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 by the GROMACS development team.
7 * Copyright (c) 2018,2019,2020,2021, 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.
38 /* IMPORTANT FOR DEVELOPERS:
40 * Triclinic pme stuff isn't entirely trivial, and we've experienced
41 * some bugs during development (many of them due to me). To avoid
42 * this in the future, please check the following things if you make
43 * changes in this file:
45 * 1. You should obtain identical (at least to the PME precision)
46 * energies, forces, and virial for
47 * a rectangular box and a triclinic one where the z (or y) axis is
48 * tilted a whole box side. For instance you could use these boxes:
50 * rectangular triclinic
55 * 2. You should check the energy conservation in a triclinic box.
57 * It might seem an overkill, but better safe than sorry.
77 #include "gromacs/domdec/domdec.h"
78 #include "gromacs/ewald/pme.h"
79 #include "gromacs/ewald/pme_coordinate_receiver_gpu.h"
80 #include "gromacs/ewald/pme_force_sender_gpu.h"
81 #include "gromacs/fft/parallel_3dfft.h"
82 #include "gromacs/fileio/pdbio.h"
83 #include "gromacs/gmxlib/network.h"
84 #include "gromacs/gmxlib/nrnb.h"
85 #include "gromacs/gpu_utils/device_stream_manager.h"
86 #include "gromacs/gpu_utils/hostallocator.h"
87 #include "gromacs/math/gmxcomplex.h"
88 #include "gromacs/math/units.h"
89 #include "gromacs/math/vec.h"
90 #include "gromacs/mdtypes/commrec.h"
91 #include "gromacs/mdtypes/forceoutput.h"
92 #include "gromacs/mdtypes/inputrec.h"
93 #include "gromacs/mdtypes/simulation_workload.h"
94 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
95 #include "gromacs/timing/cyclecounter.h"
96 #include "gromacs/timing/wallcycle.h"
97 #include "gromacs/utility/fatalerror.h"
98 #include "gromacs/utility/futil.h"
99 #include "gromacs/utility/gmxmpi.h"
100 #include "gromacs/utility/gmxomp.h"
101 #include "gromacs/utility/smalloc.h"
103 #include "pme_gpu_internal.h"
104 #include "pme_internal.h"
105 #include "pme_output.h"
106 #include "pme_pp_communication.h"
108 /*! \brief Master PP-PME communication data structure */
111 MPI_Comm mpi_comm_mysim; /**< MPI communicator for this simulation */
112 std::vector<PpRanks> ppRanks; /**< The PP partner ranks */
113 int peerRankId; /**< The peer PP rank id */
115 /**< Vectors of A- and B-state parameters used to transfer vectors to PME ranks */
116 gmx::PaddedHostVector<real> chargeA;
117 std::vector<real> chargeB;
118 std::vector<real> sqrt_c6A;
119 std::vector<real> sqrt_c6B;
120 std::vector<real> sigmaA;
121 std::vector<real> sigmaB;
123 gmx::HostVector<gmx::RVec> x; /**< Vector of atom coordinates to transfer to PME ranks */
124 std::vector<gmx::RVec> f; /**< Vector of atom forces received from PME ranks */
126 /**< Vectors of MPI objects used in non-blocking communication between multiple PP ranks per PME rank */
127 std::vector<MPI_Request> req;
128 std::vector<MPI_Status> stat;
131 /*! \brief object for receiving coordinates using communications operating on GPU memory space */
132 std::unique_ptr<gmx::PmeCoordinateReceiverGpu> pmeCoordinateReceiverGpu;
133 /*! \brief object for sending PME force using communications operating on GPU memory space */
134 std::unique_ptr<gmx::PmeForceSenderGpu> pmeForceSenderGpu;
136 /*! \brief whether GPU direct communications are active for PME-PP transfers */
137 bool useGpuDirectComm = false;
138 /*! \brief whether GPU direct communications should send forces directly to remote GPU memory */
139 bool sendForcesDirectToPpGpu = false;
142 /*! \brief Initialize the PME-only side of the PME <-> PP communication */
143 static std::unique_ptr<gmx_pme_pp> gmx_pme_pp_init(const t_commrec* cr)
145 auto pme_pp = std::make_unique<gmx_pme_pp>();
150 pme_pp->mpi_comm_mysim = cr->mpi_comm_mysim;
151 MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
152 auto ppRanks = get_pme_ddranks(cr, rank);
153 pme_pp->ppRanks.reserve(ppRanks.size());
154 for (const auto& ppRankId : ppRanks)
156 pme_pp->ppRanks.push_back({ ppRankId, 0 });
158 // The peer PP rank is the last one.
159 pme_pp->peerRankId = pme_pp->ppRanks.back().rankId;
160 pme_pp->req.resize(eCommType_NR * pme_pp->ppRanks.size());
161 pme_pp->stat.resize(eCommType_NR * pme_pp->ppRanks.size());
163 GMX_UNUSED_VALUE(cr);
169 static void reset_pmeonly_counters(gmx_wallcycle* wcycle,
170 gmx_walltime_accounting_t walltime_accounting,
175 /* Reset all the counters related to performance over the run */
176 wallcycle_stop(wcycle, WallCycleCounter::Run);
177 wallcycle_reset_all(wcycle);
179 wallcycle_start(wcycle, WallCycleCounter::Run);
180 walltime_accounting_reset_time(walltime_accounting, step);
188 static gmx_pme_t* gmx_pmeonly_switch(std::vector<gmx_pme_t*>* pmedata,
189 const ivec grid_size,
193 const t_inputrec* ir)
195 GMX_ASSERT(pmedata, "Bad PME tuning list pointer");
196 for (auto& pme : *pmedata)
198 GMX_ASSERT(pme, "Bad PME tuning list element pointer");
199 if (gmx_pme_grid_matches(*pme, grid_size))
201 /* Here we have found an existing PME data structure that suits us.
202 * However, in the GPU case, we have to reinitialize it - there's only one GPU structure.
203 * This should not cause actual GPU reallocations, at least (the allocated buffers are never shrunk).
204 * So, just some grid size updates in the GPU kernel parameters.
205 * TODO: this should be something like gmx_pme_update_split_params()
207 gmx_pme_reinit(&pme, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
212 const auto& pme = pmedata->back();
213 gmx_pme_t* newStructure = nullptr;
214 // Copy last structure with new grid params
215 gmx_pme_reinit(&newStructure, cr, pme, ir, grid_size, ewaldcoeff_q, ewaldcoeff_lj);
216 pmedata->push_back(newStructure);
220 /*! \brief Called by PME-only ranks to receive coefficients and coordinates
222 * \param[in] pme PME data structure.
223 * \param[in,out] pme_pp PME-PP communication structure.
224 * \param[out] natoms Number of received atoms.
225 * \param[out] box System box, if received.
226 * \param[out] maxshift_x Maximum shift in X direction, if received.
227 * \param[out] maxshift_y Maximum shift in Y direction, if received.
228 * \param[out] lambda_q Free-energy lambda for electrostatics, if received.
229 * \param[out] lambda_lj Free-energy lambda for Lennard-Jones, if received.
230 * \param[out] computeEnergyAndVirial Set to true if this is an energy/virial calculation
231 * step, otherwise set to false.
232 * \param[out] step MD integration step number.
233 * \param[out] grid_size PME grid size, if received.
234 * \param[out] ewaldcoeff_q Ewald cut-off parameter for electrostatics, if received.
235 * \param[out] ewaldcoeff_lj Ewald cut-off parameter for Lennard-Jones, if received.
236 * \param[in] useGpuForPme Flag on whether PME is on GPU.
237 * \param[in] stateGpu GPU state propagator object.
238 * \param[in] runMode PME run mode.
240 * \retval pmerecvqxX All parameters were set, chargeA and chargeB can be NULL.
241 * \retval pmerecvqxFINISH No parameters were set.
242 * \retval pmerecvqxSWITCHGRID Only grid_size and *ewaldcoeff were set.
243 * \retval pmerecvqxRESETCOUNTERS *step was set.
245 static int gmx_pme_recv_coeffs_coords(struct gmx_pme_t* pme,
253 gmx_bool* computeEnergyAndVirial,
259 gmx::StatePropagatorDataGpu* stateGpu,
260 PmeRunMode gmx_unused runMode)
267 bool atomSetChanged = false;
271 gmx_pme_comm_n_box_t cnb;
274 /* Receive the send count, box and time step from the peer PP node */
275 MPI_Recv(&cnb, sizeof(cnb), MPI_BYTE, pme_pp->peerRankId, eCommType_CNB, pme_pp->mpi_comm_mysim, MPI_STATUS_IGNORE);
282 "PME only rank receiving:%s%s%s%s%s\n",
283 (cnb.flags & PP_PME_CHARGE) ? " charges" : "",
284 (cnb.flags & PP_PME_COORD) ? " coordinates" : "",
285 (cnb.flags & PP_PME_FINISH) ? " finish" : "",
286 (cnb.flags & PP_PME_SWITCHGRID) ? " switch grid" : "",
287 (cnb.flags & PP_PME_RESETCOUNTERS) ? " reset counters" : "");
290 pme_pp->useGpuDirectComm = ((cnb.flags & PP_PME_GPUCOMMS) != 0);
291 GMX_ASSERT(!pme_pp->useGpuDirectComm || (pme_pp->pmeForceSenderGpu != nullptr),
292 "The use of GPU direct communication for PME-PP is enabled, "
293 "but the PME GPU force reciever object does not exist");
294 pme_pp->sendForcesDirectToPpGpu = ((cnb.flags & PP_PME_RECVFTOGPU) != 0);
296 if (cnb.flags & PP_PME_FINISH)
298 status = pmerecvqxFINISH;
301 if (cnb.flags & PP_PME_SWITCHGRID)
303 /* Special case, receive the new parameters and return */
304 copy_ivec(cnb.grid_size, *grid_size);
305 *ewaldcoeff_q = cnb.ewaldcoeff_q;
306 *ewaldcoeff_lj = cnb.ewaldcoeff_lj;
308 status = pmerecvqxSWITCHGRID;
311 if (cnb.flags & PP_PME_RESETCOUNTERS)
313 /* Special case, receive the step (set above) and return */
314 status = pmerecvqxRESETCOUNTERS;
317 if (cnb.flags & (PP_PME_CHARGE | PP_PME_SQRTC6 | PP_PME_SIGMA))
319 atomSetChanged = true;
321 /* Receive the send counts from the other PP nodes */
322 for (auto& sender : pme_pp->ppRanks)
324 if (sender.rankId == pme_pp->peerRankId)
326 sender.numAtoms = cnb.natoms;
330 MPI_Irecv(&sender.numAtoms,
331 sizeof(sender.numAtoms),
335 pme_pp->mpi_comm_mysim,
336 &pme_pp->req[messages++]);
339 MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
343 for (const auto& sender : pme_pp->ppRanks)
345 nat += sender.numAtoms;
348 if (cnb.flags & PP_PME_CHARGE)
350 pme_pp->chargeA.resizeWithPadding(nat);
352 if (cnb.flags & PP_PME_CHARGEB)
354 pme_pp->chargeB.resize(nat);
356 if (cnb.flags & PP_PME_SQRTC6)
358 pme_pp->sqrt_c6A.resize(nat);
360 if (cnb.flags & PP_PME_SQRTC6B)
362 pme_pp->sqrt_c6B.resize(nat);
364 if (cnb.flags & PP_PME_SIGMA)
366 pme_pp->sigmaA.resize(nat);
368 if (cnb.flags & PP_PME_SIGMAB)
370 pme_pp->sigmaB.resize(nat);
372 pme_pp->x.resize(nat);
373 pme_pp->f.resize(nat);
375 /* maxshift is sent when the charges are sent */
376 *maxshift_x = cnb.maxshift_x;
377 *maxshift_y = cnb.maxshift_y;
379 /* Receive the charges in place */
380 for (int q = 0; q < eCommType_NR; q++)
384 if (!(cnb.flags & (PP_PME_CHARGE << q)))
390 case eCommType_ChargeA: bufferPtr = pme_pp->chargeA.data(); break;
391 case eCommType_ChargeB: bufferPtr = pme_pp->chargeB.data(); break;
392 case eCommType_SQRTC6A: bufferPtr = pme_pp->sqrt_c6A.data(); break;
393 case eCommType_SQRTC6B: bufferPtr = pme_pp->sqrt_c6B.data(); break;
394 case eCommType_SigmaA: bufferPtr = pme_pp->sigmaA.data(); break;
395 case eCommType_SigmaB: bufferPtr = pme_pp->sigmaB.data(); break;
396 default: gmx_incons("Wrong eCommType");
399 for (const auto& sender : pme_pp->ppRanks)
401 if (sender.numAtoms > 0)
403 MPI_Irecv(bufferPtr + nat,
404 sender.numAtoms * sizeof(real),
408 pme_pp->mpi_comm_mysim,
409 &pme_pp->req[messages++]);
410 nat += sender.numAtoms;
414 "Received from PP rank %d: %d %s\n",
417 (q == eCommType_ChargeA || q == eCommType_ChargeB) ? "charges"
425 if (cnb.flags & PP_PME_COORD)
429 gmx_pme_reinit_atoms(pme, nat, pme_pp->chargeA, pme_pp->chargeB);
432 stateGpu->reinit(nat, nat);
433 pme_gpu_set_device_x(pme, stateGpu->getCoordinates());
435 if (pme_pp->useGpuDirectComm)
437 GMX_ASSERT(runMode == PmeRunMode::GPU,
438 "GPU Direct PME-PP communication has been enabled, "
439 "but PME run mode is not PmeRunMode::GPU\n");
441 // This rank will have its data accessed directly by PP rank, so needs to send the remote addresses.
442 pme_pp->pmeCoordinateReceiverGpu->sendCoordinateBufferAddressToPpRanks(
443 stateGpu->getCoordinates());
444 pme_pp->pmeForceSenderGpu->setForceSendBuffer(pme_gpu_get_device_f(pme));
449 /* The box, FE flag and lambda are sent along with the coordinates
451 copy_mat(cnb.box, box);
452 *lambda_q = cnb.lambda_q;
453 *lambda_lj = cnb.lambda_lj;
454 *computeEnergyAndVirial = ((cnb.flags & PP_PME_ENER_VIR) != 0U);
457 /* Receive the coordinates in place */
459 for (const auto& sender : pme_pp->ppRanks)
461 if (sender.numAtoms > 0)
463 if (pme_pp->useGpuDirectComm)
467 pme_pp->pmeCoordinateReceiverGpu->receiveCoordinatesSynchronizerFromPpCudaDirect(
472 pme_pp->pmeCoordinateReceiverGpu->launchReceiveCoordinatesFromPpCudaMpi(
473 stateGpu->getCoordinates(), nat, sender.numAtoms * sizeof(rvec), sender.rankId);
478 MPI_Irecv(pme_pp->x[nat],
479 sender.numAtoms * sizeof(rvec),
483 pme_pp->mpi_comm_mysim,
484 &pme_pp->req[messages++]);
486 nat += sender.numAtoms;
490 "Received from PP rank %d: %d "
498 if (pme_pp->useGpuDirectComm)
500 pme_pp->pmeCoordinateReceiverGpu->synchronizeOnCoordinatesFromPpRanks();
506 /* Wait for the coordinates and/or charges to arrive */
507 MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
509 } while (status == -1);
511 GMX_UNUSED_VALUE(pme);
512 GMX_UNUSED_VALUE(pme_pp);
513 GMX_UNUSED_VALUE(box);
514 GMX_UNUSED_VALUE(maxshift_x);
515 GMX_UNUSED_VALUE(maxshift_y);
516 GMX_UNUSED_VALUE(lambda_q);
517 GMX_UNUSED_VALUE(lambda_lj);
518 GMX_UNUSED_VALUE(computeEnergyAndVirial);
519 GMX_UNUSED_VALUE(step);
520 GMX_UNUSED_VALUE(grid_size);
521 GMX_UNUSED_VALUE(ewaldcoeff_q);
522 GMX_UNUSED_VALUE(ewaldcoeff_lj);
523 GMX_UNUSED_VALUE(useGpuForPme);
524 GMX_UNUSED_VALUE(stateGpu);
529 if (status == pmerecvqxX)
537 /*! \brief Send the PME mesh force, virial and energy to the PP-only ranks. */
538 static void gmx_pme_send_force_vir_ener(const gmx_pme_t& pme,
540 const PmeOutput& output,
546 gmx_pme_comm_vir_ene_t cve;
547 int messages, ind_start, ind_end;
550 if (pme_pp->useGpuDirectComm)
552 GMX_ASSERT((pme_pp->pmeForceSenderGpu != nullptr),
553 "The use of GPU direct communication for PME-PP is enabled, "
554 "but the PME GPU force reciever object does not exist");
560 /* Now the evaluated forces have to be transferred to the PP ranks */
561 if (pme_pp->useGpuDirectComm && GMX_THREAD_MPI)
563 int numPpRanks = static_cast<int>(pme_pp->ppRanks.size());
564 # pragma omp parallel for num_threads(std::min(numPpRanks, pme.nthread)) schedule(static)
565 for (int i = 0; i < numPpRanks; i++)
567 auto& receiver = pme_pp->ppRanks[i];
568 pme_pp->pmeForceSenderGpu->sendFToPpCudaDirect(
569 receiver.rankId, receiver.numAtoms, pme_pp->sendForcesDirectToPpGpu);
574 for (const auto& receiver : pme_pp->ppRanks)
577 ind_end = ind_start + receiver.numAtoms;
578 if (pme_pp->useGpuDirectComm)
580 pme_pp->pmeForceSenderGpu->sendFToPpCudaMpi(pme_gpu_get_device_f(&pme),
582 receiver.numAtoms * sizeof(rvec),
584 &pme_pp->req[messages]);
588 void* sendbuf = const_cast<void*>(static_cast<const void*>(output.forces_[ind_start]));
591 receiver.numAtoms * sizeof(rvec),
595 pme_pp->mpi_comm_mysim,
596 &pme_pp->req[messages]);
602 /* send virial and energy to our last PP node */
603 copy_mat(output.coulombVirial_, cve.vir_q);
604 copy_mat(output.lennardJonesVirial_, cve.vir_lj);
605 cve.energy_q = output.coulombEnergy_;
606 cve.energy_lj = output.lennardJonesEnergy_;
607 cve.dvdlambda_q = dvdlambda_q;
608 cve.dvdlambda_lj = dvdlambda_lj;
609 /* check for the signals to send back to a PP node */
610 cve.stop_cond = gmx_get_stop_condition();
616 fprintf(debug, "PME rank sending to PP rank %d: virial and energy\n", pme_pp->peerRankId);
618 MPI_Isend(&cve, sizeof(cve), MPI_BYTE, pme_pp->peerRankId, 1, pme_pp->mpi_comm_mysim, &pme_pp->req[messages++]);
620 /* Wait for the forces to arrive */
621 MPI_Waitall(messages, pme_pp->req.data(), pme_pp->stat.data());
623 GMX_RELEASE_ASSERT(false, "Invalid call to gmx_pme_send_force_vir_ener");
624 GMX_UNUSED_VALUE(pme);
625 GMX_UNUSED_VALUE(pme_pp);
626 GMX_UNUSED_VALUE(output);
627 GMX_UNUSED_VALUE(dvdlambda_q);
628 GMX_UNUSED_VALUE(dvdlambda_lj);
629 GMX_UNUSED_VALUE(cycles);
633 int gmx_pmeonly(struct gmx_pme_t* pme,
636 gmx_wallcycle* wcycle,
637 gmx_walltime_accounting_t walltime_accounting,
640 bool useGpuPmePpCommunication,
641 const gmx::DeviceStreamManager* deviceStreamManager)
648 int maxshift_x = 0, maxshift_y = 0;
649 real dvdlambda_q, dvdlambda_lj;
652 bool computeEnergyAndVirial = false;
655 /* This data will only use with PME tuning, i.e. switching PME grids */
656 std::vector<gmx_pme_t*> pmedata;
657 pmedata.push_back(pme);
659 auto pme_pp = gmx_pme_pp_init(cr);
661 std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
662 // TODO the variable below should be queried from the task assignment info
663 const bool useGpuForPme = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
667 deviceStreamManager != nullptr,
668 "Device stream manager can not be nullptr when using GPU in PME-only rank.");
669 GMX_RELEASE_ASSERT(deviceStreamManager->streamIsValid(gmx::DeviceStreamType::Pme),
670 "Device stream can not be nullptr when using GPU in PME-only rank");
671 changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
672 changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
673 if (useGpuPmePpCommunication)
675 pme_pp->pmeCoordinateReceiverGpu = std::make_unique<gmx::PmeCoordinateReceiverGpu>(
676 deviceStreamManager->stream(gmx::DeviceStreamType::Pme),
677 pme_pp->mpi_comm_mysim,
679 pme_pp->pmeForceSenderGpu =
680 std::make_unique<gmx::PmeForceSenderGpu>(pme_gpu_get_f_ready_synchronizer(pme),
681 pme_pp->mpi_comm_mysim,
682 deviceStreamManager->context(),
685 // TODO: Special PME-only constructor is used here. There is no mechanism to prevent from using the other constructor here.
686 // This should be made safer.
687 stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
688 &deviceStreamManager->stream(gmx::DeviceStreamType::Pme),
689 deviceStreamManager->context(),
690 GpuApiCallBehavior::Async,
691 pme_gpu_get_block_size(pme),
698 do /****** this is a quasi-loop over time steps! */
700 /* The reason for having a loop here is PME grid tuning/switching */
703 /* Domain decomposition */
705 real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
706 ret = gmx_pme_recv_coeffs_coords(pme,
714 &computeEnergyAndVirial,
723 if (ret == pmerecvqxSWITCHGRID)
725 /* Switch the PME grid to newGridSize */
726 pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
729 if (ret == pmerecvqxRESETCOUNTERS)
731 /* Reset the cycle and flop counters */
732 reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
734 } while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
736 if (ret == pmerecvqxFINISH)
738 /* We should stop: break out of the loop */
744 wallcycle_start(wcycle, WallCycleCounter::Run);
745 walltime_accounting_start_time(walltime_accounting);
748 wallcycle_start(wcycle, WallCycleCounter::PmeMesh);
753 // TODO Make a struct of array refs onto these per-atom fields
754 // of pme_pp (maybe box, energy and virial, too; and likewise
755 // from mdatoms for the other call to gmx_pme_do), so we have
756 // fewer lines of code and less parameter passing.
757 gmx::StepWorkload stepWork;
758 stepWork.computeVirial = computeEnergyAndVirial;
759 stepWork.computeEnergy = computeEnergyAndVirial;
760 stepWork.computeForces = true;
761 PmeOutput output = { {}, false, 0, { { 0 } }, 0, 0, { { 0 } }, 0 };
764 stepWork.haveDynamicBox = false;
765 stepWork.useGpuPmeFReduction = pme_pp->useGpuDirectComm;
766 // TODO this should be set properly by gmx_pme_recv_coeffs_coords,
767 // or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
768 pme_gpu_prepare_computation(pme, box, wcycle, stepWork);
769 if (!pme_pp->useGpuDirectComm)
771 stateGpu->copyCoordinatesToGpu(gmx::ArrayRef<gmx::RVec>(pme_pp->x),
772 gmx::AtomLocality::Local);
774 // On the separate PME rank we do not need a synchronizer as we schedule everything in a single stream
775 // TODO: with pme on GPU the receive should make a list of synchronizers and pass it here #3157
776 auto xReadyOnDevice = nullptr;
778 pme_gpu_launch_spread(pme, xReadyOnDevice, wcycle, lambda_q);
779 pme_gpu_launch_complex_transforms(pme, wcycle, stepWork);
780 pme_gpu_launch_gather(pme, wcycle, lambda_q);
781 output = pme_gpu_wait_finish_task(pme, computeEnergyAndVirial, lambda_q, wcycle);
782 pme_gpu_reinit_computation(pme, wcycle);
786 GMX_ASSERT(pme_pp->x.size() == static_cast<size_t>(natoms),
787 "The coordinate buffer should have size natoms");
804 output.coulombVirial_,
805 output.lennardJonesVirial_,
806 &output.coulombEnergy_,
807 &output.lennardJonesEnergy_,
813 output.forces_ = pme_pp->f;
816 cycles = wallcycle_stop(wcycle, WallCycleCounter::PmeMesh);
817 gmx_pme_send_force_vir_ener(*pme, pme_pp.get(), output, dvdlambda_q, dvdlambda_lj, cycles);
820 } /***** end of quasi-loop, we stop with the break above */
823 walltime_accounting_end_time(walltime_accounting);