (and possibly more along y and/or z) by specifying -dd manually.",
3072 nkx/(double)nnodes_major, pme_order);
3075 if (bValidSettings != NULL)
3077 *bValidSettings = TRUE;
3083 int gmx_pme_init(gmx_pme_t * pmedata,
3089 gmx_bool bFreeEnergy,
3090 gmx_bool bReproducible,
3093 gmx_pme_t pme = NULL;
3095 int use_threads, sum_use_threads;
3100 fprintf(debug, "Creating PME data structures.\n");
3104 pme->redist_init = FALSE;
3105 pme->sum_qgrid_tmp = NULL;
3106 pme->sum_qgrid_dd_tmp = NULL;
3107 pme->buf_nalloc = 0;
3108 pme->redist_buf_nalloc = 0;
3111 pme->bPPnode = TRUE;
3113 pme->nnodes_major = nnodes_major;
3114 pme->nnodes_minor = nnodes_minor;
3117 if (nnodes_major*nnodes_minor > 1)
3119 pme->mpi_comm = cr->mpi_comm_mygroup;
3121 MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
3122 MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
3123 if (pme->nnodes != nnodes_major*nnodes_minor)
3125 gmx_incons("PME node count mismatch");
3130 pme->mpi_comm = MPI_COMM_NULL;
3134 if (pme->nnodes == 1)
3137 pme->mpi_comm_d[0] = MPI_COMM_NULL;
3138 pme->mpi_comm_d[1] = MPI_COMM_NULL;
3140 pme->ndecompdim = 0;
3141 pme->nodeid_major = 0;
3142 pme->nodeid_minor = 0;
3144 pme->mpi_comm_d[0] = pme->mpi_comm_d[1] = MPI_COMM_NULL;
3149 if (nnodes_minor == 1)
3152 pme->mpi_comm_d[0] = pme->mpi_comm;
3153 pme->mpi_comm_d[1] = MPI_COMM_NULL;
3155 pme->ndecompdim = 1;
3156 pme->nodeid_major = pme->nodeid;
3157 pme->nodeid_minor = 0;
3160 else if (nnodes_major == 1)
3163 pme->mpi_comm_d[0] = MPI_COMM_NULL;
3164 pme->mpi_comm_d[1] = pme->mpi_comm;
3166 pme->ndecompdim = 1;
3167 pme->nodeid_major = 0;
3168 pme->nodeid_minor = pme->nodeid;
3172 if (pme->nnodes % nnodes_major != 0)
3174 gmx_incons("For 2D PME decomposition, #PME nodes must be divisible by the number of nodes in the major dimension");
3176 pme->ndecompdim = 2;
3179 MPI_Comm_split(pme->mpi_comm, pme->nodeid % nnodes_minor,
3180 pme->nodeid, &pme->mpi_comm_d[0]); /* My communicator along major dimension */
3181 MPI_Comm_split(pme->mpi_comm, pme->nodeid/nnodes_minor,
3182 pme->nodeid, &pme->mpi_comm_d[1]); /* My communicator along minor dimension */
3184 MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
3185 MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
3186 MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
3187 MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
3190 pme->bPPnode = (cr->duty & DUTY_PP);
3193 pme->nthread = nthread;
3195 /* Check if any of the PME MPI ranks uses threads */
3196 use_threads = (pme->nthread > 1 ? 1 : 0);
3198 if (pme->nnodes > 1)
3200 MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT,
3201 MPI_SUM, pme->mpi_comm);
3206 sum_use_threads = use_threads;
3208 pme->bUseThreads = (sum_use_threads > 0);
3210 if (ir->ePBC == epbcSCREW)
3212 gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
3215 pme->bFEP = ((ir->efep != efepNO) && bFreeEnergy);
3219 pme->bP3M = (ir->coulombtype == eelP3M_AD || getenv("GMX_PME_P3M") != NULL);
3220 pme->pme_order = ir->pme_order;
3221 pme->epsilon_r = ir->epsilon_r;
3223 /* If we violate restrictions, generate a fatal error here */
3224 gmx_pme_check_restrictions(pme->pme_order,
3225 pme->nkx, pme->nky, pme->nkz,
3232 if (pme->nnodes > 1)
3237 MPI_Type_contiguous(DIM, mpi_type, &(pme->rvec_mpi));
3238 MPI_Type_commit(&(pme->rvec_mpi));
3241 /* Note that the charge spreading and force gathering, which usually
3242 * takes about the same amount of time as FFT+solve_pme,
3243 * is always fully load balanced
3244 * (unless the charge distribution is inhomogeneous).
3247 imbal = pme_load_imbalance(pme);
3248 if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
3252 "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
3253 " For optimal PME load balancing\n"
3254 " PME grid_x (%d) and grid_y (%d) should be divisible by #PME_nodes_x (%d)\n"
3255 " and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_nodes_y (%d)\n"
3257 (int)((imbal-1)*100 + 0.5),
3258 pme->nkx, pme->nky, pme->nnodes_major,
3259 pme->nky, pme->nkz, pme->nnodes_minor);
3263 /* For non-divisible grid we need pme_order iso pme_order-1 */
3264 /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
3265 * y is always copied through a buffer: we don't need padding in z,
3266 * but we do need the overlap in x because of the communication order.
3268 init_overlap_comm(&pme->overlap[0], pme->pme_order,
3272 pme->nnodes_major, pme->nodeid_major,
3274 (div_round_up(pme->nky, pme->nnodes_minor)+pme->pme_order)*(pme->nkz+pme->pme_order-1));
3276 /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
3277 * We do this with an offset buffer of equal size, so we need to allocate
3278 * extra for the offset. That's what the (+1)*pme->nkz is for.
3280 init_overlap_comm(&pme->overlap[1], pme->pme_order,
3284 pme->nnodes_minor, pme->nodeid_minor,
3286 (div_round_up(pme->nkx, pme->nnodes_major)+pme->pme_order+1)*pme->nkz);
3288 /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
3289 * Note that gmx_pme_check_restrictions checked for this already.
3291 if (pme->bUseThreads && pme->overlap[0].noverlap_nodes > 1)
3293 gmx_incons("More than one communication pulse required for grid overlap communication along the major dimension while using threads");
3296 snew(pme->bsp_mod[XX], pme->nkx);
3297 snew(pme->bsp_mod[YY], pme->nky);
3298 snew(pme->bsp_mod[ZZ], pme->nkz);
3300 /* The required size of the interpolation grid, including overlap.
3301 * The allocated size (pmegrid_n?) might be slightly larger.
3303 pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] -
3304 pme->overlap[0].s2g0[pme->nodeid_major];
3305 pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] -
3306 pme->overlap[1].s2g0[pme->nodeid_minor];
3307 pme->pmegrid_nz_base = pme->nkz;
3308 pme->pmegrid_nz = pme->pmegrid_nz_base + pme->pme_order - 1;
3309 set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
3311 pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
3312 pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
3313 pme->pmegrid_start_iz = 0;
3315 make_gridindex5_to_localindex(pme->nkx,
3316 pme->pmegrid_start_ix,
3317 pme->pmegrid_nx - (pme->pme_order-1),
3318 &pme->nnx, &pme->fshx);
3319 make_gridindex5_to_localindex(pme->nky,
3320 pme->pmegrid_start_iy,
3321 pme->pmegrid_ny - (pme->pme_order-1),
3322 &pme->nny, &pme->fshy);
3323 make_gridindex5_to_localindex(pme->nkz,
3324 pme->pmegrid_start_iz,
3325 pme->pmegrid_nz_base,
3326 &pme->nnz, &pme->fshz);
3328 pmegrids_init(&pme->pmegridA,
3329 pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
3330 pme->pmegrid_nz_base,
3334 pme->overlap[0].s2g1[pme->nodeid_major]-pme->overlap[0].s2g0[pme->nodeid_major+1],
3335 pme->overlap[1].s2g1[pme->nodeid_minor]-pme->overlap[1].s2g0[pme->nodeid_minor+1]);
3337 pme->spline_work = make_pme_spline_work(pme->pme_order);
3339 ndata[0] = pme->nkx;
3340 ndata[1] = pme->nky;
3341 ndata[2] = pme->nkz;
3343 /* This routine will allocate the grid data to fit the FFTs */
3344 gmx_parallel_3dfft_init(&pme->pfft_setupA, ndata,
3345 &pme->fftgridA, &pme->cfftgridA,
3347 bReproducible, pme->nthread);
3351 pmegrids_init(&pme->pmegridB,
3352 pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
3353 pme->pmegrid_nz_base,
3357 pme->nkx % pme->nnodes_major != 0,
3358 pme->nky % pme->nnodes_minor != 0);
3360 gmx_parallel_3dfft_init(&pme->pfft_setupB, ndata,
3361 &pme->fftgridB, &pme->cfftgridB,
3363 bReproducible, pme->nthread);
3367 pme->pmegridB.grid.grid = NULL;
3368 pme->fftgridB = NULL;
3369 pme->cfftgridB = NULL;
3374 /* Use plain SPME B-spline interpolation */
3375 make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
3379 /* Use the P3M grid-optimized influence function */
3380 make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
3383 /* Use atc[0] for spreading */
3384 init_atomcomm(pme, &pme->atc[0], nnodes_major > 1 ? 0 : 1, TRUE);
3385 if (pme->ndecompdim >= 2)
3387 init_atomcomm(pme, &pme->atc[1], 1, FALSE);
3390 if (pme->nnodes == 1)
3392 pme->atc[0].n = homenr;
3393 pme_realloc_atomcomm_things(&pme->atc[0]);
3399 /* Use fft5d, order after FFT is y major, z, x minor */
3401 snew(pme->work, pme->nthread);
3402 for (thread = 0; thread < pme->nthread; thread++)
3404 realloc_work(&pme->work[thread], pme->nkx);
3413 static void reuse_pmegrids(const pmegrids_t *old, pmegrids_t *new)
3417 for (d = 0; d < DIM; d++)
3419 if (new->grid.n[d] > old->grid.n[d])
3425 sfree_aligned(new->grid.grid);
3426 new->grid.grid = old->grid.grid;
3428 if (new->grid_th != NULL && new->nthread == old->nthread)
3430 sfree_aligned(new->grid_all);
3431 for (t = 0; t < new->nthread; t++)
3433 new->grid_th[t].grid = old->grid_th[t].grid;
3438 int gmx_pme_reinit(gmx_pme_t * pmedata,
3441 const t_inputrec * ir,
3449 irc.nkx = grid_size[XX];
3450 irc.nky = grid_size[YY];
3451 irc.nkz = grid_size[ZZ];
3453 if (pme_src->nnodes == 1)
3455 homenr = pme_src->atc[0].n;
3462 ret = gmx_pme_init(pmedata, cr, pme_src->nnodes_major, pme_src->nnodes_minor,
3463 &irc, homenr, pme_src->bFEP, FALSE, pme_src->nthread);
3467 /* We can easily reuse the allocated pme grids in pme_src */
3468 reuse_pmegrids(&pme_src->pmegridA, &(*pmedata)->pmegridA);
3469 /* We would like to reuse the fft grids, but that's harder */
3476 static void copy_local_grid(gmx_pme_t pme,
3477 pmegrids_t *pmegrids, int thread, real *fftgrid)
3479 ivec local_fft_ndata, local_fft_offset, local_fft_size;
3483 int offx, offy, offz, x, y, z, i0, i0t;
3488 gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
3492 fft_my = local_fft_size[YY];
3493 fft_mz = local_fft_size[ZZ];
3495 pmegrid = &pmegrids->grid_th[thread];
3497 nsx = pmegrid->s[XX];
3498 nsy = pmegrid->s[YY];
3499 nsz = pmegrid->s[ZZ];
3501 for (d = 0; d < DIM; d++)
3503 nf[d] = min(pmegrid->n[d] - (pmegrid->order - 1),
3504 local_fft_ndata[d] - pmegrid->offset[d]);
3507 offx = pmegrid->offset[XX];
3508 offy = pmegrid->offset[YY];
3509 offz = pmegrid->offset[ZZ];
3511 /* Directly copy the non-overlapping parts of the local grids.
3512 * This also initializes the full grid.
3514 grid_th = pmegrid->grid;
3515 for (x = 0; x < nf[XX]; x++)
3517 for (y = 0; y < nf[YY]; y++)
3519 i0 = ((offx + x)*fft_my + (offy + y))*fft_mz + offz;
3520 i0t = (x*nsy + y)*nsz;
3521 for (z = 0; z < nf[ZZ]; z++)
3523 fftgrid[i0+z] = grid_th[i0t+z];
3530 reduce_threadgrid_overlap(gmx_pme_t pme,
3531 const pmegrids_t *pmegrids, int thread,
3532 real *fftgrid, real *commbuf_x, real *commbuf_y)
3534 ivec local_fft_ndata, local_fft_offset, local_fft_size;
3535 int fft_nx, fft_ny, fft_nz;
3540 int offx, offy, offz, x, y, z, i0, i0t;
3541 int sx, sy, sz, fx, fy, fz, tx1, ty1, tz1, ox, oy, oz;
3542 gmx_bool bClearBufX, bClearBufY, bClearBufXY, bClearBuf;
3543 gmx_bool bCommX, bCommY;
3546 const pmegrid_t *pmegrid, *pmegrid_g, *pmegrid_f;
3547 const real *grid_th;
3548 real *commbuf = NULL;
3550 gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
3554 fft_nx = local_fft_ndata[XX];
3555 fft_ny = local_fft_ndata[YY];
3556 fft_nz = local_fft_ndata[ZZ];
3558 fft_my = local_fft_size[YY];
3559 fft_mz = local_fft_size[ZZ];
3561 /* This routine is called when all thread have finished spreading.
3562 * Here each thread sums grid contributions calculated by other threads
3563 * to the thread local grid volume.
3564 * To minimize the number of grid copying operations,
3565 * this routines sums immediately from the pmegrid to the fftgrid.
3568 /* Determine which part of the full node grid we should operate on,
3569 * this is our thread local part of the full grid.
3571 pmegrid = &pmegrids->grid_th[thread];
3573 for (d = 0; d < DIM; d++)
3575 ne[d] = min(pmegrid->offset[d]+pmegrid->n[d]-(pmegrid->order-1),
3576 local_fft_ndata[d]);
3579 offx = pmegrid->offset[XX];
3580 offy = pmegrid->offset[YY];
3581 offz = pmegrid->offset[ZZ];
3588 /* Now loop over all the thread data blocks that contribute
3589 * to the grid region we (our thread) are operating on.
3591 /* Note that ffy_nx/y is equal to the number of grid points
3592 * between the first point of our node grid and the one of the next node.
3594 for (sx = 0; sx >= -pmegrids->nthread_comm[XX]; sx--)
3596 fx = pmegrid->ci[XX] + sx;
3601 fx += pmegrids->nc[XX];
3603 bCommX = (pme->nnodes_major > 1);
3605 pmegrid_g = &pmegrids->grid_th[fx*pmegrids->nc[YY]*pmegrids->nc[ZZ]];
3606 ox += pmegrid_g->offset[XX];
3609 tx1 = min(ox + pmegrid_g->n[XX], ne[XX]);
3613 tx1 = min(ox + pmegrid_g->n[XX], pme->pme_order);
3616 for (sy = 0; sy >= -pmegrids->nthread_comm[YY]; sy--)
3618 fy = pmegrid->ci[YY] + sy;
3623 fy += pmegrids->nc[YY];
3625 bCommY = (pme->nnodes_minor > 1);
3627 pmegrid_g = &pmegrids->grid_th[fy*pmegrids->nc[ZZ]];
3628 oy += pmegrid_g->offset[YY];
3631 ty1 = min(oy + pmegrid_g->n[YY], ne[YY]);
3635 ty1 = min(oy + pmegrid_g->n[YY], pme->pme_order);
3638 for (sz = 0; sz >= -pmegrids->nthread_comm[ZZ]; sz--)
3640 fz = pmegrid->ci[ZZ] + sz;
3644 fz += pmegrids->nc[ZZ];
3647 pmegrid_g = &pmegrids->grid_th[fz];
3648 oz += pmegrid_g->offset[ZZ];
3649 tz1 = min(oz + pmegrid_g->n[ZZ], ne[ZZ]);
3651 if (sx == 0 && sy == 0 && sz == 0)
3653 /* We have already added our local contribution
3654 * before calling this routine, so skip it here.
3659 thread_f = (fx*pmegrids->nc[YY] + fy)*pmegrids->nc[ZZ] + fz;
3661 pmegrid_f = &pmegrids->grid_th[thread_f];
3663 grid_th = pmegrid_f->grid;
3665 nsx = pmegrid_f->s[XX];
3666 nsy = pmegrid_f->s[YY];
3667 nsz = pmegrid_f->s[ZZ];
3669 #ifdef DEBUG_PME_REDUCE
3670 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",
3671 pme->nodeid, thread, thread_f,
3672 pme->pmegrid_start_ix,
3673 pme->pmegrid_start_iy,
3674 pme->pmegrid_start_iz,
3676 offx-ox, tx1-ox, offx, tx1,
3677 offy-oy, ty1-oy, offy, ty1,
3678 offz-oz, tz1-oz, offz, tz1);
3681 if (!(bCommX || bCommY))
3683 /* Copy from the thread local grid to the node grid */
3684 for (x = offx; x < tx1; x++)
3686 for (y = offy; y < ty1; y++)
3688 i0 = (x*fft_my + y)*fft_mz;
3689 i0t = ((x - ox)*nsy + (y - oy))*nsz - oz;
3690 for (z = offz; z < tz1; z++)
3692 fftgrid[i0+z] += grid_th[i0t+z];
3699 /* The order of this conditional decides
3700 * where the corner volume gets stored with x+y decomp.
3704 commbuf = commbuf_y;
3705 buf_my = ty1 - offy;
3708 /* We index commbuf modulo the local grid size */
3709 commbuf += buf_my*fft_nx*fft_nz;
3711 bClearBuf = bClearBufXY;
3712 bClearBufXY = FALSE;
3716 bClearBuf = bClearBufY;
3722 commbuf = commbuf_x;
3724 bClearBuf = bClearBufX;
3728 /* Copy to the communication buffer */
3729 for (x = offx; x < tx1; x++)
3731 for (y = offy; y < ty1; y++)
3733 i0 = (x*buf_my + y)*fft_nz;
3734 i0t = ((x - ox)*nsy + (y - oy))*nsz - oz;
3738 /* First access of commbuf, initialize it */
3739 for (z = offz; z < tz1; z++)
3741 commbuf[i0+z] = grid_th[i0t+z];
3746 for (z = offz; z < tz1; z++)
3748 commbuf[i0+z] += grid_th[i0t+z];
3760 static void sum_fftgrid_dd(gmx_pme_t pme, real *fftgrid)
3762 ivec local_fft_ndata, local_fft_offset, local_fft_size;
3763 pme_overlap_t *overlap;
3764 int send_index0, send_nindex;
3769 int send_size_y, recv_size_y;
3770 int ipulse, send_id, recv_id, datasize, gridsize, size_yx;
3771 real *sendptr, *recvptr;
3772 int x, y, z, indg, indb;
3774 /* Note that this routine is only used for forward communication.
3775 * Since the force gathering, unlike the charge spreading,
3776 * can be trivially parallelized over the particles,
3777 * the backwards process is much simpler and can use the "old"
3778 * communication setup.
3781 gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
3786 if (pme->nnodes_minor > 1)
3788 /* Major dimension */
3789 overlap = &pme->overlap[1];
3791 if (pme->nnodes_major > 1)
3793 size_yx = pme->overlap[0].comm_data[0].send_nindex;
3799 datasize = (local_fft_ndata[XX] + size_yx)*local_fft_ndata[ZZ];
3801 send_size_y = overlap->send_size;
3803 for (ipulse = 0; ipulse < overlap->noverlap_nodes; ipulse++)
3805 send_id = overlap->send_id[ipulse];
3806 recv_id = overlap->recv_id[ipulse];
3808 overlap->comm_data[ipulse].send_index0 -
3809 overlap->comm_data[0].send_index0;
3810 send_nindex = overlap->comm_data[ipulse].send_nindex;
3811 /* We don't use recv_index0, as we always receive starting at 0 */
3812 recv_nindex = overlap->comm_data[ipulse].recv_nindex;
3813 recv_size_y = overlap->comm_data[ipulse].recv_size;
3815 sendptr = overlap->sendbuf + send_index0*local_fft_ndata[ZZ];
3816 recvptr = overlap->recvbuf;
3819 MPI_Sendrecv(sendptr, send_size_y*datasize, GMX_MPI_REAL,
3821 recvptr, recv_size_y*datasize, GMX_MPI_REAL,
3823 overlap->mpi_comm, &stat);
3826 for (x = 0; x < local_fft_ndata[XX]; x++)
3828 for (y = 0; y < recv_nindex; y++)
3830 indg = (x*local_fft_size[YY] + y)*local_fft_size[ZZ];
3831 indb = (x*recv_size_y + y)*local_fft_ndata[ZZ];
3832 for (z = 0; z < local_fft_ndata[ZZ]; z++)
3834 fftgrid[indg+z] += recvptr[indb+z];
3839 if (pme->nnodes_major > 1)
3841 /* Copy from the received buffer to the send buffer for dim 0 */
3842 sendptr = pme->overlap[0].sendbuf;
3843 for (x = 0; x < size_yx; x++)
3845 for (y = 0; y < recv_nindex; y++)
3847 indg = (x*local_fft_ndata[YY] + y)*local_fft_ndata[ZZ];
3848 indb = ((local_fft_ndata[XX] + x)*recv_size_y + y)*local_fft_ndata[ZZ];
3849 for (z = 0; z < local_fft_ndata[ZZ]; z++)
3851 sendptr[indg+z] += recvptr[indb+z];
3859 /* We only support a single pulse here.
3860 * This is not a severe limitation, as this code is only used
3861 * with OpenMP and with OpenMP the (PME) domains can be larger.
3863 if (pme->nnodes_major > 1)
3865 /* Major dimension */
3866 overlap = &pme->overlap[0];
3868 datasize = local_fft_ndata[YY]*local_fft_ndata[ZZ];
3869 gridsize = local_fft_size[YY] *local_fft_size[ZZ];
3873 send_id = overlap->send_id[ipulse];
3874 recv_id = overlap->recv_id[ipulse];
3875 send_nindex = overlap->comm_data[ipulse].send_nindex;
3876 /* We don't use recv_index0, as we always receive starting at 0 */
3877 recv_nindex = overlap->comm_data[ipulse].recv_nindex;
3879 sendptr = overlap->sendbuf;
3880 recvptr = overlap->recvbuf;
3884 fprintf(debug, "PME fftgrid comm %2d x %2d x %2d\n",
3885 send_nindex, local_fft_ndata[YY], local_fft_ndata[ZZ]);
3889 MPI_Sendrecv(sendptr, send_nindex*datasize, GMX_MPI_REAL,
3891 recvptr, recv_nindex*datasize, GMX_MPI_REAL,
3893 overlap->mpi_comm, &stat);
3896 for (x = 0; x < recv_nindex; x++)
3898 for (y = 0; y < local_fft_ndata[YY]; y++)
3900 indg = (x*local_fft_size[YY] + y)*local_fft_size[ZZ];
3901 indb = (x*local_fft_ndata[YY] + y)*local_fft_ndata[ZZ];
3902 for (z = 0; z < local_fft_ndata[ZZ]; z++)
3904 fftgrid[indg+z] += recvptr[indb+z];
3912 static void spread_on_grid(gmx_pme_t pme,
3913 pme_atomcomm_t *atc, pmegrids_t *grids,
3914 gmx_bool bCalcSplines, gmx_bool bSpread,
3917 int nthread, thread;
3918 #ifdef PME_TIME_THREADS
3919 gmx_cycles_t c1, c2, c3, ct1a, ct1b, ct1c;
3920 static double cs1 = 0, cs2 = 0, cs3 = 0;
3921 static double cs1a[6] = {0, 0, 0, 0, 0, 0};
3925 nthread = pme->nthread;
3926 assert(nthread > 0);
3928 #ifdef PME_TIME_THREADS
3929 c1 = omp_cyc_start();
3933 #pragma omp parallel for num_threads(nthread) schedule(static)
3934 for (thread = 0; thread < nthread; thread++)
3938 start = atc->n* thread /nthread;
3939 end = atc->n*(thread+1)/nthread;
3941 /* Compute fftgrid index for all atoms,
3942 * with help of some extra variables.
3944 calc_interpolation_idx(pme, atc, start, end, thread);
3947 #ifdef PME_TIME_THREADS
3948 c1 = omp_cyc_end(c1);
3952 #ifdef PME_TIME_THREADS
3953 c2 = omp_cyc_start();
3955 #pragma omp parallel for num_threads(nthread) schedule(static)
3956 for (thread = 0; thread < nthread; thread++)
3958 splinedata_t *spline;
3959 pmegrid_t *grid = NULL;
3961 /* make local bsplines */
3962 if (grids == NULL || !pme->bUseThreads)
3964 spline = &atc->spline[0];
3970 grid = &grids->grid;
3975 spline = &atc->spline[thread];
3977 if (grids->nthread == 1)
3979 /* One thread, we operate on all charges */
3984 /* Get the indices our thread should operate on */
3985 make_thread_local_ind(atc, thread, spline);
3988 grid = &grids->grid_th[thread];
3993 make_bsplines(spline->theta, spline->dtheta, pme->pme_order,
3994 atc->fractx, spline->n, spline->ind, atc->q, pme->bFEP);
3999 /* put local atoms on grid. */
4000 #ifdef PME_TIME_SPREAD
4001 ct1a = omp_cyc_start();
4003 spread_q_bsplines_thread(grid, atc, spline, pme->spline_work);
4005 if (pme->bUseThreads)
4007 copy_local_grid(pme, grids, thread, fftgrid);
4009 #ifdef PME_TIME_SPREAD
4010 ct1a = omp_cyc_end(ct1a);
4011 cs1a[thread] += (double)ct1a;
4015 #ifdef PME_TIME_THREADS
4016 c2 = omp_cyc_end(c2);
4020 if (bSpread && pme->bUseThreads)
4022 #ifdef PME_TIME_THREADS
4023 c3 = omp_cyc_start();
4025 #pragma omp parallel for num_threads(grids->nthread) schedule(static)
4026 for (thread = 0; thread < grids->nthread; thread++)
4028 reduce_threadgrid_overlap(pme, grids, thread,
4030 pme->overlap[0].sendbuf,
4031 pme->overlap[1].sendbuf);
4033 #ifdef PME_TIME_THREADS
4034 c3 = omp_cyc_end(c3);
4038 if (pme->nnodes > 1)
4040 /* Communicate the overlapping part of the fftgrid.
4041 * For this communication call we need to check pme->bUseThreads
4042 * to have all ranks communicate here, regardless of pme->nthread.
4044 sum_fftgrid_dd(pme, fftgrid);
4048 #ifdef PME_TIME_THREADS
4052 printf("idx %.2f spread %.2f red %.2f",
4053 cs1*1e-9, cs2*1e-9, cs3*1e-9);
4054 #ifdef PME_TIME_SPREAD
4055 for (thread = 0; thread < nthread; thread++)
4057 printf(" %.2f", cs1a[thread]*1e-9);
4066 static void dump_grid(FILE *fp,
4067 int sx, int sy, int sz, int nx, int ny, int nz,
4068 int my, int mz, const real *g)
4072 for (x = 0; x < nx; x++)
4074 for (y = 0; y < ny; y++)
4076 for (z = 0; z < nz; z++)
4078 fprintf(fp, "%2d %2d %2d %6.3f\n",
4079 sx+x, sy+y, sz+z, g[(x*my + y)*mz + z]);
4085 static void dump_local_fftgrid(gmx_pme_t pme, const real *fftgrid)
4087 ivec local_fft_ndata, local_fft_offset, local_fft_size;
4089 gmx_parallel_3dfft_real_limits(pme->pfft_setupA,
4095 pme->pmegrid_start_ix,
4096 pme->pmegrid_start_iy,
4097 pme->pmegrid_start_iz,
4098 pme->pmegrid_nx-pme->pme_order+1,
4099 pme->pmegrid_ny-pme->pme_order+1,
4100 pme->pmegrid_nz-pme->pme_order+1,
4107 void gmx_pme_calc_energy(gmx_pme_t pme, int n, rvec *x, real *q, real *V)
4109 pme_atomcomm_t *atc;
4112 if (pme->nnodes > 1)
4114 gmx_incons("gmx_pme_calc_energy called in parallel");
4118 gmx_incons("gmx_pme_calc_energy with free energy");
4121 atc = &pme->atc_energy;
4123 if (atc->spline == NULL)
4125 snew(atc->spline, atc->nthread);
4128 atc->bSpread = TRUE;
4129 atc->pme_order = pme->pme_order;
4131 pme_realloc_atomcomm_things(atc);
4135 /* We only use the A-charges grid */
4136 grid = &pme->pmegridA;
4138 /* Only calculate the spline coefficients, don't actually spread */
4139 spread_on_grid(pme, atc, NULL, TRUE, FALSE, pme->fftgridA);
4141 *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
4145 static void reset_pmeonly_counters(gmx_wallcycle_t wcycle,
4146 gmx_walltime_accounting_t walltime_accounting,
4147 t_nrnb *nrnb, t_inputrec *ir,
4148 gmx_large_int_t step)
4150 /* Reset all the counters related to performance over the run */
4151 wallcycle_stop(wcycle, ewcRUN);
4152 wallcycle_reset_all(wcycle);
4154 if (ir->nsteps >= 0)
4156 /* ir->nsteps is not used here, but we update it for consistency */
4157 ir->nsteps -= step - ir->init_step;
4159 ir->init_step = step;
4160 wallcycle_start(wcycle, ewcRUN);
4161 walltime_accounting_start(walltime_accounting);
4165 static void gmx_pmeonly_switch(int *npmedata, gmx_pme_t **pmedata,
4167 t_commrec *cr, t_inputrec *ir,
4171 gmx_pme_t pme = NULL;
4174 while (ind < *npmedata)
4176 pme = (*pmedata)[ind];
4177 if (pme->nkx == grid_size[XX] &&
4178 pme->nky == grid_size[YY] &&
4179 pme->nkz == grid_size[ZZ])
4190 srenew(*pmedata, *npmedata);
4192 /* Generate a new PME data structure, copying part of the old pointers */
4193 gmx_pme_reinit(&((*pmedata)[ind]), cr, pme, ir, grid_size);
4195 *pme_ret = (*pmedata)[ind];
4199 int gmx_pmeonly(gmx_pme_t pme,
4200 t_commrec *cr, t_nrnb *nrnb,
4201 gmx_wallcycle_t wcycle,
4202 gmx_walltime_accounting_t walltime_accounting,
4208 gmx_pme_pp_t pme_pp;
4212 rvec *x_pp = NULL, *f_pp = NULL;
4213 real *chargeA = NULL, *chargeB = NULL;
4215 int maxshift_x = 0, maxshift_y = 0;
4216 real energy, dvdlambda;
4221 gmx_large_int_t step, step_rel;
4224 /* This data will only use with PME tuning, i.e. switching PME grids */
4226 snew(pmedata, npmedata);
4229 pme_pp = gmx_pme_pp_init(cr);
4234 do /****** this is a quasi-loop over time steps! */
4236 /* The reason for having a loop here is PME grid tuning/switching */
4239 /* Domain decomposition */
4240 ret = gmx_pme_recv_q_x(pme_pp,
4242 &chargeA, &chargeB, box, &x_pp, &f_pp,
4243 &maxshift_x, &maxshift_y,
4244 &pme->bFEP, &lambda,
4247 grid_switch, &ewaldcoeff);
4249 if (ret == pmerecvqxSWITCHGRID)
4251 /* Switch the PME grid to grid_switch */
4252 gmx_pmeonly_switch(&npmedata, &pmedata, grid_switch, cr, ir, &pme);
4255 if (ret == pmerecvqxRESETCOUNTERS)
4257 /* Reset the cycle and flop counters */
4258 reset_pmeonly_counters(wcycle, walltime_accounting, nrnb, ir, step);
4261 while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
4263 if (ret == pmerecvqxFINISH)
4265 /* We should stop: break out of the loop */
4269 step_rel = step - ir->init_step;
4273 wallcycle_start(wcycle, ewcRUN);
4274 walltime_accounting_start(walltime_accounting);
4277 wallcycle_start(wcycle, ewcPMEMESH);
4281 gmx_pme_do(pme, 0, natoms, x_pp, f_pp, chargeA, chargeB, box,
4282 cr, maxshift_x, maxshift_y, nrnb, wcycle, vir, ewaldcoeff,
4283 &energy, lambda, &dvdlambda,
4284 GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0));
4286 cycles = wallcycle_stop(wcycle, ewcPMEMESH);
4288 gmx_pme_send_force_vir_ener(pme_pp,
4289 f_pp, vir, energy, dvdlambda,
4293 } /***** end of quasi-loop, we stop with the break above */
4296 walltime_accounting_end(walltime_accounting);
4301 int gmx_pme_do(gmx_pme_t pme,
4302 int start, int homenr,
4304 real *chargeA, real *chargeB,
4305 matrix box, t_commrec *cr,
4306 int maxshift_x, int maxshift_y,
4307 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
4308 matrix vir, real ewaldcoeff,
4309 real *energy, real lambda,
4310 real *dvdlambda, int flags)
4312 int q, d, i, j, ntot, npme;
4315 pme_atomcomm_t *atc = NULL;
4316 pmegrids_t *pmegrid = NULL;
4320 real *charge = NULL, *q_d;
4324 gmx_parallel_3dfft_t pfft_setup;
4326 t_complex * cfftgrid;
4328 const gmx_bool bCalcEnerVir = flags & GMX_PME_CALC_ENER_VIR;
4329 const gmx_bool bCalcF = flags & GMX_PME_CALC_F;
4331 assert(pme->nnodes > 0);
4332 assert(pme->nnodes == 1 || pme->ndecompdim > 0);
4334 if (pme->nnodes > 1)
4338 if (atc->npd > atc->pd_nalloc)
4340 atc->pd_nalloc = over_alloc_dd(atc->npd);
4341 srenew(atc->pd, atc->pd_nalloc);
4343 atc->maxshift = (atc->dimind == 0 ? maxshift_x : maxshift_y);
4347 /* This could be necessary for TPI */
4348 pme->atc[0].n = homenr;
4351 for (q = 0; q < (pme->bFEP ? 2 : 1); q++)
4355 pmegrid = &pme->pmegridA;
4356 fftgrid = pme->fftgridA;
4357 cfftgrid = pme->cfftgridA;
4358 pfft_setup = pme->pfft_setupA;
4359 charge = chargeA+start;
4363 pmegrid = &pme->pmegridB;
4364 fftgrid = pme->fftgridB;
4365 cfftgrid = pme->cfftgridB;
4366 pfft_setup = pme->pfft_setupB;
4367 charge = chargeB+start;
4369 grid = pmegrid->grid.grid;
4370 /* Unpack structure */
4373 fprintf(debug, "PME: nnodes = %d, nodeid = %d\n",
4374 cr->nnodes, cr->nodeid);
4375 fprintf(debug, "Grid = %p\n", (void*)grid);
4378 gmx_fatal(FARGS, "No grid!");
4383 m_inv_ur0(box, pme->recipbox);
4385 if (pme->nnodes == 1)
4388 if (DOMAINDECOMP(cr))
4391 pme_realloc_atomcomm_things(atc);
4399 wallcycle_start(wcycle, ewcPME_REDISTXF);
4400 for (d = pme->ndecompdim-1; d >= 0; d--)
4402 if (d == pme->ndecompdim-1)
4410 n_d = pme->atc[d+1].n;
4416 if (atc->npd > atc->pd_nalloc)
4418 atc->pd_nalloc = over_alloc_dd(atc->npd);
4419 srenew(atc->pd, atc->pd_nalloc);
4421 atc->maxshift = (atc->dimind == 0 ? maxshift_x : maxshift_y);
4422 pme_calc_pidx_wrapper(n_d, pme->recipbox, x_d, atc);
4425 /* Redistribute x (only once) and qA or qB */
4426 if (DOMAINDECOMP(cr))
4428 dd_pmeredist_x_q(pme, n_d, q == 0, x_d, q_d, atc);
4432 pmeredist_pd(pme, TRUE, n_d, q == 0, x_d, q_d, atc);
4437 wallcycle_stop(wcycle, ewcPME_REDISTXF);
4442 fprintf(debug, "Node= %6d, pme local particles=%6d\n",
4443 cr->nodeid, atc->n);
4446 if (flags & GMX_PME_SPREAD_Q)
4448 wallcycle_start(wcycle, ewcPME_SPREADGATHER);
4450 /* Spread the charges on a grid */
4451 spread_on_grid(pme, &pme->atc[0], pmegrid, q == 0, TRUE, fftgrid);
4455 inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
4457 inc_nrnb(nrnb, eNR_SPREADQBSP,
4458 pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
4460 if (!pme->bUseThreads)
4462 wrap_periodic_pmegrid(pme, grid);
4464 /* sum contributions to local grid from other nodes */
4466 if (pme->nnodes > 1)
4468 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_QGRID_FORWARD);
4473 copy_pmegrid_to_fftgrid(pme, grid, fftgrid);
4476 wallcycle_stop(wcycle, ewcPME_SPREADGATHER);
4479 dump_local_fftgrid(pme,fftgrid);
4484 /* Here we start a large thread parallel region */
4485 #pragma omp parallel num_threads(pme->nthread) private(thread)
4487 thread = gmx_omp_get_thread_num();
4488 if (flags & GMX_PME_SOLVE)
4495 wallcycle_start(wcycle, ewcPME_FFT);
4497 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
4501 wallcycle_stop(wcycle, ewcPME_FFT);
4505 /* solve in k-space for our local cells */
4508 wallcycle_start(wcycle, ewcPME_SOLVE);
4511 solve_pme_yzx(pme, cfftgrid, ewaldcoeff,
4512 box[XX][XX]*box[YY][YY]*box[ZZ][ZZ],
4514 pme->nthread, thread);
4517 wallcycle_stop(wcycle, ewcPME_SOLVE);
4519 inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
4529 wallcycle_start(wcycle, ewcPME_FFT);
4531 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
4535 wallcycle_stop(wcycle, ewcPME_FFT);
4539 if (pme->nodeid == 0)
4541 ntot = pme->nkx*pme->nky*pme->nkz;
4542 npme = ntot*log((real)ntot)/log(2.0);
4543 inc_nrnb(nrnb, eNR_FFT, 2*npme);
4546 wallcycle_start(wcycle, ewcPME_SPREADGATHER);
4549 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, pme->nthread, thread);
4552 /* End of thread parallel section.
4553 * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
4558 /* distribute local grid to all nodes */
4560 if (pme->nnodes > 1)
4562 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_QGRID_BACKWARD);
4567 unwrap_periodic_pmegrid(pme, grid);
4569 /* interpolate forces for our local atoms */
4573 /* If we are running without parallelization,
4574 * atc->f is the actual force array, not a buffer,
4575 * therefore we should not clear it.
4577 bClearF = (q == 0 && PAR(cr));
4578 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
4579 for (thread = 0; thread < pme->nthread; thread++)
4581 gather_f_bsplines(pme, grid, bClearF, atc,
4582 &atc->spline[thread],
4583 pme->bFEP ? (q == 0 ? 1.0-lambda : lambda) : 1.0);
4588 inc_nrnb(nrnb, eNR_GATHERFBSP,
4589 pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
4590 wallcycle_stop(wcycle, ewcPME_SPREADGATHER);
4595 /* This should only be called on the master thread
4596 * and after the threads have synchronized.
4598 get_pme_ener_vir(pme, pme->nthread, &energy_AB[q], vir_AB[q]);
4602 if (bCalcF && pme->nnodes > 1)
4604 wallcycle_start(wcycle, ewcPME_REDISTXF);
4605 for (d = 0; d < pme->ndecompdim; d++)
4608 if (d == pme->ndecompdim - 1)
4615 n_d = pme->atc[d+1].n;
4616 f_d = pme->atc[d+1].f;
4618 if (DOMAINDECOMP(cr))
4620 dd_pmeredist_f(pme, atc, n_d, f_d,
4621 d == pme->ndecompdim-1 && pme->bPPnode);
4625 pmeredist_pd(pme, FALSE, n_d, TRUE, f_d, NULL, atc);
4629 wallcycle_stop(wcycle, ewcPME_REDISTXF);
4637 *energy = energy_AB[0];
4638 m_add(vir, vir_AB[0], vir);
4642 *energy = (1.0-lambda)*energy_AB[0] + lambda*energy_AB[1];
4643 *dvdlambda += energy_AB[1] - energy_AB[0];
4644 for (i = 0; i < DIM; i++)
4646 for (j = 0; j < DIM; j++)
4648 vir[i][j] += (1.0-lambda)*vir_AB[0][i][j] +
4649 lambda*vir_AB[1][i][j];
4661 fprintf(debug, "PME mesh energy: %g\n", *energy);