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 The GROMACS development team.
7 * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
8 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
9 * and including many others, as listed in the AUTHORS file in the
10 * top-level source directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
40 * \brief This file contains function definitions necessary for
41 * computing energies and forces for the PME long-ranged part (Coulomb
44 * \author Erik Lindahl <erik@kth.se>
45 * \author Berk Hess <hess@kth.se>
46 * \ingroup module_ewald
48 /* IMPORTANT FOR DEVELOPERS:
50 * Triclinic pme stuff isn't entirely trivial, and we've experienced
51 * some bugs during development (many of them due to me). To avoid
52 * this in the future, please check the following things if you make
53 * changes in this file:
55 * 1. You should obtain identical (at least to the PME precision)
56 * energies, forces, and virial for
57 * a rectangular box and a triclinic one where the z (or y) axis is
58 * tilted a whole box side. For instance you could use these boxes:
60 * rectangular triclinic
65 * 2. You should check the energy conservation in a triclinic box.
67 * It might seem an overkill, but better safe than sorry.
86 #include "gromacs/domdec/domdec.h"
87 #include "gromacs/ewald/ewald_utils.h"
88 #include "gromacs/fft/parallel_3dfft.h"
89 #include "gromacs/fileio/pdbio.h"
90 #include "gromacs/gmxlib/network.h"
91 #include "gromacs/gmxlib/nrnb.h"
92 #include "gromacs/hardware/hw_info.h"
93 #include "gromacs/math/gmxcomplex.h"
94 #include "gromacs/math/invertmatrix.h"
95 #include "gromacs/math/units.h"
96 #include "gromacs/math/vec.h"
97 #include "gromacs/math/vectypes.h"
98 #include "gromacs/mdtypes/commrec.h"
99 #include "gromacs/mdtypes/forcerec.h"
100 #include "gromacs/mdtypes/inputrec.h"
101 #include "gromacs/mdtypes/md_enums.h"
102 #include "gromacs/mdtypes/simulation_workload.h"
103 #include "gromacs/pbcutil/pbc.h"
104 #include "gromacs/timing/cyclecounter.h"
105 #include "gromacs/timing/wallcycle.h"
106 #include "gromacs/timing/walltime_accounting.h"
107 #include "gromacs/topology/topology.h"
108 #include "gromacs/utility/basedefinitions.h"
109 #include "gromacs/utility/cstringutil.h"
110 #include "gromacs/utility/exceptions.h"
111 #include "gromacs/utility/fatalerror.h"
112 #include "gromacs/utility/gmxmpi.h"
113 #include "gromacs/utility/gmxomp.h"
114 #include "gromacs/utility/logger.h"
115 #include "gromacs/utility/real.h"
116 #include "gromacs/utility/smalloc.h"
117 #include "gromacs/utility/stringutil.h"
118 #include "gromacs/utility/unique_cptr.h"
120 #include "calculate_spline_moduli.h"
121 #include "pme_gather.h"
122 #include "pme_gpu_internal.h"
123 #include "pme_grid.h"
124 #include "pme_internal.h"
125 #include "pme_redistribute.h"
126 #include "pme_solve.h"
127 #include "pme_spline_work.h"
128 #include "pme_spread.h"
130 /*! \brief Help build a descriptive message in \c error if there are
131 * \c errorReasons why PME on GPU is not supported.
133 * \returns Whether the lack of errorReasons indicate there is support. */
134 static bool addMessageIfNotSupported(const std::list<std::string>& errorReasons, std::string* error)
136 bool isSupported = errorReasons.empty();
137 if (!isSupported && error)
139 std::string regressionTestMarker = "PME GPU does not support";
140 // this prefix is tested for in the regression tests script gmxtest.pl
141 *error = regressionTestMarker;
142 if (errorReasons.size() == 1)
144 *error += " " + errorReasons.back();
148 *error += ": " + gmx::joinStrings(errorReasons, "; ");
155 bool pme_gpu_supports_build(std::string* error)
157 std::list<std::string> errorReasons;
160 errorReasons.emplace_back("a double-precision build");
162 if (GMX_GPU == GMX_GPU_NONE)
164 errorReasons.emplace_back("a non-GPU build");
166 return addMessageIfNotSupported(errorReasons, error);
169 bool pme_gpu_supports_hardware(const gmx_hw_info_t gmx_unused& hwinfo, std::string* error)
171 std::list<std::string> errorReasons;
173 if (GMX_GPU == GMX_GPU_OPENCL)
176 errorReasons.emplace_back("Apple OS X operating system");
179 return addMessageIfNotSupported(errorReasons, error);
182 bool pme_gpu_supports_input(const t_inputrec& ir, const gmx_mtop_t& mtop, std::string* error)
184 std::list<std::string> errorReasons;
185 if (!EEL_PME(ir.coulombtype))
187 errorReasons.emplace_back("systems that do not use PME for electrostatics");
189 if (ir.pme_order != 4)
191 errorReasons.emplace_back("interpolation orders other than 4");
193 if (ir.efep != efepNO)
195 if (gmx_mtop_has_perturbed_charges(mtop))
197 errorReasons.emplace_back(
198 "free energy calculations with perturbed charges (multiple grids)");
201 if (EVDW_PME(ir.vdwtype))
203 errorReasons.emplace_back("Lennard-Jones PME");
205 if (!EI_DYNAMICS(ir.eI))
207 errorReasons.emplace_back(
208 "Cannot compute PME interactions on a GPU, because PME GPU requires a dynamical "
209 "integrator (md, sd, etc).");
211 return addMessageIfNotSupported(errorReasons, error);
214 /*! \brief \libinternal
215 * Finds out if PME with given inputs is possible to run on GPU.
216 * This function is an internal final check, validating the whole PME structure on creation,
217 * but it still duplicates the preliminary checks from the above (externally exposed) pme_gpu_supports_input() - just in case.
219 * \param[in] pme The PME structure.
220 * \param[out] error The error message if the input is not supported on GPU.
221 * \returns True if this PME input is possible to run on GPU, false otherwise.
223 static bool pme_gpu_check_restrictions(const gmx_pme_t* pme, std::string* error)
225 std::list<std::string> errorReasons;
226 if (pme->nnodes != 1)
228 errorReasons.emplace_back("PME decomposition");
230 if (pme->pme_order != 4)
232 errorReasons.emplace_back("interpolation orders other than 4");
236 errorReasons.emplace_back("free energy calculations (multiple grids)");
240 errorReasons.emplace_back("Lennard-Jones PME");
244 errorReasons.emplace_back("double precision");
246 if (GMX_GPU == GMX_GPU_NONE)
248 errorReasons.emplace_back("non-GPU build of GROMACS");
251 return addMessageIfNotSupported(errorReasons, error);
254 PmeRunMode pme_run_mode(const gmx_pme_t* pme)
256 GMX_ASSERT(pme != nullptr, "Expecting valid PME data pointer");
260 gmx::PinningPolicy pme_get_pinning_policy()
262 return gmx::PinningPolicy::PinnedIfSupported;
265 /*! \brief Number of bytes in a cache line.
267 * Must also be a multiple of the SIMD and SIMD4 register size, to
268 * preserve alignment.
270 const int gmxCacheLineSize = 64;
272 //! Set up coordinate communication
273 static void setup_coordinate_communication(PmeAtomComm* atc)
281 for (i = 1; i <= nslab / 2; i++)
283 fw = (atc->nodeid + i) % nslab;
284 bw = (atc->nodeid - i + nslab) % nslab;
287 atc->slabCommSetup[n].node_dest = fw;
288 atc->slabCommSetup[n].node_src = bw;
293 atc->slabCommSetup[n].node_dest = bw;
294 atc->slabCommSetup[n].node_src = fw;
300 /*! \brief Round \p n up to the next multiple of \p f */
301 static int mult_up(int n, int f)
303 return ((n + f - 1) / f) * f;
306 /*! \brief Return estimate of the load imbalance from the PME grid not being a good match for the number of PME ranks */
307 static double estimate_pme_load_imbalance(struct gmx_pme_t* pme)
312 nma = pme->nnodes_major;
313 nmi = pme->nnodes_minor;
315 n1 = mult_up(pme->nkx, nma) * mult_up(pme->nky, nmi) * pme->nkz;
316 n2 = mult_up(pme->nkx, nma) * mult_up(pme->nkz, nmi) * pme->nky;
317 n3 = mult_up(pme->nky, nma) * mult_up(pme->nkz, nmi) * pme->nkx;
319 /* pme_solve is roughly double the cost of an fft */
321 return (n1 + n2 + 3 * n3) / static_cast<double>(6 * pme->nkx * pme->nky * pme->nkz);
326 PmeAtomComm::PmeAtomComm(MPI_Comm PmeMpiCommunicator,
327 const int numThreads,
330 const bool doSpread) :
337 if (PmeMpiCommunicator != MPI_COMM_NULL)
339 mpi_comm = PmeMpiCommunicator;
341 MPI_Comm_size(mpi_comm, &nslab);
342 MPI_Comm_rank(mpi_comm, &nodeid);
347 fprintf(debug, "For PME atom communication in dimind %d: nslab %d rank %d\n", dimind, nslab, nodeid);
352 slabCommSetup.resize(nslab);
353 setup_coordinate_communication(this);
355 count_thread.resize(nthread);
356 for (auto& countThread : count_thread)
358 countThread.resize(nslab);
364 threadMap.resize(nthread);
366 # pragma omp parallel for num_threads(nthread) schedule(static)
367 for (int thread = 0; thread < nthread; thread++)
371 /* Allocate buffer with padding to avoid cache polution */
372 threadMap[thread].nBuffer.resize(nthread + 2 * gmxCacheLineSize);
373 threadMap[thread].n = threadMap[thread].nBuffer.data() + gmxCacheLineSize;
375 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
382 /*! \brief Initialize data structure for communication */
383 static void init_overlap_comm(pme_overlap_t* ol, int norder, MPI_Comm comm, int nnodes, int nodeid, int ndata, int commplainsize)
391 /* Linear translation of the PME grid won't affect reciprocal space
392 * calculations, so to optimize we only interpolate "upwards",
393 * which also means we only have to consider overlap in one direction.
394 * I.e., particles on this node might also be spread to grid indices
395 * that belong to higher nodes (modulo nnodes)
398 ol->s2g0.resize(ol->nnodes + 1);
399 ol->s2g1.resize(ol->nnodes);
402 fprintf(debug, "PME slab boundaries:");
404 for (int i = 0; i < nnodes; i++)
406 /* s2g0 the local interpolation grid start.
407 * s2g1 the local interpolation grid end.
408 * Since in calc_pidx we divide particles, and not grid lines,
409 * spatially uniform along dimension x or y, we need to round
410 * s2g0 down and s2g1 up.
412 ol->s2g0[i] = (i * ndata + 0) / nnodes;
413 ol->s2g1[i] = ((i + 1) * ndata + nnodes - 1) / nnodes + norder - 1;
417 fprintf(debug, " %3d %3d", ol->s2g0[i], ol->s2g1[i]);
420 ol->s2g0[nnodes] = ndata;
423 fprintf(debug, "\n");
426 /* Determine with how many nodes we need to communicate the grid overlap */
427 int testRankCount = 0;
432 for (int i = 0; i < nnodes; i++)
434 if ((i + testRankCount < nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount])
435 || (i + testRankCount >= nnodes && ol->s2g1[i] > ol->s2g0[i + testRankCount - nnodes] + ndata))
440 } while (bCont && testRankCount < nnodes);
442 ol->comm_data.resize(testRankCount - 1);
445 for (size_t b = 0; b < ol->comm_data.size(); b++)
447 pme_grid_comm_t* pgc = &ol->comm_data[b];
450 pgc->send_id = (ol->nodeid + (b + 1)) % ol->nnodes;
451 int fft_start = ol->s2g0[pgc->send_id];
452 int fft_end = ol->s2g0[pgc->send_id + 1];
453 if (pgc->send_id < nodeid)
458 int send_index1 = ol->s2g1[nodeid];
459 send_index1 = std::min(send_index1, fft_end);
460 pgc->send_index0 = fft_start;
461 pgc->send_nindex = std::max(0, send_index1 - pgc->send_index0);
462 ol->send_size += pgc->send_nindex;
464 /* We always start receiving to the first index of our slab */
465 pgc->recv_id = (ol->nodeid - (b + 1) + ol->nnodes) % ol->nnodes;
466 fft_start = ol->s2g0[ol->nodeid];
467 fft_end = ol->s2g0[ol->nodeid + 1];
468 int recv_index1 = ol->s2g1[pgc->recv_id];
469 if (pgc->recv_id > nodeid)
471 recv_index1 -= ndata;
473 recv_index1 = std::min(recv_index1, fft_end);
474 pgc->recv_index0 = fft_start;
475 pgc->recv_nindex = std::max(0, recv_index1 - pgc->recv_index0);
479 /* Communicate the buffer sizes to receive */
481 for (size_t b = 0; b < ol->comm_data.size(); b++)
483 MPI_Sendrecv(&ol->send_size, 1, MPI_INT, ol->comm_data[b].send_id, b, &ol->comm_data[b].recv_size,
484 1, MPI_INT, ol->comm_data[b].recv_id, b, ol->mpi_comm, &stat);
488 /* For non-divisible grid we need pme_order iso pme_order-1 */
489 ol->sendbuf.resize(norder * commplainsize);
490 ol->recvbuf.resize(norder * commplainsize);
493 int minimalPmeGridSize(int pmeOrder)
495 /* The actual grid size limitations are:
496 * serial: >= pme_order
497 * DD, no OpenMP: >= 2*(pme_order - 1)
498 * DD, OpenMP: >= pme_order + 1
499 * But we use the maximum for simplicity since in practice there is not
500 * much performance difference between pme_order and 2*(pme_order -1).
502 int minimalSize = 2 * (pmeOrder - 1);
504 GMX_RELEASE_ASSERT(pmeOrder >= 3, "pmeOrder has to be >= 3");
505 GMX_RELEASE_ASSERT(minimalSize >= pmeOrder + 1, "The grid size should be >= pmeOrder + 1");
510 bool gmx_pme_check_restrictions(int pme_order, int nkx, int nky, int nkz, int numPmeDomainsAlongX, bool useThreads, bool errorsAreFatal)
512 if (pme_order > PME_ORDER_MAX)
519 std::string message = gmx::formatString(
520 "pme_order (%d) is larger than the maximum allowed value (%d). Modify and "
521 "recompile the code if you really need such a high order.",
522 pme_order, PME_ORDER_MAX);
523 GMX_THROW(gmx::InconsistentInputError(message));
526 const int minGridSize = minimalPmeGridSize(pme_order);
527 if (nkx < minGridSize || nky < minGridSize || nkz < minGridSize)
533 std::string message =
534 gmx::formatString("The PME grid sizes need to be >= 2*(pme_order-1) (%d)", minGridSize);
535 GMX_THROW(gmx::InconsistentInputError(message));
538 /* Check for a limitation of the (current) sum_fftgrid_dd code.
539 * We only allow multiple communication pulses in dim 1, not in dim 0.
542 && (nkx < numPmeDomainsAlongX * pme_order && nkx != numPmeDomainsAlongX * (pme_order - 1)))
549 "The number of PME grid lines per rank along x is %g. But when using OpenMP "
550 "threads, the number of grid lines per rank along x should be >= pme_order (%d) "
551 "or = pmeorder-1. To resolve this issue, use fewer ranks along x (and possibly "
552 "more along y and/or z) by specifying -dd manually.",
553 nkx / static_cast<double>(numPmeDomainsAlongX), pme_order);
559 /*! \brief Round \p enumerator */
560 static int div_round_up(int enumerator, int denominator)
562 return (enumerator + denominator - 1) / denominator;
565 gmx_pme_t* gmx_pme_init(const t_commrec* cr,
566 const NumPmeDomains& numPmeDomains,
567 const t_inputrec* ir,
568 gmx_bool bFreeEnergy_q,
569 gmx_bool bFreeEnergy_lj,
570 gmx_bool bReproducible,
576 const DeviceContext* deviceContext,
577 const DeviceStream* deviceStream,
578 const PmeGpuProgram* pmeGpuProgram,
579 const gmx::MDLogger& /*mdlog*/)
581 int use_threads, sum_use_threads, i;
586 fprintf(debug, "Creating PME data structures.\n");
589 gmx::unique_cptr<gmx_pme_t, gmx_pme_destroy> pme(new gmx_pme_t());
591 pme->sum_qgrid_tmp = nullptr;
592 pme->sum_qgrid_dd_tmp = nullptr;
599 pme->nnodes_major = numPmeDomains.x;
600 pme->nnodes_minor = numPmeDomains.y;
602 if (numPmeDomains.x * numPmeDomains.y > 1)
604 pme->mpi_comm = cr->mpi_comm_mygroup;
607 MPI_Comm_rank(pme->mpi_comm, &pme->nodeid);
608 MPI_Comm_size(pme->mpi_comm, &pme->nnodes);
610 if (pme->nnodes != numPmeDomains.x * numPmeDomains.y)
612 gmx_incons("PME rank count mismatch");
617 pme->mpi_comm = MPI_COMM_NULL;
620 if (pme->nnodes == 1)
622 pme->mpi_comm_d[0] = MPI_COMM_NULL;
623 pme->mpi_comm_d[1] = MPI_COMM_NULL;
625 pme->nodeid_major = 0;
626 pme->nodeid_minor = 0;
630 if (numPmeDomains.y == 1)
632 pme->mpi_comm_d[0] = pme->mpi_comm;
633 pme->mpi_comm_d[1] = MPI_COMM_NULL;
635 pme->nodeid_major = pme->nodeid;
636 pme->nodeid_minor = 0;
638 else if (numPmeDomains.x == 1)
640 pme->mpi_comm_d[0] = MPI_COMM_NULL;
641 pme->mpi_comm_d[1] = pme->mpi_comm;
643 pme->nodeid_major = 0;
644 pme->nodeid_minor = pme->nodeid;
648 if (pme->nnodes % numPmeDomains.x != 0)
651 "For 2D PME decomposition, #PME ranks must be divisible by the number of "
657 MPI_Comm_split(pme->mpi_comm, pme->nodeid % numPmeDomains.y, pme->nodeid,
658 &pme->mpi_comm_d[0]); /* My communicator along major dimension */
659 MPI_Comm_split(pme->mpi_comm, pme->nodeid / numPmeDomains.y, pme->nodeid,
660 &pme->mpi_comm_d[1]); /* My communicator along minor dimension */
662 MPI_Comm_rank(pme->mpi_comm_d[0], &pme->nodeid_major);
663 MPI_Comm_size(pme->mpi_comm_d[0], &pme->nnodes_major);
664 MPI_Comm_rank(pme->mpi_comm_d[1], &pme->nodeid_minor);
665 MPI_Comm_size(pme->mpi_comm_d[1], &pme->nnodes_minor);
669 // cr is always initialized if there is a a PP rank, so we can safely assume
670 // that when it is not, like in ewald tests, we not on a PP rank.
671 pme->bPPnode = ((cr != nullptr && cr->duty != 0) && thisRankHasDuty(cr, DUTY_PP));
673 pme->nthread = nthread;
675 /* Check if any of the PME MPI ranks uses threads */
676 use_threads = (pme->nthread > 1 ? 1 : 0);
680 MPI_Allreduce(&use_threads, &sum_use_threads, 1, MPI_INT, MPI_SUM, pme->mpi_comm);
685 sum_use_threads = use_threads;
687 pme->bUseThreads = (sum_use_threads > 0);
689 if (ir->pbcType == PbcType::Screw)
691 gmx_fatal(FARGS, "pme does not (yet) work with pbc = screw");
695 * It is likely that the current gmx_pme_do() routine supports calculating
696 * only Coulomb or LJ while gmx_pme_init() configures for both,
697 * but that has never been tested.
698 * It is likely that the current gmx_pme_do() routine supports calculating,
699 * not calculating free-energy for Coulomb and/or LJ while gmx_pme_init()
700 * configures with free-energy, but that has never been tested.
702 pme->doCoulomb = EEL_PME(ir->coulombtype);
703 pme->doLJ = EVDW_PME(ir->vdwtype);
704 pme->bFEP_q = ((ir->efep != efepNO) && bFreeEnergy_q);
705 pme->bFEP_lj = ((ir->efep != efepNO) && bFreeEnergy_lj);
706 pme->bFEP = (pme->bFEP_q || pme->bFEP_lj);
710 pme->bP3M = (ir->coulombtype == eelP3M_AD || getenv("GMX_PME_P3M") != nullptr);
711 pme->pme_order = ir->pme_order;
712 pme->ewaldcoeff_q = ewaldcoeff_q;
713 pme->ewaldcoeff_lj = ewaldcoeff_lj;
715 /* Always constant electrostatics coefficients */
716 pme->epsilon_r = ir->epsilon_r;
718 /* Always constant LJ coefficients */
719 pme->ljpme_combination_rule = ir->ljpme_combination_rule;
721 // The box requires scaling with nwalls = 2, we store that condition as well
722 // as the scaling factor
723 delete pme->boxScaler;
724 pme->boxScaler = new EwaldBoxZScaler(*ir);
726 /* If we violate restrictions, generate a fatal error here */
727 gmx_pme_check_restrictions(pme->pme_order, pme->nkx, pme->nky, pme->nkz, pme->nnodes_major,
728 pme->bUseThreads, true);
735 MPI_Type_contiguous(DIM, GMX_MPI_REAL, &(pme->rvec_mpi));
736 MPI_Type_commit(&(pme->rvec_mpi));
739 /* Note that the coefficient spreading and force gathering, which usually
740 * takes about the same amount of time as FFT+solve_pme,
741 * is always fully load balanced
742 * (unless the coefficient distribution is inhomogeneous).
745 imbal = estimate_pme_load_imbalance(pme.get());
746 if (imbal >= 1.2 && pme->nodeid_major == 0 && pme->nodeid_minor == 0)
750 "NOTE: The load imbalance in PME FFT and solve is %d%%.\n"
751 " For optimal PME load balancing\n"
752 " PME grid_x (%d) and grid_y (%d) should be divisible by #PME_ranks_x "
754 " and PME grid_y (%d) and grid_z (%d) should be divisible by #PME_ranks_y "
757 gmx::roundToInt((imbal - 1) * 100), pme->nkx, pme->nky, pme->nnodes_major,
758 pme->nky, pme->nkz, pme->nnodes_minor);
762 /* For non-divisible grid we need pme_order iso pme_order-1 */
763 /* In sum_qgrid_dd x overlap is copied in place: take padding into account.
764 * y is always copied through a buffer: we don't need padding in z,
765 * but we do need the overlap in x because of the communication order.
767 init_overlap_comm(&pme->overlap[0], pme->pme_order, pme->mpi_comm_d[0], pme->nnodes_major,
768 pme->nodeid_major, pme->nkx,
769 (div_round_up(pme->nky, pme->nnodes_minor) + pme->pme_order)
770 * (pme->nkz + pme->pme_order - 1));
772 /* Along overlap dim 1 we can send in multiple pulses in sum_fftgrid_dd.
773 * We do this with an offset buffer of equal size, so we need to allocate
774 * extra for the offset. That's what the (+1)*pme->nkz is for.
776 init_overlap_comm(&pme->overlap[1], pme->pme_order, pme->mpi_comm_d[1], pme->nnodes_minor,
777 pme->nodeid_minor, pme->nky,
778 (div_round_up(pme->nkx, pme->nnodes_major) + pme->pme_order + 1) * pme->nkz);
780 /* Double-check for a limitation of the (current) sum_fftgrid_dd code.
781 * Note that gmx_pme_check_restrictions checked for this already.
783 if (pme->bUseThreads && (pme->overlap[0].comm_data.size() > 1))
786 "More than one communication pulse required for grid overlap communication along "
787 "the major dimension while using threads");
790 snew(pme->bsp_mod[XX], pme->nkx);
791 snew(pme->bsp_mod[YY], pme->nky);
792 snew(pme->bsp_mod[ZZ], pme->nkz);
794 pme->gpu = pmeGpu; /* Carrying over the single GPU structure */
795 pme->runMode = runMode;
797 /* The required size of the interpolation grid, including overlap.
798 * The allocated size (pmegrid_n?) might be slightly larger.
800 pme->pmegrid_nx = pme->overlap[0].s2g1[pme->nodeid_major] - pme->overlap[0].s2g0[pme->nodeid_major];
801 pme->pmegrid_ny = pme->overlap[1].s2g1[pme->nodeid_minor] - pme->overlap[1].s2g0[pme->nodeid_minor];
802 pme->pmegrid_nz_base = pme->nkz;
803 pme->pmegrid_nz = pme->pmegrid_nz_base + pme->pme_order - 1;
804 set_grid_alignment(&pme->pmegrid_nz, pme->pme_order);
805 pme->pmegrid_start_ix = pme->overlap[0].s2g0[pme->nodeid_major];
806 pme->pmegrid_start_iy = pme->overlap[1].s2g0[pme->nodeid_minor];
807 pme->pmegrid_start_iz = 0;
809 make_gridindex_to_localindex(pme->nkx, pme->pmegrid_start_ix,
810 pme->pmegrid_nx - (pme->pme_order - 1), &pme->nnx, &pme->fshx);
811 make_gridindex_to_localindex(pme->nky, pme->pmegrid_start_iy,
812 pme->pmegrid_ny - (pme->pme_order - 1), &pme->nny, &pme->fshy);
813 make_gridindex_to_localindex(pme->nkz, pme->pmegrid_start_iz, pme->pmegrid_nz_base, &pme->nnz,
816 pme->spline_work = make_pme_spline_work(pme->pme_order);
821 /* It doesn't matter if we allocate too many grids here,
822 * we only allocate and use the ones we need.
826 pme->ngrids = ((ir->ljpme_combination_rule == eljpmeLB) ? DO_Q_AND_LJ_LB : DO_Q_AND_LJ);
832 snew(pme->fftgrid, pme->ngrids);
833 snew(pme->cfftgrid, pme->ngrids);
834 snew(pme->pfft_setup, pme->ngrids);
836 for (i = 0; i < pme->ngrids; ++i)
838 if ((i < DO_Q && pme->doCoulomb && (i == 0 || bFreeEnergy_q))
839 || (i >= DO_Q && pme->doLJ
840 && (i == 2 || bFreeEnergy_lj || ir->ljpme_combination_rule == eljpmeLB)))
842 pmegrids_init(&pme->pmegrid[i], pme->pmegrid_nx, pme->pmegrid_ny, pme->pmegrid_nz,
843 pme->pmegrid_nz_base, pme->pme_order, pme->bUseThreads, pme->nthread,
844 pme->overlap[0].s2g1[pme->nodeid_major]
845 - pme->overlap[0].s2g0[pme->nodeid_major + 1],
846 pme->overlap[1].s2g1[pme->nodeid_minor]
847 - pme->overlap[1].s2g0[pme->nodeid_minor + 1]);
848 /* This routine will allocate the grid data to fit the FFTs */
849 const auto allocateRealGridForGpu = (pme->runMode == PmeRunMode::Mixed)
850 ? gmx::PinningPolicy::PinnedIfSupported
851 : gmx::PinningPolicy::CannotBePinned;
852 gmx_parallel_3dfft_init(&pme->pfft_setup[i], ndata, &pme->fftgrid[i], &pme->cfftgrid[i],
853 pme->mpi_comm_d, bReproducible, pme->nthread, allocateRealGridForGpu);
859 /* Use plain SPME B-spline interpolation */
860 make_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
864 /* Use the P3M grid-optimized influence function */
865 make_p3m_bspline_moduli(pme->bsp_mod, pme->nkx, pme->nky, pme->nkz, pme->pme_order);
868 /* Use atc[0] for spreading */
869 const int firstDimIndex = (numPmeDomains.x > 1 ? 0 : 1);
870 MPI_Comm mpiCommFirstDim = (pme->nnodes > 1 ? pme->mpi_comm_d[firstDimIndex] : MPI_COMM_NULL);
871 bool doSpread = true;
872 pme->atc.emplace_back(mpiCommFirstDim, pme->nthread, pme->pme_order, firstDimIndex, doSpread);
873 if (pme->ndecompdim >= 2)
875 const int secondDimIndex = 1;
877 pme->atc.emplace_back(pme->mpi_comm_d[1], pme->nthread, pme->pme_order, secondDimIndex, doSpread);
880 // Initial check of validity of the input for running on the GPU
881 if (pme->runMode != PmeRunMode::CPU)
883 std::string errorString;
884 bool canRunOnGpu = pme_gpu_check_restrictions(pme.get(), &errorString);
887 GMX_THROW(gmx::NotImplementedError(errorString));
889 pme_gpu_reinit(pme.get(), deviceContext, deviceStream, pmeGpuProgram);
893 GMX_ASSERT(pme->gpu == nullptr, "Should not have PME GPU object when PME is on a CPU.");
897 pme_init_all_work(&pme->solve_work, pme->nthread, pme->nkx);
899 // no exception was thrown during the init, so we hand over the PME structure handle
900 return pme.release();
903 void gmx_pme_reinit(struct gmx_pme_t** pmedata,
905 struct gmx_pme_t* pme_src,
906 const t_inputrec* ir,
907 const ivec grid_size,
911 // Create a copy of t_inputrec fields that are used in gmx_pme_init().
912 // TODO: This would be better as just copying a sub-structure that contains
913 // all the PME parameters and nothing else.
915 irc.pbcType = ir->pbcType;
916 irc.coulombtype = ir->coulombtype;
917 irc.vdwtype = ir->vdwtype;
919 irc.pme_order = ir->pme_order;
920 irc.epsilon_r = ir->epsilon_r;
921 irc.ljpme_combination_rule = ir->ljpme_combination_rule;
922 irc.nkx = grid_size[XX];
923 irc.nky = grid_size[YY];
924 irc.nkz = grid_size[ZZ];
928 const gmx::MDLogger dummyLogger;
929 // This is reinit which is currently only changing grid size/coefficients,
930 // so we don't expect the actual logging.
931 // TODO: when PME is an object, it should take reference to mdlog on construction and save it.
932 GMX_ASSERT(pmedata, "Invalid PME pointer");
933 NumPmeDomains numPmeDomains = { pme_src->nnodes_major, pme_src->nnodes_minor };
934 *pmedata = gmx_pme_init(cr, numPmeDomains, &irc, pme_src->bFEP_q, pme_src->bFEP_lj, FALSE,
935 ewaldcoeff_q, ewaldcoeff_lj, pme_src->nthread, pme_src->runMode,
936 pme_src->gpu, nullptr, nullptr, nullptr, dummyLogger);
937 /* When running PME on the CPU not using domain decomposition,
938 * the atom data is allocated once only in gmx_pme_(re)init().
940 if (!pme_src->gpu && pme_src->nnodes == 1)
942 gmx_pme_reinit_atoms(*pmedata, pme_src->atc[0].numAtoms(), nullptr);
944 // TODO this is mostly passing around current values
946 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
948 /* We can easily reuse the allocated pme grids in pme_src */
949 reuse_pmegrids(&pme_src->pmegrid[PME_GRID_QA], &(*pmedata)->pmegrid[PME_GRID_QA]);
950 /* We would like to reuse the fft grids, but that's harder */
953 void gmx_pme_calc_energy(gmx_pme_t* pme, gmx::ArrayRef<const gmx::RVec> x, gmx::ArrayRef<const real> q, real* V)
959 gmx_incons("gmx_pme_calc_energy called in parallel");
963 gmx_incons("gmx_pme_calc_energy with free energy");
966 if (!pme->atc_energy)
968 pme->atc_energy = std::make_unique<PmeAtomComm>(MPI_COMM_NULL, 1, pme->pme_order, 0, true);
970 PmeAtomComm* atc = pme->atc_energy.get();
971 atc->setNumAtoms(x.ssize());
973 atc->coefficient = q;
975 /* We only use the A-charges grid */
976 grid = &pme->pmegrid[PME_GRID_QA];
978 /* Only calculate the spline coefficients, don't actually spread */
979 spread_on_grid(pme, atc, nullptr, TRUE, FALSE, pme->fftgrid[PME_GRID_QA], FALSE, PME_GRID_QA);
981 *V = gather_energy_bsplines(pme, grid->grid.grid, atc);
984 /*! \brief Calculate initial Lorentz-Berthelot coefficients for LJ-PME */
985 static void calc_initial_lb_coeffs(gmx::ArrayRef<real> coefficient, const real* local_c6, const real* local_sigma)
987 for (gmx::index i = 0; i < coefficient.ssize(); ++i)
989 real sigma4 = local_sigma[i];
990 sigma4 = sigma4 * sigma4;
991 sigma4 = sigma4 * sigma4;
992 coefficient[i] = local_c6[i] / sigma4;
996 /*! \brief Calculate next Lorentz-Berthelot coefficients for LJ-PME */
997 static void calc_next_lb_coeffs(gmx::ArrayRef<real> coefficient, const real* local_sigma)
999 for (gmx::index i = 0; i < coefficient.ssize(); ++i)
1001 coefficient[i] *= local_sigma[i];
1005 int gmx_pme_do(struct gmx_pme_t* pme,
1006 gmx::ArrayRef<const gmx::RVec> coordinates,
1007 gmx::ArrayRef<gmx::RVec> forces,
1015 const t_commrec* cr,
1019 gmx_wallcycle* wcycle,
1028 const gmx::StepWorkload& stepWork)
1030 GMX_ASSERT(pme->runMode == PmeRunMode::CPU,
1031 "gmx_pme_do should not be called on the GPU PME run.");
1033 int d, npme, grid_index, max_grid_index;
1034 PmeAtomComm& atc = pme->atc[0];
1035 pmegrids_t* pmegrid = nullptr;
1036 real* grid = nullptr;
1037 real* coefficient = nullptr;
1038 PmeOutput output[2]; // The second is used for the B state with FEP
1041 gmx_parallel_3dfft_t pfft_setup;
1043 t_complex* cfftgrid;
1045 gmx_bool bFirst, bDoSplines;
1047 int fep_states_lj = pme->bFEP_lj ? 2 : 1;
1048 // There's no support for computing energy without virial, or vice versa
1049 const bool computeEnergyAndVirial = (stepWork.computeEnergy || stepWork.computeVirial);
1051 /* We could be passing lambda!=0 while no q or LJ is actually perturbed */
1061 assert(pme->nnodes > 0);
1062 assert(pme->nnodes == 1 || pme->ndecompdim > 0);
1064 if (pme->nnodes > 1)
1066 atc.pd.resize(coordinates.ssize());
1067 for (int d = pme->ndecompdim - 1; d >= 0; d--)
1069 PmeAtomComm& atc = pme->atc[d];
1070 atc.maxshift = (atc.dimind == 0 ? maxshift_x : maxshift_y);
1075 GMX_ASSERT(coordinates.ssize() == atc.numAtoms(), "We expect atc.numAtoms() coordinates");
1076 GMX_ASSERT(forces.ssize() >= atc.numAtoms(),
1077 "We need a force buffer with at least atc.numAtoms() elements");
1079 atc.x = coordinates;
1084 pme->boxScaler->scaleBox(box, scaledBox);
1086 gmx::invertBoxMatrix(scaledBox, pme->recipbox);
1089 /* For simplicity, we construct the splines for all particles if
1090 * more than one PME calculations is needed. Some optimization
1091 * could be done by keeping track of which atoms have splines
1092 * constructed, and construct new splines on each pass for atoms
1093 * that don't yet have them.
1096 bDoSplines = pme->bFEP || (pme->doCoulomb && pme->doLJ);
1098 /* We need a maximum of four separate PME calculations:
1099 * grid_index=0: Coulomb PME with charges from state A
1100 * grid_index=1: Coulomb PME with charges from state B
1101 * grid_index=2: LJ PME with C6 from state A
1102 * grid_index=3: LJ PME with C6 from state B
1103 * For Lorentz-Berthelot combination rules, a separate loop is used to
1104 * calculate all the terms
1107 /* If we are doing LJ-PME with LB, we only do Q here */
1108 max_grid_index = (pme->ljpme_combination_rule == eljpmeLB) ? DO_Q : DO_Q_AND_LJ;
1110 for (grid_index = 0; grid_index < max_grid_index; ++grid_index)
1112 /* Check if we should do calculations at this grid_index
1113 * If grid_index is odd we should be doing FEP
1114 * If grid_index < 2 we should be doing electrostatic PME
1115 * If grid_index >= 2 we should be doing LJ-PME
1117 if ((grid_index < DO_Q && (!pme->doCoulomb || (grid_index == 1 && !pme->bFEP_q)))
1118 || (grid_index >= DO_Q && (!pme->doLJ || (grid_index == 3 && !pme->bFEP_lj))))
1122 /* Unpack structure */
1123 pmegrid = &pme->pmegrid[grid_index];
1124 fftgrid = pme->fftgrid[grid_index];
1125 cfftgrid = pme->cfftgrid[grid_index];
1126 pfft_setup = pme->pfft_setup[grid_index];
1129 case 0: coefficient = chargeA; break;
1130 case 1: coefficient = chargeB; break;
1131 case 2: coefficient = c6A; break;
1132 case 3: coefficient = c6B; break;
1135 grid = pmegrid->grid.grid;
1139 fprintf(debug, "PME: number of ranks = %d, rank = %d\n", cr->nnodes, cr->nodeid);
1140 fprintf(debug, "Grid = %p\n", static_cast<void*>(grid));
1141 if (grid == nullptr)
1143 gmx_fatal(FARGS, "No grid!");
1147 if (pme->nnodes == 1)
1149 atc.coefficient = gmx::arrayRefFromArray(coefficient, coordinates.size());
1153 wallcycle_start(wcycle, ewcPME_REDISTXF);
1154 do_redist_pos_coeffs(pme, cr, bFirst, coordinates, coefficient);
1156 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1161 fprintf(debug, "Rank= %6d, pme local particles=%6d\n", cr->nodeid, atc.numAtoms());
1164 wallcycle_start(wcycle, ewcPME_SPREAD);
1166 /* Spread the coefficients on a grid */
1167 spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1171 inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1173 inc_nrnb(nrnb, eNR_SPREADBSP, pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1175 if (!pme->bUseThreads)
1177 wrap_periodic_pmegrid(pme, grid);
1179 /* sum contributions to local grid from other nodes */
1180 if (pme->nnodes > 1)
1182 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1185 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1188 wallcycle_stop(wcycle, ewcPME_SPREAD);
1190 /* TODO If the OpenMP and single-threaded implementations
1191 converge, then spread_on_grid() and
1192 copy_pmegrid_to_fftgrid() will perhaps live in the same
1196 /* Here we start a large thread parallel region */
1197 #pragma omp parallel num_threads(pme->nthread) private(thread)
1201 thread = gmx_omp_get_thread_num();
1207 wallcycle_start(wcycle, ewcPME_FFT);
1209 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1212 wallcycle_stop(wcycle, ewcPME_FFT);
1215 /* solve in k-space for our local cells */
1218 wallcycle_start(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1220 if (grid_index < DO_Q)
1222 loop_count = solve_pme_yzx(
1223 pme, cfftgrid, scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1224 computeEnergyAndVirial, pme->nthread, thread);
1229 solve_pme_lj_yzx(pme, &cfftgrid, FALSE,
1230 scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1231 computeEnergyAndVirial, pme->nthread, thread);
1236 wallcycle_stop(wcycle, (grid_index < DO_Q ? ewcPME_SOLVE : ewcLJPME));
1237 inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1243 wallcycle_start(wcycle, ewcPME_FFT);
1245 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1248 wallcycle_stop(wcycle, ewcPME_FFT);
1251 if (pme->nodeid == 0)
1253 real ntot = pme->nkx * pme->nky * pme->nkz;
1254 npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1255 inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1258 /* Note: this wallcycle region is closed below
1259 outside an OpenMP region, so take care if
1260 refactoring code here. */
1261 wallcycle_start(wcycle, ewcPME_GATHER);
1264 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1266 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1268 /* End of thread parallel section.
1269 * With MPI we have to synchronize here before gmx_sum_qgrid_dd.
1272 /* distribute local grid to all nodes */
1273 if (pme->nnodes > 1)
1275 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1278 unwrap_periodic_pmegrid(pme, grid);
1280 if (stepWork.computeForces)
1282 /* interpolate forces for our local atoms */
1285 /* If we are running without parallelization,
1286 * atc->f is the actual force array, not a buffer,
1287 * therefore we should not clear it.
1289 lambda = grid_index < DO_Q ? lambda_q : lambda_lj;
1290 bClearF = (bFirst && PAR(cr));
1291 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1292 for (thread = 0; thread < pme->nthread; thread++)
1296 gather_f_bsplines(pme, grid, bClearF, &atc, &atc.spline[thread],
1297 pme->bFEP ? (grid_index % 2 == 0 ? 1.0 - lambda : lambda) : 1.0);
1299 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1303 inc_nrnb(nrnb, eNR_GATHERFBSP,
1304 pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1305 /* Note: this wallcycle region is opened above inside an OpenMP
1306 region, so take care if refactoring code here. */
1307 wallcycle_stop(wcycle, ewcPME_GATHER);
1310 if (computeEnergyAndVirial)
1312 /* This should only be called on the master thread
1313 * and after the threads have synchronized.
1317 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1321 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[grid_index % 2]);
1325 } /* of grid_index-loop */
1327 /* For Lorentz-Berthelot combination rules in LJ-PME, we need to calculate
1330 if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB)
1332 /* Loop over A- and B-state if we are doing FEP */
1333 for (fep_state = 0; fep_state < fep_states_lj; ++fep_state)
1335 real *local_c6 = nullptr, *local_sigma = nullptr, *RedistC6 = nullptr, *RedistSigma = nullptr;
1336 gmx::ArrayRef<real> coefficientBuffer;
1337 if (pme->nnodes == 1)
1339 pme->lb_buf1.resize(atc.numAtoms());
1340 coefficientBuffer = pme->lb_buf1;
1345 local_sigma = sigmaA;
1349 local_sigma = sigmaB;
1351 default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1356 coefficientBuffer = atc.coefficientBuffer;
1361 RedistSigma = sigmaA;
1365 RedistSigma = sigmaB;
1367 default: gmx_incons("Trying to access wrong FEP-state in LJ-PME routine");
1369 wallcycle_start(wcycle, ewcPME_REDISTXF);
1371 do_redist_pos_coeffs(pme, cr, bFirst, coordinates, RedistC6);
1372 pme->lb_buf1.resize(atc.numAtoms());
1373 pme->lb_buf2.resize(atc.numAtoms());
1374 local_c6 = pme->lb_buf1.data();
1375 for (int i = 0; i < atc.numAtoms(); ++i)
1377 local_c6[i] = atc.coefficient[i];
1380 do_redist_pos_coeffs(pme, cr, FALSE, coordinates, RedistSigma);
1381 local_sigma = pme->lb_buf2.data();
1382 for (int i = 0; i < atc.numAtoms(); ++i)
1384 local_sigma[i] = atc.coefficient[i];
1387 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1389 atc.coefficient = coefficientBuffer;
1390 calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1392 /*Seven terms in LJ-PME with LB, grid_index < 2 reserved for electrostatics*/
1393 for (grid_index = 2; grid_index < 9; ++grid_index)
1395 /* Unpack structure */
1396 pmegrid = &pme->pmegrid[grid_index];
1397 fftgrid = pme->fftgrid[grid_index];
1398 pfft_setup = pme->pfft_setup[grid_index];
1399 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1400 grid = pmegrid->grid.grid;
1402 wallcycle_start(wcycle, ewcPME_SPREAD);
1403 /* Spread the c6 on a grid */
1404 spread_on_grid(pme, &atc, pmegrid, bFirst, TRUE, fftgrid, bDoSplines, grid_index);
1408 inc_nrnb(nrnb, eNR_WEIGHTS, DIM * atc.numAtoms());
1411 inc_nrnb(nrnb, eNR_SPREADBSP,
1412 pme->pme_order * pme->pme_order * pme->pme_order * atc.numAtoms());
1413 if (pme->nthread == 1)
1415 wrap_periodic_pmegrid(pme, grid);
1416 /* sum contributions to local grid from other nodes */
1417 if (pme->nnodes > 1)
1419 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_FORWARD);
1421 copy_pmegrid_to_fftgrid(pme, grid, fftgrid, grid_index);
1423 wallcycle_stop(wcycle, ewcPME_SPREAD);
1425 /*Here we start a large thread parallel region*/
1426 #pragma omp parallel num_threads(pme->nthread) private(thread)
1430 thread = gmx_omp_get_thread_num();
1434 wallcycle_start(wcycle, ewcPME_FFT);
1437 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_REAL_TO_COMPLEX, thread, wcycle);
1440 wallcycle_stop(wcycle, ewcPME_FFT);
1443 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1447 /* solve in k-space for our local cells */
1448 #pragma omp parallel num_threads(pme->nthread) private(thread)
1453 thread = gmx_omp_get_thread_num();
1456 wallcycle_start(wcycle, ewcLJPME);
1460 solve_pme_lj_yzx(pme, &pme->cfftgrid[2], TRUE,
1461 scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ],
1462 computeEnergyAndVirial, pme->nthread, thread);
1465 wallcycle_stop(wcycle, ewcLJPME);
1466 inc_nrnb(nrnb, eNR_SOLVEPME, loop_count);
1469 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1472 if (computeEnergyAndVirial)
1474 /* This should only be called on the master thread and
1475 * after the threads have synchronized.
1477 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output[fep_state]);
1480 bFirst = !pme->doCoulomb;
1481 calc_initial_lb_coeffs(coefficientBuffer, local_c6, local_sigma);
1482 for (grid_index = 8; grid_index >= 2; --grid_index)
1484 /* Unpack structure */
1485 pmegrid = &pme->pmegrid[grid_index];
1486 fftgrid = pme->fftgrid[grid_index];
1487 pfft_setup = pme->pfft_setup[grid_index];
1488 grid = pmegrid->grid.grid;
1489 calc_next_lb_coeffs(coefficientBuffer, local_sigma);
1490 #pragma omp parallel num_threads(pme->nthread) private(thread)
1494 thread = gmx_omp_get_thread_num();
1498 wallcycle_start(wcycle, ewcPME_FFT);
1501 gmx_parallel_3dfft_execute(pfft_setup, GMX_FFT_COMPLEX_TO_REAL, thread, wcycle);
1504 wallcycle_stop(wcycle, ewcPME_FFT);
1507 if (pme->nodeid == 0)
1509 real ntot = pme->nkx * pme->nky * pme->nkz;
1510 npme = static_cast<int>(ntot * std::log(ntot) / std::log(2.0));
1511 inc_nrnb(nrnb, eNR_FFT, 2 * npme);
1513 wallcycle_start(wcycle, ewcPME_GATHER);
1516 copy_fftgrid_to_pmegrid(pme, fftgrid, grid, grid_index, pme->nthread, thread);
1518 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1519 } /*#pragma omp parallel*/
1521 /* distribute local grid to all nodes */
1522 if (pme->nnodes > 1)
1524 gmx_sum_qgrid_dd(pme, grid, GMX_SUM_GRID_BACKWARD);
1527 unwrap_periodic_pmegrid(pme, grid);
1529 if (stepWork.computeForces)
1531 /* interpolate forces for our local atoms */
1532 bClearF = (bFirst && PAR(cr));
1533 scale = pme->bFEP ? (fep_state < 1 ? 1.0 - lambda_lj : lambda_lj) : 1.0;
1534 scale *= lb_scale_factor[grid_index - 2];
1536 #pragma omp parallel for num_threads(pme->nthread) schedule(static)
1537 for (thread = 0; thread < pme->nthread; thread++)
1541 gather_f_bsplines(pme, grid, bClearF, &pme->atc[0],
1542 &pme->atc[0].spline[thread], scale);
1544 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1548 inc_nrnb(nrnb, eNR_GATHERFBSP,
1549 pme->pme_order * pme->pme_order * pme->pme_order * pme->atc[0].numAtoms());
1551 wallcycle_stop(wcycle, ewcPME_GATHER);
1554 } /* for (grid_index = 8; grid_index >= 2; --grid_index) */
1555 } /* for (fep_state = 0; fep_state < fep_states_lj; ++fep_state) */
1556 } /* if (pme->doLJ && pme->ljpme_combination_rule == eljpmeLB) */
1558 if (stepWork.computeForces && pme->nnodes > 1)
1560 wallcycle_start(wcycle, ewcPME_REDISTXF);
1561 for (d = 0; d < pme->ndecompdim; d++)
1563 gmx::ArrayRef<gmx::RVec> forcesRef;
1564 if (d == pme->ndecompdim - 1)
1566 const size_t numAtoms = coordinates.size();
1567 GMX_ASSERT(forces.size() >= numAtoms, "Need at least numAtoms forces");
1568 forcesRef = forces.subArray(0, numAtoms);
1572 forcesRef = pme->atc[d + 1].f;
1574 if (DOMAINDECOMP(cr))
1576 dd_pmeredist_f(pme, &pme->atc[d], forcesRef, d == pme->ndecompdim - 1 && pme->bPPnode);
1580 wallcycle_stop(wcycle, ewcPME_REDISTXF);
1583 if (computeEnergyAndVirial)
1589 *energy_q = output[0].coulombEnergy_;
1590 m_add(vir_q, output[0].coulombVirial_, vir_q);
1594 *energy_q = (1.0 - lambda_q) * output[0].coulombEnergy_ + lambda_q * output[1].coulombEnergy_;
1595 *dvdlambda_q += output[1].coulombEnergy_ - output[0].coulombEnergy_;
1596 for (int i = 0; i < DIM; i++)
1598 for (int j = 0; j < DIM; j++)
1600 vir_q[i][j] += (1.0 - lambda_q) * output[0].coulombVirial_[i][j]
1601 + lambda_q * output[1].coulombVirial_[i][j];
1607 fprintf(debug, "Electrostatic PME mesh energy: %g\n", *energy_q);
1619 *energy_lj = output[0].lennardJonesEnergy_;
1620 m_add(vir_lj, output[0].lennardJonesVirial_, vir_lj);
1624 *energy_lj = (1.0 - lambda_lj) * output[0].lennardJonesEnergy_
1625 + lambda_lj * output[1].lennardJonesEnergy_;
1626 *dvdlambda_lj += output[1].lennardJonesEnergy_ - output[0].lennardJonesEnergy_;
1627 for (int i = 0; i < DIM; i++)
1629 for (int j = 0; j < DIM; j++)
1631 vir_lj[i][j] += (1.0 - lambda_lj) * output[0].lennardJonesVirial_[i][j]
1632 + lambda_lj * output[1].lennardJonesVirial_[i][j];
1638 fprintf(debug, "Lennard-Jones PME mesh energy: %g\n", *energy_lj);
1649 void gmx_pme_destroy(gmx_pme_t* pme)
1656 delete pme->boxScaler;
1665 for (int i = 0; i < pme->ngrids; ++i)
1667 pmegrids_destroy(&pme->pmegrid[i]);
1669 if (pme->pfft_setup)
1671 for (int i = 0; i < pme->ngrids; ++i)
1673 gmx_parallel_3dfft_destroy(pme->pfft_setup[i]);
1676 sfree(pme->fftgrid);
1677 sfree(pme->cfftgrid);
1678 sfree(pme->pfft_setup);
1680 for (int i = 0; i < DIM; i++)
1682 sfree(pme->bsp_mod[i]);
1688 if (pme->solve_work)
1690 pme_free_all_work(&pme->solve_work, pme->nthread);
1693 sfree(pme->sum_qgrid_tmp);
1694 sfree(pme->sum_qgrid_dd_tmp);
1696 destroy_pme_spline_work(pme->spline_work);
1698 if (pme->gpu != nullptr)
1700 pme_gpu_destroy(pme->gpu);
1706 void gmx_pme_reinit_atoms(gmx_pme_t* pme, const int numAtoms, const real* charges)
1708 if (pme->gpu != nullptr)
1710 pme_gpu_reinit_atoms(pme->gpu, numAtoms, charges);
1714 pme->atc[0].setNumAtoms(numAtoms);
1715 // TODO: set the charges here as well
1719 bool gmx_pme_grid_matches(const gmx_pme_t& pme, const ivec grid_size)
1721 return (pme.nkx == grid_size[XX] && pme.nky == grid_size[YY] && pme.nkz == grid_size[ZZ]);