static void assign_factors(gmx_domdec_t *dd,
real limit,real cutoff,
matrix box,gmx_ddbox_t *ddbox,t_inputrec *ir,
float pbcdxr,int npme,
int ndiv,int *div,int *mdiv,ivec ir_try,ivec opt)
{
int x,y,z,i;
float ce;
if (ndiv == 0)
{
ce = comm_cost_est(dd,limit,cutoff,box,ddbox,ir,pbcdxr,npme,ir_try);
if (ce >= 0 && (opt[XX] == 0 ||
ce < comm_cost_est(dd,limit,cutoff,box,ddbox,ir,pbcdxr,
npme,opt)))
{
copy_ivec(ir_try,opt);
}
return;
}
for(x=mdiv[0]; x>=0; x--)
{
for(i=0; i<x; i++)
{
ir_try[XX] *= div[0];
}
for(y=mdiv[0]-x; y>=0; y--)
{
for(i=0; i<y; i++)
{
ir_try[YY] *= div[0];
}
for(i=0; i<mdiv[0]-x-y; i++)
{
ir_try[ZZ] *= div[0];
}
/* recurse */
assign_factors(dd,limit,cutoff,box,ddbox,ir,pbcdxr,npme,
ndiv-1,div+1,mdiv+1,ir_try,opt);
for(i=0; i<mdiv[0]-x-y; i++)
{
ir_try[ZZ] /= div[0];
}
for(i=0; i<y; i++)
{
ir_try[YY] /= div[0];
}
}
for(i=0; i<x; i++)
{
ir_try[XX] /= div[0];
}
}
}
/*! \brief Determine the optimal distribution of DD cells for the simulation system and number of MPI ranks */
static real optimize_ncells(FILE *fplog,
int nnodes_tot, int npme_only,
gmx_bool bDynLoadBal, real dlb_scale,
gmx_mtop_t *mtop, matrix box, gmx_ddbox_t *ddbox,
t_inputrec *ir,
gmx_domdec_t *dd,
real cellsize_limit, real cutoff,
gmx_bool bInterCGBondeds,
ivec nc)
{
int npp, npme, ndiv, *div, *mdiv, d, nmax;
double pbcdxr;
real limit;
ivec itry;
limit = cellsize_limit;
dd->nc[XX] = 1;
dd->nc[YY] = 1;
dd->nc[ZZ] = 1;
npp = nnodes_tot - npme_only;
if (EEL_PME(ir->coulombtype))
{
npme = (npme_only > 0 ? npme_only : npp);
}
else
{
npme = 0;
}
if (bInterCGBondeds)
{
/* If we can skip PBC for distance calculations in plain-C bondeds,
* we can save some time (e.g. 3D DD with pbc=xyz).
* Here we ignore SIMD bondeds as they always do (fast) PBC.
*/
count_bonded_distances(mtop, ir, &pbcdxr, NULL);
pbcdxr /= (double)mtop->natoms;
}
else
{
/* Every molecule is a single charge group: no pbc required */
pbcdxr = 0;
}
/* Add a margin for DLB and/or pressure scaling */
if (bDynLoadBal)
{
if (dlb_scale >= 1.0)
{
gmx_fatal(FARGS, "The value for option -dds should be smaller than 1");
}
if (fplog)
{
fprintf(fplog, "Scaling the initial minimum size with 1/%g (option -dds) = %g\n", dlb_scale, 1/dlb_scale);
}
limit /= dlb_scale;
}
else if (ir->epc != epcNO)
{
if (fplog)
{
fprintf(fplog, "To account for pressure scaling, scaling the initial minimum size with %g\n", DD_GRID_MARGIN_PRES_SCALE);
limit *= DD_GRID_MARGIN_PRES_SCALE;
}
}
if (fplog)
{
fprintf(fplog, "Optimizing the DD grid for %d cells with a minimum initial size of %.3f nm\n", npp, limit);
if (inhomogeneous_z(ir))
{
fprintf(fplog, "Ewald_geometry=%s: assuming inhomogeneous particle distribution in z, will not decompose in z.\n", eewg_names[ir->ewald_geometry]);
}
if (limit > 0)
{
fprintf(fplog, "The maximum allowed number of cells is:");
for (d = 0; d < DIM; d++)
{
nmax = (int)(ddbox->box_size[d]*ddbox->skew_fac[d]/limit);
if (d >= ddbox->npbcdim && nmax < 2)
{
nmax = 2;
}
if (d == ZZ && inhomogeneous_z(ir))
{
nmax = 1;
}
fprintf(fplog, " %c %d", 'X' + d, nmax);
}
fprintf(fplog, "\n");
}
}
if (debug)
{
fprintf(debug, "Average nr of pbc_dx calls per atom %.2f\n", pbcdxr);
}
/* Decompose npp in factors */
ndiv = factorize(npp, &div, &mdiv);
itry[XX] = 1;
itry[YY] = 1;
itry[ZZ] = 1;
clear_ivec(nc);
assign_factors(dd, limit, cutoff, box, ddbox, mtop->natoms, ir, pbcdxr,
npme, ndiv, div, mdiv, itry, nc);
sfree(div);
sfree(mdiv);
return limit;
}
static real optimize_ncells(FILE *fplog,
int nnodes_tot,int npme_only,
bool bDynLoadBal,real dlb_scale,
gmx_mtop_t *mtop,matrix box,gmx_ddbox_t *ddbox,
t_inputrec *ir,
gmx_domdec_t *dd,
real cellsize_limit,real cutoff,
bool bInterCGBondeds,bool bInterCGMultiBody,
ivec nc)
{
int npp,npme,ndiv,*div,*mdiv,d,nmax;
bool bExcl_pbcdx;
float pbcdxr;
real limit;
ivec itry;
limit = cellsize_limit;
dd->nc[XX] = 1;
dd->nc[YY] = 1;
dd->nc[ZZ] = 1;
npp = nnodes_tot - npme_only;
if (EEL_PME(ir->coulombtype))
{
npme = (npme_only > 0 ? npme_only : npp);
}
else
{
npme = 0;
}
if (bInterCGBondeds)
{
/* For Ewald exclusions pbc_dx is not called */
bExcl_pbcdx =
(EEL_EXCL_FORCES(ir->coulombtype) && !EEL_FULL(ir->coulombtype));
pbcdxr = (double)n_bonded_dx(mtop,bExcl_pbcdx)/(double)mtop->natoms;
}
else
{
/* Every molecule is a single charge group: no pbc required */
pbcdxr = 0;
}
/* Add a margin for DLB and/or pressure scaling */
if (bDynLoadBal)
{
if (dlb_scale >= 1.0)
{
gmx_fatal(FARGS,"The value for option -dds should be smaller than 1");
}
if (fplog)
{
fprintf(fplog,"Scaling the initial minimum size with 1/%g (option -dds) = %g\n",dlb_scale,1/dlb_scale);
}
limit /= dlb_scale;
}
else if (ir->epc != epcNO)
{
if (fplog)
{
fprintf(fplog,"To account for pressure scaling, scaling the initial minimum size with %g\n",DD_GRID_MARGIN_PRES_SCALE);
limit *= DD_GRID_MARGIN_PRES_SCALE;
}
}
if (fplog)
{
fprintf(fplog,"Optimizing the DD grid for %d cells with a minimum initial size of %.3f nm\n",npp,limit);
if (limit > 0)
{
fprintf(fplog,"The maximum allowed number of cells is:");
for(d=0; d<DIM; d++)
{
nmax = (int)(ddbox->box_size[d]*ddbox->skew_fac[d]/limit);
if (d >= ddbox->npbcdim && nmax < 2)
{
nmax = 2;
}
fprintf(fplog," %c %d",'X' + d,nmax);
}
fprintf(fplog,"\n");
}
}
if (debug)
{
fprintf(debug,"Average nr of pbc_dx calls per atom %.2f\n",pbcdxr);
}
/* Decompose npp in factors */
ndiv = factorize(npp,&div,&mdiv);
itry[XX] = 1;
itry[YY] = 1;
itry[ZZ] = 1;
clear_ivec(nc);
assign_factors(dd,limit,cutoff,box,ddbox,ir,pbcdxr,
npme,ndiv,div,mdiv,itry,nc);
sfree(div);
sfree(mdiv);
return limit;
}
