2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2005,2006,2007,2008,2009 by the GROMACS development team.
5 * Copyright (c) 2010,2011,2012,2013,2014 by the GROMACS development team.
6 * Copyright (c) 2015,2016,2017,2018,2019 by the GROMACS development team.
7 * Copyright (c) 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.
56 #include "gromacs/domdec/builder.h"
57 #include "gromacs/domdec/collect.h"
58 #include "gromacs/domdec/dlb.h"
59 #include "gromacs/domdec/dlbtiming.h"
60 #include "gromacs/domdec/domdec_network.h"
61 #include "gromacs/domdec/ga2la.h"
62 #include "gromacs/domdec/gpuhaloexchange.h"
63 #include "gromacs/domdec/options.h"
64 #include "gromacs/domdec/partition.h"
65 #include "gromacs/gmxlib/network.h"
66 #include "gromacs/gmxlib/nrnb.h"
67 #include "gromacs/gpu_utils/gpu_utils.h"
68 #include "gromacs/hardware/hw_info.h"
69 #include "gromacs/listed_forces/manage_threading.h"
70 #include "gromacs/math/vec.h"
71 #include "gromacs/math/vectypes.h"
72 #include "gromacs/mdlib/calc_verletbuf.h"
73 #include "gromacs/mdlib/constr.h"
74 #include "gromacs/mdlib/constraintrange.h"
75 #include "gromacs/mdlib/updategroups.h"
76 #include "gromacs/mdlib/vsite.h"
77 #include "gromacs/mdtypes/commrec.h"
78 #include "gromacs/mdtypes/forceoutput.h"
79 #include "gromacs/mdtypes/inputrec.h"
80 #include "gromacs/mdtypes/mdrunoptions.h"
81 #include "gromacs/mdtypes/state.h"
82 #include "gromacs/pbcutil/ishift.h"
83 #include "gromacs/pbcutil/pbc.h"
84 #include "gromacs/pulling/pull.h"
85 #include "gromacs/timing/wallcycle.h"
86 #include "gromacs/topology/block.h"
87 #include "gromacs/topology/idef.h"
88 #include "gromacs/topology/ifunc.h"
89 #include "gromacs/topology/mtop_lookup.h"
90 #include "gromacs/topology/mtop_util.h"
91 #include "gromacs/topology/topology.h"
92 #include "gromacs/utility/basedefinitions.h"
93 #include "gromacs/utility/basenetwork.h"
94 #include "gromacs/utility/cstringutil.h"
95 #include "gromacs/utility/exceptions.h"
96 #include "gromacs/utility/fatalerror.h"
97 #include "gromacs/utility/gmxmpi.h"
98 #include "gromacs/utility/logger.h"
99 #include "gromacs/utility/real.h"
100 #include "gromacs/utility/smalloc.h"
101 #include "gromacs/utility/strconvert.h"
102 #include "gromacs/utility/stringstream.h"
103 #include "gromacs/utility/stringutil.h"
104 #include "gromacs/utility/textwriter.h"
106 #include "atomdistribution.h"
108 #include "cellsizes.h"
109 #include "distribute.h"
110 #include "domdec_constraints.h"
111 #include "domdec_internal.h"
112 #include "domdec_setup.h"
113 #include "domdec_vsite.h"
114 #include "redistribute.h"
117 // TODO remove this when moving domdec into gmx namespace
118 using gmx::DdRankOrder;
119 using gmx::DlbOption;
120 using gmx::DomdecOptions;
122 static const char* edlbs_names[int(DlbState::nr)] = { "off", "auto", "locked", "on", "on" };
124 /* The size per atom group of the cggl_flag buffer in gmx_domdec_comm_t */
127 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
128 #define DD_FLAG_NRCG 65535
129 #define DD_FLAG_FW(d) (1 << (16 + (d)*2))
130 #define DD_FLAG_BW(d) (1 << (16 + (d)*2 + 1))
132 /* The DD zone order */
133 static const ivec dd_zo[DD_MAXZONE] = { { 0, 0, 0 }, { 1, 0, 0 }, { 1, 1, 0 }, { 0, 1, 0 },
134 { 0, 1, 1 }, { 0, 0, 1 }, { 1, 0, 1 }, { 1, 1, 1 } };
136 /* The non-bonded zone-pair setup for domain decomposition
137 * The first number is the i-zone, the second number the first j-zone seen by
138 * this i-zone, the third number the last+1 j-zone seen by this i-zone.
139 * As is, this is for 3D decomposition, where there are 4 i-zones.
140 * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
141 * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
143 static const int ddNonbondedZonePairRanges[DD_MAXIZONE][3] = { { 0, 0, 8 },
148 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
150 xyz[XX] = ind / (nc[YY] * nc[ZZ]);
151 xyz[YY] = (ind / nc[ZZ]) % nc[YY];
152 xyz[ZZ] = ind % nc[ZZ];
155 static int ddcoord2ddnodeid(gmx_domdec_t* dd, ivec c)
159 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
160 const int ddindex = dd_index(dd->numCells, c);
161 if (cartSetup.bCartesianPP_PME)
163 ddnodeid = cartSetup.ddindex2ddnodeid[ddindex];
165 else if (cartSetup.bCartesianPP)
168 MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
179 int ddglatnr(const gmx_domdec_t* dd, int i)
189 if (i >= dd->comm->atomRanges.numAtomsTotal())
192 "glatnr called with %d, which is larger than the local number of atoms (%d)",
193 i, dd->comm->atomRanges.numAtomsTotal());
195 atnr = dd->globalAtomIndices[i] + 1;
201 gmx::ArrayRef<const gmx::RangePartitioning> getUpdateGroupingPerMoleculetype(const gmx_domdec_t& dd)
203 GMX_RELEASE_ASSERT(dd.comm, "Need a valid dd.comm");
204 return dd.comm->systemInfo.updateGroupingPerMoleculetype;
207 void dd_store_state(gmx_domdec_t* dd, t_state* state)
211 if (state->ddp_count != dd->ddp_count)
213 gmx_incons("The MD state does not match the domain decomposition state");
216 state->cg_gl.resize(dd->ncg_home);
217 for (i = 0; i < dd->ncg_home; i++)
219 state->cg_gl[i] = dd->globalAtomGroupIndices[i];
222 state->ddp_count_cg_gl = dd->ddp_count;
225 gmx_domdec_zones_t* domdec_zones(gmx_domdec_t* dd)
227 return &dd->comm->zones;
230 int dd_numAtomsZones(const gmx_domdec_t& dd)
232 return dd.comm->atomRanges.end(DDAtomRanges::Type::Zones);
235 int dd_numHomeAtoms(const gmx_domdec_t& dd)
237 return dd.comm->atomRanges.numHomeAtoms();
240 int dd_natoms_mdatoms(const gmx_domdec_t* dd)
242 /* We currently set mdatoms entries for all atoms:
243 * local + non-local + communicated for vsite + constraints
246 return dd->comm->atomRanges.numAtomsTotal();
249 int dd_natoms_vsite(const gmx_domdec_t* dd)
251 return dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
254 void dd_get_constraint_range(const gmx_domdec_t* dd, int* at_start, int* at_end)
256 *at_start = dd->comm->atomRanges.start(DDAtomRanges::Type::Constraints);
257 *at_end = dd->comm->atomRanges.end(DDAtomRanges::Type::Constraints);
260 void dd_move_x(gmx_domdec_t* dd, const matrix box, gmx::ArrayRef<gmx::RVec> x, gmx_wallcycle* wcycle)
262 wallcycle_start(wcycle, ewcMOVEX);
265 gmx_domdec_comm_t* comm;
266 gmx_domdec_comm_dim_t* cd;
267 rvec shift = { 0, 0, 0 };
268 gmx_bool bPBC, bScrew;
273 nat_tot = comm->atomRanges.numHomeAtoms();
274 for (int d = 0; d < dd->ndim; d++)
276 bPBC = (dd->ci[dd->dim[d]] == 0);
277 bScrew = (bPBC && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
280 copy_rvec(box[dd->dim[d]], shift);
283 for (const gmx_domdec_ind_t& ind : cd->ind)
285 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
286 gmx::ArrayRef<gmx::RVec>& sendBuffer = sendBufferAccess.buffer;
290 for (int j : ind.index)
292 sendBuffer[n] = x[j];
298 for (int j : ind.index)
300 /* We need to shift the coordinates */
301 for (int d = 0; d < DIM; d++)
303 sendBuffer[n][d] = x[j][d] + shift[d];
310 for (int j : ind.index)
313 sendBuffer[n][XX] = x[j][XX] + shift[XX];
315 * This operation requires a special shift force
316 * treatment, which is performed in calc_vir.
318 sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
319 sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
324 DDBufferAccess<gmx::RVec> receiveBufferAccess(
325 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
327 gmx::ArrayRef<gmx::RVec> receiveBuffer;
328 if (cd->receiveInPlace)
330 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
334 receiveBuffer = receiveBufferAccess.buffer;
336 /* Send and receive the coordinates */
337 ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
339 if (!cd->receiveInPlace)
342 for (int zone = 0; zone < nzone; zone++)
344 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
346 x[i] = receiveBuffer[j++];
350 nat_tot += ind.nrecv[nzone + 1];
355 wallcycle_stop(wcycle, ewcMOVEX);
358 void dd_move_f(gmx_domdec_t* dd, gmx::ForceWithShiftForces* forceWithShiftForces, gmx_wallcycle* wcycle)
360 wallcycle_start(wcycle, ewcMOVEF);
362 gmx::ArrayRef<gmx::RVec> f = forceWithShiftForces->force();
363 gmx::ArrayRef<gmx::RVec> fshift = forceWithShiftForces->shiftForces();
365 gmx_domdec_comm_t& comm = *dd->comm;
366 int nzone = comm.zones.n / 2;
367 int nat_tot = comm.atomRanges.end(DDAtomRanges::Type::Zones);
368 for (int d = dd->ndim - 1; d >= 0; d--)
370 /* Only forces in domains near the PBC boundaries need to
371 consider PBC in the treatment of fshift */
372 const bool shiftForcesNeedPbc =
373 (forceWithShiftForces->computeVirial() && dd->ci[dd->dim[d]] == 0);
374 const bool applyScrewPbc =
375 (shiftForcesNeedPbc && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
376 /* Determine which shift vector we need */
377 ivec vis = { 0, 0, 0 };
379 const int is = IVEC2IS(vis);
381 /* Loop over the pulses */
382 const gmx_domdec_comm_dim_t& cd = comm.cd[d];
383 for (int p = cd.numPulses() - 1; p >= 0; p--)
385 const gmx_domdec_ind_t& ind = cd.ind[p];
386 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm.rvecBuffer, ind.nsend[nzone + 1]);
387 gmx::ArrayRef<gmx::RVec>& receiveBuffer = receiveBufferAccess.buffer;
389 nat_tot -= ind.nrecv[nzone + 1];
391 DDBufferAccess<gmx::RVec> sendBufferAccess(
392 comm.rvecBuffer2, cd.receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
394 gmx::ArrayRef<gmx::RVec> sendBuffer;
395 if (cd.receiveInPlace)
397 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
401 sendBuffer = sendBufferAccess.buffer;
403 for (int zone = 0; zone < nzone; zone++)
405 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
407 sendBuffer[j++] = f[i];
411 /* Communicate the forces */
412 ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
413 /* Add the received forces */
415 if (!shiftForcesNeedPbc)
417 for (int j : ind.index)
419 for (int d = 0; d < DIM; d++)
421 f[j][d] += receiveBuffer[n][d];
426 else if (!applyScrewPbc)
428 for (int j : ind.index)
430 for (int d = 0; d < DIM; d++)
432 f[j][d] += receiveBuffer[n][d];
434 /* Add this force to the shift force */
435 for (int d = 0; d < DIM; d++)
437 fshift[is][d] += receiveBuffer[n][d];
444 for (int j : ind.index)
446 /* Rotate the force */
447 f[j][XX] += receiveBuffer[n][XX];
448 f[j][YY] -= receiveBuffer[n][YY];
449 f[j][ZZ] -= receiveBuffer[n][ZZ];
450 if (shiftForcesNeedPbc)
452 /* Add this force to the shift force */
453 for (int d = 0; d < DIM; d++)
455 fshift[is][d] += receiveBuffer[n][d];
464 wallcycle_stop(wcycle, ewcMOVEF);
467 /* Convenience function for extracting a real buffer from an rvec buffer
469 * To reduce the number of temporary communication buffers and avoid
470 * cache polution, we reuse gmx::RVec buffers for storing reals.
471 * This functions return a real buffer reference with the same number
472 * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
474 static gmx::ArrayRef<real> realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
476 return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0], arrayRef.size());
479 void dd_atom_spread_real(gmx_domdec_t* dd, real v[])
482 gmx_domdec_comm_t* comm;
483 gmx_domdec_comm_dim_t* cd;
488 nat_tot = comm->atomRanges.numHomeAtoms();
489 for (int d = 0; d < dd->ndim; d++)
492 for (const gmx_domdec_ind_t& ind : cd->ind)
494 /* Note: We provision for RVec instead of real, so a factor of 3
495 * more than needed. The buffer actually already has this size
496 * and we pass a plain pointer below, so this does not matter.
498 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
499 gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
501 for (int j : ind.index)
503 sendBuffer[n++] = v[j];
506 DDBufferAccess<gmx::RVec> receiveBufferAccess(
507 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
509 gmx::ArrayRef<real> receiveBuffer;
510 if (cd->receiveInPlace)
512 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
516 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
518 /* Send and receive the data */
519 ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
520 if (!cd->receiveInPlace)
523 for (int zone = 0; zone < nzone; zone++)
525 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
527 v[i] = receiveBuffer[j++];
531 nat_tot += ind.nrecv[nzone + 1];
537 void dd_atom_sum_real(gmx_domdec_t* dd, real v[])
540 gmx_domdec_comm_t* comm;
541 gmx_domdec_comm_dim_t* cd;
545 nzone = comm->zones.n / 2;
546 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
547 for (int d = dd->ndim - 1; d >= 0; d--)
550 for (int p = cd->numPulses() - 1; p >= 0; p--)
552 const gmx_domdec_ind_t& ind = cd->ind[p];
554 /* Note: We provision for RVec instead of real, so a factor of 3
555 * more than needed. The buffer actually already has this size
556 * and we typecast, so this works as intended.
558 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
559 gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
560 nat_tot -= ind.nrecv[nzone + 1];
562 DDBufferAccess<gmx::RVec> sendBufferAccess(
563 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
565 gmx::ArrayRef<real> sendBuffer;
566 if (cd->receiveInPlace)
568 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
572 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
574 for (int zone = 0; zone < nzone; zone++)
576 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
578 sendBuffer[j++] = v[i];
582 /* Communicate the forces */
583 ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
584 /* Add the received forces */
586 for (int j : ind.index)
588 v[j] += receiveBuffer[n];
596 real dd_cutoff_multibody(const gmx_domdec_t* dd)
598 const gmx_domdec_comm_t& comm = *dd->comm;
599 const DDSystemInfo& systemInfo = comm.systemInfo;
602 if (systemInfo.haveInterDomainMultiBodyBondeds)
604 if (comm.cutoff_mbody > 0)
606 r = comm.cutoff_mbody;
610 /* cutoff_mbody=0 means we do not have DLB */
611 r = comm.cellsize_min[dd->dim[0]];
612 for (int di = 1; di < dd->ndim; di++)
614 r = std::min(r, comm.cellsize_min[dd->dim[di]]);
616 if (comm.systemInfo.filterBondedCommunication)
618 r = std::max(r, comm.cutoff_mbody);
622 r = std::min(r, systemInfo.cutoff);
630 real dd_cutoff_twobody(const gmx_domdec_t* dd)
634 r_mb = dd_cutoff_multibody(dd);
636 return std::max(dd->comm->systemInfo.cutoff, r_mb);
640 static void dd_cart_coord2pmecoord(const DDRankSetup& ddRankSetup,
641 const CartesianRankSetup& cartSetup,
645 const int nc = ddRankSetup.numPPCells[cartSetup.cartpmedim];
646 const int ntot = cartSetup.ntot[cartSetup.cartpmedim];
647 copy_ivec(coord, coord_pme);
648 coord_pme[cartSetup.cartpmedim] =
649 nc + (coord[cartSetup.cartpmedim] * (ntot - nc) + (ntot - nc) / 2) / nc;
652 /* Returns the PME rank index in 0...npmenodes-1 for the PP cell with index ddCellIndex */
653 static int ddindex2pmeindex(const DDRankSetup& ddRankSetup, const int ddCellIndex)
655 const int npp = ddRankSetup.numPPRanks;
656 const int npme = ddRankSetup.numRanksDoingPme;
658 /* Here we assign a PME node to communicate with this DD node
659 * by assuming that the major index of both is x.
660 * We add npme/2 to obtain an even distribution.
662 return (ddCellIndex * npme + npme / 2) / npp;
665 static std::vector<int> dd_interleaved_pme_ranks(const DDRankSetup& ddRankSetup)
667 std::vector<int> pmeRanks(ddRankSetup.numRanksDoingPme);
670 for (int i = 0; i < ddRankSetup.numPPRanks; i++)
672 const int p0 = ddindex2pmeindex(ddRankSetup, i);
673 const int p1 = ddindex2pmeindex(ddRankSetup, i + 1);
674 if (i + 1 == ddRankSetup.numPPRanks || p1 > p0)
678 fprintf(debug, "pme_rank[%d] = %d\n", n, i + 1 + n);
680 pmeRanks[n] = i + 1 + n;
688 static int gmx_ddcoord2pmeindex(const t_commrec* cr, int x, int y, int z)
698 slab = ddindex2pmeindex(dd->comm->ddRankSetup, dd_index(dd->numCells, coords));
703 static int ddcoord2simnodeid(const t_commrec* cr, int x, int y, int z)
705 const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
706 ivec coords = { x, y, z };
709 if (cartSetup.bCartesianPP_PME)
712 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
717 int ddindex = dd_index(cr->dd->numCells, coords);
718 if (cartSetup.bCartesianPP)
720 nodeid = cartSetup.ddindex2simnodeid[ddindex];
724 if (cr->dd->comm->ddRankSetup.usePmeOnlyRanks)
726 nodeid = ddindex + gmx_ddcoord2pmeindex(cr, x, y, z);
738 static int dd_simnode2pmenode(const DDRankSetup& ddRankSetup,
739 const CartesianRankSetup& cartSetup,
740 gmx::ArrayRef<const int> pmeRanks,
741 const t_commrec gmx_unused* cr,
742 const int sim_nodeid)
746 /* This assumes a uniform x domain decomposition grid cell size */
747 if (cartSetup.bCartesianPP_PME)
750 ivec coord, coord_pme;
751 MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
752 if (coord[cartSetup.cartpmedim] < ddRankSetup.numPPCells[cartSetup.cartpmedim])
754 /* This is a PP rank */
755 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
756 MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
760 else if (cartSetup.bCartesianPP)
762 if (sim_nodeid < ddRankSetup.numPPRanks)
764 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
769 /* This assumes DD cells with identical x coordinates
770 * are numbered sequentially.
772 if (pmeRanks.empty())
774 if (sim_nodeid < ddRankSetup.numPPRanks)
776 /* The DD index equals the nodeid */
777 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
783 while (sim_nodeid > pmeRanks[i])
787 if (sim_nodeid < pmeRanks[i])
789 pmenode = pmeRanks[i];
797 NumPmeDomains getNumPmeDomains(const gmx_domdec_t* dd)
801 return { dd->comm->ddRankSetup.npmenodes_x, dd->comm->ddRankSetup.npmenodes_y };
809 std::vector<int> get_pme_ddranks(const t_commrec* cr, const int pmenodeid)
811 const DDRankSetup& ddRankSetup = cr->dd->comm->ddRankSetup;
812 const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
813 GMX_RELEASE_ASSERT(ddRankSetup.usePmeOnlyRanks,
814 "This function should only be called when PME-only ranks are in use");
815 std::vector<int> ddranks;
816 ddranks.reserve((ddRankSetup.numPPRanks + ddRankSetup.numRanksDoingPme - 1) / ddRankSetup.numRanksDoingPme);
818 for (int x = 0; x < ddRankSetup.numPPCells[XX]; x++)
820 for (int y = 0; y < ddRankSetup.numPPCells[YY]; y++)
822 for (int z = 0; z < ddRankSetup.numPPCells[ZZ]; z++)
824 if (cartSetup.bCartesianPP_PME)
826 ivec coord = { x, y, z };
828 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
829 if (cr->dd->ci[XX] == coord_pme[XX] && cr->dd->ci[YY] == coord_pme[YY]
830 && cr->dd->ci[ZZ] == coord_pme[ZZ])
832 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
837 /* The slab corresponds to the nodeid in the PME group */
838 if (gmx_ddcoord2pmeindex(cr, x, y, z) == pmenodeid)
840 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
849 static gmx_bool receive_vir_ener(const gmx_domdec_t* dd, gmx::ArrayRef<const int> pmeRanks, const t_commrec* cr)
851 gmx_bool bReceive = TRUE;
853 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
854 if (ddRankSetup.usePmeOnlyRanks)
856 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
857 if (cartSetup.bCartesianPP_PME)
860 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
862 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
863 coords[cartSetup.cartpmedim]++;
864 if (coords[cartSetup.cartpmedim] < dd->numCells[cartSetup.cartpmedim])
867 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
868 if (dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, rank) == pmenode)
870 /* This is not the last PP node for pmenode */
877 "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
882 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
883 if (cr->sim_nodeid + 1 < cr->nnodes
884 && dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid + 1) == pmenode)
886 /* This is not the last PP node for pmenode */
895 static void set_slb_pme_dim_f(gmx_domdec_t* dd, int dim, real** dim_f)
897 gmx_domdec_comm_t* comm;
902 snew(*dim_f, dd->numCells[dim] + 1);
904 for (i = 1; i < dd->numCells[dim]; i++)
906 if (comm->slb_frac[dim])
908 (*dim_f)[i] = (*dim_f)[i - 1] + comm->slb_frac[dim][i - 1];
912 (*dim_f)[i] = static_cast<real>(i) / static_cast<real>(dd->numCells[dim]);
915 (*dim_f)[dd->numCells[dim]] = 1;
918 static void init_ddpme(gmx_domdec_t* dd, gmx_ddpme_t* ddpme, int dimind)
920 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
922 if (dimind == 0 && dd->dim[0] == YY && ddRankSetup.npmenodes_x == 1)
930 ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
932 ddpme->nslab = (ddpme->dim == 0 ? ddRankSetup.npmenodes_x : ddRankSetup.npmenodes_y);
934 if (ddpme->nslab <= 1)
939 const int nso = ddRankSetup.numRanksDoingPme / ddpme->nslab;
940 /* Determine for each PME slab the PP location range for dimension dim */
941 snew(ddpme->pp_min, ddpme->nslab);
942 snew(ddpme->pp_max, ddpme->nslab);
943 for (int slab = 0; slab < ddpme->nslab; slab++)
945 ddpme->pp_min[slab] = dd->numCells[dd->dim[dimind]] - 1;
946 ddpme->pp_max[slab] = 0;
948 for (int i = 0; i < dd->nnodes; i++)
951 ddindex2xyz(dd->numCells, i, xyz);
952 /* For y only use our y/z slab.
953 * This assumes that the PME x grid size matches the DD grid size.
955 if (dimind == 0 || xyz[XX] == dd->ci[XX])
957 const int pmeindex = ddindex2pmeindex(ddRankSetup, i);
961 slab = pmeindex / nso;
965 slab = pmeindex % ddpme->nslab;
967 ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
968 ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
972 set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
975 int dd_pme_maxshift_x(const gmx_domdec_t* dd)
977 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
979 if (ddRankSetup.ddpme[0].dim == XX)
981 return ddRankSetup.ddpme[0].maxshift;
989 int dd_pme_maxshift_y(const gmx_domdec_t* dd)
991 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
993 if (ddRankSetup.ddpme[0].dim == YY)
995 return ddRankSetup.ddpme[0].maxshift;
997 else if (ddRankSetup.npmedecompdim >= 2 && ddRankSetup.ddpme[1].dim == YY)
999 return ddRankSetup.ddpme[1].maxshift;
1007 bool ddHaveSplitConstraints(const gmx_domdec_t& dd)
1009 return dd.comm->systemInfo.haveSplitConstraints;
1012 bool ddUsesUpdateGroups(const gmx_domdec_t& dd)
1014 return dd.comm->systemInfo.useUpdateGroups;
1017 void dd_cycles_add(const gmx_domdec_t* dd, float cycles, int ddCycl)
1019 /* Note that the cycles value can be incorrect, either 0 or some
1020 * extremely large value, when our thread migrated to another core
1021 * with an unsynchronized cycle counter. If this happens less often
1022 * that once per nstlist steps, this will not cause issues, since
1023 * we later subtract the maximum value from the sum over nstlist steps.
1024 * A zero count will slightly lower the total, but that's a small effect.
1025 * Note that the main purpose of the subtraction of the maximum value
1026 * is to avoid throwing off the load balancing when stalls occur due
1027 * e.g. system activity or network congestion.
1029 dd->comm->cycl[ddCycl] += cycles;
1030 dd->comm->cycl_n[ddCycl]++;
1031 if (cycles > dd->comm->cycl_max[ddCycl])
1033 dd->comm->cycl_max[ddCycl] = cycles;
1038 static void make_load_communicator(gmx_domdec_t* dd, int dim_ind, ivec loc)
1043 gmx_bool bPartOfGroup = FALSE;
1045 dim = dd->dim[dim_ind];
1046 copy_ivec(loc, loc_c);
1047 for (i = 0; i < dd->numCells[dim]; i++)
1050 rank = dd_index(dd->numCells, loc_c);
1051 if (rank == dd->rank)
1053 /* This process is part of the group */
1054 bPartOfGroup = TRUE;
1057 MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank, &c_row);
1060 dd->comm->mpi_comm_load[dim_ind] = c_row;
1061 if (!isDlbDisabled(dd->comm))
1063 DDCellsizesWithDlb& cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
1065 if (dd->ci[dim] == dd->master_ci[dim])
1067 /* This is the root process of this row */
1068 cellsizes.rowMaster = std::make_unique<RowMaster>();
1070 RowMaster& rowMaster = *cellsizes.rowMaster;
1071 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
1072 rowMaster.oldCellFrac.resize(dd->numCells[dim] + 1);
1073 rowMaster.isCellMin.resize(dd->numCells[dim]);
1076 rowMaster.bounds.resize(dd->numCells[dim]);
1078 rowMaster.buf_ncd.resize(dd->numCells[dim]);
1082 /* This is not a root process, we only need to receive cell_f */
1083 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
1086 if (dd->ci[dim] == dd->master_ci[dim])
1088 snew(dd->comm->load[dim_ind].load, dd->numCells[dim] * DD_NLOAD_MAX);
1094 void dd_setup_dlb_resource_sharing(const t_commrec* cr, int gpu_id)
1097 int physicalnode_id_hash;
1099 MPI_Comm mpi_comm_pp_physicalnode;
1101 if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
1103 /* Only ranks with short-ranged tasks (currently) use GPUs.
1104 * If we don't have GPUs assigned, there are no resources to share.
1109 physicalnode_id_hash = gmx_physicalnode_id_hash();
1115 fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
1116 fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n", dd->rank,
1117 physicalnode_id_hash, gpu_id);
1119 /* Split the PP communicator over the physical nodes */
1120 /* TODO: See if we should store this (before), as it's also used for
1121 * for the nodecomm summation.
1123 // TODO PhysicalNodeCommunicator could be extended/used to handle
1124 // the need for per-node per-group communicators.
1125 MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank, &mpi_comm_pp_physicalnode);
1126 MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank, &dd->comm->mpi_comm_gpu_shared);
1127 MPI_Comm_free(&mpi_comm_pp_physicalnode);
1128 MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
1132 fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
1135 /* Note that some ranks could share a GPU, while others don't */
1137 if (dd->comm->nrank_gpu_shared == 1)
1139 MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
1142 GMX_UNUSED_VALUE(cr);
1143 GMX_UNUSED_VALUE(gpu_id);
1147 static void make_load_communicators(gmx_domdec_t gmx_unused* dd)
1150 int dim0, dim1, i, j;
1155 fprintf(debug, "Making load communicators\n");
1158 dd->comm->load = new domdec_load_t[std::max(dd->ndim, 1)];
1159 snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
1167 make_load_communicator(dd, 0, loc);
1171 for (i = 0; i < dd->numCells[dim0]; i++)
1174 make_load_communicator(dd, 1, loc);
1180 for (i = 0; i < dd->numCells[dim0]; i++)
1184 for (j = 0; j < dd->numCells[dim1]; j++)
1187 make_load_communicator(dd, 2, loc);
1194 fprintf(debug, "Finished making load communicators\n");
1199 /*! \brief Sets up the relation between neighboring domains and zones */
1200 static void setup_neighbor_relations(gmx_domdec_t* dd)
1204 gmx_domdec_zones_t* zones;
1205 GMX_ASSERT((dd->ndim >= 0) && (dd->ndim <= DIM), "Must have valid number of dimensions for DD");
1207 for (d = 0; d < dd->ndim; d++)
1210 copy_ivec(dd->ci, tmp);
1211 tmp[dim] = (tmp[dim] + 1) % dd->numCells[dim];
1212 dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
1213 copy_ivec(dd->ci, tmp);
1214 tmp[dim] = (tmp[dim] - 1 + dd->numCells[dim]) % dd->numCells[dim];
1215 dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
1218 fprintf(debug, "DD rank %d neighbor ranks in dir %d are + %d - %d\n", dd->rank, dim,
1219 dd->neighbor[d][0], dd->neighbor[d][1]);
1223 int nzone = (1 << dd->ndim);
1224 int nizone = (1 << std::max(dd->ndim - 1, 0));
1225 assert(nizone >= 1 && nizone <= DD_MAXIZONE);
1227 zones = &dd->comm->zones;
1229 for (int i = 0; i < nzone; i++)
1232 clear_ivec(zones->shift[i]);
1233 for (d = 0; d < dd->ndim; d++)
1235 zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
1240 for (int i = 0; i < nzone; i++)
1242 for (d = 0; d < DIM; d++)
1244 s[d] = dd->ci[d] - zones->shift[i][d];
1247 s[d] += dd->numCells[d];
1249 else if (s[d] >= dd->numCells[d])
1251 s[d] -= dd->numCells[d];
1255 for (int iZoneIndex = 0; iZoneIndex < nizone; iZoneIndex++)
1258 ddNonbondedZonePairRanges[iZoneIndex][0] == iZoneIndex,
1259 "The first element for each ddNonbondedZonePairRanges should match its index");
1261 DDPairInteractionRanges iZone;
1262 iZone.iZoneIndex = iZoneIndex;
1263 /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
1264 * j-zones up to nzone.
1266 iZone.jZoneRange = gmx::Range<int>(std::min(ddNonbondedZonePairRanges[iZoneIndex][1], nzone),
1267 std::min(ddNonbondedZonePairRanges[iZoneIndex][2], nzone));
1268 for (dim = 0; dim < DIM; dim++)
1270 if (dd->numCells[dim] == 1)
1272 /* All shifts should be allowed */
1273 iZone.shift0[dim] = -1;
1274 iZone.shift1[dim] = 1;
1278 /* Determine the min/max j-zone shift wrt the i-zone */
1279 iZone.shift0[dim] = 1;
1280 iZone.shift1[dim] = -1;
1281 for (int jZone : iZone.jZoneRange)
1283 int shift_diff = zones->shift[jZone][dim] - zones->shift[iZoneIndex][dim];
1284 if (shift_diff < iZone.shift0[dim])
1286 iZone.shift0[dim] = shift_diff;
1288 if (shift_diff > iZone.shift1[dim])
1290 iZone.shift1[dim] = shift_diff;
1296 zones->iZones.push_back(iZone);
1299 if (!isDlbDisabled(dd->comm))
1301 dd->comm->cellsizesWithDlb.resize(dd->ndim);
1304 if (dd->comm->ddSettings.recordLoad)
1306 make_load_communicators(dd);
1310 static void make_pp_communicator(const gmx::MDLogger& mdlog,
1312 t_commrec gmx_unused* cr,
1313 bool gmx_unused reorder)
1316 gmx_domdec_comm_t* comm = dd->comm;
1317 CartesianRankSetup& cartSetup = comm->cartesianRankSetup;
1319 if (cartSetup.bCartesianPP)
1321 /* Set up cartesian communication for the particle-particle part */
1323 .appendTextFormatted("Will use a Cartesian communicator: %d x %d x %d",
1324 dd->numCells[XX], dd->numCells[YY], dd->numCells[ZZ]);
1327 for (int i = 0; i < DIM; i++)
1332 MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->numCells, periods, static_cast<int>(reorder),
1334 /* We overwrite the old communicator with the new cartesian one */
1335 cr->mpi_comm_mygroup = comm_cart;
1338 dd->mpi_comm_all = cr->mpi_comm_mygroup;
1339 MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
1341 if (cartSetup.bCartesianPP_PME)
1343 /* Since we want to use the original cartesian setup for sim,
1344 * and not the one after split, we need to make an index.
1346 cartSetup.ddindex2ddnodeid.resize(dd->nnodes);
1347 cartSetup.ddindex2ddnodeid[dd_index(dd->numCells, dd->ci)] = dd->rank;
1348 gmx_sumi(dd->nnodes, cartSetup.ddindex2ddnodeid.data(), cr);
1349 /* Get the rank of the DD master,
1350 * above we made sure that the master node is a PP node.
1361 MPI_Allreduce(&rank, &dd->masterrank, 1, MPI_INT, MPI_SUM, dd->mpi_comm_all);
1363 else if (cartSetup.bCartesianPP)
1365 if (!comm->ddRankSetup.usePmeOnlyRanks)
1367 /* The PP communicator is also
1368 * the communicator for this simulation
1370 cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
1372 cr->nodeid = dd->rank;
1374 MPI_Cart_coords(dd->mpi_comm_all, dd->rank, DIM, dd->ci);
1376 /* We need to make an index to go from the coordinates
1377 * to the nodeid of this simulation.
1379 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1380 std::vector<int> buf(dd->nnodes);
1381 if (thisRankHasDuty(cr, DUTY_PP))
1383 buf[dd_index(dd->numCells, dd->ci)] = cr->sim_nodeid;
1385 /* Communicate the ddindex to simulation nodeid index */
1386 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1387 cr->mpi_comm_mysim);
1389 /* Determine the master coordinates and rank.
1390 * The DD master should be the same node as the master of this sim.
1392 for (int i = 0; i < dd->nnodes; i++)
1394 if (cartSetup.ddindex2simnodeid[i] == 0)
1396 ddindex2xyz(dd->numCells, i, dd->master_ci);
1397 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
1402 fprintf(debug, "The master rank is %d\n", dd->masterrank);
1407 /* No Cartesian communicators */
1408 /* We use the rank in dd->comm->all as DD index */
1409 ddindex2xyz(dd->numCells, dd->rank, dd->ci);
1410 /* The simulation master nodeid is 0, so the DD master rank is also 0 */
1412 clear_ivec(dd->master_ci);
1417 .appendTextFormatted("Domain decomposition rank %d, coordinates %d %d %d\n", dd->rank,
1418 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1421 fprintf(debug, "Domain decomposition rank %d, coordinates %d %d %d\n\n", dd->rank,
1422 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1426 static void receive_ddindex2simnodeid(gmx_domdec_t* dd, t_commrec* cr)
1429 CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1431 if (!cartSetup.bCartesianPP_PME && cartSetup.bCartesianPP)
1433 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1434 std::vector<int> buf(dd->nnodes);
1435 if (thisRankHasDuty(cr, DUTY_PP))
1437 buf[dd_index(dd->numCells, dd->ci)] = cr->sim_nodeid;
1439 /* Communicate the ddindex to simulation nodeid index */
1440 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1441 cr->mpi_comm_mysim);
1444 GMX_UNUSED_VALUE(dd);
1445 GMX_UNUSED_VALUE(cr);
1449 static CartesianRankSetup split_communicator(const gmx::MDLogger& mdlog,
1451 const DdRankOrder ddRankOrder,
1452 bool gmx_unused reorder,
1453 const DDRankSetup& ddRankSetup,
1455 std::vector<int>* pmeRanks)
1457 CartesianRankSetup cartSetup;
1459 cartSetup.bCartesianPP = (ddRankOrder == DdRankOrder::cartesian);
1460 cartSetup.bCartesianPP_PME = false;
1462 const ivec& numDDCells = ddRankSetup.numPPCells;
1463 /* Initially we set ntot to the number of PP cells */
1464 copy_ivec(numDDCells, cartSetup.ntot);
1466 if (cartSetup.bCartesianPP)
1468 const int numDDCellsTot = ddRankSetup.numPPRanks;
1470 for (int i = 1; i < DIM; i++)
1472 bDiv[i] = ((ddRankSetup.numRanksDoingPme * numDDCells[i]) % numDDCellsTot == 0);
1474 if (bDiv[YY] || bDiv[ZZ])
1476 cartSetup.bCartesianPP_PME = TRUE;
1477 /* If we have 2D PME decomposition, which is always in x+y,
1478 * we stack the PME only nodes in z.
1479 * Otherwise we choose the direction that provides the thinnest slab
1480 * of PME only nodes as this will have the least effect
1481 * on the PP communication.
1482 * But for the PME communication the opposite might be better.
1484 if (bDiv[ZZ] && (ddRankSetup.npmenodes_y > 1 || !bDiv[YY] || numDDCells[YY] > numDDCells[ZZ]))
1486 cartSetup.cartpmedim = ZZ;
1490 cartSetup.cartpmedim = YY;
1492 cartSetup.ntot[cartSetup.cartpmedim] +=
1493 (ddRankSetup.numRanksDoingPme * numDDCells[cartSetup.cartpmedim]) / numDDCellsTot;
1498 .appendTextFormatted(
1499 "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or "
1501 ddRankSetup.numRanksDoingPme, numDDCells[XX], numDDCells[YY],
1502 numDDCells[XX], numDDCells[ZZ]);
1504 .appendText("Will not use a Cartesian communicator for PP <-> PME\n");
1508 if (cartSetup.bCartesianPP_PME)
1515 .appendTextFormatted(
1516 "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d",
1517 cartSetup.ntot[XX], cartSetup.ntot[YY], cartSetup.ntot[ZZ]);
1519 for (int i = 0; i < DIM; i++)
1524 MPI_Cart_create(cr->mpi_comm_mysim, DIM, cartSetup.ntot, periods, static_cast<int>(reorder),
1526 MPI_Comm_rank(comm_cart, &rank);
1527 if (MASTER(cr) && rank != 0)
1529 gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
1532 /* With this assigment we loose the link to the original communicator
1533 * which will usually be MPI_COMM_WORLD, unless have multisim.
1535 cr->mpi_comm_mysim = comm_cart;
1536 cr->sim_nodeid = rank;
1538 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, ddCellIndex);
1541 .appendTextFormatted("Cartesian rank %d, coordinates %d %d %d\n", cr->sim_nodeid,
1542 ddCellIndex[XX], ddCellIndex[YY], ddCellIndex[ZZ]);
1544 if (ddCellIndex[cartSetup.cartpmedim] < numDDCells[cartSetup.cartpmedim])
1548 if (!ddRankSetup.usePmeOnlyRanks
1549 || ddCellIndex[cartSetup.cartpmedim] >= numDDCells[cartSetup.cartpmedim])
1551 cr->duty = DUTY_PME;
1554 /* Split the sim communicator into PP and PME only nodes */
1555 MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr),
1556 dd_index(cartSetup.ntot, ddCellIndex), &cr->mpi_comm_mygroup);
1558 GMX_UNUSED_VALUE(ddCellIndex);
1563 switch (ddRankOrder)
1565 case DdRankOrder::pp_pme:
1566 GMX_LOG(mdlog.info).appendText("Order of the ranks: PP first, PME last");
1568 case DdRankOrder::interleave:
1569 /* Interleave the PP-only and PME-only ranks */
1570 GMX_LOG(mdlog.info).appendText("Interleaving PP and PME ranks");
1571 *pmeRanks = dd_interleaved_pme_ranks(ddRankSetup);
1573 case DdRankOrder::cartesian: break;
1574 default: gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(ddRankOrder));
1577 if (dd_simnode2pmenode(ddRankSetup, cartSetup, *pmeRanks, cr, cr->sim_nodeid) == -1)
1579 cr->duty = DUTY_PME;
1586 /* Split the sim communicator into PP and PME only nodes */
1587 MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr), cr->nodeid, &cr->mpi_comm_mygroup);
1588 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
1593 .appendTextFormatted("This rank does only %s work.\n",
1594 thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
1599 /*! \brief Makes the PP communicator and the PME communicator, when needed
1601 * Returns the Cartesian rank setup.
1602 * Sets \p cr->mpi_comm_mygroup
1603 * For PP ranks, sets the DD PP cell index in \p ddCellIndex.
1604 * With separate PME ranks in interleaved order, set the PME ranks in \p pmeRanks.
1606 static CartesianRankSetup makeGroupCommunicators(const gmx::MDLogger& mdlog,
1607 const DDSettings& ddSettings,
1608 const DdRankOrder ddRankOrder,
1609 const DDRankSetup& ddRankSetup,
1612 std::vector<int>* pmeRanks)
1614 CartesianRankSetup cartSetup;
1616 if (ddRankSetup.usePmeOnlyRanks)
1618 /* Split the communicator into a PP and PME part */
1619 cartSetup = split_communicator(mdlog, cr, ddRankOrder, ddSettings.useCartesianReorder,
1620 ddRankSetup, ddCellIndex, pmeRanks);
1624 /* All nodes do PP and PME */
1625 /* We do not require separate communicators */
1626 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
1628 cartSetup.bCartesianPP = false;
1629 cartSetup.bCartesianPP_PME = false;
1635 /*! \brief For PP ranks, sets or makes the communicator
1637 * For PME ranks get the rank id.
1638 * For PP only ranks, sets the PME-only rank.
1640 static void setupGroupCommunication(const gmx::MDLogger& mdlog,
1641 const DDSettings& ddSettings,
1642 gmx::ArrayRef<const int> pmeRanks,
1644 const int numAtomsInSystem,
1647 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1648 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1650 if (thisRankHasDuty(cr, DUTY_PP))
1652 /* Copy or make a new PP communicator */
1654 /* We (possibly) reordered the nodes in split_communicator,
1655 * so it is no longer required in make_pp_communicator.
1657 const bool useCartesianReorder = (ddSettings.useCartesianReorder && !cartSetup.bCartesianPP_PME);
1659 make_pp_communicator(mdlog, dd, cr, useCartesianReorder);
1663 receive_ddindex2simnodeid(dd, cr);
1666 if (!thisRankHasDuty(cr, DUTY_PME))
1668 /* Set up the commnuication to our PME node */
1669 dd->pme_nodeid = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
1670 dd->pme_receive_vir_ener = receive_vir_ener(dd, pmeRanks, cr);
1673 fprintf(debug, "My pme_nodeid %d receive ener %s\n", dd->pme_nodeid,
1674 gmx::boolToString(dd->pme_receive_vir_ener));
1679 dd->pme_nodeid = -1;
1682 /* We can not use DDMASTER(dd), because dd->masterrank is set later */
1685 dd->ma = std::make_unique<AtomDistribution>(dd->numCells, numAtomsInSystem, numAtomsInSystem);
1689 static real* get_slb_frac(const gmx::MDLogger& mdlog, const char* dir, int nc, const char* size_string)
1691 real * slb_frac, tot;
1696 if (nc > 1 && size_string != nullptr)
1698 GMX_LOG(mdlog.info).appendTextFormatted("Using static load balancing for the %s direction", dir);
1701 for (i = 0; i < nc; i++)
1704 sscanf(size_string, "%20lf%n", &dbl, &n);
1708 "Incorrect or not enough DD cell size entries for direction %s: '%s'",
1716 std::string relativeCellSizes = "Relative cell sizes:";
1717 for (i = 0; i < nc; i++)
1720 relativeCellSizes += gmx::formatString(" %5.3f", slb_frac[i]);
1722 GMX_LOG(mdlog.info).appendText(relativeCellSizes);
1728 static int multi_body_bondeds_count(const gmx_mtop_t* mtop)
1731 gmx_mtop_ilistloop_t iloop = gmx_mtop_ilistloop_init(mtop);
1733 while (const InteractionLists* ilists = gmx_mtop_ilistloop_next(iloop, &nmol))
1735 for (auto& ilist : extractILists(*ilists, IF_BOND))
1737 if (NRAL(ilist.functionType) > 2)
1739 n += nmol * (ilist.iatoms.size() / ilistStride(ilist));
1747 static int dd_getenv(const gmx::MDLogger& mdlog, const char* env_var, int def)
1753 val = getenv(env_var);
1756 if (sscanf(val, "%20d", &nst) <= 0)
1760 GMX_LOG(mdlog.info).appendTextFormatted("Found env.var. %s = %s, using value %d", env_var, val, nst);
1766 static void check_dd_restrictions(const gmx_domdec_t* dd, const t_inputrec* ir, const gmx::MDLogger& mdlog)
1768 if (ir->pbcType == PbcType::Screw
1769 && (dd->numCells[XX] == 1 || dd->numCells[YY] > 1 || dd->numCells[ZZ] > 1))
1771 gmx_fatal(FARGS, "With pbc=%s can only do domain decomposition in the x-direction",
1772 c_pbcTypeNames[ir->pbcType].c_str());
1775 if (ir->nstlist == 0)
1777 gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
1780 if (ir->comm_mode == ecmANGULAR && ir->pbcType != PbcType::No)
1782 GMX_LOG(mdlog.warning)
1784 "comm-mode angular will give incorrect results when the comm group "
1785 "partially crosses a periodic boundary");
1789 static real average_cellsize_min(const gmx_ddbox_t& ddbox, const ivec numDomains)
1791 real r = ddbox.box_size[XX];
1792 for (int d = 0; d < DIM; d++)
1794 if (numDomains[d] > 1)
1796 /* Check using the initial average cell size */
1797 r = std::min(r, ddbox.box_size[d] * ddbox.skew_fac[d] / numDomains[d]);
1804 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
1806 static DlbState forceDlbOffOrBail(DlbState cmdlineDlbState,
1807 const std::string& reasonStr,
1808 const gmx::MDLogger& mdlog)
1810 std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
1811 std::string dlbDisableNote = "NOTE: disabling dynamic load balancing as ";
1813 if (cmdlineDlbState == DlbState::onUser)
1815 gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
1817 else if (cmdlineDlbState == DlbState::offCanTurnOn)
1819 GMX_LOG(mdlog.info).appendText(dlbDisableNote + reasonStr);
1821 return DlbState::offForever;
1824 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
1826 * This function parses the parameters of "-dlb" command line option setting
1827 * corresponding state values. Then it checks the consistency of the determined
1828 * state with other run parameters and settings. As a result, the initial state
1829 * may be altered or an error may be thrown if incompatibility of options is detected.
1831 * \param [in] mdlog Logger.
1832 * \param [in] dlbOption Enum value for the DLB option.
1833 * \param [in] bRecordLoad True if the load balancer is recording load information.
1834 * \param [in] mdrunOptions Options for mdrun.
1835 * \param [in] ir Pointer mdrun to input parameters.
1836 * \returns DLB initial/startup state.
1838 static DlbState determineInitialDlbState(const gmx::MDLogger& mdlog,
1839 DlbOption dlbOption,
1840 gmx_bool bRecordLoad,
1841 const gmx::MdrunOptions& mdrunOptions,
1842 const t_inputrec* ir)
1844 DlbState dlbState = DlbState::offCanTurnOn;
1848 case DlbOption::turnOnWhenUseful: dlbState = DlbState::offCanTurnOn; break;
1849 case DlbOption::no: dlbState = DlbState::offUser; break;
1850 case DlbOption::yes: dlbState = DlbState::onUser; break;
1851 default: gmx_incons("Invalid dlbOption enum value");
1854 /* Reruns don't support DLB: bail or override auto mode */
1855 if (mdrunOptions.rerun)
1857 std::string reasonStr = "it is not supported in reruns.";
1858 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1861 /* Unsupported integrators */
1862 if (!EI_DYNAMICS(ir->eI))
1864 auto reasonStr = gmx::formatString(
1865 "it is only supported with dynamics, not with integrator '%s'.", EI(ir->eI));
1866 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1869 /* Without cycle counters we can't time work to balance on */
1872 std::string reasonStr =
1873 "cycle counters unsupported or not enabled in the operating system kernel.";
1874 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1877 if (mdrunOptions.reproducible)
1879 std::string reasonStr = "you started a reproducible run.";
1882 case DlbState::offUser: break;
1883 case DlbState::offForever:
1884 GMX_RELEASE_ASSERT(false, "DlbState::offForever is not a valid initial state");
1886 case DlbState::offCanTurnOn: return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1887 case DlbState::onCanTurnOff:
1888 GMX_RELEASE_ASSERT(false, "DlbState::offCanTurnOff is not a valid initial state");
1890 case DlbState::onUser:
1891 return forceDlbOffOrBail(
1894 + " In load balanced runs binary reproducibility cannot be "
1898 gmx_fatal(FARGS, "Death horror: undefined case (%d) for load balancing choice",
1899 static_cast<int>(dlbState));
1906 static gmx_domdec_comm_t* init_dd_comm()
1908 gmx_domdec_comm_t* comm = new gmx_domdec_comm_t;
1910 comm->n_load_have = 0;
1911 comm->n_load_collect = 0;
1913 comm->haveTurnedOffDlb = false;
1915 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
1917 comm->sum_nat[i] = 0;
1921 comm->load_step = 0;
1924 clear_ivec(comm->load_lim);
1928 /* This should be replaced by a unique pointer */
1929 comm->balanceRegion = ddBalanceRegionAllocate();
1934 /* Returns whether mtop contains constraints and/or vsites */
1935 static bool systemHasConstraintsOrVsites(const gmx_mtop_t& mtop)
1937 auto ilistLoop = gmx_mtop_ilistloop_init(mtop);
1939 while (const InteractionLists* ilists = gmx_mtop_ilistloop_next(ilistLoop, &nmol))
1941 if (!extractILists(*ilists, IF_CONSTRAINT | IF_VSITE).empty())
1950 static void setupUpdateGroups(const gmx::MDLogger& mdlog,
1951 const gmx_mtop_t& mtop,
1952 const t_inputrec& inputrec,
1953 const real cutoffMargin,
1954 DDSystemInfo* systemInfo)
1956 /* When we have constraints and/or vsites, it is beneficial to use
1957 * update groups (when possible) to allow independent update of groups.
1959 if (!systemHasConstraintsOrVsites(mtop))
1961 /* No constraints or vsites, atoms can be updated independently */
1965 systemInfo->updateGroupingPerMoleculetype = gmx::makeUpdateGroups(mtop);
1966 systemInfo->useUpdateGroups = (!systemInfo->updateGroupingPerMoleculetype.empty()
1967 && getenv("GMX_NO_UPDATEGROUPS") == nullptr);
1969 if (systemInfo->useUpdateGroups)
1971 int numUpdateGroups = 0;
1972 for (const auto& molblock : mtop.molblock)
1974 numUpdateGroups += molblock.nmol
1975 * systemInfo->updateGroupingPerMoleculetype[molblock.type].numBlocks();
1978 systemInfo->maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1979 mtop, systemInfo->updateGroupingPerMoleculetype, maxReferenceTemperature(inputrec));
1981 /* To use update groups, the large domain-to-domain cutoff distance
1982 * should be compatible with the box size.
1984 systemInfo->useUpdateGroups = (atomToAtomIntoDomainToDomainCutoff(*systemInfo, 0) < cutoffMargin);
1986 if (systemInfo->useUpdateGroups)
1989 .appendTextFormatted(
1990 "Using update groups, nr %d, average size %.1f atoms, max. radius %.3f "
1992 numUpdateGroups, mtop.natoms / static_cast<double>(numUpdateGroups),
1993 systemInfo->maxUpdateGroupRadius);
1998 .appendTextFormatted(
1999 "The combination of rlist and box size prohibits the use of update "
2001 systemInfo->updateGroupingPerMoleculetype.clear();
2006 UnitCellInfo::UnitCellInfo(const t_inputrec& ir) :
2007 npbcdim(numPbcDimensions(ir.pbcType)),
2008 numBoundedDimensions(inputrec2nboundeddim(&ir)),
2009 ddBoxIsDynamic(numBoundedDimensions < DIM || inputrecDynamicBox(&ir)),
2010 haveScrewPBC(ir.pbcType == PbcType::Screw)
2014 /* Returns whether molecules are always whole, i.e. not broken by PBC */
2015 static bool moleculesAreAlwaysWhole(const gmx_mtop_t& mtop,
2016 const bool useUpdateGroups,
2017 gmx::ArrayRef<const gmx::RangePartitioning> updateGroupingPerMoleculetype)
2019 if (useUpdateGroups)
2021 GMX_RELEASE_ASSERT(updateGroupingPerMoleculetype.size() == mtop.moltype.size(),
2022 "Need one grouping per moltype");
2023 for (size_t mol = 0; mol < mtop.moltype.size(); mol++)
2025 if (updateGroupingPerMoleculetype[mol].numBlocks() > 1)
2033 for (const auto& moltype : mtop.moltype)
2035 if (moltype.atoms.nr > 1)
2045 /*! \brief Generate the simulation system information */
2046 static DDSystemInfo getSystemInfo(const gmx::MDLogger& mdlog,
2047 const t_commrec* cr,
2048 const DomdecOptions& options,
2049 const gmx_mtop_t& mtop,
2050 const t_inputrec& ir,
2052 gmx::ArrayRef<const gmx::RVec> xGlobal)
2054 const real tenPercentMargin = 1.1;
2056 DDSystemInfo systemInfo;
2058 /* We need to decide on update groups early, as this affects communication distances */
2059 systemInfo.useUpdateGroups = false;
2060 if (ir.cutoff_scheme == ecutsVERLET)
2062 real cutoffMargin = std::sqrt(max_cutoff2(ir.pbcType, box)) - ir.rlist;
2063 setupUpdateGroups(mdlog, mtop, ir, cutoffMargin, &systemInfo);
2066 systemInfo.moleculesAreAlwaysWhole = moleculesAreAlwaysWhole(
2067 mtop, systemInfo.useUpdateGroups, systemInfo.updateGroupingPerMoleculetype);
2068 systemInfo.haveInterDomainBondeds =
2069 (!systemInfo.moleculesAreAlwaysWhole || mtop.bIntermolecularInteractions);
2070 systemInfo.haveInterDomainMultiBodyBondeds =
2071 (systemInfo.haveInterDomainBondeds && multi_body_bondeds_count(&mtop) > 0);
2073 if (systemInfo.useUpdateGroups)
2075 systemInfo.haveSplitConstraints = false;
2076 systemInfo.haveSplitSettles = false;
2080 systemInfo.haveSplitConstraints = (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0
2081 || gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0);
2082 systemInfo.haveSplitSettles = (gmx_mtop_ftype_count(mtop, F_SETTLE) > 0);
2087 /* Set the cut-off to some very large value,
2088 * so we don't need if statements everywhere in the code.
2089 * We use sqrt, since the cut-off is squared in some places.
2091 systemInfo.cutoff = GMX_CUTOFF_INF;
2095 systemInfo.cutoff = atomToAtomIntoDomainToDomainCutoff(systemInfo, ir.rlist);
2097 systemInfo.minCutoffForMultiBody = 0;
2099 /* Determine the minimum cell size limit, affected by many factors */
2100 systemInfo.cellsizeLimit = 0;
2101 systemInfo.filterBondedCommunication = false;
2103 /* We do not allow home atoms to move beyond the neighboring domain
2104 * between domain decomposition steps, which limits the cell size.
2105 * Get an estimate of cell size limit due to atom displacement.
2106 * In most cases this is a large overestimate, because it assumes
2107 * non-interaction atoms.
2108 * We set the chance to 1 in a trillion steps.
2110 constexpr real c_chanceThatAtomMovesBeyondDomain = 1e-12;
2111 const real limitForAtomDisplacement = minCellSizeForAtomDisplacement(
2112 mtop, ir, systemInfo.updateGroupingPerMoleculetype, c_chanceThatAtomMovesBeyondDomain);
2113 GMX_LOG(mdlog.info).appendTextFormatted("Minimum cell size due to atom displacement: %.3f nm", limitForAtomDisplacement);
2115 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, limitForAtomDisplacement);
2117 /* TODO: PME decomposition currently requires atoms not to be more than
2118 * 2/3 of comm->cutoff, which is >=rlist, outside of their domain.
2119 * In nearly all cases, limitForAtomDisplacement will be smaller
2120 * than 2/3*rlist, so the PME requirement is satisfied.
2121 * But it would be better for both correctness and performance
2122 * to use limitForAtomDisplacement instead of 2/3*comm->cutoff.
2123 * Note that we would need to improve the pairlist buffer case.
2126 if (systemInfo.haveInterDomainBondeds)
2128 if (options.minimumCommunicationRange > 0)
2130 systemInfo.minCutoffForMultiBody =
2131 atomToAtomIntoDomainToDomainCutoff(systemInfo, options.minimumCommunicationRange);
2132 if (options.useBondedCommunication)
2134 systemInfo.filterBondedCommunication =
2135 (systemInfo.minCutoffForMultiBody > systemInfo.cutoff);
2139 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2142 else if (ir.bPeriodicMols)
2144 /* Can not easily determine the required cut-off */
2145 GMX_LOG(mdlog.warning)
2147 "NOTE: Periodic molecules are present in this system. Because of this, "
2148 "the domain decomposition algorithm cannot easily determine the "
2149 "minimum cell size that it requires for treating bonded interactions. "
2150 "Instead, domain decomposition will assume that half the non-bonded "
2151 "cut-off will be a suitable lower bound.");
2152 systemInfo.minCutoffForMultiBody = systemInfo.cutoff / 2;
2160 dd_bonded_cg_distance(mdlog, &mtop, &ir, as_rvec_array(xGlobal.data()), box,
2161 options.checkBondedInteractions, &r_2b, &r_mb);
2163 gmx_bcast(sizeof(r_2b), &r_2b, cr);
2164 gmx_bcast(sizeof(r_mb), &r_mb, cr);
2166 /* We use an initial margin of 10% for the minimum cell size,
2167 * except when we are just below the non-bonded cut-off.
2169 if (options.useBondedCommunication)
2171 if (std::max(r_2b, r_mb) > systemInfo.cutoff)
2173 const real r_bonded = std::max(r_2b, r_mb);
2174 systemInfo.minCutoffForMultiBody = tenPercentMargin * r_bonded;
2175 /* This is the (only) place where we turn on the filtering */
2176 systemInfo.filterBondedCommunication = true;
2180 const real r_bonded = r_mb;
2181 systemInfo.minCutoffForMultiBody =
2182 std::min(tenPercentMargin * r_bonded, systemInfo.cutoff);
2184 /* We determine cutoff_mbody later */
2185 systemInfo.increaseMultiBodyCutoff = true;
2189 /* No special bonded communication,
2190 * simply increase the DD cut-off.
2192 systemInfo.minCutoffForMultiBody = tenPercentMargin * std::max(r_2b, r_mb);
2193 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2197 .appendTextFormatted("Minimum cell size due to bonded interactions: %.3f nm",
2198 systemInfo.minCutoffForMultiBody);
2200 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.minCutoffForMultiBody);
2203 systemInfo.constraintCommunicationRange = 0;
2204 if (systemInfo.haveSplitConstraints && options.constraintCommunicationRange <= 0)
2206 /* There is a cell size limit due to the constraints (P-LINCS) */
2207 systemInfo.constraintCommunicationRange = gmx::constr_r_max(mdlog, &mtop, &ir);
2209 .appendTextFormatted("Estimated maximum distance required for P-LINCS: %.3f nm",
2210 systemInfo.constraintCommunicationRange);
2211 if (systemInfo.constraintCommunicationRange > systemInfo.cellsizeLimit)
2215 "This distance will limit the DD cell size, you can override this with "
2219 else if (options.constraintCommunicationRange > 0)
2221 /* Here we do not check for dd->splitConstraints.
2222 * because one can also set a cell size limit for virtual sites only
2223 * and at this point we don't know yet if there are intercg v-sites.
2226 .appendTextFormatted("User supplied maximum distance required for P-LINCS: %.3f nm",
2227 options.constraintCommunicationRange);
2228 systemInfo.constraintCommunicationRange = options.constraintCommunicationRange;
2230 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.constraintCommunicationRange);
2235 /*! \brief Exit with a fatal error if the DDGridSetup cannot be
2237 static void checkDDGridSetup(const DDGridSetup& ddGridSetup,
2238 const t_commrec* cr,
2239 const DomdecOptions& options,
2240 const DDSettings& ddSettings,
2241 const DDSystemInfo& systemInfo,
2242 const real cellsizeLimit,
2243 const gmx_ddbox_t& ddbox)
2245 if (options.numCells[XX] <= 0 && (ddGridSetup.numDomains[XX] == 0))
2248 gmx_bool bC = (systemInfo.haveSplitConstraints
2249 && systemInfo.constraintCommunicationRange > systemInfo.minCutoffForMultiBody);
2250 sprintf(buf, "Change the number of ranks or mdrun option %s%s%s", !bC ? "-rdd" : "-rcon",
2251 ddSettings.initialDlbState != DlbState::offUser ? " or -dds" : "",
2252 bC ? " or your LINCS settings" : "");
2254 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2255 "There is no domain decomposition for %d ranks that is compatible "
2256 "with the given box and a minimum cell size of %g nm\n"
2258 "Look in the log file for details on the domain decomposition",
2259 cr->nnodes - ddGridSetup.numPmeOnlyRanks, cellsizeLimit, buf);
2262 const real acs = average_cellsize_min(ddbox, ddGridSetup.numDomains);
2263 if (acs < cellsizeLimit)
2265 if (options.numCells[XX] <= 0)
2269 "dd_choose_grid() should return a grid that satisfies the cell size limits");
2273 gmx_fatal_collective(
2274 FARGS, cr->mpi_comm_mysim, MASTER(cr),
2275 "The initial cell size (%f) is smaller than the cell size limit (%f), change "
2276 "options -dd, -rdd or -rcon, see the log file for details",
2277 acs, cellsizeLimit);
2281 const int numPPRanks =
2282 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2283 if (cr->nnodes - numPPRanks != ddGridSetup.numPmeOnlyRanks)
2285 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2286 "The size of the domain decomposition grid (%d) does not match the "
2287 "number of PP ranks (%d). The total number of ranks is %d",
2288 numPPRanks, cr->nnodes - ddGridSetup.numPmeOnlyRanks, cr->nnodes);
2290 if (ddGridSetup.numPmeOnlyRanks > numPPRanks)
2292 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2293 "The number of separate PME ranks (%d) is larger than the number of "
2294 "PP ranks (%d), this is not supported.",
2295 ddGridSetup.numPmeOnlyRanks, numPPRanks);
2299 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
2300 static DDRankSetup getDDRankSetup(const gmx::MDLogger& mdlog,
2302 const DDGridSetup& ddGridSetup,
2303 const t_inputrec& ir)
2306 .appendTextFormatted("Domain decomposition grid %d x %d x %d, separate PME ranks %d",
2307 ddGridSetup.numDomains[XX], ddGridSetup.numDomains[YY],
2308 ddGridSetup.numDomains[ZZ], ddGridSetup.numPmeOnlyRanks);
2310 DDRankSetup ddRankSetup;
2312 ddRankSetup.numPPRanks = cr->nnodes - ddGridSetup.numPmeOnlyRanks;
2313 copy_ivec(ddGridSetup.numDomains, ddRankSetup.numPPCells);
2315 ddRankSetup.usePmeOnlyRanks = (ddGridSetup.numPmeOnlyRanks > 0);
2316 if (ddRankSetup.usePmeOnlyRanks)
2318 ddRankSetup.numRanksDoingPme = ddGridSetup.numPmeOnlyRanks;
2322 ddRankSetup.numRanksDoingPme =
2323 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2326 if (EEL_PME(ir.coulombtype) || EVDW_PME(ir.vdwtype))
2328 /* The following choices should match those
2329 * in comm_cost_est in domdec_setup.c.
2330 * Note that here the checks have to take into account
2331 * that the decomposition might occur in a different order than xyz
2332 * (for instance through the env.var. GMX_DD_ORDER_ZYX),
2333 * in which case they will not match those in comm_cost_est,
2334 * but since that is mainly for testing purposes that's fine.
2336 if (ddGridSetup.numDDDimensions >= 2 && ddGridSetup.ddDimensions[0] == XX
2337 && ddGridSetup.ddDimensions[1] == YY
2338 && ddRankSetup.numRanksDoingPme > ddGridSetup.numDomains[XX]
2339 && ddRankSetup.numRanksDoingPme % ddGridSetup.numDomains[XX] == 0
2340 && getenv("GMX_PMEONEDD") == nullptr)
2342 ddRankSetup.npmedecompdim = 2;
2343 ddRankSetup.npmenodes_x = ddGridSetup.numDomains[XX];
2344 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme / ddRankSetup.npmenodes_x;
2348 /* In case nc is 1 in both x and y we could still choose to
2349 * decompose pme in y instead of x, but we use x for simplicity.
2351 ddRankSetup.npmedecompdim = 1;
2352 if (ddGridSetup.ddDimensions[0] == YY)
2354 ddRankSetup.npmenodes_x = 1;
2355 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme;
2359 ddRankSetup.npmenodes_x = ddRankSetup.numRanksDoingPme;
2360 ddRankSetup.npmenodes_y = 1;
2364 .appendTextFormatted("PME domain decomposition: %d x %d x %d",
2365 ddRankSetup.npmenodes_x, ddRankSetup.npmenodes_y, 1);
2369 ddRankSetup.npmedecompdim = 0;
2370 ddRankSetup.npmenodes_x = 0;
2371 ddRankSetup.npmenodes_y = 0;
2377 /*! \brief Set the cell size and interaction limits */
2378 static void set_dd_limits(const gmx::MDLogger& mdlog,
2381 const DomdecOptions& options,
2382 const DDSettings& ddSettings,
2383 const DDSystemInfo& systemInfo,
2384 const DDGridSetup& ddGridSetup,
2385 const int numPPRanks,
2386 const gmx_mtop_t* mtop,
2387 const t_inputrec* ir,
2388 const gmx_ddbox_t& ddbox)
2390 gmx_domdec_comm_t* comm = dd->comm;
2391 comm->ddSettings = ddSettings;
2393 /* Initialize to GPU share count to 0, might change later */
2394 comm->nrank_gpu_shared = 0;
2396 comm->dlbState = comm->ddSettings.initialDlbState;
2397 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
2398 /* To consider turning DLB on after 2*nstlist steps we need to check
2399 * at partitioning count 3. Thus we need to increase the first count by 2.
2401 comm->ddPartioningCountFirstDlbOff += 2;
2403 comm->bPMELoadBalDLBLimits = FALSE;
2405 /* Allocate the charge group/atom sorting struct */
2406 comm->sort = std::make_unique<gmx_domdec_sort_t>();
2408 comm->systemInfo = systemInfo;
2410 if (systemInfo.useUpdateGroups)
2412 /* Note: We would like to use dd->nnodes for the atom count estimate,
2413 * but that is not yet available here. But this anyhow only
2414 * affect performance up to the second dd_partition_system call.
2416 const int homeAtomCountEstimate = mtop->natoms / numPPRanks;
2417 comm->updateGroupsCog = std::make_unique<gmx::UpdateGroupsCog>(
2418 *mtop, systemInfo.updateGroupingPerMoleculetype, maxReferenceTemperature(*ir),
2419 homeAtomCountEstimate);
2422 /* Set the DD setup given by ddGridSetup */
2423 copy_ivec(ddGridSetup.numDomains, dd->numCells);
2424 dd->ndim = ddGridSetup.numDDDimensions;
2425 copy_ivec(ddGridSetup.ddDimensions, dd->dim);
2427 dd->nnodes = dd->numCells[XX] * dd->numCells[YY] * dd->numCells[ZZ];
2429 snew(comm->slb_frac, DIM);
2430 if (isDlbDisabled(comm))
2432 comm->slb_frac[XX] = get_slb_frac(mdlog, "x", dd->numCells[XX], options.cellSizeX);
2433 comm->slb_frac[YY] = get_slb_frac(mdlog, "y", dd->numCells[YY], options.cellSizeY);
2434 comm->slb_frac[ZZ] = get_slb_frac(mdlog, "z", dd->numCells[ZZ], options.cellSizeZ);
2437 /* Set the multi-body cut-off and cellsize limit for DLB */
2438 comm->cutoff_mbody = systemInfo.minCutoffForMultiBody;
2439 comm->cellsize_limit = systemInfo.cellsizeLimit;
2440 if (systemInfo.haveInterDomainBondeds && systemInfo.increaseMultiBodyCutoff)
2442 if (systemInfo.filterBondedCommunication || !isDlbDisabled(comm))
2444 /* Set the bonded communication distance to halfway
2445 * the minimum and the maximum,
2446 * since the extra communication cost is nearly zero.
2448 real acs = average_cellsize_min(ddbox, dd->numCells);
2449 comm->cutoff_mbody = 0.5 * (systemInfo.minCutoffForMultiBody + acs);
2450 if (!isDlbDisabled(comm))
2452 /* Check if this does not limit the scaling */
2453 comm->cutoff_mbody = std::min(comm->cutoff_mbody, options.dlbScaling * acs);
2455 if (!systemInfo.filterBondedCommunication)
2457 /* Without bBondComm do not go beyond the n.b. cut-off */
2458 comm->cutoff_mbody = std::min(comm->cutoff_mbody, systemInfo.cutoff);
2459 if (comm->cellsize_limit >= systemInfo.cutoff)
2461 /* We don't loose a lot of efficieny
2462 * when increasing it to the n.b. cut-off.
2463 * It can even be slightly faster, because we need
2464 * less checks for the communication setup.
2466 comm->cutoff_mbody = systemInfo.cutoff;
2469 /* Check if we did not end up below our original limit */
2470 comm->cutoff_mbody = std::max(comm->cutoff_mbody, systemInfo.minCutoffForMultiBody);
2472 if (comm->cutoff_mbody > comm->cellsize_limit)
2474 comm->cellsize_limit = comm->cutoff_mbody;
2477 /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
2483 "Bonded atom communication beyond the cut-off: %s\n"
2484 "cellsize limit %f\n",
2485 gmx::boolToString(systemInfo.filterBondedCommunication), comm->cellsize_limit);
2490 check_dd_restrictions(dd, ir, mdlog);
2494 void dd_init_bondeds(FILE* fplog,
2496 const gmx_mtop_t& mtop,
2497 const gmx_vsite_t* vsite,
2498 const t_inputrec* ir,
2500 gmx::ArrayRef<cginfo_mb_t> cginfo_mb)
2502 gmx_domdec_comm_t* comm;
2504 dd_make_reverse_top(fplog, dd, &mtop, vsite, ir, bBCheck);
2508 if (comm->systemInfo.filterBondedCommunication)
2510 /* Communicate atoms beyond the cut-off for bonded interactions */
2511 comm->bondedLinks = makeBondedLinks(mtop, cginfo_mb);
2515 /* Only communicate atoms based on cut-off */
2516 comm->bondedLinks = nullptr;
2520 static void writeSettings(gmx::TextWriter* log,
2522 const gmx_mtop_t* mtop,
2523 const t_inputrec* ir,
2524 gmx_bool bDynLoadBal,
2526 const gmx_ddbox_t* ddbox)
2528 gmx_domdec_comm_t* comm;
2537 log->writeString("The maximum number of communication pulses is:");
2538 for (d = 0; d < dd->ndim; d++)
2540 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
2542 log->ensureLineBreak();
2543 log->writeLineFormatted("The minimum size for domain decomposition cells is %.3f nm",
2544 comm->cellsize_limit);
2545 log->writeLineFormatted("The requested allowed shrink of DD cells (option -dds) is: %.2f", dlb_scale);
2546 log->writeString("The allowed shrink of domain decomposition cells is:");
2547 for (d = 0; d < DIM; d++)
2549 if (dd->numCells[d] > 1)
2551 if (d >= ddbox->npbcdim && dd->numCells[d] == 2)
2557 shrink = comm->cellsize_min_dlb[d]
2558 / (ddbox->box_size[d] * ddbox->skew_fac[d] / dd->numCells[d]);
2560 log->writeStringFormatted(" %c %.2f", dim2char(d), shrink);
2563 log->ensureLineBreak();
2567 set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
2568 log->writeString("The initial number of communication pulses is:");
2569 for (d = 0; d < dd->ndim; d++)
2571 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
2573 log->ensureLineBreak();
2574 log->writeString("The initial domain decomposition cell size is:");
2575 for (d = 0; d < DIM; d++)
2577 if (dd->numCells[d] > 1)
2579 log->writeStringFormatted(" %c %.2f nm", dim2char(d), dd->comm->cellsize_min[d]);
2582 log->ensureLineBreak();
2586 const bool haveInterDomainVsites =
2587 (countInterUpdategroupVsites(*mtop, comm->systemInfo.updateGroupingPerMoleculetype) != 0);
2589 if (comm->systemInfo.haveInterDomainBondeds || haveInterDomainVsites
2590 || comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2592 std::string decompUnits;
2593 if (comm->systemInfo.useUpdateGroups)
2595 decompUnits = "atom groups";
2599 decompUnits = "atoms";
2602 log->writeLineFormatted("The maximum allowed distance for %s involved in interactions is:",
2603 decompUnits.c_str());
2604 log->writeLineFormatted("%40s %-7s %6.3f nm", "non-bonded interactions", "",
2605 comm->systemInfo.cutoff);
2609 limit = dd->comm->cellsize_limit;
2613 if (dd->unitCellInfo.ddBoxIsDynamic)
2616 "(the following are initial values, they could change due to box "
2619 limit = dd->comm->cellsize_min[XX];
2620 for (d = 1; d < DIM; d++)
2622 limit = std::min(limit, dd->comm->cellsize_min[d]);
2626 if (comm->systemInfo.haveInterDomainBondeds)
2628 log->writeLineFormatted("%40s %-7s %6.3f nm", "two-body bonded interactions", "(-rdd)",
2629 std::max(comm->systemInfo.cutoff, comm->cutoff_mbody));
2630 log->writeLineFormatted("%40s %-7s %6.3f nm", "multi-body bonded interactions",
2632 (comm->systemInfo.filterBondedCommunication || isDlbOn(dd->comm))
2633 ? comm->cutoff_mbody
2634 : std::min(comm->systemInfo.cutoff, limit));
2636 if (haveInterDomainVsites)
2638 log->writeLineFormatted("%40s %-7s %6.3f nm", "virtual site constructions", "(-rcon)", limit);
2640 if (comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2642 std::string separation =
2643 gmx::formatString("atoms separated by up to %d constraints", 1 + ir->nProjOrder);
2644 log->writeLineFormatted("%40s %-7s %6.3f nm\n", separation.c_str(), "(-rcon)", limit);
2646 log->ensureLineBreak();
2650 static void logSettings(const gmx::MDLogger& mdlog,
2652 const gmx_mtop_t* mtop,
2653 const t_inputrec* ir,
2655 const gmx_ddbox_t* ddbox)
2657 gmx::StringOutputStream stream;
2658 gmx::TextWriter log(&stream);
2659 writeSettings(&log, dd, mtop, ir, isDlbOn(dd->comm), dlb_scale, ddbox);
2660 if (dd->comm->dlbState == DlbState::offCanTurnOn)
2663 log.ensureEmptyLine();
2665 "When dynamic load balancing gets turned on, these settings will change to:");
2667 writeSettings(&log, dd, mtop, ir, true, dlb_scale, ddbox);
2669 GMX_LOG(mdlog.info).asParagraph().appendText(stream.toString());
2672 static void set_cell_limits_dlb(const gmx::MDLogger& mdlog,
2675 const t_inputrec* ir,
2676 const gmx_ddbox_t* ddbox)
2678 gmx_domdec_comm_t* comm;
2679 int d, dim, npulse, npulse_d_max, npulse_d;
2684 bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
2686 /* Determine the maximum number of comm. pulses in one dimension */
2688 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2690 /* Determine the maximum required number of grid pulses */
2691 if (comm->cellsize_limit >= comm->systemInfo.cutoff)
2693 /* Only a single pulse is required */
2696 else if (!bNoCutOff && comm->cellsize_limit > 0)
2698 /* We round down slightly here to avoid overhead due to the latency
2699 * of extra communication calls when the cut-off
2700 * would be only slightly longer than the cell size.
2701 * Later cellsize_limit is redetermined,
2702 * so we can not miss interactions due to this rounding.
2704 npulse = static_cast<int>(0.96 + comm->systemInfo.cutoff / comm->cellsize_limit);
2708 /* There is no cell size limit */
2709 npulse = std::max(dd->numCells[XX] - 1, std::max(dd->numCells[YY] - 1, dd->numCells[ZZ] - 1));
2712 if (!bNoCutOff && npulse > 1)
2714 /* See if we can do with less pulses, based on dlb_scale */
2716 for (d = 0; d < dd->ndim; d++)
2719 npulse_d = static_cast<int>(
2721 + dd->numCells[dim] * comm->systemInfo.cutoff
2722 / (ddbox->box_size[dim] * ddbox->skew_fac[dim] * dlb_scale));
2723 npulse_d_max = std::max(npulse_d_max, npulse_d);
2725 npulse = std::min(npulse, npulse_d_max);
2728 /* This env var can override npulse */
2729 d = dd_getenv(mdlog, "GMX_DD_NPULSE", 0);
2736 comm->bVacDLBNoLimit = (ir->pbcType == PbcType::No);
2737 for (d = 0; d < dd->ndim; d++)
2739 comm->cd[d].np_dlb = std::min(npulse, dd->numCells[dd->dim[d]] - 1);
2740 comm->maxpulse = std::max(comm->maxpulse, comm->cd[d].np_dlb);
2741 if (comm->cd[d].np_dlb < dd->numCells[dd->dim[d]] - 1)
2743 comm->bVacDLBNoLimit = FALSE;
2747 /* cellsize_limit is set for LINCS in init_domain_decomposition */
2748 if (!comm->bVacDLBNoLimit)
2750 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->systemInfo.cutoff / comm->maxpulse);
2752 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2753 /* Set the minimum cell size for each DD dimension */
2754 for (d = 0; d < dd->ndim; d++)
2756 if (comm->bVacDLBNoLimit || comm->cd[d].np_dlb * comm->cellsize_limit >= comm->systemInfo.cutoff)
2758 comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
2762 comm->cellsize_min_dlb[dd->dim[d]] = comm->systemInfo.cutoff / comm->cd[d].np_dlb;
2765 if (comm->cutoff_mbody <= 0)
2767 comm->cutoff_mbody = std::min(comm->systemInfo.cutoff, comm->cellsize_limit);
2775 bool dd_moleculesAreAlwaysWhole(const gmx_domdec_t& dd)
2777 return dd.comm->systemInfo.moleculesAreAlwaysWhole;
2780 gmx_bool dd_bonded_molpbc(const gmx_domdec_t* dd, PbcType pbcType)
2782 /* If each molecule is a single charge group
2783 * or we use domain decomposition for each periodic dimension,
2784 * we do not need to take pbc into account for the bonded interactions.
2786 return (pbcType != PbcType::No && dd->comm->systemInfo.haveInterDomainBondeds
2787 && !(dd->numCells[XX] > 1 && dd->numCells[YY] > 1
2788 && (dd->numCells[ZZ] > 1 || pbcType == PbcType::XY)));
2791 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
2792 static void set_ddgrid_parameters(const gmx::MDLogger& mdlog,
2795 const gmx_mtop_t* mtop,
2796 const t_inputrec* ir,
2797 const gmx_ddbox_t* ddbox)
2799 gmx_domdec_comm_t* comm = dd->comm;
2800 DDRankSetup& ddRankSetup = comm->ddRankSetup;
2802 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
2804 init_ddpme(dd, &ddRankSetup.ddpme[0], 0);
2805 if (ddRankSetup.npmedecompdim >= 2)
2807 init_ddpme(dd, &ddRankSetup.ddpme[1], 1);
2812 ddRankSetup.numRanksDoingPme = 0;
2813 if (dd->pme_nodeid >= 0)
2815 gmx_fatal_collective(FARGS, dd->mpi_comm_all, DDMASTER(dd),
2816 "Can not have separate PME ranks without PME electrostatics");
2822 fprintf(debug, "The DD cut-off is %f\n", comm->systemInfo.cutoff);
2824 if (!isDlbDisabled(comm))
2826 set_cell_limits_dlb(mdlog, dd, dlb_scale, ir, ddbox);
2829 logSettings(mdlog, dd, mtop, ir, dlb_scale, ddbox);
2832 if (ir->pbcType == PbcType::No)
2834 vol_frac = 1 - 1 / static_cast<double>(dd->nnodes);
2838 vol_frac = (1 + comm_box_frac(dd->numCells, comm->systemInfo.cutoff, *ddbox))
2839 / static_cast<double>(dd->nnodes);
2843 fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
2845 int natoms_tot = mtop->natoms;
2847 dd->ga2la = new gmx_ga2la_t(natoms_tot, static_cast<int>(vol_frac * natoms_tot));
2850 /*! \brief Get some important DD parameters which can be modified by env.vars */
2851 static DDSettings getDDSettings(const gmx::MDLogger& mdlog,
2852 const DomdecOptions& options,
2853 const gmx::MdrunOptions& mdrunOptions,
2854 const t_inputrec& ir)
2856 DDSettings ddSettings;
2858 ddSettings.useSendRecv2 = (dd_getenv(mdlog, "GMX_DD_USE_SENDRECV2", 0) != 0);
2859 ddSettings.dlb_scale_lim = dd_getenv(mdlog, "GMX_DLB_MAX_BOX_SCALING", 10);
2860 ddSettings.request1D = bool(dd_getenv(mdlog, "GMX_DD_1D", 0));
2861 ddSettings.useDDOrderZYX = bool(dd_getenv(mdlog, "GMX_DD_ORDER_ZYX", 0));
2862 ddSettings.useCartesianReorder = bool(dd_getenv(mdlog, "GMX_NO_CART_REORDER", 1));
2863 ddSettings.eFlop = dd_getenv(mdlog, "GMX_DLB_BASED_ON_FLOPS", 0);
2864 const int recload = dd_getenv(mdlog, "GMX_DD_RECORD_LOAD", 1);
2865 ddSettings.nstDDDump = dd_getenv(mdlog, "GMX_DD_NST_DUMP", 0);
2866 ddSettings.nstDDDumpGrid = dd_getenv(mdlog, "GMX_DD_NST_DUMP_GRID", 0);
2867 ddSettings.DD_debug = dd_getenv(mdlog, "GMX_DD_DEBUG", 0);
2869 if (ddSettings.useSendRecv2)
2873 "Will use two sequential MPI_Sendrecv calls instead of two simultaneous "
2874 "non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite "
2878 if (ddSettings.eFlop)
2880 GMX_LOG(mdlog.info).appendText("Will load balance based on FLOP count");
2881 ddSettings.recordLoad = true;
2885 ddSettings.recordLoad = (wallcycle_have_counter() && recload > 0);
2888 ddSettings.initialDlbState = determineInitialDlbState(mdlog, options.dlbOption,
2889 ddSettings.recordLoad, mdrunOptions, &ir);
2891 .appendTextFormatted("Dynamic load balancing: %s",
2892 edlbs_names[static_cast<int>(ddSettings.initialDlbState)]);
2897 gmx_domdec_t::gmx_domdec_t(const t_inputrec& ir) : unitCellInfo(ir) {}
2899 /*! \brief Return whether the simulation described can run a 1D DD.
2901 * The GPU halo exchange code requires 1D DD. Such a DD
2902 * generally requires a larger box than other possible decompositions
2903 * with the same rank count, so the calling code might need to decide
2904 * what is the most appropriate way to run the simulation based on
2905 * whether such a DD is possible.
2907 * This function works like init_domain_decomposition(), but will not
2908 * give a fatal error, and only uses \c cr for communicating between
2911 * It is safe to call before thread-MPI spawns ranks, so that
2912 * thread-MPI can decide whether and how to trigger the GPU halo
2913 * exchange code path. The number of PME ranks, if any, should be set
2914 * in \c options.numPmeRanks.
2916 static bool canMake1DDomainDecomposition(const DDSettings& ddSettingsOriginal,
2917 const t_commrec* cr,
2918 const int numRanksRequested,
2919 const DomdecOptions& options,
2920 const gmx_mtop_t& mtop,
2921 const t_inputrec& ir,
2923 gmx::ArrayRef<const gmx::RVec> xGlobal)
2925 // Ensure we don't write any output from this checking routine
2926 gmx::MDLogger dummyLogger;
2928 DDSystemInfo systemInfo = getSystemInfo(dummyLogger, cr, options, mtop, ir, box, xGlobal);
2930 DDSettings ddSettings = ddSettingsOriginal;
2931 ddSettings.request1D = true;
2932 const real gridSetupCellsizeLimit =
2933 getDDGridSetupCellSizeLimit(dummyLogger, !isDlbDisabled(ddSettings.initialDlbState),
2934 options.dlbScaling, ir, systemInfo.cellsizeLimit);
2935 gmx_ddbox_t ddbox = { 0 };
2936 DDGridSetup ddGridSetup =
2937 getDDGridSetup(dummyLogger, cr, numRanksRequested, options, ddSettings, systemInfo,
2938 gridSetupCellsizeLimit, mtop, ir, box, xGlobal, &ddbox);
2940 const bool canMake1DDD = (ddGridSetup.numDomains[XX] != 0);
2945 bool is1D(const gmx_domdec_t& dd)
2947 const int maxDimensionSize = std::max(dd.numCells[XX], std::max(dd.numCells[YY], dd.numCells[ZZ]));
2948 const int productOfDimensionSizes = dd.numCells[XX] * dd.numCells[YY] * dd.numCells[ZZ];
2949 const bool decompositionHasOneDimension = (maxDimensionSize == productOfDimensionSizes);
2951 return decompositionHasOneDimension;
2957 // TODO once the functionality stablizes, move this class and
2958 // supporting functionality into builder.cpp
2959 /*! \brief Impl class for DD builder */
2960 class DomainDecompositionBuilder::Impl
2964 Impl(const MDLogger& mdlog,
2966 const DomdecOptions& options,
2967 const MdrunOptions& mdrunOptions,
2969 const gmx_mtop_t& mtop,
2970 const t_inputrec& ir,
2972 ArrayRef<const RVec> xGlobal);
2974 //! Build the resulting DD manager
2975 gmx_domdec_t* build(LocalAtomSetManager* atomSets);
2977 //! Objects used in constructing and configuring DD
2980 const MDLogger& mdlog_;
2981 //! Communication object
2983 //! User-supplied options configuring DD behavior
2984 const DomdecOptions options_;
2985 //! Global system topology
2986 const gmx_mtop_t& mtop_;
2987 //! User input values from the tpr file
2988 const t_inputrec& ir_;
2991 //! Internal objects used in constructing DD
2993 //! Settings combined from the user input
2994 DDSettings ddSettings_;
2995 //! Information derived from the simulation system
2996 DDSystemInfo systemInfo_;
2998 gmx_ddbox_t ddbox_ = { 0 };
2999 //! Organization of the DD grids
3000 DDGridSetup ddGridSetup_;
3001 //! Organzation of the DD ranks
3002 DDRankSetup ddRankSetup_;
3003 //! Number of DD cells in each dimension
3004 ivec ddCellIndex_ = { 0, 0, 0 };
3005 //! IDs of PME-only ranks
3006 std::vector<int> pmeRanks_;
3007 //! Contains a valid Cartesian-communicator-based setup, or defaults.
3008 CartesianRankSetup cartSetup_;
3012 DomainDecompositionBuilder::Impl::Impl(const MDLogger& mdlog,
3014 const DomdecOptions& options,
3015 const MdrunOptions& mdrunOptions,
3016 const bool prefer1D,
3017 const gmx_mtop_t& mtop,
3018 const t_inputrec& ir,
3020 ArrayRef<const RVec> xGlobal) :
3027 GMX_LOG(mdlog_.info).appendTextFormatted("\nInitializing Domain Decomposition on %d ranks", cr_->nnodes);
3029 ddSettings_ = getDDSettings(mdlog_, options_, mdrunOptions, ir_);
3032 && canMake1DDomainDecomposition(ddSettings_, cr_, cr_->nnodes, options_, mtop_, ir_, box, xGlobal))
3034 ddSettings_.request1D = true;
3037 if (ddSettings_.eFlop > 1)
3039 /* Ensure that we have different random flop counts on different ranks */
3040 srand(1 + cr_->nodeid);
3043 systemInfo_ = getSystemInfo(mdlog_, cr_, options_, mtop_, ir_, box, xGlobal);
3045 const int numRanksRequested = cr_->nnodes;
3046 const bool checkForLargePrimeFactors = (options_.numCells[0] <= 0);
3047 checkForValidRankCountRequests(numRanksRequested, EEL_PME(ir_.coulombtype),
3048 options_.numPmeRanks, checkForLargePrimeFactors);
3050 // DD grid setup uses a more different cell size limit for
3051 // automated setup than the one in systemInfo_. The latter is used
3052 // in set_dd_limits() to configure DLB, for example.
3053 const real gridSetupCellsizeLimit =
3054 getDDGridSetupCellSizeLimit(mdlog_, !isDlbDisabled(ddSettings_.initialDlbState),
3055 options_.dlbScaling, ir_, systemInfo_.cellsizeLimit);
3056 ddGridSetup_ = getDDGridSetup(mdlog_, cr_, numRanksRequested, options_, ddSettings_, systemInfo_,
3057 gridSetupCellsizeLimit, mtop_, ir_, box, xGlobal, &ddbox_);
3058 checkDDGridSetup(ddGridSetup_, cr_, options_, ddSettings_, systemInfo_, gridSetupCellsizeLimit, ddbox_);
3060 cr_->npmenodes = ddGridSetup_.numPmeOnlyRanks;
3062 ddRankSetup_ = getDDRankSetup(mdlog_, cr_, ddGridSetup_, ir_);
3064 /* Generate the group communicator, also decides the duty of each rank */
3065 cartSetup_ = makeGroupCommunicators(mdlog_, ddSettings_, options_.rankOrder, ddRankSetup_, cr_,
3066 ddCellIndex_, &pmeRanks_);
3069 gmx_domdec_t* DomainDecompositionBuilder::Impl::build(LocalAtomSetManager* atomSets)
3071 gmx_domdec_t* dd = new gmx_domdec_t(ir_);
3073 copy_ivec(ddCellIndex_, dd->ci);
3075 dd->comm = init_dd_comm();
3077 dd->comm->ddRankSetup = ddRankSetup_;
3078 dd->comm->cartesianRankSetup = cartSetup_;
3080 set_dd_limits(mdlog_, cr_, dd, options_, ddSettings_, systemInfo_, ddGridSetup_,
3081 ddRankSetup_.numPPRanks, &mtop_, &ir_, ddbox_);
3083 setupGroupCommunication(mdlog_, ddSettings_, pmeRanks_, cr_, mtop_.natoms, dd);
3085 if (thisRankHasDuty(cr_, DUTY_PP))
3087 set_ddgrid_parameters(mdlog_, dd, options_.dlbScaling, &mtop_, &ir_, &ddbox_);
3089 setup_neighbor_relations(dd);
3092 /* Set overallocation to avoid frequent reallocation of arrays */
3093 set_over_alloc_dd(TRUE);
3095 dd->atomSets = atomSets;
3100 DomainDecompositionBuilder::DomainDecompositionBuilder(const MDLogger& mdlog,
3102 const DomdecOptions& options,
3103 const MdrunOptions& mdrunOptions,
3104 const bool prefer1D,
3105 const gmx_mtop_t& mtop,
3106 const t_inputrec& ir,
3108 ArrayRef<const RVec> xGlobal) :
3109 impl_(new Impl(mdlog, cr, options, mdrunOptions, prefer1D, mtop, ir, box, xGlobal))
3113 gmx_domdec_t* DomainDecompositionBuilder::build(LocalAtomSetManager* atomSets)
3115 return impl_->build(atomSets);
3118 DomainDecompositionBuilder::~DomainDecompositionBuilder() = default;
3122 static gmx_bool test_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3132 set_ddbox(*dd, false, box, true, x, &ddbox);
3136 for (d = 0; d < dd->ndim; d++)
3140 inv_cell_size = DD_CELL_MARGIN * dd->numCells[dim] / ddbox.box_size[dim];
3141 if (dd->unitCellInfo.ddBoxIsDynamic)
3143 inv_cell_size *= DD_PRES_SCALE_MARGIN;
3146 np = 1 + static_cast<int>(cutoffRequested * inv_cell_size * ddbox.skew_fac[dim]);
3148 if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
3150 if (np > dd->comm->cd[d].np_dlb)
3155 /* If a current local cell size is smaller than the requested
3156 * cut-off, we could still fix it, but this gets very complicated.
3157 * Without fixing here, we might actually need more checks.
3159 real cellSizeAlongDim =
3160 (dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim]) * ddbox.skew_fac[dim];
3161 if (cellSizeAlongDim * dd->comm->cd[d].np_dlb < cutoffRequested)
3168 if (!isDlbDisabled(dd->comm))
3170 /* If DLB is not active yet, we don't need to check the grid jumps.
3171 * Actually we shouldn't, because then the grid jump data is not set.
3173 if (isDlbOn(dd->comm) && check_grid_jump(0, dd, cutoffRequested, &ddbox, FALSE))
3178 gmx_sumi(1, &LocallyLimited, cr);
3180 if (LocallyLimited > 0)
3189 gmx_bool change_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3191 gmx_bool bCutoffAllowed;
3193 bCutoffAllowed = test_dd_cutoff(cr, box, x, cutoffRequested);
3197 cr->dd->comm->systemInfo.cutoff = cutoffRequested;
3200 return bCutoffAllowed;
3203 void constructGpuHaloExchange(const gmx::MDLogger& mdlog,
3204 const t_commrec& cr,
3205 const DeviceContext& deviceContext,
3206 const DeviceStream& streamLocal,
3207 const DeviceStream& streamNonLocal)
3210 int gpuHaloExchangeSize = 0;
3212 if (cr.dd->gpuHaloExchange.empty())
3214 GMX_LOG(mdlog.warning)
3216 .appendTextFormatted(
3217 "NOTE: Activating the 'GPU halo exchange' feature, enabled "
3219 "GMX_GPU_DD_COMMS environment variable.");
3223 gpuHaloExchangeSize = static_cast<int>(cr.dd->gpuHaloExchange.size());
3224 pulseStart = gpuHaloExchangeSize - 1;
3226 if (cr.dd->comm->cd[0].numPulses() > gpuHaloExchangeSize)
3228 for (int pulse = pulseStart; pulse < cr.dd->comm->cd[0].numPulses(); pulse++)
3230 cr.dd->gpuHaloExchange.push_back(std::make_unique<gmx::GpuHaloExchange>(
3231 cr.dd, cr.mpi_comm_mysim, deviceContext, streamLocal, streamNonLocal, pulse));
3236 void reinitGpuHaloExchange(const t_commrec& cr,
3237 const DeviceBuffer<gmx::RVec> d_coordinatesBuffer,
3238 const DeviceBuffer<gmx::RVec> d_forcesBuffer)
3240 for (int pulse = 0; pulse < cr.dd->comm->cd[0].numPulses(); pulse++)
3242 cr.dd->gpuHaloExchange[pulse]->reinitHalo(d_coordinatesBuffer, d_forcesBuffer);
3246 void communicateGpuHaloCoordinates(const t_commrec& cr,
3248 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
3250 for (int pulse = 0; pulse < cr.dd->comm->cd[0].numPulses(); pulse++)
3252 cr.dd->gpuHaloExchange[pulse]->communicateHaloCoordinates(box, coordinatesReadyOnDeviceEvent);
3256 void communicateGpuHaloForces(const t_commrec& cr, bool accumulateForces)
3258 for (int pulse = cr.dd->comm->cd[0].numPulses() - 1; pulse >= 0; pulse--)
3260 cr.dd->gpuHaloExchange[pulse]->communicateHaloForces(accumulateForces);