/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2012,2013, by the GROMACS development team, led by
+ * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
#include <config.h>
#endif
-#include "smalloc.h"
+#include "gromacs/utility/smalloc.h"
+#include "types/commrec.h"
#include "network.h"
#include "calcgrid.h"
#include "pme.h"
real spacing; /* (largest) PME grid spacing */
ivec grid; /* the PME grid dimensions */
real grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
- real ewaldcoeff; /* the Ewald coefficient */
+ real ewaldcoeff_q; /* Electrostatic Ewald coefficient */
+ real ewaldcoeff_lj; /* LJ Ewald coefficient, only for the call to send_switchgrid */
gmx_pme_t pmedata; /* the data structure used in the PME code */
-
int count; /* number of times this setup has been timed */
double cycles; /* the fastest time for this setup in cycles */
} pme_setup_t;
pme_lb->n = 1;
snew(pme_lb->setup, pme_lb->n);
- pme_lb->rcut_vdw = ic->rvdw;
- pme_lb->rcut_coulomb_start = ir->rcoulomb;
- pme_lb->nstcalclr_start = ir->nstcalclr;
-
- pme_lb->cur = 0;
- pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
- pme_lb->setup[0].rlist = ic->rlist;
- pme_lb->setup[0].rlistlong = ic->rlistlong;
- pme_lb->setup[0].nstcalclr = ir->nstcalclr;
- pme_lb->setup[0].grid[XX] = ir->nkx;
- pme_lb->setup[0].grid[YY] = ir->nky;
- pme_lb->setup[0].grid[ZZ] = ir->nkz;
- pme_lb->setup[0].ewaldcoeff = ic->ewaldcoeff;
+ pme_lb->rcut_vdw = ic->rvdw;
+ pme_lb->rcut_coulomb_start = ir->rcoulomb;
+ pme_lb->nstcalclr_start = ir->nstcalclr;
+
+ pme_lb->cur = 0;
+ pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
+ pme_lb->setup[0].rlist = ic->rlist;
+ pme_lb->setup[0].rlistlong = ic->rlistlong;
+ pme_lb->setup[0].nstcalclr = ir->nstcalclr;
+ pme_lb->setup[0].grid[XX] = ir->nkx;
+ pme_lb->setup[0].grid[YY] = ir->nky;
+ pme_lb->setup[0].grid[ZZ] = ir->nkz;
+ pme_lb->setup[0].ewaldcoeff_q = ic->ewaldcoeff_q;
+ pme_lb->setup[0].ewaldcoeff_lj = ic->ewaldcoeff_lj;
pme_lb->setup[0].pmedata = pmedata;
*pme_lb_p = pme_lb;
}
-static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
- int pme_order,
+static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
+ int pme_order,
const gmx_domdec_t *dd)
{
pme_setup_t *set;
while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
set->rcut_coulomb = pme_lb->cut_spacing*sp;
+ if (set->rcut_coulomb < pme_lb->rcut_coulomb_start)
+ {
+ /* This is unlikely, but can happen when e.g. continuing from
+ * a checkpoint after equilibration where the box shrank a lot.
+ * We want to avoid rcoulomb getting smaller than rvdw
+ * and there might be more issues with decreasing rcoulomb.
+ */
+ set->rcut_coulomb = pme_lb->rcut_coulomb_start;
+ }
if (pme_lb->cutoff_scheme == ecutsVERLET)
{
set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
}
/* The Ewald coefficient is inversly proportional to the cut-off */
- set->ewaldcoeff =
- pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
+ set->ewaldcoeff_q =
+ pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
+ /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
+ set->ewaldcoeff_lj =
+ pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
set->count = 0;
set->cycles = 0;
}
static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
- gmx_large_int_t step,
+ gmx_int64_t step,
pme_load_balancing_t pme_lb)
{
char buf[STRLEN], sbuf[22];
interaction_const_t *ic,
nonbonded_verlet_t *nbv,
gmx_pme_t *pmedata,
- gmx_large_int_t step)
+ gmx_int64_t step)
{
gmx_bool OK;
pme_setup_t *set;
set = &pme_lb->setup[pme_lb->cur];
- ic->rcoulomb = set->rcut_coulomb;
- ic->rlist = set->rlist;
- ic->rlistlong = set->rlistlong;
- ir->nstcalclr = set->nstcalclr;
- ic->ewaldcoeff = set->ewaldcoeff;
+ ic->rcoulomb = set->rcut_coulomb;
+ ic->rlist = set->rlist;
+ ic->rlistlong = set->rlistlong;
+ ir->nstcalclr = set->nstcalclr;
+ ic->ewaldcoeff_q = set->ewaldcoeff_q;
+ /* TODO: centralize the code that sets the potentials shifts */
+ if (ic->coulomb_modifier == eintmodPOTSHIFT)
+ {
+ ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
+ }
+ if (EVDW_PME(ic->vdwtype))
+ {
+ /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
+ ic->rvdw = set->rcut_coulomb;
+ ic->ewaldcoeff_lj = set->ewaldcoeff_lj;
+ if (ic->vdw_modifier == eintmodPOTSHIFT)
+ {
+ real crc2;
+
+ ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
+ ic->repulsion_shift.cpot = -pow(ic->rvdw, -12.0);
+ ic->sh_invrc6 = -ic->dispersion_shift.cpot;
+ crc2 = sqr(ic->ewaldcoeff_lj*ic->rvdw);
+ ic->sh_lj_ewald = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
+ }
+ }
bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
if (pme_lb->cutoff_scheme == ecutsVERLET &&
{
gmx_barrier(cr);
}
-#endif /* GMX_THREAD_MPI */
- }
- else
- {
- init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
- rtab);
+#endif /* GMX_THREAD_MPI */
}
- if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
- {
- init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
- rtab);
- }
+ /* Usually we won't need the simple tables with GPUs.
+ * But we do with hybrid acceleration and with free energy.
+ * To avoid bugs, we always re-initialize the simple tables here.
+ */
+ init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
if (cr->duty & DUTY_PME)
{
else
{
/* Tell our PME-only node to switch grid */
- gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff);
+ gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
}
if (debug)
name,
setup->rcut_coulomb, pme_loadbal_rlist(setup),
setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
- setup->spacing, 1/setup->ewaldcoeff);
+ setup->spacing, 1/setup->ewaldcoeff_q);
}
static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
{
md_print_warn(cr, fplog,
"NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
- " For better performance use (more) PME nodes (mdrun -npme),\n"
- " or in case you are beyond the scaling limit, use less nodes in total.\n");
+ " For better performance, use (more) PME ranks (mdrun -npme),\n"
+ " or if you are beyond the scaling limit, use fewer total ranks (or nodes).\n");
}
else
{