2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018,2019,2020, 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_pp.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/mdatom.h"
79 #include "gromacs/mdtypes/nblist.h"
80 #include "gromacs/mdtypes/state.h"
81 #include "gromacs/nbnxm/nbnxm.h"
82 #include "gromacs/pulling/pull.h"
83 #include "gromacs/timing/wallcycle.h"
84 #include "gromacs/topology/mtop_util.h"
85 #include "gromacs/topology/topology.h"
86 #include "gromacs/utility/cstringutil.h"
87 #include "gromacs/utility/fatalerror.h"
88 #include "gromacs/utility/logger.h"
89 #include "gromacs/utility/real.h"
90 #include "gromacs/utility/smalloc.h"
91 #include "gromacs/utility/strconvert.h"
92 #include "gromacs/utility/stringstream.h"
93 #include "gromacs/utility/stringutil.h"
94 #include "gromacs/utility/textwriter.h"
97 #include "cellsizes.h"
98 #include "distribute.h"
99 #include "domdec_constraints.h"
100 #include "domdec_internal.h"
101 #include "domdec_vsite.h"
103 #include "redistribute.h"
106 /*! \brief Turn on DLB when the load imbalance causes this amount of total loss.
108 * There is a bit of overhead with DLB and it's difficult to achieve
109 * a load imbalance of less than 2% with DLB.
111 #define DD_PERF_LOSS_DLB_ON 0.02
113 //! Warn about imbalance due to PP or PP/PME load imbalance at this loss.
114 #define DD_PERF_LOSS_WARN 0.05
117 //! Debug helper printing a DD zone
118 static void print_ddzone(FILE* fp, int d, int i, int j, gmx_ddzone_t* zone)
121 "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 "
123 d, i, j, zone->min0, zone->max1, zone->mch0, zone->mch0, 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, const gmx_ddbox_t* ddbox, rvec cell_ns_x0, rvec cell_ns_x1)
133 constexpr int c_ddZoneCommMaxNumZones = 5;
134 gmx_ddzone_t buf_s[c_ddZoneCommMaxNumZones];
135 gmx_ddzone_t buf_r[c_ddZoneCommMaxNumZones];
136 gmx_ddzone_t buf_e[c_ddZoneCommMaxNumZones];
137 gmx_domdec_comm_t* comm = dd->comm;
141 for (int d = 1; d < dd->ndim; d++)
143 int dim = dd->dim[d];
144 gmx_ddzone_t& zp = (d == 1) ? comm->zone_d1[0] : comm->zone_d2[0][0];
146 /* Copy the base sizes of the home zone */
147 zp.min0 = cell_ns_x0[dim];
148 zp.max1 = cell_ns_x1[dim];
149 zp.min1 = cell_ns_x1[dim];
150 zp.mch0 = cell_ns_x0[dim];
151 zp.mch1 = cell_ns_x1[dim];
152 zp.p1_0 = cell_ns_x0[dim];
153 zp.p1_1 = cell_ns_x1[dim];
157 gmx::ArrayRef<DDCellsizesWithDlb> cellsizes = comm->cellsizesWithDlb;
159 /* Loop backward over the dimensions and aggregate the extremes
162 for (int d = dd->ndim - 2; d >= 0; d--)
164 const int dim = dd->dim[d];
165 const bool applyPbc = (dim < ddbox->npbcdim);
167 /* Use an rvec to store two reals */
168 extr_s[d][0] = cellsizes[d + 1].fracLower;
169 extr_s[d][1] = cellsizes[d + 1].fracUpper;
170 extr_s[d][2] = cellsizes[d + 1].fracUpper;
173 GMX_ASSERT(pos < c_ddZoneCommMaxNumZones, "The buffers should be sufficiently large");
174 /* Store the extremes in the backward sending buffer,
175 * so they get updated separately from the forward communication.
177 for (int d1 = d; d1 < dd->ndim - 1; d1++)
179 gmx_ddzone_t& buf = buf_s[pos];
181 /* We invert the order to be able to use the same loop for buf_e */
182 buf.min0 = extr_s[d1][1];
183 buf.max1 = extr_s[d1][0];
184 buf.min1 = extr_s[d1][2];
187 /* Store the cell corner of the dimension we communicate along */
188 buf.p1_0 = comm->cell_x0[dim];
194 buf_s[pos] = (dd->ndim == 2) ? comm->zone_d1[0] : comm->zone_d2[0][0];
197 if (dd->ndim == 3 && d == 0)
199 buf_s[pos] = comm->zone_d2[0][1];
201 buf_s[pos] = comm->zone_d1[0];
205 /* We only need to communicate the extremes
206 * in the forward direction
208 int numPulses = comm->cd[d].numPulses();
212 /* Take the minimum to avoid double communication */
213 numPulsesMin = std::min(numPulses, dd->numCells[dim] - 1 - numPulses);
217 /* Without PBC we should really not communicate over
218 * the boundaries, but implementing that complicates
219 * the communication setup and therefore we simply
220 * do all communication, but ignore some data.
222 numPulsesMin = numPulses;
224 for (int pulse = 0; pulse < numPulsesMin; pulse++)
226 /* Communicate the extremes forward */
227 bool receiveValidData = (applyPbc || dd->ci[dim] > 0);
229 int numElements = dd->ndim - d - 1;
230 ddSendrecv(dd, d, dddirForward, extr_s + d, numElements, extr_r + d, numElements);
232 if (receiveValidData)
234 for (int d1 = d; d1 < dd->ndim - 1; d1++)
236 extr_s[d1][0] = std::max(extr_s[d1][0], extr_r[d1][0]);
237 extr_s[d1][1] = std::min(extr_s[d1][1], extr_r[d1][1]);
238 extr_s[d1][2] = std::min(extr_s[d1][2], extr_r[d1][2]);
243 const int numElementsInBuffer = pos;
244 for (int pulse = 0; pulse < numPulses; pulse++)
246 /* Communicate all the zone information backward */
247 bool receiveValidData = (applyPbc || dd->ci[dim] < dd->numCells[dim] - 1);
250 sizeof(gmx_ddzone_t) == c_ddzoneNumReals * sizeof(real),
251 "Here we expect gmx_ddzone_t to consist of c_ddzoneNumReals reals (only)");
253 int numReals = numElementsInBuffer * c_ddzoneNumReals;
254 ddSendrecv(dd, d, dddirBackward, gmx::arrayRefFromArray(&buf_s[0].min0, numReals),
255 gmx::arrayRefFromArray(&buf_r[0].min0, numReals));
260 for (int d1 = d + 1; d1 < dd->ndim; d1++)
262 /* Determine the decrease of maximum required
263 * communication height along d1 due to the distance along d,
264 * this avoids a lot of useless atom communication.
266 real dist_d = comm->cell_x1[dim] - buf_r[0].p1_0;
269 if (ddbox->tric_dir[dim])
271 /* c is the off-diagonal coupling between the cell planes
272 * along directions d and d1.
274 c = ddbox->v[dim][dd->dim[d1]][dim];
280 real det = (1 + c * c) * gmx::square(comm->systemInfo.cutoff) - dist_d * dist_d;
283 dh[d1] = comm->systemInfo.cutoff - (c * dist_d + std::sqrt(det)) / (1 + c * c);
287 /* A negative value signals out of range */
293 /* Accumulate the extremes over all pulses */
294 for (int i = 0; i < numElementsInBuffer; i++)
302 if (receiveValidData)
304 buf_e[i].min0 = std::min(buf_e[i].min0, buf_r[i].min0);
305 buf_e[i].max1 = std::max(buf_e[i].max1, buf_r[i].max1);
306 buf_e[i].min1 = std::min(buf_e[i].min1, buf_r[i].min1);
310 if (dd->ndim == 3 && d == 0 && i == numElementsInBuffer - 1)
318 if (receiveValidData && dh[d1] >= 0)
320 buf_e[i].mch0 = std::max(buf_e[i].mch0, buf_r[i].mch0 - dh[d1]);
321 buf_e[i].mch1 = std::max(buf_e[i].mch1, buf_r[i].mch1 - dh[d1]);
324 /* Copy the received buffer to the send buffer,
325 * to pass the data through with the next pulse.
329 if (((applyPbc || dd->ci[dim] + numPulses < dd->numCells[dim]) && pulse == numPulses - 1)
330 || (!applyPbc && dd->ci[dim] + 1 + pulse == dd->numCells[dim] - 1))
332 /* Store the extremes */
335 for (int d1 = d; d1 < dd->ndim - 1; d1++)
337 extr_s[d1][1] = std::min(extr_s[d1][1], buf_e[pos].min0);
338 extr_s[d1][0] = std::max(extr_s[d1][0], buf_e[pos].max1);
339 extr_s[d1][2] = std::min(extr_s[d1][2], buf_e[pos].min1);
343 if (d == 1 || (d == 0 && dd->ndim == 3))
345 for (int i = d; i < 2; i++)
347 comm->zone_d2[1 - d][i] = buf_e[pos];
353 comm->zone_d1[1] = buf_e[pos];
359 if (d == 1 || (d == 0 && dd->ndim == 3))
361 for (int i = d; i < 2; i++)
363 comm->zone_d2[1 - d][i].dataSet = 0;
368 comm->zone_d1[1].dataSet = 0;
376 int dim = dd->dim[1];
377 for (int i = 0; i < 2; i++)
379 if (comm->zone_d1[i].dataSet != 0)
383 print_ddzone(debug, 1, i, 0, &comm->zone_d1[i]);
385 cell_ns_x0[dim] = std::min(cell_ns_x0[dim], comm->zone_d1[i].min0);
386 cell_ns_x1[dim] = std::max(cell_ns_x1[dim], comm->zone_d1[i].max1);
392 int dim = dd->dim[2];
393 for (int i = 0; i < 2; i++)
395 for (int j = 0; j < 2; j++)
397 if (comm->zone_d2[i][j].dataSet != 0)
401 print_ddzone(debug, 2, i, j, &comm->zone_d2[i][j]);
403 cell_ns_x0[dim] = std::min(cell_ns_x0[dim], comm->zone_d2[i][j].min0);
404 cell_ns_x1[dim] = std::max(cell_ns_x1[dim], comm->zone_d2[i][j].max1);
409 for (int d = 1; d < dd->ndim; d++)
411 cellsizes[d].fracLowerMax = extr_s[d - 1][0];
412 cellsizes[d].fracUpperMin = extr_s[d - 1][1];
415 fprintf(debug, "Cell fraction d %d, max0 %f, min1 %f\n", d, cellsizes[d].fracLowerMax,
416 cellsizes[d].fracUpperMin);
421 //! Sets the charge-group zones to be equal to the home zone.
422 static void set_zones_ncg_home(gmx_domdec_t* dd)
424 gmx_domdec_zones_t* zones;
427 zones = &dd->comm->zones;
429 zones->cg_range[0] = 0;
430 for (i = 1; i < zones->n + 1; i++)
432 zones->cg_range[i] = dd->ncg_home;
434 /* zone_ncg1[0] should always be equal to ncg_home */
435 dd->comm->zone_ncg1[0] = dd->ncg_home;
438 //! Restore atom groups for the charge groups.
439 static void restoreAtomGroups(gmx_domdec_t* dd, const t_state* state)
441 gmx::ArrayRef<const int> atomGroupsState = state->cg_gl;
443 std::vector<int>& globalAtomGroupIndices = dd->globalAtomGroupIndices;
445 globalAtomGroupIndices.resize(atomGroupsState.size());
447 /* Copy back the global charge group indices from state
448 * and rebuild the local charge group to atom index.
450 for (gmx::index i = 0; i < atomGroupsState.ssize(); i++)
452 globalAtomGroupIndices[i] = atomGroupsState[i];
455 dd->ncg_home = atomGroupsState.size();
456 dd->comm->atomRanges.setEnd(DDAtomRanges::Type::Home, atomGroupsState.ssize());
458 set_zones_ncg_home(dd);
461 //! Sets the cginfo structures.
462 static void dd_set_cginfo(gmx::ArrayRef<const int> index_gl, int cg0, int cg1, t_forcerec* fr)
466 gmx::ArrayRef<cginfo_mb_t> cginfo_mb = fr->cginfo_mb;
467 gmx::ArrayRef<int> cginfo = fr->cginfo;
469 for (int cg = cg0; cg < cg1; cg++)
471 cginfo[cg] = ddcginfo(cginfo_mb, index_gl[cg]);
476 //! Makes the mappings between global and local atom indices during DD repartioning.
477 static void make_dd_indices(gmx_domdec_t* dd, const int atomStart)
479 const int numZones = dd->comm->zones.n;
480 const int* zone2cg = dd->comm->zones.cg_range;
481 const int* zone_ncg1 = dd->comm->zone_ncg1;
482 gmx::ArrayRef<const int> globalAtomGroupIndices = dd->globalAtomGroupIndices;
484 std::vector<int>& globalAtomIndices = dd->globalAtomIndices;
485 gmx_ga2la_t& ga2la = *dd->ga2la;
487 if (zone2cg[1] != dd->ncg_home)
489 gmx_incons("dd->ncg_zone is not up to date");
492 /* Make the local to global and global to local atom index */
494 globalAtomIndices.resize(a);
495 for (int zone = 0; zone < numZones; zone++)
506 int cg1 = zone2cg[zone + 1];
507 int cg1_p1 = cg0 + zone_ncg1[zone];
509 for (int cg = cg0; cg < cg1; cg++)
514 /* Signal that this cg is from more than one pulse away */
517 int cg_gl = globalAtomGroupIndices[cg];
518 globalAtomIndices.push_back(cg_gl);
519 ga2la.insert(cg_gl, { a, zone1 });
525 //! Checks whether global and local atom indices are consistent.
526 static void check_index_consistency(const gmx_domdec_t* dd, int natoms_sys, const char* where)
530 const int numAtomsInZones = dd->comm->atomRanges.end(DDAtomRanges::Type::Zones);
532 if (dd->comm->ddSettings.DD_debug > 1)
534 std::vector<int> have(natoms_sys);
535 for (int a = 0; a < numAtomsInZones; a++)
537 int globalAtomIndex = dd->globalAtomIndices[a];
538 if (have[globalAtomIndex] > 0)
540 fprintf(stderr, "DD rank %d: global atom %d occurs twice: index %d and %d\n",
541 dd->rank, globalAtomIndex + 1, have[globalAtomIndex], a + 1);
545 have[globalAtomIndex] = a + 1;
550 std::vector<int> have(numAtomsInZones);
553 for (int i = 0; i < natoms_sys; i++)
555 if (const auto entry = dd->ga2la->find(i))
557 const int a = entry->la;
558 if (a >= numAtomsInZones)
561 "DD rank %d: global atom %d marked as local atom %d, which is larger than "
563 dd->rank, i + 1, a + 1, numAtomsInZones);
569 if (dd->globalAtomIndices[a] != i)
572 "DD rank %d: global atom %d marked as local atom %d, which has global "
574 dd->rank, i + 1, a + 1, dd->globalAtomIndices[a] + 1);
581 if (ngl != numAtomsInZones)
583 fprintf(stderr, "DD rank %d, %s: %d global atom indices, %d local atoms\n", dd->rank, where,
584 ngl, numAtomsInZones);
586 for (int a = 0; a < numAtomsInZones; a++)
590 fprintf(stderr, "DD rank %d, %s: local atom %d, global %d has no global index\n",
591 dd->rank, where, a + 1, dd->globalAtomIndices[a] + 1);
597 gmx_fatal(FARGS, "DD rank %d, %s: %d atom(group) index inconsistencies", dd->rank, where, nerr);
601 //! Clear all DD global state indices
602 static void clearDDStateIndices(gmx_domdec_t* dd, const bool keepLocalAtomIndices)
604 gmx_ga2la_t& ga2la = *dd->ga2la;
606 if (!keepLocalAtomIndices)
608 /* Clear the whole list without the overhead of searching */
613 const int numAtomsInZones = dd->comm->atomRanges.end(DDAtomRanges::Type::Zones);
614 for (int i = 0; i < numAtomsInZones; i++)
616 ga2la.erase(dd->globalAtomIndices[i]);
620 dd_clear_local_vsite_indices(dd);
624 dd_clear_local_constraint_indices(dd);
628 bool check_grid_jump(int64_t step, const gmx_domdec_t* dd, real cutoff, const gmx_ddbox_t* ddbox, gmx_bool bFatal)
630 gmx_domdec_comm_t* comm = dd->comm;
631 bool invalid = false;
633 for (int d = 1; d < dd->ndim; d++)
635 const DDCellsizesWithDlb& cellsizes = comm->cellsizesWithDlb[d];
636 const int dim = dd->dim[d];
637 const real limit = grid_jump_limit(comm, cutoff, d);
638 real bfac = ddbox->box_size[dim];
639 if (ddbox->tric_dir[dim])
641 bfac *= ddbox->skew_fac[dim];
643 if ((cellsizes.fracUpper - cellsizes.fracLowerMax) * bfac < limit
644 || (cellsizes.fracLower - cellsizes.fracUpperMin) * bfac > -limit)
652 /* This error should never be triggered under normal
653 * circumstances, but you never know ...
656 "step %s: The domain decomposition grid has shifted too much in the "
657 "%c-direction around cell %d %d %d. This should not have happened. "
658 "Running with fewer ranks might avoid this issue.",
659 gmx_step_str(step, buf), dim2char(dim), dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
666 //! Return the duration of force calculations on this rank.
667 static float dd_force_load(gmx_domdec_comm_t* comm)
671 if (comm->ddSettings.eFlop)
674 if (comm->ddSettings.eFlop > 1)
676 load *= 1.0 + (comm->ddSettings.eFlop - 1) * (0.1 * rand() / RAND_MAX - 0.05);
681 load = comm->cycl[ddCyclF];
682 if (comm->cycl_n[ddCyclF] > 1)
684 /* Subtract the maximum of the last n cycle counts
685 * to get rid of possible high counts due to other sources,
686 * for instance system activity, that would otherwise
687 * affect the dynamic load balancing.
689 load -= comm->cycl_max[ddCyclF];
693 if (comm->cycl_n[ddCyclWaitGPU] && comm->nrank_gpu_shared > 1)
695 float gpu_wait, gpu_wait_sum;
697 gpu_wait = comm->cycl[ddCyclWaitGPU];
698 if (comm->cycl_n[ddCyclF] > 1)
700 /* We should remove the WaitGPU time of the same MD step
701 * as the one with the maximum F time, since the F time
702 * and the wait time are not independent.
703 * Furthermore, the step for the max F time should be chosen
704 * the same on all ranks that share the same GPU.
705 * But to keep the code simple, we remove the average instead.
706 * The main reason for artificially long times at some steps
707 * is spurious CPU activity or MPI time, so we don't expect
708 * that changes in the GPU wait time matter a lot here.
710 gpu_wait *= (comm->cycl_n[ddCyclF] - 1) / static_cast<float>(comm->cycl_n[ddCyclF]);
712 /* Sum the wait times over the ranks that share the same GPU */
713 MPI_Allreduce(&gpu_wait, &gpu_wait_sum, 1, MPI_FLOAT, MPI_SUM, comm->mpi_comm_gpu_shared);
714 /* Replace the wait time by the average over the ranks */
715 load += -gpu_wait + gpu_wait_sum / comm->nrank_gpu_shared;
723 //! Runs cell size checks and communicates the boundaries.
724 static void comm_dd_ns_cell_sizes(gmx_domdec_t* dd, gmx_ddbox_t* ddbox, rvec cell_ns_x0, rvec cell_ns_x1, int64_t step)
726 gmx_domdec_comm_t* comm;
731 for (dim_ind = 0; dim_ind < dd->ndim; dim_ind++)
733 dim = dd->dim[dim_ind];
735 /* Without PBC we don't have restrictions on the outer cells */
736 if (!(dim >= ddbox->npbcdim && (dd->ci[dim] == 0 || dd->ci[dim] == dd->numCells[dim] - 1))
738 && (comm->cell_x1[dim] - comm->cell_x0[dim]) * ddbox->skew_fac[dim] < comm->cellsize_min[dim])
742 "step %s: The %c-size (%f) times the triclinic skew factor (%f) is smaller "
743 "than the smallest allowed cell size (%f) for domain decomposition grid cell "
745 gmx_step_str(step, buf), dim2char(dim),
746 comm->cell_x1[dim] - comm->cell_x0[dim], ddbox->skew_fac[dim],
747 dd->comm->cellsize_min[dim], dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
751 if ((isDlbOn(dd->comm) && dd->ndim > 1) || ddbox->nboundeddim < DIM)
753 /* Communicate the boundaries and update cell_ns_x0/1 */
754 dd_move_cellx(dd, ddbox, cell_ns_x0, cell_ns_x1);
755 if (isDlbOn(dd->comm) && dd->ndim > 1)
757 check_grid_jump(step, dd, dd->comm->systemInfo.cutoff, ddbox, TRUE);
762 //! Compute and communicate to determine the load distribution across PP ranks.
763 static void get_load_distribution(gmx_domdec_t* dd, gmx_wallcycle_t wcycle)
765 gmx_domdec_comm_t* comm;
767 float cell_frac = 0, sbuf[DD_NLOAD_MAX];
772 fprintf(debug, "get_load_distribution start\n");
775 wallcycle_start(wcycle, ewcDDCOMMLOAD);
779 bSepPME = (dd->pme_nodeid >= 0);
781 if (dd->ndim == 0 && bSepPME)
783 /* Without decomposition, but with PME nodes, we need the load */
784 comm->load[0].mdf = comm->cycl[ddCyclPPduringPME];
785 comm->load[0].pme = comm->cycl[ddCyclPME];
788 for (int d = dd->ndim - 1; d >= 0; d--)
790 const DDCellsizesWithDlb* cellsizes = (isDlbOn(dd->comm) ? &comm->cellsizesWithDlb[d] : nullptr);
791 const int dim = dd->dim[d];
792 /* Check if we participate in the communication in this dimension */
793 if (d == dd->ndim - 1 || (dd->ci[dd->dim[d + 1]] == 0 && dd->ci[dd->dim[dd->ndim - 1]] == 0))
795 load = &comm->load[d];
796 if (isDlbOn(dd->comm))
798 cell_frac = cellsizes->fracUpper - cellsizes->fracLower;
801 if (d == dd->ndim - 1)
803 sbuf[pos++] = dd_force_load(comm);
804 sbuf[pos++] = sbuf[0];
805 if (isDlbOn(dd->comm))
807 sbuf[pos++] = sbuf[0];
808 sbuf[pos++] = cell_frac;
811 sbuf[pos++] = cellsizes->fracLowerMax;
812 sbuf[pos++] = cellsizes->fracUpperMin;
817 sbuf[pos++] = comm->cycl[ddCyclPPduringPME];
818 sbuf[pos++] = comm->cycl[ddCyclPME];
823 sbuf[pos++] = comm->load[d + 1].sum;
824 sbuf[pos++] = comm->load[d + 1].max;
825 if (isDlbOn(dd->comm))
827 sbuf[pos++] = comm->load[d + 1].sum_m;
828 sbuf[pos++] = comm->load[d + 1].cvol_min * cell_frac;
829 sbuf[pos++] = comm->load[d + 1].flags;
832 sbuf[pos++] = cellsizes->fracLowerMax;
833 sbuf[pos++] = cellsizes->fracUpperMin;
838 sbuf[pos++] = comm->load[d + 1].mdf;
839 sbuf[pos++] = comm->load[d + 1].pme;
843 /* Communicate a row in DD direction d.
844 * The communicators are setup such that the root always has rank 0.
847 MPI_Gather(sbuf, load->nload * sizeof(float), MPI_BYTE, load->load,
848 load->nload * sizeof(float), MPI_BYTE, 0, comm->mpi_comm_load[d]);
850 if (dd->ci[dim] == dd->master_ci[dim])
852 /* We are the master along this row, process this row */
853 RowMaster* rowMaster = nullptr;
857 rowMaster = cellsizes->rowMaster.get();
867 for (int i = 0; i < dd->numCells[dim]; i++)
869 load->sum += load->load[pos++];
870 load->max = std::max(load->max, load->load[pos]);
872 if (isDlbOn(dd->comm))
874 if (rowMaster->dlbIsLimited)
876 /* This direction could not be load balanced properly,
877 * therefore we need to use the maximum iso the average load.
879 load->sum_m = std::max(load->sum_m, load->load[pos]);
883 load->sum_m += load->load[pos];
886 load->cvol_min = std::min(load->cvol_min, load->load[pos]);
888 if (d < dd->ndim - 1)
890 load->flags = gmx::roundToInt(load->load[pos++]);
894 rowMaster->bounds[i].cellFracLowerMax = load->load[pos++];
895 rowMaster->bounds[i].cellFracUpperMin = load->load[pos++];
900 load->mdf = std::max(load->mdf, load->load[pos]);
902 load->pme = std::max(load->pme, load->load[pos]);
906 if (isDlbOn(comm) && rowMaster->dlbIsLimited)
908 load->sum_m *= dd->numCells[dim];
909 load->flags |= (1 << d);
917 comm->nload += dd_load_count(comm);
918 comm->load_step += comm->cycl[ddCyclStep];
919 comm->load_sum += comm->load[0].sum;
920 comm->load_max += comm->load[0].max;
923 for (int d = 0; d < dd->ndim; d++)
925 if (comm->load[0].flags & (1 << d))
933 comm->load_mdf += comm->load[0].mdf;
934 comm->load_pme += comm->load[0].pme;
938 wallcycle_stop(wcycle, ewcDDCOMMLOAD);
942 fprintf(debug, "get_load_distribution finished\n");
946 /*! \brief Return the relative performance loss on the total run time
947 * due to the force calculation load imbalance. */
948 static float dd_force_load_fraction(gmx_domdec_t* dd)
950 if (dd->comm->nload > 0 && dd->comm->load_step > 0)
952 return dd->comm->load_sum / (dd->comm->load_step * dd->nnodes);
960 /*! \brief Return the relative performance loss on the total run time
961 * due to the force calculation load imbalance. */
962 static float dd_force_imb_perf_loss(gmx_domdec_t* dd)
964 if (dd->comm->nload > 0 && dd->comm->load_step > 0)
966 return (dd->comm->load_max * dd->nnodes - dd->comm->load_sum) / (dd->comm->load_step * dd->nnodes);
974 //! Print load-balance report e.g. at the end of a run.
975 static void print_dd_load_av(FILE* fplog, gmx_domdec_t* dd)
977 gmx_domdec_comm_t* comm = dd->comm;
979 /* Only the master rank prints loads and only if we measured loads */
980 if (!DDMASTER(dd) || comm->nload == 0)
986 int numPpRanks = dd->nnodes;
987 int numPmeRanks = (comm->ddRankSetup.usePmeOnlyRanks ? comm->ddRankSetup.numRanksDoingPme : 0);
988 int numRanks = numPpRanks + numPmeRanks;
989 float lossFraction = 0;
991 /* Print the average load imbalance and performance loss */
992 if (dd->nnodes > 1 && comm->load_sum > 0)
994 float imbalance = comm->load_max * numPpRanks / comm->load_sum - 1;
995 lossFraction = dd_force_imb_perf_loss(dd);
997 std::string msg = "\nDynamic load balancing report:\n";
998 std::string dlbStateStr;
1000 switch (dd->comm->dlbState)
1002 case DlbState::offUser:
1003 dlbStateStr = "DLB was off during the run per user request.";
1005 case DlbState::offForever:
1006 /* Currectly this can happen due to performance loss observed, cell size
1007 * limitations or incompatibility with other settings observed during
1008 * determineInitialDlbState(). */
1009 dlbStateStr = "DLB got disabled because it was unsuitable to use.";
1011 case DlbState::offCanTurnOn:
1012 dlbStateStr = "DLB was off during the run due to low measured imbalance.";
1014 case DlbState::offTemporarilyLocked:
1016 "DLB was locked at the end of the run due to unfinished PP-PME "
1019 case DlbState::onCanTurnOff:
1020 dlbStateStr = "DLB was turned on during the run due to measured imbalance.";
1022 case DlbState::onUser:
1023 dlbStateStr = "DLB was permanently on during the run per user request.";
1025 default: GMX_ASSERT(false, "Undocumented DLB state");
1028 msg += " " + dlbStateStr + "\n";
1029 msg += gmx::formatString(" Average load imbalance: %.1f%%.\n", imbalance * 100);
1030 msg += gmx::formatString(
1031 " The balanceable part of the MD step is %d%%, load imbalance is computed from "
1033 gmx::roundToInt(dd_force_load_fraction(dd) * 100));
1034 msg += gmx::formatString(
1035 " Part of the total run time spent waiting due to load imbalance: %.1f%%.\n",
1036 lossFraction * 100);
1037 fprintf(fplog, "%s", msg.c_str());
1038 fprintf(stderr, "\n%s", msg.c_str());
1041 /* Print during what percentage of steps the load balancing was limited */
1042 bool dlbWasLimited = false;
1045 sprintf(buf, " Steps where the load balancing was limited by -rdd, -rcon and/or -dds:");
1046 for (int d = 0; d < dd->ndim; d++)
1048 int limitPercentage = (200 * comm->load_lim[d] + 1) / (2 * comm->nload);
1049 sprintf(buf + strlen(buf), " %c %d %%", dim2char(dd->dim[d]), limitPercentage);
1050 if (limitPercentage >= 50)
1052 dlbWasLimited = true;
1055 sprintf(buf + strlen(buf), "\n");
1056 fprintf(fplog, "%s", buf);
1057 fprintf(stderr, "%s", buf);
1060 /* Print the performance loss due to separate PME - PP rank imbalance */
1061 float lossFractionPme = 0;
1062 if (numPmeRanks > 0 && comm->load_mdf > 0 && comm->load_step > 0)
1064 float pmeForceRatio = comm->load_pme / comm->load_mdf;
1065 lossFractionPme = (comm->load_pme - comm->load_mdf) / comm->load_step;
1066 if (lossFractionPme <= 0)
1068 lossFractionPme *= numPmeRanks / static_cast<float>(numRanks);
1072 lossFractionPme *= numPpRanks / static_cast<float>(numRanks);
1074 sprintf(buf, " Average PME mesh/force load: %5.3f\n", pmeForceRatio);
1075 fprintf(fplog, "%s", buf);
1076 fprintf(stderr, "%s", buf);
1077 sprintf(buf, " Part of the total run time spent waiting due to PP/PME imbalance: %.1f %%\n",
1078 std::fabs(lossFractionPme) * 100);
1079 fprintf(fplog, "%s", buf);
1080 fprintf(stderr, "%s", buf);
1082 fprintf(fplog, "\n");
1083 fprintf(stderr, "\n");
1085 if (lossFraction >= DD_PERF_LOSS_WARN)
1087 std::string message = gmx::formatString(
1088 "NOTE: %.1f %% of the available CPU time was lost due to load imbalance\n"
1089 " in the domain decomposition.\n",
1090 lossFraction * 100);
1092 bool hadSuggestion = false;
1095 message += " You might want to use dynamic load balancing (option -dlb.)\n";
1096 hadSuggestion = true;
1098 else if (dlbWasLimited)
1101 " You might want to decrease the cell size limit (options -rdd, -rcon "
1103 hadSuggestion = true;
1105 message += gmx::formatString(
1106 " You can %sconsider manually changing the decomposition (option -dd);\n"
1107 " e.g. by using fewer domains along the box dimension in which there is\n"
1108 " considerable inhomogeneity in the simulated system.",
1109 hadSuggestion ? "also " : "");
1112 fprintf(fplog, "%s\n", message.c_str());
1113 fprintf(stderr, "%s\n", message.c_str());
1115 if (numPmeRanks > 0 && std::fabs(lossFractionPme) >= DD_PERF_LOSS_WARN)
1118 "NOTE: %.1f %% performance was lost because the PME ranks\n"
1119 " had %s work to do than the PP ranks.\n"
1120 " You might want to %s the number of PME ranks\n"
1121 " or %s the cut-off and the grid spacing.\n",
1122 std::fabs(lossFractionPme * 100), (lossFractionPme < 0) ? "less" : "more",
1123 (lossFractionPme < 0) ? "decrease" : "increase",
1124 (lossFractionPme < 0) ? "decrease" : "increase");
1125 fprintf(fplog, "%s\n", buf);
1126 fprintf(stderr, "%s\n", buf);
1130 //! Return the minimum communication volume.
1131 static float dd_vol_min(gmx_domdec_t* dd)
1133 return dd->comm->load[0].cvol_min * dd->nnodes;
1136 //! Return the DD load flags.
1137 static int dd_load_flags(gmx_domdec_t* dd)
1139 return dd->comm->load[0].flags;
1142 //! Return the reported load imbalance in force calculations.
1143 static float dd_f_imbal(gmx_domdec_t* dd)
1145 if (dd->comm->load[0].sum > 0)
1147 return dd->comm->load[0].max * dd->nnodes / dd->comm->load[0].sum - 1.0F;
1151 /* Something is wrong in the cycle counting, report no load imbalance */
1156 //! Returns DD load balance report.
1157 static std::string dd_print_load(gmx_domdec_t* dd, int64_t step)
1159 gmx::StringOutputStream stream;
1160 gmx::TextWriter log(&stream);
1162 int flags = dd_load_flags(dd);
1165 log.writeString("DD load balancing is limited by minimum cell size in dimension");
1166 for (int d = 0; d < dd->ndim; d++)
1168 if (flags & (1 << d))
1170 log.writeStringFormatted(" %c", dim2char(dd->dim[d]));
1173 log.ensureLineBreak();
1175 log.writeString("DD step " + gmx::toString(step));
1176 if (isDlbOn(dd->comm))
1178 log.writeStringFormatted(" vol min/aver %5.3f%c", dd_vol_min(dd), flags ? '!' : ' ');
1182 log.writeStringFormatted(" load imb.: force %4.1f%%", dd_f_imbal(dd) * 100);
1184 if (dd->comm->cycl_n[ddCyclPME])
1186 log.writeStringFormatted(" pme mesh/force %5.3f", dd_pme_f_ratio(dd));
1188 log.ensureLineBreak();
1189 return stream.toString();
1192 //! Prints DD load balance report in mdrun verbose mode.
1193 static void dd_print_load_verbose(gmx_domdec_t* dd)
1195 if (isDlbOn(dd->comm))
1197 fprintf(stderr, "vol %4.2f%c ", dd_vol_min(dd), dd_load_flags(dd) ? '!' : ' ');
1201 fprintf(stderr, "imb F %2d%% ", gmx::roundToInt(dd_f_imbal(dd) * 100));
1203 if (dd->comm->cycl_n[ddCyclPME])
1205 fprintf(stderr, "pme/F %4.2f ", dd_pme_f_ratio(dd));
1209 //! Turns on dynamic load balancing if possible and needed.
1210 static void turn_on_dlb(const gmx::MDLogger& mdlog, gmx_domdec_t* dd, int64_t step)
1212 gmx_domdec_comm_t* comm = dd->comm;
1214 real cellsize_min = comm->cellsize_min[dd->dim[0]];
1215 for (int d = 1; d < dd->ndim; d++)
1217 cellsize_min = std::min(cellsize_min, comm->cellsize_min[dd->dim[d]]);
1220 /* Turn off DLB if we're too close to the cell size limit. */
1221 if (cellsize_min < comm->cellsize_limit * 1.05)
1224 .appendTextFormatted(
1225 "step %s Measured %.1f %% performance loss due to load imbalance, "
1226 "but the minimum cell size is smaller than 1.05 times the cell size limit. "
1227 "Will no longer try dynamic load balancing.",
1228 gmx::toString(step).c_str(), dd_force_imb_perf_loss(dd) * 100);
1230 comm->dlbState = DlbState::offForever;
1235 .appendTextFormatted(
1236 "step %s Turning on dynamic load balancing, because the performance loss due "
1237 "to load imbalance is %.1f %%.",
1238 gmx::toString(step).c_str(), dd_force_imb_perf_loss(dd) * 100);
1239 comm->dlbState = DlbState::onCanTurnOff;
1241 /* Store the non-DLB performance, so we can check if DLB actually
1242 * improves performance.
1244 GMX_RELEASE_ASSERT(comm->cycl_n[ddCyclStep] > 0,
1245 "When we turned on DLB, we should have measured cycles");
1246 comm->cyclesPerStepBeforeDLB = comm->cycl[ddCyclStep] / comm->cycl_n[ddCyclStep];
1250 /* We can set the required cell size info here,
1251 * so we do not need to communicate this.
1252 * The grid is completely uniform.
1254 for (int d = 0; d < dd->ndim; d++)
1256 RowMaster* rowMaster = comm->cellsizesWithDlb[d].rowMaster.get();
1260 comm->load[d].sum_m = comm->load[d].sum;
1262 int nc = dd->numCells[dd->dim[d]];
1263 for (int i = 0; i < nc; i++)
1265 rowMaster->cellFrac[i] = i / static_cast<real>(nc);
1268 rowMaster->bounds[i].cellFracLowerMax = i / static_cast<real>(nc);
1269 rowMaster->bounds[i].cellFracUpperMin = (i + 1) / static_cast<real>(nc);
1272 rowMaster->cellFrac[nc] = 1.0;
1277 //! Turns off dynamic load balancing (but leave it able to turn back on).
1278 static void turn_off_dlb(const gmx::MDLogger& mdlog, gmx_domdec_t* dd, int64_t step)
1282 "step " + gmx::toString(step)
1283 + " Turning off dynamic load balancing, because it is degrading performance.");
1284 dd->comm->dlbState = DlbState::offCanTurnOn;
1285 dd->comm->haveTurnedOffDlb = true;
1286 dd->comm->ddPartioningCountFirstDlbOff = dd->ddp_count;
1289 //! Turns off dynamic load balancing permanently.
1290 static void turn_off_dlb_forever(const gmx::MDLogger& mdlog, gmx_domdec_t* dd, int64_t step)
1292 GMX_RELEASE_ASSERT(dd->comm->dlbState == DlbState::offCanTurnOn,
1293 "Can only turn off DLB forever when it was in the can-turn-on state");
1296 "step " + gmx::toString(step)
1297 + " Will no longer try dynamic load balancing, as it degraded performance.");
1298 dd->comm->dlbState = DlbState::offForever;
1301 void set_dd_dlb_max_cutoff(t_commrec* cr, real cutoff)
1303 gmx_domdec_comm_t* comm;
1305 comm = cr->dd->comm;
1307 /* Turn on the DLB limiting (might have been on already) */
1308 comm->bPMELoadBalDLBLimits = TRUE;
1310 /* Change the cut-off limit */
1311 comm->PMELoadBal_max_cutoff = cutoff;
1316 "PME load balancing set a limit to the DLB staggering such that a %f cut-off will "
1317 "continue to fit\n",
1318 comm->PMELoadBal_max_cutoff);
1322 //! Merge atom buffers.
1323 static void merge_cg_buffers(int ncell,
1324 gmx_domdec_comm_dim_t* cd,
1327 gmx::ArrayRef<int> index_gl,
1329 gmx::ArrayRef<gmx::RVec> x,
1330 gmx::ArrayRef<const gmx::RVec> recv_vr,
1331 gmx::ArrayRef<cginfo_mb_t> cginfo_mb,
1332 gmx::ArrayRef<int> cginfo)
1334 gmx_domdec_ind_t *ind, *ind_p;
1335 int p, cell, c, cg, cg0, cg1, cg_gl;
1338 ind = &cd->ind[pulse];
1340 /* First correct the already stored data */
1341 shift = ind->nrecv[ncell];
1342 for (cell = ncell - 1; cell >= 0; cell--)
1344 shift -= ind->nrecv[cell];
1347 /* Move the cg's present from previous grid pulses */
1348 cg0 = ncg_cell[ncell + cell];
1349 cg1 = ncg_cell[ncell + cell + 1];
1350 for (cg = cg1 - 1; cg >= cg0; cg--)
1352 index_gl[cg + shift] = index_gl[cg];
1353 x[cg + shift] = x[cg];
1354 cginfo[cg + shift] = cginfo[cg];
1356 /* Correct the already stored send indices for the shift */
1357 for (p = 1; p <= pulse; p++)
1359 ind_p = &cd->ind[p];
1361 for (c = 0; c < cell; c++)
1363 cg0 += ind_p->nsend[c];
1365 cg1 = cg0 + ind_p->nsend[cell];
1366 for (cg = cg0; cg < cg1; cg++)
1368 ind_p->index[cg] += shift;
1374 /* Merge in the communicated buffers */
1377 for (cell = 0; cell < ncell; cell++)
1379 cg1 = ncg_cell[ncell + cell + 1] + shift;
1380 for (cg = 0; cg < ind->nrecv[cell]; cg++)
1382 /* Copy this atom from the buffer */
1383 index_gl[cg1] = recv_i[cg0];
1384 x[cg1] = recv_vr[cg0];
1385 /* Copy information */
1386 cg_gl = index_gl[cg1];
1387 cginfo[cg1] = ddcginfo(cginfo_mb, cg_gl);
1391 shift += ind->nrecv[cell];
1392 ncg_cell[ncell + cell + 1] = cg1;
1396 //! Makes a range partitioning for the atom groups wthin a cell
1397 static void make_cell2at_index(gmx_domdec_comm_dim_t* cd, int nzone, int atomGroupStart)
1399 /* Store the atom block boundaries for easy copying of communication buffers
1401 int g = atomGroupStart;
1402 for (int zone = 0; zone < nzone; zone++)
1404 for (gmx_domdec_ind_t& ind : cd->ind)
1406 ind.cell2at0[zone] = g;
1407 g += ind.nrecv[zone];
1408 ind.cell2at1[zone] = g;
1413 //! Returns whether a link is missing.
1414 static gmx_bool missing_link(const t_blocka& link, const int globalAtomIndex, const gmx_ga2la_t& ga2la)
1416 for (int i = link.index[globalAtomIndex]; i < link.index[globalAtomIndex + 1]; i++)
1418 if (!ga2la.findHome(link.a[i]))
1427 //! Domain corners for communication, a maximum of 4 i-zones see a j domain
1430 //! The corners for the non-bonded communication.
1432 //! Corner for rounding.
1434 //! Corners for rounding.
1436 //! Corners for bounded communication.
1438 //! Corner for rounding for bonded communication.
1442 //! Determine the corners of the domain(s) we are communicating with.
1443 static void set_dd_corners(const gmx_domdec_t* dd, int dim0, int dim1, int dim2, gmx_bool bDistMB, dd_corners_t* c)
1445 const gmx_domdec_comm_t* comm;
1446 const gmx_domdec_zones_t* zones;
1450 zones = &comm->zones;
1452 /* Keep the compiler happy */
1456 /* The first dimension is equal for all cells */
1457 c->c[0][0] = comm->cell_x0[dim0];
1460 c->bc[0] = c->c[0][0];
1465 /* This cell row is only seen from the first row */
1466 c->c[1][0] = comm->cell_x0[dim1];
1467 /* All rows can see this row */
1468 c->c[1][1] = comm->cell_x0[dim1];
1469 if (isDlbOn(dd->comm))
1471 c->c[1][1] = std::max(comm->cell_x0[dim1], comm->zone_d1[1].mch0);
1474 /* For the multi-body distance we need the maximum */
1475 c->bc[1] = std::max(comm->cell_x0[dim1], comm->zone_d1[1].p1_0);
1478 /* Set the upper-right corner for rounding */
1479 c->cr0 = comm->cell_x1[dim0];
1484 for (int j = 0; j < 4; j++)
1486 c->c[2][j] = comm->cell_x0[dim2];
1488 if (isDlbOn(dd->comm))
1490 /* Use the maximum of the i-cells that see a j-cell */
1491 for (const auto& iZone : zones->iZones)
1493 const int iZoneIndex = iZone.iZoneIndex;
1494 for (int jZone : iZone.jZoneRange)
1498 c->c[2][jZone - 4] = std::max(
1500 comm->zone_d2[zones->shift[iZoneIndex][dim0]][zones->shift[iZoneIndex][dim1]]
1507 /* For the multi-body distance we need the maximum */
1508 c->bc[2] = comm->cell_x0[dim2];
1509 for (int i = 0; i < 2; i++)
1511 for (int j = 0; j < 2; j++)
1513 c->bc[2] = std::max(c->bc[2], comm->zone_d2[i][j].p1_0);
1519 /* Set the upper-right corner for rounding */
1520 /* Cell (0,0,0) and cell (1,0,0) can see cell 4 (0,1,1)
1521 * Only cell (0,0,0) can see cell 7 (1,1,1)
1523 c->cr1[0] = comm->cell_x1[dim1];
1524 c->cr1[3] = comm->cell_x1[dim1];
1525 if (isDlbOn(dd->comm))
1527 c->cr1[0] = std::max(comm->cell_x1[dim1], comm->zone_d1[1].mch1);
1530 /* For the multi-body distance we need the maximum */
1531 c->bcr1 = std::max(comm->cell_x1[dim1], comm->zone_d1[1].p1_1);
1538 /*! \brief Add the atom groups we need to send in this pulse from this
1539 * zone to \p localAtomGroups and \p work. */
1540 static void get_zone_pulse_cgs(gmx_domdec_t* dd,
1545 gmx::ArrayRef<const int> globalAtomGroupIndices,
1554 bool distanceIsTriclinic,
1561 const dd_corners_t* c,
1562 const rvec sf2_round,
1563 gmx_bool bDistBonded,
1568 gmx::ArrayRef<const int> cginfo,
1569 std::vector<int>* localAtomGroups,
1570 dd_comm_setup_work_t* work)
1572 gmx_domdec_comm_t* comm;
1574 gmx_bool bDistMB_pulse;
1576 real r2, rb2, r, tric_sh;
1583 bScrew = (dd->unitCellInfo.haveScrewPBC && dim == XX);
1585 bDistMB_pulse = (bDistMB && bDistBonded);
1587 /* Unpack the work data */
1588 std::vector<int>& ibuf = work->atomGroupBuffer;
1589 std::vector<gmx::RVec>& vbuf = work->positionBuffer;
1593 for (cg = cg0; cg < cg1; cg++)
1597 if (!distanceIsTriclinic)
1599 /* Rectangular direction, easy */
1600 r = cg_cm[cg][dim] - c->c[dim_ind][zone];
1607 r = cg_cm[cg][dim] - c->bc[dim_ind];
1613 /* Rounding gives at most a 16% reduction
1614 * in communicated atoms
1616 if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
1618 r = cg_cm[cg][dim0] - c->cr0;
1619 /* This is the first dimension, so always r >= 0 */
1626 if (dim_ind == 2 && (zonei == 2 || zonei == 3))
1628 r = cg_cm[cg][dim1] - c->cr1[zone];
1635 r = cg_cm[cg][dim1] - c->bcr1;
1645 /* Triclinic direction, more complicated */
1648 /* Rounding, conservative as the skew_fac multiplication
1649 * will slightly underestimate the distance.
1651 if (dim_ind >= 1 && (zonei == 1 || zonei == 2))
1653 rn[dim0] = cg_cm[cg][dim0] - c->cr0;
1654 for (i = dim0 + 1; i < DIM; i++)
1656 rn[dim0] -= cg_cm[cg][i] * v_0[i][dim0];
1658 r2 = rn[dim0] * rn[dim0] * sf2_round[dim0];
1661 rb[dim0] = rn[dim0];
1664 /* Take care that the cell planes along dim0 might not
1665 * be orthogonal to those along dim1 and dim2.
1667 for (i = 1; i <= dim_ind; i++)
1670 if (normal[dim0][dimd] > 0)
1672 rn[dimd] -= rn[dim0] * normal[dim0][dimd];
1675 rb[dimd] -= rb[dim0] * normal[dim0][dimd];
1680 if (dim_ind == 2 && (zonei == 2 || zonei == 3))
1682 GMX_ASSERT(dim1 >= 0 && dim1 < DIM, "Must have a valid dimension index");
1683 rn[dim1] += cg_cm[cg][dim1] - c->cr1[zone];
1685 for (i = dim1 + 1; i < DIM; i++)
1687 tric_sh -= cg_cm[cg][i] * v_1[i][dim1];
1689 rn[dim1] += tric_sh;
1692 r2 += rn[dim1] * rn[dim1] * sf2_round[dim1];
1693 /* Take care of coupling of the distances
1694 * to the planes along dim0 and dim1 through dim2.
1696 r2 -= rn[dim0] * rn[dim1] * skew_fac_01;
1697 /* Take care that the cell planes along dim1
1698 * might not be orthogonal to that along dim2.
1700 if (normal[dim1][dim2] > 0)
1702 rn[dim2] -= rn[dim1] * normal[dim1][dim2];
1707 rb[dim1] += cg_cm[cg][dim1] - c->bcr1 + tric_sh;
1710 rb2 += rb[dim1] * rb[dim1] * sf2_round[dim1];
1711 /* Take care of coupling of the distances
1712 * to the planes along dim0 and dim1 through dim2.
1714 rb2 -= rb[dim0] * rb[dim1] * skew_fac_01;
1715 /* Take care that the cell planes along dim1
1716 * might not be orthogonal to that along dim2.
1718 if (normal[dim1][dim2] > 0)
1720 rb[dim2] -= rb[dim1] * normal[dim1][dim2];
1725 /* The distance along the communication direction */
1726 rn[dim] += cg_cm[cg][dim] - c->c[dim_ind][zone];
1728 for (i = dim + 1; i < DIM; i++)
1730 tric_sh -= cg_cm[cg][i] * v_d[i][dim];
1735 r2 += rn[dim] * rn[dim] * skew_fac2_d;
1736 /* Take care of coupling of the distances
1737 * to the planes along dim0 and dim1 through dim2.
1739 if (dim_ind == 1 && zonei == 1)
1741 r2 -= rn[dim0] * rn[dim] * skew_fac_01;
1747 GMX_ASSERT(dim >= 0 && dim < DIM, "Must have a valid dimension index");
1748 rb[dim] += cg_cm[cg][dim] - c->bc[dim_ind] + tric_sh;
1751 rb2 += rb[dim] * rb[dim] * skew_fac2_d;
1752 /* Take care of coupling of the distances
1753 * to the planes along dim0 and dim1 through dim2.
1755 if (dim_ind == 1 && zonei == 1)
1757 rb2 -= rb[dim0] * rb[dim] * skew_fac_01;
1764 || (bDistBonded && ((bDistMB && rb2 < r_bcomm2) || (bDist2B && r2 < r_bcomm2))
1766 || (GET_CGINFO_BOND_INTER(cginfo[cg])
1767 && missing_link(*comm->bondedLinks, globalAtomGroupIndices[cg], *dd->ga2la)))))
1769 /* Store the local and global atom group indices and position */
1770 localAtomGroups->push_back(cg);
1771 ibuf.push_back(globalAtomGroupIndices[cg]);
1775 if (dd->ci[dim] == 0)
1777 /* Correct cg_cm for pbc */
1778 rvec_add(cg_cm[cg], box[dim], posPbc);
1781 posPbc[YY] = box[YY][YY] - posPbc[YY];
1782 posPbc[ZZ] = box[ZZ][ZZ] - posPbc[ZZ];
1787 copy_rvec(cg_cm[cg], posPbc);
1789 vbuf.emplace_back(posPbc[XX], posPbc[YY], posPbc[ZZ]);
1796 work->nsend_zone = nsend_z;
1800 static void clearCommSetupData(dd_comm_setup_work_t* work)
1802 work->localAtomGroupBuffer.clear();
1803 work->atomGroupBuffer.clear();
1804 work->positionBuffer.clear();
1806 work->nsend_zone = 0;
1809 //! Prepare DD communication.
1810 static void setup_dd_communication(gmx_domdec_t* dd, matrix box, gmx_ddbox_t* ddbox, t_forcerec* fr, t_state* state)
1812 int dim_ind, dim, dim0, dim1, dim2, dimd, nat_tot;
1813 int nzone, nzone_send, zone, zonei, cg0, cg1;
1815 int * zone_cg_range, pos_cg;
1816 gmx_domdec_comm_t* comm;
1817 gmx_domdec_zones_t* zones;
1818 gmx_domdec_comm_dim_t* cd;
1819 gmx_bool bBondComm, bDist2B, bDistMB, bDistBonded;
1820 dd_corners_t corners;
1821 rvec * normal, *v_d, *v_0 = nullptr, *v_1 = nullptr;
1822 real skew_fac2_d, skew_fac_01;
1827 fprintf(debug, "Setting up DD communication\n");
1832 if (comm->dth.empty())
1834 /* Initialize the thread data.
1835 * This can not be done in init_domain_decomposition,
1836 * as the numbers of threads is determined later.
1838 int numThreads = gmx_omp_nthreads_get(emntDomdec);
1839 comm->dth.resize(numThreads);
1842 bBondComm = comm->systemInfo.filterBondedCommunication;
1844 /* Do we need to determine extra distances for multi-body bondeds? */
1845 bDistMB = (comm->systemInfo.haveInterDomainMultiBodyBondeds && isDlbOn(dd->comm) && dd->ndim > 1);
1847 /* Do we need to determine extra distances for only two-body bondeds? */
1848 bDist2B = (bBondComm && !bDistMB);
1850 const real r_comm2 =
1851 gmx::square(domainToDomainIntoAtomToDomainCutoff(comm->systemInfo, comm->systemInfo.cutoff));
1852 const real r_bcomm2 =
1853 gmx::square(domainToDomainIntoAtomToDomainCutoff(comm->systemInfo, comm->cutoff_mbody));
1857 fprintf(debug, "bBondComm %s, r_bc %f\n", gmx::boolToString(bBondComm), std::sqrt(r_bcomm2));
1860 zones = &comm->zones;
1863 dim1 = (dd->ndim >= 2 ? dd->dim[1] : -1);
1864 dim2 = (dd->ndim >= 3 ? dd->dim[2] : -1);
1866 set_dd_corners(dd, dim0, dim1, dim2, bDistMB, &corners);
1868 /* Triclinic stuff */
1869 normal = ddbox->normal;
1873 v_0 = ddbox->v[dim0];
1874 if (ddbox->tric_dir[dim0] && ddbox->tric_dir[dim1])
1876 /* Determine the coupling coefficient for the distances
1877 * to the cell planes along dim0 and dim1 through dim2.
1878 * This is required for correct rounding.
1880 skew_fac_01 = ddbox->v[dim0][dim1 + 1][dim0] * ddbox->v[dim1][dim1 + 1][dim1];
1883 fprintf(debug, "\nskew_fac_01 %f\n", skew_fac_01);
1889 v_1 = ddbox->v[dim1];
1892 zone_cg_range = zones->cg_range;
1893 gmx::ArrayRef<cginfo_mb_t> cginfo_mb = fr->cginfo_mb;
1895 zone_cg_range[0] = 0;
1896 zone_cg_range[1] = dd->ncg_home;
1897 comm->zone_ncg1[0] = dd->ncg_home;
1898 pos_cg = dd->ncg_home;
1900 nat_tot = comm->atomRanges.numHomeAtoms();
1902 for (dim_ind = 0; dim_ind < dd->ndim; dim_ind++)
1904 dim = dd->dim[dim_ind];
1905 cd = &comm->cd[dim_ind];
1907 /* Check if we need to compute triclinic distances along this dim */
1908 bool distanceIsTriclinic = false;
1909 for (int i = 0; i <= dim_ind; i++)
1911 if (ddbox->tric_dir[dd->dim[i]])
1913 distanceIsTriclinic = true;
1917 if (dim >= ddbox->npbcdim && dd->ci[dim] == 0)
1919 /* No pbc in this dimension, the first node should not comm. */
1927 v_d = ddbox->v[dim];
1928 skew_fac2_d = gmx::square(ddbox->skew_fac[dim]);
1930 cd->receiveInPlace = true;
1931 for (int p = 0; p < cd->numPulses(); p++)
1933 /* Only atoms communicated in the first pulse are used
1934 * for multi-body bonded interactions or for bBondComm.
1936 bDistBonded = ((bDistMB || bDist2B) && p == 0);
1938 gmx_domdec_ind_t* ind = &cd->ind[p];
1940 /* Thread 0 writes in the global index array */
1942 clearCommSetupData(&comm->dth[0]);
1944 for (zone = 0; zone < nzone_send; zone++)
1946 if (dim_ind > 0 && distanceIsTriclinic)
1948 /* Determine slightly more optimized skew_fac's
1950 * This reduces the number of communicated atoms
1951 * by about 10% for 3D DD of rhombic dodecahedra.
1953 for (dimd = 0; dimd < dim; dimd++)
1955 sf2_round[dimd] = 1;
1956 if (ddbox->tric_dir[dimd])
1958 for (int i = dd->dim[dimd] + 1; i < DIM; i++)
1960 /* If we are shifted in dimension i
1961 * and the cell plane is tilted forward
1962 * in dimension i, skip this coupling.
1964 if (!(zones->shift[nzone + zone][i] && ddbox->v[dimd][i][dimd] >= 0))
1966 sf2_round[dimd] += gmx::square(ddbox->v[dimd][i][dimd]);
1969 sf2_round[dimd] = 1 / sf2_round[dimd];
1974 zonei = zone_perm[dim_ind][zone];
1977 /* Here we permutate the zones to obtain a convenient order
1978 * for neighbor searching
1980 cg0 = zone_cg_range[zonei];
1981 cg1 = zone_cg_range[zonei + 1];
1985 /* Look only at the cg's received in the previous grid pulse
1987 cg1 = zone_cg_range[nzone + zone + 1];
1988 cg0 = cg1 - cd->ind[p - 1].nrecv[zone];
1991 const int numThreads = gmx::ssize(comm->dth);
1992 #pragma omp parallel for num_threads(numThreads) schedule(static)
1993 for (int th = 0; th < numThreads; th++)
1997 dd_comm_setup_work_t& work = comm->dth[th];
1999 /* Retain data accumulated into buffers of thread 0 */
2002 clearCommSetupData(&work);
2005 int cg0_th = cg0 + ((cg1 - cg0) * th) / numThreads;
2006 int cg1_th = cg0 + ((cg1 - cg0) * (th + 1)) / numThreads;
2008 /* Get the cg's for this pulse in this zone */
2009 get_zone_pulse_cgs(dd, zonei, zone, cg0_th, cg1_th, dd->globalAtomGroupIndices,
2010 dim, dim_ind, dim0, dim1, dim2, r_comm2, r_bcomm2, box,
2011 distanceIsTriclinic, normal, skew_fac2_d, skew_fac_01,
2012 v_d, v_0, v_1, &corners, sf2_round, bDistBonded, bBondComm,
2013 bDist2B, bDistMB, state->x.rvec_array(), fr->cginfo,
2014 th == 0 ? &ind->index : &work.localAtomGroupBuffer, &work);
2016 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
2019 std::vector<int>& atomGroups = comm->dth[0].atomGroupBuffer;
2020 std::vector<gmx::RVec>& positions = comm->dth[0].positionBuffer;
2021 ind->nsend[zone] = comm->dth[0].nsend_zone;
2022 /* Append data of threads>=1 to the communication buffers */
2023 for (int th = 1; th < numThreads; th++)
2025 const dd_comm_setup_work_t& dth = comm->dth[th];
2027 ind->index.insert(ind->index.end(), dth.localAtomGroupBuffer.begin(),
2028 dth.localAtomGroupBuffer.end());
2029 atomGroups.insert(atomGroups.end(), dth.atomGroupBuffer.begin(),
2030 dth.atomGroupBuffer.end());
2031 positions.insert(positions.end(), dth.positionBuffer.begin(),
2032 dth.positionBuffer.end());
2033 comm->dth[0].nat += dth.nat;
2034 ind->nsend[zone] += dth.nsend_zone;
2037 /* Clear the counts in case we do not have pbc */
2038 for (zone = nzone_send; zone < nzone; zone++)
2040 ind->nsend[zone] = 0;
2042 ind->nsend[nzone] = ind->index.size();
2043 ind->nsend[nzone + 1] = comm->dth[0].nat;
2044 /* Communicate the number of cg's and atoms to receive */
2045 ddSendrecv(dd, dim_ind, dddirBackward, ind->nsend, nzone + 2, ind->nrecv, nzone + 2);
2049 /* We can receive in place if only the last zone is not empty */
2050 for (zone = 0; zone < nzone - 1; zone++)
2052 if (ind->nrecv[zone] > 0)
2054 cd->receiveInPlace = false;
2059 int receiveBufferSize = 0;
2060 if (!cd->receiveInPlace)
2062 receiveBufferSize = ind->nrecv[nzone];
2064 /* These buffer are actually only needed with in-place */
2065 DDBufferAccess<int> globalAtomGroupBuffer(comm->intBuffer, receiveBufferSize);
2066 DDBufferAccess<gmx::RVec> rvecBuffer(comm->rvecBuffer, receiveBufferSize);
2068 dd_comm_setup_work_t& work = comm->dth[0];
2070 /* Make space for the global cg indices */
2071 int numAtomGroupsNew = pos_cg + ind->nrecv[nzone];
2072 dd->globalAtomGroupIndices.resize(numAtomGroupsNew);
2073 /* Communicate the global cg indices */
2074 gmx::ArrayRef<int> integerBufferRef;
2075 if (cd->receiveInPlace)
2077 integerBufferRef = gmx::arrayRefFromArray(
2078 dd->globalAtomGroupIndices.data() + pos_cg, ind->nrecv[nzone]);
2082 integerBufferRef = globalAtomGroupBuffer.buffer;
2084 ddSendrecv<int>(dd, dim_ind, dddirBackward, work.atomGroupBuffer, integerBufferRef);
2086 /* Make space for cg_cm */
2087 dd_resize_atominfo_and_state(fr, state, pos_cg + ind->nrecv[nzone]);
2089 /* Communicate the coordinates */
2090 gmx::ArrayRef<gmx::RVec> rvecBufferRef;
2091 if (cd->receiveInPlace)
2093 rvecBufferRef = gmx::makeArrayRef(state->x).subArray(pos_cg, ind->nrecv[nzone]);
2097 rvecBufferRef = rvecBuffer.buffer;
2099 ddSendrecv<gmx::RVec>(dd, dim_ind, dddirBackward, work.positionBuffer, rvecBufferRef);
2101 /* Make the charge group index */
2102 if (cd->receiveInPlace)
2104 zone = (p == 0 ? 0 : nzone - 1);
2105 while (zone < nzone)
2107 for (int i = 0; i < ind->nrecv[zone]; i++)
2109 int globalAtomIndex = dd->globalAtomGroupIndices[pos_cg];
2110 fr->cginfo[pos_cg] = ddcginfo(cginfo_mb, globalAtomIndex);
2115 comm->zone_ncg1[nzone + zone] = ind->nrecv[zone];
2118 zone_cg_range[nzone + zone] = pos_cg;
2123 /* This part of the code is never executed with bBondComm. */
2124 merge_cg_buffers(nzone, cd, p, zone_cg_range, dd->globalAtomGroupIndices,
2125 integerBufferRef.data(), state->x, rvecBufferRef, fr->cginfo_mb,
2127 pos_cg += ind->nrecv[nzone];
2129 nat_tot += ind->nrecv[nzone + 1];
2131 if (!cd->receiveInPlace)
2133 /* Store the atom block for easy copying of communication buffers */
2134 make_cell2at_index(cd, nzone, zone_cg_range[nzone]);
2139 comm->atomRanges.setEnd(DDAtomRanges::Type::Zones, nat_tot);
2143 /* We don't need to update cginfo, since that was alrady done above.
2144 * So we pass NULL for the forcerec.
2146 dd_set_cginfo(dd->globalAtomGroupIndices, dd->ncg_home, dd->globalAtomGroupIndices.size(), nullptr);
2151 fprintf(debug, "Finished setting up DD communication, zones:");
2152 for (c = 0; c < zones->n; c++)
2154 fprintf(debug, " %d", zones->cg_range[c + 1] - zones->cg_range[c]);
2156 fprintf(debug, "\n");
2160 //! Set boundaries for the charge group range.
2161 static void set_cg_boundaries(gmx_domdec_zones_t* zones)
2163 for (auto& iZone : zones->iZones)
2165 iZone.iAtomRange = gmx::Range<int>(0, zones->cg_range[iZone.iZoneIndex + 1]);
2166 iZone.jAtomRange = gmx::Range<int>(zones->cg_range[iZone.jZoneRange.begin()],
2167 zones->cg_range[iZone.jZoneRange.end()]);
2171 /*! \brief Set zone dimensions for zones \p zone_start to \p zone_end-1
2173 * Also sets the atom density for the home zone when \p zone_start=0.
2174 * For this \p numMovedChargeGroupsInHomeZone needs to be passed to tell
2175 * how many charge groups will move but are still part of the current range.
2176 * \todo When converting domdec to use proper classes, all these variables
2177 * should be private and a method should return the correct count
2178 * depending on an internal state.
2180 * \param[in,out] dd The domain decomposition struct
2181 * \param[in] box The box
2182 * \param[in] ddbox The domain decomposition box struct
2183 * \param[in] zone_start The start of the zone range to set sizes for
2184 * \param[in] zone_end The end of the zone range to set sizes for
2185 * \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
2187 static void set_zones_size(gmx_domdec_t* dd,
2189 const gmx_ddbox_t* ddbox,
2192 int numMovedChargeGroupsInHomeZone)
2194 gmx_domdec_comm_t* comm;
2195 gmx_domdec_zones_t* zones;
2204 zones = &comm->zones;
2206 /* Do we need to determine extra distances for multi-body bondeds? */
2207 bDistMB = (comm->systemInfo.haveInterDomainMultiBodyBondeds && isDlbOn(dd->comm) && dd->ndim > 1);
2209 for (z = zone_start; z < zone_end; z++)
2211 /* Copy cell limits to zone limits.
2212 * Valid for non-DD dims and non-shifted dims.
2214 copy_rvec(comm->cell_x0, zones->size[z].x0);
2215 copy_rvec(comm->cell_x1, zones->size[z].x1);
2218 for (d = 0; d < dd->ndim; d++)
2222 for (z = 0; z < zones->n; z++)
2224 /* With a staggered grid we have different sizes
2225 * for non-shifted dimensions.
2227 if (isDlbOn(dd->comm) && zones->shift[z][dim] == 0)
2231 zones->size[z].x0[dim] = comm->zone_d1[zones->shift[z][dd->dim[d - 1]]].min0;
2232 zones->size[z].x1[dim] = comm->zone_d1[zones->shift[z][dd->dim[d - 1]]].max1;
2236 zones->size[z].x0[dim] =
2237 comm->zone_d2[zones->shift[z][dd->dim[d - 2]]][zones->shift[z][dd->dim[d - 1]]]
2239 zones->size[z].x1[dim] =
2240 comm->zone_d2[zones->shift[z][dd->dim[d - 2]]][zones->shift[z][dd->dim[d - 1]]]
2246 rcs = comm->systemInfo.cutoff;
2247 rcmbs = comm->cutoff_mbody;
2248 if (ddbox->tric_dir[dim])
2250 rcs /= ddbox->skew_fac[dim];
2251 rcmbs /= ddbox->skew_fac[dim];
2254 /* Set the lower limit for the shifted zone dimensions */
2255 for (z = zone_start; z < zone_end; z++)
2257 if (zones->shift[z][dim] > 0)
2260 if (!isDlbOn(dd->comm) || d == 0)
2262 zones->size[z].x0[dim] = comm->cell_x1[dim];
2263 zones->size[z].x1[dim] = comm->cell_x1[dim] + rcs;
2267 /* Here we take the lower limit of the zone from
2268 * the lowest domain of the zone below.
2272 zones->size[z].x0[dim] = comm->zone_d1[zones->shift[z][dd->dim[d - 1]]].min1;
2278 zones->size[z].x0[dim] = zones->size[zone_perm[2][z - 4]].x0[dim];
2282 zones->size[z].x0[dim] =
2283 comm->zone_d2[zones->shift[z][dd->dim[d - 2]]][zones->shift[z][dd->dim[d - 1]]]
2287 /* A temporary limit, is updated below */
2288 zones->size[z].x1[dim] = zones->size[z].x0[dim];
2292 for (size_t zi = 0; zi < zones->iZones.size(); zi++)
2294 if (zones->shift[zi][dim] == 0)
2296 /* This takes the whole zone into account.
2297 * With multiple pulses this will lead
2298 * to a larger zone then strictly necessary.
2300 zones->size[z].x1[dim] = std::max(zones->size[z].x1[dim],
2301 zones->size[zi].x1[dim] + rcmbs);
2309 /* Loop over the i-zones to set the upper limit of each
2312 for (const auto& iZone : zones->iZones)
2314 const int zi = iZone.iZoneIndex;
2315 if (zones->shift[zi][dim] == 0)
2317 /* We should only use zones up to zone_end */
2318 const auto& jZoneRangeFull = iZone.jZoneRange;
2319 if (zone_end <= *jZoneRangeFull.begin())
2323 const gmx::Range<int> jZoneRange(*jZoneRangeFull.begin(),
2324 std::min(*jZoneRangeFull.end(), zone_end));
2325 for (int jZone : jZoneRange)
2327 if (zones->shift[jZone][dim] > 0)
2329 zones->size[jZone].x1[dim] =
2330 std::max(zones->size[jZone].x1[dim], zones->size[zi].x1[dim] + rcs);
2337 for (z = zone_start; z < zone_end; z++)
2339 /* Initialization only required to keep the compiler happy */
2340 rvec corner_min = { 0, 0, 0 }, corner_max = { 0, 0, 0 }, corner;
2343 /* To determine the bounding box for a zone we need to find
2344 * the extreme corners of 4, 2 or 1 corners.
2346 nc = 1 << (ddbox->nboundeddim - 1);
2348 for (c = 0; c < nc; c++)
2350 /* Set up a zone corner at x=0, ignoring trilinic couplings */
2354 corner[YY] = zones->size[z].x0[YY];
2358 corner[YY] = zones->size[z].x1[YY];
2362 corner[ZZ] = zones->size[z].x0[ZZ];
2366 corner[ZZ] = zones->size[z].x1[ZZ];
2368 if (dd->ndim == 1 && dd->dim[0] < ZZ && ZZ < dd->unitCellInfo.npbcdim
2369 && box[ZZ][1 - dd->dim[0]] != 0)
2371 /* With 1D domain decomposition the cg's are not in
2372 * the triclinic box, but triclinic x-y and rectangular y/x-z.
2373 * Shift the corner of the z-vector back to along the box
2374 * vector of dimension d, so it will later end up at 0 along d.
2375 * This can affect the location of this corner along dd->dim[0]
2376 * through the matrix operation below if box[d][dd->dim[0]]!=0.
2378 int d = 1 - dd->dim[0];
2380 corner[d] -= corner[ZZ] * box[ZZ][d] / box[ZZ][ZZ];
2382 /* Apply the triclinic couplings */
2383 for (i = YY; i < ddbox->npbcdim && i < DIM; i++)
2385 for (j = XX; j < i; j++)
2387 corner[j] += corner[i] * box[i][j] / box[i][i];
2392 copy_rvec(corner, corner_min);
2393 copy_rvec(corner, corner_max);
2397 for (i = 0; i < DIM; i++)
2399 corner_min[i] = std::min(corner_min[i], corner[i]);
2400 corner_max[i] = std::max(corner_max[i], corner[i]);
2404 /* Copy the extreme cornes without offset along x */
2405 for (i = 0; i < DIM; i++)
2407 zones->size[z].bb_x0[i] = corner_min[i];
2408 zones->size[z].bb_x1[i] = corner_max[i];
2410 /* Add the offset along x */
2411 zones->size[z].bb_x0[XX] += zones->size[z].x0[XX];
2412 zones->size[z].bb_x1[XX] += zones->size[z].x1[XX];
2415 if (zone_start == 0)
2418 for (dim = 0; dim < DIM; dim++)
2420 vol *= zones->size[0].x1[dim] - zones->size[0].x0[dim];
2423 (zones->cg_range[1] - zones->cg_range[0] - numMovedChargeGroupsInHomeZone) / vol;
2428 for (z = zone_start; z < zone_end; z++)
2430 fprintf(debug, "zone %d %6.3f - %6.3f %6.3f - %6.3f %6.3f - %6.3f\n", z,
2431 zones->size[z].x0[XX], zones->size[z].x1[XX], zones->size[z].x0[YY],
2432 zones->size[z].x1[YY], zones->size[z].x0[ZZ], zones->size[z].x1[ZZ]);
2433 fprintf(debug, "zone %d bb %6.3f - %6.3f %6.3f - %6.3f %6.3f - %6.3f\n", z,
2434 zones->size[z].bb_x0[XX], zones->size[z].bb_x1[XX], zones->size[z].bb_x0[YY],
2435 zones->size[z].bb_x1[YY], zones->size[z].bb_x0[ZZ], zones->size[z].bb_x1[ZZ]);
2440 /*! \brief Order data in \p dataToSort according to \p sort
2442 * Note: both buffers should have at least \p sort.size() elements.
2444 template<typename T>
2445 static void orderVector(gmx::ArrayRef<const gmx_cgsort_t> sort,
2446 gmx::ArrayRef<T> dataToSort,
2447 gmx::ArrayRef<T> sortBuffer)
2449 GMX_ASSERT(dataToSort.size() >= sort.size(), "The vector needs to be sufficiently large");
2450 GMX_ASSERT(sortBuffer.size() >= sort.size(),
2451 "The sorting buffer needs to be sufficiently large");
2453 /* Order the data into the temporary buffer */
2455 for (const gmx_cgsort_t& entry : sort)
2457 sortBuffer[i++] = dataToSort[entry.ind];
2460 /* Copy back to the original array */
2461 std::copy(sortBuffer.begin(), sortBuffer.begin() + sort.size(), dataToSort.begin());
2464 /*! \brief Order data in \p dataToSort according to \p sort
2466 * Note: \p vectorToSort should have at least \p sort.size() elements,
2467 * \p workVector is resized when it is too small.
2469 template<typename T>
2470 static void orderVector(gmx::ArrayRef<const gmx_cgsort_t> sort,
2471 gmx::ArrayRef<T> vectorToSort,
2472 std::vector<T>* workVector)
2474 if (gmx::index(workVector->size()) < sort.ssize())
2476 workVector->resize(sort.size());
2478 orderVector<T>(sort, vectorToSort, *workVector);
2481 //! Returns the sorting order for atoms based on the nbnxn grid order in sort
2482 static void dd_sort_order_nbnxn(const t_forcerec* fr, std::vector<gmx_cgsort_t>* sort)
2484 gmx::ArrayRef<const int> atomOrder = fr->nbv->getLocalAtomOrder();
2486 /* Using push_back() instead of this resize results in much slower code */
2487 sort->resize(atomOrder.size());
2488 gmx::ArrayRef<gmx_cgsort_t> buffer = *sort;
2489 size_t numSorted = 0;
2490 for (int i : atomOrder)
2494 /* The values of nsc and ind_gl are not used in this case */
2495 buffer[numSorted++].ind = i;
2498 sort->resize(numSorted);
2501 //! Returns the sorting state for DD.
2502 static void dd_sort_state(gmx_domdec_t* dd, t_forcerec* fr, t_state* state)
2504 gmx_domdec_sort_t* sort = dd->comm->sort.get();
2506 dd_sort_order_nbnxn(fr, &sort->sorted);
2508 /* We alloc with the old size, since cgindex is still old */
2509 DDBufferAccess<gmx::RVec> rvecBuffer(dd->comm->rvecBuffer, dd->ncg_home);
2511 /* Set the new home atom/charge group count */
2512 dd->ncg_home = sort->sorted.size();
2515 fprintf(debug, "Set the new home atom count to %d\n", dd->ncg_home);
2518 /* Reorder the state */
2519 gmx::ArrayRef<const gmx_cgsort_t> cgsort = sort->sorted;
2520 GMX_RELEASE_ASSERT(cgsort.ssize() == dd->ncg_home, "We should sort all the home atom groups");
2522 if (state->flags & (1 << estX))
2524 orderVector(cgsort, makeArrayRef(state->x), rvecBuffer.buffer);
2526 if (state->flags & (1 << estV))
2528 orderVector(cgsort, makeArrayRef(state->v), rvecBuffer.buffer);
2530 if (state->flags & (1 << estCGP))
2532 orderVector(cgsort, makeArrayRef(state->cg_p), rvecBuffer.buffer);
2535 /* Reorder the global cg index */
2536 orderVector<int>(cgsort, dd->globalAtomGroupIndices, &sort->intBuffer);
2537 /* Reorder the cginfo */
2538 orderVector<int>(cgsort, fr->cginfo, &sort->intBuffer);
2539 /* Set the home atom number */
2540 dd->comm->atomRanges.setEnd(DDAtomRanges::Type::Home, dd->ncg_home);
2542 /* The atoms are now exactly in grid order, update the grid order */
2543 fr->nbv->setLocalAtomOrder();
2546 //! Accumulates load statistics.
2547 static void add_dd_statistics(gmx_domdec_t* dd)
2549 gmx_domdec_comm_t* comm = dd->comm;
2551 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
2553 auto range = static_cast<DDAtomRanges::Type>(i);
2554 comm->sum_nat[i] += comm->atomRanges.end(range) - comm->atomRanges.start(range);
2559 void reset_dd_statistics_counters(gmx_domdec_t* dd)
2561 gmx_domdec_comm_t* comm = dd->comm;
2563 /* Reset all the statistics and counters for total run counting */
2564 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
2566 comm->sum_nat[i] = 0;
2570 comm->load_step = 0;
2573 clear_ivec(comm->load_lim);
2578 void print_dd_statistics(const t_commrec* cr, const t_inputrec* ir, FILE* fplog)
2580 gmx_domdec_comm_t* comm = cr->dd->comm;
2582 const int numRanges = static_cast<int>(DDAtomRanges::Type::Number);
2583 gmx_sumd(numRanges, comm->sum_nat, cr);
2585 if (fplog == nullptr)
2590 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");
2592 for (int i = static_cast<int>(DDAtomRanges::Type::Zones); i < numRanges; i++)
2594 auto range = static_cast<DDAtomRanges::Type>(i);
2595 double av = comm->sum_nat[i] / comm->ndecomp;
2598 case DDAtomRanges::Type::Zones:
2599 fprintf(fplog, " av. #atoms communicated per step for force: %d x %.1f\n", 2, av);
2601 case DDAtomRanges::Type::Vsites:
2602 if (cr->dd->vsite_comm)
2604 fprintf(fplog, " av. #atoms communicated per step for vsites: %d x %.1f\n",
2605 (EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD) ? 3 : 2, av);
2608 case DDAtomRanges::Type::Constraints:
2609 if (cr->dd->constraint_comm)
2611 fprintf(fplog, " av. #atoms communicated per step for LINCS: %d x %.1f\n",
2612 1 + ir->nLincsIter, av);
2615 default: gmx_incons(" Unknown type for DD statistics");
2618 fprintf(fplog, "\n");
2620 if (comm->ddSettings.recordLoad && EI_DYNAMICS(ir->eI))
2622 print_dd_load_av(fplog, cr->dd);
2626 //!\brief TODO Remove fplog when group scheme and charge groups are gone
2627 void dd_partition_system(FILE* fplog,
2628 const gmx::MDLogger& mdlog,
2630 const t_commrec* cr,
2631 gmx_bool bMasterState,
2633 t_state* state_global,
2634 const gmx_mtop_t& top_global,
2635 const t_inputrec* ir,
2636 gmx::ImdSession* imdSession,
2638 t_state* state_local,
2639 PaddedHostVector<gmx::RVec>* f,
2640 gmx::MDAtoms* mdAtoms,
2641 gmx_localtop_t* top_local,
2643 gmx::VirtualSitesHandler* vsite,
2644 gmx::Constraints* constr,
2646 gmx_wallcycle* wcycle,
2650 gmx_domdec_comm_t* comm;
2651 gmx_ddbox_t ddbox = { 0 };
2652 int64_t step_pcoupl;
2653 rvec cell_ns_x0, cell_ns_x1;
2654 int ncgindex_set, ncg_moved, nat_f_novirsum;
2655 gmx_bool bBoxChanged, bNStGlobalComm, bDoDLB, bCheckWhetherToTurnDlbOn, bLogLoad;
2661 wallcycle_start(wcycle, ewcDOMDEC);
2666 // TODO if the update code becomes accessible here, use
2667 // upd->deform for this logic.
2668 bBoxChanged = (bMasterState || inputrecDeform(ir));
2669 if (ir->epc != epcNO)
2671 /* With nstpcouple > 1 pressure coupling happens.
2672 * one step after calculating the pressure.
2673 * Box scaling happens at the end of the MD step,
2674 * after the DD partitioning.
2675 * We therefore have to do DLB in the first partitioning
2676 * after an MD step where P-coupling occurred.
2677 * We need to determine the last step in which p-coupling occurred.
2678 * MRS -- need to validate this for vv?
2680 int n = ir->nstpcouple;
2683 step_pcoupl = step - 1;
2687 step_pcoupl = ((step - 1) / n) * n + 1;
2689 if (step_pcoupl >= comm->partition_step)
2695 bNStGlobalComm = (step % nstglobalcomm == 0);
2703 /* Should we do dynamic load balacing this step?
2704 * Since it requires (possibly expensive) global communication,
2705 * we might want to do DLB less frequently.
2707 if (bBoxChanged || ir->epc != epcNO)
2709 bDoDLB = bBoxChanged;
2713 bDoDLB = bNStGlobalComm;
2717 /* Check if we have recorded loads on the nodes */
2718 if (comm->ddSettings.recordLoad && dd_load_count(comm) > 0)
2720 bCheckWhetherToTurnDlbOn = dd_dlb_get_should_check_whether_to_turn_dlb_on(dd);
2722 /* Print load every nstlog, first and last step to the log file */
2723 bLogLoad = ((ir->nstlog > 0 && step % ir->nstlog == 0) || comm->n_load_collect == 0
2724 || (ir->nsteps >= 0 && (step + ir->nstlist > ir->init_step + ir->nsteps)));
2726 /* Avoid extra communication due to verbose screen output
2727 * when nstglobalcomm is set.
2729 if (bDoDLB || bLogLoad || bCheckWhetherToTurnDlbOn
2730 || (bVerbose && (ir->nstlist == 0 || nstglobalcomm <= ir->nstlist)))
2732 get_load_distribution(dd, wcycle);
2737 GMX_LOG(mdlog.info).asParagraph().appendText(dd_print_load(dd, step - 1));
2741 dd_print_load_verbose(dd);
2744 comm->n_load_collect++;
2750 /* Add the measured cycles to the running average */
2751 const float averageFactor = 0.1F;
2752 comm->cyclesPerStepDlbExpAverage =
2753 (1 - averageFactor) * comm->cyclesPerStepDlbExpAverage
2754 + averageFactor * comm->cycl[ddCyclStep] / comm->cycl_n[ddCyclStep];
2756 if (comm->dlbState == DlbState::onCanTurnOff
2757 && dd->comm->n_load_have % c_checkTurnDlbOffInterval == c_checkTurnDlbOffInterval - 1)
2759 gmx_bool turnOffDlb;
2762 /* If the running averaged cycles with DLB are more
2763 * than before we turned on DLB, turn off DLB.
2764 * We will again run and check the cycles without DLB
2765 * and we can then decide if to turn off DLB forever.
2767 turnOffDlb = (comm->cyclesPerStepDlbExpAverage > comm->cyclesPerStepBeforeDLB);
2769 dd_bcast(dd, sizeof(turnOffDlb), &turnOffDlb);
2772 /* To turn off DLB, we need to redistribute the atoms */
2773 dd_collect_state(dd, state_local, state_global);
2774 bMasterState = TRUE;
2775 turn_off_dlb(mdlog, dd, step);
2779 else if (bCheckWhetherToTurnDlbOn)
2781 gmx_bool turnOffDlbForever = FALSE;
2782 gmx_bool turnOnDlb = FALSE;
2784 /* Since the timings are node dependent, the master decides */
2787 /* If we recently turned off DLB, we want to check if
2788 * performance is better without DLB. We want to do this
2789 * ASAP to minimize the chance that external factors
2790 * slowed down the DLB step are gone here and we
2791 * incorrectly conclude that DLB was causing the slowdown.
2792 * So we measure one nstlist block, no running average.
2794 if (comm->haveTurnedOffDlb
2795 && comm->cycl[ddCyclStep] / comm->cycl_n[ddCyclStep] < comm->cyclesPerStepDlbExpAverage)
2797 /* After turning off DLB we ran nstlist steps in fewer
2798 * cycles than with DLB. This likely means that DLB
2799 * in not benefical, but this could be due to a one
2800 * time unlucky fluctuation, so we require two such
2801 * observations in close succession to turn off DLB
2804 if (comm->dlbSlowerPartitioningCount > 0
2805 && dd->ddp_count < comm->dlbSlowerPartitioningCount + 10 * c_checkTurnDlbOnInterval)
2807 turnOffDlbForever = TRUE;
2809 comm->haveTurnedOffDlb = false;
2810 /* Register when we last measured DLB slowdown */
2811 comm->dlbSlowerPartitioningCount = dd->ddp_count;
2815 /* Here we check if the max PME rank load is more than 0.98
2816 * the max PP force load. If so, PP DLB will not help,
2817 * since we are (almost) limited by PME. Furthermore,
2818 * DLB will cause a significant extra x/f redistribution
2819 * cost on the PME ranks, which will then surely result
2820 * in lower total performance.
2822 if (comm->ddRankSetup.usePmeOnlyRanks && dd_pme_f_ratio(dd) > 1 - DD_PERF_LOSS_DLB_ON)
2828 turnOnDlb = (dd_force_imb_perf_loss(dd) >= DD_PERF_LOSS_DLB_ON);
2834 gmx_bool turnOffDlbForever;
2836 } bools{ turnOffDlbForever, turnOnDlb };
2837 dd_bcast(dd, sizeof(bools), &bools);
2838 if (bools.turnOffDlbForever)
2840 turn_off_dlb_forever(mdlog, dd, step);
2842 else if (bools.turnOnDlb)
2844 turn_on_dlb(mdlog, dd, step);
2849 comm->n_load_have++;
2855 /* Clear the old state */
2856 clearDDStateIndices(dd, false);
2859 auto xGlobal = positionsFromStatePointer(state_global);
2861 set_ddbox(*dd, true, DDMASTER(dd) ? state_global->box : nullptr, true, xGlobal, &ddbox);
2863 distributeState(mdlog, dd, top_global, state_global, ddbox, state_local);
2865 /* Ensure that we have space for the new distribution */
2866 dd_resize_atominfo_and_state(fr, state_local, dd->ncg_home);
2868 inc_nrnb(nrnb, eNR_CGCM, comm->atomRanges.numHomeAtoms());
2870 dd_set_cginfo(dd->globalAtomGroupIndices, 0, dd->ncg_home, fr);
2872 else if (state_local->ddp_count != dd->ddp_count)
2874 if (state_local->ddp_count > dd->ddp_count)
2877 "Internal inconsistency state_local->ddp_count (%d) > dd->ddp_count (%" PRId64
2879 state_local->ddp_count, dd->ddp_count);
2882 if (state_local->ddp_count_cg_gl != state_local->ddp_count)
2885 "Internal inconsistency state_local->ddp_count_cg_gl (%d) != "
2886 "state_local->ddp_count (%d)",
2887 state_local->ddp_count_cg_gl, state_local->ddp_count);
2890 /* Clear the old state */
2891 clearDDStateIndices(dd, false);
2893 /* Restore the atom group indices from state_local */
2894 restoreAtomGroups(dd, state_local);
2895 make_dd_indices(dd, 0);
2896 ncgindex_set = dd->ncg_home;
2898 inc_nrnb(nrnb, eNR_CGCM, comm->atomRanges.numHomeAtoms());
2900 dd_set_cginfo(dd->globalAtomGroupIndices, 0, dd->ncg_home, fr);
2902 set_ddbox(*dd, bMasterState, state_local->box, true, state_local->x, &ddbox);
2904 bRedist = isDlbOn(comm);
2908 /* We have the full state, only redistribute the cgs */
2910 /* Clear the non-home indices */
2911 clearDDStateIndices(dd, true);
2914 /* To avoid global communication, we do not recompute the extent
2915 * of the system for dims without pbc. Therefore we need to copy
2916 * the previously computed values when we do not communicate.
2918 if (!bNStGlobalComm)
2920 copy_rvec(comm->box0, ddbox.box0);
2921 copy_rvec(comm->box_size, ddbox.box_size);
2923 set_ddbox(*dd, bMasterState, state_local->box, bNStGlobalComm, state_local->x, &ddbox);
2928 /* Copy needed for dim's without pbc when avoiding communication */
2929 copy_rvec(ddbox.box0, comm->box0);
2930 copy_rvec(ddbox.box_size, comm->box_size);
2932 set_dd_cell_sizes(dd, &ddbox, dd->unitCellInfo.ddBoxIsDynamic, bMasterState, bDoDLB, step, wcycle);
2934 if (comm->ddSettings.nstDDDumpGrid > 0 && step % comm->ddSettings.nstDDDumpGrid == 0)
2936 write_dd_grid_pdb("dd_grid", step, dd, state_local->box, &ddbox);
2939 if (comm->systemInfo.useUpdateGroups)
2941 comm->updateGroupsCog->addCogs(
2942 gmx::arrayRefFromArray(dd->globalAtomGroupIndices.data(), dd->ncg_home), state_local->x);
2945 /* Check if we should sort the charge groups */
2946 const bool bSortCG = (bMasterState || bRedist);
2948 /* When repartitioning we mark atom groups that will move to neighboring
2949 * DD cells, but we do not move them right away for performance reasons.
2950 * Thus we need to keep track of how many charge groups will move for
2951 * obtaining correct local charge group / atom counts.
2956 wallcycle_sub_start(wcycle, ewcsDD_REDIST);
2958 ncgindex_set = dd->ncg_home;
2959 dd_redistribute_cg(fplog, step, dd, ddbox.tric_dir, state_local, fr, nrnb, &ncg_moved);
2961 GMX_RELEASE_ASSERT(bSortCG, "Sorting is required after redistribution");
2963 if (comm->systemInfo.useUpdateGroups)
2965 comm->updateGroupsCog->addCogs(
2966 gmx::arrayRefFromArray(dd->globalAtomGroupIndices.data(), dd->ncg_home),
2970 wallcycle_sub_stop(wcycle, ewcsDD_REDIST);
2973 // TODO: Integrate this code in the nbnxm module
2974 get_nsgrid_boundaries(ddbox.nboundeddim, state_local->box, dd, &ddbox, &comm->cell_x0,
2975 &comm->cell_x1, dd->ncg_home, as_rvec_array(state_local->x.data()),
2976 cell_ns_x0, cell_ns_x1, &grid_density);
2980 comm_dd_ns_cell_sizes(dd, &ddbox, cell_ns_x0, cell_ns_x1, step);
2985 wallcycle_sub_start(wcycle, ewcsDD_GRID);
2987 /* Sort the state on charge group position.
2988 * This enables exact restarts from this step.
2989 * It also improves performance by about 15% with larger numbers
2990 * of atoms per node.
2993 /* Fill the ns grid with the home cell,
2994 * so we can sort with the indices.
2996 set_zones_ncg_home(dd);
2998 set_zones_size(dd, state_local->box, &ddbox, 0, 1, ncg_moved);
3000 nbnxn_put_on_grid(fr->nbv.get(), state_local->box, 0, comm->zones.size[0].bb_x0,
3001 comm->zones.size[0].bb_x1, comm->updateGroupsCog.get(),
3002 { 0, dd->ncg_home }, comm->zones.dens_zone0, fr->cginfo, state_local->x,
3003 ncg_moved, bRedist ? comm->movedBuffer.data() : nullptr);
3007 fprintf(debug, "Step %s, sorting the %d home charge groups\n", gmx_step_str(step, sbuf),
3010 dd_sort_state(dd, fr, state_local);
3012 /* After sorting and compacting we set the correct size */
3013 state_change_natoms(state_local, comm->atomRanges.numHomeAtoms());
3015 /* Rebuild all the indices */
3019 wallcycle_sub_stop(wcycle, ewcsDD_GRID);
3023 /* With the group scheme the sorting array is part of the DD state,
3024 * but it just got out of sync, so mark as invalid by emptying it.
3026 if (ir->cutoff_scheme == ecutsGROUP)
3028 comm->sort->sorted.clear();
3032 if (comm->systemInfo.useUpdateGroups)
3034 /* The update groups cog's are invalid after sorting
3035 * and need to be cleared before the next partitioning anyhow.
3037 comm->updateGroupsCog->clear();
3040 wallcycle_sub_start(wcycle, ewcsDD_SETUPCOMM);
3042 /* Set the induces for the home atoms */
3043 set_zones_ncg_home(dd);
3044 make_dd_indices(dd, ncgindex_set);
3046 /* Setup up the communication and communicate the coordinates */
3047 setup_dd_communication(dd, state_local->box, &ddbox, fr, state_local);
3049 /* Set the indices for the halo atoms */
3050 make_dd_indices(dd, dd->ncg_home);
3052 /* Set the charge group boundaries for neighbor searching */
3053 set_cg_boundaries(&comm->zones);
3055 if (fr->cutoff_scheme == ecutsVERLET)
3057 /* When bSortCG=true, we have already set the size for zone 0 */
3058 set_zones_size(dd, state_local->box, &ddbox, bSortCG ? 1 : 0, comm->zones.n, 0);
3061 wallcycle_sub_stop(wcycle, ewcsDD_SETUPCOMM);
3064 write_dd_pdb("dd_home",step,"dump",top_global,cr,
3065 -1,state_local->x.rvec_array(),state_local->box);
3068 wallcycle_sub_start(wcycle, ewcsDD_MAKETOP);
3070 /* Extract a local topology from the global topology */
3071 for (int i = 0; i < dd->ndim; i++)
3073 np[dd->dim[i]] = comm->cd[i].numPulses();
3075 dd_make_local_top(dd, &comm->zones, dd->unitCellInfo.npbcdim, state_local->box,
3076 comm->cellsize_min, np, fr, state_local->x.rvec_array(), top_global, top_local);
3078 wallcycle_sub_stop(wcycle, ewcsDD_MAKETOP);
3080 wallcycle_sub_start(wcycle, ewcsDD_MAKECONSTR);
3082 /* Set up the special atom communication */
3083 int n = comm->atomRanges.end(DDAtomRanges::Type::Zones);
3084 for (int i = static_cast<int>(DDAtomRanges::Type::Zones) + 1;
3085 i < static_cast<int>(DDAtomRanges::Type::Number); i++)
3087 auto range = static_cast<DDAtomRanges::Type>(i);
3090 case DDAtomRanges::Type::Vsites:
3091 if (vsite && vsite->numInterUpdategroupVirtualSites())
3093 n = dd_make_local_vsites(dd, n, top_local->idef.il);
3096 case DDAtomRanges::Type::Constraints:
3097 if (dd->comm->systemInfo.haveSplitConstraints || dd->comm->systemInfo.haveSplitSettles)
3099 /* Only for inter-cg constraints we need special code */
3100 n = dd_make_local_constraints(dd, n, &top_global, fr->cginfo.data(), constr,
3101 ir->nProjOrder, top_local->idef.il);
3104 default: gmx_incons("Unknown special atom type setup");
3106 comm->atomRanges.setEnd(range, n);
3109 wallcycle_sub_stop(wcycle, ewcsDD_MAKECONSTR);
3111 wallcycle_sub_start(wcycle, ewcsDD_TOPOTHER);
3113 /* Make space for the extra coordinates for virtual site
3114 * or constraint communication.
3116 state_local->natoms = comm->atomRanges.numAtomsTotal();
3118 state_change_natoms(state_local, state_local->natoms);
3120 if (fr->haveDirectVirialContributions)
3122 if (vsite && vsite->numInterUpdategroupVirtualSites())
3124 nat_f_novirsum = comm->atomRanges.end(DDAtomRanges::Type::Vsites);
3128 if (EEL_FULL(ir->coulombtype) && dd->haveExclusions)
3130 nat_f_novirsum = comm->atomRanges.end(DDAtomRanges::Type::Zones);
3134 nat_f_novirsum = comm->atomRanges.numHomeAtoms();
3143 /* Set the number of atoms required for the force calculation.
3144 * Forces need to be constrained when doing energy
3145 * minimization. For simple simulations we could avoid some
3146 * allocation, zeroing and copying, but this is probably not worth
3147 * the complications and checking.
3149 forcerec_set_ranges(fr, comm->atomRanges.end(DDAtomRanges::Type::Zones),
3150 comm->atomRanges.end(DDAtomRanges::Type::Constraints), nat_f_novirsum);
3152 /* Update atom data for mdatoms and several algorithms */
3153 mdAlgorithmsSetupAtomData(cr, ir, top_global, top_local, fr, f, mdAtoms, constr, vsite, nullptr);
3155 auto mdatoms = mdAtoms->mdatoms();
3156 if (!thisRankHasDuty(cr, DUTY_PME))
3158 /* Send the charges and/or c6/sigmas to our PME only node */
3159 gmx_pme_send_parameters(cr, fr->ic, mdatoms->nChargePerturbed != 0,
3160 mdatoms->nTypePerturbed != 0, mdatoms->chargeA, mdatoms->chargeB,
3161 mdatoms->sqrt_c6A, mdatoms->sqrt_c6B, mdatoms->sigmaA,
3162 mdatoms->sigmaB, dd_pme_maxshift_x(dd), dd_pme_maxshift_y(dd));
3167 /* Update the local pull groups */
3168 dd_make_local_pull_groups(cr, pull_work);
3171 if (dd->atomSets != nullptr)
3173 /* Update the local atom sets */
3174 dd->atomSets->setIndicesInDomainDecomposition(*(dd->ga2la));
3177 /* Update the local atoms to be communicated via the IMD protocol if bIMD is TRUE. */
3178 imdSession->dd_make_local_IMD_atoms(dd);
3180 add_dd_statistics(dd);
3182 /* Make sure we only count the cycles for this DD partitioning */
3183 clear_dd_cycle_counts(dd);
3185 /* Because the order of the atoms might have changed since
3186 * the last vsite construction, we need to communicate the constructing
3187 * atom coordinates again (for spreading the forces this MD step).
3189 dd_move_x_vsites(*dd, state_local->box, state_local->x.rvec_array());
3191 wallcycle_sub_stop(wcycle, ewcsDD_TOPOTHER);
3193 if (comm->ddSettings.nstDDDump > 0 && step % comm->ddSettings.nstDDDump == 0)
3195 dd_move_x(dd, state_local->box, state_local->x, nullWallcycle);
3196 write_dd_pdb("dd_dump", step, "dump", &top_global, cr, -1, state_local->x.rvec_array(),
3200 /* Store the partitioning step */
3201 comm->partition_step = step;
3203 /* Increase the DD partitioning counter */
3205 /* The state currently matches this DD partitioning count, store it */
3206 state_local->ddp_count = dd->ddp_count;
3209 /* The DD master node knows the complete cg distribution,
3210 * store the count so we can possibly skip the cg info communication.
3212 comm->master_cg_ddp_count = (bSortCG ? 0 : dd->ddp_count);
3215 if (comm->ddSettings.DD_debug > 0)
3217 /* Set the env var GMX_DD_DEBUG if you suspect corrupted indices */
3218 check_index_consistency(dd, top_global.natoms, "after partitioning");
3221 wallcycle_stop(wcycle, ewcDOMDEC);
3224 /*! \brief Check whether bonded interactions are missing, if appropriate */
3225 void checkNumberOfBondedInteractions(const gmx::MDLogger& mdlog,
3227 int totalNumberOfBondedInteractions,
3228 const gmx_mtop_t* top_global,
3229 const gmx_localtop_t* top_local,
3230 gmx::ArrayRef<const gmx::RVec> x,
3232 bool* shouldCheckNumberOfBondedInteractions)
3234 if (*shouldCheckNumberOfBondedInteractions)
3236 if (totalNumberOfBondedInteractions != cr->dd->nbonded_global)
3238 dd_print_missing_interactions(mdlog, cr, totalNumberOfBondedInteractions, top_global,
3239 top_local, x, box); // Does not return
3241 *shouldCheckNumberOfBondedInteractions = false;