static void low_set_ddbox(t_inputrec *ir, ivec *dd_nc, matrix box,
gmx_bool bCalcUnboundedSize, int ncg, t_block *cgs, rvec *x,
t_commrec *cr_sum,
gmx_ddbox_t *ddbox)
{
rvec av, stddev;
real b0, b1;
int d;
ddbox->npbcdim = ePBC2npbcdim(ir->ePBC);
ddbox->nboundeddim = inputrec2nboundeddim(ir);
for (d = 0; d < ddbox->nboundeddim; d++)
{
ddbox->box0[d] = 0;
ddbox->box_size[d] = box[d][d];
}
if (ddbox->nboundeddim < DIM && bCalcUnboundedSize)
{
calc_cgcm_av_stddev(cgs, ncg, x, av, stddev, cr_sum);
/* GRID_STDDEV_FAC * stddev
* gives a uniform load for a rectangular block of cg's.
* For a sphere it is not a bad approximation for 4x1x1 up to 4x2x2.
*/
for (d = ddbox->nboundeddim; d < DIM; d++)
{
b0 = av[d] - GRID_STDDEV_FAC*stddev[d];
b1 = av[d] + GRID_STDDEV_FAC*stddev[d];
if (debug)
{
fprintf(debug, "Setting global DD grid boundaries to %f - %f\n",
b0, b1);
}
ddbox->box0[d] = b0;
ddbox->box_size[d] = b1 - b0;
}
}
set_tric_dir(dd_nc, ddbox, box);
}
/* Try to increase nstlist when using the Verlet cut-off scheme */
static void increase_nstlist(FILE *fp, t_commrec *cr,
t_inputrec *ir, int nstlist_cmdline,
const gmx_mtop_t *mtop, matrix box,
gmx_bool bGPU)
{
float listfac_ok, listfac_max;
int nstlist_orig, nstlist_prev;
verletbuf_list_setup_t ls;
real rlistWithReferenceNstlist, rlist_inc, rlist_ok, rlist_max;
real rlist_new, rlist_prev;
size_t nstlist_ind = 0;
t_state state_tmp;
gmx_bool bBox, bDD, bCont;
const char *nstl_gpu = "\nFor optimal performance with a GPU nstlist (now %d) should be larger.\nThe optimum depends on your CPU and GPU resources.\nYou might want to try several nstlist values.\n";
const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
const char *box_err = "Can not increase nstlist because the box is too small";
const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
char buf[STRLEN];
const float oneThird = 1.0f / 3.0f;
if (nstlist_cmdline <= 0)
{
if (ir->nstlist == 1)
{
/* The user probably set nstlist=1 for a reason,
* don't mess with the settings.
*/
return;
}
if (fp != NULL && bGPU && ir->nstlist < nstlist_try[0])
{
fprintf(fp, nstl_gpu, ir->nstlist);
}
nstlist_ind = 0;
while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
{
nstlist_ind++;
}
if (nstlist_ind == NNSTL)
{
/* There are no larger nstlist value to try */
return;
}
}
if (EI_MD(ir->eI) && ir->etc == etcNO)
{
if (MASTER(cr))
{
fprintf(stderr, "%s\n", nve_err);
}
if (fp != NULL)
{
fprintf(fp, "%s\n", nve_err);
}
return;
}
if (ir->verletbuf_tol == 0 && bGPU)
{
gmx_fatal(FARGS, "You are using an old tpr file with a GPU, please generate a new tpr file with an up to date version of grompp");
}
if (ir->verletbuf_tol < 0)
{
if (MASTER(cr))
{
fprintf(stderr, "%s\n", vbd_err);
}
if (fp != NULL)
{
fprintf(fp, "%s\n", vbd_err);
}
return;
}
if (bGPU)
{
listfac_ok = nbnxn_gpu_listfac_ok;
listfac_max = nbnxn_gpu_listfac_max;
}
else
{
listfac_ok = nbnxn_cpu_listfac_ok;
listfac_max = nbnxn_cpu_listfac_max;
}
nstlist_orig = ir->nstlist;
if (nstlist_cmdline > 0)
{
if (fp)
{
sprintf(buf, "Getting nstlist=%d from command line option",
nstlist_cmdline);
}
ir->nstlist = nstlist_cmdline;
}
verletbuf_get_list_setup(TRUE, bGPU, &ls);
/* Allow rlist to make the list a given factor larger than the list
* would be with the reference value for nstlist (10).
*/
nstlist_prev = ir->nstlist;
ir->nstlist = nbnxnReferenceNstlist;
calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL,
&rlistWithReferenceNstlist);
ir->nstlist = nstlist_prev;
/* Determine the pair list size increase due to zero interactions */
rlist_inc = nbnxn_get_rlist_effective_inc(ls.cluster_size_j,
mtop->natoms/det(box));
rlist_ok = (rlistWithReferenceNstlist + rlist_inc)*pow(listfac_ok, oneThird) - rlist_inc;
rlist_max = (rlistWithReferenceNstlist + rlist_inc)*pow(listfac_max, oneThird) - rlist_inc;
if (debug)
{
fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
rlist_inc, rlist_ok, rlist_max);
}
nstlist_prev = nstlist_orig;
rlist_prev = ir->rlist;
do
{
if (nstlist_cmdline <= 0)
{
ir->nstlist = nstlist_try[nstlist_ind];
}
/* Set the pair-list buffer size in ir */
calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
/* Does rlist fit in the box? */
bBox = (sqr(rlist_new) < max_cutoff2(ir->ePBC, box));
bDD = TRUE;
if (bBox && DOMAINDECOMP(cr))
{
/* Check if rlist fits in the domain decomposition */
if (inputrec2nboundeddim(ir) < DIM)
{
gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
}
copy_mat(box, state_tmp.box);
bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
}
if (debug)
{
fprintf(debug, "nstlist %d rlist %.3f bBox %d bDD %d\n",
ir->nstlist, rlist_new, bBox, bDD);
}
bCont = FALSE;
if (nstlist_cmdline <= 0)
{
if (bBox && bDD && rlist_new <= rlist_max)
{
/* Increase nstlist */
nstlist_prev = ir->nstlist;
rlist_prev = rlist_new;
bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
}
else
{
/* Stick with the previous nstlist */
ir->nstlist = nstlist_prev;
rlist_new = rlist_prev;
bBox = TRUE;
bDD = TRUE;
}
}
nstlist_ind++;
}
while (bCont);
if (!bBox || !bDD)
{
gmx_warning(!bBox ? box_err : dd_err);
if (fp != NULL)
{
fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
}
ir->nstlist = nstlist_orig;
}
else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
{
sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
nstlist_orig, ir->nstlist,
ir->rlist, rlist_new);
if (MASTER(cr))
{
fprintf(stderr, "%s\n\n", buf);
}
if (fp != NULL)
{
fprintf(fp, "%s\n\n", buf);
}
ir->rlist = rlist_new;
ir->rlistlong = rlist_new;
}
}
