2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013,2014,2015,2016,2017,2018, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
39 * \brief This file contains function definitions necessary for
40 * computing energies and forces for the PME long-ranged part (Coulomb
43 * \author Erik Lindahl <erik@kth.se>
44 * \author Berk Hess <hess@kth.se>
45 * \ingroup module_ewald
47 /* IMPORTANT FOR DEVELOPERS:
49 * Triclinic pme stuff isn't entirely trivial, and we've experienced
50 * some bugs during development (many of them due to me). To avoid
51 * this in the future, please check the following things if you make
52 * changes in this file:
54 * 1. You should obtain identical (at least to the PME precision)
55 * energies, forces, and virial for
56 * a rectangular box and a triclinic one where the z (or y) axis is
57 * tilted a whole box side. For instance you could use these boxes:
59 * rectangular triclinic
64 * 2. You should check the energy conservation in a triclinic box.
66 * It might seem an overkill, but better safe than sorry.
85 #include "gromacs/ewald/ewald-utils.h"
86 #include "gromacs/fft/parallel_3dfft.h"
87 #include "gromacs/fileio/pdbio.h"
88 #include "gromacs/gmxlib/network.h"
89 #include "gromacs/gmxlib/nrnb.h"
90 #include "gromacs/math/gmxcomplex.h"
91 #include "gromacs/math/invertmatrix.h"
92 #include "gromacs/math/units.h"
93 #include "gromacs/math/vec.h"
94 #include "gromacs/math/vectypes.h"
95 #include "gromacs/mdtypes/commrec.h"
96 #include "gromacs/mdtypes/forcerec.h"
97 #include "gromacs/mdtypes/inputrec.h"
98 #include "gromacs/mdtypes/md_enums.h"
99 #include "gromacs/pbcutil/pbc.h"
100 #include "gromacs/timing/cyclecounter.h"
101 #include "gromacs/timing/wallcycle.h"
102 #include "gromacs/timing/walltime_accounting.h"
103 #include "gromacs/utility/basedefinitions.h"
104 #include "gromacs/utility/exceptions.h"
105 #include "gromacs/utility/fatalerror.h"
106 #include "gromacs/utility/gmxmpi.h"
107 #include "gromacs/utility/gmxomp.h"
108 #include "gromacs/utility/logger.h"
109 #include "gromacs/utility/real.h"
110 #include "gromacs/utility/smalloc.h"
111 #include "gromacs/utility/stringutil.h"
112 #include "gromacs/utility/unique_cptr.h"
114 #include "calculate-spline-moduli.h"
115 #include "pme-gather.h"
116 #include "pme-gpu-internal.h"
117 #include "pme-grid.h"
118 #include "pme-internal.h"
119 #include "pme-redistribute.h"
120 #include "pme-solve.h"
121 #include "pme-spline-work.h"
122 #include "pme-spread.h"
124 /*! \brief Number of bytes in a cache line.
126 * Must also be a multiple of the SIMD and SIMD4 register size, to
127 * preserve alignment.
129 const int gmxCacheLineSize = 64;
131 //! Set up coordinate communication
132 static void setup_coordinate_communication(pme_atomcomm_t *atc)
140 for (i = 1; i <= nslab/2; i++)
142 fw = (atc->nodeid + i) % nslab;
143 bw = (atc->nodeid - i + nslab) % nslab;
146 atc->node_dest[n] = fw;
147 atc->node_src[n] = bw;
152 atc->node_dest[n] = bw;
153 atc->node_src[n] = fw;
159 /*! \brief Round \p n up to the next multiple of \p f */
160 static int mult_up(int n, int f)
162 return ((n + f - 1)/f)*f;
165 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
166 static double estimate_pme_load_imbalance(struct gmx_pme_t *pme)
171 nma = pme->nnodes_major;
172 nmi = pme->nnodes_minor;
174 n1 = mult_up(pme->nkx, nma)*mult_up(pme->nky, nmi)*pme->nkz;
175 n2 = mult_up(pme->nkx, nma)*mult_up(pme->nkz, nmi)*pme->nky;
176 n3 = mult_up(pme->nky, nma)*mult_up(pme->nkz, nmi)*pme->nkx;
178 /* pme_solve is roughly double the cost of an fft */
180 return (n1 + n2 + 3*n3)/(double)(6*pme->nkx*pme->nky*pme->nkz);
183 /*! \brief Initialize atom communication data structure */
184 static void init_atomcomm(struct gmx_pme_t *pme, pme_atomcomm_t *atc,
185 int dimind, gmx_bool bSpread)
189 atc->dimind = dimind;
196 atc->mpi_comm = pme->mpi_comm_d[dimind];
197 MPI_Comm_size(atc->mpi_comm, &atc->nslab);
198 MPI_Comm_rank(atc->mpi_comm, &atc->nodeid);
202 fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", atc->dimind, atc->nslab, atc->nodeid);
206 atc->bSpread = bSpread;
207 atc->pme_order = pme->pme_order;
211 snew(atc->node_dest, atc->nslab);
212 snew(atc->node_src, atc->nslab);
213 setup_coordinate_communication(atc);
215 snew(atc->count_thread, pme->nthread);
216 for (thread = 0; thread < pme->nthread; thread++)
218 snew(atc->count_thread[thread], atc->nslab);
220 atc->count = atc->count_thread[0];
221 snew(atc->rcount, atc->nslab);
222 snew(atc->buf_index, atc->nslab);
225 atc->nthread = pme->nthread;
226 if (atc->nthread > 1)
228 snew(atc->thread_plist, atc->nthread);
230 snew(atc->spline, atc->nthread);
231 for (thread = 0; thread < atc->nthread; thread++)
233 if (atc->nthread > 1)
235 snew(atc->thread_plist[thread].n, atc->nthread+2*gmxCacheLineSize);
236 atc->thread_plist[thread].n += gmxCacheLineSize;
241 /*! \brief Destroy an atom communication data structure and its child structs */
242 static void destroy_atomcomm(pme_atomcomm_t *atc)
247 sfree(atc->node_dest);
248 sfree(atc->node_src);
249 for (int i = 0; i < atc->nthread; i++)
251 sfree(atc->count_thread[i]);
253 sfree(atc->count_thread);
255 sfree(atc->buf_index);
258 sfree(atc->coefficient);
264 sfree(atc->thread_idx);
265 for (int i = 0; i < atc->nthread; i++)
267 if (atc->nthread > 1)
269 int *n_ptr = atc->thread_plist[i].n - gmxCacheLineSize;
271 sfree(atc->thread_plist[i].i);
273 sfree(atc->spline[i].ind);
274 for (int d = 0; d < ZZ; d++)
276 sfree(atc->spline[i].theta[d]);
277 sfree(atc->spline[i].dtheta[d]);
279 sfree_aligned(atc->spline[i].ptr_dtheta_z);
280 sfree_aligned(atc->spline[i].ptr_theta_z);
282 if (atc->nthread > 1)
284 sfree(atc->thread_plist);
289 /*! \brief Initialize data structure for communication */
291 init_overlap_comm(pme_overlap_t * ol,
311 /* Linear translation of the PME grid won't affect reciprocal space
312 * calculations, so to optimize we only interpolate "upwards",
313 * which also means we only have to consider overlap in one direction.
314 * I.e., particles on this node might also be spread to grid indices
315 * that belong to higher nodes (modulo nnodes)
318 ol->s2g0.resize(ol->nnodes + 1);
319 ol->s2g1.resize(ol->nnodes);
322 fprintf(debug, "PME slab boundaries:");
324 for (int i = 0; i < nnodes; i++)
326 /* s2g0 the local interpolation grid start.
327 * s2g1 the local interpolation grid end.
328 * Since in calc_pidx we divide particles, and not grid lines,
329 * spatially uniform along dimension x or y, we need to round
330 * s2g0 down and s2g1 up.
332 ol->s2g0[i] = (i * ndata + 0) / nnodes;
333 ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
337 fprintf(debug, " %3d %3d", ol->s2g0[i], ol->s2g1[i]);
340 ol->s2g0[nnodes] = ndata;
343 fprintf(debug, "\n");
346 /* Determine with how many nodes we need to communicate the grid overlap */
347 int testRankCount = 0;
352 for (int i = 0; i < nnodes; i++)
354 if ((i + testRankCount < nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount]) ||
355 (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
361 while (bCont && testRankCount < nnodes);
363 ol->comm_data.resize(testRankCount - 1);
366 for (size_t b = 0; b < ol->comm_data.size(); b++)
368 pme_grid_comm_t *pgc = &ol->comm_data[b];
371 pgc->send_id = (ol->nodeid + (b + 1)) % ol->nnodes;
372 int fft_start = ol->s2g0[pgc->send_id];
373 int fft_end = ol->s2g0[pgc->send_id + 1];
374 if (pgc->send_id < nodeid)
379 int send_index1 = ol->s2g1[nodeid];
380 send_index1 = std::min(send_index1, fft_end);
381 pgc->send_index0 = fft_start;
382 pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
383 ol->send_size += pgc->send_nindex;
385 /* We always start receiving to the first index of our slab */
386 pgc->recv_id = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
387 fft_start = ol->s2g0[ol->nodeid];
388 fft_end = ol->s2g0[ol->nodeid + 1];
389 int recv_index1 = ol->s2g1[pgc->recv_id];
390 if (pgc->recv_id > nodeid)
392 recv_index1 -= ndata;
394 recv_index1 = std::min(recv_index1, fft_end);
395 pgc->recv_index0 = fft_start;
396 pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
400 /* Communicate the buffer sizes to receive */
401 for (size_t b = 0; b < ol->comm_data.size(); b++)
403 MPI_Sendrecv(&ol->send_size, 1, MPI_INT, ol->comm_data[b].send_id, b,
404 &ol->comm_data[b].recv_size, 1, MPI_INT, ol->comm_data[b].recv_id, b,
405 ol->mpi_comm, &stat);
409 /* For non-divisible grid we need pme_order iso pme_order-1 */
410 ol->sendbuf.resize(norder * commplainsize);
411 ol->recvbuf.resize(norder * commplainsize);
414 int minimalPmeGridSize(int pmeOrder)
416 /* The actual grid size limitations are:
417 * serial: >= pme_order
418 * DD, no OpenMP: >= 2*(pme_order - 1)
419 * DD, OpenMP: >= pme_order + 1
420 * But we use the maximum for simplicity since in practice there is not
421 * much performance difference between pme_order and 2*(pme_order -1).
423 int minimalSize = 2*(pmeOrder - 1);
425 GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
426 GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
431 bool gmx_pme_check_restrictions(int pme_order,
432 int nkx, int nky, int nkz,
437 if (pme_order > PME_ORDER_MAX)
444 std::string message = gmx::formatString(
445 "pme_order (%d) is larger than the maximum allowed value (%d). Modify and recompile the code if you really need such a high order.",
446 pme_order, PME_ORDER_MAX);
447 GMX_THROW(InconsistentInputError(message));
450 const int minGridSize = minimalPmeGridSize(pme_order);
451 if (nkx < minGridSize ||
459 std::string message = gmx::formatString(
460 "The PME grid sizes need to be >= 2*(pme_order-1) (%d)",
462 GMX_THROW(InconsistentInputError(message));
465 /* Check for a limitation of the (current) sum_fftgrid_dd code.
466 * We only allow multiple communication pulses in dim 1, not in dim 0.
468 if (useThreads && (nkx < nnodes_major*pme_order &&
469 nkx != nnodes_major*(pme_order - 1)))
475 gmx_fatal(FARGS, "The number of PME grid lines per rank along x is %g. But when using OpenMP threads, the number of grid lines per rank along x should be >= pme_order (%d) or = pmeorder-1. To resolve this issue, use fewer ranks along x (and possibly more along y and/or z) by specifying -dd manually.",
476 nkx/(double)nnodes_major, pme_order);
482 /*! \brief Round \p enumerator */
483 static int div_round_up(int enumerator, int denominator)
485 return (enumerator + denominator - 1)/denominator;
488 gmx_pme_t *gmx_pme_init(const t_commrec *cr,
491 const t_inputrec *ir,
493 gmx_bool bFreeEnergy_q,
494 gmx_bool bFreeEnergy_lj,
495 gmx_bool bReproducible,
501 gmx_device_info_t *gpuInfo,
502 const gmx::MDLogger & /*mdlog*/)
504 int use_threads, sum_use_threads, i;
509 fprintf(debug, "Creating PME data structures.\n");
512 unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
514 pme->sum_qgrid_tmp = nullptr;
515 pme->sum_qgrid_dd_tmp = nullptr;
522 pme->nnodes_major = nnodes_major;
523 pme->nnodes_minor = nnodes_minor;
526 if (nnodes_major*nnodes_minor > 1)
528 pme->mpi_comm = cr->mpi_comm_mygroup;
530 MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
531 MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
532 if (pme->nnodes != nnodes_major*nnodes_minor)
534 gmx_incons("PME rank count mismatch");
539 pme->mpi_comm = MPI_COMM_NULL;
543 if (pme->nnodes == 1)
546 pme->mpi_comm_d[0] = MPI_COMM_NULL;
547 pme->mpi_comm_d[1] = MPI_COMM_NULL;
550 pme->nodeid_major = 0;
551 pme->nodeid_minor = 0;
553 pme->mpi_comm_d[0] = pme->mpi_comm_d[1] = MPI_COMM_NULL;
558 if (nnodes_minor == 1)
561 pme->mpi_comm_d[0] = pme->mpi_comm;
562 pme->mpi_comm_d[1] = MPI_COMM_NULL;
565 pme->nodeid_major = pme->nodeid;
566 pme->nodeid_minor = 0;
569 else if (nnodes_major == 1)
572 pme->mpi_comm_d[0] = MPI_COMM_NULL;
573 pme->mpi_comm_d[1] = pme->mpi_comm;
576 pme->nodeid_major = 0;
577 pme->nodeid_minor = pme->nodeid;
581 if (pme->nnodes % nnodes_major != 0)
583 gmx_incons("For 2D PME decomposition, #PME ranks must be divisible by the number of ranks in the major dimension");
588 MPI_Comm_split(pme->mpi_comm, pme->nodeid % nnodes_minor,
589 pme->nodeid, &pme->mpi_comm_d[0]); /* My communicator along major dimension */
590 MPI_Comm_split(pme->mpi_comm, pme->nodeid/nnodes_minor,
591 pme->nodeid, &pme->mpi_comm_d[1]); /* My communicator along minor dimension */
593 MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
594 MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
595 MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
596 MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
599 pme->bPPnode = thisRankHasDuty(cr, DUTY_PP);
602 pme->nthread = nthread;
604 /* Check if any of the PME MPI ranks uses threads */
605 use_threads = (pme->nthread > 1 ? 1 : 0);
609 MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT,
610 MPI_SUM, pme->mpi_comm);
615 sum_use_threads = use_threads;
617 pme->bUseThreads = (sum_use_threads > 0);
619 if (ir->ePBC == epbcSCREW)
621 gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
625 * It is likely that the current gmx_pme_do() routine supports calculating
626 * only Coulomb or LJ while gmx_pme_init() configures for both,
627 * but that has never been tested.
628 * It is likely that the current gmx_pme_do() routine supports calculating,
629 * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
630 * configures with free-energy, but that has never been tested.
632 pme->doCoulomb = EEL_PME(ir->coulombtype);
633 pme->doLJ = EVDW_PME(ir->vdwtype);
634 pme->bFEP_q = ((ir->efep != efepNO) && bFreeEnergy_q);
635 pme->bFEP_lj = ((ir->efep != efepNO) && bFreeEnergy_lj);
636 pme->bFEP = (pme->bFEP_q || pme->bFEP_lj);
640 pme->bP3M = (ir->coulombtype == eelP3M_AD || getenv("GMX_PME_P3M") != nullptr);
641 pme->pme_order = ir->pme_order;
642 pme->ewaldcoeff_q = ewaldcoeff_q;
643 pme->ewaldcoeff_lj = ewaldcoeff_lj;
645 /* Always constant electrostatics coefficients */
646 pme->epsilon_r = ir->epsilon_r;
648 /* Always constant LJ coefficients */
649 pme->ljpme_combination_rule = ir->ljpme_combination_rule;
651 // The box requires scaling with nwalls = 2, we store that condition as well
652 // as the scaling factor
653 delete pme->boxScaler;
654 pme->boxScaler = new EwaldBoxZScaler(*ir);
656 /* If we violate restrictions, generate a fatal error here */
657 gmx_pme_check_restrictions(pme->pme_order,
658 pme->nkx, pme->nky, pme->nkz,
668 MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
669 MPI_Type_commit(&(pme->rvec_mpi));
672 /* Note that the coefficient spreading and force gathering, which usually
673 * takes about the same amount of time as FFT+solve_pme,
674 * is always fully load balanced
675 * (unless the coefficient distribution is inhomogeneous).
678 imbal = estimate_pme_load_imbalance(pme.get());
679 if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
683 "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
684 " For optimal PME load balancing\n"
685 " PME grid_x (%d) and grid_y (%d) should be divisible by #PME_ranks_x (%d)\n"
686 " and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_ranks_y (%d)\n"
688 (int)((imbal-1)*100 + 0.5),
689 pme->nkx, pme->nky, pme->nnodes_major,
690 pme->nky, pme->nkz, pme->nnodes_minor);
694 /* For non-divisible grid we need pme_order iso pme_order-1 */
695 /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
696 * y is always copied through a buffer: we don't need padding in z,
697 * but we do need the overlap in x because of the communication order.
699 init_overlap_comm(&pme->overlap[0], pme->pme_order,
703 pme->nnodes_major, pme->nodeid_major,
705 (div_round_up(pme->nky, pme->nnodes_minor)+pme->pme_order)*(pme->nkz+pme->pme_order-1));
707 /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
708 * We do this with an offset buffer of equal size, so we need to allocate
709 * extra for the offset. That's what the (+1)*pme->nkz is for.
711 init_overlap_comm(&pme->overlap[1], pme->pme_order,
715 pme->nnodes_minor, pme->nodeid_minor,
717 (div_round_up(pme->nkx, pme->nnodes_major)+pme->pme_order+1)*pme->nkz);
719 /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
720 * Note that gmx_pme_check_restrictions checked for this already.
722 if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
724 gmx_incons("More than one communication pulse required for grid overlap communication along the major dimension while using threads");
727 snew(pme->bsp_mod[XX], pme->nkx);
728 snew(pme->bsp_mod[YY], pme->nky);
729 snew(pme->bsp_mod[ZZ], pme->nkz);
731 pme->gpu = pmeGpu; /* Carrying over the single GPU structure */
732 pme->runMode = runMode;
734 /* The required size of the interpolation grid, including overlap.
735 * The allocated size (pmegrid_n?) might be slightly larger.
737 pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] -
738 pme->overlap[0].s2g0[pme->nodeid_major];
739 pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] -
740 pme->overlap[1].s2g0[pme->nodeid_minor];
741 pme->pmegrid_nz_base = pme->nkz;
742 pme->pmegrid_nz = pme->pmegrid_nz_base + pme->pme_order - 1;
743 set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
744 pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
745 pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
746 pme->pmegrid_start_iz = 0;
748 make_gridindex_to_localindex(pme->nkx,
749 pme->pmegrid_start_ix,
750 pme->pmegrid_nx - (pme->pme_order-1),
751 &pme->nnx, &pme->fshx);
752 make_gridindex_to_localindex(pme->nky,
753 pme->pmegrid_start_iy,
754 pme->pmegrid_ny - (pme->pme_order-1),
755 &pme->nny, &pme->fshy);
756 make_gridindex_to_localindex(pme->nkz,
757 pme->pmegrid_start_iz,
758 pme->pmegrid_nz_base,
759 &pme->nnz, &pme->fshz);
761 pme->spline_work = make_pme_spline_work(pme->pme_order);
766 /* It doesn't matter if we allocate too many grids here,
767 * we only allocate and use the ones we need.
771 pme->ngrids = ((ir->ljpme_combination_rule == eljpmeLB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
777 snew(pme->fftgrid, pme->ngrids);
778 snew(pme->cfftgrid, pme->ngrids);
779 snew(pme->pfft_setup, pme->ngrids);
781 for (i = 0; i < pme->ngrids; ++i)
783 if ((i < DO_Q && pme->doCoulomb && (i == 0 ||
785 (i >= DO_Q && pme->doLJ && (i == 2 ||
787 ir->ljpme_combination_rule == eljpmeLB)))
789 pmegrids_init(&pme->pmegrid[i],
790 pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
791 pme->pmegrid_nz_base,
795 pme->overlap[0].s2g1[pme->nodeid_major]-pme->overlap[0].s2g0[pme->nodeid_major+1],
796 pme->overlap[1].s2g1[pme->nodeid_minor]-pme->overlap[1].s2g0[pme->nodeid_minor+1]);
797 /* This routine will allocate the grid data to fit the FFTs */
798 const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed) ? gmx::PinningPolicy::CanBePinned : gmx::PinningPolicy::CannotBePinned;
799 gmx_parallel_3dfft_init(&pme->pfft_setup[i], ndata,
800 &pme->fftgrid[i], &pme->cfftgrid[i],
802 bReproducible, pme->nthread, allocateRealGridForGpu);
809 /* Use plain SPME B-spline interpolation */
810 make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
814 /* Use the P3M grid-optimized influence function */
815 make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
818 /* Use atc[0] for spreading */
819 init_atomcomm(pme.get(), &pme->atc[0], nnodes_major > 1 ? 0 : 1, TRUE);
820 if (pme->ndecompdim >= 2)
822 init_atomcomm(pme.get(), &pme->atc[1], 1, FALSE);
825 if (pme->nnodes == 1)
827 pme->atc[0].n = homenr;
828 pme_realloc_atomcomm_things(&pme->atc[0]);
831 pme->lb_buf1 = nullptr;
832 pme->lb_buf2 = nullptr;
833 pme->lb_buf_nalloc = 0;
835 pme_gpu_reinit(pme.get(), gpuInfo);
837 pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
839 // no exception was thrown during the init, so we hand over the PME structure handle
840 return pme.release();
843 void gmx_pme_reinit(struct gmx_pme_t **pmedata,
845 struct gmx_pme_t * pme_src,
846 const t_inputrec * ir,
847 const ivec grid_size,
853 // Create a copy of t_inputrec fields that are used in gmx_pme_init().
854 // TODO: This would be better as just copying a sub-structure that contains
855 // all the PME parameters and nothing else.
858 irc.coulombtype = ir->coulombtype;
859 irc.vdwtype = ir->vdwtype;
861 irc.pme_order = ir->pme_order;
862 irc.epsilon_r = ir->epsilon_r;
863 irc.ljpme_combination_rule = ir->ljpme_combination_rule;
864 irc.nkx = grid_size[XX];
865 irc.nky = grid_size[YY];
866 irc.nkz = grid_size[ZZ];
868 if (pme_src->nnodes == 1)
870 homenr = pme_src->atc[0].n;
879 const gmx::MDLogger dummyLogger;
880 // This is reinit which is currently only changing grid size/coefficients,
881 // so we don't expect the actual logging.
882 // TODO: when PME is an object, it should take reference to mdlog on construction and save it.
883 GMX_ASSERT(pmedata, "Invalid PME pointer");
884 *pmedata = gmx_pme_init(cr, pme_src->nnodes_major, pme_src->nnodes_minor,
885 &irc, homenr, pme_src->bFEP_q, pme_src->bFEP_lj, FALSE, ewaldcoeff_q, ewaldcoeff_lj,
886 pme_src->nthread, pme_src->runMode, pme_src->gpu, nullptr, dummyLogger);
887 //TODO this is mostly passing around current values
889 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
891 /* We can easily reuse the allocated pme grids in pme_src */
892 reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
893 /* We would like to reuse the fft grids, but that's harder */
896 void gmx_pme_calc_energy(struct gmx_pme_t *pme, int n, rvec *x, real *q, real *V)
903 gmx_incons("gmx_pme_calc_energy called in parallel");
907 gmx_incons("gmx_pme_calc_energy with free energy");
910 atc = &pme->atc_energy;
912 if (atc->spline == nullptr)
914 snew(atc->spline, atc->nthread);
918 atc->pme_order = pme->pme_order;
920 pme_realloc_atomcomm_things(atc);
922 atc->coefficient = q;
924 /* We only use the A-charges grid */
925 grid = &pme->pmegrid[PME_GRID_QA];
927 /* Only calculate the spline coefficients, don't actually spread */
928 spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
930 *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
933 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
935 calc_initial_lb_coeffs(struct gmx_pme_t *pme, real *local_c6, real *local_sigma)
938 for (i = 0; i < pme->atc[0].n; ++i)
941 sigma4 = local_sigma[i];
942 sigma4 = sigma4*sigma4;
943 sigma4 = sigma4*sigma4;
944 pme->atc[0].coefficient[i] = local_c6[i] / sigma4;
948 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
950 calc_next_lb_coeffs(struct gmx_pme_t *pme, real *local_sigma)
954 for (i = 0; i < pme->atc[0].n; ++i)
956 pme->atc[0].coefficient[i] *= local_sigma[i];
960 int gmx_pme_do(struct gmx_pme_t *pme,
961 int start, int homenr,
963 real chargeA[], real chargeB[],
964 real c6A[], real c6B[],
965 real sigmaA[], real sigmaB[],
966 matrix box, t_commrec *cr,
967 int maxshift_x, int maxshift_y,
968 t_nrnb *nrnb, gmx_wallcycle *wcycle,
969 matrix vir_q, matrix vir_lj,
970 real *energy_q, real *energy_lj,
971 real lambda_q, real lambda_lj,
972 real *dvdlambda_q, real *dvdlambda_lj,
975 GMX_ASSERT(pme->runMode == PmeRunMode::CPU, "gmx_pme_do should not be called on the GPU PME run.");
977 int d, i, j, npme, grid_index, max_grid_index;
979 pme_atomcomm_t *atc = nullptr;
980 pmegrids_t *pmegrid = nullptr;
981 real *grid = nullptr;
983 real *coefficient = nullptr;
988 gmx_parallel_3dfft_t pfft_setup;
990 t_complex * cfftgrid;
992 gmx_bool bFirst, bDoSplines;
994 int fep_states_lj = pme->bFEP_lj ? 2 : 1;
995 const gmx_bool bCalcEnerVir = flags & GMX_PME_CALC_ENER_VIR;
996 const gmx_bool bBackFFT = flags & (GMX_PME_CALC_F | GMX_PME_CALC_POT);
997 const gmx_bool bCalcF = flags & GMX_PME_CALC_F;
999 assert(pme->nnodes > 0);
1000 assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1002 if (pme->nnodes > 1)
1006 if (atc->npd > atc->pd_nalloc)
1008 atc->pd_nalloc = over_alloc_dd(atc->npd);
1009 srenew(atc->pd, atc->pd_nalloc);
1011 for (d = pme->ndecompdim-1; d >= 0; d--)
1014 atc->maxshift = (atc->dimind == 0 ? maxshift_x : maxshift_y);
1020 /* This could be necessary for TPI */
1021 pme->atc[0].n = homenr;
1022 if (DOMAINDECOMP(cr))
1024 pme_realloc_atomcomm_things(atc);
1031 pme->boxScaler->scaleBox(box, scaledBox);
1033 gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1036 /* For simplicity, we construct the splines for all particles if
1037 * more than one PME calculations is needed. Some optimization
1038 * could be done by keeping track of which atoms have splines
1039 * constructed, and construct new splines on each pass for atoms
1040 * that don't yet have them.
1043 bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1045 /* We need a maximum of four separate PME calculations:
1046 * grid_index=0: Coulomb PME with charges from state A
1047 * grid_index=1: Coulomb PME with charges from state B
1048 * grid_index=2: LJ PME with C6 from state A
1049 * grid_index=3: LJ PME with C6 from state B
1050 * For Lorentz-Berthelot combination rules, a separate loop is used to
1051 * calculate all the terms
1054 /* If we are doing LJ-PME with LB, we only do Q here */
1055 max_grid_index = (pme->ljpme_combination_rule == eljpmeLB) ? DO_Q : DO_Q_AND_LJ;
1057 for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1059 /* Check if we should do calculations at this grid_index
1060 * If grid_index is odd we should be doing FEP
1061 * If grid_index < 2 we should be doing electrostatic PME
1062 * If grid_index >= 2 we should be doing LJ-PME
1064 if ((grid_index < DO_Q && (!pme->doCoulomb ||
1065 (grid_index == 1 && !pme->bFEP_q))) ||
1066 (grid_index >= DO_Q && (!pme->doLJ ||
1067 (grid_index == 3 && !pme->bFEP_lj))))
1071 /* Unpack structure */
1072 pmegrid = &pme->pmegrid[grid_index];
1073 fftgrid = pme->fftgrid[grid_index];
1074 cfftgrid = pme->cfftgrid[grid_index];
1075 pfft_setup = pme->pfft_setup[grid_index];
1078 case 0: coefficient = chargeA + start; break;
1079 case 1: coefficient = chargeB + start; break;
1080 case 2: coefficient = c6A + start; break;
1081 case 3: coefficient = c6B + start; break;
1084 grid = pmegrid->grid.grid;
1088 fprintf(debug, "PME: number of ranks = %d, rank = %d\n",
1089 cr->nnodes, cr->nodeid);
1090 fprintf(debug, "Grid = %p\n", (void*)grid);
1091 if (grid == nullptr)
1093 gmx_fatal(FARGS, "No grid!");
1098 if (pme->nnodes == 1)
1100 atc->coefficient = coefficient;
1104 wallcycle_start(wcycle, ewcPME_REDISTXF);
1105 do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, coefficient);
1108 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1113 fprintf(debug, "Rank= %6d, pme local particles=%6d\n",
1114 cr->nodeid, atc->n);
1117 if (flags & GMX_PME_SPREAD)
1119 wallcycle_start(wcycle, ewcPME_SPREAD);
1121 /* Spread the coefficients on a grid */
1122 spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1126 inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1128 inc_nrnb(nrnb, eNR_SPREADBSP,
1129 pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1131 if (!pme->bUseThreads)
1133 wrap_periodic_pmegrid(pme, grid);
1135 /* sum contributions to local grid from other nodes */
1137 if (pme->nnodes > 1)
1139 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1144 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1147 wallcycle_stop(wcycle, ewcPME_SPREAD);
1149 /* TODO If the OpenMP and single-threaded implementations
1150 converge, then spread_on_grid() and
1151 copy_pmegrid_to_fftgrid() will perhaps live in the same
1156 /* Here we start a large thread parallel region */
1157 #pragma omp parallel num_threads(pme->nthread) private(thread)
1161 thread = gmx_omp_get_thread_num();
1162 if (flags & GMX_PME_SOLVE)
1169 wallcycle_start(wcycle, ewcPME_FFT);
1171 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1175 wallcycle_stop(wcycle, ewcPME_FFT);
1179 /* solve in k-space for our local cells */
1182 wallcycle_start(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1184 if (grid_index < DO_Q)
1187 solve_pme_yzx(pme, cfftgrid,
1188 scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1190 pme->nthread, thread);
1195 solve_pme_lj_yzx(pme, &cfftgrid, FALSE,
1196 scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1198 pme->nthread, thread);
1203 wallcycle_stop(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1205 inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1215 wallcycle_start(wcycle, ewcPME_FFT);
1217 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1221 wallcycle_stop(wcycle, ewcPME_FFT);
1225 if (pme->nodeid == 0)
1227 real ntot = pme->nkx*pme->nky*pme->nkz;
1228 npme = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1229 inc_nrnb(nrnb, eNR_FFT, 2*npme);
1232 /* Note: this wallcycle region is closed below
1233 outside an OpenMP region, so take care if
1234 refactoring code here. */
1235 wallcycle_start(wcycle, ewcPME_GATHER);
1238 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1240 } GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1242 /* End of thread parallel section.
1243 * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1248 /* distribute local grid to all nodes */
1250 if (pme->nnodes > 1)
1252 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1257 unwrap_periodic_pmegrid(pme, grid);
1262 /* interpolate forces for our local atoms */
1266 /* If we are running without parallelization,
1267 * atc->f is the actual force array, not a buffer,
1268 * therefore we should not clear it.
1270 lambda = grid_index < DO_Q ? lambda_q : lambda_lj;
1271 bClearF = (bFirst && PAR(cr));
1272 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1273 for (thread = 0; thread < pme->nthread; thread++)
1277 gather_f_bsplines(pme, grid, bClearF, atc,
1278 &atc->spline[thread],
1279 pme->bFEP ? (grid_index % 2 == 0 ? 1.0-lambda : lambda) : 1.0);
1281 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1286 inc_nrnb(nrnb, eNR_GATHERFBSP,
1287 pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1288 /* Note: this wallcycle region is opened above inside an OpenMP
1289 region, so take care if refactoring code here. */
1290 wallcycle_stop(wcycle, ewcPME_GATHER);
1295 /* This should only be called on the master thread
1296 * and after the threads have synchronized.
1300 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &energy_AB[grid_index], vir_AB[grid_index]);
1304 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &energy_AB[grid_index], vir_AB[grid_index]);
1308 } /* of grid_index-loop */
1310 /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1313 if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB)
1315 /* Loop over A- and B-state if we are doing FEP */
1316 for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1318 real *local_c6 = nullptr, *local_sigma = nullptr, *RedistC6 = nullptr, *RedistSigma = nullptr;
1319 if (pme->nnodes == 1)
1321 if (pme->lb_buf1 == nullptr)
1323 pme->lb_buf_nalloc = pme->atc[0].n;
1324 snew(pme->lb_buf1, pme->lb_buf_nalloc);
1326 pme->atc[0].coefficient = pme->lb_buf1;
1331 local_sigma = sigmaA;
1335 local_sigma = sigmaB;
1338 gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1348 RedistSigma = sigmaA;
1352 RedistSigma = sigmaB;
1355 gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1357 wallcycle_start(wcycle, ewcPME_REDISTXF);
1359 do_redist_pos_coeffs(pme, cr, start, homenr, bFirst, x, RedistC6);
1360 if (pme->lb_buf_nalloc < atc->n)
1362 pme->lb_buf_nalloc = atc->nalloc;
1363 srenew(pme->lb_buf1, pme->lb_buf_nalloc);
1364 srenew(pme->lb_buf2, pme->lb_buf_nalloc);
1366 local_c6 = pme->lb_buf1;
1367 for (i = 0; i < atc->n; ++i)
1369 local_c6[i] = atc->coefficient[i];
1373 do_redist_pos_coeffs(pme, cr, start, homenr, FALSE, x, RedistSigma);
1374 local_sigma = pme->lb_buf2;
1375 for (i = 0; i < atc->n; ++i)
1377 local_sigma[i] = atc->coefficient[i];
1381 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1383 calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1385 /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1386 for (grid_index = 2; grid_index < 9; ++grid_index)
1388 /* Unpack structure */
1389 pmegrid = &pme->pmegrid[grid_index];
1390 fftgrid = pme->fftgrid[grid_index];
1391 pfft_setup = pme->pfft_setup[grid_index];
1392 calc_next_lb_coeffs(pme, local_sigma);
1393 grid = pmegrid->grid.grid;
1396 if (flags & GMX_PME_SPREAD)
1398 wallcycle_start(wcycle, ewcPME_SPREAD);
1399 /* Spread the c6 on a grid */
1400 spread_on_grid(pme, &pme->atc[0], pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1404 inc_nrnb(nrnb, eNR_WEIGHTS, DIM*atc->n);
1407 inc_nrnb(nrnb, eNR_SPREADBSP,
1408 pme->pme_order*pme->pme_order*pme->pme_order*atc->n);
1409 if (pme->nthread == 1)
1411 wrap_periodic_pmegrid(pme, grid);
1412 /* sum contributions to local grid from other nodes */
1414 if (pme->nnodes > 1)
1416 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1420 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1422 wallcycle_stop(wcycle, ewcPME_SPREAD);
1424 /*Here we start a large thread parallel region*/
1425 #pragma omp parallel num_threads(pme->nthread) private(thread)
1429 thread = gmx_omp_get_thread_num();
1430 if (flags & GMX_PME_SOLVE)
1435 wallcycle_start(wcycle, ewcPME_FFT);
1438 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX,
1442 wallcycle_stop(wcycle, ewcPME_FFT);
1447 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1451 if (flags & GMX_PME_SOLVE)
1453 /* solve in k-space for our local cells */
1454 #pragma omp parallel num_threads(pme->nthread) private(thread)
1459 thread = gmx_omp_get_thread_num();
1462 wallcycle_start(wcycle, ewcLJPME);
1466 solve_pme_lj_yzx(pme, &pme->cfftgrid[2], TRUE,
1467 scaledBox[XX][XX]*scaledBox[YY][YY]*scaledBox[ZZ][ZZ],
1469 pme->nthread, thread);
1472 wallcycle_stop(wcycle, ewcLJPME);
1474 inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1477 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1483 /* This should only be called on the master thread and
1484 * after the threads have synchronized.
1486 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &energy_AB[2+fep_state], vir_AB[2+fep_state]);
1491 bFirst = !pme->doCoulomb;
1492 calc_initial_lb_coeffs(pme, local_c6, local_sigma);
1493 for (grid_index = 8; grid_index >= 2; --grid_index)
1495 /* Unpack structure */
1496 pmegrid = &pme->pmegrid[grid_index];
1497 fftgrid = pme->fftgrid[grid_index];
1498 pfft_setup = pme->pfft_setup[grid_index];
1499 grid = pmegrid->grid.grid;
1500 calc_next_lb_coeffs(pme, local_sigma);
1502 #pragma omp parallel num_threads(pme->nthread) private(thread)
1506 thread = gmx_omp_get_thread_num();
1511 wallcycle_start(wcycle, ewcPME_FFT);
1514 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL,
1518 wallcycle_stop(wcycle, ewcPME_FFT);
1522 if (pme->nodeid == 0)
1524 real ntot = pme->nkx*pme->nky*pme->nkz;
1525 npme = static_cast<int>(ntot*std::log(ntot)/std::log(2.0));
1526 inc_nrnb(nrnb, eNR_FFT, 2*npme);
1528 wallcycle_start(wcycle, ewcPME_GATHER);
1531 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1533 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1534 } /*#pragma omp parallel*/
1536 /* distribute local grid to all nodes */
1538 if (pme->nnodes > 1)
1540 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1545 unwrap_periodic_pmegrid(pme, grid);
1549 /* interpolate forces for our local atoms */
1551 bClearF = (bFirst && PAR(cr));
1552 scale = pme->bFEP ? (fep_state < 1 ? 1.0-lambda_lj : lambda_lj) : 1.0;
1553 scale *= lb_scale_factor[grid_index-2];
1555 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1556 for (thread = 0; thread < pme->nthread; thread++)
1560 gather_f_bsplines(pme, grid, bClearF, &pme->atc[0],
1561 &pme->atc[0].spline[thread],
1564 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1569 inc_nrnb(nrnb, eNR_GATHERFBSP,
1570 pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
1572 wallcycle_stop(wcycle, ewcPME_GATHER);
1575 } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1577 } /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1578 } /* if ((flags & GMX_PME_DO_LJ) && pme->ljpme_combination_rule == eljpmeLB) */
1580 if (bCalcF && pme->nnodes > 1)
1582 wallcycle_start(wcycle, ewcPME_REDISTXF);
1583 for (d = 0; d < pme->ndecompdim; d++)
1586 if (d == pme->ndecompdim - 1)
1593 n_d = pme->atc[d+1].n;
1594 f_d = pme->atc[d+1].f;
1596 if (DOMAINDECOMP(cr))
1598 dd_pmeredist_f(pme, atc, n_d, f_d,
1599 d == pme->ndecompdim-1 && pme->bPPnode);
1603 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1613 *energy_q = energy_AB[0];
1614 m_add(vir_q, vir_AB[0], vir_q);
1618 *energy_q = (1.0-lambda_q)*energy_AB[0] + lambda_q*energy_AB[1];
1619 *dvdlambda_q += energy_AB[1] - energy_AB[0];
1620 for (i = 0; i < DIM; i++)
1622 for (j = 0; j < DIM; j++)
1624 vir_q[i][j] += (1.0-lambda_q)*vir_AB[0][i][j] +
1625 lambda_q*vir_AB[1][i][j];
1631 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1643 *energy_lj = energy_AB[2];
1644 m_add(vir_lj, vir_AB[2], vir_lj);
1648 *energy_lj = (1.0-lambda_lj)*energy_AB[2] + lambda_lj*energy_AB[3];
1649 *dvdlambda_lj += energy_AB[3] - energy_AB[2];
1650 for (i = 0; i < DIM; i++)
1652 for (j = 0; j < DIM; j++)
1654 vir_lj[i][j] += (1.0-lambda_lj)*vir_AB[2][i][j] + lambda_lj*vir_AB[3][i][j];
1660 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1671 void gmx_pme_destroy(gmx_pme_t *pme)
1678 delete pme->boxScaler;
1687 for (int i = 0; i < pme->ngrids; ++i)
1689 pmegrids_destroy(&pme->pmegrid[i]);
1691 if (pme->pfft_setup)
1693 for (int i = 0; i < pme->ngrids; ++i)
1695 gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1698 sfree(pme->fftgrid);
1699 sfree(pme->cfftgrid);
1700 sfree(pme->pfft_setup);
1702 for (int i = 0; i < std::max(1, pme->ndecompdim); i++) //pme->atc[0] is always allocated
1704 destroy_atomcomm(&pme->atc[i]);
1707 for (int i = 0; i < DIM; i++)
1709 sfree(pme->bsp_mod[i]);
1712 sfree(pme->lb_buf1);
1713 sfree(pme->lb_buf2);
1718 if (pme->solve_work)
1720 pme_free_all_work(&pme->solve_work, pme->nthread);
1723 sfree(pme->sum_qgrid_tmp);
1724 sfree(pme->sum_qgrid_dd_tmp);
1726 if (pme_gpu_active(pme) && pme->gpu)
1728 pme_gpu_destroy(pme->gpu);
1734 void gmx_pme_reinit_atoms(const gmx_pme_t *pme, const int nAtoms, const real *charges)
1736 if (pme_gpu_active(pme))
1738 pme_gpu_reinit_atoms(pme->gpu, nAtoms, charges);
1740 // TODO: handle the CPU case here; handle the whole t_mdatoms