2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2005,2006,2007,2008,2009,2010,2011,2012,2013,2014,2015,2016,2017,2018, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
52 #include "gromacs/compat/make_unique.h"
53 #include "gromacs/domdec/collect.h"
54 #include "gromacs/domdec/dlbtiming.h"
55 #include "gromacs/domdec/domdec_network.h"
56 #include "gromacs/domdec/ga2la.h"
57 #include "gromacs/domdec/localatomsetmanager.h"
58 #include "gromacs/ewald/pme.h"
59 #include "gromacs/fileio/gmxfio.h"
60 #include "gromacs/fileio/pdbio.h"
61 #include "gromacs/gmxlib/chargegroup.h"
62 #include "gromacs/gmxlib/network.h"
63 #include "gromacs/gmxlib/nrnb.h"
64 #include "gromacs/gpu_utils/gpu_utils.h"
65 #include "gromacs/hardware/hw_info.h"
66 #include "gromacs/imd/imd.h"
67 #include "gromacs/listed-forces/manage-threading.h"
68 #include "gromacs/math/functions.h"
69 #include "gromacs/math/vec.h"
70 #include "gromacs/math/vectypes.h"
71 #include "gromacs/mdlib/constr.h"
72 #include "gromacs/mdlib/constraintrange.h"
73 #include "gromacs/mdlib/forcerec.h"
74 #include "gromacs/mdlib/gmx_omp_nthreads.h"
75 #include "gromacs/mdlib/lincs.h"
76 #include "gromacs/mdlib/mdatoms.h"
77 #include "gromacs/mdlib/mdrun.h"
78 #include "gromacs/mdlib/mdsetup.h"
79 #include "gromacs/mdlib/nb_verlet.h"
80 #include "gromacs/mdlib/nbnxn_grid.h"
81 #include "gromacs/mdlib/nsgrid.h"
82 #include "gromacs/mdlib/vsite.h"
83 #include "gromacs/mdtypes/commrec.h"
84 #include "gromacs/mdtypes/df_history.h"
85 #include "gromacs/mdtypes/forcerec.h"
86 #include "gromacs/mdtypes/inputrec.h"
87 #include "gromacs/mdtypes/md_enums.h"
88 #include "gromacs/mdtypes/mdatom.h"
89 #include "gromacs/mdtypes/nblist.h"
90 #include "gromacs/mdtypes/state.h"
91 #include "gromacs/pbcutil/ishift.h"
92 #include "gromacs/pbcutil/pbc.h"
93 #include "gromacs/pulling/pull.h"
94 #include "gromacs/timing/wallcycle.h"
95 #include "gromacs/topology/block.h"
96 #include "gromacs/topology/idef.h"
97 #include "gromacs/topology/ifunc.h"
98 #include "gromacs/topology/mtop_lookup.h"
99 #include "gromacs/topology/mtop_util.h"
100 #include "gromacs/topology/topology.h"
101 #include "gromacs/utility/basedefinitions.h"
102 #include "gromacs/utility/basenetwork.h"
103 #include "gromacs/utility/cstringutil.h"
104 #include "gromacs/utility/exceptions.h"
105 #include "gromacs/utility/fatalerror.h"
106 #include "gromacs/utility/gmxmpi.h"
107 #include "gromacs/utility/qsort_threadsafe.h"
108 #include "gromacs/utility/real.h"
109 #include "gromacs/utility/smalloc.h"
110 #include "gromacs/utility/strconvert.h"
111 #include "gromacs/utility/stringutil.h"
113 #include "atomdistribution.h"
114 #include "cellsizes.h"
115 #include "distribute.h"
116 #include "domdec_constraints.h"
117 #include "domdec_internal.h"
118 #include "domdec_vsite.h"
119 #include "redistribute.h"
122 #define DD_NLOAD_MAX 9
124 static const char *edlbs_names[edlbsNR] = { "off", "auto", "locked", "on", "on" };
126 /* The size per charge group of the cggl_flag buffer in gmx_domdec_comm_t */
129 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
130 #define DD_FLAG_NRCG 65535
131 #define DD_FLAG_FW(d) (1<<(16+(d)*2))
132 #define DD_FLAG_BW(d) (1<<(16+(d)*2+1))
134 /* The DD zone order */
135 static const ivec dd_zo[DD_MAXZONE] =
136 {{0, 0, 0}, {1, 0, 0}, {1, 1, 0}, {0, 1, 0}, {0, 1, 1}, {0, 0, 1}, {1, 0, 1}, {1, 1, 1}};
138 /* The non-bonded zone-pair setup for domain decomposition
139 * The first number is the i-zone, the second number the first j-zone seen by
140 * this i-zone, the third number the last+1 j-zone seen by this i-zone.
141 * As is, this is for 3D decomposition, where there are 4 i-zones.
142 * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
143 * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
146 ddNonbondedZonePairRanges[DD_MAXIZONE][3] = {{0, 0, 8},
151 /* Turn on DLB when the load imbalance causes this amount of total loss.
152 * There is a bit of overhead with DLB and it's difficult to achieve
153 * a load imbalance of less than 2% with DLB.
155 #define DD_PERF_LOSS_DLB_ON 0.02
157 /* Warn about imbalance due to PP or PP/PME load imbalance at this loss */
158 #define DD_PERF_LOSS_WARN 0.05
161 /* We check if to turn on DLB at the first and every 100 DD partitionings.
162 * With large imbalance DLB will turn on at the first step, so we can
163 * make the interval so large that the MPI overhead of the check is negligible.
165 static const int c_checkTurnDlbOnInterval = 100;
166 /* We need to check if DLB results in worse performance and then turn it off.
167 * We check this more often then for turning DLB on, because the DLB can scale
168 * the domains very rapidly, so if unlucky the load imbalance can go up quickly
169 * and furthermore, we are already synchronizing often with DLB, so
170 * the overhead of the MPI Bcast is not that high.
172 static const int c_checkTurnDlbOffInterval = 20;
174 /* Forward declaration */
175 static void dd_dlb_set_should_check_whether_to_turn_dlb_on(gmx_domdec_t *dd, gmx_bool bValue);
179 #define dd_index(n,i) ((((i)[ZZ]*(n)[YY] + (i)[YY])*(n)[XX]) + (i)[XX])
181 static void index2xyz(ivec nc,int ind,ivec xyz)
183 xyz[XX] = ind % nc[XX];
184 xyz[YY] = (ind / nc[XX]) % nc[YY];
185 xyz[ZZ] = ind / (nc[YY]*nc[XX]);
189 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
191 xyz[XX] = ind / (nc[YY]*nc[ZZ]);
192 xyz[YY] = (ind / nc[ZZ]) % nc[YY];
193 xyz[ZZ] = ind % nc[ZZ];
196 static int ddcoord2ddnodeid(gmx_domdec_t *dd, ivec c)
201 ddindex = dd_index(dd->nc, c);
202 if (dd->comm->bCartesianPP_PME)
204 ddnodeid = dd->comm->ddindex2ddnodeid[ddindex];
206 else if (dd->comm->bCartesianPP)
209 MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
220 static gmx_bool dynamic_dd_box(const gmx_ddbox_t *ddbox, const t_inputrec *ir)
222 return (ddbox->nboundeddim < DIM || inputrecDynamicBox(ir));
225 int ddglatnr(const gmx_domdec_t *dd, int i)
235 if (i >= dd->comm->atomRanges.numAtomsTotal())
237 gmx_fatal(FARGS, "glatnr called with %d, which is larger than the local number of atoms (%d)", i, dd->comm->atomRanges.numAtomsTotal());
239 atnr = dd->globalAtomIndices[i] + 1;
245 t_block *dd_charge_groups_global(gmx_domdec_t *dd)
247 return &dd->comm->cgs_gl;
250 void dd_store_state(gmx_domdec_t *dd, t_state *state)
254 if (state->ddp_count != dd->ddp_count)
256 gmx_incons("The MD state does not match the domain decomposition state");
259 state->cg_gl.resize(dd->ncg_home);
260 for (i = 0; i < dd->ncg_home; i++)
262 state->cg_gl[i] = dd->globalAtomGroupIndices[i];
265 state->ddp_count_cg_gl = dd->ddp_count;
268 gmx_domdec_zones_t *domdec_zones(gmx_domdec_t *dd)
270 return &dd->comm->zones;
273 void dd_get_ns_ranges(const gmx_domdec_t *dd, int icg,
274 int *jcg0, int *jcg1, ivec shift0, ivec shift1)
276 gmx_domdec_zones_t *zones;
279 zones = &dd->comm->zones;
282 while (icg >= zones->izone[izone].cg1)
291 else if (izone < zones->nizone)
293 *jcg0 = zones->izone[izone].jcg0;
297 gmx_fatal(FARGS, "DD icg %d out of range: izone (%d) >= nizone (%d)",
298 icg, izone, zones->nizone);
301 *jcg1 = zones->izone[izone].jcg1;
303 for (d = 0; d < dd->ndim; d++)
306 shift0[dim] = zones->izone[izone].shift0[dim];
307 shift1[dim] = zones->izone[izone].shift1[dim];
308 if (dd->comm->tric_dir[dim] || (isDlbOn(dd->comm) && d > 0))
310 /* A conservative approach, this can be optimized */
317 int dd_numHomeAtoms(const gmx_domdec_t &dd)
319 return dd.comm->atomRanges.numHomeAtoms();
322 int dd_natoms_mdatoms(const gmx_domdec_t *dd)
324 /* We currently set mdatoms entries for all atoms:
325 * local + non-local + communicated for vsite + constraints
328 return dd->comm->atomRanges.numAtomsTotal();
331 int dd_natoms_vsite(const gmx_domdec_t *dd)
333 return dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
336 void dd_get_constraint_range(const gmx_domdec_t *dd, int *at_start, int *at_end)
338 *at_start = dd->comm->atomRanges.start(DDAtomRanges::Type::Constraints);
339 *at_end = dd->comm->atomRanges.end(DDAtomRanges::Type::Constraints);
342 void dd_move_x(gmx_domdec_t *dd,
344 gmx::ArrayRef<gmx::RVec> x,
345 gmx_wallcycle *wcycle)
347 wallcycle_start(wcycle, ewcMOVEX);
350 gmx_domdec_comm_t *comm;
351 gmx_domdec_comm_dim_t *cd;
352 rvec shift = {0, 0, 0};
353 gmx_bool bPBC, bScrew;
357 const gmx::RangePartitioning &atomGrouping = dd->atomGrouping();
360 nat_tot = comm->atomRanges.numHomeAtoms();
361 for (int d = 0; d < dd->ndim; d++)
363 bPBC = (dd->ci[dd->dim[d]] == 0);
364 bScrew = (bPBC && dd->bScrewPBC && dd->dim[d] == XX);
367 copy_rvec(box[dd->dim[d]], shift);
370 for (const gmx_domdec_ind_t &ind : cd->ind)
372 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
373 gmx::ArrayRef<gmx::RVec> &sendBuffer = sendBufferAccess.buffer;
377 for (int g : ind.index)
379 for (int j : atomGrouping.block(g))
381 sendBuffer[n] = x[j];
388 for (int g : ind.index)
390 for (int j : atomGrouping.block(g))
392 /* We need to shift the coordinates */
393 for (int d = 0; d < DIM; d++)
395 sendBuffer[n][d] = x[j][d] + shift[d];
403 for (int g : ind.index)
405 for (int j : atomGrouping.block(g))
408 sendBuffer[n][XX] = x[j][XX] + shift[XX];
410 * This operation requires a special shift force
411 * treatment, which is performed in calc_vir.
413 sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
414 sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
420 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
422 gmx::ArrayRef<gmx::RVec> receiveBuffer;
423 if (cd->receiveInPlace)
425 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
429 receiveBuffer = receiveBufferAccess.buffer;
431 /* Send and receive the coordinates */
432 ddSendrecv(dd, d, dddirBackward,
433 sendBuffer, receiveBuffer);
435 if (!cd->receiveInPlace)
438 for (int zone = 0; zone < nzone; zone++)
440 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
442 x[i] = receiveBuffer[j++];
446 nat_tot += ind.nrecv[nzone+1];
451 wallcycle_stop(wcycle, ewcMOVEX);
454 void dd_move_f(gmx_domdec_t *dd,
455 gmx::ArrayRef<gmx::RVec> f,
457 gmx_wallcycle *wcycle)
459 wallcycle_start(wcycle, ewcMOVEF);
462 gmx_domdec_comm_t *comm;
463 gmx_domdec_comm_dim_t *cd;
466 gmx_bool bShiftForcesNeedPbc, bScrew;
470 const gmx::RangePartitioning &atomGrouping = dd->atomGrouping();
472 nzone = comm->zones.n/2;
473 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
474 for (int d = dd->ndim-1; d >= 0; d--)
476 /* Only forces in domains near the PBC boundaries need to
477 consider PBC in the treatment of fshift */
478 bShiftForcesNeedPbc = (dd->ci[dd->dim[d]] == 0);
479 bScrew = (bShiftForcesNeedPbc && dd->bScrewPBC && dd->dim[d] == XX);
480 if (fshift == nullptr && !bScrew)
482 bShiftForcesNeedPbc = FALSE;
484 /* Determine which shift vector we need */
490 for (int p = cd->numPulses() - 1; p >= 0; p--)
492 const gmx_domdec_ind_t &ind = cd->ind[p];
493 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
494 gmx::ArrayRef<gmx::RVec> &receiveBuffer = receiveBufferAccess.buffer;
496 nat_tot -= ind.nrecv[nzone+1];
498 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
500 gmx::ArrayRef<gmx::RVec> sendBuffer;
501 if (cd->receiveInPlace)
503 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
507 sendBuffer = sendBufferAccess.buffer;
509 for (int zone = 0; zone < nzone; zone++)
511 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
513 sendBuffer[j++] = f[i];
517 /* Communicate the forces */
518 ddSendrecv(dd, d, dddirForward,
519 sendBuffer, receiveBuffer);
520 /* Add the received forces */
522 if (!bShiftForcesNeedPbc)
524 for (int g : ind.index)
526 for (int j : atomGrouping.block(g))
528 for (int d = 0; d < DIM; d++)
530 f[j][d] += receiveBuffer[n][d];
538 /* fshift should always be defined if this function is
539 * called when bShiftForcesNeedPbc is true */
540 assert(nullptr != fshift);
541 for (int g : ind.index)
543 for (int j : atomGrouping.block(g))
545 for (int d = 0; d < DIM; d++)
547 f[j][d] += receiveBuffer[n][d];
549 /* Add this force to the shift force */
550 for (int d = 0; d < DIM; d++)
552 fshift[is][d] += receiveBuffer[n][d];
560 for (int g : ind.index)
562 for (int j : atomGrouping.block(g))
564 /* Rotate the force */
565 f[j][XX] += receiveBuffer[n][XX];
566 f[j][YY] -= receiveBuffer[n][YY];
567 f[j][ZZ] -= receiveBuffer[n][ZZ];
570 /* Add this force to the shift force */
571 for (int d = 0; d < DIM; d++)
573 fshift[is][d] += receiveBuffer[n][d];
583 wallcycle_stop(wcycle, ewcMOVEF);
586 /* Convenience function for extracting a real buffer from an rvec buffer
588 * To reduce the number of temporary communication buffers and avoid
589 * cache polution, we reuse gmx::RVec buffers for storing reals.
590 * This functions return a real buffer reference with the same number
591 * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
593 static gmx::ArrayRef<real>
594 realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
596 return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0],
600 void dd_atom_spread_real(gmx_domdec_t *dd, real v[])
603 gmx_domdec_comm_t *comm;
604 gmx_domdec_comm_dim_t *cd;
608 const gmx::RangePartitioning &atomGrouping = dd->atomGrouping();
611 nat_tot = comm->atomRanges.numHomeAtoms();
612 for (int d = 0; d < dd->ndim; d++)
615 for (const gmx_domdec_ind_t &ind : cd->ind)
617 /* Note: We provision for RVec instead of real, so a factor of 3
618 * more than needed. The buffer actually already has this size
619 * and we pass a plain pointer below, so this does not matter.
621 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
622 gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
624 for (int g : ind.index)
626 for (int j : atomGrouping.block(g))
628 sendBuffer[n++] = v[j];
632 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
634 gmx::ArrayRef<real> receiveBuffer;
635 if (cd->receiveInPlace)
637 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
641 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
643 /* Send and receive the data */
644 ddSendrecv(dd, d, dddirBackward,
645 sendBuffer, receiveBuffer);
646 if (!cd->receiveInPlace)
649 for (int zone = 0; zone < nzone; zone++)
651 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
653 v[i] = receiveBuffer[j++];
657 nat_tot += ind.nrecv[nzone+1];
663 void dd_atom_sum_real(gmx_domdec_t *dd, real v[])
666 gmx_domdec_comm_t *comm;
667 gmx_domdec_comm_dim_t *cd;
671 const gmx::RangePartitioning &atomGrouping = dd->atomGrouping();
673 nzone = comm->zones.n/2;
674 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
675 for (int d = dd->ndim-1; d >= 0; d--)
678 for (int p = cd->numPulses() - 1; p >= 0; p--)
680 const gmx_domdec_ind_t &ind = cd->ind[p];
682 /* Note: We provision for RVec instead of real, so a factor of 3
683 * more than needed. The buffer actually already has this size
684 * and we typecast, so this works as intended.
686 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
687 gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
688 nat_tot -= ind.nrecv[nzone + 1];
690 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
692 gmx::ArrayRef<real> sendBuffer;
693 if (cd->receiveInPlace)
695 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
699 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
701 for (int zone = 0; zone < nzone; zone++)
703 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
705 sendBuffer[j++] = v[i];
709 /* Communicate the forces */
710 ddSendrecv(dd, d, dddirForward,
711 sendBuffer, receiveBuffer);
712 /* Add the received forces */
714 for (int g : ind.index)
716 for (int j : atomGrouping.block(g))
718 v[j] += receiveBuffer[n];
727 static void print_ddzone(FILE *fp, int d, int i, int j, gmx_ddzone_t *zone)
729 fprintf(fp, "zone d0 %d d1 %d d2 %d min0 %6.3f max1 %6.3f mch0 %6.3f mch1 %6.3f p1_0 %6.3f p1_1 %6.3f\n",
731 zone->min0, zone->max1,
732 zone->mch0, zone->mch0,
733 zone->p1_0, zone->p1_1);
736 /* Using the home grid size as input in cell_ns_x0 and cell_ns_x1
737 * takes the extremes over all home and remote zones in the halo
738 * and returns the results in cell_ns_x0 and cell_ns_x1.
739 * Note: only used with the group cut-off scheme.
741 static void dd_move_cellx(gmx_domdec_t *dd,
742 const gmx_ddbox_t *ddbox,
746 constexpr int c_ddZoneCommMaxNumZones = 5;
747 gmx_ddzone_t buf_s[c_ddZoneCommMaxNumZones];
748 gmx_ddzone_t buf_r[c_ddZoneCommMaxNumZones];
749 gmx_ddzone_t buf_e[c_ddZoneCommMaxNumZones];
750 gmx_domdec_comm_t *comm = dd->comm;
754 for (int d = 1; d < dd->ndim; d++)
756 int dim = dd->dim[d];
757 gmx_ddzone_t &zp = (d == 1) ? comm->zone_d1[0] : comm->zone_d2[0][0];
759 /* Copy the base sizes of the home zone */
760 zp.min0 = cell_ns_x0[dim];
761 zp.max1 = cell_ns_x1[dim];
762 zp.min1 = cell_ns_x1[dim];
763 zp.mch0 = cell_ns_x0[dim];
764 zp.mch1 = cell_ns_x1[dim];
765 zp.p1_0 = cell_ns_x0[dim];
766 zp.p1_1 = cell_ns_x1[dim];
769 gmx::ArrayRef<DDCellsizesWithDlb> cellsizes = comm->cellsizesWithDlb;
771 /* Loop backward over the dimensions and aggregate the extremes
774 for (int d = dd->ndim - 2; d >= 0; d--)
776 const int dim = dd->dim[d];
777 const bool applyPbc = (dim < ddbox->npbcdim);
779 /* Use an rvec to store two reals */
780 extr_s[d][0] = cellsizes[d + 1].fracLower;
781 extr_s[d][1] = cellsizes[d + 1].fracUpper;
782 extr_s[d][2] = cellsizes[d + 1].fracUpper;
785 GMX_ASSERT(pos < c_ddZoneCommMaxNumZones, "The buffers should be sufficiently large");
786 /* Store the extremes in the backward sending buffer,
787 * so they get updated separately from the forward communication.
789 for (int d1 = d; d1 < dd->ndim-1; d1++)
791 /* We invert the order to be able to use the same loop for buf_e */
792 buf_s[pos].min0 = extr_s[d1][1];
793 buf_s[pos].max1 = extr_s[d1][0];
794 buf_s[pos].min1 = extr_s[d1][2];
797 /* Store the cell corner of the dimension we communicate along */
798 buf_s[pos].p1_0 = comm->cell_x0[dim];
803 buf_s[pos] = (dd->ndim == 2) ? comm->zone_d1[0] : comm->zone_d2[0][0];
806 if (dd->ndim == 3 && d == 0)
808 buf_s[pos] = comm->zone_d2[0][1];
810 buf_s[pos] = comm->zone_d1[0];
814 /* We only need to communicate the extremes
815 * in the forward direction
817 int numPulses = comm->cd[d].numPulses();
821 /* Take the minimum to avoid double communication */
822 numPulsesMin = std::min(numPulses, dd->nc[dim] - 1 - numPulses);
826 /* Without PBC we should really not communicate over
827 * the boundaries, but implementing that complicates
828 * the communication setup and therefore we simply
829 * do all communication, but ignore some data.
831 numPulsesMin = numPulses;
833 for (int pulse = 0; pulse < numPulsesMin; pulse++)
835 /* Communicate the extremes forward */
836 bool receiveValidData = (applyPbc || dd->ci[dim] > 0);
838 int numElements = dd->ndim - d - 1;
839 ddSendrecv(dd, d, dddirForward,
840 extr_s + d, numElements,
841 extr_r + d, numElements);
843 if (receiveValidData)
845 for (int d1 = d; d1 < dd->ndim - 1; d1++)
847 extr_s[d1][0] = std::max(extr_s[d1][0], extr_r[d1][0]);
848 extr_s[d1][1] = std::min(extr_s[d1][1], extr_r[d1][1]);
849 extr_s[d1][2] = std::min(extr_s[d1][2], extr_r[d1][2]);
854 const int numElementsInBuffer = pos;
855 for (int pulse = 0; pulse < numPulses; pulse++)
857 /* Communicate all the zone information backward */
858 bool receiveValidData = (applyPbc || dd->ci[dim] < dd->nc[dim] - 1);
860 static_assert(sizeof(gmx_ddzone_t) == c_ddzoneNumReals*sizeof(real), "Here we expect gmx_ddzone_t to consist of c_ddzoneNumReals reals (only)");
862 int numReals = numElementsInBuffer*c_ddzoneNumReals;
863 ddSendrecv(dd, d, dddirBackward,
864 gmx::arrayRefFromArray(&buf_s[0].min0, numReals),
865 gmx::arrayRefFromArray(&buf_r[0].min0, numReals));
870 for (int d1 = d + 1; d1 < dd->ndim; d1++)
872 /* Determine the decrease of maximum required
873 * communication height along d1 due to the distance along d,
874 * this avoids a lot of useless atom communication.
876 real dist_d = comm->cell_x1[dim] - buf_r[0].p1_0;
879 if (ddbox->tric_dir[dim])
881 /* c is the off-diagonal coupling between the cell planes
882 * along directions d and d1.
884 c = ddbox->v[dim][dd->dim[d1]][dim];
890 real det = (1 + c*c)*comm->cutoff*comm->cutoff - dist_d*dist_d;
893 dh[d1] = comm->cutoff - (c*dist_d + std::sqrt(det))/(1 + c*c);
897 /* A negative value signals out of range */
903 /* Accumulate the extremes over all pulses */
904 for (int i = 0; i < numElementsInBuffer; i++)
912 if (receiveValidData)
914 buf_e[i].min0 = std::min(buf_e[i].min0, buf_r[i].min0);
915 buf_e[i].max1 = std::max(buf_e[i].max1, buf_r[i].max1);
916 buf_e[i].min1 = std::min(buf_e[i].min1, buf_r[i].min1);
920 if (dd->ndim == 3 && d == 0 && i == numElementsInBuffer - 1)
928 if (receiveValidData && dh[d1] >= 0)
930 buf_e[i].mch0 = std::max(buf_e[i].mch0, buf_r[i].mch0-dh[d1]);
931 buf_e[i].mch1 = std::max(buf_e[i].mch1, buf_r[i].mch1-dh[d1]);
934 /* Copy the received buffer to the send buffer,
935 * to pass the data through with the next pulse.
939 if (((applyPbc || dd->ci[dim] + numPulses < dd->nc[dim]) && pulse == numPulses - 1) ||
940 (!applyPbc && dd->ci[dim] + 1 + pulse == dd->nc[dim] - 1))
942 /* Store the extremes */
945 for (int d1 = d; d1 < dd->ndim-1; d1++)
947 extr_s[d1][1] = std::min(extr_s[d1][1], buf_e[pos].min0);
948 extr_s[d1][0] = std::max(extr_s[d1][0], buf_e[pos].max1);
949 extr_s[d1][2] = std::min(extr_s[d1][2], buf_e[pos].min1);
953 if (d == 1 || (d == 0 && dd->ndim == 3))
955 for (int i = d; i < 2; i++)
957 comm->zone_d2[1-d][i] = buf_e[pos];
963 comm->zone_d1[1] = buf_e[pos];
972 int dim = dd->dim[1];
973 for (int i = 0; i < 2; i++)
977 print_ddzone(debug, 1, i, 0, &comm->zone_d1[i]);
979 cell_ns_x0[dim] = std::min(cell_ns_x0[dim], comm->zone_d1[i].min0);
980 cell_ns_x1[dim] = std::max(cell_ns_x1[dim], comm->zone_d1[i].max1);
985 int dim = dd->dim[2];
986 for (int i = 0; i < 2; i++)
988 for (int j = 0; j < 2; j++)
992 print_ddzone(debug, 2, i, j, &comm->zone_d2[i][j]);
994 cell_ns_x0[dim] = std::min(cell_ns_x0[dim], comm->zone_d2[i][j].min0);
995 cell_ns_x1[dim] = std::max(cell_ns_x1[dim], comm->zone_d2[i][j].max1);
999 for (int d = 1; d < dd->ndim; d++)
1001 cellsizes[d].fracLowerMax = extr_s[d-1][0];
1002 cellsizes[d].fracUpperMin = extr_s[d-1][1];
1005 fprintf(debug, "Cell fraction d %d, max0 %f, min1 %f\n",
1006 d, cellsizes[d].fracLowerMax, cellsizes[d].fracUpperMin);
1011 static void write_dd_grid_pdb(const char *fn, int64_t step,
1012 gmx_domdec_t *dd, matrix box, gmx_ddbox_t *ddbox)
1014 rvec grid_s[2], *grid_r = nullptr, cx, r;
1015 char fname[STRLEN], buf[22];
1017 int a, i, d, z, y, x;
1021 copy_rvec(dd->comm->cell_x0, grid_s[0]);
1022 copy_rvec(dd->comm->cell_x1, grid_s[1]);
1026 snew(grid_r, 2*dd->nnodes);
1029 dd_gather(dd, 2*sizeof(rvec), grid_s, DDMASTER(dd) ? grid_r : nullptr);
1033 for (d = 0; d < DIM; d++)
1035 for (i = 0; i < DIM; i++)
1043 if (d < ddbox->npbcdim && dd->nc[d] > 1)
1045 tric[d][i] = box[i][d]/box[i][i];
1054 sprintf(fname, "%s_%s.pdb", fn, gmx_step_str(step, buf));
1055 out = gmx_fio_fopen(fname, "w");
1056 gmx_write_pdb_box(out, dd->bScrewPBC ? epbcSCREW : epbcXYZ, box);
1058 for (i = 0; i < dd->nnodes; i++)
1060 vol = dd->nnodes/(box[XX][XX]*box[YY][YY]*box[ZZ][ZZ]);
1061 for (d = 0; d < DIM; d++)
1063 vol *= grid_r[i*2+1][d] - grid_r[i*2][d];
1065 for (z = 0; z < 2; z++)
1067 for (y = 0; y < 2; y++)
1069 for (x = 0; x < 2; x++)
1071 cx[XX] = grid_r[i*2+x][XX];
1072 cx[YY] = grid_r[i*2+y][YY];
1073 cx[ZZ] = grid_r[i*2+z][ZZ];
1075 gmx_fprintf_pdb_atomline(out, epdbATOM, a++, "CA", ' ', "GLY", ' ', i+1, ' ',
1076 10*r[XX], 10*r[YY], 10*r[ZZ], 1.0, vol, "");
1080 for (d = 0; d < DIM; d++)
1082 for (x = 0; x < 4; x++)
1086 case 0: y = 1 + i*8 + 2*x; break;
1087 case 1: y = 1 + i*8 + 2*x - (x % 2); break;
1088 case 2: y = 1 + i*8 + x; break;
1090 fprintf(out, "%6s%5d%5d\n", "CONECT", y, y+(1<<d));
1094 gmx_fio_fclose(out);
1099 void write_dd_pdb(const char *fn, int64_t step, const char *title,
1100 const gmx_mtop_t *mtop, const t_commrec *cr,
1101 int natoms, const rvec x[], const matrix box)
1103 char fname[STRLEN], buf[22];
1106 const char *atomname, *resname;
1112 natoms = dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
1115 sprintf(fname, "%s_%s_n%d.pdb", fn, gmx_step_str(step, buf), cr->sim_nodeid);
1117 out = gmx_fio_fopen(fname, "w");
1119 fprintf(out, "TITLE %s\n", title);
1120 gmx_write_pdb_box(out, dd->bScrewPBC ? epbcSCREW : epbcXYZ, box);
1122 for (int i = 0; i < natoms; i++)
1124 int ii = dd->globalAtomIndices[i];
1125 mtopGetAtomAndResidueName(mtop, ii, &molb, &atomname, &resnr, &resname, nullptr);
1128 if (i < dd->comm->atomRanges.end(DDAtomRanges::Type::Zones))
1131 while (i >= dd->atomGrouping().subRange(0, dd->comm->zones.cg_range[c + 1]).end())
1137 else if (i < dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites))
1139 b = dd->comm->zones.n;
1143 b = dd->comm->zones.n + 1;
1145 gmx_fprintf_pdb_atomline(out, epdbATOM, ii+1, atomname, ' ', resname, ' ', resnr, ' ',
1146 10*x[i][XX], 10*x[i][YY], 10*x[i][ZZ], 1.0, b, "");
1148 fprintf(out, "TER\n");
1150 gmx_fio_fclose(out);
1153 real dd_cutoff_multibody(const gmx_domdec_t *dd)
1155 gmx_domdec_comm_t *comm;
1162 if (comm->bInterCGBondeds)
1164 if (comm->cutoff_mbody > 0)
1166 r = comm->cutoff_mbody;
1170 /* cutoff_mbody=0 means we do not have DLB */
1171 r = comm->cellsize_min[dd->dim[0]];
1172 for (di = 1; di < dd->ndim; di++)
1174 r = std::min(r, comm->cellsize_min[dd->dim[di]]);
1176 if (comm->bBondComm)
1178 r = std::max(r, comm->cutoff_mbody);
1182 r = std::min(r, comm->cutoff);
1190 real dd_cutoff_twobody(const gmx_domdec_t *dd)
1194 r_mb = dd_cutoff_multibody(dd);
1196 return std::max(dd->comm->cutoff, r_mb);
1200 static void dd_cart_coord2pmecoord(const gmx_domdec_t *dd, const ivec coord,
1205 nc = dd->nc[dd->comm->cartpmedim];
1206 ntot = dd->comm->ntot[dd->comm->cartpmedim];
1207 copy_ivec(coord, coord_pme);
1208 coord_pme[dd->comm->cartpmedim] =
1209 nc + (coord[dd->comm->cartpmedim]*(ntot - nc) + (ntot - nc)/2)/nc;
1212 static int ddindex2pmeindex(const gmx_domdec_t *dd, int ddindex)
1217 npme = dd->comm->npmenodes;
1219 /* Here we assign a PME node to communicate with this DD node
1220 * by assuming that the major index of both is x.
1221 * We add cr->npmenodes/2 to obtain an even distribution.
1223 return (ddindex*npme + npme/2)/npp;
1226 static int *dd_interleaved_pme_ranks(const gmx_domdec_t *dd)
1231 snew(pme_rank, dd->comm->npmenodes);
1233 for (i = 0; i < dd->nnodes; i++)
1235 p0 = ddindex2pmeindex(dd, i);
1236 p1 = ddindex2pmeindex(dd, i+1);
1237 if (i+1 == dd->nnodes || p1 > p0)
1241 fprintf(debug, "pme_rank[%d] = %d\n", n, i+1+n);
1243 pme_rank[n] = i + 1 + n;
1251 static int gmx_ddcoord2pmeindex(const t_commrec *cr, int x, int y, int z)
1259 if (dd->comm->bCartesian) {
1260 gmx_ddindex2xyz(dd->nc,ddindex,coords);
1261 dd_coords2pmecoords(dd,coords,coords_pme);
1262 copy_ivec(dd->ntot,nc);
1263 nc[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
1264 coords_pme[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
1266 slab = (coords_pme[XX]*nc[YY] + coords_pme[YY])*nc[ZZ] + coords_pme[ZZ];
1268 slab = (ddindex*cr->npmenodes + cr->npmenodes/2)/dd->nnodes;
1274 slab = ddindex2pmeindex(dd, dd_index(dd->nc, coords));
1279 static int ddcoord2simnodeid(const t_commrec *cr, int x, int y, int z)
1281 gmx_domdec_comm_t *comm;
1283 int ddindex, nodeid = -1;
1285 comm = cr->dd->comm;
1290 if (comm->bCartesianPP_PME)
1293 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
1298 ddindex = dd_index(cr->dd->nc, coords);
1299 if (comm->bCartesianPP)
1301 nodeid = comm->ddindex2simnodeid[ddindex];
1307 nodeid = ddindex + gmx_ddcoord2pmeindex(cr, x, y, z);
1319 static int dd_simnode2pmenode(const gmx_domdec_t *dd,
1320 const t_commrec gmx_unused *cr,
1325 const gmx_domdec_comm_t *comm = dd->comm;
1327 /* This assumes a uniform x domain decomposition grid cell size */
1328 if (comm->bCartesianPP_PME)
1331 ivec coord, coord_pme;
1332 MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
1333 if (coord[comm->cartpmedim] < dd->nc[comm->cartpmedim])
1335 /* This is a PP node */
1336 dd_cart_coord2pmecoord(dd, coord, coord_pme);
1337 MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
1341 else if (comm->bCartesianPP)
1343 if (sim_nodeid < dd->nnodes)
1345 pmenode = dd->nnodes + ddindex2pmeindex(dd, sim_nodeid);
1350 /* This assumes DD cells with identical x coordinates
1351 * are numbered sequentially.
1353 if (dd->comm->pmenodes == nullptr)
1355 if (sim_nodeid < dd->nnodes)
1357 /* The DD index equals the nodeid */
1358 pmenode = dd->nnodes + ddindex2pmeindex(dd, sim_nodeid);
1364 while (sim_nodeid > dd->comm->pmenodes[i])
1368 if (sim_nodeid < dd->comm->pmenodes[i])
1370 pmenode = dd->comm->pmenodes[i];
1378 NumPmeDomains getNumPmeDomains(const gmx_domdec_t *dd)
1383 dd->comm->npmenodes_x, dd->comm->npmenodes_y
1394 std::vector<int> get_pme_ddranks(const t_commrec *cr, int pmenodeid)
1398 ivec coord, coord_pme;
1402 std::vector<int> ddranks;
1403 ddranks.reserve((dd->nnodes+cr->npmenodes-1)/cr->npmenodes);
1405 for (x = 0; x < dd->nc[XX]; x++)
1407 for (y = 0; y < dd->nc[YY]; y++)
1409 for (z = 0; z < dd->nc[ZZ]; z++)
1411 if (dd->comm->bCartesianPP_PME)
1416 dd_cart_coord2pmecoord(dd, coord, coord_pme);
1417 if (dd->ci[XX] == coord_pme[XX] &&
1418 dd->ci[YY] == coord_pme[YY] &&
1419 dd->ci[ZZ] == coord_pme[ZZ])
1421 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
1426 /* The slab corresponds to the nodeid in the PME group */
1427 if (gmx_ddcoord2pmeindex(cr, x, y, z) == pmenodeid)
1429 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
1438 static gmx_bool receive_vir_ener(const gmx_domdec_t *dd, const t_commrec *cr)
1440 gmx_bool bReceive = TRUE;
1442 if (cr->npmenodes < dd->nnodes)
1444 gmx_domdec_comm_t *comm = dd->comm;
1445 if (comm->bCartesianPP_PME)
1448 int pmenode = dd_simnode2pmenode(dd, cr, cr->sim_nodeid);
1450 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
1451 coords[comm->cartpmedim]++;
1452 if (coords[comm->cartpmedim] < dd->nc[comm->cartpmedim])
1455 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
1456 if (dd_simnode2pmenode(dd, cr, rank) == pmenode)
1458 /* This is not the last PP node for pmenode */
1463 GMX_RELEASE_ASSERT(false, "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
1468 int pmenode = dd_simnode2pmenode(dd, cr, cr->sim_nodeid);
1469 if (cr->sim_nodeid+1 < cr->nnodes &&
1470 dd_simnode2pmenode(dd, cr, cr->sim_nodeid+1) == pmenode)
1472 /* This is not the last PP node for pmenode */
1481 static void set_zones_ncg_home(gmx_domdec_t *dd)
1483 gmx_domdec_zones_t *zones;
1486 zones = &dd->comm->zones;
1488 zones->cg_range[0] = 0;
1489 for (i = 1; i < zones->n+1; i++)
1491 zones->cg_range[i] = dd->ncg_home;
1493 /* zone_ncg1[0] should always be equal to ncg_home */
1494 dd->comm->zone_ncg1[0] = dd->ncg_home;
1497 static void restoreAtomGroups(gmx_domdec_t *dd,
1498 const int *gcgs_index, const t_state *state)
1500 gmx::ArrayRef<const int> atomGroupsState = state->cg_gl;
1502 std::vector<int> &globalAtomGroupIndices = dd->globalAtomGroupIndices;
1503 gmx::RangePartitioning &atomGrouping = dd->atomGrouping_;
1505 globalAtomGroupIndices.resize(atomGroupsState.size());
1506 atomGrouping.clear();
1508 /* Copy back the global charge group indices from state
1509 * and rebuild the local charge group to atom index.
1511 for (gmx::index i = 0; i < atomGroupsState.size(); i++)
1513 const int atomGroupGlobal = atomGroupsState[i];
1514 const int groupSize = gcgs_index[atomGroupGlobal + 1] - gcgs_index[atomGroupGlobal];
1515 globalAtomGroupIndices[i] = atomGroupGlobal;
1516 atomGrouping.appendBlock(groupSize);
1519 dd->ncg_home = atomGroupsState.size();
1520 dd->comm->atomRanges.setEnd(DDAtomRanges::Type::Home, atomGrouping.fullRange().end());
1522 set_zones_ncg_home(dd);
1525 static void dd_set_cginfo(gmx::ArrayRef<const int> index_gl, int cg0, int cg1,
1526 t_forcerec *fr, char *bLocalCG)
1528 cginfo_mb_t *cginfo_mb;
1534 cginfo_mb = fr->cginfo_mb;
1535 cginfo = fr->cginfo;
1537 for (cg = cg0; cg < cg1; cg++)
1539 cginfo[cg] = ddcginfo(cginfo_mb, index_gl[cg]);
1543 if (bLocalCG != nullptr)
1545 for (cg = cg0; cg < cg1; cg++)
1547 bLocalCG[index_gl[cg]] = TRUE;
1552 static void make_dd_indices(gmx_domdec_t *dd,
1553 const int *gcgs_index, int cg_start)
1555 const int numZones = dd->comm->zones.n;
1556 const int *zone2cg = dd->comm->zones.cg_range;
1557 const int *zone_ncg1 = dd->comm->zone_ncg1;
1558 gmx::ArrayRef<const int> globalAtomGroupIndices = dd->globalAtomGroupIndices;
1559 const gmx_bool bCGs = dd->comm->bCGs;
1561 std::vector<int> &globalAtomIndices = dd->globalAtomIndices;
1563 if (zone2cg[1] != dd->ncg_home)
1565 gmx_incons("dd->ncg_zone is not up to date");
1568 /* Make the local to global and global to local atom index */
1569 int a = dd->atomGrouping().subRange(cg_start, cg_start).begin();
1570 globalAtomIndices.resize(a);
1571 for (int zone = 0; zone < numZones; zone++)
1580 cg0 = zone2cg[zone];
1582 int cg1 = zone2cg[zone+1];
1583 int cg1_p1 = cg0 + zone_ncg1[zone];
1585 for (int cg = cg0; cg < cg1; cg++)
1590 /* Signal that this cg is from more than one pulse away */
1593 int cg_gl = globalAtomGroupIndices[cg];
1596 for (int a_gl = gcgs_index[cg_gl]; a_gl < gcgs_index[cg_gl+1]; a_gl++)
1598 globalAtomIndices.push_back(a_gl);
1599 ga2la_set(dd->ga2la, a_gl, a, zone1);
1605 globalAtomIndices.push_back(cg_gl);
1606 ga2la_set(dd->ga2la, cg_gl, a, zone1);
1613 static int check_bLocalCG(gmx_domdec_t *dd, int ncg_sys, const char *bLocalCG,
1617 if (bLocalCG == nullptr)
1621 for (size_t i = 0; i < dd->globalAtomGroupIndices.size(); i++)
1623 if (!bLocalCG[dd->globalAtomGroupIndices[i]])
1626 "DD rank %d, %s: atom group %zu, global atom group %d is not marked in bLocalCG (ncg_home %d)\n", dd->rank, where, i + 1, dd->globalAtomGroupIndices[i] + 1, dd->ncg_home);
1631 for (int i = 0; i < ncg_sys; i++)
1638 if (ngl != dd->globalAtomGroupIndices.size())
1640 fprintf(stderr, "DD rank %d, %s: In bLocalCG %zu atom groups are marked as local, whereas there are %zu\n", dd->rank, where, ngl, dd->globalAtomGroupIndices.size());
1647 static void check_index_consistency(gmx_domdec_t *dd,
1648 int natoms_sys, int ncg_sys,
1653 const int numAtomsInZones = dd->comm->atomRanges.end(DDAtomRanges::Type::Zones);
1655 if (dd->comm->DD_debug > 1)
1657 std::vector<int> have(natoms_sys);
1658 for (int a = 0; a < numAtomsInZones; a++)
1660 int globalAtomIndex = dd->globalAtomIndices[a];
1661 if (have[globalAtomIndex] > 0)
1663 fprintf(stderr, "DD rank %d: global atom %d occurs twice: index %d and %d\n", dd->rank, globalAtomIndex + 1, have[globalAtomIndex], a+1);
1667 have[globalAtomIndex] = a + 1;
1672 std::vector<int> have(numAtomsInZones);
1675 for (int i = 0; i < natoms_sys; i++)
1679 if (ga2la_get(dd->ga2la, i, &a, &cell))
1681 if (a >= numAtomsInZones)
1683 fprintf(stderr, "DD rank %d: global atom %d marked as local atom %d, which is larger than nat_tot (%d)\n", dd->rank, i+1, a+1, numAtomsInZones);
1689 if (dd->globalAtomIndices[a] != i)
1691 fprintf(stderr, "DD rank %d: global atom %d marked as local atom %d, which has global atom index %d\n", dd->rank, i+1, a+1, dd->globalAtomIndices[a]+1);
1698 if (ngl != numAtomsInZones)
1701 "DD rank %d, %s: %d global atom indices, %d local atoms\n",
1702 dd->rank, where, ngl, numAtomsInZones);
1704 for (int a = 0; a < numAtomsInZones; a++)
1709 "DD rank %d, %s: local atom %d, global %d has no global index\n",
1710 dd->rank, where, a + 1, dd->globalAtomIndices[a] + 1);
1714 nerr += check_bLocalCG(dd, ncg_sys, dd->comm->bLocalCG, where);
1718 gmx_fatal(FARGS, "DD rank %d, %s: %d atom(group) index inconsistencies",
1719 dd->rank, where, nerr);
1723 /* Clear all DD global state indices, starting from \p atomGroupStart and \p atomStart */
1724 static void clearDDStateIndices(gmx_domdec_t *dd,
1730 /* Clear the whole list without searching */
1731 ga2la_clear(dd->ga2la);
1735 const int numAtomsInZones = dd->comm->atomRanges.end(DDAtomRanges::Type::Zones);
1736 for (int i = 0; i < numAtomsInZones; i++)
1738 ga2la_del(dd->ga2la, dd->globalAtomIndices[i]);
1742 char *bLocalCG = dd->comm->bLocalCG;
1745 for (size_t atomGroup = atomGroupStart; atomGroup < dd->globalAtomGroupIndices.size(); atomGroup++)
1747 bLocalCG[dd->globalAtomGroupIndices[atomGroup]] = FALSE;
1751 dd_clear_local_vsite_indices(dd);
1753 if (dd->constraints)
1755 dd_clear_local_constraint_indices(dd);
1759 static bool check_grid_jump(int64_t step,
1760 const gmx_domdec_t *dd,
1762 const gmx_ddbox_t *ddbox,
1765 gmx_domdec_comm_t *comm = dd->comm;
1766 bool invalid = false;
1768 for (int d = 1; d < dd->ndim; d++)
1770 const DDCellsizesWithDlb &cellsizes = comm->cellsizesWithDlb[d];
1771 const int dim = dd->dim[d];
1772 const real limit = grid_jump_limit(comm, cutoff, d);
1773 real bfac = ddbox->box_size[dim];
1774 if (ddbox->tric_dir[dim])
1776 bfac *= ddbox->skew_fac[dim];
1778 if ((cellsizes.fracUpper - cellsizes.fracLowerMax)*bfac < limit ||
1779 (cellsizes.fracLower - cellsizes.fracUpperMin)*bfac > -limit)
1787 /* This error should never be triggered under normal
1788 * circumstances, but you never know ...
1790 gmx_fatal(FARGS, "step %s: The domain decomposition grid has shifted too much in the %c-direction around cell %d %d %d. This should not have happened. Running with fewer ranks might avoid this issue.",
1791 gmx_step_str(step, buf),
1792 dim2char(dim), dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1800 static float dd_force_load(gmx_domdec_comm_t *comm)
1807 if (comm->eFlop > 1)
1809 load *= 1.0 + (comm->eFlop - 1)*(0.1*rand()/RAND_MAX - 0.05);
1814 load = comm->cycl[ddCyclF];
1815 if (comm->cycl_n[ddCyclF] > 1)
1817 /* Subtract the maximum of the last n cycle counts
1818 * to get rid of possible high counts due to other sources,
1819 * for instance system activity, that would otherwise
1820 * affect the dynamic load balancing.
1822 load -= comm->cycl_max[ddCyclF];
1826 if (comm->cycl_n[ddCyclWaitGPU] && comm->nrank_gpu_shared > 1)
1828 float gpu_wait, gpu_wait_sum;
1830 gpu_wait = comm->cycl[ddCyclWaitGPU];
1831 if (comm->cycl_n[ddCyclF] > 1)
1833 /* We should remove the WaitGPU time of the same MD step
1834 * as the one with the maximum F time, since the F time
1835 * and the wait time are not independent.
1836 * Furthermore, the step for the max F time should be chosen
1837 * the same on all ranks that share the same GPU.
1838 * But to keep the code simple, we remove the average instead.
1839 * The main reason for artificially long times at some steps
1840 * is spurious CPU activity or MPI time, so we don't expect
1841 * that changes in the GPU wait time matter a lot here.
1843 gpu_wait *= (comm->cycl_n[ddCyclF] - 1)/static_cast<float>(comm->cycl_n[ddCyclF]);
1845 /* Sum the wait times over the ranks that share the same GPU */
1846 MPI_Allreduce(&gpu_wait, &gpu_wait_sum, 1, MPI_FLOAT, MPI_SUM,
1847 comm->mpi_comm_gpu_shared);
1848 /* Replace the wait time by the average over the ranks */
1849 load += -gpu_wait + gpu_wait_sum/comm->nrank_gpu_shared;
1857 static void set_slb_pme_dim_f(gmx_domdec_t *dd, int dim, real **dim_f)
1859 gmx_domdec_comm_t *comm;
1864 snew(*dim_f, dd->nc[dim]+1);
1866 for (i = 1; i < dd->nc[dim]; i++)
1868 if (comm->slb_frac[dim])
1870 (*dim_f)[i] = (*dim_f)[i-1] + comm->slb_frac[dim][i-1];
1874 (*dim_f)[i] = static_cast<real>(i)/static_cast<real>(dd->nc[dim]);
1877 (*dim_f)[dd->nc[dim]] = 1;
1880 static void init_ddpme(gmx_domdec_t *dd, gmx_ddpme_t *ddpme, int dimind)
1882 int pmeindex, slab, nso, i;
1885 if (dimind == 0 && dd->dim[0] == YY && dd->comm->npmenodes_x == 1)
1891 ddpme->dim = dimind;
1893 ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
1895 ddpme->nslab = (ddpme->dim == 0 ?
1896 dd->comm->npmenodes_x :
1897 dd->comm->npmenodes_y);
1899 if (ddpme->nslab <= 1)
1904 nso = dd->comm->npmenodes/ddpme->nslab;
1905 /* Determine for each PME slab the PP location range for dimension dim */
1906 snew(ddpme->pp_min, ddpme->nslab);
1907 snew(ddpme->pp_max, ddpme->nslab);
1908 for (slab = 0; slab < ddpme->nslab; slab++)
1910 ddpme->pp_min[slab] = dd->nc[dd->dim[dimind]] - 1;
1911 ddpme->pp_max[slab] = 0;
1913 for (i = 0; i < dd->nnodes; i++)
1915 ddindex2xyz(dd->nc, i, xyz);
1916 /* For y only use our y/z slab.
1917 * This assumes that the PME x grid size matches the DD grid size.
1919 if (dimind == 0 || xyz[XX] == dd->ci[XX])
1921 pmeindex = ddindex2pmeindex(dd, i);
1924 slab = pmeindex/nso;
1928 slab = pmeindex % ddpme->nslab;
1930 ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
1931 ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
1935 set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
1938 int dd_pme_maxshift_x(const gmx_domdec_t *dd)
1940 if (dd->comm->ddpme[0].dim == XX)
1942 return dd->comm->ddpme[0].maxshift;
1950 int dd_pme_maxshift_y(const gmx_domdec_t *dd)
1952 if (dd->comm->ddpme[0].dim == YY)
1954 return dd->comm->ddpme[0].maxshift;
1956 else if (dd->comm->npmedecompdim >= 2 && dd->comm->ddpme[1].dim == YY)
1958 return dd->comm->ddpme[1].maxshift;
1966 static void comm_dd_ns_cell_sizes(gmx_domdec_t *dd,
1968 rvec cell_ns_x0, rvec cell_ns_x1,
1971 gmx_domdec_comm_t *comm;
1976 for (dim_ind = 0; dim_ind < dd->ndim; dim_ind++)
1978 dim = dd->dim[dim_ind];
1980 /* Without PBC we don't have restrictions on the outer cells */
1981 if (!(dim >= ddbox->npbcdim &&
1982 (dd->ci[dim] == 0 || dd->ci[dim] == dd->nc[dim] - 1)) &&
1984 (comm->cell_x1[dim] - comm->cell_x0[dim])*ddbox->skew_fac[dim] <
1985 comm->cellsize_min[dim])
1988 gmx_fatal(FARGS, "step %s: The %c-size (%f) times the triclinic skew factor (%f) is smaller than the smallest allowed cell size (%f) for domain decomposition grid cell %d %d %d",
1989 gmx_step_str(step, buf), dim2char(dim),
1990 comm->cell_x1[dim] - comm->cell_x0[dim],
1991 ddbox->skew_fac[dim],
1992 dd->comm->cellsize_min[dim],
1993 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1997 if ((isDlbOn(dd->comm) && dd->ndim > 1) || ddbox->nboundeddim < DIM)
1999 /* Communicate the boundaries and update cell_ns_x0/1 */
2000 dd_move_cellx(dd, ddbox, cell_ns_x0, cell_ns_x1);
2001 if (isDlbOn(dd->comm) && dd->ndim > 1)
2003 check_grid_jump(step, dd, dd->comm->cutoff, ddbox, TRUE);
2008 void dd_cycles_add(const gmx_domdec_t *dd, float cycles, int ddCycl)
2010 /* Note that the cycles value can be incorrect, either 0 or some
2011 * extremely large value, when our thread migrated to another core
2012 * with an unsynchronized cycle counter. If this happens less often
2013 * that once per nstlist steps, this will not cause issues, since
2014 * we later subtract the maximum value from the sum over nstlist steps.
2015 * A zero count will slightly lower the total, but that's a small effect.
2016 * Note that the main purpose of the subtraction of the maximum value
2017 * is to avoid throwing off the load balancing when stalls occur due
2018 * e.g. system activity or network congestion.
2020 dd->comm->cycl[ddCycl] += cycles;
2021 dd->comm->cycl_n[ddCycl]++;
2022 if (cycles > dd->comm->cycl_max[ddCycl])
2024 dd->comm->cycl_max[ddCycl] = cycles;
2028 static double force_flop_count(t_nrnb *nrnb)
2035 for (i = 0; i < eNR_NBKERNEL_FREE_ENERGY; i++)
2037 /* To get closer to the real timings, we half the count
2038 * for the normal loops and again half it for water loops.
2041 if (strstr(name, "W3") != nullptr || strstr(name, "W4") != nullptr)
2043 sum += nrnb->n[i]*0.25*cost_nrnb(i);
2047 sum += nrnb->n[i]*0.50*cost_nrnb(i);
2050 for (i = eNR_NBKERNEL_FREE_ENERGY; i <= eNR_NB14; i++)
2053 if (strstr(name, "W3") != nullptr || strstr(name, "W4") != nullptr)
2055 sum += nrnb->n[i]*cost_nrnb(i);
2058 for (i = eNR_BONDS; i <= eNR_WALLS; i++)
2060 sum += nrnb->n[i]*cost_nrnb(i);
2066 void dd_force_flop_start(gmx_domdec_t *dd, t_nrnb *nrnb)
2068 if (dd->comm->eFlop)
2070 dd->comm->flop -= force_flop_count(nrnb);
2073 void dd_force_flop_stop(gmx_domdec_t *dd, t_nrnb *nrnb)
2075 if (dd->comm->eFlop)
2077 dd->comm->flop += force_flop_count(nrnb);
2082 static void clear_dd_cycle_counts(gmx_domdec_t *dd)
2086 for (i = 0; i < ddCyclNr; i++)
2088 dd->comm->cycl[i] = 0;
2089 dd->comm->cycl_n[i] = 0;
2090 dd->comm->cycl_max[i] = 0;
2093 dd->comm->flop_n = 0;
2096 static void get_load_distribution(gmx_domdec_t *dd, gmx_wallcycle_t wcycle)
2098 gmx_domdec_comm_t *comm;
2099 domdec_load_t *load;
2100 float cell_frac = 0, sbuf[DD_NLOAD_MAX];
2105 fprintf(debug, "get_load_distribution start\n");
2108 wallcycle_start(wcycle, ewcDDCOMMLOAD);
2112 bSepPME = (dd->pme_nodeid >= 0);
2114 if (dd->ndim == 0 && bSepPME)
2116 /* Without decomposition, but with PME nodes, we need the load */
2117 comm->load[0].mdf = comm->cycl[ddCyclPPduringPME];
2118 comm->load[0].pme = comm->cycl[ddCyclPME];
2121 for (int d = dd->ndim - 1; d >= 0; d--)
2123 const DDCellsizesWithDlb &cellsizes = comm->cellsizesWithDlb[d];
2124 const int dim = dd->dim[d];
2125 /* Check if we participate in the communication in this dimension */
2126 if (d == dd->ndim-1 ||
2127 (dd->ci[dd->dim[d+1]] == 0 && dd->ci[dd->dim[dd->ndim-1]] == 0))
2129 load = &comm->load[d];
2130 if (isDlbOn(dd->comm))
2132 cell_frac = cellsizes.fracUpper - cellsizes.fracLower;
2135 if (d == dd->ndim-1)
2137 sbuf[pos++] = dd_force_load(comm);
2138 sbuf[pos++] = sbuf[0];
2139 if (isDlbOn(dd->comm))
2141 sbuf[pos++] = sbuf[0];
2142 sbuf[pos++] = cell_frac;
2145 sbuf[pos++] = cellsizes.fracLowerMax;
2146 sbuf[pos++] = cellsizes.fracUpperMin;
2151 sbuf[pos++] = comm->cycl[ddCyclPPduringPME];
2152 sbuf[pos++] = comm->cycl[ddCyclPME];
2157 sbuf[pos++] = comm->load[d+1].sum;
2158 sbuf[pos++] = comm->load[d+1].max;
2159 if (isDlbOn(dd->comm))
2161 sbuf[pos++] = comm->load[d+1].sum_m;
2162 sbuf[pos++] = comm->load[d+1].cvol_min*cell_frac;
2163 sbuf[pos++] = comm->load[d+1].flags;
2166 sbuf[pos++] = cellsizes.fracLowerMax;
2167 sbuf[pos++] = cellsizes.fracUpperMin;
2172 sbuf[pos++] = comm->load[d+1].mdf;
2173 sbuf[pos++] = comm->load[d+1].pme;
2177 /* Communicate a row in DD direction d.
2178 * The communicators are setup such that the root always has rank 0.
2181 MPI_Gather(sbuf, load->nload*sizeof(float), MPI_BYTE,
2182 load->load, load->nload*sizeof(float), MPI_BYTE,
2183 0, comm->mpi_comm_load[d]);
2185 if (dd->ci[dim] == dd->master_ci[dim])
2187 /* We are the master along this row, process this row */
2188 RowMaster *rowMaster = nullptr;
2192 rowMaster = cellsizes.rowMaster.get();
2202 for (int i = 0; i < dd->nc[dim]; i++)
2204 load->sum += load->load[pos++];
2205 load->max = std::max(load->max, load->load[pos]);
2207 if (isDlbOn(dd->comm))
2209 if (rowMaster->dlbIsLimited)
2211 /* This direction could not be load balanced properly,
2212 * therefore we need to use the maximum iso the average load.
2214 load->sum_m = std::max(load->sum_m, load->load[pos]);
2218 load->sum_m += load->load[pos];
2221 load->cvol_min = std::min(load->cvol_min, load->load[pos]);
2225 load->flags = static_cast<int>(load->load[pos++] + 0.5);
2229 rowMaster->bounds[i].cellFracLowerMax = load->load[pos++];
2230 rowMaster->bounds[i].cellFracUpperMin = load->load[pos++];
2235 load->mdf = std::max(load->mdf, load->load[pos]);
2237 load->pme = std::max(load->pme, load->load[pos]);
2241 if (isDlbOn(comm) && rowMaster->dlbIsLimited)
2243 load->sum_m *= dd->nc[dim];
2244 load->flags |= (1<<d);
2252 comm->nload += dd_load_count(comm);
2253 comm->load_step += comm->cycl[ddCyclStep];
2254 comm->load_sum += comm->load[0].sum;
2255 comm->load_max += comm->load[0].max;
2258 for (int d = 0; d < dd->ndim; d++)
2260 if (comm->load[0].flags & (1<<d))
2262 comm->load_lim[d]++;
2268 comm->load_mdf += comm->load[0].mdf;
2269 comm->load_pme += comm->load[0].pme;
2273 wallcycle_stop(wcycle, ewcDDCOMMLOAD);
2277 fprintf(debug, "get_load_distribution finished\n");
2281 static float dd_force_load_fraction(gmx_domdec_t *dd)
2283 /* Return the relative performance loss on the total run time
2284 * due to the force calculation load imbalance.
2286 if (dd->comm->nload > 0 && dd->comm->load_step > 0)
2288 return dd->comm->load_sum/(dd->comm->load_step*dd->nnodes);
2296 static float dd_force_imb_perf_loss(gmx_domdec_t *dd)
2298 /* Return the relative performance loss on the total run time
2299 * due to the force calculation load imbalance.
2301 if (dd->comm->nload > 0 && dd->comm->load_step > 0)
2304 (dd->comm->load_max*dd->nnodes - dd->comm->load_sum)/
2305 (dd->comm->load_step*dd->nnodes);
2313 static void print_dd_load_av(FILE *fplog, gmx_domdec_t *dd)
2315 gmx_domdec_comm_t *comm = dd->comm;
2317 /* Only the master rank prints loads and only if we measured loads */
2318 if (!DDMASTER(dd) || comm->nload == 0)
2324 int numPpRanks = dd->nnodes;
2325 int numPmeRanks = (dd->pme_nodeid >= 0) ? comm->npmenodes : 0;
2326 int numRanks = numPpRanks + numPmeRanks;
2327 float lossFraction = 0;
2329 /* Print the average load imbalance and performance loss */
2330 if (dd->nnodes > 1 && comm->load_sum > 0)
2332 float imbalance = comm->load_max*numPpRanks/comm->load_sum - 1;
2333 lossFraction = dd_force_imb_perf_loss(dd);
2335 std::string msg = "\n Dynamic load balancing report:\n";
2336 std::string dlbStateStr;
2338 switch (dd->comm->dlbState)
2341 dlbStateStr = "DLB was off during the run per user request.";
2343 case edlbsOffForever:
2344 /* Currectly this can happen due to performance loss observed, cell size
2345 * limitations or incompatibility with other settings observed during
2346 * determineInitialDlbState(). */
2347 dlbStateStr = "DLB got disabled because it was unsuitable to use.";
2349 case edlbsOffCanTurnOn:
2350 dlbStateStr = "DLB was off during the run due to low measured imbalance.";
2352 case edlbsOffTemporarilyLocked:
2353 dlbStateStr = "DLB was locked at the end of the run due to unfinished PP-PME balancing.";
2355 case edlbsOnCanTurnOff:
2356 dlbStateStr = "DLB was turned on during the run due to measured imbalance.";
2359 dlbStateStr = "DLB was permanently on during the run per user request.";
2362 GMX_ASSERT(false, "Undocumented DLB state");
2365 msg += " " + dlbStateStr + "\n";
2366 msg += gmx::formatString(" Average load imbalance: %.1f%%.\n", imbalance*100);
2367 msg += gmx::formatString(" The balanceable part of the MD step is %d%%, load imbalance is computed from this.\n",
2368 static_cast<int>(dd_force_load_fraction(dd)*100 + 0.5));
2369 msg += gmx::formatString(" Part of the total run time spent waiting due to load imbalance: %.1f%%.\n",
2371 fprintf(fplog, "%s", msg.c_str());
2372 fprintf(stderr, "%s", msg.c_str());
2375 /* Print during what percentage of steps the load balancing was limited */
2376 bool dlbWasLimited = false;
2379 sprintf(buf, " Steps where the load balancing was limited by -rdd, -rcon and/or -dds:");
2380 for (int d = 0; d < dd->ndim; d++)
2382 int limitPercentage = (200*comm->load_lim[d] + 1)/(2*comm->nload);
2383 sprintf(buf+strlen(buf), " %c %d %%",
2384 dim2char(dd->dim[d]), limitPercentage);
2385 if (limitPercentage >= 50)
2387 dlbWasLimited = true;
2390 sprintf(buf + strlen(buf), "\n");
2391 fprintf(fplog, "%s", buf);
2392 fprintf(stderr, "%s", buf);
2395 /* Print the performance loss due to separate PME - PP rank imbalance */
2396 float lossFractionPme = 0;
2397 if (numPmeRanks > 0 && comm->load_mdf > 0 && comm->load_step > 0)
2399 float pmeForceRatio = comm->load_pme/comm->load_mdf;
2400 lossFractionPme = (comm->load_pme - comm->load_mdf)/comm->load_step;
2401 if (lossFractionPme <= 0)
2403 lossFractionPme *= numPmeRanks/static_cast<float>(numRanks);
2407 lossFractionPme *= numPpRanks/static_cast<float>(numRanks);
2409 sprintf(buf, " Average PME mesh/force load: %5.3f\n", pmeForceRatio);
2410 fprintf(fplog, "%s", buf);
2411 fprintf(stderr, "%s", buf);
2412 sprintf(buf, " Part of the total run time spent waiting due to PP/PME imbalance: %.1f %%\n", std::fabs(lossFractionPme)*100);
2413 fprintf(fplog, "%s", buf);
2414 fprintf(stderr, "%s", buf);
2416 fprintf(fplog, "\n");
2417 fprintf(stderr, "\n");
2419 if (lossFraction >= DD_PERF_LOSS_WARN)
2422 "NOTE: %.1f %% of the available CPU time was lost due to load imbalance\n"
2423 " in the domain decomposition.\n", lossFraction*100);
2426 sprintf(buf+strlen(buf), " You might want to use dynamic load balancing (option -dlb.)\n");
2428 else if (dlbWasLimited)
2430 sprintf(buf+strlen(buf), " You might want to decrease the cell size limit (options -rdd, -rcon and/or -dds).\n");
2432 fprintf(fplog, "%s\n", buf);
2433 fprintf(stderr, "%s\n", buf);
2435 if (numPmeRanks > 0 && std::fabs(lossFractionPme) >= DD_PERF_LOSS_WARN)
2438 "NOTE: %.1f %% performance was lost because the PME ranks\n"
2439 " had %s work to do than the PP ranks.\n"
2440 " You might want to %s the number of PME ranks\n"
2441 " or %s the cut-off and the grid spacing.\n",
2442 std::fabs(lossFractionPme*100),
2443 (lossFractionPme < 0) ? "less" : "more",
2444 (lossFractionPme < 0) ? "decrease" : "increase",
2445 (lossFractionPme < 0) ? "decrease" : "increase");
2446 fprintf(fplog, "%s\n", buf);
2447 fprintf(stderr, "%s\n", buf);
2451 static float dd_vol_min(gmx_domdec_t *dd)
2453 return dd->comm->load[0].cvol_min*dd->nnodes;
2456 static gmx_bool dd_load_flags(gmx_domdec_t *dd)
2458 return dd->comm->load[0].flags;
2461 static float dd_f_imbal(gmx_domdec_t *dd)
2463 if (dd->comm->load[0].sum > 0)
2465 return dd->comm->load[0].max*dd->nnodes/dd->comm->load[0].sum - 1.0f;
2469 /* Something is wrong in the cycle counting, report no load imbalance */
2474 float dd_pme_f_ratio(gmx_domdec_t *dd)
2476 /* Should only be called on the DD master rank */
2477 assert(DDMASTER(dd));
2479 if (dd->comm->load[0].mdf > 0 && dd->comm->cycl_n[ddCyclPME] > 0)
2481 return dd->comm->load[0].pme/dd->comm->load[0].mdf;
2489 static void dd_print_load(FILE *fplog, gmx_domdec_t *dd, int64_t step)
2494 flags = dd_load_flags(dd);
2498 "DD load balancing is limited by minimum cell size in dimension");
2499 for (d = 0; d < dd->ndim; d++)
2503 fprintf(fplog, " %c", dim2char(dd->dim[d]));
2506 fprintf(fplog, "\n");
2508 fprintf(fplog, "DD step %s", gmx_step_str(step, buf));
2509 if (isDlbOn(dd->comm))
2511 fprintf(fplog, " vol min/aver %5.3f%c",
2512 dd_vol_min(dd), flags ? '!' : ' ');
2516 fprintf(fplog, " load imb.: force %4.1f%%", dd_f_imbal(dd)*100);
2518 if (dd->comm->cycl_n[ddCyclPME])
2520 fprintf(fplog, " pme mesh/force %5.3f", dd_pme_f_ratio(dd));
2522 fprintf(fplog, "\n\n");
2525 static void dd_print_load_verbose(gmx_domdec_t *dd)
2527 if (isDlbOn(dd->comm))
2529 fprintf(stderr, "vol %4.2f%c ",
2530 dd_vol_min(dd), dd_load_flags(dd) ? '!' : ' ');
2534 fprintf(stderr, "imb F %2d%% ", static_cast<int>(dd_f_imbal(dd)*100+0.5));
2536 if (dd->comm->cycl_n[ddCyclPME])
2538 fprintf(stderr, "pme/F %4.2f ", dd_pme_f_ratio(dd));
2543 static void make_load_communicator(gmx_domdec_t *dd, int dim_ind, ivec loc)
2548 gmx_bool bPartOfGroup = FALSE;
2550 dim = dd->dim[dim_ind];
2551 copy_ivec(loc, loc_c);
2552 for (i = 0; i < dd->nc[dim]; i++)
2555 rank = dd_index(dd->nc, loc_c);
2556 if (rank == dd->rank)
2558 /* This process is part of the group */
2559 bPartOfGroup = TRUE;
2562 MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank,
2566 dd->comm->mpi_comm_load[dim_ind] = c_row;
2567 if (!isDlbDisabled(dd->comm))
2569 DDCellsizesWithDlb &cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
2571 if (dd->ci[dim] == dd->master_ci[dim])
2573 /* This is the root process of this row */
2574 cellsizes.rowMaster = gmx::compat::make_unique<RowMaster>();
2576 RowMaster &rowMaster = *cellsizes.rowMaster;
2577 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
2578 rowMaster.oldCellFrac.resize(dd->nc[dim] + 1);
2579 rowMaster.isCellMin.resize(dd->nc[dim]);
2582 rowMaster.bounds.resize(dd->nc[dim]);
2584 rowMaster.buf_ncd.resize(dd->nc[dim]);
2588 /* This is not a root process, we only need to receive cell_f */
2589 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
2592 if (dd->ci[dim] == dd->master_ci[dim])
2594 snew(dd->comm->load[dim_ind].load, dd->nc[dim]*DD_NLOAD_MAX);
2600 void dd_setup_dlb_resource_sharing(t_commrec *cr,
2604 int physicalnode_id_hash;
2606 MPI_Comm mpi_comm_pp_physicalnode;
2608 if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
2610 /* Only ranks with short-ranged tasks (currently) use GPUs.
2611 * If we don't have GPUs assigned, there are no resources to share.
2616 physicalnode_id_hash = gmx_physicalnode_id_hash();
2622 fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
2623 fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n",
2624 dd->rank, physicalnode_id_hash, gpu_id);
2626 /* Split the PP communicator over the physical nodes */
2627 /* TODO: See if we should store this (before), as it's also used for
2628 * for the nodecomm summation.
2630 // TODO PhysicalNodeCommunicator could be extended/used to handle
2631 // the need for per-node per-group communicators.
2632 MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank,
2633 &mpi_comm_pp_physicalnode);
2634 MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank,
2635 &dd->comm->mpi_comm_gpu_shared);
2636 MPI_Comm_free(&mpi_comm_pp_physicalnode);
2637 MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
2641 fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
2644 /* Note that some ranks could share a GPU, while others don't */
2646 if (dd->comm->nrank_gpu_shared == 1)
2648 MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
2651 GMX_UNUSED_VALUE(cr);
2652 GMX_UNUSED_VALUE(gpu_id);
2656 static void make_load_communicators(gmx_domdec_t gmx_unused *dd)
2659 int dim0, dim1, i, j;
2664 fprintf(debug, "Making load communicators\n");
2667 snew(dd->comm->load, std::max(dd->ndim, 1));
2668 snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
2676 make_load_communicator(dd, 0, loc);
2680 for (i = 0; i < dd->nc[dim0]; i++)
2683 make_load_communicator(dd, 1, loc);
2689 for (i = 0; i < dd->nc[dim0]; i++)
2693 for (j = 0; j < dd->nc[dim1]; j++)
2696 make_load_communicator(dd, 2, loc);
2703 fprintf(debug, "Finished making load communicators\n");
2708 /*! \brief Sets up the relation between neighboring domains and zones */
2709 static void setup_neighbor_relations(gmx_domdec_t *dd)
2711 int d, dim, i, j, m;
2713 gmx_domdec_zones_t *zones;
2714 gmx_domdec_ns_ranges_t *izone;
2716 for (d = 0; d < dd->ndim; d++)
2719 copy_ivec(dd->ci, tmp);
2720 tmp[dim] = (tmp[dim] + 1) % dd->nc[dim];
2721 dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
2722 copy_ivec(dd->ci, tmp);
2723 tmp[dim] = (tmp[dim] - 1 + dd->nc[dim]) % dd->nc[dim];
2724 dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
2727 fprintf(debug, "DD rank %d neighbor ranks in dir %d are + %d - %d\n",
2730 dd->neighbor[d][1]);
2734 int nzone = (1 << dd->ndim);
2735 int nizone = (1 << std::max(dd->ndim - 1, 0));
2736 assert(nizone >= 1 && nizone <= DD_MAXIZONE);
2738 zones = &dd->comm->zones;
2740 for (i = 0; i < nzone; i++)
2743 clear_ivec(zones->shift[i]);
2744 for (d = 0; d < dd->ndim; d++)
2746 zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
2751 for (i = 0; i < nzone; i++)
2753 for (d = 0; d < DIM; d++)
2755 s[d] = dd->ci[d] - zones->shift[i][d];
2760 else if (s[d] >= dd->nc[d])
2766 zones->nizone = nizone;
2767 for (i = 0; i < zones->nizone; i++)
2769 assert(ddNonbondedZonePairRanges[i][0] == i);
2771 izone = &zones->izone[i];
2772 /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
2773 * j-zones up to nzone.
2775 izone->j0 = std::min(ddNonbondedZonePairRanges[i][1], nzone);
2776 izone->j1 = std::min(ddNonbondedZonePairRanges[i][2], nzone);
2777 for (dim = 0; dim < DIM; dim++)
2779 if (dd->nc[dim] == 1)
2781 /* All shifts should be allowed */
2782 izone->shift0[dim] = -1;
2783 izone->shift1[dim] = 1;
2787 /* Determine the min/max j-zone shift wrt the i-zone */
2788 izone->shift0[dim] = 1;
2789 izone->shift1[dim] = -1;
2790 for (j = izone->j0; j < izone->j1; j++)
2792 int shift_diff = zones->shift[j][dim] - zones->shift[i][dim];
2793 if (shift_diff < izone->shift0[dim])
2795 izone->shift0[dim] = shift_diff;
2797 if (shift_diff > izone->shift1[dim])
2799 izone->shift1[dim] = shift_diff;
2806 if (!isDlbDisabled(dd->comm))
2808 dd->comm->cellsizesWithDlb.resize(dd->ndim);
2811 if (dd->comm->bRecordLoad)
2813 make_load_communicators(dd);
2817 static void make_pp_communicator(FILE *fplog,
2819 t_commrec gmx_unused *cr,
2820 int gmx_unused reorder)
2823 gmx_domdec_comm_t *comm;
2830 if (comm->bCartesianPP)
2832 /* Set up cartesian communication for the particle-particle part */
2835 fprintf(fplog, "Will use a Cartesian communicator: %d x %d x %d\n",
2836 dd->nc[XX], dd->nc[YY], dd->nc[ZZ]);
2839 for (int i = 0; i < DIM; i++)
2843 MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->nc, periods, reorder,
2845 /* We overwrite the old communicator with the new cartesian one */
2846 cr->mpi_comm_mygroup = comm_cart;
2849 dd->mpi_comm_all = cr->mpi_comm_mygroup;
2850 MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
2852 if (comm->bCartesianPP_PME)
2854 /* Since we want to use the original cartesian setup for sim,
2855 * and not the one after split, we need to make an index.
2857 snew(comm->ddindex2ddnodeid, dd->nnodes);
2858 comm->ddindex2ddnodeid[dd_index(dd->nc, dd->ci)] = dd->rank;
2859 gmx_sumi(dd->nnodes, comm->ddindex2ddnodeid, cr);
2860 /* Get the rank of the DD master,
2861 * above we made sure that the master node is a PP node.
2871 MPI_Allreduce(&rank, &dd->masterrank, 1, MPI_INT, MPI_SUM, dd->mpi_comm_all);
2873 else if (comm->bCartesianPP)
2875 if (cr->npmenodes == 0)
2877 /* The PP communicator is also
2878 * the communicator for this simulation
2880 cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
2882 cr->nodeid = dd->rank;
2884 MPI_Cart_coords(dd->mpi_comm_all, dd->rank, DIM, dd->ci);
2886 /* We need to make an index to go from the coordinates
2887 * to the nodeid of this simulation.
2889 snew(comm->ddindex2simnodeid, dd->nnodes);
2890 snew(buf, dd->nnodes);
2891 if (thisRankHasDuty(cr, DUTY_PP))
2893 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
2895 /* Communicate the ddindex to simulation nodeid index */
2896 MPI_Allreduce(buf, comm->ddindex2simnodeid, dd->nnodes, MPI_INT, MPI_SUM,
2897 cr->mpi_comm_mysim);
2900 /* Determine the master coordinates and rank.
2901 * The DD master should be the same node as the master of this sim.
2903 for (int i = 0; i < dd->nnodes; i++)
2905 if (comm->ddindex2simnodeid[i] == 0)
2907 ddindex2xyz(dd->nc, i, dd->master_ci);
2908 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
2913 fprintf(debug, "The master rank is %d\n", dd->masterrank);
2918 /* No Cartesian communicators */
2919 /* We use the rank in dd->comm->all as DD index */
2920 ddindex2xyz(dd->nc, dd->rank, dd->ci);
2921 /* The simulation master nodeid is 0, so the DD master rank is also 0 */
2923 clear_ivec(dd->master_ci);
2930 "Domain decomposition rank %d, coordinates %d %d %d\n\n",
2931 dd->rank, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
2936 "Domain decomposition rank %d, coordinates %d %d %d\n\n",
2937 dd->rank, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
2941 static void receive_ddindex2simnodeid(gmx_domdec_t *dd,
2945 gmx_domdec_comm_t *comm = dd->comm;
2947 if (!comm->bCartesianPP_PME && comm->bCartesianPP)
2950 snew(comm->ddindex2simnodeid, dd->nnodes);
2951 snew(buf, dd->nnodes);
2952 if (thisRankHasDuty(cr, DUTY_PP))
2954 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
2956 /* Communicate the ddindex to simulation nodeid index */
2957 MPI_Allreduce(buf, comm->ddindex2simnodeid, dd->nnodes, MPI_INT, MPI_SUM,
2958 cr->mpi_comm_mysim);
2962 GMX_UNUSED_VALUE(dd);
2963 GMX_UNUSED_VALUE(cr);
2967 static void split_communicator(FILE *fplog, t_commrec *cr, gmx_domdec_t *dd,
2968 DdRankOrder gmx_unused rankOrder,
2969 int gmx_unused reorder)
2971 gmx_domdec_comm_t *comm;
2980 if (comm->bCartesianPP)
2982 for (i = 1; i < DIM; i++)
2984 bDiv[i] = ((cr->npmenodes*dd->nc[i]) % (dd->nnodes) == 0);
2986 if (bDiv[YY] || bDiv[ZZ])
2988 comm->bCartesianPP_PME = TRUE;
2989 /* If we have 2D PME decomposition, which is always in x+y,
2990 * we stack the PME only nodes in z.
2991 * Otherwise we choose the direction that provides the thinnest slab
2992 * of PME only nodes as this will have the least effect
2993 * on the PP communication.
2994 * But for the PME communication the opposite might be better.
2996 if (bDiv[ZZ] && (comm->npmenodes_y > 1 ||
2998 dd->nc[YY] > dd->nc[ZZ]))
3000 comm->cartpmedim = ZZ;
3004 comm->cartpmedim = YY;
3006 comm->ntot[comm->cartpmedim]
3007 += (cr->npmenodes*dd->nc[comm->cartpmedim])/dd->nnodes;
3011 fprintf(fplog, "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or nx*nz (%d*%d)\n", cr->npmenodes, dd->nc[XX], dd->nc[YY], dd->nc[XX], dd->nc[ZZ]);
3013 "Will not use a Cartesian communicator for PP <-> PME\n\n");
3017 if (comm->bCartesianPP_PME)
3025 fprintf(fplog, "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d\n", comm->ntot[XX], comm->ntot[YY], comm->ntot[ZZ]);
3028 for (i = 0; i < DIM; i++)
3032 MPI_Cart_create(cr->mpi_comm_mysim, DIM, comm->ntot, periods, reorder,
3034 MPI_Comm_rank(comm_cart, &rank);
3035 if (MASTER(cr) && rank != 0)
3037 gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
3040 /* With this assigment we loose the link to the original communicator
3041 * which will usually be MPI_COMM_WORLD, unless have multisim.
3043 cr->mpi_comm_mysim = comm_cart;
3044 cr->sim_nodeid = rank;
3046 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, dd->ci);
3050 fprintf(fplog, "Cartesian rank %d, coordinates %d %d %d\n\n",
3051 cr->sim_nodeid, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
3054 if (dd->ci[comm->cartpmedim] < dd->nc[comm->cartpmedim])
3058 if (cr->npmenodes == 0 ||
3059 dd->ci[comm->cartpmedim] >= dd->nc[comm->cartpmedim])
3061 cr->duty = DUTY_PME;
3064 /* Split the sim communicator into PP and PME only nodes */
3065 MPI_Comm_split(cr->mpi_comm_mysim,
3066 getThisRankDuties(cr),
3067 dd_index(comm->ntot, dd->ci),
3068 &cr->mpi_comm_mygroup);
3075 case DdRankOrder::pp_pme:
3078 fprintf(fplog, "Order of the ranks: PP first, PME last\n");
3081 case DdRankOrder::interleave:
3082 /* Interleave the PP-only and PME-only ranks */
3085 fprintf(fplog, "Interleaving PP and PME ranks\n");
3087 comm->pmenodes = dd_interleaved_pme_ranks(dd);
3089 case DdRankOrder::cartesian:
3092 gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(rankOrder));
3095 if (dd_simnode2pmenode(dd, cr, cr->sim_nodeid) == -1)
3097 cr->duty = DUTY_PME;
3104 /* Split the sim communicator into PP and PME only nodes */
3105 MPI_Comm_split(cr->mpi_comm_mysim,
3106 getThisRankDuties(cr),
3108 &cr->mpi_comm_mygroup);
3109 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
3115 fprintf(fplog, "This rank does only %s work.\n\n",
3116 thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
3120 /*! \brief Generates the MPI communicators for domain decomposition */
3121 static void make_dd_communicators(FILE *fplog, t_commrec *cr,
3122 gmx_domdec_t *dd, DdRankOrder ddRankOrder)
3124 gmx_domdec_comm_t *comm;
3129 copy_ivec(dd->nc, comm->ntot);
3131 comm->bCartesianPP = (ddRankOrder == DdRankOrder::cartesian);
3132 comm->bCartesianPP_PME = FALSE;
3134 /* Reorder the nodes by default. This might change the MPI ranks.
3135 * Real reordering is only supported on very few architectures,
3136 * Blue Gene is one of them.
3138 CartReorder = (getenv("GMX_NO_CART_REORDER") == nullptr);
3140 if (cr->npmenodes > 0)
3142 /* Split the communicator into a PP and PME part */
3143 split_communicator(fplog, cr, dd, ddRankOrder, CartReorder);
3144 if (comm->bCartesianPP_PME)
3146 /* We (possibly) reordered the nodes in split_communicator,
3147 * so it is no longer required in make_pp_communicator.
3149 CartReorder = FALSE;
3154 /* All nodes do PP and PME */
3156 /* We do not require separate communicators */
3157 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
3161 if (thisRankHasDuty(cr, DUTY_PP))
3163 /* Copy or make a new PP communicator */
3164 make_pp_communicator(fplog, dd, cr, CartReorder);
3168 receive_ddindex2simnodeid(dd, cr);
3171 if (!thisRankHasDuty(cr, DUTY_PME))
3173 /* Set up the commnuication to our PME node */
3174 dd->pme_nodeid = dd_simnode2pmenode(dd, cr, cr->sim_nodeid);
3175 dd->pme_receive_vir_ener = receive_vir_ener(dd, cr);
3178 fprintf(debug, "My pme_nodeid %d receive ener %s\n",
3179 dd->pme_nodeid, gmx::boolToString(dd->pme_receive_vir_ener));
3184 dd->pme_nodeid = -1;
3189 dd->ma = gmx::compat::make_unique<AtomDistribution>(dd->nc,
3191 comm->cgs_gl.index[comm->cgs_gl.nr]);
3195 static real *get_slb_frac(FILE *fplog, const char *dir, int nc, const char *size_string)
3197 real *slb_frac, tot;
3202 if (nc > 1 && size_string != nullptr)
3206 fprintf(fplog, "Using static load balancing for the %s direction\n",
3211 for (i = 0; i < nc; i++)
3214 sscanf(size_string, "%20lf%n", &dbl, &n);
3217 gmx_fatal(FARGS, "Incorrect or not enough DD cell size entries for direction %s: '%s'", dir, size_string);
3226 fprintf(fplog, "Relative cell sizes:");
3228 for (i = 0; i < nc; i++)
3233 fprintf(fplog, " %5.3f", slb_frac[i]);
3238 fprintf(fplog, "\n");
3245 static int multi_body_bondeds_count(const gmx_mtop_t *mtop)
3248 gmx_mtop_ilistloop_t iloop;
3252 iloop = gmx_mtop_ilistloop_init(mtop);
3253 while (gmx_mtop_ilistloop_next(iloop, &il, &nmol))
3255 for (ftype = 0; ftype < F_NRE; ftype++)
3257 if ((interaction_function[ftype].flags & IF_BOND) &&
3260 n += nmol*il[ftype].nr/(1 + NRAL(ftype));
3268 static int dd_getenv(FILE *fplog, const char *env_var, int def)
3274 val = getenv(env_var);
3277 if (sscanf(val, "%20d", &nst) <= 0)
3283 fprintf(fplog, "Found env.var. %s = %s, using value %d\n",
3291 static void dd_warning(const t_commrec *cr, FILE *fplog, const char *warn_string)
3295 fprintf(stderr, "\n%s\n", warn_string);
3299 fprintf(fplog, "\n%s\n", warn_string);
3303 static void check_dd_restrictions(t_commrec *cr, const gmx_domdec_t *dd,
3304 const t_inputrec *ir, FILE *fplog)
3306 if (ir->ePBC == epbcSCREW &&
3307 (dd->nc[XX] == 1 || dd->nc[YY] > 1 || dd->nc[ZZ] > 1))
3309 gmx_fatal(FARGS, "With pbc=%s can only do domain decomposition in the x-direction", epbc_names[ir->ePBC]);
3312 if (ir->ns_type == ensSIMPLE)
3314 gmx_fatal(FARGS, "Domain decomposition does not support simple neighbor searching, use grid searching or run with one MPI rank");
3317 if (ir->nstlist == 0)
3319 gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
3322 if (ir->comm_mode == ecmANGULAR && ir->ePBC != epbcNONE)
3324 dd_warning(cr, fplog, "comm-mode angular will give incorrect results when the comm group partially crosses a periodic boundary");
3328 static real average_cellsize_min(gmx_domdec_t *dd, gmx_ddbox_t *ddbox)
3333 r = ddbox->box_size[XX];
3334 for (di = 0; di < dd->ndim; di++)
3337 /* Check using the initial average cell size */
3338 r = std::min(r, ddbox->box_size[d]*ddbox->skew_fac[d]/dd->nc[d]);
3344 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
3346 static int forceDlbOffOrBail(int cmdlineDlbState,
3347 const std::string &reasonStr,
3351 std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
3352 std::string dlbDisableNote = "NOTE: disabling dynamic load balancing as ";
3354 if (cmdlineDlbState == edlbsOnUser)
3356 gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
3358 else if (cmdlineDlbState == edlbsOffCanTurnOn)
3360 dd_warning(cr, fplog, (dlbDisableNote + reasonStr + "\n").c_str());
3362 return edlbsOffForever;
3365 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
3367 * This function parses the parameters of "-dlb" command line option setting
3368 * corresponding state values. Then it checks the consistency of the determined
3369 * state with other run parameters and settings. As a result, the initial state
3370 * may be altered or an error may be thrown if incompatibility of options is detected.
3372 * \param [in] fplog Pointer to mdrun log file.
3373 * \param [in] cr Pointer to MPI communication object.
3374 * \param [in] dlbOption Enum value for the DLB option.
3375 * \param [in] bRecordLoad True if the load balancer is recording load information.
3376 * \param [in] mdrunOptions Options for mdrun.
3377 * \param [in] ir Pointer mdrun to input parameters.
3378 * \returns DLB initial/startup state.
3380 static int determineInitialDlbState(FILE *fplog, t_commrec *cr,
3381 DlbOption dlbOption, gmx_bool bRecordLoad,
3382 const MdrunOptions &mdrunOptions,
3383 const t_inputrec *ir)
3385 int dlbState = edlbsOffCanTurnOn;
3389 case DlbOption::turnOnWhenUseful: dlbState = edlbsOffCanTurnOn; break;
3390 case DlbOption::no: dlbState = edlbsOffUser; break;
3391 case DlbOption::yes: dlbState = edlbsOnUser; break;
3392 default: gmx_incons("Invalid dlbOption enum value");
3395 /* Reruns don't support DLB: bail or override auto mode */
3396 if (mdrunOptions.rerun)
3398 std::string reasonStr = "it is not supported in reruns.";
3399 return forceDlbOffOrBail(dlbState, reasonStr, cr, fplog);
3402 /* Unsupported integrators */
3403 if (!EI_DYNAMICS(ir->eI))
3405 auto reasonStr = gmx::formatString("it is only supported with dynamics, not with integrator '%s'.", EI(ir->eI));
3406 return forceDlbOffOrBail(dlbState, reasonStr, cr, fplog);
3409 /* Without cycle counters we can't time work to balance on */
3412 std::string reasonStr = "cycle counters unsupported or not enabled in the operating system kernel.";
3413 return forceDlbOffOrBail(dlbState, reasonStr, cr, fplog);
3416 if (mdrunOptions.reproducible)
3418 std::string reasonStr = "you started a reproducible run.";
3423 case edlbsOffForever:
3424 GMX_RELEASE_ASSERT(false, "edlbsOffForever is not a valid initial state");
3426 case edlbsOffCanTurnOn:
3427 return forceDlbOffOrBail(dlbState, reasonStr, cr, fplog);
3428 case edlbsOnCanTurnOff:
3429 GMX_RELEASE_ASSERT(false, "edlbsOffCanTurnOff is not a valid initial state");
3432 return forceDlbOffOrBail(dlbState, reasonStr + " In load balanced runs binary reproducibility cannot be ensured.", cr, fplog);
3434 gmx_fatal(FARGS, "Death horror: undefined case (%d) for load balancing choice", dlbState);
3441 static void set_dd_dim(FILE *fplog, gmx_domdec_t *dd)
3444 if (getenv("GMX_DD_ORDER_ZYX") != nullptr)
3446 /* Decomposition order z,y,x */
3449 fprintf(fplog, "Using domain decomposition order z, y, x\n");
3451 for (int dim = DIM-1; dim >= 0; dim--)
3453 if (dd->nc[dim] > 1)
3455 dd->dim[dd->ndim++] = dim;
3461 /* Decomposition order x,y,z */
3462 for (int dim = 0; dim < DIM; dim++)
3464 if (dd->nc[dim] > 1)
3466 dd->dim[dd->ndim++] = dim;
3473 /* Set dim[0] to avoid extra checks on ndim in several places */
3478 static gmx_domdec_comm_t *init_dd_comm()
3480 gmx_domdec_comm_t *comm = new gmx_domdec_comm_t;
3482 comm->n_load_have = 0;
3483 comm->n_load_collect = 0;
3485 comm->haveTurnedOffDlb = false;
3487 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
3489 comm->sum_nat[i] = 0;
3493 comm->load_step = 0;
3496 clear_ivec(comm->load_lim);
3500 /* This should be replaced by a unique pointer */
3501 comm->balanceRegion = ddBalanceRegionAllocate();
3506 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
3507 static void set_dd_limits_and_grid(FILE *fplog, t_commrec *cr, gmx_domdec_t *dd,
3508 const DomdecOptions &options,
3509 const MdrunOptions &mdrunOptions,
3510 const gmx_mtop_t *mtop,
3511 const t_inputrec *ir,
3513 gmx::ArrayRef<const gmx::RVec> xGlobal,
3517 real r_bonded_limit = -1;
3518 const real tenPercentMargin = 1.1;
3519 gmx_domdec_comm_t *comm = dd->comm;
3521 dd->npbcdim = ePBC2npbcdim(ir->ePBC);
3522 dd->bScrewPBC = (ir->ePBC == epbcSCREW);
3524 dd->pme_recv_f_alloc = 0;
3525 dd->pme_recv_f_buf = nullptr;
3527 /* Initialize to GPU share count to 0, might change later */
3528 comm->nrank_gpu_shared = 0;
3530 comm->dlbState = determineInitialDlbState(fplog, cr, options.dlbOption, comm->bRecordLoad, mdrunOptions, ir);
3531 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
3532 /* To consider turning DLB on after 2*nstlist steps we need to check
3533 * at partitioning count 3. Thus we need to increase the first count by 2.
3535 comm->ddPartioningCountFirstDlbOff += 2;
3539 fprintf(fplog, "Dynamic load balancing: %s\n",
3540 edlbs_names[comm->dlbState]);
3542 comm->bPMELoadBalDLBLimits = FALSE;
3544 /* Allocate the charge group/atom sorting struct */
3545 comm->sort = gmx::compat::make_unique<gmx_domdec_sort_t>();
3547 comm->bCGs = (ncg_mtop(mtop) < mtop->natoms);
3549 comm->bInterCGBondeds = ((ncg_mtop(mtop) > gmx_mtop_num_molecules(*mtop)) ||
3550 mtop->bIntermolecularInteractions);
3551 if (comm->bInterCGBondeds)
3553 comm->bInterCGMultiBody = (multi_body_bondeds_count(mtop) > 0);
3557 comm->bInterCGMultiBody = FALSE;
3560 dd->bInterCGcons = gmx::inter_charge_group_constraints(*mtop);
3561 dd->bInterCGsettles = gmx::inter_charge_group_settles(*mtop);
3565 /* Set the cut-off to some very large value,
3566 * so we don't need if statements everywhere in the code.
3567 * We use sqrt, since the cut-off is squared in some places.
3569 comm->cutoff = GMX_CUTOFF_INF;
3573 comm->cutoff = ir->rlist;
3575 comm->cutoff_mbody = 0;
3577 comm->cellsize_limit = 0;
3578 comm->bBondComm = FALSE;
3580 /* Atoms should be able to move by up to half the list buffer size (if > 0)
3581 * within nstlist steps. Since boundaries are allowed to displace by half
3582 * a cell size, DD cells should be at least the size of the list buffer.
3584 comm->cellsize_limit = std::max(comm->cellsize_limit,
3585 ir->rlist - std::max(ir->rvdw, ir->rcoulomb));
3587 if (comm->bInterCGBondeds)
3589 if (options.minimumCommunicationRange > 0)
3591 comm->cutoff_mbody = options.minimumCommunicationRange;
3592 if (options.useBondedCommunication)
3594 comm->bBondComm = (comm->cutoff_mbody > comm->cutoff);
3598 comm->cutoff = std::max(comm->cutoff, comm->cutoff_mbody);
3600 r_bonded_limit = comm->cutoff_mbody;
3602 else if (ir->bPeriodicMols)
3604 /* Can not easily determine the required cut-off */
3605 dd_warning(cr, fplog, "NOTE: Periodic molecules are present in this system. Because of this, the domain decomposition algorithm cannot easily determine the minimum cell size that it requires for treating bonded interactions. Instead, domain decomposition will assume that half the non-bonded cut-off will be a suitable lower bound.\n");
3606 comm->cutoff_mbody = comm->cutoff/2;
3607 r_bonded_limit = comm->cutoff_mbody;
3615 dd_bonded_cg_distance(fplog, mtop, ir, as_rvec_array(xGlobal.data()), box,
3616 options.checkBondedInteractions,
3619 gmx_bcast(sizeof(r_2b), &r_2b, cr);
3620 gmx_bcast(sizeof(r_mb), &r_mb, cr);
3622 /* We use an initial margin of 10% for the minimum cell size,
3623 * except when we are just below the non-bonded cut-off.
3625 if (options.useBondedCommunication)
3627 if (std::max(r_2b, r_mb) > comm->cutoff)
3629 r_bonded = std::max(r_2b, r_mb);
3630 r_bonded_limit = tenPercentMargin*r_bonded;
3631 comm->bBondComm = TRUE;
3636 r_bonded_limit = std::min(tenPercentMargin*r_bonded, comm->cutoff);
3638 /* We determine cutoff_mbody later */
3642 /* No special bonded communication,
3643 * simply increase the DD cut-off.
3645 r_bonded_limit = tenPercentMargin*std::max(r_2b, r_mb);
3646 comm->cutoff_mbody = r_bonded_limit;
3647 comm->cutoff = std::max(comm->cutoff, comm->cutoff_mbody);
3653 "Minimum cell size due to bonded interactions: %.3f nm\n",
3656 comm->cellsize_limit = std::max(comm->cellsize_limit, r_bonded_limit);
3660 if (dd->bInterCGcons && options.constraintCommunicationRange <= 0)
3662 /* There is a cell size limit due to the constraints (P-LINCS) */
3663 rconstr = gmx::constr_r_max(fplog, mtop, ir);
3667 "Estimated maximum distance required for P-LINCS: %.3f nm\n",
3669 if (rconstr > comm->cellsize_limit)
3671 fprintf(fplog, "This distance will limit the DD cell size, you can override this with -rcon\n");
3675 else if (options.constraintCommunicationRange > 0 && fplog)
3677 /* Here we do not check for dd->bInterCGcons,
3678 * because one can also set a cell size limit for virtual sites only
3679 * and at this point we don't know yet if there are intercg v-sites.
3682 "User supplied maximum distance required for P-LINCS: %.3f nm\n",
3683 options.constraintCommunicationRange);
3684 rconstr = options.constraintCommunicationRange;
3686 comm->cellsize_limit = std::max(comm->cellsize_limit, rconstr);
3688 comm->cgs_gl = gmx_mtop_global_cgs(mtop);
3690 if (options.numCells[XX] > 0)
3692 copy_ivec(options.numCells, dd->nc);
3693 set_dd_dim(fplog, dd);
3694 set_ddbox_cr(cr, &dd->nc, ir, box, xGlobal, ddbox);
3696 if (options.numPmeRanks >= 0)
3698 cr->npmenodes = options.numPmeRanks;
3702 /* When the DD grid is set explicitly and -npme is set to auto,
3703 * don't use PME ranks. We check later if the DD grid is
3704 * compatible with the total number of ranks.
3709 real acs = average_cellsize_min(dd, ddbox);
3710 if (acs < comm->cellsize_limit)
3714 fprintf(fplog, "ERROR: The initial cell size (%f) is smaller than the cell size limit (%f)\n", acs, comm->cellsize_limit);
3716 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
3717 "The initial cell size (%f) is smaller than the cell size limit (%f), change options -dd, -rdd or -rcon, see the log file for details",
3718 acs, comm->cellsize_limit);
3723 set_ddbox_cr(cr, nullptr, ir, box, xGlobal, ddbox);
3725 /* We need to choose the optimal DD grid and possibly PME nodes */
3727 dd_choose_grid(fplog, cr, dd, ir, mtop, box, ddbox,
3728 options.numPmeRanks,
3729 !isDlbDisabled(comm),
3731 comm->cellsize_limit, comm->cutoff,
3732 comm->bInterCGBondeds);
3734 if (dd->nc[XX] == 0)
3737 gmx_bool bC = (dd->bInterCGcons && rconstr > r_bonded_limit);
3738 sprintf(buf, "Change the number of ranks or mdrun option %s%s%s",
3739 !bC ? "-rdd" : "-rcon",
3740 comm->dlbState != edlbsOffUser ? " or -dds" : "",
3741 bC ? " or your LINCS settings" : "");
3743 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
3744 "There is no domain decomposition for %d ranks that is compatible with the given box and a minimum cell size of %g nm\n"
3746 "Look in the log file for details on the domain decomposition",
3747 cr->nnodes-cr->npmenodes, limit, buf);
3749 set_dd_dim(fplog, dd);
3755 "Domain decomposition grid %d x %d x %d, separate PME ranks %d\n",
3756 dd->nc[XX], dd->nc[YY], dd->nc[ZZ], cr->npmenodes);
3759 dd->nnodes = dd->nc[XX]*dd->nc[YY]*dd->nc[ZZ];
3760 if (cr->nnodes - dd->nnodes != cr->npmenodes)
3762 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
3763 "The size of the domain decomposition grid (%d) does not match the number of ranks (%d). The total number of ranks is %d",
3764 dd->nnodes, cr->nnodes - cr->npmenodes, cr->nnodes);
3766 if (cr->npmenodes > dd->nnodes)
3768 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
3769 "The number of separate PME ranks (%d) is larger than the number of PP ranks (%d), this is not supported.", cr->npmenodes, dd->nnodes);
3771 if (cr->npmenodes > 0)
3773 comm->npmenodes = cr->npmenodes;
3777 comm->npmenodes = dd->nnodes;
3780 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
3782 /* The following choices should match those
3783 * in comm_cost_est in domdec_setup.c.
3784 * Note that here the checks have to take into account
3785 * that the decomposition might occur in a different order than xyz
3786 * (for instance through the env.var. GMX_DD_ORDER_ZYX),
3787 * in which case they will not match those in comm_cost_est,
3788 * but since that is mainly for testing purposes that's fine.
3790 if (dd->ndim >= 2 && dd->dim[0] == XX && dd->dim[1] == YY &&
3791 comm->npmenodes > dd->nc[XX] && comm->npmenodes % dd->nc[XX] == 0 &&
3792 getenv("GMX_PMEONEDD") == nullptr)
3794 comm->npmedecompdim = 2;
3795 comm->npmenodes_x = dd->nc[XX];
3796 comm->npmenodes_y = comm->npmenodes/comm->npmenodes_x;
3800 /* In case nc is 1 in both x and y we could still choose to
3801 * decompose pme in y instead of x, but we use x for simplicity.
3803 comm->npmedecompdim = 1;
3804 if (dd->dim[0] == YY)
3806 comm->npmenodes_x = 1;
3807 comm->npmenodes_y = comm->npmenodes;
3811 comm->npmenodes_x = comm->npmenodes;
3812 comm->npmenodes_y = 1;
3817 fprintf(fplog, "PME domain decomposition: %d x %d x %d\n",
3818 comm->npmenodes_x, comm->npmenodes_y, 1);
3823 comm->npmedecompdim = 0;
3824 comm->npmenodes_x = 0;
3825 comm->npmenodes_y = 0;
3828 snew(comm->slb_frac, DIM);
3829 if (isDlbDisabled(comm))
3831 comm->slb_frac[XX] = get_slb_frac(fplog, "x", dd->nc[XX], options.cellSizeX);
3832 comm->slb_frac[YY] = get_slb_frac(fplog, "y", dd->nc[YY], options.cellSizeY);
3833 comm->slb_frac[ZZ] = get_slb_frac(fplog, "z", dd->nc[ZZ], options.cellSizeZ);
3836 if (comm->bInterCGBondeds && comm->cutoff_mbody == 0)
3838 if (comm->bBondComm || !isDlbDisabled(comm))
3840 /* Set the bonded communication distance to halfway
3841 * the minimum and the maximum,
3842 * since the extra communication cost is nearly zero.
3844 real acs = average_cellsize_min(dd, ddbox);
3845 comm->cutoff_mbody = 0.5*(r_bonded + acs);
3846 if (!isDlbDisabled(comm))
3848 /* Check if this does not limit the scaling */
3849 comm->cutoff_mbody = std::min(comm->cutoff_mbody,
3850 options.dlbScaling*acs);
3852 if (!comm->bBondComm)
3854 /* Without bBondComm do not go beyond the n.b. cut-off */
3855 comm->cutoff_mbody = std::min(comm->cutoff_mbody, comm->cutoff);
3856 if (comm->cellsize_limit >= comm->cutoff)
3858 /* We don't loose a lot of efficieny
3859 * when increasing it to the n.b. cut-off.
3860 * It can even be slightly faster, because we need
3861 * less checks for the communication setup.
3863 comm->cutoff_mbody = comm->cutoff;
3866 /* Check if we did not end up below our original limit */
3867 comm->cutoff_mbody = std::max(comm->cutoff_mbody, r_bonded_limit);
3869 if (comm->cutoff_mbody > comm->cellsize_limit)
3871 comm->cellsize_limit = comm->cutoff_mbody;
3874 /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
3879 fprintf(debug, "Bonded atom communication beyond the cut-off: %s\n"
3880 "cellsize limit %f\n",
3881 gmx::boolToString(comm->bBondComm), comm->cellsize_limit);
3886 check_dd_restrictions(cr, dd, ir, fplog);
3890 static void set_dlb_limits(gmx_domdec_t *dd)
3893 for (int d = 0; d < dd->ndim; d++)
3895 /* Set the number of pulses to the value for DLB */
3896 dd->comm->cd[d].ind.resize(dd->comm->cd[d].np_dlb);
3898 dd->comm->cellsize_min[dd->dim[d]] =
3899 dd->comm->cellsize_min_dlb[dd->dim[d]];
3904 static void turn_on_dlb(FILE *fplog, const t_commrec *cr, int64_t step)
3906 gmx_domdec_t *dd = cr->dd;
3907 gmx_domdec_comm_t *comm = dd->comm;
3909 real cellsize_min = comm->cellsize_min[dd->dim[0]];
3910 for (int d = 1; d < dd->ndim; d++)
3912 cellsize_min = std::min(cellsize_min, comm->cellsize_min[dd->dim[d]]);
3915 /* Turn off DLB if we're too close to the cell size limit. */
3916 if (cellsize_min < comm->cellsize_limit*1.05)
3918 auto str = gmx::formatString("step %" PRId64 " Measured %.1f %% performance loss due to load imbalance, "
3919 "but the minimum cell size is smaller than 1.05 times the cell size limit."
3920 "Will no longer try dynamic load balancing.\n", step, dd_force_imb_perf_loss(dd)*100);
3921 dd_warning(cr, fplog, str.c_str());
3923 comm->dlbState = edlbsOffForever;
3928 sprintf(buf, "step %" PRId64 " Turning on dynamic load balancing, because the performance loss due to load imbalance is %.1f %%.\n", step, dd_force_imb_perf_loss(dd)*100);
3929 dd_warning(cr, fplog, buf);
3930 comm->dlbState = edlbsOnCanTurnOff;
3932 /* Store the non-DLB performance, so we can check if DLB actually
3933 * improves performance.
3935 GMX_RELEASE_ASSERT(comm->cycl_n[ddCyclStep] > 0, "When we turned on DLB, we should have measured cycles");
3936 comm->cyclesPerStepBeforeDLB = comm->cycl[ddCyclStep]/comm->cycl_n[ddCyclStep];
3940 /* We can set the required cell size info here,
3941 * so we do not need to communicate this.
3942 * The grid is completely uniform.
3944 for (int d = 0; d < dd->ndim; d++)
3946 RowMaster *rowMaster = comm->cellsizesWithDlb[d].rowMaster.get();
3950 comm->load[d].sum_m = comm->load[d].sum;
3952 int nc = dd->nc[dd->dim[d]];
3953 for (int i = 0; i < nc; i++)
3955 rowMaster->cellFrac[i] = i/static_cast<real>(nc);
3958 rowMaster->bounds[i].cellFracLowerMax = i /static_cast<real>(nc);
3959 rowMaster->bounds[i].cellFracUpperMin = (i + 1)/static_cast<real>(nc);
3962 rowMaster->cellFrac[nc] = 1.0;
3967 static void turn_off_dlb(FILE *fplog, const t_commrec *cr, int64_t step)
3969 gmx_domdec_t *dd = cr->dd;
3972 sprintf(buf, "step %" PRId64 " Turning off dynamic load balancing, because it is degrading performance.\n", step);
3973 dd_warning(cr, fplog, buf);
3974 dd->comm->dlbState = edlbsOffCanTurnOn;
3975 dd->comm->haveTurnedOffDlb = true;
3976 dd->comm->ddPartioningCountFirstDlbOff = dd->ddp_count;
3979 static void turn_off_dlb_forever(FILE *fplog, const t_commrec *cr, int64_t step)
3981 GMX_RELEASE_ASSERT(cr->dd->comm->dlbState == edlbsOffCanTurnOn, "Can only turn off DLB forever when it was in the can-turn-on state");
3983 sprintf(buf, "step %" PRId64 " Will no longer try dynamic load balancing, as it degraded performance.\n", step);
3984 dd_warning(cr, fplog, buf);
3985 cr->dd->comm->dlbState = edlbsOffForever;
3988 static char *init_bLocalCG(const gmx_mtop_t *mtop)
3993 ncg = ncg_mtop(mtop);
3994 snew(bLocalCG, ncg);
3995 for (cg = 0; cg < ncg; cg++)
3997 bLocalCG[cg] = FALSE;
4003 void dd_init_bondeds(FILE *fplog,
4005 const gmx_mtop_t *mtop,
4006 const gmx_vsite_t *vsite,
4007 const t_inputrec *ir,
4008 gmx_bool bBCheck, cginfo_mb_t *cginfo_mb)
4010 gmx_domdec_comm_t *comm;
4012 dd_make_reverse_top(fplog, dd, mtop, vsite, ir, bBCheck);
4016 if (comm->bBondComm)
4018 /* Communicate atoms beyond the cut-off for bonded interactions */
4021 comm->cglink = make_charge_group_links(mtop, dd, cginfo_mb);
4023 comm->bLocalCG = init_bLocalCG(mtop);
4027 /* Only communicate atoms based on cut-off */
4028 comm->cglink = nullptr;
4029 comm->bLocalCG = nullptr;
4033 static void print_dd_settings(FILE *fplog, gmx_domdec_t *dd,
4034 const gmx_mtop_t *mtop, const t_inputrec *ir,
4035 gmx_bool bDynLoadBal, real dlb_scale,
4036 const gmx_ddbox_t *ddbox)
4038 gmx_domdec_comm_t *comm;
4044 if (fplog == nullptr)
4053 fprintf(fplog, "The maximum number of communication pulses is:");
4054 for (d = 0; d < dd->ndim; d++)
4056 fprintf(fplog, " %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
4058 fprintf(fplog, "\n");
4059 fprintf(fplog, "The minimum size for domain decomposition cells is %.3f nm\n", comm->cellsize_limit);
4060 fprintf(fplog, "The requested allowed shrink of DD cells (option -dds) is: %.2f\n", dlb_scale);
4061 fprintf(fplog, "The allowed shrink of domain decomposition cells is:");
4062 for (d = 0; d < DIM; d++)
4066 if (d >= ddbox->npbcdim && dd->nc[d] == 2)
4073 comm->cellsize_min_dlb[d]/
4074 (ddbox->box_size[d]*ddbox->skew_fac[d]/dd->nc[d]);
4076 fprintf(fplog, " %c %.2f", dim2char(d), shrink);
4079 fprintf(fplog, "\n");
4083 set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
4084 fprintf(fplog, "The initial number of communication pulses is:");
4085 for (d = 0; d < dd->ndim; d++)
4087 fprintf(fplog, " %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
4089 fprintf(fplog, "\n");
4090 fprintf(fplog, "The initial domain decomposition cell size is:");
4091 for (d = 0; d < DIM; d++)
4095 fprintf(fplog, " %c %.2f nm",
4096 dim2char(d), dd->comm->cellsize_min[d]);
4099 fprintf(fplog, "\n\n");
4102 gmx_bool bInterCGVsites = count_intercg_vsites(mtop);
4104 if (comm->bInterCGBondeds ||
4106 dd->bInterCGcons || dd->bInterCGsettles)
4108 fprintf(fplog, "The maximum allowed distance for charge groups involved in interactions is:\n");
4109 fprintf(fplog, "%40s %-7s %6.3f nm\n",
4110 "non-bonded interactions", "", comm->cutoff);
4114 limit = dd->comm->cellsize_limit;
4118 if (dynamic_dd_box(ddbox, ir))
4120 fprintf(fplog, "(the following are initial values, they could change due to box deformation)\n");
4122 limit = dd->comm->cellsize_min[XX];
4123 for (d = 1; d < DIM; d++)
4125 limit = std::min(limit, dd->comm->cellsize_min[d]);
4129 if (comm->bInterCGBondeds)
4131 fprintf(fplog, "%40s %-7s %6.3f nm\n",
4132 "two-body bonded interactions", "(-rdd)",
4133 std::max(comm->cutoff, comm->cutoff_mbody));
4134 fprintf(fplog, "%40s %-7s %6.3f nm\n",
4135 "multi-body bonded interactions", "(-rdd)",
4136 (comm->bBondComm || isDlbOn(dd->comm)) ? comm->cutoff_mbody : std::min(comm->cutoff, limit));
4140 fprintf(fplog, "%40s %-7s %6.3f nm\n",
4141 "virtual site constructions", "(-rcon)", limit);
4143 if (dd->bInterCGcons || dd->bInterCGsettles)
4145 sprintf(buf, "atoms separated by up to %d constraints",
4147 fprintf(fplog, "%40s %-7s %6.3f nm\n",
4148 buf, "(-rcon)", limit);
4150 fprintf(fplog, "\n");
4156 static void set_cell_limits_dlb(gmx_domdec_t *dd,
4158 const t_inputrec *ir,
4159 const gmx_ddbox_t *ddbox)
4161 gmx_domdec_comm_t *comm;
4162 int d, dim, npulse, npulse_d_max, npulse_d;
4167 bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
4169 /* Determine the maximum number of comm. pulses in one dimension */
4171 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
4173 /* Determine the maximum required number of grid pulses */
4174 if (comm->cellsize_limit >= comm->cutoff)
4176 /* Only a single pulse is required */
4179 else if (!bNoCutOff && comm->cellsize_limit > 0)
4181 /* We round down slightly here to avoid overhead due to the latency
4182 * of extra communication calls when the cut-off
4183 * would be only slightly longer than the cell size.
4184 * Later cellsize_limit is redetermined,
4185 * so we can not miss interactions due to this rounding.
4187 npulse = static_cast<int>(0.96 + comm->cutoff/comm->cellsize_limit);
4191 /* There is no cell size limit */
4192 npulse = std::max(dd->nc[XX]-1, std::max(dd->nc[YY]-1, dd->nc[ZZ]-1));
4195 if (!bNoCutOff && npulse > 1)
4197 /* See if we can do with less pulses, based on dlb_scale */
4199 for (d = 0; d < dd->ndim; d++)
4202 npulse_d = static_cast<int>(1 + dd->nc[dim]*comm->cutoff
4203 /(ddbox->box_size[dim]*ddbox->skew_fac[dim]*dlb_scale));
4204 npulse_d_max = std::max(npulse_d_max, npulse_d);
4206 npulse = std::min(npulse, npulse_d_max);
4209 /* This env var can override npulse */
4210 d = dd_getenv(debug, "GMX_DD_NPULSE", 0);
4217 comm->bVacDLBNoLimit = (ir->ePBC == epbcNONE);
4218 for (d = 0; d < dd->ndim; d++)
4220 comm->cd[d].np_dlb = std::min(npulse, dd->nc[dd->dim[d]]-1);
4221 comm->maxpulse = std::max(comm->maxpulse, comm->cd[d].np_dlb);
4222 if (comm->cd[d].np_dlb < dd->nc[dd->dim[d]]-1)
4224 comm->bVacDLBNoLimit = FALSE;
4228 /* cellsize_limit is set for LINCS in init_domain_decomposition */
4229 if (!comm->bVacDLBNoLimit)
4231 comm->cellsize_limit = std::max(comm->cellsize_limit,
4232 comm->cutoff/comm->maxpulse);
4234 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
4235 /* Set the minimum cell size for each DD dimension */
4236 for (d = 0; d < dd->ndim; d++)
4238 if (comm->bVacDLBNoLimit ||
4239 comm->cd[d].np_dlb*comm->cellsize_limit >= comm->cutoff)
4241 comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
4245 comm->cellsize_min_dlb[dd->dim[d]] =
4246 comm->cutoff/comm->cd[d].np_dlb;
4249 if (comm->cutoff_mbody <= 0)
4251 comm->cutoff_mbody = std::min(comm->cutoff, comm->cellsize_limit);
4259 gmx_bool dd_bonded_molpbc(const gmx_domdec_t *dd, int ePBC)
4261 /* If each molecule is a single charge group
4262 * or we use domain decomposition for each periodic dimension,
4263 * we do not need to take pbc into account for the bonded interactions.
4265 return (ePBC != epbcNONE && dd->comm->bInterCGBondeds &&
4268 (dd->nc[ZZ] > 1 || ePBC == epbcXY)));
4271 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
4272 static void set_ddgrid_parameters(FILE *fplog, gmx_domdec_t *dd, real dlb_scale,
4273 const gmx_mtop_t *mtop, const t_inputrec *ir,
4274 const gmx_ddbox_t *ddbox)
4276 gmx_domdec_comm_t *comm;
4282 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
4284 init_ddpme(dd, &comm->ddpme[0], 0);
4285 if (comm->npmedecompdim >= 2)
4287 init_ddpme(dd, &comm->ddpme[1], 1);
4292 comm->npmenodes = 0;
4293 if (dd->pme_nodeid >= 0)
4295 gmx_fatal_collective(FARGS, dd->mpi_comm_all, DDMASTER(dd),
4296 "Can not have separate PME ranks without PME electrostatics");
4302 fprintf(debug, "The DD cut-off is %f\n", comm->cutoff);
4304 if (!isDlbDisabled(comm))
4306 set_cell_limits_dlb(dd, dlb_scale, ir, ddbox);
4309 print_dd_settings(fplog, dd, mtop, ir, isDlbOn(comm), dlb_scale, ddbox);
4310 if (comm->dlbState == edlbsOffCanTurnOn)
4314 fprintf(fplog, "When dynamic load balancing gets turned on, these settings will change to:\n");
4316 print_dd_settings(fplog, dd, mtop, ir, TRUE, dlb_scale, ddbox);
4319 if (ir->ePBC == epbcNONE)
4321 vol_frac = 1 - 1/static_cast<double>(dd->nnodes);
4326 (1 + comm_box_frac(dd->nc, comm->cutoff, ddbox))/static_cast<double>(dd->nnodes);
4330 fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
4332 natoms_tot = comm->cgs_gl.index[comm->cgs_gl.nr];
4334 dd->ga2la = ga2la_init(natoms_tot, static_cast<int>(vol_frac*natoms_tot));
4337 /*! \brief Set some important DD parameters that can be modified by env.vars */
4338 static void set_dd_envvar_options(FILE *fplog, gmx_domdec_t *dd, int rank_mysim)
4340 gmx_domdec_comm_t *comm = dd->comm;
4342 dd->bSendRecv2 = dd_getenv(fplog, "GMX_DD_USE_SENDRECV2", 0);
4343 comm->dlb_scale_lim = dd_getenv(fplog, "GMX_DLB_MAX_BOX_SCALING", 10);
4344 comm->eFlop = dd_getenv(fplog, "GMX_DLB_BASED_ON_FLOPS", 0);
4345 int recload = dd_getenv(fplog, "GMX_DD_RECORD_LOAD", 1);
4346 comm->nstDDDump = dd_getenv(fplog, "GMX_DD_NST_DUMP", 0);
4347 comm->nstDDDumpGrid = dd_getenv(fplog, "GMX_DD_NST_DUMP_GRID", 0);
4348 comm->DD_debug = dd_getenv(fplog, "GMX_DD_DEBUG", 0);
4350 if (dd->bSendRecv2 && fplog)
4352 fprintf(fplog, "Will use two sequential MPI_Sendrecv calls instead of two simultaneous non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite communication\n");
4359 fprintf(fplog, "Will load balance based on FLOP count\n");
4361 if (comm->eFlop > 1)
4363 srand(1 + rank_mysim);
4365 comm->bRecordLoad = TRUE;
4369 comm->bRecordLoad = (wallcycle_have_counter() && recload > 0);
4373 DomdecOptions::DomdecOptions() :
4374 checkBondedInteractions(TRUE),
4375 useBondedCommunication(TRUE),
4377 rankOrder(DdRankOrder::pp_pme),
4378 minimumCommunicationRange(0),
4379 constraintCommunicationRange(0),
4380 dlbOption(DlbOption::turnOnWhenUseful),
4386 clear_ivec(numCells);
4389 gmx_domdec_t *init_domain_decomposition(FILE *fplog, t_commrec *cr,
4390 const DomdecOptions &options,
4391 const MdrunOptions &mdrunOptions,
4392 const gmx_mtop_t *mtop,
4393 const t_inputrec *ir,
4395 gmx::ArrayRef<const gmx::RVec> xGlobal,
4396 gmx::LocalAtomSetManager *atomSets)
4403 "\nInitializing Domain Decomposition on %d ranks\n", cr->nnodes);
4406 dd = new gmx_domdec_t;
4408 dd->comm = init_dd_comm();
4410 /* Initialize DD paritioning counters */
4411 dd->comm->partition_step = INT_MIN;
4414 set_dd_envvar_options(fplog, dd, cr->nodeid);
4416 gmx_ddbox_t ddbox = {0};
4417 set_dd_limits_and_grid(fplog, cr, dd, options, mdrunOptions,
4422 make_dd_communicators(fplog, cr, dd, options.rankOrder);
4424 if (thisRankHasDuty(cr, DUTY_PP))
4426 set_ddgrid_parameters(fplog, dd, options.dlbScaling, mtop, ir, &ddbox);
4428 setup_neighbor_relations(dd);
4431 /* Set overallocation to avoid frequent reallocation of arrays */
4432 set_over_alloc_dd(TRUE);
4434 clear_dd_cycle_counts(dd);
4436 dd->atomSets = atomSets;
4441 static gmx_bool test_dd_cutoff(t_commrec *cr,
4442 t_state *state, const t_inputrec *ir,
4453 set_ddbox(dd, false, ir, state->box, true, state->x, &ddbox);
4457 for (d = 0; d < dd->ndim; d++)
4461 inv_cell_size = DD_CELL_MARGIN*dd->nc[dim]/ddbox.box_size[dim];
4462 if (dynamic_dd_box(&ddbox, ir))
4464 inv_cell_size *= DD_PRES_SCALE_MARGIN;
4467 np = 1 + static_cast<int>(cutoff_req*inv_cell_size*ddbox.skew_fac[dim]);
4469 if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
4471 if (np > dd->comm->cd[d].np_dlb)
4476 /* If a current local cell size is smaller than the requested
4477 * cut-off, we could still fix it, but this gets very complicated.
4478 * Without fixing here, we might actually need more checks.
4480 if ((dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim])*ddbox.skew_fac[dim]*dd->comm->cd[d].np_dlb < cutoff_req)
4487 if (!isDlbDisabled(dd->comm))
4489 /* If DLB is not active yet, we don't need to check the grid jumps.
4490 * Actually we shouldn't, because then the grid jump data is not set.
4492 if (isDlbOn(dd->comm) &&
4493 check_grid_jump(0, dd, cutoff_req, &ddbox, FALSE))
4498 gmx_sumi(1, &LocallyLimited, cr);
4500 if (LocallyLimited > 0)
4509 gmx_bool change_dd_cutoff(t_commrec *cr, t_state *state, const t_inputrec *ir,
4512 gmx_bool bCutoffAllowed;
4514 bCutoffAllowed = test_dd_cutoff(cr, state, ir, cutoff_req);
4518 cr->dd->comm->cutoff = cutoff_req;
4521 return bCutoffAllowed;
4524 void set_dd_dlb_max_cutoff(t_commrec *cr, real cutoff)
4526 gmx_domdec_comm_t *comm;
4528 comm = cr->dd->comm;
4530 /* Turn on the DLB limiting (might have been on already) */
4531 comm->bPMELoadBalDLBLimits = TRUE;
4533 /* Change the cut-off limit */
4534 comm->PMELoadBal_max_cutoff = cutoff;
4538 fprintf(debug, "PME load balancing set a limit to the DLB staggering such that a %f cut-off will continue to fit\n",
4539 comm->PMELoadBal_max_cutoff);
4543 /* Sets whether we should later check the load imbalance data, so that
4544 * we can trigger dynamic load balancing if enough imbalance has
4547 * Used after PME load balancing unlocks DLB, so that the check
4548 * whether DLB will be useful can happen immediately.
4550 static void dd_dlb_set_should_check_whether_to_turn_dlb_on(gmx_domdec_t *dd, gmx_bool bValue)
4552 if (dd->comm->dlbState == edlbsOffCanTurnOn)
4554 dd->comm->bCheckWhetherToTurnDlbOn = bValue;
4558 /* Store the DD partitioning count, so we can ignore cycle counts
4559 * over the next nstlist steps, which are often slower.
4561 dd->comm->ddPartioningCountFirstDlbOff = dd->ddp_count;
4566 /* Returns if we should check whether there has been enough load
4567 * imbalance to trigger dynamic load balancing.
4569 static gmx_bool dd_dlb_get_should_check_whether_to_turn_dlb_on(gmx_domdec_t *dd)
4571 if (dd->comm->dlbState != edlbsOffCanTurnOn)
4576 if (dd->ddp_count <= dd->comm->ddPartioningCountFirstDlbOff)
4578 /* We ignore the first nstlist steps at the start of the run
4579 * or after PME load balancing or after turning DLB off, since
4580 * these often have extra allocation or cache miss overhead.
4585 if (dd->comm->cycl_n[ddCyclStep] == 0)
4587 /* We can have zero timed steps when dd_partition_system is called
4588 * more than once at the same step, e.g. with replica exchange.
4589 * Turning on DLB would trigger an assertion failure later, but is
4590 * also useless right after exchanging replicas.
4595 /* We should check whether we should use DLB directly after
4597 if (dd->comm->bCheckWhetherToTurnDlbOn)
4599 /* This flag was set when the PME load-balancing routines
4600 unlocked DLB, and should now be cleared. */
4601 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, FALSE);
4604 /* We check whether we should use DLB every c_checkTurnDlbOnInterval
4605 * partitionings (we do not do this every partioning, so that we
4606 * avoid excessive communication). */
4607 return dd->comm->n_load_have % c_checkTurnDlbOnInterval == c_checkTurnDlbOnInterval - 1;
4610 gmx_bool dd_dlb_is_on(const gmx_domdec_t *dd)
4612 return isDlbOn(dd->comm);
4615 gmx_bool dd_dlb_is_locked(const gmx_domdec_t *dd)
4617 return (dd->comm->dlbState == edlbsOffTemporarilyLocked);
4620 void dd_dlb_lock(gmx_domdec_t *dd)
4622 /* We can only lock the DLB when it is set to auto, otherwise don't do anything */
4623 if (dd->comm->dlbState == edlbsOffCanTurnOn)
4625 dd->comm->dlbState = edlbsOffTemporarilyLocked;
4629 void dd_dlb_unlock(gmx_domdec_t *dd)
4631 /* We can only lock the DLB when it is set to auto, otherwise don't do anything */
4632 if (dd->comm->dlbState == edlbsOffTemporarilyLocked)
4634 dd->comm->dlbState = edlbsOffCanTurnOn;
4635 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
4639 static void merge_cg_buffers(int ncell,
4640 gmx_domdec_comm_dim_t *cd, int pulse,
4642 gmx::ArrayRef<int> index_gl,
4644 rvec *cg_cm, rvec *recv_vr,
4645 gmx::ArrayRef<int> cgindex,
4646 cginfo_mb_t *cginfo_mb, int *cginfo)
4648 gmx_domdec_ind_t *ind, *ind_p;
4649 int p, cell, c, cg, cg0, cg1, cg_gl, nat;
4650 int shift, shift_at;
4652 ind = &cd->ind[pulse];
4654 /* First correct the already stored data */
4655 shift = ind->nrecv[ncell];
4656 for (cell = ncell-1; cell >= 0; cell--)
4658 shift -= ind->nrecv[cell];
4661 /* Move the cg's present from previous grid pulses */
4662 cg0 = ncg_cell[ncell+cell];
4663 cg1 = ncg_cell[ncell+cell+1];
4664 cgindex[cg1+shift] = cgindex[cg1];
4665 for (cg = cg1-1; cg >= cg0; cg--)
4667 index_gl[cg+shift] = index_gl[cg];
4668 copy_rvec(cg_cm[cg], cg_cm[cg+shift]);
4669 cgindex[cg+shift] = cgindex[cg];
4670 cginfo[cg+shift] = cginfo[cg];
4672 /* Correct the already stored send indices for the shift */
4673 for (p = 1; p <= pulse; p++)
4675 ind_p = &cd->ind[p];
4677 for (c = 0; c < cell; c++)
4679 cg0 += ind_p->nsend[c];
4681 cg1 = cg0 + ind_p->nsend[cell];
4682 for (cg = cg0; cg < cg1; cg++)
4684 ind_p->index[cg] += shift;
4690 /* Merge in the communicated buffers */
4694 for (cell = 0; cell < ncell; cell++)
4696 cg1 = ncg_cell[ncell+cell+1] + shift;
4699 /* Correct the old cg indices */
4700 for (cg = ncg_cell[ncell+cell]; cg < cg1; cg++)
4702 cgindex[cg+1] += shift_at;
4705 for (cg = 0; cg < ind->nrecv[cell]; cg++)
4707 /* Copy this charge group from the buffer */
4708 index_gl[cg1] = recv_i[cg0];
4709 copy_rvec(recv_vr[cg0], cg_cm[cg1]);
4710 /* Add it to the cgindex */
4711 cg_gl = index_gl[cg1];
4712 cginfo[cg1] = ddcginfo(cginfo_mb, cg_gl);
4713 nat = GET_CGINFO_NATOMS(cginfo[cg1]);
4714 cgindex[cg1+1] = cgindex[cg1] + nat;
4719 shift += ind->nrecv[cell];
4720 ncg_cell[ncell+cell+1] = cg1;
4724 static void make_cell2at_index(gmx_domdec_comm_dim_t *cd,
4727 const gmx::RangePartitioning &atomGroups)
4729 /* Store the atom block boundaries for easy copying of communication buffers
4731 int g = atomGroupStart;
4732 for (int zone = 0; zone < nzone; zone++)
4734 for (gmx_domdec_ind_t &ind : cd->ind)
4736 const auto range = atomGroups.subRange(g, g + ind.nrecv[zone]);
4737 ind.cell2at0[zone] = range.begin();
4738 ind.cell2at1[zone] = range.end();
4739 g += ind.nrecv[zone];
4744 static gmx_bool missing_link(t_blocka *link, int cg_gl, const char *bLocalCG)
4750 for (i = link->index[cg_gl]; i < link->index[cg_gl+1]; i++)
4752 if (!bLocalCG[link->a[i]])
4761 /* Domain corners for communication, a maximum of 4 i-zones see a j domain */
4763 real c[DIM][4]; /* the corners for the non-bonded communication */
4764 real cr0; /* corner for rounding */
4765 real cr1[4]; /* corners for rounding */
4766 real bc[DIM]; /* corners for bounded communication */
4767 real bcr1; /* corner for rounding for bonded communication */
4770 /* Determine the corners of the domain(s) we are communicating with */
4772 set_dd_corners(const gmx_domdec_t *dd,
4773 int dim0, int dim1, int dim2,
4777 const gmx_domdec_comm_t *comm;
4778 const gmx_domdec_zones_t *zones;
4783 zones = &comm->zones;
4785 /* Keep the compiler happy */
4789 /* The first dimension is equal for all cells */
4790 c->c[0][0] = comm->cell_x0[dim0];
4793 c->bc[0] = c->c[0][0];
4798 /* This cell row is only seen from the first row */
4799 c->c[1][0] = comm->cell_x0[dim1];
4800 /* All rows can see this row */
4801 c->c[1][1] = comm->cell_x0[dim1];
4802 if (isDlbOn(dd->comm))
4804 c->c[1][1] = std::max(comm->cell_x0[dim1], comm->zone_d1[1].mch0);
4807 /* For the multi-body distance we need the maximum */
4808 c->bc[1] = std::max(comm->cell_x0[dim1], comm->zone_d1[1].p1_0);
4811 /* Set the upper-right corner for rounding */
4812 c->cr0 = comm->cell_x1[dim0];
4817 for (j = 0; j < 4; j++)
4819 c->c[2][j] = comm->cell_x0[dim2];
4821 if (isDlbOn(dd->comm))
4823 /* Use the maximum of the i-cells that see a j-cell */
4824 for (i = 0; i < zones->nizone; i++)
4826 for (j = zones->izone[i].j0; j < zones->izone[i].j1; j++)
4831 std::max(c->c[2][j-4],
4832 comm->zone_d2[zones->shift[i][dim0]][zones->shift[i][dim1]].mch0);
4838 /* For the multi-body distance we need the maximum */
4839 c->bc[2] = comm->cell_x0[dim2];
4840 for (i = 0; i < 2; i++)
4842 for (j = 0; j < 2; j++)
4844 c->bc[2] = std::max(c->bc[2], comm->zone_d2[i][j].p1_0);
4850 /* Set the upper-right corner for rounding */
4851 /* Cell (0,0,0) and cell (1,0,0) can see cell 4 (0,1,1)
4852 * Only cell (0,0,0) can see cell 7 (1,1,1)
4854 c->cr1[0] = comm->cell_x1[dim1];
4855 c->cr1[3] = comm->cell_x1[dim1];
4856 if (isDlbOn(dd->comm))
4858 c->cr1[0] = std::max(comm->cell_x1[dim1], comm->zone_d1[1].mch1);
4861 /* For the multi-body distance we need the maximum */
4862 c->bcr1 = std::max(comm->cell_x1[dim1], comm->zone_d1[1].p1_1);
4869 /* Add the atom groups we need to send in this pulse from this zone to
4870 * \p localAtomGroups and \p work
4873 get_zone_pulse_cgs(gmx_domdec_t *dd,
4874 int zonei, int zone,
4876 gmx::ArrayRef<const int> globalAtomGroupIndices,
4877 const gmx::RangePartitioning &atomGroups,
4878 int dim, int dim_ind,
4879 int dim0, int dim1, int dim2,
4880 real r_comm2, real r_bcomm2,
4882 bool distanceIsTriclinic,
4884 real skew_fac2_d, real skew_fac_01,
4885 rvec *v_d, rvec *v_0, rvec *v_1,
4886 const dd_corners_t *c,
4887 const rvec sf2_round,
4888 gmx_bool bDistBonded,
4894 std::vector<int> *localAtomGroups,
4895 dd_comm_setup_work_t *work)
4897 gmx_domdec_comm_t *comm;
4899 gmx_bool bDistMB_pulse;
4901 real r2, rb2, r, tric_sh;
4908 bScrew = (dd->bScrewPBC && dim == XX);
4910 bDistMB_pulse = (bDistMB && bDistBonded);
4912 /* Unpack the work data */
4913 std::vector<int> &ibuf = work->atomGroupBuffer;
4914 std::vector<gmx::RVec> &vbuf = work->positionBuffer;
4918 for (cg = cg0; cg < cg1; cg++)
4922 if (!distanceIsTriclinic)
4924 /* Rectangular direction, easy */
4925 r = cg_cm[cg][dim] - c->c[dim_ind][zone];
4932 r = cg_cm[cg][dim] - c->bc[dim_ind];
4938 /* Rounding gives at most a 16% reduction
4939 * in communicated atoms
4941 if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
4943 r = cg_cm[cg][dim0] - c->cr0;
4944 /* This is the first dimension, so always r >= 0 */
4951 if (dim_ind == 2 && (zonei == 2 || zonei == 3))
4953 r = cg_cm[cg][dim1] - c->cr1[zone];
4960 r = cg_cm[cg][dim1] - c->bcr1;
4970 /* Triclinic direction, more complicated */
4973 /* Rounding, conservative as the skew_fac multiplication
4974 * will slightly underestimate the distance.
4976 if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
4978 rn[dim0] = cg_cm[cg][dim0] - c->cr0;
4979 for (i = dim0+1; i < DIM; i++)
4981 rn[dim0] -= cg_cm[cg][i]*v_0[i][dim0];
4983 r2 = rn[dim0]*rn[dim0]*sf2_round[dim0];
4986 rb[dim0] = rn[dim0];
4989 /* Take care that the cell planes along dim0 might not
4990 * be orthogonal to those along dim1 and dim2.
4992 for (i = 1; i <= dim_ind; i++)
4995 if (normal[dim0][dimd] > 0)
4997 rn[dimd] -= rn[dim0]*normal[dim0][dimd];
5000 rb[dimd] -= rb[dim0]*normal[dim0][dimd];
5005 if (dim_ind == 2 && (zonei == 2 || zonei == 3))
5007 rn[dim1] += cg_cm[cg][dim1] - c->cr1[zone];
5009 for (i = dim1+1; i < DIM; i++)
5011 tric_sh -= cg_cm[cg][i]*v_1[i][dim1];
5013 rn[dim1] += tric_sh;
5016 r2 += rn[dim1]*rn[dim1]*sf2_round[dim1];
5017 /* Take care of coupling of the distances
5018 * to the planes along dim0 and dim1 through dim2.
5020 r2 -= rn[dim0]*rn[dim1]*skew_fac_01;
5021 /* Take care that the cell planes along dim1
5022 * might not be orthogonal to that along dim2.
5024 if (normal[dim1][dim2] > 0)
5026 rn[dim2] -= rn[dim1]*normal[dim1][dim2];
5032 cg_cm[cg][dim1] - c->bcr1 + tric_sh;
5035 rb2 += rb[dim1]*rb[dim1]*sf2_round[dim1];
5036 /* Take care of coupling of the distances
5037 * to the planes along dim0 and dim1 through dim2.
5039 rb2 -= rb[dim0]*rb[dim1]*skew_fac_01;
5040 /* Take care that the cell planes along dim1
5041 * might not be orthogonal to that along dim2.
5043 if (normal[dim1][dim2] > 0)
5045 rb[dim2] -= rb[dim1]*normal[dim1][dim2];
5050 /* The distance along the communication direction */
5051 rn[dim] += cg_cm[cg][dim] - c->c[dim_ind][zone];
5053 for (i = dim+1; i < DIM; i++)
5055 tric_sh -= cg_cm[cg][i]*v_d[i][dim];
5060 r2 += rn[dim]*rn[dim]*skew_fac2_d;
5061 /* Take care of coupling of the distances
5062 * to the planes along dim0 and dim1 through dim2.
5064 if (dim_ind == 1 && zonei == 1)
5066 r2 -= rn[dim0]*rn[dim]*skew_fac_01;
5072 rb[dim] += cg_cm[cg][dim] - c->bc[dim_ind] + tric_sh;
5075 rb2 += rb[dim]*rb[dim]*skew_fac2_d;
5076 /* Take care of coupling of the distances
5077 * to the planes along dim0 and dim1 through dim2.
5079 if (dim_ind == 1 && zonei == 1)
5081 rb2 -= rb[dim0]*rb[dim]*skew_fac_01;
5089 ((bDistMB && rb2 < r_bcomm2) ||
5090 (bDist2B && r2 < r_bcomm2)) &&
5092 (GET_CGINFO_BOND_INTER(cginfo[cg]) &&
5093 missing_link(comm->cglink, globalAtomGroupIndices[cg],
5096 /* Store the local and global atom group indices and position */
5097 localAtomGroups->push_back(cg);
5098 ibuf.push_back(globalAtomGroupIndices[cg]);
5102 if (dd->ci[dim] == 0)
5104 /* Correct cg_cm for pbc */
5105 rvec_add(cg_cm[cg], box[dim], posPbc);
5108 posPbc[YY] = box[YY][YY] - posPbc[YY];
5109 posPbc[ZZ] = box[ZZ][ZZ] - posPbc[ZZ];
5114 copy_rvec(cg_cm[cg], posPbc);
5116 vbuf.emplace_back(posPbc[XX], posPbc[YY], posPbc[ZZ]);
5118 nat += atomGroups.block(cg).size();
5123 work->nsend_zone = nsend_z;
5126 static void clearCommSetupData(dd_comm_setup_work_t *work)
5128 work->localAtomGroupBuffer.clear();
5129 work->atomGroupBuffer.clear();
5130 work->positionBuffer.clear();
5132 work->nsend_zone = 0;
5135 static void setup_dd_communication(gmx_domdec_t *dd,
5136 matrix box, gmx_ddbox_t *ddbox,
5138 t_state *state, PaddedRVecVector *f)
5140 int dim_ind, dim, dim0, dim1, dim2, dimd, nat_tot;
5141 int nzone, nzone_send, zone, zonei, cg0, cg1;
5142 int c, i, cg, cg_gl, nrcg;
5143 int *zone_cg_range, pos_cg;
5144 gmx_domdec_comm_t *comm;
5145 gmx_domdec_zones_t *zones;
5146 gmx_domdec_comm_dim_t *cd;
5147 cginfo_mb_t *cginfo_mb;
5148 gmx_bool bBondComm, bDist2B, bDistMB, bDistBonded;
5149 real r_comm2, r_bcomm2;
5150 dd_corners_t corners;
5151 rvec *cg_cm, *normal, *v_d, *v_0 = nullptr, *v_1 = nullptr;
5152 real skew_fac2_d, skew_fac_01;
5157 fprintf(debug, "Setting up DD communication\n");
5162 if (comm->dth.empty())
5164 /* Initialize the thread data.
5165 * This can not be done in init_domain_decomposition,
5166 * as the numbers of threads is determined later.
5168 int numThreads = gmx_omp_nthreads_get(emntDomdec);
5169 comm->dth.resize(numThreads);
5172 switch (fr->cutoff_scheme)
5178 cg_cm = as_rvec_array(state->x.data());
5181 gmx_incons("unimplemented");
5184 bBondComm = comm->bBondComm;
5186 /* Do we need to determine extra distances for multi-body bondeds? */
5187 bDistMB = (comm->bInterCGMultiBody && isDlbOn(dd->comm) && dd->ndim > 1);
5189 /* Do we need to determine extra distances for only two-body bondeds? */
5190 bDist2B = (bBondComm && !bDistMB);
5192 r_comm2 = gmx::square(comm->cutoff);
5193 r_bcomm2 = gmx::square(comm->cutoff_mbody);
5197 fprintf(debug, "bBondComm %s, r_bc %f\n", gmx::boolToString(bBondComm), std::sqrt(r_bcomm2));
5200 zones = &comm->zones;
5203 dim1 = (dd->ndim >= 2 ? dd->dim[1] : -1);
5204 dim2 = (dd->ndim >= 3 ? dd->dim[2] : -1);
5206 set_dd_corners(dd, dim0, dim1, dim2, bDistMB, &corners);
5208 /* Triclinic stuff */
5209 normal = ddbox->normal;
5213 v_0 = ddbox->v[dim0];
5214 if (ddbox->tric_dir[dim0] && ddbox->tric_dir[dim1])
5216 /* Determine the coupling coefficient for the distances
5217 * to the cell planes along dim0 and dim1 through dim2.
5218 * This is required for correct rounding.
5221 ddbox->v[dim0][dim1+1][dim0]*ddbox->v[dim1][dim1+1][dim1];
5224 fprintf(debug, "\nskew_fac_01 %f\n", skew_fac_01);
5230 v_1 = ddbox->v[dim1];
5233 zone_cg_range = zones->cg_range;
5234 cginfo_mb = fr->cginfo_mb;
5236 zone_cg_range[0] = 0;
5237 zone_cg_range[1] = dd->ncg_home;
5238 comm->zone_ncg1[0] = dd->ncg_home;
5239 pos_cg = dd->ncg_home;
5241 nat_tot = comm->atomRanges.numHomeAtoms();
5243 for (dim_ind = 0; dim_ind < dd->ndim; dim_ind++)
5245 dim = dd->dim[dim_ind];
5246 cd = &comm->cd[dim_ind];
5248 /* Check if we need to compute triclinic distances along this dim */
5249 bool distanceIsTriclinic = false;
5250 for (i = 0; i <= dim_ind; i++)
5252 if (ddbox->tric_dir[dd->dim[i]])
5254 distanceIsTriclinic = true;
5258 if (dim >= ddbox->npbcdim && dd->ci[dim] == 0)
5260 /* No pbc in this dimension, the first node should not comm. */
5268 v_d = ddbox->v[dim];
5269 skew_fac2_d = gmx::square(ddbox->skew_fac[dim]);
5271 cd->receiveInPlace = true;
5272 for (int p = 0; p < cd->numPulses(); p++)
5274 /* Only atoms communicated in the first pulse are used
5275 * for multi-body bonded interactions or for bBondComm.
5277 bDistBonded = ((bDistMB || bDist2B) && p == 0);
5279 gmx_domdec_ind_t *ind = &cd->ind[p];
5281 /* Thread 0 writes in the global index array */
5283 clearCommSetupData(&comm->dth[0]);
5285 for (zone = 0; zone < nzone_send; zone++)
5287 if (dim_ind > 0 && distanceIsTriclinic)
5289 /* Determine slightly more optimized skew_fac's
5291 * This reduces the number of communicated atoms
5292 * by about 10% for 3D DD of rhombic dodecahedra.
5294 for (dimd = 0; dimd < dim; dimd++)
5296 sf2_round[dimd] = 1;
5297 if (ddbox->tric_dir[dimd])
5299 for (i = dd->dim[dimd]+1; i < DIM; i++)
5301 /* If we are shifted in dimension i
5302 * and the cell plane is tilted forward
5303 * in dimension i, skip this coupling.
5305 if (!(zones->shift[nzone+zone][i] &&
5306 ddbox->v[dimd][i][dimd] >= 0))
5309 gmx::square(ddbox->v[dimd][i][dimd]);
5312 sf2_round[dimd] = 1/sf2_round[dimd];
5317 zonei = zone_perm[dim_ind][zone];
5320 /* Here we permutate the zones to obtain a convenient order
5321 * for neighbor searching
5323 cg0 = zone_cg_range[zonei];
5324 cg1 = zone_cg_range[zonei+1];
5328 /* Look only at the cg's received in the previous grid pulse
5330 cg1 = zone_cg_range[nzone+zone+1];
5331 cg0 = cg1 - cd->ind[p-1].nrecv[zone];
5334 const int numThreads = static_cast<int>(comm->dth.size());
5335 #pragma omp parallel for num_threads(numThreads) schedule(static)
5336 for (int th = 0; th < numThreads; th++)
5340 dd_comm_setup_work_t &work = comm->dth[th];
5342 /* Retain data accumulated into buffers of thread 0 */
5345 clearCommSetupData(&work);
5348 int cg0_th = cg0 + ((cg1 - cg0)* th )/numThreads;
5349 int cg1_th = cg0 + ((cg1 - cg0)*(th+1))/numThreads;
5351 /* Get the cg's for this pulse in this zone */
5352 get_zone_pulse_cgs(dd, zonei, zone, cg0_th, cg1_th,
5353 dd->globalAtomGroupIndices,
5355 dim, dim_ind, dim0, dim1, dim2,
5357 box, distanceIsTriclinic,
5358 normal, skew_fac2_d, skew_fac_01,
5359 v_d, v_0, v_1, &corners, sf2_round,
5360 bDistBonded, bBondComm,
5363 th == 0 ? &ind->index : &work.localAtomGroupBuffer,
5366 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
5369 std::vector<int> &atomGroups = comm->dth[0].atomGroupBuffer;
5370 std::vector<gmx::RVec> &positions = comm->dth[0].positionBuffer;
5371 ind->nsend[zone] = comm->dth[0].nsend_zone;
5372 /* Append data of threads>=1 to the communication buffers */
5373 for (int th = 1; th < numThreads; th++)
5375 const dd_comm_setup_work_t &dth = comm->dth[th];
5377 ind->index.insert(ind->index.end(), dth.localAtomGroupBuffer.begin(), dth.localAtomGroupBuffer.end());
5378 atomGroups.insert(atomGroups.end(), dth.atomGroupBuffer.begin(), dth.atomGroupBuffer.end());
5379 positions.insert(positions.end(), dth.positionBuffer.begin(), dth.positionBuffer.end());
5380 comm->dth[0].nat += dth.nat;
5381 ind->nsend[zone] += dth.nsend_zone;
5384 /* Clear the counts in case we do not have pbc */
5385 for (zone = nzone_send; zone < nzone; zone++)
5387 ind->nsend[zone] = 0;
5389 ind->nsend[nzone] = ind->index.size();
5390 ind->nsend[nzone + 1] = comm->dth[0].nat;
5391 /* Communicate the number of cg's and atoms to receive */
5392 ddSendrecv(dd, dim_ind, dddirBackward,
5393 ind->nsend, nzone+2,
5394 ind->nrecv, nzone+2);
5398 /* We can receive in place if only the last zone is not empty */
5399 for (zone = 0; zone < nzone-1; zone++)
5401 if (ind->nrecv[zone] > 0)
5403 cd->receiveInPlace = false;
5408 int receiveBufferSize = 0;
5409 if (!cd->receiveInPlace)
5411 receiveBufferSize = ind->nrecv[nzone];
5413 /* These buffer are actually only needed with in-place */
5414 DDBufferAccess<int> globalAtomGroupBuffer(comm->intBuffer, receiveBufferSize);
5415 DDBufferAccess<gmx::RVec> rvecBuffer(comm->rvecBuffer, receiveBufferSize);
5417 dd_comm_setup_work_t &work = comm->dth[0];
5419 /* Make space for the global cg indices */
5420 int numAtomGroupsNew = pos_cg + ind->nrecv[nzone];
5421 dd->globalAtomGroupIndices.resize(numAtomGroupsNew);
5422 /* Communicate the global cg indices */
5423 gmx::ArrayRef<int> integerBufferRef;
5424 if (cd->receiveInPlace)
5426 integerBufferRef = gmx::arrayRefFromArray(dd->globalAtomGroupIndices.data() + pos_cg, ind->nrecv[nzone]);
5430 integerBufferRef = globalAtomGroupBuffer.buffer;
5432 ddSendrecv<int>(dd, dim_ind, dddirBackward,
5433 work.atomGroupBuffer, integerBufferRef);
5435 /* Make space for cg_cm */
5436 dd_check_alloc_ncg(fr, state, f, pos_cg + ind->nrecv[nzone]);
5437 if (fr->cutoff_scheme == ecutsGROUP)
5443 cg_cm = as_rvec_array(state->x.data());
5445 /* Communicate cg_cm */
5446 gmx::ArrayRef<gmx::RVec> rvecBufferRef;
5447 if (cd->receiveInPlace)
5449 rvecBufferRef = gmx::arrayRefFromArray(reinterpret_cast<gmx::RVec *>(cg_cm + pos_cg), ind->nrecv[nzone]);
5453 rvecBufferRef = rvecBuffer.buffer;
5455 ddSendrecv<gmx::RVec>(dd, dim_ind, dddirBackward,
5456 work.positionBuffer, rvecBufferRef);
5458 /* Make the charge group index */
5459 if (cd->receiveInPlace)
5461 zone = (p == 0 ? 0 : nzone - 1);
5462 while (zone < nzone)
5464 for (cg = 0; cg < ind->nrecv[zone]; cg++)
5466 cg_gl = dd->globalAtomGroupIndices[pos_cg];
5467 fr->cginfo[pos_cg] = ddcginfo(cginfo_mb, cg_gl);
5468 nrcg = GET_CGINFO_NATOMS(fr->cginfo[pos_cg]);
5469 dd->atomGrouping_.appendBlock(nrcg);
5472 /* Update the charge group presence,
5473 * so we can use it in the next pass of the loop.
5475 comm->bLocalCG[cg_gl] = TRUE;
5481 comm->zone_ncg1[nzone+zone] = ind->nrecv[zone];
5484 zone_cg_range[nzone+zone] = pos_cg;
5489 /* This part of the code is never executed with bBondComm. */
5490 std::vector<int> &atomGroupsIndex = dd->atomGrouping_.rawIndex();
5491 atomGroupsIndex.resize(numAtomGroupsNew + 1);
5493 merge_cg_buffers(nzone, cd, p, zone_cg_range,
5494 dd->globalAtomGroupIndices, integerBufferRef.data(),
5495 cg_cm, as_rvec_array(rvecBufferRef.data()),
5497 fr->cginfo_mb, fr->cginfo);
5498 pos_cg += ind->nrecv[nzone];
5500 nat_tot += ind->nrecv[nzone+1];
5502 if (!cd->receiveInPlace)
5504 /* Store the atom block for easy copying of communication buffers */
5505 make_cell2at_index(cd, nzone, zone_cg_range[nzone], dd->atomGrouping());
5510 comm->atomRanges.setEnd(DDAtomRanges::Type::Zones, nat_tot);
5514 /* We don't need to update cginfo, since that was alrady done above.
5515 * So we pass NULL for the forcerec.
5517 dd_set_cginfo(dd->globalAtomGroupIndices,
5518 dd->ncg_home, dd->globalAtomGroupIndices.size(),
5519 nullptr, comm->bLocalCG);
5524 fprintf(debug, "Finished setting up DD communication, zones:");
5525 for (c = 0; c < zones->n; c++)
5527 fprintf(debug, " %d", zones->cg_range[c+1]-zones->cg_range[c]);
5529 fprintf(debug, "\n");
5533 static void set_cg_boundaries(gmx_domdec_zones_t *zones)
5537 for (c = 0; c < zones->nizone; c++)
5539 zones->izone[c].cg1 = zones->cg_range[c+1];
5540 zones->izone[c].jcg0 = zones->cg_range[zones->izone[c].j0];
5541 zones->izone[c].jcg1 = zones->cg_range[zones->izone[c].j1];
5545 /* \brief Set zone dimensions for zones \p zone_start to \p zone_end-1
5547 * Also sets the atom density for the home zone when \p zone_start=0.
5548 * For this \p numMovedChargeGroupsInHomeZone needs to be passed to tell
5549 * how many charge groups will move but are still part of the current range.
5550 * \todo When converting domdec to use proper classes, all these variables
5551 * should be private and a method should return the correct count
5552 * depending on an internal state.
5554 * \param[in,out] dd The domain decomposition struct
5555 * \param[in] box The box
5556 * \param[in] ddbox The domain decomposition box struct
5557 * \param[in] zone_start The start of the zone range to set sizes for
5558 * \param[in] zone_end The end of the zone range to set sizes for
5559 * \param[in] numMovedChargeGroupsInHomeZone The number of charge groups in the home zone that should moved but are still present in dd->comm->zones.cg_range
5561 static void set_zones_size(gmx_domdec_t *dd,
5562 matrix box, const gmx_ddbox_t *ddbox,
5563 int zone_start, int zone_end,
5564 int numMovedChargeGroupsInHomeZone)
5566 gmx_domdec_comm_t *comm;
5567 gmx_domdec_zones_t *zones;
5576 zones = &comm->zones;
5578 /* Do we need to determine extra distances for multi-body bondeds? */
5579 bDistMB = (comm->bInterCGMultiBody && isDlbOn(dd->comm) && dd->ndim > 1);
5581 for (z = zone_start; z < zone_end; z++)
5583 /* Copy cell limits to zone limits.
5584 * Valid for non-DD dims and non-shifted dims.
5586 copy_rvec(comm->cell_x0, zones->size[z].x0);
5587 copy_rvec(comm->cell_x1, zones->size[z].x1);
5590 for (d = 0; d < dd->ndim; d++)
5594 for (z = 0; z < zones->n; z++)
5596 /* With a staggered grid we have different sizes
5597 * for non-shifted dimensions.
5599 if (isDlbOn(dd->comm) && zones->shift[z][dim] == 0)
5603 zones->size[z].x0[dim] = comm->zone_d1[zones->shift[z][dd->dim[d-1]]].min0;
5604 zones->size[z].x1[dim] = comm->zone_d1[zones->shift[z][dd->dim[d-1]]].max1;
5608 zones->size[z].x0[dim] = comm->zone_d2[zones->shift[z][dd->dim[d-2]]][zones->shift[z][dd->dim[d-1]]].min0;
5609 zones->size[z].x1[dim] = comm->zone_d2[zones->shift[z][dd->dim[d-2]]][zones->shift[z][dd->dim[d-1]]].max1;
5615 rcmbs = comm->cutoff_mbody;
5616 if (ddbox->tric_dir[dim])
5618 rcs /= ddbox->skew_fac[dim];
5619 rcmbs /= ddbox->skew_fac[dim];
5622 /* Set the lower limit for the shifted zone dimensions */
5623 for (z = zone_start; z < zone_end; z++)
5625 if (zones->shift[z][dim] > 0)
5628 if (!isDlbOn(dd->comm) || d == 0)
5630 zones->size[z].x0[dim] = comm->cell_x1[dim];
5631 zones->size[z].x1[dim] = comm->cell_x1[dim] + rcs;
5635 /* Here we take the lower limit of the zone from
5636 * the lowest domain of the zone below.
5640 zones->size[z].x0[dim] =
5641 comm->zone_d1[zones->shift[z][dd->dim[d-1]]].min1;
5647 zones->size[z].x0[dim] =
5648 zones->size[zone_perm[2][z-4]].x0[dim];
5652 zones->size[z].x0[dim] =
5653 comm->zone_d2[zones->shift[z][dd->dim[d-2]]][zones->shift[z][dd->dim[d-1]]].min1;
5656 /* A temporary limit, is updated below */
5657 zones->size[z].x1[dim] = zones->size[z].x0[dim];
5661 for (zi = 0; zi < zones->nizone; zi++)
5663 if (zones->shift[zi][dim] == 0)
5665 /* This takes the whole zone into account.
5666 * With multiple pulses this will lead
5667 * to a larger zone then strictly necessary.
5669 zones->size[z].x1[dim] = std::max(zones->size[z].x1[dim],
5670 zones->size[zi].x1[dim]+rcmbs);
5678 /* Loop over the i-zones to set the upper limit of each
5681 for (zi = 0; zi < zones->nizone; zi++)
5683 if (zones->shift[zi][dim] == 0)
5685 /* We should only use zones up to zone_end */
5686 int jZoneEnd = std::min(zones->izone[zi].j1, zone_end);
5687 for (z = zones->izone[zi].j0; z < jZoneEnd; z++)
5689 if (zones->shift[z][dim] > 0)
5691 zones->size[z].x1[dim] = std::max(zones->size[z].x1[dim],
5692 zones->size[zi].x1[dim]+rcs);
5699 for (z = zone_start; z < zone_end; z++)
5701 /* Initialization only required to keep the compiler happy */
5702 rvec corner_min = {0, 0, 0}, corner_max = {0, 0, 0}, corner;
5705 /* To determine the bounding box for a zone we need to find
5706 * the extreme corners of 4, 2 or 1 corners.
5708 nc = 1 << (ddbox->nboundeddim - 1);
5710 for (c = 0; c < nc; c++)
5712 /* Set up a zone corner at x=0, ignoring trilinic couplings */
5716 corner[YY] = zones->size[z].x0[YY];
5720 corner[YY] = zones->size[z].x1[YY];
5724 corner[ZZ] = zones->size[z].x0[ZZ];
5728 corner[ZZ] = zones->size[z].x1[ZZ];
5730 if (dd->ndim == 1 && dd->dim[0] < ZZ && ZZ < dd->npbcdim &&
5731 box[ZZ][1 - dd->dim[0]] != 0)
5733 /* With 1D domain decomposition the cg's are not in
5734 * the triclinic box, but triclinic x-y and rectangular y/x-z.
5735 * Shift the corner of the z-vector back to along the box
5736 * vector of dimension d, so it will later end up at 0 along d.
5737 * This can affect the location of this corner along dd->dim[0]
5738 * through the matrix operation below if box[d][dd->dim[0]]!=0.
5740 int d = 1 - dd->dim[0];
5742 corner[d] -= corner[ZZ]*box[ZZ][d]/box[ZZ][ZZ];
5744 /* Apply the triclinic couplings */
5745 assert(ddbox->npbcdim <= DIM);
5746 for (i = YY; i < ddbox->npbcdim; i++)
5748 for (j = XX; j < i; j++)
5750 corner[j] += corner[i]*box[i][j]/box[i][i];
5755 copy_rvec(corner, corner_min);
5756 copy_rvec(corner, corner_max);
5760 for (i = 0; i < DIM; i++)
5762 corner_min[i] = std::min(corner_min[i], corner[i]);
5763 corner_max[i] = std::max(corner_max[i], corner[i]);
5767 /* Copy the extreme cornes without offset along x */
5768 for (i = 0; i < DIM; i++)
5770 zones->size[z].bb_x0[i] = corner_min[i];
5771 zones->size[z].bb_x1[i] = corner_max[i];
5773 /* Add the offset along x */
5774 zones->size[z].bb_x0[XX] += zones->size[z].x0[XX];
5775 zones->size[z].bb_x1[XX] += zones->size[z].x1[XX];
5778 if (zone_start == 0)
5781 for (dim = 0; dim < DIM; dim++)
5783 vol *= zones->size[0].x1[dim] - zones->size[0].x0[dim];
5785 zones->dens_zone0 = (zones->cg_range[1] - zones->cg_range[0] - numMovedChargeGroupsInHomeZone)/vol;
5790 for (z = zone_start; z < zone_end; z++)
5792 fprintf(debug, "zone %d %6.3f - %6.3f %6.3f - %6.3f %6.3f - %6.3f\n",
5794 zones->size[z].x0[XX], zones->size[z].x1[XX],
5795 zones->size[z].x0[YY], zones->size[z].x1[YY],
5796 zones->size[z].x0[ZZ], zones->size[z].x1[ZZ]);
5797 fprintf(debug, "zone %d bb %6.3f - %6.3f %6.3f - %6.3f %6.3f - %6.3f\n",
5799 zones->size[z].bb_x0[XX], zones->size[z].bb_x1[XX],
5800 zones->size[z].bb_x0[YY], zones->size[z].bb_x1[YY],
5801 zones->size[z].bb_x0[ZZ], zones->size[z].bb_x1[ZZ]);
5806 static int comp_cgsort(const void *a, const void *b)
5810 gmx_cgsort_t *cga, *cgb;
5811 cga = (gmx_cgsort_t *)a;
5812 cgb = (gmx_cgsort_t *)b;
5814 comp = cga->nsc - cgb->nsc;
5817 comp = cga->ind_gl - cgb->ind_gl;
5823 /* Order data in \p dataToSort according to \p sort
5825 * Note: both buffers should have at least \p sort.size() elements.
5827 template <typename T>
5829 orderVector(gmx::ArrayRef<const gmx_cgsort_t> sort,
5830 gmx::ArrayRef<T> dataToSort,
5831 gmx::ArrayRef<T> sortBuffer)
5833 GMX_ASSERT(dataToSort.size() >= sort.size(), "The vector needs to be sufficiently large");
5834 GMX_ASSERT(sortBuffer.size() >= sort.size(), "The sorting buffer needs to be sufficiently large");
5836 /* Order the data into the temporary buffer */
5838 for (const gmx_cgsort_t &entry : sort)
5840 sortBuffer[i++] = dataToSort[entry.ind];
5843 /* Copy back to the original array */
5844 std::copy(sortBuffer.begin(), sortBuffer.begin() + sort.size(),
5845 dataToSort.begin());
5848 /* Order data in \p dataToSort according to \p sort
5850 * Note: \p vectorToSort should have at least \p sort.size() elements,
5851 * \p workVector is resized when it is too small.
5853 template <typename T>
5855 orderVector(gmx::ArrayRef<const gmx_cgsort_t> sort,
5856 gmx::ArrayRef<T> vectorToSort,
5857 std::vector<T> *workVector)
5859 if (gmx::index(workVector->size()) < sort.size())
5861 workVector->resize(sort.size());
5863 orderVector<T>(sort, vectorToSort, *workVector);
5866 static void order_vec_atom(const gmx::RangePartitioning *atomGroups,
5867 gmx::ArrayRef<const gmx_cgsort_t> sort,
5868 gmx::ArrayRef<gmx::RVec> v,
5869 gmx::ArrayRef<gmx::RVec> buf)
5871 if (atomGroups == nullptr)
5873 /* Avoid the useless loop of the atoms within a cg */
5874 orderVector(sort, v, buf);
5879 /* Order the data */
5881 for (const gmx_cgsort_t &entry : sort)
5883 for (int i : atomGroups->block(entry.ind))
5885 copy_rvec(v[i], buf[a]);
5891 /* Copy back to the original array */
5892 for (int a = 0; a < atot; a++)
5894 copy_rvec(buf[a], v[a]);
5898 /* Returns whether a < b */
5899 static bool compareCgsort(const gmx_cgsort_t &a,
5900 const gmx_cgsort_t &b)
5902 return (a.nsc < b.nsc ||
5903 (a.nsc == b.nsc && a.ind_gl < b.ind_gl));
5906 static void orderedSort(gmx::ArrayRef<const gmx_cgsort_t> stationary,
5907 gmx::ArrayRef<gmx_cgsort_t> moved,
5908 std::vector<gmx_cgsort_t> *sort1)
5910 /* The new indices are not very ordered, so we qsort them */
5911 gmx_qsort_threadsafe(moved.data(), moved.size(), sizeof(moved[0]), comp_cgsort);
5913 /* stationary is already ordered, so now we can merge the two arrays */
5914 sort1->resize(stationary.size() + moved.size());
5915 std::merge(stationary.begin(), stationary.end(),
5916 moved.begin(), moved.end(),
5921 /* Set the sorting order for systems with charge groups, returned in sort->sort.
5922 * The order is according to the global charge group index.
5923 * This adds and removes charge groups that moved between domains.
5925 static void dd_sort_order(const gmx_domdec_t *dd,
5926 const t_forcerec *fr,
5928 gmx_domdec_sort_t *sort)
5930 const int *a = fr->ns->grid->cell_index;
5932 const int movedValue = NSGRID_SIGNAL_MOVED_FAC*fr->ns->grid->ncells;
5934 if (ncg_home_old >= 0)
5936 std::vector<gmx_cgsort_t> &stationary = sort->stationary;
5937 std::vector<gmx_cgsort_t> &moved = sort->moved;
5939 /* The charge groups that remained in the same ns grid cell
5940 * are completely ordered. So we can sort efficiently by sorting
5941 * the charge groups that did move into the stationary list.
5942 * Note: push_back() seems to be slightly slower than direct access.
5946 for (int i = 0; i < dd->ncg_home; i++)
5948 /* Check if this cg did not move to another node */
5949 if (a[i] < movedValue)
5953 entry.ind_gl = dd->globalAtomGroupIndices[i];
5956 if (i >= ncg_home_old || a[i] != sort->sorted[i].nsc)
5958 /* This cg is new on this node or moved ns grid cell */
5959 moved.push_back(entry);
5963 /* This cg did not move */
5964 stationary.push_back(entry);
5971 fprintf(debug, "ordered sort cgs: stationary %zu moved %zu\n",
5972 stationary.size(), moved.size());
5974 /* Sort efficiently */
5975 orderedSort(stationary, moved, &sort->sorted);
5979 std::vector<gmx_cgsort_t> &cgsort = sort->sorted;
5981 cgsort.reserve(dd->ncg_home);
5983 for (int i = 0; i < dd->ncg_home; i++)
5985 /* Sort on the ns grid cell indices
5986 * and the global topology index
5990 entry.ind_gl = dd->globalAtomGroupIndices[i];
5992 cgsort.push_back(entry);
5993 if (cgsort[i].nsc < movedValue)
6000 fprintf(debug, "qsort cgs: %d new home %d\n", dd->ncg_home, numCGNew);
6002 /* Determine the order of the charge groups using qsort */
6003 gmx_qsort_threadsafe(cgsort.data(), dd->ncg_home, sizeof(cgsort[0]), comp_cgsort);
6005 /* Remove the charge groups which are no longer at home here */
6006 cgsort.resize(numCGNew);
6010 /* Returns the sorting order for atoms based on the nbnxn grid order in sort */
6011 static void dd_sort_order_nbnxn(const t_forcerec *fr,
6012 std::vector<gmx_cgsort_t> *sort)
6014 gmx::ArrayRef<const int> atomOrder = nbnxn_get_atomorder(fr->nbv->nbs.get());
6016 /* Using push_back() instead of this resize results in much slower code */
6017 sort->resize(atomOrder.size());
6018 gmx::ArrayRef<gmx_cgsort_t> buffer = *sort;
6019 size_t numSorted = 0;
6020 for (int i : atomOrder)
6024 /* The values of nsc and ind_gl are not used in this case */
6025 buffer[numSorted++].ind = i;
6028 sort->resize(numSorted);
6031 static void dd_sort_state(gmx_domdec_t *dd, rvec *cgcm, t_forcerec *fr, t_state *state,
6034 gmx_domdec_sort_t *sort = dd->comm->sort.get();
6036 switch (fr->cutoff_scheme)
6039 dd_sort_order(dd, fr, ncg_home_old, sort);
6042 dd_sort_order_nbnxn(fr, &sort->sorted);
6045 gmx_incons("unimplemented");
6048 const gmx::RangePartitioning &atomGrouping = dd->atomGrouping();
6050 /* We alloc with the old size, since cgindex is still old */
6051 GMX_ASSERT(atomGrouping.numBlocks() == dd->ncg_home, "atomGroups and dd should be consistent");
6052 DDBufferAccess<gmx::RVec> rvecBuffer(dd->comm->rvecBuffer, atomGrouping.fullRange().end());
6054 const gmx::RangePartitioning *atomGroupsPtr = (dd->comm->bCGs ? &atomGrouping : nullptr);
6056 /* Set the new home atom/charge group count */
6057 dd->ncg_home = sort->sorted.size();
6060 fprintf(debug, "Set the new home charge group count to %d\n",
6064 /* Reorder the state */
6065 gmx::ArrayRef<const gmx_cgsort_t> cgsort = sort->sorted;
6066 GMX_RELEASE_ASSERT(cgsort.size() == dd->ncg_home, "We should sort all the home atom groups");
6068 if (state->flags & (1 << estX))
6070 order_vec_atom(atomGroupsPtr, cgsort, state->x, rvecBuffer.buffer);
6072 if (state->flags & (1 << estV))
6074 order_vec_atom(atomGroupsPtr, cgsort, state->v, rvecBuffer.buffer);
6076 if (state->flags & (1 << estCGP))
6078 order_vec_atom(atomGroupsPtr, cgsort, state->cg_p, rvecBuffer.buffer);
6081 if (fr->cutoff_scheme == ecutsGROUP)
6084 gmx::ArrayRef<gmx::RVec> cgcmRef = gmx::arrayRefFromArray(reinterpret_cast<gmx::RVec *>(cgcm[0]), cgsort.size());
6085 orderVector(cgsort, cgcmRef, rvecBuffer.buffer);
6088 /* Reorder the global cg index */
6089 orderVector<int>(cgsort, dd->globalAtomGroupIndices, &sort->intBuffer);
6090 /* Reorder the cginfo */
6091 orderVector<int>(cgsort, gmx::arrayRefFromArray(fr->cginfo, cgsort.size()), &sort->intBuffer);
6092 /* Rebuild the local cg index */
6095 /* We make a new, ordered atomGroups object and assign it to
6096 * the old one. This causes some allocation overhead, but saves
6097 * a copy back of the whole index.
6099 gmx::RangePartitioning ordered;
6100 for (const gmx_cgsort_t &entry : cgsort)
6102 ordered.appendBlock(atomGrouping.block(entry.ind).size());
6104 dd->atomGrouping_ = ordered;
6108 dd->atomGrouping_.setAllBlocksSizeOne(dd->ncg_home);
6110 /* Set the home atom number */
6111 dd->comm->atomRanges.setEnd(DDAtomRanges::Type::Home, dd->atomGrouping().fullRange().end());
6113 if (fr->cutoff_scheme == ecutsVERLET)
6115 /* The atoms are now exactly in grid order, update the grid order */
6116 nbnxn_set_atomorder(fr->nbv->nbs.get());
6120 /* Copy the sorted ns cell indices back to the ns grid struct */
6121 for (gmx::index i = 0; i < cgsort.size(); i++)
6123 fr->ns->grid->cell_index[i] = cgsort[i].nsc;
6125 fr->ns->grid->nr = cgsort.size();
6129 static void add_dd_statistics(gmx_domdec_t *dd)
6131 gmx_domdec_comm_t *comm = dd->comm;
6133 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
6135 auto range = static_cast<DDAtomRanges::Type>(i);
6137 comm->atomRanges.end(range) - comm->atomRanges.start(range);
6142 void reset_dd_statistics_counters(gmx_domdec_t *dd)
6144 gmx_domdec_comm_t *comm = dd->comm;
6146 /* Reset all the statistics and counters for total run counting */
6147 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
6149 comm->sum_nat[i] = 0;
6153 comm->load_step = 0;
6156 clear_ivec(comm->load_lim);
6161 void print_dd_statistics(const t_commrec *cr, const t_inputrec *ir, FILE *fplog)
6163 gmx_domdec_comm_t *comm = cr->dd->comm;
6165 const int numRanges = static_cast<int>(DDAtomRanges::Type::Number);
6166 gmx_sumd(numRanges, comm->sum_nat, cr);
6168 if (fplog == nullptr)
6173 fprintf(fplog, "\n D O M A I N D E C O M P O S I T I O N S T A T I S T I C S\n\n");
6175 for (int i = static_cast<int>(DDAtomRanges::Type::Zones); i < numRanges; i++)
6177 auto range = static_cast<DDAtomRanges::Type>(i);
6178 double av = comm->sum_nat[i]/comm->ndecomp;
6181 case DDAtomRanges::Type::Zones:
6183 " av. #atoms communicated per step for force: %d x %.1f\n",
6186 case DDAtomRanges::Type::Vsites:
6187 if (cr->dd->vsite_comm)
6190 " av. #atoms communicated per step for vsites: %d x %.1f\n",
6191 (EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD) ? 3 : 2,
6195 case DDAtomRanges::Type::Constraints:
6196 if (cr->dd->constraint_comm)
6199 " av. #atoms communicated per step for LINCS: %d x %.1f\n",
6200 1 + ir->nLincsIter, av);
6204 gmx_incons(" Unknown type for DD statistics");
6207 fprintf(fplog, "\n");
6209 if (comm->bRecordLoad && EI_DYNAMICS(ir->eI))
6211 print_dd_load_av(fplog, cr->dd);
6215 void dd_partition_system(FILE *fplog,
6217 const t_commrec *cr,
6218 gmx_bool bMasterState,
6220 t_state *state_global,
6221 const gmx_mtop_t *top_global,
6222 const t_inputrec *ir,
6223 t_state *state_local,
6224 PaddedRVecVector *f,
6225 gmx::MDAtoms *mdAtoms,
6226 gmx_localtop_t *top_local,
6229 gmx::Constraints *constr,
6231 gmx_wallcycle *wcycle,
6235 gmx_domdec_comm_t *comm;
6236 gmx_ddbox_t ddbox = {0};
6238 int64_t step_pcoupl;
6239 rvec cell_ns_x0, cell_ns_x1;
6240 int ncgindex_set, ncg_home_old = -1, ncg_moved, nat_f_novirsum;
6241 gmx_bool bBoxChanged, bNStGlobalComm, bDoDLB, bCheckWhetherToTurnDlbOn, bLogLoad;
6242 gmx_bool bRedist, bSortCG, bResortAll;
6243 ivec ncells_old = {0, 0, 0}, ncells_new = {0, 0, 0}, np;
6247 wallcycle_start(wcycle, ewcDOMDEC);
6252 // TODO if the update code becomes accessible here, use
6253 // upd->deform for this logic.
6254 bBoxChanged = (bMasterState || inputrecDeform(ir));
6255 if (ir->epc != epcNO)
6257 /* With nstpcouple > 1 pressure coupling happens.
6258 * one step after calculating the pressure.
6259 * Box scaling happens at the end of the MD step,
6260 * after the DD partitioning.
6261 * We therefore have to do DLB in the first partitioning
6262 * after an MD step where P-coupling occurred.
6263 * We need to determine the last step in which p-coupling occurred.
6264 * MRS -- need to validate this for vv?
6266 int n = ir->nstpcouple;
6269 step_pcoupl = step - 1;
6273 step_pcoupl = ((step - 1)/n)*n + 1;
6275 if (step_pcoupl >= comm->partition_step)
6281 bNStGlobalComm = (step % nstglobalcomm == 0);
6289 /* Should we do dynamic load balacing this step?
6290 * Since it requires (possibly expensive) global communication,
6291 * we might want to do DLB less frequently.
6293 if (bBoxChanged || ir->epc != epcNO)
6295 bDoDLB = bBoxChanged;
6299 bDoDLB = bNStGlobalComm;
6303 /* Check if we have recorded loads on the nodes */
6304 if (comm->bRecordLoad && dd_load_count(comm) > 0)
6306 bCheckWhetherToTurnDlbOn = dd_dlb_get_should_check_whether_to_turn_dlb_on(dd);
6308 /* Print load every nstlog, first and last step to the log file */
6309 bLogLoad = ((ir->nstlog > 0 && step % ir->nstlog == 0) ||
6310 comm->n_load_collect == 0 ||
6312 (step + ir->nstlist > ir->init_step + ir->nsteps)));
6314 /* Avoid extra communication due to verbose screen output
6315 * when nstglobalcomm is set.
6317 if (bDoDLB || bLogLoad || bCheckWhetherToTurnDlbOn ||
6318 (bVerbose && (ir->nstlist == 0 || nstglobalcomm <= ir->nstlist)))
6320 get_load_distribution(dd, wcycle);
6325 dd_print_load(fplog, dd, step-1);
6329 dd_print_load_verbose(dd);
6332 comm->n_load_collect++;
6338 /* Add the measured cycles to the running average */
6339 const float averageFactor = 0.1f;
6340 comm->cyclesPerStepDlbExpAverage =
6341 (1 - averageFactor)*comm->cyclesPerStepDlbExpAverage +
6342 averageFactor*comm->cycl[ddCyclStep]/comm->cycl_n[ddCyclStep];
6344 if (comm->dlbState == edlbsOnCanTurnOff &&
6345 dd->comm->n_load_have % c_checkTurnDlbOffInterval == c_checkTurnDlbOffInterval - 1)
6347 gmx_bool turnOffDlb;
6350 /* If the running averaged cycles with DLB are more
6351 * than before we turned on DLB, turn off DLB.
6352 * We will again run and check the cycles without DLB
6353 * and we can then decide if to turn off DLB forever.
6355 turnOffDlb = (comm->cyclesPerStepDlbExpAverage >
6356 comm->cyclesPerStepBeforeDLB);
6358 dd_bcast(dd, sizeof(turnOffDlb), &turnOffDlb);
6361 /* To turn off DLB, we need to redistribute the atoms */
6362 dd_collect_state(dd, state_local, state_global);
6363 bMasterState = TRUE;
6364 turn_off_dlb(fplog, cr, step);
6368 else if (bCheckWhetherToTurnDlbOn)
6370 gmx_bool turnOffDlbForever = FALSE;
6371 gmx_bool turnOnDlb = FALSE;
6373 /* Since the timings are node dependent, the master decides */
6376 /* If we recently turned off DLB, we want to check if
6377 * performance is better without DLB. We want to do this
6378 * ASAP to minimize the chance that external factors
6379 * slowed down the DLB step are gone here and we
6380 * incorrectly conclude that DLB was causing the slowdown.
6381 * So we measure one nstlist block, no running average.
6383 if (comm->haveTurnedOffDlb &&
6384 comm->cycl[ddCyclStep]/comm->cycl_n[ddCyclStep] <
6385 comm->cyclesPerStepDlbExpAverage)
6387 /* After turning off DLB we ran nstlist steps in fewer
6388 * cycles than with DLB. This likely means that DLB
6389 * in not benefical, but this could be due to a one
6390 * time unlucky fluctuation, so we require two such
6391 * observations in close succession to turn off DLB
6394 if (comm->dlbSlowerPartitioningCount > 0 &&
6395 dd->ddp_count < comm->dlbSlowerPartitioningCount + 10*c_checkTurnDlbOnInterval)
6397 turnOffDlbForever = TRUE;
6399 comm->haveTurnedOffDlb = false;
6400 /* Register when we last measured DLB slowdown */
6401 comm->dlbSlowerPartitioningCount = dd->ddp_count;
6405 /* Here we check if the max PME rank load is more than 0.98
6406 * the max PP force load. If so, PP DLB will not help,
6407 * since we are (almost) limited by PME. Furthermore,
6408 * DLB will cause a significant extra x/f redistribution
6409 * cost on the PME ranks, which will then surely result
6410 * in lower total performance.
6412 if (cr->npmenodes > 0 &&
6413 dd_pme_f_ratio(dd) > 1 - DD_PERF_LOSS_DLB_ON)
6419 turnOnDlb = (dd_force_imb_perf_loss(dd) >= DD_PERF_LOSS_DLB_ON);
6425 gmx_bool turnOffDlbForever;
6429 turnOffDlbForever, turnOnDlb
6431 dd_bcast(dd, sizeof(bools), &bools);
6432 if (bools.turnOffDlbForever)
6434 turn_off_dlb_forever(fplog, cr, step);
6436 else if (bools.turnOnDlb)
6438 turn_on_dlb(fplog, cr, step);
6443 comm->n_load_have++;
6446 cgs_gl = &comm->cgs_gl;
6451 /* Clear the old state */
6452 clearDDStateIndices(dd, 0, 0);
6455 auto xGlobal = positionsFromStatePointer(state_global);
6457 set_ddbox(dd, true, ir,
6458 DDMASTER(dd) ? state_global->box : nullptr,
6462 distributeState(fplog, dd, state_global, ddbox, state_local, f);
6464 dd_make_local_cgs(dd, &top_local->cgs);
6466 /* Ensure that we have space for the new distribution */
6467 dd_check_alloc_ncg(fr, state_local, f, dd->ncg_home);
6469 if (fr->cutoff_scheme == ecutsGROUP)
6471 calc_cgcm(fplog, 0, dd->ncg_home,
6472 &top_local->cgs, as_rvec_array(state_local->x.data()), fr->cg_cm);
6475 inc_nrnb(nrnb, eNR_CGCM, comm->atomRanges.numHomeAtoms());
6477 dd_set_cginfo(dd->globalAtomGroupIndices, 0, dd->ncg_home, fr, comm->bLocalCG);
6479 else if (state_local->ddp_count != dd->ddp_count)
6481 if (state_local->ddp_count > dd->ddp_count)
6483 gmx_fatal(FARGS, "Internal inconsistency state_local->ddp_count (%d) > dd->ddp_count (%ld)", state_local->ddp_count, dd->ddp_count);
6486 if (state_local->ddp_count_cg_gl != state_local->ddp_count)
6488 gmx_fatal(FARGS, "Internal inconsistency state_local->ddp_count_cg_gl (%d) != state_local->ddp_count (%d)", state_local->ddp_count_cg_gl, state_local->ddp_count);
6491 /* Clear the old state */
6492 clearDDStateIndices(dd, 0, 0);
6494 /* Restore the atom group indices from state_local */
6495 restoreAtomGroups(dd, cgs_gl->index, state_local);
6496 make_dd_indices(dd, cgs_gl->index, 0);
6497 ncgindex_set = dd->ncg_home;
6499 if (fr->cutoff_scheme == ecutsGROUP)
6501 /* Redetermine the cg COMs */
6502 calc_cgcm(fplog, 0, dd->ncg_home,
6503 &top_local->cgs, as_rvec_array(state_local->x.data()), fr->cg_cm);
6506 inc_nrnb(nrnb, eNR_CGCM, comm->atomRanges.numHomeAtoms());
6508 dd_set_cginfo(dd->globalAtomGroupIndices, 0, dd->ncg_home, fr, comm->bLocalCG);
6510 set_ddbox(dd, bMasterState, ir, state_local->box,
6511 true, state_local->x, &ddbox);
6513 bRedist = isDlbOn(comm);
6517 /* We have the full state, only redistribute the cgs */
6519 /* Clear the non-home indices */
6520 clearDDStateIndices(dd, dd->ncg_home, comm->atomRanges.numHomeAtoms());
6523 /* Avoid global communication for dim's without pbc and -gcom */
6524 if (!bNStGlobalComm)
6526 copy_rvec(comm->box0, ddbox.box0 );
6527 copy_rvec(comm->box_size, ddbox.box_size);
6529 set_ddbox(dd, bMasterState, ir, state_local->box,
6530 bNStGlobalComm, state_local->x, &ddbox);
6535 /* For dim's without pbc and -gcom */
6536 copy_rvec(ddbox.box0, comm->box0 );
6537 copy_rvec(ddbox.box_size, comm->box_size);
6539 set_dd_cell_sizes(dd, &ddbox, dynamic_dd_box(&ddbox, ir), bMasterState, bDoDLB,
6542 if (comm->nstDDDumpGrid > 0 && step % comm->nstDDDumpGrid == 0)
6544 write_dd_grid_pdb("dd_grid", step, dd, state_local->box, &ddbox);
6547 /* Check if we should sort the charge groups */
6548 bSortCG = (bMasterState || bRedist);
6550 ncg_home_old = dd->ncg_home;
6552 /* When repartitioning we mark charge groups that will move to neighboring
6553 * DD cells, but we do not move them right away for performance reasons.
6554 * Thus we need to keep track of how many charge groups will move for
6555 * obtaining correct local charge group / atom counts.
6560 wallcycle_sub_start(wcycle, ewcsDD_REDIST);
6562 dd_redistribute_cg(fplog, step, dd, ddbox.tric_dir,
6564 !bSortCG, nrnb, &ncgindex_set, &ncg_moved);
6566 wallcycle_sub_stop(wcycle, ewcsDD_REDIST);
6569 get_nsgrid_boundaries(ddbox.nboundeddim, state_local->box,
6571 &comm->cell_x0, &comm->cell_x1,
6572 dd->ncg_home, fr->cg_cm,
6573 cell_ns_x0, cell_ns_x1, &grid_density);
6577 comm_dd_ns_cell_sizes(dd, &ddbox, cell_ns_x0, cell_ns_x1, step);
6580 switch (fr->cutoff_scheme)
6583 copy_ivec(fr->ns->grid->n, ncells_old);
6584 grid_first(fplog, fr->ns->grid, dd, &ddbox,
6585 state_local->box, cell_ns_x0, cell_ns_x1,
6586 fr->rlist, grid_density);
6589 nbnxn_get_ncells(fr->nbv->nbs.get(), &ncells_old[XX], &ncells_old[YY]);
6592 gmx_incons("unimplemented");
6594 /* We need to store tric_dir for dd_get_ns_ranges called from ns.c */
6595 copy_ivec(ddbox.tric_dir, comm->tric_dir);
6599 wallcycle_sub_start(wcycle, ewcsDD_GRID);
6601 /* Sort the state on charge group position.
6602 * This enables exact restarts from this step.
6603 * It also improves performance by about 15% with larger numbers
6604 * of atoms per node.
6607 /* Fill the ns grid with the home cell,
6608 * so we can sort with the indices.
6610 set_zones_ncg_home(dd);
6612 switch (fr->cutoff_scheme)
6615 set_zones_size(dd, state_local->box, &ddbox, 0, 1, ncg_moved);
6617 nbnxn_put_on_grid(fr->nbv->nbs.get(), fr->ePBC, state_local->box,
6619 comm->zones.size[0].bb_x0,
6620 comm->zones.size[0].bb_x1,
6622 comm->zones.dens_zone0,
6624 as_rvec_array(state_local->x.data()),
6625 ncg_moved, bRedist ? comm->movedBuffer.data() : nullptr,
6626 fr->nbv->grp[eintLocal].kernel_type,
6629 nbnxn_get_ncells(fr->nbv->nbs.get(), &ncells_new[XX], &ncells_new[YY]);
6632 fill_grid(&comm->zones, fr->ns->grid, dd->ncg_home,
6633 0, dd->ncg_home, fr->cg_cm);
6635 copy_ivec(fr->ns->grid->n, ncells_new);
6638 gmx_incons("unimplemented");
6641 bResortAll = bMasterState;
6643 /* Check if we can user the old order and ns grid cell indices
6644 * of the charge groups to sort the charge groups efficiently.
6646 if (ncells_new[XX] != ncells_old[XX] ||
6647 ncells_new[YY] != ncells_old[YY] ||
6648 ncells_new[ZZ] != ncells_old[ZZ])
6655 fprintf(debug, "Step %s, sorting the %d home charge groups\n",
6656 gmx_step_str(step, sbuf), dd->ncg_home);
6658 dd_sort_state(dd, fr->cg_cm, fr, state_local,
6659 bResortAll ? -1 : ncg_home_old);
6661 /* After sorting and compacting we set the correct size */
6662 dd_resize_state(state_local, f, comm->atomRanges.numHomeAtoms());
6664 /* Rebuild all the indices */
6665 ga2la_clear(dd->ga2la);
6668 wallcycle_sub_stop(wcycle, ewcsDD_GRID);
6671 wallcycle_sub_start(wcycle, ewcsDD_SETUPCOMM);
6673 /* Setup up the communication and communicate the coordinates */
6674 setup_dd_communication(dd, state_local->box, &ddbox, fr, state_local, f);
6676 /* Set the indices */
6677 make_dd_indices(dd, cgs_gl->index, ncgindex_set);
6679 /* Set the charge group boundaries for neighbor searching */
6680 set_cg_boundaries(&comm->zones);
6682 if (fr->cutoff_scheme == ecutsVERLET)
6684 /* When bSortCG=true, we have already set the size for zone 0 */
6685 set_zones_size(dd, state_local->box, &ddbox,
6686 bSortCG ? 1 : 0, comm->zones.n,
6690 wallcycle_sub_stop(wcycle, ewcsDD_SETUPCOMM);
6693 write_dd_pdb("dd_home",step,"dump",top_global,cr,
6694 -1,as_rvec_array(state_local->x.data()),state_local->box);
6697 wallcycle_sub_start(wcycle, ewcsDD_MAKETOP);
6699 /* Extract a local topology from the global topology */
6700 for (int i = 0; i < dd->ndim; i++)
6702 np[dd->dim[i]] = comm->cd[i].numPulses();
6704 dd_make_local_top(dd, &comm->zones, dd->npbcdim, state_local->box,
6705 comm->cellsize_min, np,
6707 fr->cutoff_scheme == ecutsGROUP ? fr->cg_cm : as_rvec_array(state_local->x.data()),
6708 vsite, top_global, top_local);
6710 wallcycle_sub_stop(wcycle, ewcsDD_MAKETOP);
6712 wallcycle_sub_start(wcycle, ewcsDD_MAKECONSTR);
6714 /* Set up the special atom communication */
6715 int n = comm->atomRanges.end(DDAtomRanges::Type::Zones);
6716 for (int i = static_cast<int>(DDAtomRanges::Type::Zones) + 1; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
6718 auto range = static_cast<DDAtomRanges::Type>(i);
6721 case DDAtomRanges::Type::Vsites:
6722 if (vsite && vsite->n_intercg_vsite)
6724 n = dd_make_local_vsites(dd, n, top_local->idef.il);
6727 case DDAtomRanges::Type::Constraints:
6728 if (dd->bInterCGcons || dd->bInterCGsettles)
6730 /* Only for inter-cg constraints we need special code */
6731 n = dd_make_local_constraints(dd, n, top_global, fr->cginfo,
6732 constr, ir->nProjOrder,
6733 top_local->idef.il);
6737 gmx_incons("Unknown special atom type setup");
6739 comm->atomRanges.setEnd(range, n);
6742 wallcycle_sub_stop(wcycle, ewcsDD_MAKECONSTR);
6744 wallcycle_sub_start(wcycle, ewcsDD_TOPOTHER);
6746 /* Make space for the extra coordinates for virtual site
6747 * or constraint communication.
6749 state_local->natoms = comm->atomRanges.numAtomsTotal();
6751 dd_resize_state(state_local, f, state_local->natoms);
6753 if (fr->haveDirectVirialContributions)
6755 if (vsite && vsite->n_intercg_vsite)
6757 nat_f_novirsum = comm->atomRanges.end(DDAtomRanges::Type::Vsites);
6761 if (EEL_FULL(ir->coulombtype) && dd->n_intercg_excl > 0)
6763 nat_f_novirsum = comm->atomRanges.end(DDAtomRanges::Type::Zones);
6767 nat_f_novirsum = comm->atomRanges.numHomeAtoms();
6776 /* Set the number of atoms required for the force calculation.
6777 * Forces need to be constrained when doing energy
6778 * minimization. For simple simulations we could avoid some
6779 * allocation, zeroing and copying, but this is probably not worth
6780 * the complications and checking.
6782 forcerec_set_ranges(fr, dd->ncg_home, dd->globalAtomGroupIndices.size(),
6783 comm->atomRanges.end(DDAtomRanges::Type::Zones),
6784 comm->atomRanges.end(DDAtomRanges::Type::Constraints),
6787 /* Update atom data for mdatoms and several algorithms */
6788 mdAlgorithmsSetupAtomData(cr, ir, top_global, top_local, fr,
6789 nullptr, mdAtoms, constr, vsite, nullptr);
6791 auto mdatoms = mdAtoms->mdatoms();
6792 if (!thisRankHasDuty(cr, DUTY_PME))
6794 /* Send the charges and/or c6/sigmas to our PME only node */
6795 gmx_pme_send_parameters(cr,
6797 mdatoms->nChargePerturbed, mdatoms->nTypePerturbed,
6798 mdatoms->chargeA, mdatoms->chargeB,
6799 mdatoms->sqrt_c6A, mdatoms->sqrt_c6B,
6800 mdatoms->sigmaA, mdatoms->sigmaB,
6801 dd_pme_maxshift_x(dd), dd_pme_maxshift_y(dd));
6806 /* Update the local pull groups */
6807 dd_make_local_pull_groups(cr, ir->pull_work);
6810 if (dd->atomSets != nullptr)
6812 /* Update the local atom sets */
6813 dd->atomSets->setIndicesInDomainDecomposition(*(dd->ga2la));
6816 /* Update the local atoms to be communicated via the IMD protocol if bIMD is TRUE. */
6817 dd_make_local_IMD_atoms(ir->bIMD, dd, ir->imd);
6819 add_dd_statistics(dd);
6821 /* Make sure we only count the cycles for this DD partitioning */
6822 clear_dd_cycle_counts(dd);
6824 /* Because the order of the atoms might have changed since
6825 * the last vsite construction, we need to communicate the constructing
6826 * atom coordinates again (for spreading the forces this MD step).
6828 dd_move_x_vsites(dd, state_local->box, as_rvec_array(state_local->x.data()));
6830 wallcycle_sub_stop(wcycle, ewcsDD_TOPOTHER);
6832 if (comm->nstDDDump > 0 && step % comm->nstDDDump == 0)
6834 dd_move_x(dd, state_local->box, state_local->x, nullWallcycle);
6835 write_dd_pdb("dd_dump", step, "dump", top_global, cr,
6836 -1, as_rvec_array(state_local->x.data()), state_local->box);
6839 /* Store the partitioning step */
6840 comm->partition_step = step;
6842 /* Increase the DD partitioning counter */
6844 /* The state currently matches this DD partitioning count, store it */
6845 state_local->ddp_count = dd->ddp_count;
6848 /* The DD master node knows the complete cg distribution,
6849 * store the count so we can possibly skip the cg info communication.
6851 comm->master_cg_ddp_count = (bSortCG ? 0 : dd->ddp_count);
6854 if (comm->DD_debug > 0)
6856 /* Set the env var GMX_DD_DEBUG if you suspect corrupted indices */
6857 check_index_consistency(dd, top_global->natoms, ncg_mtop(top_global),
6858 "after partitioning");
6861 wallcycle_stop(wcycle, ewcDOMDEC);
6864 /*! \brief Check whether bonded interactions are missing, if appropriate */
6865 void checkNumberOfBondedInteractions(FILE *fplog,
6867 int totalNumberOfBondedInteractions,
6868 const gmx_mtop_t *top_global,
6869 const gmx_localtop_t *top_local,
6870 const t_state *state,
6871 bool *shouldCheckNumberOfBondedInteractions)
6873 if (*shouldCheckNumberOfBondedInteractions)
6875 if (totalNumberOfBondedInteractions != cr->dd->nbonded_global)
6877 dd_print_missing_interactions(fplog, cr, totalNumberOfBondedInteractions, top_global, top_local, state); // Does not return
6879 *shouldCheckNumberOfBondedInteractions = false;