2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018,2019, 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.
37 * \brief This file defines functions for mdrun to call to make a new
38 * domain decomposition, and check it.
40 * \author Berk Hess <hess@kth.se>
41 * \ingroup module_domdec
46 #include "partition.h"
55 #include "gromacs/domdec/collect.h"
56 #include "gromacs/domdec/dlb.h"
57 #include "gromacs/domdec/dlbtiming.h"
58 #include "gromacs/domdec/domdec.h"
59 #include "gromacs/domdec/domdec_network.h"
60 #include "gromacs/domdec/ga2la.h"
61 #include "gromacs/domdec/localatomsetmanager.h"
62 #include "gromacs/domdec/mdsetup.h"
63 #include "gromacs/ewald/pme.h"
64 #include "gromacs/gmxlib/network.h"
65 #include "gromacs/gmxlib/nrnb.h"
66 #include "gromacs/imd/imd.h"
67 #include "gromacs/math/functions.h"
68 #include "gromacs/math/vec.h"
69 #include "gromacs/mdlib/forcerec.h"
70 #include "gromacs/mdlib/gmx_omp_nthreads.h"
71 #include "gromacs/mdlib/mdatoms.h"
72 #include "gromacs/mdlib/nsgrid.h"
73 #include "gromacs/mdlib/vsite.h"
74 #include "gromacs/mdtypes/commrec.h"
75 #include "gromacs/mdtypes/forcerec.h"
76 #include "gromacs/mdtypes/inputrec.h"
77 #include "gromacs/mdtypes/md_enums.h"
78 #include "gromacs/mdtypes/nblist.h"
79 #include "gromacs/mdtypes/state.h"
80 #include "gromacs/nbnxm/nbnxm.h"
81 #include "gromacs/pulling/pull.h"
82 #include "gromacs/timing/wallcycle.h"
83 #include "gromacs/topology/mtop_util.h"
84 #include "gromacs/topology/topology.h"
85 #include "gromacs/utility/cstringutil.h"
86 #include "gromacs/utility/fatalerror.h"
87 #include "gromacs/utility/logger.h"
88 #include "gromacs/utility/real.h"
89 #include "gromacs/utility/smalloc.h"
90 #include "gromacs/utility/strconvert.h"
91 #include "gromacs/utility/stringstream.h"
92 #include "gromacs/utility/stringutil.h"
93 #include "gromacs/utility/textwriter.h"
96 #include "cellsizes.h"
97 #include "distribute.h"
98 #include "domdec_constraints.h"
99 #include "domdec_internal.h"
100 #include "domdec_vsite.h"
102 #include "redistribute.h"
105 /*! \brief Turn on DLB when the load imbalance causes this amount of total loss.
107 * There is a bit of overhead with DLB and it's difficult to achieve
108 * a load imbalance of less than 2% with DLB.
110 #define DD_PERF_LOSS_DLB_ON 0.02
112 //! Warn about imbalance due to PP or PP/PME load imbalance at this loss.
113 #define DD_PERF_LOSS_WARN 0.05
116 //! Debug helper printing a DD zone
117 static void print_ddzone(FILE *fp, int d, int i, int j, gmx_ddzone_t *zone)
119 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",
121 zone->min0, zone->max1,
122 zone->mch0, zone->mch0,
123 zone->p1_0, zone->p1_1);
126 /*! \brief Using the home grid size as input in cell_ns_x0 and cell_ns_x1
127 * takes the extremes over all home and remote zones in the halo
128 * and returns the results in cell_ns_x0 and cell_ns_x1.
129 * Note: only used with the group cut-off scheme.
131 static void dd_move_cellx(gmx_domdec_t *dd,
132 const gmx_ddbox_t *ddbox,
136 constexpr int c_ddZoneCommMaxNumZones = 5;
137 gmx_ddzone_t buf_s[c_ddZoneCommMaxNumZones];
138 gmx_ddzone_t buf_r[c_ddZoneCommMaxNumZones];
139 gmx_ddzone_t buf_e[c_ddZoneCommMaxNumZones];
140 gmx_domdec_comm_t *comm = dd->comm;
144 for (int d = 1; d < dd->ndim; d++)
146 int dim = dd->dim[d];
147 gmx_ddzone_t &zp = (d == 1) ? comm->zone_d1[0] : comm->zone_d2[0][0];
149 /* Copy the base sizes of the home zone */
150 zp.min0 = cell_ns_x0[dim];
151 zp.max1 = cell_ns_x1[dim];
152 zp.min1 = cell_ns_x1[dim];
153 zp.mch0 = cell_ns_x0[dim];
154 zp.mch1 = cell_ns_x1[dim];
155 zp.p1_0 = cell_ns_x0[dim];
156 zp.p1_1 = cell_ns_x1[dim];
160 gmx::ArrayRef<DDCellsizesWithDlb> cellsizes = comm->cellsizesWithDlb;
162 /* Loop backward over the dimensions and aggregate the extremes
165 for (int d = dd->ndim - 2; d >= 0; d--)
167 const int dim = dd->dim[d];
168 const bool applyPbc = (dim < ddbox->npbcdim);
170 /* Use an rvec to store two reals */
171 extr_s[d][0] = cellsizes[d + 1].fracLower;
172 extr_s[d][1] = cellsizes[d + 1].fracUpper;
173 extr_s[d][2] = cellsizes[d + 1].fracUpper;
176 GMX_ASSERT(pos < c_ddZoneCommMaxNumZones, "The buffers should be sufficiently large");
177 /* Store the extremes in the backward sending buffer,
178 * so they get updated separately from the forward communication.
180 for (int d1 = d; d1 < dd->ndim-1; d1++)
182 gmx_ddzone_t &buf = buf_s[pos];
184 /* We invert the order to be able to use the same loop for buf_e */
185 buf.min0 = extr_s[d1][1];
186 buf.max1 = extr_s[d1][0];
187 buf.min1 = extr_s[d1][2];
190 /* Store the cell corner of the dimension we communicate along */
191 buf.p1_0 = comm->cell_x0[dim];
197 buf_s[pos] = (dd->ndim == 2) ? comm->zone_d1[0] : comm->zone_d2[0][0];
200 if (dd->ndim == 3 && d == 0)
202 buf_s[pos] = comm->zone_d2[0][1];
204 buf_s[pos] = comm->zone_d1[0];
208 /* We only need to communicate the extremes
209 * in the forward direction
211 int numPulses = comm->cd[d].numPulses();
215 /* Take the minimum to avoid double communication */
216 numPulsesMin = std::min(numPulses, dd->nc[dim] - 1 - numPulses);
220 /* Without PBC we should really not communicate over
221 * the boundaries, but implementing that complicates
222 * the communication setup and therefore we simply
223 * do all communication, but ignore some data.
225 numPulsesMin = numPulses;
227 for (int pulse = 0; pulse < numPulsesMin; pulse++)
229 /* Communicate the extremes forward */
230 bool receiveValidData = (applyPbc || dd->ci[dim] > 0);
232 int numElements = dd->ndim - d - 1;
233 ddSendrecv(dd, d, dddirForward,
234 extr_s + d, numElements,
235 extr_r + d, numElements);
237 if (receiveValidData)
239 for (int d1 = d; d1 < dd->ndim - 1; d1++)
241 extr_s[d1][0] = std::max(extr_s[d1][0], extr_r[d1][0]);
242 extr_s[d1][1] = std::min(extr_s[d1][1], extr_r[d1][1]);
243 extr_s[d1][2] = std::min(extr_s[d1][2], extr_r[d1][2]);
248 const int numElementsInBuffer = pos;
249 for (int pulse = 0; pulse < numPulses; pulse++)
251 /* Communicate all the zone information backward */
252 bool receiveValidData = (applyPbc || dd->ci[dim] < dd->nc[dim] - 1);
254 static_assert(sizeof(gmx_ddzone_t) == c_ddzoneNumReals*sizeof(real), "Here we expect gmx_ddzone_t to consist of c_ddzoneNumReals reals (only)");
256 int numReals = numElementsInBuffer*c_ddzoneNumReals;
257 ddSendrecv(dd, d, dddirBackward,
258 gmx::arrayRefFromArray(&buf_s[0].min0, numReals),
259 gmx::arrayRefFromArray(&buf_r[0].min0, numReals));
264 for (int d1 = d + 1; d1 < dd->ndim; d1++)
266 /* Determine the decrease of maximum required
267 * communication height along d1 due to the distance along d,
268 * this avoids a lot of useless atom communication.
270 real dist_d = comm->cell_x1[dim] - buf_r[0].p1_0;
273 if (ddbox->tric_dir[dim])
275 /* c is the off-diagonal coupling between the cell planes
276 * along directions d and d1.
278 c = ddbox->v[dim][dd->dim[d1]][dim];
284 real det = (1 + c*c)*gmx::square(comm->systemInfo.cutoff) - dist_d*dist_d;
287 dh[d1] = comm->systemInfo.cutoff - (c*dist_d + std::sqrt(det))/(1 + c*c);
291 /* A negative value signals out of range */
297 /* Accumulate the extremes over all pulses */
298 for (int i = 0; i < numElementsInBuffer; i++)
306 if (receiveValidData)
308 buf_e[i].min0 = std::min(buf_e[i].min0, buf_r[i].min0);
309 buf_e[i].max1 = std::max(buf_e[i].max1, buf_r[i].max1);
310 buf_e[i].min1 = std::min(buf_e[i].min1, buf_r[i].min1);
314 if (dd->ndim == 3 && d == 0 && i == numElementsInBuffer - 1)
322 if (receiveValidData && dh[d1] >= 0)
324 buf_e[i].mch0 = std::max(buf_e[i].mch0, buf_r[i].mch0-dh[d1]);
325 buf_e[i].mch1 = std::max(buf_e[i].mch1, buf_r[i].mch1-dh[d1]);
328 /* Copy the received buffer to the send buffer,
329 * to pass the data through with the next pulse.
333 if (((applyPbc || dd->ci[dim] + numPulses < dd->nc[dim]) && pulse == numPulses - 1) ||
334 (!applyPbc && dd->ci[dim] + 1 + pulse == dd->nc[dim] - 1))
336 /* Store the extremes */
339 for (int d1 = d; d1 < dd->ndim-1; d1++)
341 extr_s[d1][1] = std::min(extr_s[d1][1], buf_e[pos].min0);
342 extr_s[d1][0] = std::max(extr_s[d1][0], buf_e[pos].max1);
343 extr_s[d1][2] = std::min(extr_s[d1][2], buf_e[pos].min1);
347 if (d == 1 || (d == 0 && dd->ndim == 3))
349 for (int i = d; i < 2; i++)
351 comm->zone_d2[1-d][i] = buf_e[pos];
357 comm->zone_d1[1] = buf_e[pos];
363 if (d == 1 || (d == 0 && dd->ndim == 3))
365 for (int i = d; i < 2; i++)
367 comm->zone_d2[1 - d][i].dataSet = 0;
372 comm->zone_d1[1].dataSet = 0;
380 int dim = dd->dim[1];
381 for (int i = 0; i < 2; i++)
383 if (comm->zone_d1[i].dataSet != 0)
387 print_ddzone(debug, 1, i, 0, &comm->zone_d1[i]);
389 cell_ns_x0[dim] = std::min(cell_ns_x0[dim], comm->zone_d1[i].min0);
390 cell_ns_x1[dim] = std::max(cell_ns_x1[dim], comm->zone_d1[i].max1);
396 int dim = dd->dim[2];
397 for (int i = 0; i < 2; i++)
399 for (int j = 0; j < 2; j++)
401 if (comm->zone_d2[i][j].dataSet != 0)
405 print_ddzone(debug, 2, i, j, &comm->zone_d2[i][j]);
407 cell_ns_x0[dim] = std::min(cell_ns_x0[dim], comm->zone_d2[i][j].min0);
408 cell_ns_x1[dim] = std::max(cell_ns_x1[dim], comm->zone_d2[i][j].max1);
413 for (int d = 1; d < dd->ndim; d++)
415 cellsizes[d].fracLowerMax = extr_s[d-1][0];
416 cellsizes[d].fracUpperMin = extr_s[d-1][1];
419 fprintf(debug, "Cell fraction d %d, max0 %f, min1 %f\n",
420 d, cellsizes[d].fracLowerMax, cellsizes[d].fracUpperMin);
425 //! Sets the charge-group zones to be equal to the home zone.
426 static void set_zones_ncg_home(gmx_domdec_t *dd)
428 gmx_domdec_zones_t *zones;
431 zones = &dd->comm->zones;
433 zones->cg_range[0] = 0;
434 for (i = 1; i < zones->n+1; i++)
436 zones->cg_range[i] = dd->ncg_home;
438 /* zone_ncg1[0] should always be equal to ncg_home */
439 dd->comm->zone_ncg1[0] = dd->ncg_home;
442 //! Restore atom groups for the charge groups.
443 static void restoreAtomGroups(gmx_domdec_t *dd,
444 const t_state *state)
446 gmx::ArrayRef<const int> atomGroupsState = state->cg_gl;
448 std::vector<int> &globalAtomGroupIndices = dd->globalAtomGroupIndices;
450 globalAtomGroupIndices.resize(atomGroupsState.size());
452 /* Copy back the global charge group indices from state
453 * and rebuild the local charge group to atom index.
455 for (gmx::index i = 0; i < atomGroupsState.ssize(); i++)
457 globalAtomGroupIndices[i] = atomGroupsState[i];
460 dd->ncg_home = atomGroupsState.size();
461 dd->comm->atomRanges.setEnd(DDAtomRanges::Type::Home, atomGroupsState.ssize());
463 set_zones_ncg_home(dd);
466 //! Sets the cginfo structures.
467 static void dd_set_cginfo(gmx::ArrayRef<const int> index_gl, int cg0, int cg1,
472 const cginfo_mb_t *cginfo_mb = fr->cginfo_mb;
473 gmx::ArrayRef<int> cginfo = fr->cginfo;
475 for (int cg = cg0; cg < cg1; cg++)
477 cginfo[cg] = ddcginfo(cginfo_mb, index_gl[cg]);
482 //! Makes the mappings between global and local atom indices during DD repartioning.
483 static void make_dd_indices(gmx_domdec_t *dd,
486 const int numZones = dd->comm->zones.n;
487 const int *zone2cg = dd->comm->zones.cg_range;
488 const int *zone_ncg1 = dd->comm->zone_ncg1;
489 gmx::ArrayRef<const int> globalAtomGroupIndices = dd->globalAtomGroupIndices;
491 std::vector<int> &globalAtomIndices = dd->globalAtomIndices;
492 gmx_ga2la_t &ga2la = *dd->ga2la;
494 if (zone2cg[1] != dd->ncg_home)
496 gmx_incons("dd->ncg_zone is not up to date");
499 /* Make the local to global and global to local atom index */
501 globalAtomIndices.resize(a);
502 for (int zone = 0; zone < numZones; zone++)
513 int cg1 = zone2cg[zone+1];
514 int cg1_p1 = cg0 + zone_ncg1[zone];
516 for (int cg = cg0; cg < cg1; cg++)
521 /* Signal that this cg is from more than one pulse away */
524 int cg_gl = globalAtomGroupIndices[cg];
525 globalAtomIndices.push_back(cg_gl);
526 ga2la.insert(cg_gl, {a, zone1});
532 //! Checks whether global and local atom indices are consistent.
533 static void check_index_consistency(const gmx_domdec_t *dd,
539 const int numAtomsInZones = dd->comm->atomRanges.end(DDAtomRanges::Type::Zones);
541 if (dd->comm->ddSettings.DD_debug > 1)
543 std::vector<int> have(natoms_sys);
544 for (int a = 0; a < numAtomsInZones; a++)
546 int globalAtomIndex = dd->globalAtomIndices[a];
547 if (have[globalAtomIndex] > 0)
549 fprintf(stderr, "DD rank %d: global atom %d occurs twice: index %d and %d\n", dd->rank, globalAtomIndex + 1, have[globalAtomIndex], a+1);
553 have[globalAtomIndex] = a + 1;
558 std::vector<int> have(numAtomsInZones);
561 for (int i = 0; i < natoms_sys; i++)
563 if (const auto entry = dd->ga2la->find(i))
565 const int a = entry->la;
566 if (a >= numAtomsInZones)
568 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);
574 if (dd->globalAtomIndices[a] != i)
576 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);
583 if (ngl != numAtomsInZones)
586 "DD rank %d, %s: %d global atom indices, %d local atoms\n",
587 dd->rank, where, ngl, numAtomsInZones);
589 for (int a = 0; a < numAtomsInZones; a++)
594 "DD rank %d, %s: local atom %d, global %d has no global index\n",
595 dd->rank, where, a + 1, dd->globalAtomIndices[a] + 1);
601 gmx_fatal(FARGS, "DD rank %d, %s: %d atom(group) index inconsistencies",
602 dd->rank, where, nerr);
606 //! Clear all DD global state indices
607 static void clearDDStateIndices(gmx_domdec_t *dd,
608 const bool keepLocalAtomIndices)
610 gmx_ga2la_t &ga2la = *dd->ga2la;
612 if (!keepLocalAtomIndices)
614 /* Clear the whole list without the overhead of searching */
619 const int numAtomsInZones = dd->comm->atomRanges.end(DDAtomRanges::Type::Zones);
620 for (int i = 0; i < numAtomsInZones; i++)
622 ga2la.erase(dd->globalAtomIndices[i]);
626 dd_clear_local_vsite_indices(dd);
630 dd_clear_local_constraint_indices(dd);
634 bool check_grid_jump(int64_t step,
635 const gmx_domdec_t *dd,
637 const gmx_ddbox_t *ddbox,
640 gmx_domdec_comm_t *comm = dd->comm;
641 bool invalid = false;
643 for (int d = 1; d < dd->ndim; d++)
645 const DDCellsizesWithDlb &cellsizes = comm->cellsizesWithDlb[d];
646 const int dim = dd->dim[d];
647 const real limit = grid_jump_limit(comm, cutoff, d);
648 real bfac = ddbox->box_size[dim];
649 if (ddbox->tric_dir[dim])
651 bfac *= ddbox->skew_fac[dim];
653 if ((cellsizes.fracUpper - cellsizes.fracLowerMax)*bfac < limit ||
654 (cellsizes.fracLower - cellsizes.fracUpperMin)*bfac > -limit)
662 /* This error should never be triggered under normal
663 * circumstances, but you never know ...
665 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.",
666 gmx_step_str(step, buf),
667 dim2char(dim), dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
674 //! Return the duration of force calculations on this rank.
675 static float dd_force_load(gmx_domdec_comm_t *comm)
679 if (comm->ddSettings.eFlop)
682 if (comm->ddSettings.eFlop > 1)
684 load *= 1.0 + (comm->ddSettings.eFlop - 1)*(0.1*rand()/RAND_MAX - 0.05);
689 load = comm->cycl[ddCyclF];
690 if (comm->cycl_n[ddCyclF] > 1)
692 /* Subtract the maximum of the last n cycle counts
693 * to get rid of possible high counts due to other sources,
694 * for instance system activity, that would otherwise
695 * affect the dynamic load balancing.
697 load -= comm->cycl_max[ddCyclF];
701 if (comm->cycl_n[ddCyclWaitGPU] && comm->nrank_gpu_shared > 1)
703 float gpu_wait, gpu_wait_sum;
705 gpu_wait = comm->cycl[ddCyclWaitGPU];
706 if (comm->cycl_n[ddCyclF] > 1)
708 /* We should remove the WaitGPU time of the same MD step
709 * as the one with the maximum F time, since the F time
710 * and the wait time are not independent.
711 * Furthermore, the step for the max F time should be chosen
712 * the same on all ranks that share the same GPU.
713 * But to keep the code simple, we remove the average instead.
714 * The main reason for artificially long times at some steps
715 * is spurious CPU activity or MPI time, so we don't expect
716 * that changes in the GPU wait time matter a lot here.
718 gpu_wait *= (comm->cycl_n[ddCyclF] - 1)/static_cast<float>(comm->cycl_n[ddCyclF]);
720 /* Sum the wait times over the ranks that share the same GPU */
721 MPI_Allreduce(&gpu_wait, &gpu_wait_sum, 1, MPI_FLOAT, MPI_SUM,
722 comm->mpi_comm_gpu_shared);
723 /* Replace the wait time by the average over the ranks */
724 load += -gpu_wait + gpu_wait_sum/comm->nrank_gpu_shared;
732 //! Runs cell size checks and communicates the boundaries.
733 static void comm_dd_ns_cell_sizes(gmx_domdec_t *dd,
735 rvec cell_ns_x0, rvec cell_ns_x1,
738 gmx_domdec_comm_t *comm;
743 for (dim_ind = 0; dim_ind < dd->ndim; dim_ind++)
745 dim = dd->dim[dim_ind];
747 /* Without PBC we don't have restrictions on the outer cells */
748 if (!(dim >= ddbox->npbcdim &&
749 (dd->ci[dim] == 0 || dd->ci[dim] == dd->nc[dim] - 1)) &&
751 (comm->cell_x1[dim] - comm->cell_x0[dim])*ddbox->skew_fac[dim] <
752 comm->cellsize_min[dim])
755 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",
756 gmx_step_str(step, buf), dim2char(dim),
757 comm->cell_x1[dim] - comm->cell_x0[dim],
758 ddbox->skew_fac[dim],
759 dd->comm->cellsize_min[dim],
760 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
764 if ((isDlbOn(dd->comm) && dd->ndim > 1) || ddbox->nboundeddim < DIM)
766 /* Communicate the boundaries and update cell_ns_x0/1 */
767 dd_move_cellx(dd, ddbox, cell_ns_x0, cell_ns_x1);
768 if (isDlbOn(dd->comm) && dd->ndim > 1)
770 check_grid_jump(step, dd, dd->comm->systemInfo.cutoff, ddbox, TRUE);
775 //! Compute and communicate to determine the load distribution across PP ranks.
776 static void get_load_distribution(gmx_domdec_t *dd, gmx_wallcycle_t wcycle)
778 gmx_domdec_comm_t *comm;
780 float cell_frac = 0, sbuf[DD_NLOAD_MAX];
785 fprintf(debug, "get_load_distribution start\n");
788 wallcycle_start(wcycle, ewcDDCOMMLOAD);
792 bSepPME = (dd->pme_nodeid >= 0);
794 if (dd->ndim == 0 && bSepPME)
796 /* Without decomposition, but with PME nodes, we need the load */
797 comm->load[0].mdf = comm->cycl[ddCyclPPduringPME];
798 comm->load[0].pme = comm->cycl[ddCyclPME];
801 for (int d = dd->ndim - 1; d >= 0; d--)
803 const DDCellsizesWithDlb *cellsizes = (isDlbOn(dd->comm) ? &comm->cellsizesWithDlb[d] : nullptr);
804 const int dim = dd->dim[d];
805 /* Check if we participate in the communication in this dimension */
806 if (d == dd->ndim-1 ||
807 (dd->ci[dd->dim[d+1]] == 0 && dd->ci[dd->dim[dd->ndim-1]] == 0))
809 load = &comm->load[d];
810 if (isDlbOn(dd->comm))
812 cell_frac = cellsizes->fracUpper - cellsizes->fracLower;
817 sbuf[pos++] = dd_force_load(comm);
818 sbuf[pos++] = sbuf[0];
819 if (isDlbOn(dd->comm))
821 sbuf[pos++] = sbuf[0];
822 sbuf[pos++] = cell_frac;
825 sbuf[pos++] = cellsizes->fracLowerMax;
826 sbuf[pos++] = cellsizes->fracUpperMin;
831 sbuf[pos++] = comm->cycl[ddCyclPPduringPME];
832 sbuf[pos++] = comm->cycl[ddCyclPME];
837 sbuf[pos++] = comm->load[d+1].sum;
838 sbuf[pos++] = comm->load[d+1].max;
839 if (isDlbOn(dd->comm))
841 sbuf[pos++] = comm->load[d+1].sum_m;
842 sbuf[pos++] = comm->load[d+1].cvol_min*cell_frac;
843 sbuf[pos++] = comm->load[d+1].flags;
846 sbuf[pos++] = cellsizes->fracLowerMax;
847 sbuf[pos++] = cellsizes->fracUpperMin;
852 sbuf[pos++] = comm->load[d+1].mdf;
853 sbuf[pos++] = comm->load[d+1].pme;
857 /* Communicate a row in DD direction d.
858 * The communicators are setup such that the root always has rank 0.
861 MPI_Gather(sbuf, load->nload*sizeof(float), MPI_BYTE,
862 load->load, load->nload*sizeof(float), MPI_BYTE,
863 0, comm->mpi_comm_load[d]);
865 if (dd->ci[dim] == dd->master_ci[dim])
867 /* We are the master along this row, process this row */
868 RowMaster *rowMaster = nullptr;
872 rowMaster = cellsizes->rowMaster.get();
882 for (int i = 0; i < dd->nc[dim]; i++)
884 load->sum += load->load[pos++];
885 load->max = std::max(load->max, load->load[pos]);
887 if (isDlbOn(dd->comm))
889 if (rowMaster->dlbIsLimited)
891 /* This direction could not be load balanced properly,
892 * therefore we need to use the maximum iso the average load.
894 load->sum_m = std::max(load->sum_m, load->load[pos]);
898 load->sum_m += load->load[pos];
901 load->cvol_min = std::min(load->cvol_min, load->load[pos]);
905 load->flags = gmx::roundToInt(load->load[pos++]);
909 rowMaster->bounds[i].cellFracLowerMax = load->load[pos++];
910 rowMaster->bounds[i].cellFracUpperMin = load->load[pos++];
915 load->mdf = std::max(load->mdf, load->load[pos]);
917 load->pme = std::max(load->pme, load->load[pos]);
921 if (isDlbOn(comm) && rowMaster->dlbIsLimited)
923 load->sum_m *= dd->nc[dim];
924 load->flags |= (1<<d);
932 comm->nload += dd_load_count(comm);
933 comm->load_step += comm->cycl[ddCyclStep];
934 comm->load_sum += comm->load[0].sum;
935 comm->load_max += comm->load[0].max;
938 for (int d = 0; d < dd->ndim; d++)
940 if (comm->load[0].flags & (1<<d))
948 comm->load_mdf += comm->load[0].mdf;
949 comm->load_pme += comm->load[0].pme;
953 wallcycle_stop(wcycle, ewcDDCOMMLOAD);
957 fprintf(debug, "get_load_distribution finished\n");
961 /*! \brief Return the relative performance loss on the total run time
962 * due to the force calculation load imbalance. */
963 static float dd_force_load_fraction(gmx_domdec_t *dd)
965 if (dd->comm->nload > 0 && dd->comm->load_step > 0)
967 return dd->comm->load_sum/(dd->comm->load_step*dd->nnodes);
975 /*! \brief Return the relative performance loss on the total run time
976 * due to the force calculation load imbalance. */
977 static float dd_force_imb_perf_loss(gmx_domdec_t *dd)
979 if (dd->comm->nload > 0 && dd->comm->load_step > 0)
982 (dd->comm->load_max*dd->nnodes - dd->comm->load_sum)/
983 (dd->comm->load_step*dd->nnodes);
991 //! Print load-balance report e.g. at the end of a run.
992 static void print_dd_load_av(FILE *fplog, gmx_domdec_t *dd)
994 gmx_domdec_comm_t *comm = dd->comm;
996 /* Only the master rank prints loads and only if we measured loads */
997 if (!DDMASTER(dd) || comm->nload == 0)
1003 int numPpRanks = dd->nnodes;
1004 int numPmeRanks = (comm->ddRankSetup.usePmeOnlyRanks ? comm->ddRankSetup.numRanksDoingPme : 0);
1005 int numRanks = numPpRanks + numPmeRanks;
1006 float lossFraction = 0;
1008 /* Print the average load imbalance and performance loss */
1009 if (dd->nnodes > 1 && comm->load_sum > 0)
1011 float imbalance = comm->load_max*numPpRanks/comm->load_sum - 1;
1012 lossFraction = dd_force_imb_perf_loss(dd);
1014 std::string msg = "\nDynamic load balancing report:\n";
1015 std::string dlbStateStr;
1017 switch (dd->comm->dlbState)
1019 case DlbState::offUser:
1020 dlbStateStr = "DLB was off during the run per user request.";
1022 case DlbState::offForever:
1023 /* Currectly this can happen due to performance loss observed, cell size
1024 * limitations or incompatibility with other settings observed during
1025 * determineInitialDlbState(). */
1026 dlbStateStr = "DLB got disabled because it was unsuitable to use.";
1028 case DlbState::offCanTurnOn:
1029 dlbStateStr = "DLB was off during the run due to low measured imbalance.";
1031 case DlbState::offTemporarilyLocked:
1032 dlbStateStr = "DLB was locked at the end of the run due to unfinished PP-PME balancing.";
1034 case DlbState::onCanTurnOff:
1035 dlbStateStr = "DLB was turned on during the run due to measured imbalance.";
1037 case DlbState::onUser:
1038 dlbStateStr = "DLB was permanently on during the run per user request.";
1041 GMX_ASSERT(false, "Undocumented DLB state");
1044 msg += " " + dlbStateStr + "\n";
1045 msg += gmx::formatString(" Average load imbalance: %.1f%%.\n", imbalance*100);
1046 msg += gmx::formatString(" The balanceable part of the MD step is %d%%, load imbalance is computed from this.\n",
1047 gmx::roundToInt(dd_force_load_fraction(dd)*100));
1048 msg += gmx::formatString(" Part of the total run time spent waiting due to load imbalance: %.1f%%.\n",
1050 fprintf(fplog, "%s", msg.c_str());
1051 fprintf(stderr, "\n%s", msg.c_str());
1054 /* Print during what percentage of steps the load balancing was limited */
1055 bool dlbWasLimited = false;
1058 sprintf(buf, " Steps where the load balancing was limited by -rdd, -rcon and/or -dds:");
1059 for (int d = 0; d < dd->ndim; d++)
1061 int limitPercentage = (200*comm->load_lim[d] + 1)/(2*comm->nload);
1062 sprintf(buf+strlen(buf), " %c %d %%",
1063 dim2char(dd->dim[d]), limitPercentage);
1064 if (limitPercentage >= 50)
1066 dlbWasLimited = true;
1069 sprintf(buf + strlen(buf), "\n");
1070 fprintf(fplog, "%s", buf);
1071 fprintf(stderr, "%s", buf);
1074 /* Print the performance loss due to separate PME - PP rank imbalance */
1075 float lossFractionPme = 0;
1076 if (numPmeRanks > 0 && comm->load_mdf > 0 && comm->load_step > 0)
1078 float pmeForceRatio = comm->load_pme/comm->load_mdf;
1079 lossFractionPme = (comm->load_pme - comm->load_mdf)/comm->load_step;
1080 if (lossFractionPme <= 0)
1082 lossFractionPme *= numPmeRanks/static_cast<float>(numRanks);
1086 lossFractionPme *= numPpRanks/static_cast<float>(numRanks);
1088 sprintf(buf, " Average PME mesh/force load: %5.3f\n", pmeForceRatio);
1089 fprintf(fplog, "%s", buf);
1090 fprintf(stderr, "%s", buf);
1091 sprintf(buf, " Part of the total run time spent waiting due to PP/PME imbalance: %.1f %%\n", std::fabs(lossFractionPme)*100);
1092 fprintf(fplog, "%s", buf);
1093 fprintf(stderr, "%s", buf);
1095 fprintf(fplog, "\n");
1096 fprintf(stderr, "\n");
1098 if (lossFraction >= DD_PERF_LOSS_WARN)
1100 std::string message = gmx::formatString(
1101 "NOTE: %.1f %% of the available CPU time was lost due to load imbalance\n"
1102 " in the domain decomposition.\n", lossFraction*100);
1104 bool hadSuggestion = false;
1107 message += " You might want to use dynamic load balancing (option -dlb.)\n";
1108 hadSuggestion = true;
1110 else if (dlbWasLimited)
1112 message += " You might want to decrease the cell size limit (options -rdd, -rcon and/or -dds).\n";
1113 hadSuggestion = true;
1115 message += gmx::formatString(
1116 " You can %sconsider manually changing the decomposition (option -dd);\n"
1117 " e.g. by using fewer domains along the box dimension in which there is\n"
1118 " considerable inhomogeneity in the simulated system.",
1119 hadSuggestion ? "also " : "");
1122 fprintf(fplog, "%s\n", message.c_str());
1123 fprintf(stderr, "%s\n", message.c_str());
1125 if (numPmeRanks > 0 && std::fabs(lossFractionPme) >= DD_PERF_LOSS_WARN)
1128 "NOTE: %.1f %% performance was lost because the PME ranks\n"
1129 " had %s work to do than the PP ranks.\n"
1130 " You might want to %s the number of PME ranks\n"
1131 " or %s the cut-off and the grid spacing.\n",
1132 std::fabs(lossFractionPme*100),
1133 (lossFractionPme < 0) ? "less" : "more",
1134 (lossFractionPme < 0) ? "decrease" : "increase",
1135 (lossFractionPme < 0) ? "decrease" : "increase");
1136 fprintf(fplog, "%s\n", buf);
1137 fprintf(stderr, "%s\n", buf);
1141 //! Return the minimum communication volume.
1142 static float dd_vol_min(gmx_domdec_t *dd)
1144 return dd->comm->load[0].cvol_min*dd->nnodes;
1147 //! Return the DD load flags.
1148 static int dd_load_flags(gmx_domdec_t *dd)
1150 return dd->comm->load[0].flags;
1153 //! Return the reported load imbalance in force calculations.
1154 static float dd_f_imbal(gmx_domdec_t *dd)
1156 if (dd->comm->load[0].sum > 0)
1158 return dd->comm->load[0].max*dd->nnodes/dd->comm->load[0].sum - 1.0F;
1162 /* Something is wrong in the cycle counting, report no load imbalance */
1167 //! Returns DD load balance report.
1169 dd_print_load(gmx_domdec_t *dd,
1172 gmx::StringOutputStream stream;
1173 gmx::TextWriter log(&stream);
1175 int flags = dd_load_flags(dd);
1178 log.writeString("DD load balancing is limited by minimum cell size in dimension");
1179 for (int d = 0; d < dd->ndim; d++)
1183 log.writeStringFormatted(" %c", dim2char(dd->dim[d]));
1186 log.ensureLineBreak();
1188 log.writeString("DD step " + gmx::toString(step));
1189 if (isDlbOn(dd->comm))
1191 log.writeStringFormatted(" vol min/aver %5.3f%c",
1192 dd_vol_min(dd), flags ? '!' : ' ');
1196 log.writeStringFormatted(" load imb.: force %4.1f%%", dd_f_imbal(dd)*100);
1198 if (dd->comm->cycl_n[ddCyclPME])
1200 log.writeStringFormatted(" pme mesh/force %5.3f", dd_pme_f_ratio(dd));
1202 log.ensureLineBreak();
1203 return stream.toString();
1206 //! Prints DD load balance report in mdrun verbose mode.
1207 static void dd_print_load_verbose(gmx_domdec_t *dd)
1209 if (isDlbOn(dd->comm))
1211 fprintf(stderr, "vol %4.2f%c ",
1212 dd_vol_min(dd), dd_load_flags(dd) ? '!' : ' ');
1216 fprintf(stderr, "imb F %2d%% ", gmx::roundToInt(dd_f_imbal(dd)*100));
1218 if (dd->comm->cycl_n[ddCyclPME])
1220 fprintf(stderr, "pme/F %4.2f ", dd_pme_f_ratio(dd));
1224 //! Turns on dynamic load balancing if possible and needed.
1225 static void turn_on_dlb(const gmx::MDLogger &mdlog,
1229 gmx_domdec_comm_t *comm = dd->comm;
1231 real cellsize_min = comm->cellsize_min[dd->dim[0]];
1232 for (int d = 1; d < dd->ndim; d++)
1234 cellsize_min = std::min(cellsize_min, comm->cellsize_min[dd->dim[d]]);
1237 /* Turn off DLB if we're too close to the cell size limit. */
1238 if (cellsize_min < comm->cellsize_limit*1.05)
1240 GMX_LOG(mdlog.info).appendTextFormatted(
1241 "step %s Measured %.1f %% performance loss due to load imbalance, "
1242 "but the minimum cell size is smaller than 1.05 times the cell size limit. "
1243 "Will no longer try dynamic load balancing.",
1244 gmx::toString(step).c_str(), dd_force_imb_perf_loss(dd)*100);
1246 comm->dlbState = DlbState::offForever;
1250 GMX_LOG(mdlog.info).appendTextFormatted(
1251 "step %s Turning on dynamic load balancing, because the performance loss due to load imbalance is %.1f %%.",
1252 gmx::toString(step).c_str(), dd_force_imb_perf_loss(dd)*100);
1253 comm->dlbState = DlbState::onCanTurnOff;
1255 /* Store the non-DLB performance, so we can check if DLB actually
1256 * improves performance.
1258 GMX_RELEASE_ASSERT(comm->cycl_n[ddCyclStep] > 0, "When we turned on DLB, we should have measured cycles");
1259 comm->cyclesPerStepBeforeDLB = comm->cycl[ddCyclStep]/comm->cycl_n[ddCyclStep];
1263 /* We can set the required cell size info here,
1264 * so we do not need to communicate this.
1265 * The grid is completely uniform.
1267 for (int d = 0; d < dd->ndim; d++)
1269 RowMaster *rowMaster = comm->cellsizesWithDlb[d].rowMaster.get();
1273 comm->load[d].sum_m = comm->load[d].sum;
1275 int nc = dd->nc[dd->dim[d]];
1276 for (int i = 0; i < nc; i++)
1278 rowMaster->cellFrac[i] = i/static_cast<real>(nc);
1281 rowMaster->bounds[i].cellFracLowerMax = i /static_cast<real>(nc);
1282 rowMaster->bounds[i].cellFracUpperMin = (i + 1)/static_cast<real>(nc);
1285 rowMaster->cellFrac[nc] = 1.0;
1290 //! Turns off dynamic load balancing (but leave it able to turn back on).
1291 static void turn_off_dlb(const gmx::MDLogger &mdlog,
1295 GMX_LOG(mdlog.info).appendText(
1296 "step " + gmx::toString(step) + " Turning off dynamic load balancing, because it is degrading performance.");
1297 dd->comm->dlbState = DlbState::offCanTurnOn;
1298 dd->comm->haveTurnedOffDlb = true;
1299 dd->comm->ddPartioningCountFirstDlbOff = dd->ddp_count;
1302 //! Turns off dynamic load balancing permanently.
1303 static void turn_off_dlb_forever(const gmx::MDLogger &mdlog,
1307 GMX_RELEASE_ASSERT(dd->comm->dlbState == DlbState::offCanTurnOn, "Can only turn off DLB forever when it was in the can-turn-on state");
1308 GMX_LOG(mdlog.info).appendText(
1309 "step " + gmx::toString(step) + " Will no longer try dynamic load balancing, as it degraded performance.");
1310 dd->comm->dlbState = DlbState::offForever;
1313 void set_dd_dlb_max_cutoff(t_commrec *cr, real cutoff)
1315 gmx_domdec_comm_t *comm;
1317 comm = cr->dd->comm;
1319 /* Turn on the DLB limiting (might have been on already) */
1320 comm->bPMELoadBalDLBLimits = TRUE;
1322 /* Change the cut-off limit */
1323 comm->PMELoadBal_max_cutoff = cutoff;
1327 fprintf(debug, "PME load balancing set a limit to the DLB staggering such that a %f cut-off will continue to fit\n",
1328 comm->PMELoadBal_max_cutoff);
1332 //! Merge atom buffers.
1333 static void merge_cg_buffers(int ncell,
1334 gmx_domdec_comm_dim_t *cd, int pulse,
1336 gmx::ArrayRef<int> index_gl,
1338 gmx::ArrayRef<gmx::RVec> x,
1339 gmx::ArrayRef<const gmx::RVec> recv_vr,
1340 cginfo_mb_t *cginfo_mb,
1341 gmx::ArrayRef<int> cginfo)
1343 gmx_domdec_ind_t *ind, *ind_p;
1344 int p, cell, c, cg, cg0, cg1, cg_gl;
1347 ind = &cd->ind[pulse];
1349 /* First correct the already stored data */
1350 shift = ind->nrecv[ncell];
1351 for (cell = ncell-1; cell >= 0; cell--)
1353 shift -= ind->nrecv[cell];
1356 /* Move the cg's present from previous grid pulses */
1357 cg0 = ncg_cell[ncell+cell];
1358 cg1 = ncg_cell[ncell+cell+1];
1359 for (cg = cg1-1; cg >= cg0; cg--)
1361 index_gl[cg + shift] = index_gl[cg];
1362 x[cg + shift] = x[cg];
1363 cginfo[cg + shift] = cginfo[cg];
1365 /* Correct the already stored send indices for the shift */
1366 for (p = 1; p <= pulse; p++)
1368 ind_p = &cd->ind[p];
1370 for (c = 0; c < cell; c++)
1372 cg0 += ind_p->nsend[c];
1374 cg1 = cg0 + ind_p->nsend[cell];
1375 for (cg = cg0; cg < cg1; cg++)
1377 ind_p->index[cg] += shift;
1383 /* Merge in the communicated buffers */
1386 for (cell = 0; cell < ncell; cell++)
1388 cg1 = ncg_cell[ncell+cell+1] + shift;
1389 for (cg = 0; cg < ind->nrecv[cell]; cg++)
1391 /* Copy this atom from the buffer */
1392 index_gl[cg1] = recv_i[cg0];
1393 x[cg1] = recv_vr[cg0];
1394 /* Copy information */
1395 cg_gl = index_gl[cg1];
1396 cginfo[cg1] = ddcginfo(cginfo_mb, cg_gl);
1400 shift += ind->nrecv[cell];
1401 ncg_cell[ncell+cell+1] = cg1;
1405 //! Makes a range partitioning for the atom groups wthin a cell
1406 static void make_cell2at_index(gmx_domdec_comm_dim_t *cd,
1410 /* Store the atom block boundaries for easy copying of communication buffers
1412 int g = atomGroupStart;
1413 for (int zone = 0; zone < nzone; zone++)
1415 for (gmx_domdec_ind_t &ind : cd->ind)
1417 ind.cell2at0[zone] = g;
1418 g += ind.nrecv[zone];
1419 ind.cell2at1[zone] = g;
1424 //! Returns whether a link is missing.
1425 static gmx_bool missing_link(const t_blocka &link,
1426 const int globalAtomIndex,
1427 const gmx_ga2la_t &ga2la)
1429 for (int i = link.index[globalAtomIndex]; i < link.index[globalAtomIndex + 1]; i++)
1431 if (!ga2la.findHome(link.a[i]))
1440 //! Domain corners for communication, a maximum of 4 i-zones see a j domain
1442 //! The corners for the non-bonded communication.
1444 //! Corner for rounding.
1446 //! Corners for rounding.
1448 //! Corners for bounded communication.
1450 //! Corner for rounding for bonded communication.
1454 //! Determine the corners of the domain(s) we are communicating with.
1456 set_dd_corners(const gmx_domdec_t *dd,
1457 int dim0, int dim1, int dim2,
1461 const gmx_domdec_comm_t *comm;
1462 const gmx_domdec_zones_t *zones;
1466 zones = &comm->zones;
1468 /* Keep the compiler happy */
1472 /* The first dimension is equal for all cells */
1473 c->c[0][0] = comm->cell_x0[dim0];
1476 c->bc[0] = c->c[0][0];
1481 /* This cell row is only seen from the first row */
1482 c->c[1][0] = comm->cell_x0[dim1];
1483 /* All rows can see this row */
1484 c->c[1][1] = comm->cell_x0[dim1];
1485 if (isDlbOn(dd->comm))
1487 c->c[1][1] = std::max(comm->cell_x0[dim1], comm->zone_d1[1].mch0);
1490 /* For the multi-body distance we need the maximum */
1491 c->bc[1] = std::max(comm->cell_x0[dim1], comm->zone_d1[1].p1_0);
1494 /* Set the upper-right corner for rounding */
1495 c->cr0 = comm->cell_x1[dim0];
1500 for (int j = 0; j < 4; j++)
1502 c->c[2][j] = comm->cell_x0[dim2];
1504 if (isDlbOn(dd->comm))
1506 /* Use the maximum of the i-cells that see a j-cell */
1507 for (const auto &iZone : zones->iZones)
1509 const int iZoneIndex = iZone.iZoneIndex;
1510 for (int jZone : iZone.jZoneRange)
1514 c->c[2][jZone - 4] =
1515 std::max(c->c[2][jZone - 4],
1516 comm->zone_d2[zones->shift[iZoneIndex][dim0]][zones->shift[iZoneIndex][dim1]].mch0);
1522 /* For the multi-body distance we need the maximum */
1523 c->bc[2] = comm->cell_x0[dim2];
1524 for (int i = 0; i < 2; i++)
1526 for (int j = 0; j < 2; j++)
1528 c->bc[2] = std::max(c->bc[2], comm->zone_d2[i][j].p1_0);
1534 /* Set the upper-right corner for rounding */
1535 /* Cell (0,0,0) and cell (1,0,0) can see cell 4 (0,1,1)
1536 * Only cell (0,0,0) can see cell 7 (1,1,1)
1538 c->cr1[0] = comm->cell_x1[dim1];
1539 c->cr1[3] = comm->cell_x1[dim1];
1540 if (isDlbOn(dd->comm))
1542 c->cr1[0] = std::max(comm->cell_x1[dim1], comm->zone_d1[1].mch1);
1545 /* For the multi-body distance we need the maximum */
1546 c->bcr1 = std::max(comm->cell_x1[dim1], comm->zone_d1[1].p1_1);
1553 /*! \brief Add the atom groups we need to send in this pulse from this
1554 * zone to \p localAtomGroups and \p work. */
1556 get_zone_pulse_cgs(gmx_domdec_t *dd,
1557 int zonei, int zone,
1559 gmx::ArrayRef<const int> globalAtomGroupIndices,
1560 int dim, int dim_ind,
1561 int dim0, int dim1, int dim2,
1562 real r_comm2, real r_bcomm2,
1564 bool distanceIsTriclinic,
1566 real skew_fac2_d, real skew_fac_01,
1567 rvec *v_d, rvec *v_0, rvec *v_1,
1568 const dd_corners_t *c,
1569 const rvec sf2_round,
1570 gmx_bool bDistBonded,
1575 gmx::ArrayRef<const int> cginfo,
1576 std::vector<int> *localAtomGroups,
1577 dd_comm_setup_work_t *work)
1579 gmx_domdec_comm_t *comm;
1581 gmx_bool bDistMB_pulse;
1583 real r2, rb2, r, tric_sh;
1590 bScrew = (dd->unitCellInfo.haveScrewPBC && dim == XX);
1592 bDistMB_pulse = (bDistMB && bDistBonded);
1594 /* Unpack the work data */
1595 std::vector<int> &ibuf = work->atomGroupBuffer;
1596 std::vector<gmx::RVec> &vbuf = work->positionBuffer;
1600 for (cg = cg0; cg < cg1; cg++)
1604 if (!distanceIsTriclinic)
1606 /* Rectangular direction, easy */
1607 r = cg_cm[cg][dim] - c->c[dim_ind][zone];
1614 r = cg_cm[cg][dim] - c->bc[dim_ind];
1620 /* Rounding gives at most a 16% reduction
1621 * in communicated atoms
1623 if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
1625 r = cg_cm[cg][dim0] - c->cr0;
1626 /* This is the first dimension, so always r >= 0 */
1633 if (dim_ind == 2 && (zonei == 2 || zonei == 3))
1635 r = cg_cm[cg][dim1] - c->cr1[zone];
1642 r = cg_cm[cg][dim1] - c->bcr1;
1652 /* Triclinic direction, more complicated */
1655 /* Rounding, conservative as the skew_fac multiplication
1656 * will slightly underestimate the distance.
1658 if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
1660 rn[dim0] = cg_cm[cg][dim0] - c->cr0;
1661 for (i = dim0+1; i < DIM; i++)
1663 rn[dim0] -= cg_cm[cg][i]*v_0[i][dim0];
1665 r2 = rn[dim0]*rn[dim0]*sf2_round[dim0];
1668 rb[dim0] = rn[dim0];
1671 /* Take care that the cell planes along dim0 might not
1672 * be orthogonal to those along dim1 and dim2.
1674 for (i = 1; i <= dim_ind; i++)
1677 if (normal[dim0][dimd] > 0)
1679 rn[dimd] -= rn[dim0]*normal[dim0][dimd];
1682 rb[dimd] -= rb[dim0]*normal[dim0][dimd];
1687 if (dim_ind == 2 && (zonei == 2 || zonei == 3))
1689 GMX_ASSERT(dim1 >= 0 && dim1 < DIM, "Must have a valid dimension index");
1690 rn[dim1] += cg_cm[cg][dim1] - c->cr1[zone];
1692 for (i = dim1+1; i < DIM; i++)
1694 tric_sh -= cg_cm[cg][i]*v_1[i][dim1];
1696 rn[dim1] += tric_sh;
1699 r2 += rn[dim1]*rn[dim1]*sf2_round[dim1];
1700 /* Take care of coupling of the distances
1701 * to the planes along dim0 and dim1 through dim2.
1703 r2 -= rn[dim0]*rn[dim1]*skew_fac_01;
1704 /* Take care that the cell planes along dim1
1705 * might not be orthogonal to that along dim2.
1707 if (normal[dim1][dim2] > 0)
1709 rn[dim2] -= rn[dim1]*normal[dim1][dim2];
1715 cg_cm[cg][dim1] - c->bcr1 + tric_sh;
1718 rb2 += rb[dim1]*rb[dim1]*sf2_round[dim1];
1719 /* Take care of coupling of the distances
1720 * to the planes along dim0 and dim1 through dim2.
1722 rb2 -= rb[dim0]*rb[dim1]*skew_fac_01;
1723 /* Take care that the cell planes along dim1
1724 * might not be orthogonal to that along dim2.
1726 if (normal[dim1][dim2] > 0)
1728 rb[dim2] -= rb[dim1]*normal[dim1][dim2];
1733 /* The distance along the communication direction */
1734 rn[dim] += cg_cm[cg][dim] - c->c[dim_ind][zone];
1736 for (i = dim+1; i < DIM; i++)
1738 tric_sh -= cg_cm[cg][i]*v_d[i][dim];
1743 r2 += rn[dim]*rn[dim]*skew_fac2_d;
1744 /* Take care of coupling of the distances
1745 * to the planes along dim0 and dim1 through dim2.
1747 if (dim_ind == 1 && zonei == 1)
1749 r2 -= rn[dim0]*rn[dim]*skew_fac_01;
1755 GMX_ASSERT(dim >= 0 && dim < DIM, "Must have a valid dimension index");
1756 rb[dim] += cg_cm[cg][dim] - c->bc[dim_ind] + tric_sh;
1759 rb2 += rb[dim]*rb[dim]*skew_fac2_d;
1760 /* Take care of coupling of the distances
1761 * to the planes along dim0 and dim1 through dim2.
1763 if (dim_ind == 1 && zonei == 1)
1765 rb2 -= rb[dim0]*rb[dim]*skew_fac_01;
1773 ((bDistMB && rb2 < r_bcomm2) ||
1774 (bDist2B && r2 < r_bcomm2)) &&
1776 (GET_CGINFO_BOND_INTER(cginfo[cg]) &&
1777 missing_link(*comm->bondedLinks, globalAtomGroupIndices[cg],
1780 /* Store the local and global atom group indices and position */
1781 localAtomGroups->push_back(cg);
1782 ibuf.push_back(globalAtomGroupIndices[cg]);
1786 if (dd->ci[dim] == 0)
1788 /* Correct cg_cm for pbc */
1789 rvec_add(cg_cm[cg], box[dim], posPbc);
1792 posPbc[YY] = box[YY][YY] - posPbc[YY];
1793 posPbc[ZZ] = box[ZZ][ZZ] - posPbc[ZZ];
1798 copy_rvec(cg_cm[cg], posPbc);
1800 vbuf.emplace_back(posPbc[XX], posPbc[YY], posPbc[ZZ]);
1807 work->nsend_zone = nsend_z;
1811 static void clearCommSetupData(dd_comm_setup_work_t *work)
1813 work->localAtomGroupBuffer.clear();
1814 work->atomGroupBuffer.clear();
1815 work->positionBuffer.clear();
1817 work->nsend_zone = 0;
1820 //! Prepare DD communication.
1821 static void setup_dd_communication(gmx_domdec_t *dd,
1822 matrix box, gmx_ddbox_t *ddbox,
1825 PaddedHostVector<gmx::RVec> *f)
1827 int dim_ind, dim, dim0, dim1, dim2, dimd, nat_tot;
1828 int nzone, nzone_send, zone, zonei, cg0, cg1;
1830 int *zone_cg_range, pos_cg;
1831 gmx_domdec_comm_t *comm;
1832 gmx_domdec_zones_t *zones;
1833 gmx_domdec_comm_dim_t *cd;
1834 cginfo_mb_t *cginfo_mb;
1835 gmx_bool bBondComm, bDist2B, bDistMB, bDistBonded;
1836 dd_corners_t corners;
1837 rvec *normal, *v_d, *v_0 = nullptr, *v_1 = nullptr;
1838 real skew_fac2_d, skew_fac_01;
1843 fprintf(debug, "Setting up DD communication\n");
1848 if (comm->dth.empty())
1850 /* Initialize the thread data.
1851 * This can not be done in init_domain_decomposition,
1852 * as the numbers of threads is determined later.
1854 int numThreads = gmx_omp_nthreads_get(emntDomdec);
1855 comm->dth.resize(numThreads);
1858 bBondComm = comm->systemInfo.filterBondedCommunication;
1860 /* Do we need to determine extra distances for multi-body bondeds? */
1861 bDistMB = (comm->systemInfo.haveInterDomainMultiBodyBondeds && isDlbOn(dd->comm) && dd->ndim > 1);
1863 /* Do we need to determine extra distances for only two-body bondeds? */
1864 bDist2B = (bBondComm && !bDistMB);
1866 const real r_comm2 = gmx::square(domainToDomainIntoAtomToDomainCutoff(comm->systemInfo, comm->systemInfo.cutoff));
1867 const real r_bcomm2 = gmx::square(domainToDomainIntoAtomToDomainCutoff(comm->systemInfo, comm->cutoff_mbody));
1871 fprintf(debug, "bBondComm %s, r_bc %f\n", gmx::boolToString(bBondComm), std::sqrt(r_bcomm2));
1874 zones = &comm->zones;
1877 dim1 = (dd->ndim >= 2 ? dd->dim[1] : -1);
1878 dim2 = (dd->ndim >= 3 ? dd->dim[2] : -1);
1880 set_dd_corners(dd, dim0, dim1, dim2, bDistMB, &corners);
1882 /* Triclinic stuff */
1883 normal = ddbox->normal;
1887 v_0 = ddbox->v[dim0];
1888 if (ddbox->tric_dir[dim0] && ddbox->tric_dir[dim1])
1890 /* Determine the coupling coefficient for the distances
1891 * to the cell planes along dim0 and dim1 through dim2.
1892 * This is required for correct rounding.
1895 ddbox->v[dim0][dim1+1][dim0]*ddbox->v[dim1][dim1+1][dim1];
1898 fprintf(debug, "\nskew_fac_01 %f\n", skew_fac_01);
1904 v_1 = ddbox->v[dim1];
1907 zone_cg_range = zones->cg_range;
1908 cginfo_mb = fr->cginfo_mb;
1910 zone_cg_range[0] = 0;
1911 zone_cg_range[1] = dd->ncg_home;
1912 comm->zone_ncg1[0] = dd->ncg_home;
1913 pos_cg = dd->ncg_home;
1915 nat_tot = comm->atomRanges.numHomeAtoms();
1917 for (dim_ind = 0; dim_ind < dd->ndim; dim_ind++)
1919 dim = dd->dim[dim_ind];
1920 cd = &comm->cd[dim_ind];
1922 /* Check if we need to compute triclinic distances along this dim */
1923 bool distanceIsTriclinic = false;
1924 for (int i = 0; i <= dim_ind; i++)
1926 if (ddbox->tric_dir[dd->dim[i]])
1928 distanceIsTriclinic = true;
1932 if (dim >= ddbox->npbcdim && dd->ci[dim] == 0)
1934 /* No pbc in this dimension, the first node should not comm. */
1942 v_d = ddbox->v[dim];
1943 skew_fac2_d = gmx::square(ddbox->skew_fac[dim]);
1945 cd->receiveInPlace = true;
1946 for (int p = 0; p < cd->numPulses(); p++)
1948 /* Only atoms communicated in the first pulse are used
1949 * for multi-body bonded interactions or for bBondComm.
1951 bDistBonded = ((bDistMB || bDist2B) && p == 0);
1953 gmx_domdec_ind_t *ind = &cd->ind[p];
1955 /* Thread 0 writes in the global index array */
1957 clearCommSetupData(&comm->dth[0]);
1959 for (zone = 0; zone < nzone_send; zone++)
1961 if (dim_ind > 0 && distanceIsTriclinic)
1963 /* Determine slightly more optimized skew_fac's
1965 * This reduces the number of communicated atoms
1966 * by about 10% for 3D DD of rhombic dodecahedra.
1968 for (dimd = 0; dimd < dim; dimd++)
1970 sf2_round[dimd] = 1;
1971 if (ddbox->tric_dir[dimd])
1973 for (int i = dd->dim[dimd] + 1; i < DIM; i++)
1975 /* If we are shifted in dimension i
1976 * and the cell plane is tilted forward
1977 * in dimension i, skip this coupling.
1979 if (!(zones->shift[nzone+zone][i] &&
1980 ddbox->v[dimd][i][dimd] >= 0))
1983 gmx::square(ddbox->v[dimd][i][dimd]);
1986 sf2_round[dimd] = 1/sf2_round[dimd];
1991 zonei = zone_perm[dim_ind][zone];
1994 /* Here we permutate the zones to obtain a convenient order
1995 * for neighbor searching
1997 cg0 = zone_cg_range[zonei];
1998 cg1 = zone_cg_range[zonei+1];
2002 /* Look only at the cg's received in the previous grid pulse
2004 cg1 = zone_cg_range[nzone+zone+1];
2005 cg0 = cg1 - cd->ind[p-1].nrecv[zone];
2008 const int numThreads = gmx::ssize(comm->dth);
2009 #pragma omp parallel for num_threads(numThreads) schedule(static)
2010 for (int th = 0; th < numThreads; th++)
2014 dd_comm_setup_work_t &work = comm->dth[th];
2016 /* Retain data accumulated into buffers of thread 0 */
2019 clearCommSetupData(&work);
2022 int cg0_th = cg0 + ((cg1 - cg0)* th )/numThreads;
2023 int cg1_th = cg0 + ((cg1 - cg0)*(th+1))/numThreads;
2025 /* Get the cg's for this pulse in this zone */
2026 get_zone_pulse_cgs(dd, zonei, zone, cg0_th, cg1_th,
2027 dd->globalAtomGroupIndices,
2028 dim, dim_ind, dim0, dim1, dim2,
2030 box, distanceIsTriclinic,
2031 normal, skew_fac2_d, skew_fac_01,
2032 v_d, v_0, v_1, &corners, sf2_round,
2033 bDistBonded, bBondComm,
2035 state->x.rvec_array(),
2037 th == 0 ? &ind->index : &work.localAtomGroupBuffer,
2040 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
2043 std::vector<int> &atomGroups = comm->dth[0].atomGroupBuffer;
2044 std::vector<gmx::RVec> &positions = comm->dth[0].positionBuffer;
2045 ind->nsend[zone] = comm->dth[0].nsend_zone;
2046 /* Append data of threads>=1 to the communication buffers */
2047 for (int th = 1; th < numThreads; th++)
2049 const dd_comm_setup_work_t &dth = comm->dth[th];
2051 ind->index.insert(ind->index.end(), dth.localAtomGroupBuffer.begin(), dth.localAtomGroupBuffer.end());
2052 atomGroups.insert(atomGroups.end(), dth.atomGroupBuffer.begin(), dth.atomGroupBuffer.end());
2053 positions.insert(positions.end(), dth.positionBuffer.begin(), dth.positionBuffer.end());
2054 comm->dth[0].nat += dth.nat;
2055 ind->nsend[zone] += dth.nsend_zone;
2058 /* Clear the counts in case we do not have pbc */
2059 for (zone = nzone_send; zone < nzone; zone++)
2061 ind->nsend[zone] = 0;
2063 ind->nsend[nzone] = ind->index.size();
2064 ind->nsend[nzone + 1] = comm->dth[0].nat;
2065 /* Communicate the number of cg's and atoms to receive */
2066 ddSendrecv(dd, dim_ind, dddirBackward,
2067 ind->nsend, nzone+2,
2068 ind->nrecv, nzone+2);
2072 /* We can receive in place if only the last zone is not empty */
2073 for (zone = 0; zone < nzone-1; zone++)
2075 if (ind->nrecv[zone] > 0)
2077 cd->receiveInPlace = false;
2082 int receiveBufferSize = 0;
2083 if (!cd->receiveInPlace)
2085 receiveBufferSize = ind->nrecv[nzone];
2087 /* These buffer are actually only needed with in-place */
2088 DDBufferAccess<int> globalAtomGroupBuffer(comm->intBuffer, receiveBufferSize);
2089 DDBufferAccess<gmx::RVec> rvecBuffer(comm->rvecBuffer, receiveBufferSize);
2091 dd_comm_setup_work_t &work = comm->dth[0];
2093 /* Make space for the global cg indices */
2094 int numAtomGroupsNew = pos_cg + ind->nrecv[nzone];
2095 dd->globalAtomGroupIndices.resize(numAtomGroupsNew);
2096 /* Communicate the global cg indices */
2097 gmx::ArrayRef<int> integerBufferRef;
2098 if (cd->receiveInPlace)
2100 integerBufferRef = gmx::arrayRefFromArray(dd->globalAtomGroupIndices.data() + pos_cg, ind->nrecv[nzone]);
2104 integerBufferRef = globalAtomGroupBuffer.buffer;
2106 ddSendrecv<int>(dd, dim_ind, dddirBackward,
2107 work.atomGroupBuffer, integerBufferRef);
2109 /* Make space for cg_cm */
2110 dd_check_alloc_ncg(fr, state, f, pos_cg + ind->nrecv[nzone]);
2112 /* Communicate the coordinates */
2113 gmx::ArrayRef<gmx::RVec> rvecBufferRef;
2114 if (cd->receiveInPlace)
2116 rvecBufferRef = gmx::makeArrayRef(state->x).subArray(pos_cg, ind->nrecv[nzone]);
2120 rvecBufferRef = rvecBuffer.buffer;
2122 ddSendrecv<gmx::RVec>(dd, dim_ind, dddirBackward,
2123 work.positionBuffer, rvecBufferRef);
2125 /* Make the charge group index */
2126 if (cd->receiveInPlace)
2128 zone = (p == 0 ? 0 : nzone - 1);
2129 while (zone < nzone)
2131 for (int i = 0; i < ind->nrecv[zone]; i++)
2133 int globalAtomIndex = dd->globalAtomGroupIndices[pos_cg];
2134 fr->cginfo[pos_cg] = ddcginfo(cginfo_mb, globalAtomIndex);
2139 comm->zone_ncg1[nzone+zone] = ind->nrecv[zone];
2142 zone_cg_range[nzone+zone] = pos_cg;
2147 /* This part of the code is never executed with bBondComm. */
2148 merge_cg_buffers(nzone, cd, p, zone_cg_range,
2149 dd->globalAtomGroupIndices, integerBufferRef.data(),
2150 state->x, rvecBufferRef,
2151 fr->cginfo_mb, fr->cginfo);
2152 pos_cg += ind->nrecv[nzone];
2154 nat_tot += ind->nrecv[nzone+1];
2156 if (!cd->receiveInPlace)
2158 /* Store the atom block for easy copying of communication buffers */
2159 make_cell2at_index(cd, nzone, zone_cg_range[nzone]);
2164 comm->atomRanges.setEnd(DDAtomRanges::Type::Zones, nat_tot);
2168 /* We don't need to update cginfo, since that was alrady done above.
2169 * So we pass NULL for the forcerec.
2171 dd_set_cginfo(dd->globalAtomGroupIndices,
2172 dd->ncg_home, dd->globalAtomGroupIndices.size(),
2178 fprintf(debug, "Finished setting up DD communication, zones:");
2179 for (c = 0; c < zones->n; c++)
2181 fprintf(debug, " %d", zones->cg_range[c+1]-zones->cg_range[c]);
2183 fprintf(debug, "\n");
2187 //! Set boundaries for the charge group range.
2188 static void set_cg_boundaries(gmx_domdec_zones_t *zones)
2190 for (auto &iZone : zones->iZones)
2192 iZone.iAtomRange = gmx::Range<int>(0, zones->cg_range[iZone.iZoneIndex + 1]);
2194 gmx::Range<int>(zones->cg_range[iZone.jZoneRange.begin()],
2195 zones->cg_range[iZone.jZoneRange.end()]);
2199 /*! \brief Set zone dimensions for zones \p zone_start to \p zone_end-1
2201 * Also sets the atom density for the home zone when \p zone_start=0.
2202 * For this \p numMovedChargeGroupsInHomeZone needs to be passed to tell
2203 * how many charge groups will move but are still part of the current range.
2204 * \todo When converting domdec to use proper classes, all these variables
2205 * should be private and a method should return the correct count
2206 * depending on an internal state.
2208 * \param[in,out] dd The domain decomposition struct
2209 * \param[in] box The box
2210 * \param[in] ddbox The domain decomposition box struct
2211 * \param[in] zone_start The start of the zone range to set sizes for
2212 * \param[in] zone_end The end of the zone range to set sizes for
2213 * \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
2215 static void set_zones_size(gmx_domdec_t *dd,
2216 matrix box, const gmx_ddbox_t *ddbox,
2217 int zone_start, int zone_end,
2218 int numMovedChargeGroupsInHomeZone)
2220 gmx_domdec_comm_t *comm;
2221 gmx_domdec_zones_t *zones;
2230 zones = &comm->zones;
2232 /* Do we need to determine extra distances for multi-body bondeds? */
2233 bDistMB = (comm->systemInfo.haveInterDomainMultiBodyBondeds &&
2234 isDlbOn(dd->comm) &&
2237 for (z = zone_start; z < zone_end; z++)
2239 /* Copy cell limits to zone limits.
2240 * Valid for non-DD dims and non-shifted dims.
2242 copy_rvec(comm->cell_x0, zones->size[z].x0);
2243 copy_rvec(comm->cell_x1, zones->size[z].x1);
2246 for (d = 0; d < dd->ndim; d++)
2250 for (z = 0; z < zones->n; z++)
2252 /* With a staggered grid we have different sizes
2253 * for non-shifted dimensions.
2255 if (isDlbOn(dd->comm) && zones->shift[z][dim] == 0)
2259 zones->size[z].x0[dim] = comm->zone_d1[zones->shift[z][dd->dim[d-1]]].min0;
2260 zones->size[z].x1[dim] = comm->zone_d1[zones->shift[z][dd->dim[d-1]]].max1;
2264 zones->size[z].x0[dim] = comm->zone_d2[zones->shift[z][dd->dim[d-2]]][zones->shift[z][dd->dim[d-1]]].min0;
2265 zones->size[z].x1[dim] = comm->zone_d2[zones->shift[z][dd->dim[d-2]]][zones->shift[z][dd->dim[d-1]]].max1;
2270 rcs = comm->systemInfo.cutoff;
2271 rcmbs = comm->cutoff_mbody;
2272 if (ddbox->tric_dir[dim])
2274 rcs /= ddbox->skew_fac[dim];
2275 rcmbs /= ddbox->skew_fac[dim];
2278 /* Set the lower limit for the shifted zone dimensions */
2279 for (z = zone_start; z < zone_end; z++)
2281 if (zones->shift[z][dim] > 0)
2284 if (!isDlbOn(dd->comm) || d == 0)
2286 zones->size[z].x0[dim] = comm->cell_x1[dim];
2287 zones->size[z].x1[dim] = comm->cell_x1[dim] + rcs;
2291 /* Here we take the lower limit of the zone from
2292 * the lowest domain of the zone below.
2296 zones->size[z].x0[dim] =
2297 comm->zone_d1[zones->shift[z][dd->dim[d-1]]].min1;
2303 zones->size[z].x0[dim] =
2304 zones->size[zone_perm[2][z-4]].x0[dim];
2308 zones->size[z].x0[dim] =
2309 comm->zone_d2[zones->shift[z][dd->dim[d-2]]][zones->shift[z][dd->dim[d-1]]].min1;
2312 /* A temporary limit, is updated below */
2313 zones->size[z].x1[dim] = zones->size[z].x0[dim];
2317 for (size_t zi = 0; zi < zones->iZones.size(); zi++)
2319 if (zones->shift[zi][dim] == 0)
2321 /* This takes the whole zone into account.
2322 * With multiple pulses this will lead
2323 * to a larger zone then strictly necessary.
2325 zones->size[z].x1[dim] = std::max(zones->size[z].x1[dim],
2326 zones->size[zi].x1[dim]+rcmbs);
2334 /* Loop over the i-zones to set the upper limit of each
2337 for (const auto &iZone : zones->iZones)
2339 const int zi = iZone.iZoneIndex;
2340 if (zones->shift[zi][dim] == 0)
2342 /* We should only use zones up to zone_end */
2343 const auto &jZoneRangeFull = iZone.jZoneRange;
2344 if (zone_end <= *jZoneRangeFull.begin())
2348 const gmx::Range<int> jZoneRange(*jZoneRangeFull.begin(),
2349 std::min(*jZoneRangeFull.end(), zone_end));
2350 for (int jZone : jZoneRange)
2352 if (zones->shift[jZone][dim] > 0)
2354 zones->size[jZone].x1[dim] = std::max(zones->size[jZone].x1[dim],
2355 zones->size[zi].x1[dim]+rcs);
2362 for (z = zone_start; z < zone_end; z++)
2364 /* Initialization only required to keep the compiler happy */
2365 rvec corner_min = {0, 0, 0}, corner_max = {0, 0, 0}, corner;
2368 /* To determine the bounding box for a zone we need to find
2369 * the extreme corners of 4, 2 or 1 corners.
2371 nc = 1 << (ddbox->nboundeddim - 1);
2373 for (c = 0; c < nc; c++)
2375 /* Set up a zone corner at x=0, ignoring trilinic couplings */
2379 corner[YY] = zones->size[z].x0[YY];
2383 corner[YY] = zones->size[z].x1[YY];
2387 corner[ZZ] = zones->size[z].x0[ZZ];
2391 corner[ZZ] = zones->size[z].x1[ZZ];
2393 if (dd->ndim == 1 && dd->dim[0] < ZZ && ZZ < dd->unitCellInfo.npbcdim &&
2394 box[ZZ][1 - dd->dim[0]] != 0)
2396 /* With 1D domain decomposition the cg's are not in
2397 * the triclinic box, but triclinic x-y and rectangular y/x-z.
2398 * Shift the corner of the z-vector back to along the box
2399 * vector of dimension d, so it will later end up at 0 along d.
2400 * This can affect the location of this corner along dd->dim[0]
2401 * through the matrix operation below if box[d][dd->dim[0]]!=0.
2403 int d = 1 - dd->dim[0];
2405 corner[d] -= corner[ZZ]*box[ZZ][d]/box[ZZ][ZZ];
2407 /* Apply the triclinic couplings */
2408 for (i = YY; i < ddbox->npbcdim && i < DIM; i++)
2410 for (j = XX; j < i; j++)
2412 corner[j] += corner[i]*box[i][j]/box[i][i];
2417 copy_rvec(corner, corner_min);
2418 copy_rvec(corner, corner_max);
2422 for (i = 0; i < DIM; i++)
2424 corner_min[i] = std::min(corner_min[i], corner[i]);
2425 corner_max[i] = std::max(corner_max[i], corner[i]);
2429 /* Copy the extreme cornes without offset along x */
2430 for (i = 0; i < DIM; i++)
2432 zones->size[z].bb_x0[i] = corner_min[i];
2433 zones->size[z].bb_x1[i] = corner_max[i];
2435 /* Add the offset along x */
2436 zones->size[z].bb_x0[XX] += zones->size[z].x0[XX];
2437 zones->size[z].bb_x1[XX] += zones->size[z].x1[XX];
2440 if (zone_start == 0)
2443 for (dim = 0; dim < DIM; dim++)
2445 vol *= zones->size[0].x1[dim] - zones->size[0].x0[dim];
2447 zones->dens_zone0 = (zones->cg_range[1] - zones->cg_range[0] - numMovedChargeGroupsInHomeZone)/vol;
2452 for (z = zone_start; z < zone_end; z++)
2454 fprintf(debug, "zone %d %6.3f - %6.3f %6.3f - %6.3f %6.3f - %6.3f\n",
2456 zones->size[z].x0[XX], zones->size[z].x1[XX],
2457 zones->size[z].x0[YY], zones->size[z].x1[YY],
2458 zones->size[z].x0[ZZ], zones->size[z].x1[ZZ]);
2459 fprintf(debug, "zone %d bb %6.3f - %6.3f %6.3f - %6.3f %6.3f - %6.3f\n",
2461 zones->size[z].bb_x0[XX], zones->size[z].bb_x1[XX],
2462 zones->size[z].bb_x0[YY], zones->size[z].bb_x1[YY],
2463 zones->size[z].bb_x0[ZZ], zones->size[z].bb_x1[ZZ]);
2468 /*! \brief Order data in \p dataToSort according to \p sort
2470 * Note: both buffers should have at least \p sort.size() elements.
2472 template <typename T>
2474 orderVector(gmx::ArrayRef<const gmx_cgsort_t> sort,
2475 gmx::ArrayRef<T> dataToSort,
2476 gmx::ArrayRef<T> sortBuffer)
2478 GMX_ASSERT(dataToSort.size() >= sort.size(), "The vector needs to be sufficiently large");
2479 GMX_ASSERT(sortBuffer.size() >= sort.size(), "The sorting buffer needs to be sufficiently large");
2481 /* Order the data into the temporary buffer */
2483 for (const gmx_cgsort_t &entry : sort)
2485 sortBuffer[i++] = dataToSort[entry.ind];
2488 /* Copy back to the original array */
2489 std::copy(sortBuffer.begin(), sortBuffer.begin() + sort.size(),
2490 dataToSort.begin());
2493 /*! \brief Order data in \p dataToSort according to \p sort
2495 * Note: \p vectorToSort should have at least \p sort.size() elements,
2496 * \p workVector is resized when it is too small.
2498 template <typename T>
2500 orderVector(gmx::ArrayRef<const gmx_cgsort_t> sort,
2501 gmx::ArrayRef<T> vectorToSort,
2502 std::vector<T> *workVector)
2504 if (gmx::index(workVector->size()) < sort.ssize())
2506 workVector->resize(sort.size());
2508 orderVector<T>(sort, vectorToSort, *workVector);
2511 //! Returns the sorting order for atoms based on the nbnxn grid order in sort
2512 static void dd_sort_order_nbnxn(const t_forcerec *fr,
2513 std::vector<gmx_cgsort_t> *sort)
2515 gmx::ArrayRef<const int> atomOrder = fr->nbv->getLocalAtomOrder();
2517 /* Using push_back() instead of this resize results in much slower code */
2518 sort->resize(atomOrder.size());
2519 gmx::ArrayRef<gmx_cgsort_t> buffer = *sort;
2520 size_t numSorted = 0;
2521 for (int i : atomOrder)
2525 /* The values of nsc and ind_gl are not used in this case */
2526 buffer[numSorted++].ind = i;
2529 sort->resize(numSorted);
2532 //! Returns the sorting state for DD.
2533 static void dd_sort_state(gmx_domdec_t *dd, t_forcerec *fr, t_state *state)
2535 gmx_domdec_sort_t *sort = dd->comm->sort.get();
2537 dd_sort_order_nbnxn(fr, &sort->sorted);
2539 /* We alloc with the old size, since cgindex is still old */
2540 DDBufferAccess<gmx::RVec> rvecBuffer(dd->comm->rvecBuffer, dd->ncg_home);
2542 /* Set the new home atom/charge group count */
2543 dd->ncg_home = sort->sorted.size();
2546 fprintf(debug, "Set the new home atom count to %d\n",
2550 /* Reorder the state */
2551 gmx::ArrayRef<const gmx_cgsort_t> cgsort = sort->sorted;
2552 GMX_RELEASE_ASSERT(cgsort.ssize() == dd->ncg_home, "We should sort all the home atom groups");
2554 if (state->flags & (1 << estX))
2556 orderVector(cgsort, makeArrayRef(state->x), rvecBuffer.buffer);
2558 if (state->flags & (1 << estV))
2560 orderVector(cgsort, makeArrayRef(state->v), rvecBuffer.buffer);
2562 if (state->flags & (1 << estCGP))
2564 orderVector(cgsort, makeArrayRef(state->cg_p), rvecBuffer.buffer);
2567 /* Reorder the global cg index */
2568 orderVector<int>(cgsort, dd->globalAtomGroupIndices, &sort->intBuffer);
2569 /* Reorder the cginfo */
2570 orderVector<int>(cgsort, fr->cginfo, &sort->intBuffer);
2571 /* Set the home atom number */
2572 dd->comm->atomRanges.setEnd(DDAtomRanges::Type::Home, dd->ncg_home);
2574 /* The atoms are now exactly in grid order, update the grid order */
2575 fr->nbv->setLocalAtomOrder();
2578 //! Accumulates load statistics.
2579 static void add_dd_statistics(gmx_domdec_t *dd)
2581 gmx_domdec_comm_t *comm = dd->comm;
2583 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
2585 auto range = static_cast<DDAtomRanges::Type>(i);
2587 comm->atomRanges.end(range) - comm->atomRanges.start(range);
2592 void reset_dd_statistics_counters(gmx_domdec_t *dd)
2594 gmx_domdec_comm_t *comm = dd->comm;
2596 /* Reset all the statistics and counters for total run counting */
2597 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
2599 comm->sum_nat[i] = 0;
2603 comm->load_step = 0;
2606 clear_ivec(comm->load_lim);
2611 void print_dd_statistics(const t_commrec *cr, const t_inputrec *ir, FILE *fplog)
2613 gmx_domdec_comm_t *comm = cr->dd->comm;
2615 const int numRanges = static_cast<int>(DDAtomRanges::Type::Number);
2616 gmx_sumd(numRanges, comm->sum_nat, cr);
2618 if (fplog == nullptr)
2623 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");
2625 for (int i = static_cast<int>(DDAtomRanges::Type::Zones); i < numRanges; i++)
2627 auto range = static_cast<DDAtomRanges::Type>(i);
2628 double av = comm->sum_nat[i]/comm->ndecomp;
2631 case DDAtomRanges::Type::Zones:
2633 " av. #atoms communicated per step for force: %d x %.1f\n",
2636 case DDAtomRanges::Type::Vsites:
2637 if (cr->dd->vsite_comm)
2640 " av. #atoms communicated per step for vsites: %d x %.1f\n",
2641 (EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD) ? 3 : 2,
2645 case DDAtomRanges::Type::Constraints:
2646 if (cr->dd->constraint_comm)
2649 " av. #atoms communicated per step for LINCS: %d x %.1f\n",
2650 1 + ir->nLincsIter, av);
2654 gmx_incons(" Unknown type for DD statistics");
2657 fprintf(fplog, "\n");
2659 if (comm->ddSettings.recordLoad && EI_DYNAMICS(ir->eI))
2661 print_dd_load_av(fplog, cr->dd);
2665 //!\brief TODO Remove fplog when group scheme and charge groups are gone
2666 void dd_partition_system(FILE *fplog,
2667 const gmx::MDLogger &mdlog,
2669 const t_commrec *cr,
2670 gmx_bool bMasterState,
2672 t_state *state_global,
2673 const gmx_mtop_t &top_global,
2674 const t_inputrec *ir,
2675 gmx::ImdSession *imdSession,
2677 t_state *state_local,
2678 PaddedHostVector<gmx::RVec> *f,
2679 gmx::MDAtoms *mdAtoms,
2680 gmx_localtop_t *top_local,
2683 gmx::Constraints *constr,
2685 gmx_wallcycle *wcycle,
2689 gmx_domdec_comm_t *comm;
2690 gmx_ddbox_t ddbox = {0};
2691 int64_t step_pcoupl;
2692 rvec cell_ns_x0, cell_ns_x1;
2693 int ncgindex_set, ncg_moved, nat_f_novirsum;
2694 gmx_bool bBoxChanged, bNStGlobalComm, bDoDLB, bCheckWhetherToTurnDlbOn, bLogLoad;
2700 wallcycle_start(wcycle, ewcDOMDEC);
2705 // TODO if the update code becomes accessible here, use
2706 // upd->deform for this logic.
2707 bBoxChanged = (bMasterState || inputrecDeform(ir));
2708 if (ir->epc != epcNO)
2710 /* With nstpcouple > 1 pressure coupling happens.
2711 * one step after calculating the pressure.
2712 * Box scaling happens at the end of the MD step,
2713 * after the DD partitioning.
2714 * We therefore have to do DLB in the first partitioning
2715 * after an MD step where P-coupling occurred.
2716 * We need to determine the last step in which p-coupling occurred.
2717 * MRS -- need to validate this for vv?
2719 int n = ir->nstpcouple;
2722 step_pcoupl = step - 1;
2726 step_pcoupl = ((step - 1)/n)*n + 1;
2728 if (step_pcoupl >= comm->partition_step)
2734 bNStGlobalComm = (step % nstglobalcomm == 0);
2742 /* Should we do dynamic load balacing this step?
2743 * Since it requires (possibly expensive) global communication,
2744 * we might want to do DLB less frequently.
2746 if (bBoxChanged || ir->epc != epcNO)
2748 bDoDLB = bBoxChanged;
2752 bDoDLB = bNStGlobalComm;
2756 /* Check if we have recorded loads on the nodes */
2757 if (comm->ddSettings.recordLoad && dd_load_count(comm) > 0)
2759 bCheckWhetherToTurnDlbOn = dd_dlb_get_should_check_whether_to_turn_dlb_on(dd);
2761 /* Print load every nstlog, first and last step to the log file */
2762 bLogLoad = ((ir->nstlog > 0 && step % ir->nstlog == 0) ||
2763 comm->n_load_collect == 0 ||
2765 (step + ir->nstlist > ir->init_step + ir->nsteps)));
2767 /* Avoid extra communication due to verbose screen output
2768 * when nstglobalcomm is set.
2770 if (bDoDLB || bLogLoad || bCheckWhetherToTurnDlbOn ||
2771 (bVerbose && (ir->nstlist == 0 || nstglobalcomm <= ir->nstlist)))
2773 get_load_distribution(dd, wcycle);
2778 GMX_LOG(mdlog.info).asParagraph().appendText(dd_print_load(dd, step-1));
2782 dd_print_load_verbose(dd);
2785 comm->n_load_collect++;
2791 /* Add the measured cycles to the running average */
2792 const float averageFactor = 0.1F;
2793 comm->cyclesPerStepDlbExpAverage =
2794 (1 - averageFactor)*comm->cyclesPerStepDlbExpAverage +
2795 averageFactor*comm->cycl[ddCyclStep]/comm->cycl_n[ddCyclStep];
2797 if (comm->dlbState == DlbState::onCanTurnOff &&
2798 dd->comm->n_load_have % c_checkTurnDlbOffInterval == c_checkTurnDlbOffInterval - 1)
2800 gmx_bool turnOffDlb;
2803 /* If the running averaged cycles with DLB are more
2804 * than before we turned on DLB, turn off DLB.
2805 * We will again run and check the cycles without DLB
2806 * and we can then decide if to turn off DLB forever.
2808 turnOffDlb = (comm->cyclesPerStepDlbExpAverage >
2809 comm->cyclesPerStepBeforeDLB);
2811 dd_bcast(dd, sizeof(turnOffDlb), &turnOffDlb);
2814 /* To turn off DLB, we need to redistribute the atoms */
2815 dd_collect_state(dd, state_local, state_global);
2816 bMasterState = TRUE;
2817 turn_off_dlb(mdlog, dd, step);
2821 else if (bCheckWhetherToTurnDlbOn)
2823 gmx_bool turnOffDlbForever = FALSE;
2824 gmx_bool turnOnDlb = FALSE;
2826 /* Since the timings are node dependent, the master decides */
2829 /* If we recently turned off DLB, we want to check if
2830 * performance is better without DLB. We want to do this
2831 * ASAP to minimize the chance that external factors
2832 * slowed down the DLB step are gone here and we
2833 * incorrectly conclude that DLB was causing the slowdown.
2834 * So we measure one nstlist block, no running average.
2836 if (comm->haveTurnedOffDlb &&
2837 comm->cycl[ddCyclStep]/comm->cycl_n[ddCyclStep] <
2838 comm->cyclesPerStepDlbExpAverage)
2840 /* After turning off DLB we ran nstlist steps in fewer
2841 * cycles than with DLB. This likely means that DLB
2842 * in not benefical, but this could be due to a one
2843 * time unlucky fluctuation, so we require two such
2844 * observations in close succession to turn off DLB
2847 if (comm->dlbSlowerPartitioningCount > 0 &&
2848 dd->ddp_count < comm->dlbSlowerPartitioningCount + 10*c_checkTurnDlbOnInterval)
2850 turnOffDlbForever = TRUE;
2852 comm->haveTurnedOffDlb = false;
2853 /* Register when we last measured DLB slowdown */
2854 comm->dlbSlowerPartitioningCount = dd->ddp_count;
2858 /* Here we check if the max PME rank load is more than 0.98
2859 * the max PP force load. If so, PP DLB will not help,
2860 * since we are (almost) limited by PME. Furthermore,
2861 * DLB will cause a significant extra x/f redistribution
2862 * cost on the PME ranks, which will then surely result
2863 * in lower total performance.
2865 if (comm->ddRankSetup.usePmeOnlyRanks &&
2866 dd_pme_f_ratio(dd) > 1 - DD_PERF_LOSS_DLB_ON)
2872 turnOnDlb = (dd_force_imb_perf_loss(dd) >= DD_PERF_LOSS_DLB_ON);
2878 gmx_bool turnOffDlbForever;
2882 turnOffDlbForever, turnOnDlb
2884 dd_bcast(dd, sizeof(bools), &bools);
2885 if (bools.turnOffDlbForever)
2887 turn_off_dlb_forever(mdlog, dd, step);
2889 else if (bools.turnOnDlb)
2891 turn_on_dlb(mdlog, dd, step);
2896 comm->n_load_have++;
2902 /* Clear the old state */
2903 clearDDStateIndices(dd, false);
2906 auto xGlobal = positionsFromStatePointer(state_global);
2908 set_ddbox(*dd, true,
2909 DDMASTER(dd) ? state_global->box : nullptr,
2913 distributeState(mdlog, dd, top_global, state_global, ddbox, state_local, f);
2915 /* Ensure that we have space for the new distribution */
2916 dd_check_alloc_ncg(fr, state_local, f, dd->ncg_home);
2918 inc_nrnb(nrnb, eNR_CGCM, comm->atomRanges.numHomeAtoms());
2920 dd_set_cginfo(dd->globalAtomGroupIndices, 0, dd->ncg_home, fr);
2922 else if (state_local->ddp_count != dd->ddp_count)
2924 if (state_local->ddp_count > dd->ddp_count)
2926 gmx_fatal(FARGS, "Internal inconsistency state_local->ddp_count (%d) > dd->ddp_count (%" PRId64 ")", state_local->ddp_count, dd->ddp_count);
2929 if (state_local->ddp_count_cg_gl != state_local->ddp_count)
2931 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);
2934 /* Clear the old state */
2935 clearDDStateIndices(dd, false);
2937 /* Restore the atom group indices from state_local */
2938 restoreAtomGroups(dd, state_local);
2939 make_dd_indices(dd, 0);
2940 ncgindex_set = dd->ncg_home;
2942 inc_nrnb(nrnb, eNR_CGCM, comm->atomRanges.numHomeAtoms());
2944 dd_set_cginfo(dd->globalAtomGroupIndices, 0, dd->ncg_home, fr);
2946 set_ddbox(*dd, bMasterState, state_local->box,
2947 true, state_local->x, &ddbox);
2949 bRedist = isDlbOn(comm);
2953 /* We have the full state, only redistribute the cgs */
2955 /* Clear the non-home indices */
2956 clearDDStateIndices(dd, true);
2959 /* To avoid global communication, we do not recompute the extent
2960 * of the system for dims without pbc. Therefore we need to copy
2961 * the previously computed values when we do not communicate.
2963 if (!bNStGlobalComm)
2965 copy_rvec(comm->box0, ddbox.box0 );
2966 copy_rvec(comm->box_size, ddbox.box_size);
2968 set_ddbox(*dd, bMasterState, state_local->box,
2969 bNStGlobalComm, state_local->x, &ddbox);
2974 /* Copy needed for dim's without pbc when avoiding communication */
2975 copy_rvec(ddbox.box0, comm->box0 );
2976 copy_rvec(ddbox.box_size, comm->box_size);
2978 set_dd_cell_sizes(dd, &ddbox, dd->unitCellInfo.ddBoxIsDynamic, bMasterState, bDoDLB,
2981 if (comm->ddSettings.nstDDDumpGrid > 0 && step % comm->ddSettings.nstDDDumpGrid == 0)
2983 write_dd_grid_pdb("dd_grid", step, dd, state_local->box, &ddbox);
2986 if (comm->systemInfo.useUpdateGroups)
2988 comm->updateGroupsCog->addCogs(gmx::arrayRefFromArray(dd->globalAtomGroupIndices.data(), dd->ncg_home),
2992 /* Check if we should sort the charge groups */
2993 const bool bSortCG = (bMasterState || bRedist);
2995 /* When repartitioning we mark atom groups that will move to neighboring
2996 * DD cells, but we do not move them right away for performance reasons.
2997 * Thus we need to keep track of how many charge groups will move for
2998 * obtaining correct local charge group / atom counts.
3003 wallcycle_sub_start(wcycle, ewcsDD_REDIST);
3005 ncgindex_set = dd->ncg_home;
3006 dd_redistribute_cg(fplog, step, dd, ddbox.tric_dir,
3010 GMX_RELEASE_ASSERT(bSortCG, "Sorting is required after redistribution");
3012 if (comm->systemInfo.useUpdateGroups)
3014 comm->updateGroupsCog->addCogs(gmx::arrayRefFromArray(dd->globalAtomGroupIndices.data(), dd->ncg_home),
3018 wallcycle_sub_stop(wcycle, ewcsDD_REDIST);
3021 // TODO: Integrate this code in the nbnxm module
3022 get_nsgrid_boundaries(ddbox.nboundeddim, state_local->box,
3024 &comm->cell_x0, &comm->cell_x1,
3025 dd->ncg_home, as_rvec_array(state_local->x.data()),
3026 cell_ns_x0, cell_ns_x1, &grid_density);
3030 comm_dd_ns_cell_sizes(dd, &ddbox, cell_ns_x0, cell_ns_x1, step);
3035 wallcycle_sub_start(wcycle, ewcsDD_GRID);
3037 /* Sort the state on charge group position.
3038 * This enables exact restarts from this step.
3039 * It also improves performance by about 15% with larger numbers
3040 * of atoms per node.
3043 /* Fill the ns grid with the home cell,
3044 * so we can sort with the indices.
3046 set_zones_ncg_home(dd);
3048 set_zones_size(dd, state_local->box, &ddbox, 0, 1, ncg_moved);
3050 nbnxn_put_on_grid(fr->nbv.get(), state_local->box,
3052 comm->zones.size[0].bb_x0,
3053 comm->zones.size[0].bb_x1,
3054 comm->updateGroupsCog.get(),
3055 { 0, dd->ncg_home },
3056 comm->zones.dens_zone0,
3059 ncg_moved, bRedist ? comm->movedBuffer.data() : nullptr);
3063 fprintf(debug, "Step %s, sorting the %d home charge groups\n",
3064 gmx_step_str(step, sbuf), dd->ncg_home);
3066 dd_sort_state(dd, fr, state_local);
3068 /* After sorting and compacting we set the correct size */
3069 dd_resize_state(state_local, f, comm->atomRanges.numHomeAtoms());
3071 /* Rebuild all the indices */
3075 wallcycle_sub_stop(wcycle, ewcsDD_GRID);
3079 /* With the group scheme the sorting array is part of the DD state,
3080 * but it just got out of sync, so mark as invalid by emptying it.
3082 if (ir->cutoff_scheme == ecutsGROUP)
3084 comm->sort->sorted.clear();
3088 if (comm->systemInfo.useUpdateGroups)
3090 /* The update groups cog's are invalid after sorting
3091 * and need to be cleared before the next partitioning anyhow.
3093 comm->updateGroupsCog->clear();
3096 wallcycle_sub_start(wcycle, ewcsDD_SETUPCOMM);
3098 /* Setup up the communication and communicate the coordinates */
3099 setup_dd_communication(dd, state_local->box, &ddbox, fr, state_local, f);
3101 /* Set the indices */
3102 make_dd_indices(dd, ncgindex_set);
3104 /* Set the charge group boundaries for neighbor searching */
3105 set_cg_boundaries(&comm->zones);
3107 if (fr->cutoff_scheme == ecutsVERLET)
3109 /* When bSortCG=true, we have already set the size for zone 0 */
3110 set_zones_size(dd, state_local->box, &ddbox,
3111 bSortCG ? 1 : 0, comm->zones.n,
3115 wallcycle_sub_stop(wcycle, ewcsDD_SETUPCOMM);
3118 write_dd_pdb("dd_home",step,"dump",top_global,cr,
3119 -1,state_local->x.rvec_array(),state_local->box);
3122 wallcycle_sub_start(wcycle, ewcsDD_MAKETOP);
3124 /* Extract a local topology from the global topology */
3125 for (int i = 0; i < dd->ndim; i++)
3127 np[dd->dim[i]] = comm->cd[i].numPulses();
3129 dd_make_local_top(dd, &comm->zones, dd->unitCellInfo.npbcdim, state_local->box,
3130 comm->cellsize_min, np,
3132 state_local->x.rvec_array(),
3133 top_global, top_local);
3135 wallcycle_sub_stop(wcycle, ewcsDD_MAKETOP);
3137 wallcycle_sub_start(wcycle, ewcsDD_MAKECONSTR);
3139 /* Set up the special atom communication */
3140 int n = comm->atomRanges.end(DDAtomRanges::Type::Zones);
3141 for (int i = static_cast<int>(DDAtomRanges::Type::Zones) + 1; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
3143 auto range = static_cast<DDAtomRanges::Type>(i);
3146 case DDAtomRanges::Type::Vsites:
3147 if (vsite && vsite->numInterUpdategroupVsites)
3149 n = dd_make_local_vsites(dd, n, top_local->idef.il);
3152 case DDAtomRanges::Type::Constraints:
3153 if (dd->comm->systemInfo.haveSplitConstraints || dd->comm->systemInfo.haveSplitSettles)
3155 /* Only for inter-cg constraints we need special code */
3156 n = dd_make_local_constraints(dd, n, &top_global, fr->cginfo.data(),
3157 constr, ir->nProjOrder,
3158 top_local->idef.il);
3162 gmx_incons("Unknown special atom type setup");
3164 comm->atomRanges.setEnd(range, n);
3167 wallcycle_sub_stop(wcycle, ewcsDD_MAKECONSTR);
3169 wallcycle_sub_start(wcycle, ewcsDD_TOPOTHER);
3171 /* Make space for the extra coordinates for virtual site
3172 * or constraint communication.
3174 state_local->natoms = comm->atomRanges.numAtomsTotal();
3176 dd_resize_state(state_local, f, state_local->natoms);
3178 if (fr->haveDirectVirialContributions)
3180 if (vsite && vsite->numInterUpdategroupVsites)
3182 nat_f_novirsum = comm->atomRanges.end(DDAtomRanges::Type::Vsites);
3186 if (EEL_FULL(ir->coulombtype) && dd->haveExclusions)
3188 nat_f_novirsum = comm->atomRanges.end(DDAtomRanges::Type::Zones);
3192 nat_f_novirsum = comm->atomRanges.numHomeAtoms();
3201 /* Set the number of atoms required for the force calculation.
3202 * Forces need to be constrained when doing energy
3203 * minimization. For simple simulations we could avoid some
3204 * allocation, zeroing and copying, but this is probably not worth
3205 * the complications and checking.
3207 forcerec_set_ranges(fr,
3208 comm->atomRanges.end(DDAtomRanges::Type::Zones),
3209 comm->atomRanges.end(DDAtomRanges::Type::Constraints),
3212 /* Update atom data for mdatoms and several algorithms */
3213 mdAlgorithmsSetupAtomData(cr, ir, top_global, top_local, fr,
3214 nullptr, mdAtoms, constr, vsite, nullptr);
3216 auto mdatoms = mdAtoms->mdatoms();
3217 if (!thisRankHasDuty(cr, DUTY_PME))
3219 /* Send the charges and/or c6/sigmas to our PME only node */
3220 gmx_pme_send_parameters(cr,
3222 mdatoms->nChargePerturbed != 0, mdatoms->nTypePerturbed != 0,
3223 mdatoms->chargeA, mdatoms->chargeB,
3224 mdatoms->sqrt_c6A, mdatoms->sqrt_c6B,
3225 mdatoms->sigmaA, mdatoms->sigmaB,
3226 dd_pme_maxshift_x(dd), dd_pme_maxshift_y(dd));
3231 /* Update the local pull groups */
3232 dd_make_local_pull_groups(cr, pull_work);
3235 if (dd->atomSets != nullptr)
3237 /* Update the local atom sets */
3238 dd->atomSets->setIndicesInDomainDecomposition(*(dd->ga2la));
3241 /* Update the local atoms to be communicated via the IMD protocol if bIMD is TRUE. */
3242 imdSession->dd_make_local_IMD_atoms(dd);
3244 add_dd_statistics(dd);
3246 /* Make sure we only count the cycles for this DD partitioning */
3247 clear_dd_cycle_counts(dd);
3249 /* Because the order of the atoms might have changed since
3250 * the last vsite construction, we need to communicate the constructing
3251 * atom coordinates again (for spreading the forces this MD step).
3253 dd_move_x_vsites(dd, state_local->box, state_local->x.rvec_array());
3255 wallcycle_sub_stop(wcycle, ewcsDD_TOPOTHER);
3257 if (comm->ddSettings.nstDDDump > 0 && step % comm->ddSettings.nstDDDump == 0)
3259 dd_move_x(dd, state_local->box, state_local->x, nullWallcycle);
3260 write_dd_pdb("dd_dump", step, "dump", &top_global, cr,
3261 -1, state_local->x.rvec_array(), state_local->box);
3264 /* Store the partitioning step */
3265 comm->partition_step = step;
3267 /* Increase the DD partitioning counter */
3269 /* The state currently matches this DD partitioning count, store it */
3270 state_local->ddp_count = dd->ddp_count;
3273 /* The DD master node knows the complete cg distribution,
3274 * store the count so we can possibly skip the cg info communication.
3276 comm->master_cg_ddp_count = (bSortCG ? 0 : dd->ddp_count);
3279 if (comm->ddSettings.DD_debug > 0)
3281 /* Set the env var GMX_DD_DEBUG if you suspect corrupted indices */
3282 check_index_consistency(dd, top_global.natoms, "after partitioning");
3285 wallcycle_stop(wcycle, ewcDOMDEC);
3288 /*! \brief Check whether bonded interactions are missing, if appropriate */
3289 void checkNumberOfBondedInteractions(const gmx::MDLogger &mdlog,
3291 int totalNumberOfBondedInteractions,
3292 const gmx_mtop_t *top_global,
3293 const gmx_localtop_t *top_local,
3296 bool *shouldCheckNumberOfBondedInteractions)
3298 if (*shouldCheckNumberOfBondedInteractions)
3300 if (totalNumberOfBondedInteractions != cr->dd->nbonded_global)
3302 dd_print_missing_interactions(mdlog, cr, totalNumberOfBondedInteractions, top_global, top_local, x, box); // Does not return
3304 *shouldCheckNumberOfBondedInteractions = false;