static void do_lincs(rvec *x,rvec *xp,matrix box,t_pbc *pbc,
struct gmx_lincsdata *lincsd,real *invmass,
t_commrec *cr,
real wangle,int *warn,
real invdt,rvec *v,
gmx_bool bCalcVir,tensor rmdr)
{
int b,i,j,k,n,iter;
real tmp0,tmp1,tmp2,im1,im2,mvb,rlen,len,len2,dlen2,wfac,lam;
rvec dx;
int ncons,*bla,*blnr,*blbnb;
rvec *r;
real *blc,*blmf,*bllen,*blcc,*rhs1,*rhs2,*sol,*lambda;
int *nlocat;
ncons = lincsd->nc;
bla = lincsd->bla;
r = lincsd->tmpv;
blnr = lincsd->blnr;
blbnb = lincsd->blbnb;
blc = lincsd->blc;
blmf = lincsd->blmf;
bllen = lincsd->bllen;
blcc = lincsd->tmpncc;
rhs1 = lincsd->tmp1;
rhs2 = lincsd->tmp2;
sol = lincsd->tmp3;
lambda = lincsd->lambda;
if (DOMAINDECOMP(cr) && cr->dd->constraints)
{
nlocat = dd_constraints_nlocalatoms(cr->dd);
}
else if (PARTDECOMP(cr))
{
nlocat = pd_constraints_nlocalatoms(cr->pd);
}
else
{
nlocat = NULL;
}
*warn = 0;
if (pbc)
{
/* Compute normalized i-j vectors */
for(b=0; b<ncons; b++)
{
pbc_dx_aiuc(pbc,x[bla[2*b]],x[bla[2*b+1]],dx);
unitv(dx,r[b]);
}
for(b=0; b<ncons; b++)
{
for(n=blnr[b]; n<blnr[b+1]; n++)
{
blcc[n] = blmf[n]*iprod(r[b],r[blbnb[n]]);
}
pbc_dx_aiuc(pbc,xp[bla[2*b]],xp[bla[2*b+1]],dx);
mvb = blc[b]*(iprod(r[b],dx) - bllen[b]);
rhs1[b] = mvb;
sol[b] = mvb;
}
}
else
{
/* Compute normalized i-j vectors */
for(b=0; b<ncons; b++)
{
i = bla[2*b];
j = bla[2*b+1];
tmp0 = x[i][0] - x[j][0];
tmp1 = x[i][1] - x[j][1];
tmp2 = x[i][2] - x[j][2];
rlen = gmx_invsqrt(tmp0*tmp0+tmp1*tmp1+tmp2*tmp2);
r[b][0] = rlen*tmp0;
r[b][1] = rlen*tmp1;
r[b][2] = rlen*tmp2;
} /* 16 ncons flops */
for(b=0; b<ncons; b++)
{
tmp0 = r[b][0];
tmp1 = r[b][1];
tmp2 = r[b][2];
len = bllen[b];
i = bla[2*b];
j = bla[2*b+1];
for(n=blnr[b]; n<blnr[b+1]; n++)
{
k = blbnb[n];
blcc[n] = blmf[n]*(tmp0*r[k][0] + tmp1*r[k][1] + tmp2*r[k][2]);
} /* 6 nr flops */
mvb = blc[b]*(tmp0*(xp[i][0] - xp[j][0]) +
tmp1*(xp[i][1] - xp[j][1]) +
tmp2*(xp[i][2] - xp[j][2]) - len);
rhs1[b] = mvb;
sol[b] = mvb;
/* 10 flops */
}
/* Together: 26*ncons + 6*nrtot flops */
}
lincs_matrix_expand(lincsd,blcc,rhs1,rhs2,sol);
/* nrec*(ncons+2*nrtot) flops */
for(b=0; b<ncons; b++)
{
i = bla[2*b];
j = bla[2*b+1];
mvb = blc[b]*sol[b];
lambda[b] = -mvb;
im1 = invmass[i];
im2 = invmass[j];
tmp0 = r[b][0]*mvb;
tmp1 = r[b][1]*mvb;
tmp2 = r[b][2]*mvb;
xp[i][0] -= tmp0*im1;
xp[i][1] -= tmp1*im1;
xp[i][2] -= tmp2*im1;
xp[j][0] += tmp0*im2;
xp[j][1] += tmp1*im2;
xp[j][2] += tmp2*im2;
} /* 16 ncons flops */
/*
******** Correction for centripetal effects ********
*/
wfac = cos(DEG2RAD*wangle);
wfac = wfac*wfac;
for(iter=0; iter<lincsd->nIter; iter++)
{
if (DOMAINDECOMP(cr) && cr->dd->constraints)
{
/* Communicate the corrected non-local coordinates */
dd_move_x_constraints(cr->dd,box,xp,NULL);
}
else if (PARTDECOMP(cr))
{
pd_move_x_constraints(cr,xp,NULL);
}
for(b=0; b<ncons; b++)
{
len = bllen[b];
if (pbc)
{
pbc_dx_aiuc(pbc,xp[bla[2*b]],xp[bla[2*b+1]],dx);
}
else
{
rvec_sub(xp[bla[2*b]],xp[bla[2*b+1]],dx);
}
len2 = len*len;
dlen2 = 2*len2 - norm2(dx);
if (dlen2 < wfac*len2 && (nlocat==NULL || nlocat[b]))
{
*warn = b;
}
if (dlen2 > 0)
{
mvb = blc[b]*(len - dlen2*gmx_invsqrt(dlen2));
}
else
{
mvb = blc[b]*len;
}
rhs1[b] = mvb;
sol[b] = mvb;
} /* 20*ncons flops */
lincs_matrix_expand(lincsd,blcc,rhs1,rhs2,sol);
/* nrec*(ncons+2*nrtot) flops */
for(b=0; b<ncons; b++)
{
i = bla[2*b];
j = bla[2*b+1];
lam = lambda[b];
mvb = blc[b]*sol[b];
lambda[b] = lam - mvb;
im1 = invmass[i];
im2 = invmass[j];
tmp0 = r[b][0]*mvb;
tmp1 = r[b][1]*mvb;
tmp2 = r[b][2]*mvb;
xp[i][0] -= tmp0*im1;
xp[i][1] -= tmp1*im1;
xp[i][2] -= tmp2*im1;
xp[j][0] += tmp0*im2;
xp[j][1] += tmp1*im2;
xp[j][2] += tmp2*im2;
} /* 17 ncons flops */
} /* nit*ncons*(37+9*nrec) flops */
if (v)
{
/* Correct the velocities */
for(b=0; b<ncons; b++)
{
i = bla[2*b];
j = bla[2*b+1];
im1 = invmass[i]*lambda[b]*invdt;
im2 = invmass[j]*lambda[b]*invdt;
v[i][0] += im1*r[b][0];
v[i][1] += im1*r[b][1];
v[i][2] += im1*r[b][2];
v[j][0] -= im2*r[b][0];
v[j][1] -= im2*r[b][1];
v[j][2] -= im2*r[b][2];
} /* 16 ncons flops */
}
if (nlocat)
{
/* Only account for local atoms */
for(b=0; b<ncons; b++)
{
lambda[b] *= 0.5*nlocat[b];
}
}
if (bCalcVir)
{
/* Constraint virial */
for(b=0; b<ncons; b++)
{
tmp0 = bllen[b]*lambda[b];
for(i=0; i<DIM; i++)
{
tmp1 = tmp0*r[b][i];
for(j=0; j<DIM; j++)
{
rmdr[i][j] -= tmp1*r[b][j];
}
}
} /* 22 ncons flops */
}
/* Total:
* 26*ncons + 6*nrtot + nrec*(ncons+2*nrtot)
* + nit * (20*ncons + nrec*(ncons+2*nrtot) + 17 ncons)
*
* (26+nrec)*ncons + (6+2*nrec)*nrtot
* + nit * ((37+nrec)*ncons + 2*nrec*nrtot)
* if nit=1
* (63+nrec)*ncons + (6+4*nrec)*nrtot
*/
}
void set_lincs(t_idef *idef,t_mdatoms *md,
gmx_bool bDynamics,t_commrec *cr,
struct gmx_lincsdata *li)
{
int start,natoms,nflexcon;
t_blocka at2con;
t_iatom *iatom;
int i,k,ncc_alloc,ni,con,nconnect,concon;
int type,a1,a2;
real lenA=0,lenB;
gmx_bool bLocal;
li->nc = 0;
li->ncc = 0;
/* This is the local topology, so there are only F_CONSTR constraints */
if (idef->il[F_CONSTR].nr == 0)
{
/* There are no constraints,
* we do not need to fill any data structures.
*/
return;
}
if (debug)
{
fprintf(debug,"Building the LINCS connectivity\n");
}
if (DOMAINDECOMP(cr))
{
if (cr->dd->constraints)
{
dd_get_constraint_range(cr->dd,&start,&natoms);
}
else
{
natoms = cr->dd->nat_home;
}
start = 0;
}
else if(PARTDECOMP(cr))
{
pd_get_constraint_range(cr->pd,&start,&natoms);
}
else
{
start = md->start;
natoms = md->homenr;
}
at2con = make_at2con(start,natoms,idef->il,idef->iparams,bDynamics,
&nflexcon);
if (idef->il[F_CONSTR].nr/3 > li->nc_alloc || li->nc_alloc == 0)
{
li->nc_alloc = over_alloc_dd(idef->il[F_CONSTR].nr/3);
srenew(li->bllen0,li->nc_alloc);
srenew(li->ddist,li->nc_alloc);
srenew(li->bla,2*li->nc_alloc);
srenew(li->blc,li->nc_alloc);
srenew(li->blc1,li->nc_alloc);
srenew(li->blnr,li->nc_alloc+1);
srenew(li->bllen,li->nc_alloc);
srenew(li->tmpv,li->nc_alloc);
srenew(li->tmp1,li->nc_alloc);
srenew(li->tmp2,li->nc_alloc);
srenew(li->tmp3,li->nc_alloc);
srenew(li->lambda,li->nc_alloc);
if (li->ncg_triangle > 0)
{
/* This is allocating too much, but it is difficult to improve */
srenew(li->triangle,li->nc_alloc);
srenew(li->tri_bits,li->nc_alloc);
}
}
iatom = idef->il[F_CONSTR].iatoms;
ncc_alloc = li->ncc_alloc;
li->blnr[0] = 0;
ni = idef->il[F_CONSTR].nr/3;
con = 0;
nconnect = 0;
li->blnr[con] = nconnect;
for(i=0; i<ni; i++)
{
bLocal = TRUE;
type = iatom[3*i];
a1 = iatom[3*i+1];
a2 = iatom[3*i+2];
lenA = idef->iparams[type].constr.dA;
lenB = idef->iparams[type].constr.dB;
/* Skip the flexible constraints when not doing dynamics */
if (bDynamics || lenA!=0 || lenB!=0)
{
li->bllen0[con] = lenA;
li->ddist[con] = lenB - lenA;
/* Set the length to the topology A length */
li->bllen[con] = li->bllen0[con];
li->bla[2*con] = a1;
li->bla[2*con+1] = a2;
/* Construct the constraint connection matrix blbnb */
for(k=at2con.index[a1-start]; k<at2con.index[a1-start+1]; k++)
{
concon = at2con.a[k];
if (concon != i)
{
if (nconnect >= ncc_alloc)
{
ncc_alloc = over_alloc_small(nconnect+1);
srenew(li->blbnb,ncc_alloc);
}
li->blbnb[nconnect++] = concon;
}
}
for(k=at2con.index[a2-start]; k<at2con.index[a2-start+1]; k++)
{
concon = at2con.a[k];
if (concon != i)
{
if (nconnect+1 > ncc_alloc)
{
ncc_alloc = over_alloc_small(nconnect+1);
srenew(li->blbnb,ncc_alloc);
}
li->blbnb[nconnect++] = concon;
}
}
li->blnr[con+1] = nconnect;
if (cr->dd == NULL)
{
/* Order the blbnb matrix to optimize memory access */
qsort(&(li->blbnb[li->blnr[con]]),li->blnr[con+1]-li->blnr[con],
sizeof(li->blbnb[0]),int_comp);
}
/* Increase the constraint count */
con++;
}
}
done_blocka(&at2con);
/* This is the real number of constraints,
* without dynamics the flexible constraints are not present.
*/
li->nc = con;
li->ncc = li->blnr[con];
if (cr->dd == NULL)
{
/* Since the matrix is static, we can free some memory */
ncc_alloc = li->ncc;
srenew(li->blbnb,ncc_alloc);
}
if (ncc_alloc > li->ncc_alloc)
{
li->ncc_alloc = ncc_alloc;
srenew(li->blmf,li->ncc_alloc);
srenew(li->blmf1,li->ncc_alloc);
srenew(li->tmpncc,li->ncc_alloc);
}
if (debug)
{
fprintf(debug,"Number of constraints is %d, couplings %d\n",
li->nc,li->ncc);
}
set_lincs_matrix(li,md->invmass,md->lambda);
}
int relax_shell_flexcon(FILE *fplog,t_commrec *cr,gmx_bool bVerbose,
gmx_large_int_t mdstep,t_inputrec *inputrec,
gmx_bool bDoNS,int force_flags,
gmx_bool bStopCM,
gmx_localtop_t *top,
gmx_mtop_t* mtop,
gmx_constr_t constr,
gmx_enerdata_t *enerd,t_fcdata *fcd,
t_state *state,rvec f[],
tensor force_vir,
t_mdatoms *md,
t_nrnb *nrnb,gmx_wallcycle_t wcycle,
t_graph *graph,
gmx_groups_t *groups,
struct gmx_shellfc *shfc,
t_forcerec *fr,
gmx_bool bBornRadii,
double t,rvec mu_tot,
int natoms,gmx_bool *bConverged,
gmx_vsite_t *vsite,
FILE *fp_field)
{
int nshell;
t_shell *shell;
t_idef *idef;
rvec *pos[2],*force[2],*acc_dir=NULL,*x_old=NULL;
real Epot[2],df[2];
rvec dx;
real sf_dir,invdt;
real ftol,xiH,xiS,dum=0;
char sbuf[22];
gmx_bool bCont,bInit;
int nat,dd_ac0,dd_ac1=0,i;
int start=md->start,homenr=md->homenr,end=start+homenr,cg0,cg1;
int nflexcon,g,number_steps,d,Min=0,count=0;
#define Try (1-Min) /* At start Try = 1 */
bCont = (mdstep == inputrec->init_step) && inputrec->bContinuation;
bInit = (mdstep == inputrec->init_step) || shfc->bRequireInit;
ftol = inputrec->em_tol;
number_steps = inputrec->niter;
nshell = shfc->nshell;
shell = shfc->shell;
nflexcon = shfc->nflexcon;
idef = &top->idef;
if (DOMAINDECOMP(cr)) {
nat = dd_natoms_vsite(cr->dd);
if (nflexcon > 0) {
dd_get_constraint_range(cr->dd,&dd_ac0,&dd_ac1);
nat = max(nat,dd_ac1);
}
} else {
nat = state->natoms;
}
if (nat > shfc->x_nalloc) {
/* Allocate local arrays */
shfc->x_nalloc = over_alloc_dd(nat);
for(i=0; (i<2); i++) {
srenew(shfc->x[i],shfc->x_nalloc);
srenew(shfc->f[i],shfc->x_nalloc);
}
}
for(i=0; (i<2); i++) {
pos[i] = shfc->x[i];
force[i] = shfc->f[i];
}
/* With particle decomposition this code only works
* when all particles involved with each shell are in the same cg.
*/
if (bDoNS && inputrec->ePBC != epbcNONE && !DOMAINDECOMP(cr)) {
/* This is the only time where the coordinates are used
* before do_force is called, which normally puts all
* charge groups in the box.
*/
if (PARTDECOMP(cr)) {
pd_cg_range(cr,&cg0,&cg1);
} else {
cg0 = 0;
cg1 = top->cgs.nr;
}
put_charge_groups_in_box(fplog,cg0,cg1,fr->ePBC,state->box,
&(top->cgs),state->x,fr->cg_cm);
if (graph)
mk_mshift(fplog,graph,fr->ePBC,state->box,state->x);
}
/* After this all coordinate arrays will contain whole molecules */
if (graph)
shift_self(graph,state->box,state->x);
if (nflexcon) {
if (nat > shfc->flex_nalloc) {
shfc->flex_nalloc = over_alloc_dd(nat);
srenew(shfc->acc_dir,shfc->flex_nalloc);
srenew(shfc->x_old,shfc->flex_nalloc);
}
acc_dir = shfc->acc_dir;
x_old = shfc->x_old;
for(i=0; i<homenr; i++) {
for(d=0; d<DIM; d++)
shfc->x_old[i][d] =
state->x[start+i][d] - state->v[start+i][d]*inputrec->delta_t;
}
}
/* Do a prediction of the shell positions */
if (shfc->bPredict && !bCont) {
predict_shells(fplog,state->x,state->v,inputrec->delta_t,nshell,shell,
md->massT,NULL,bInit);
}
/* do_force expected the charge groups to be in the box */
if (graph)
unshift_self(graph,state->box,state->x);
/* Calculate the forces first time around */
if (gmx_debug_at) {
pr_rvecs(debug,0,"x b4 do_force",state->x + start,homenr);
}
do_force(fplog,cr,inputrec,mdstep,nrnb,wcycle,top,mtop,groups,
state->box,state->x,&state->hist,
force[Min],force_vir,md,enerd,fcd,
state->lambda,graph,
fr,vsite,mu_tot,t,fp_field,NULL,bBornRadii,
(bDoNS ? GMX_FORCE_NS : 0) | force_flags);
sf_dir = 0;
if (nflexcon) {
init_adir(fplog,shfc,
constr,idef,inputrec,cr,dd_ac1,mdstep,md,start,end,
shfc->x_old-start,state->x,state->x,force[Min],
shfc->acc_dir-start,
fr->bMolPBC,state->box,state->lambda,&dum,nrnb);
for(i=start; i<end; i++)
sf_dir += md->massT[i]*norm2(shfc->acc_dir[i-start]);
}
Epot[Min] = enerd->term[F_EPOT];
df[Min]=rms_force(cr,shfc->f[Min],nshell,shell,nflexcon,&sf_dir,&Epot[Min]);
df[Try]=0;
if (debug) {
fprintf(debug,"df = %g %g\n",df[Min],df[Try]);
}
if (gmx_debug_at) {
pr_rvecs(debug,0,"force0",force[Min],md->nr);
}
if (nshell+nflexcon > 0) {
/* Copy x to pos[Min] & pos[Try]: during minimization only the
* shell positions are updated, therefore the other particles must
* be set here.
*/
memcpy(pos[Min],state->x,nat*sizeof(state->x[0]));
memcpy(pos[Try],state->x,nat*sizeof(state->x[0]));
}
if (bVerbose && MASTER(cr))
print_epot(stdout,mdstep,0,Epot[Min],df[Min],nflexcon,sf_dir);
if (debug) {
fprintf(debug,"%17s: %14.10e\n",
interaction_function[F_EKIN].longname,enerd->term[F_EKIN]);
fprintf(debug,"%17s: %14.10e\n",
interaction_function[F_EPOT].longname,enerd->term[F_EPOT]);
fprintf(debug,"%17s: %14.10e\n",
interaction_function[F_ETOT].longname,enerd->term[F_ETOT]);
fprintf(debug,"SHELLSTEP %s\n",gmx_step_str(mdstep,sbuf));
}
/* First check whether we should do shells, or whether the force is
* low enough even without minimization.
*/
*bConverged = (df[Min] < ftol);
for(count=1; (!(*bConverged) && (count < number_steps)); count++) {
if (vsite)
construct_vsites(fplog,vsite,pos[Min],nrnb,inputrec->delta_t,state->v,
idef->iparams,idef->il,
fr->ePBC,fr->bMolPBC,graph,cr,state->box);
if (nflexcon) {
init_adir(fplog,shfc,
constr,idef,inputrec,cr,dd_ac1,mdstep,md,start,end,
x_old-start,state->x,pos[Min],force[Min],acc_dir-start,
fr->bMolPBC,state->box,state->lambda,&dum,nrnb);
directional_sd(fplog,pos[Min],pos[Try],acc_dir-start,start,end,
fr->fc_stepsize);
}
/* New positions, Steepest descent */
shell_pos_sd(fplog,pos[Min],pos[Try],force[Min],nshell,shell,count);
/* do_force expected the charge groups to be in the box */
if (graph)
unshift_self(graph,state->box,pos[Try]);
if (gmx_debug_at) {
pr_rvecs(debug,0,"RELAX: pos[Min] ",pos[Min] + start,homenr);
pr_rvecs(debug,0,"RELAX: pos[Try] ",pos[Try] + start,homenr);
}
/* Try the new positions */
do_force(fplog,cr,inputrec,1,nrnb,wcycle,
top,mtop,groups,state->box,pos[Try],&state->hist,
force[Try],force_vir,
md,enerd,fcd,state->lambda,graph,
fr,vsite,mu_tot,t,fp_field,NULL,bBornRadii,
force_flags);
if (gmx_debug_at) {
pr_rvecs(debug,0,"RELAX: force[Min]",force[Min] + start,homenr);
pr_rvecs(debug,0,"RELAX: force[Try]",force[Try] + start,homenr);
}
sf_dir = 0;
if (nflexcon) {
init_adir(fplog,shfc,
constr,idef,inputrec,cr,dd_ac1,mdstep,md,start,end,
x_old-start,state->x,pos[Try],force[Try],acc_dir-start,
fr->bMolPBC,state->box,state->lambda,&dum,nrnb);
for(i=start; i<end; i++)
sf_dir += md->massT[i]*norm2(acc_dir[i-start]);
}
Epot[Try] = enerd->term[F_EPOT];
df[Try]=rms_force(cr,force[Try],nshell,shell,nflexcon,&sf_dir,&Epot[Try]);
if (debug)
fprintf(debug,"df = %g %g\n",df[Min],df[Try]);
if (debug) {
if (gmx_debug_at)
pr_rvecs(debug,0,"F na do_force",force[Try] + start,homenr);
if (gmx_debug_at) {
fprintf(debug,"SHELL ITER %d\n",count);
dump_shells(debug,pos[Try],force[Try],ftol,nshell,shell);
}
}
if (bVerbose && MASTER(cr))
print_epot(stdout,mdstep,count,Epot[Try],df[Try],nflexcon,sf_dir);
*bConverged = (df[Try] < ftol);
if ((df[Try] < df[Min])) {
if (debug)
fprintf(debug,"Swapping Min and Try\n");
if (nflexcon) {
/* Correct the velocities for the flexible constraints */
invdt = 1/inputrec->delta_t;
for(i=start; i<end; i++) {
for(d=0; d<DIM; d++)
state->v[i][d] += (pos[Try][i][d] - pos[Min][i][d])*invdt;
}
}
Min = Try;
} else {
decrease_step_size(nshell,shell);
}
}
if (MASTER(cr) && !(*bConverged)) {
/* Note that the energies and virial are incorrect when not converged */
if (fplog)
fprintf(fplog,
"step %s: EM did not converge in %d iterations, RMS force %.3f\n",
gmx_step_str(mdstep,sbuf),number_steps,df[Min]);
fprintf(stderr,
"step %s: EM did not converge in %d iterations, RMS force %.3f\n",
gmx_step_str(mdstep,sbuf),number_steps,df[Min]);
}
/* Copy back the coordinates and the forces */
memcpy(state->x,pos[Min],nat*sizeof(state->x[0]));
memcpy(f,force[Min],nat*sizeof(f[0]));
return count;
}
gmx_bool constrain_lincs(FILE *fplog,gmx_bool bLog,gmx_bool bEner,
t_inputrec *ir,
gmx_large_int_t step,
struct gmx_lincsdata *lincsd,t_mdatoms *md,
t_commrec *cr,
rvec *x,rvec *xprime,rvec *min_proj,matrix box,
real lambda,real *dvdlambda,
real invdt,rvec *v,
gmx_bool bCalcVir,tensor rmdr,
int econq,
t_nrnb *nrnb,
int maxwarn,int *warncount)
{
char buf[STRLEN],buf2[22],buf3[STRLEN];
int i,warn,p_imax,error;
real ncons_loc,p_ssd,p_max;
t_pbc pbc,*pbc_null;
rvec dx;
gmx_bool bOK;
bOK = TRUE;
if (lincsd->nc == 0 && cr->dd == NULL)
{
if (bLog || bEner)
{
lincsd->rmsd_data[0] = 0;
if (ir->eI == eiSD2 && v == NULL)
{
i = 2;
}
else
{
i = 1;
}
lincsd->rmsd_data[i] = 0;
}
return bOK;
}
/* We do not need full pbc when constraints do not cross charge groups,
* i.e. when dd->constraint_comm==NULL
*/
if ((cr->dd || ir->bPeriodicMols) && !(cr->dd && cr->dd->constraint_comm==NULL))
{
/* With pbc=screw the screw has been changed to a shift
* by the constraint coordinate communication routine,
* so that here we can use normal pbc.
*/
pbc_null = set_pbc_dd(&pbc,ir->ePBC,cr->dd,FALSE,box);
}
else
{
pbc_null = NULL;
}
if (cr->dd)
{
/* Communicate the coordinates required for the non-local constraints */
dd_move_x_constraints(cr->dd,box,x,xprime);
/* dump_conf(dd,lincsd,NULL,"con",TRUE,xprime,box); */
}
else if (PARTDECOMP(cr))
{
pd_move_x_constraints(cr,x,xprime);
}
if (econq == econqCoord)
{
if (ir->efep != efepNO)
{
if (md->nMassPerturbed && lincsd->matlam != md->lambda)
{
set_lincs_matrix(lincsd,md->invmass,md->lambda);
}
for(i=0; i<lincsd->nc; i++)
{
lincsd->bllen[i] = lincsd->bllen0[i] + lambda*lincsd->ddist[i];
}
}
if (lincsd->ncg_flex)
{
/* Set the flexible constraint lengths to the old lengths */
if (pbc_null)
{
for(i=0; i<lincsd->nc; i++)
{
if (lincsd->bllen[i] == 0) {
pbc_dx_aiuc(pbc_null,x[lincsd->bla[2*i]],x[lincsd->bla[2*i+1]],dx);
lincsd->bllen[i] = norm(dx);
}
}
}
else
{
for(i=0; i<lincsd->nc; i++)
{
if (lincsd->bllen[i] == 0)
{
lincsd->bllen[i] =
sqrt(distance2(x[lincsd->bla[2*i]],
x[lincsd->bla[2*i+1]]));
}
}
}
}
if (bLog && fplog)
{
cconerr(cr->dd,lincsd->nc,lincsd->bla,lincsd->bllen,xprime,pbc_null,
&ncons_loc,&p_ssd,&p_max,&p_imax);
}
do_lincs(x,xprime,box,pbc_null,lincsd,md->invmass,cr,
ir->LincsWarnAngle,&warn,
invdt,v,bCalcVir,rmdr);
if (ir->efep != efepNO)
{
real dt_2,dvdl=0;
dt_2 = 1.0/(ir->delta_t*ir->delta_t);
for(i=0; (i<lincsd->nc); i++)
{
dvdl += lincsd->lambda[i]*dt_2*lincsd->ddist[i];
}
*dvdlambda += dvdl;
}
if (bLog && fplog && lincsd->nc > 0)
{
fprintf(fplog," Rel. Constraint Deviation: RMS MAX between atoms\n");
fprintf(fplog," Before LINCS %.6f %.6f %6d %6d\n",
sqrt(p_ssd/ncons_loc),p_max,
ddglatnr(cr->dd,lincsd->bla[2*p_imax]),
ddglatnr(cr->dd,lincsd->bla[2*p_imax+1]));
}
if (bLog || bEner)
{
cconerr(cr->dd,lincsd->nc,lincsd->bla,lincsd->bllen,xprime,pbc_null,
&ncons_loc,&p_ssd,&p_max,&p_imax);
/* Check if we are doing the second part of SD */
if (ir->eI == eiSD2 && v == NULL)
{
i = 2;
}
else
{
i = 1;
}
lincsd->rmsd_data[0] = ncons_loc;
lincsd->rmsd_data[i] = p_ssd;
}
else
{
lincsd->rmsd_data[0] = 0;
lincsd->rmsd_data[1] = 0;
lincsd->rmsd_data[2] = 0;
}
if (bLog && fplog && lincsd->nc > 0)
{
fprintf(fplog,
" After LINCS %.6f %.6f %6d %6d\n\n",
sqrt(p_ssd/ncons_loc),p_max,
ddglatnr(cr->dd,lincsd->bla[2*p_imax]),
ddglatnr(cr->dd,lincsd->bla[2*p_imax+1]));
}
if (warn > 0)
{
if (maxwarn >= 0)
{
cconerr(cr->dd,lincsd->nc,lincsd->bla,lincsd->bllen,xprime,pbc_null,
&ncons_loc,&p_ssd,&p_max,&p_imax);
if (MULTISIM(cr))
{
sprintf(buf3," in simulation %d", cr->ms->sim);
}
else
{
buf3[0] = 0;
}
sprintf(buf,"\nStep %s, time %g (ps) LINCS WARNING%s\n"
"relative constraint deviation after LINCS:\n"
"rms %.6f, max %.6f (between atoms %d and %d)\n",
gmx_step_str(step,buf2),ir->init_t+step*ir->delta_t,
buf3,
sqrt(p_ssd/ncons_loc),p_max,
ddglatnr(cr->dd,lincsd->bla[2*p_imax]),
ddglatnr(cr->dd,lincsd->bla[2*p_imax+1]));
if (fplog)
{
fprintf(fplog,"%s",buf);
}
fprintf(stderr,"%s",buf);
lincs_warning(fplog,cr->dd,x,xprime,pbc_null,
lincsd->nc,lincsd->bla,lincsd->bllen,
ir->LincsWarnAngle,maxwarn,warncount);
}
bOK = (p_max < 0.5);
}
if (lincsd->ncg_flex) {
for(i=0; (i<lincsd->nc); i++)
if (lincsd->bllen0[i] == 0 && lincsd->ddist[i] == 0)
lincsd->bllen[i] = 0;
}
}
else
{
do_lincsp(x,xprime,min_proj,pbc_null,lincsd,md->invmass,econq,dvdlambda,
bCalcVir,rmdr);
}
/* count assuming nit=1 */
inc_nrnb(nrnb,eNR_LINCS,lincsd->nc);
inc_nrnb(nrnb,eNR_LINCSMAT,(2+lincsd->nOrder)*lincsd->ncc);
if (lincsd->ntriangle > 0)
{
inc_nrnb(nrnb,eNR_LINCSMAT,lincsd->nOrder*lincsd->ncc_triangle);
}
if (v)
{
inc_nrnb(nrnb,eNR_CONSTR_V,lincsd->nc*2);
}
if (bCalcVir)
{
inc_nrnb(nrnb,eNR_CONSTR_VIR,lincsd->nc);
}
return bOK;
}
void finish_run(FILE *fplog,t_commrec *cr,char *confout,
t_inputrec *inputrec,
t_nrnb nrnb[],gmx_wallcycle_t wcycle,
double nodetime,double realtime,int nsteps_done,
bool bWriteStat)
{
int i,j;
t_nrnb *nrnb_all=NULL,ntot;
real delta_t;
double nbfs,mflop;
double cycles[ewcNR];
#ifdef GMX_MPI
int sender;
double nrnb_buf[4];
MPI_Status status;
#endif
wallcycle_sum(cr,wcycle,cycles);
if (cr->nnodes > 1) {
if (SIMMASTER(cr))
snew(nrnb_all,cr->nnodes);
#ifdef GMX_MPI
MPI_Gather(nrnb,sizeof(t_nrnb),MPI_BYTE,
nrnb_all,sizeof(t_nrnb),MPI_BYTE,
0,cr->mpi_comm_mysim);
#endif
} else {
nrnb_all = nrnb;
}
if (SIMMASTER(cr)) {
for(i=0; (i<eNRNB); i++)
ntot.n[i]=0;
for(i=0; (i<cr->nnodes); i++)
for(j=0; (j<eNRNB); j++)
ntot.n[j] += nrnb_all[i].n[j];
print_flop(fplog,&ntot,&nbfs,&mflop);
if (nrnb_all) {
sfree(nrnb_all);
}
}
if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr)) {
print_dd_statistics(cr,inputrec,fplog);
}
if (SIMMASTER(cr)) {
if (PARTDECOMP(cr)) {
pr_load(fplog,cr,nrnb_all);
}
wallcycle_print(fplog,cr->nnodes,cr->npmenodes,realtime,wcycle,cycles);
if (EI_DYNAMICS(inputrec->eI)) {
delta_t = inputrec->delta_t;
} else {
delta_t = 0;
}
if (fplog) {
print_perf(fplog,nodetime,realtime,cr->nnodes-cr->npmenodes,
nsteps_done,delta_t,nbfs,mflop);
}
if (bWriteStat) {
print_perf(stderr,nodetime,realtime,cr->nnodes-cr->npmenodes,
nsteps_done,delta_t,nbfs,mflop);
}
/*
runtime=inputrec->nsteps*inputrec->delta_t;
if (bWriteStat) {
if (cr->nnodes == 1)
fprintf(stderr,"\n\n");
print_perf(stderr,nodetime,realtime,runtime,&ntot,
cr->nnodes-cr->npmenodes,FALSE);
}
wallcycle_print(fplog,cr->nnodes,cr->npmenodes,realtime,wcycle,cycles);
print_perf(fplog,nodetime,realtime,runtime,&ntot,cr->nnodes-cr->npmenodes,
TRUE);
if (PARTDECOMP(cr))
pr_load(fplog,cr,nrnb_all);
if (cr->nnodes > 1)
sfree(nrnb_all);
*/
}
}
void do_force(FILE *fplog,t_commrec *cr,
t_inputrec *inputrec,
int step,t_nrnb *nrnb,gmx_wallcycle_t wcycle,
gmx_localtop_t *top,
gmx_groups_t *groups,
matrix box,rvec x[],history_t *hist,
rvec f[],rvec buf[],
tensor vir_force,
t_mdatoms *mdatoms,
gmx_enerdata_t *enerd,t_fcdata *fcd,
real lambda,t_graph *graph,
t_forcerec *fr,gmx_vsite_t *vsite,rvec mu_tot,
real t,FILE *field,gmx_edsam_t ed,
int flags)
{
static rvec box_size;
int cg0,cg1,i,j;
int start,homenr;
static double mu[2*DIM];
rvec mu_tot_AB[2];
bool bSepDVDL,bStateChanged,bNS,bFillGrid,bCalcCGCM,bBS,bDoForces;
matrix boxs;
real e,v,dvdl;
t_pbc pbc;
float cycles_ppdpme,cycles_pme,cycles_force;
start = mdatoms->start;
homenr = mdatoms->homenr;
bSepDVDL = (fr->bSepDVDL && do_per_step(step,inputrec->nstlog));
clear_mat(vir_force);
if (PARTDECOMP(cr)) {
pd_cg_range(cr,&cg0,&cg1);
} else {
cg0 = 0;
if (DOMAINDECOMP(cr))
cg1 = cr->dd->ncg_tot;
else
cg1 = top->cgs.nr;
if (fr->n_tpi > 0)
cg1--;
}
bStateChanged = (flags & GMX_FORCE_STATECHANGED);
bNS = (flags & GMX_FORCE_NS);
bFillGrid = (bNS && bStateChanged);
bCalcCGCM = (bFillGrid && !DOMAINDECOMP(cr));
bDoForces = (flags & GMX_FORCE_FORCES);
if (bStateChanged) {
update_forcerec(fplog,fr,box);
/* Calculate total (local) dipole moment in a temporary common array.
* This makes it possible to sum them over nodes faster.
*/
calc_mu(start,homenr,
x,mdatoms->chargeA,mdatoms->chargeB,mdatoms->nChargePerturbed,
mu,mu+DIM);
}
if (fr->ePBC != epbcNONE) {
/* Compute shift vectors every step,
* because of pressure coupling or box deformation!
*/
if (DYNAMIC_BOX(*inputrec) && bStateChanged)
calc_shifts(box,fr->shift_vec);
if (bCalcCGCM) {
put_charge_groups_in_box(fplog,cg0,cg1,fr->ePBC,box,
&(top->cgs),x,fr->cg_cm);
inc_nrnb(nrnb,eNR_CGCM,homenr);
inc_nrnb(nrnb,eNR_RESETX,cg1-cg0);
}
else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph) {
unshift_self(graph,box,x);
}
}
else if (bCalcCGCM) {
calc_cgcm(fplog,cg0,cg1,&(top->cgs),x,fr->cg_cm);
inc_nrnb(nrnb,eNR_CGCM,homenr);
}
if (bCalcCGCM) {
if (PAR(cr)) {
move_cgcm(fplog,cr,fr->cg_cm);
}
if (gmx_debug_at)
pr_rvecs(debug,0,"cgcm",fr->cg_cm,top->cgs.nr);
}
#ifdef GMX_MPI
if (!(cr->duty & DUTY_PME)) {
/* Send particle coordinates to the pme nodes.
* Since this is only implemented for domain decomposition
* and domain decomposition does not use the graph,
* we do not need to worry about shifting.
*/
wallcycle_start(wcycle,ewcPP_PMESENDX);
GMX_MPE_LOG(ev_send_coordinates_start);
bBS = (inputrec->nwall == 2);
if (bBS) {
copy_mat(box,boxs);
svmul(inputrec->wall_ewald_zfac,boxs[ZZ],boxs[ZZ]);
}
gmx_pme_send_x(cr,bBS ? boxs : box,x,mdatoms->nChargePerturbed,lambda);
GMX_MPE_LOG(ev_send_coordinates_finish);
wallcycle_stop(wcycle,ewcPP_PMESENDX);
}
#endif /* GMX_MPI */
/* Communicate coordinates and sum dipole if necessary */
if (PAR(cr)) {
wallcycle_start(wcycle,ewcMOVEX);
if (DOMAINDECOMP(cr)) {
dd_move_x(cr->dd,box,x,buf);
} else {
move_x(fplog,cr,GMX_LEFT,GMX_RIGHT,x,nrnb);
}
/* When we don't need the total dipole we sum it in global_stat */
if (NEED_MUTOT(*inputrec))
gmx_sumd(2*DIM,mu,cr);
wallcycle_stop(wcycle,ewcMOVEX);
}
for(i=0; i<2; i++)
for(j=0;j<DIM;j++)
mu_tot_AB[i][j] = mu[i*DIM + j];
if (fr->efep == efepNO)
copy_rvec(mu_tot_AB[0],mu_tot);
else
for(j=0; j<DIM; j++)
mu_tot[j] = (1.0 - lambda)*mu_tot_AB[0][j] + lambda*mu_tot_AB[1][j];
/* Reset energies */
reset_energies(&(inputrec->opts),fr,bNS,enerd,MASTER(cr));
if (bNS) {
wallcycle_start(wcycle,ewcNS);
if (graph && bStateChanged)
/* Calculate intramolecular shift vectors to make molecules whole */
mk_mshift(fplog,graph,fr->ePBC,box,x);
/* Reset long range forces if necessary */
if (fr->bTwinRange) {
clear_rvecs(fr->f_twin_n,fr->f_twin);
clear_rvecs(SHIFTS,fr->fshift_twin);
}
/* Do the actual neighbour searching and if twin range electrostatics
* also do the calculation of long range forces and energies.
*/
dvdl = 0;
ns(fplog,fr,x,f,box,groups,&(inputrec->opts),top,mdatoms,
cr,nrnb,step,lambda,&dvdl,&enerd->grpp,bFillGrid,bDoForces);
if (bSepDVDL)
fprintf(fplog,sepdvdlformat,"LR non-bonded",0,dvdl);
enerd->dvdl_lr = dvdl;
enerd->term[F_DVDL] += dvdl;
wallcycle_stop(wcycle,ewcNS);
}
if (DOMAINDECOMP(cr)) {
if (!(cr->duty & DUTY_PME)) {
wallcycle_start(wcycle,ewcPPDURINGPME);
dd_force_flop_start(cr->dd,nrnb);
}
}
/* Start the force cycle counter.
* This counter is stopped in do_forcelow_level.
* No parallel communication should occur while this counter is running,
* since that will interfere with the dynamic load balancing.
*/
wallcycle_start(wcycle,ewcFORCE);
if (bDoForces) {
/* Reset PME/Ewald forces if necessary */
if (fr->bF_NoVirSum)
{
GMX_BARRIER(cr->mpi_comm_mygroup);
if (fr->bDomDec)
clear_rvecs(fr->f_novirsum_n,fr->f_novirsum);
else
clear_rvecs(homenr,fr->f_novirsum+start);
GMX_BARRIER(cr->mpi_comm_mygroup);
}
/* Copy long range forces into normal buffers */
if (fr->bTwinRange) {
for(i=0; i<fr->f_twin_n; i++)
copy_rvec(fr->f_twin[i],f[i]);
for(i=0; i<SHIFTS; i++)
copy_rvec(fr->fshift_twin[i],fr->fshift[i]);
}
else {
if (DOMAINDECOMP(cr))
clear_rvecs(cr->dd->nat_tot,f);
else
clear_rvecs(mdatoms->nr,f);
clear_rvecs(SHIFTS,fr->fshift);
}
clear_rvec(fr->vir_diag_posres);
GMX_BARRIER(cr->mpi_comm_mygroup);
}
if (inputrec->ePull == epullCONSTRAINT)
clear_pull_forces(inputrec->pull);
/* update QMMMrec, if necessary */
if(fr->bQMMM)
update_QMMMrec(cr,fr,x,mdatoms,box,top);
if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0) {
/* Position restraints always require full pbc */
set_pbc(&pbc,inputrec->ePBC,box);
v = posres(top->idef.il[F_POSRES].nr,top->idef.il[F_POSRES].iatoms,
top->idef.iparams_posres,
(const rvec*)x,fr->f_novirsum,fr->vir_diag_posres,
inputrec->ePBC==epbcNONE ? NULL : &pbc,lambda,&dvdl,
fr->rc_scaling,fr->ePBC,fr->posres_com,fr->posres_comB);
if (bSepDVDL) {
fprintf(fplog,sepdvdlformat,
interaction_function[F_POSRES].longname,v,dvdl);
}
enerd->term[F_POSRES] += v;
enerd->term[F_DVDL] += dvdl;
inc_nrnb(nrnb,eNR_POSRES,top->idef.il[F_POSRES].nr/2);
}
/* Compute the bonded and non-bonded forces */
do_force_lowlevel(fplog,step,fr,inputrec,&(top->idef),
cr,nrnb,wcycle,mdatoms,&(inputrec->opts),
x,hist,f,enerd,fcd,box,lambda,graph,&(top->excls),mu_tot_AB,
flags,&cycles_force);
GMX_BARRIER(cr->mpi_comm_mygroup);
if (ed) {
do_flood(fplog,cr,x,f,ed,box,step);
}
if (DOMAINDECOMP(cr)) {
dd_force_flop_stop(cr->dd,nrnb);
if (wcycle)
dd_cycles_add(cr->dd,cycles_force,ddCyclF);
}
if (bDoForces) {
/* Compute forces due to electric field */
calc_f_el(MASTER(cr) ? field : NULL,
start,homenr,mdatoms->chargeA,x,f,inputrec->ex,inputrec->et,t);
/* When using PME/Ewald we compute the long range virial there.
* otherwise we do it based on long range forces from twin range
* cut-off based calculation (or not at all).
*/
/* Communicate the forces */
if (PAR(cr)) {
wallcycle_start(wcycle,ewcMOVEF);
if (DOMAINDECOMP(cr)) {
dd_move_f(cr->dd,f,buf,fr->fshift);
/* Position restraint do not introduce inter-cg forces */
if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl)
dd_move_f(cr->dd,fr->f_novirsum,buf,NULL);
} else {
move_f(fplog,cr,GMX_LEFT,GMX_RIGHT,f,buf,nrnb);
}
wallcycle_stop(wcycle,ewcMOVEF);
}
}
if (bDoForces) {
if (vsite) {
wallcycle_start(wcycle,ewcVSITESPREAD);
spread_vsite_f(fplog,vsite,x,f,fr->fshift,nrnb,
&top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
wallcycle_stop(wcycle,ewcVSITESPREAD);
}
/* Calculation of the virial must be done after vsites! */
calc_virial(fplog,mdatoms->start,mdatoms->homenr,x,f,
vir_force,graph,box,nrnb,fr,inputrec->ePBC);
}
if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F) {
/* Calculate the center of mass forces, this requires communication,
* which is why pull_potential is called close to other communication.
* The virial contribution is calculated directly,
* which is why we call pull_potential after calc_virial.
*/
set_pbc(&pbc,inputrec->ePBC,box);
dvdl = 0;
enerd->term[F_COM_PULL] =
pull_potential(inputrec->ePull,inputrec->pull,mdatoms,&pbc,
cr,t,lambda,x,f,vir_force,&dvdl);
if (bSepDVDL)
fprintf(fplog,sepdvdlformat,"Com pull",enerd->term[F_COM_PULL],dvdl);
enerd->term[F_DVDL] += dvdl;
}
if (!(cr->duty & DUTY_PME)) {
cycles_ppdpme = wallcycle_stop(wcycle,ewcPPDURINGPME);
dd_cycles_add(cr->dd,cycles_ppdpme,ddCyclPPduringPME);
}
#ifdef GMX_MPI
if (PAR(cr) && !(cr->duty & DUTY_PME)) {
/* In case of node-splitting, the PP nodes receive the long-range
* forces, virial and energy from the PME nodes here.
*/
wallcycle_start(wcycle,ewcPP_PMEWAITRECVF);
dvdl = 0;
gmx_pme_receive_f(cr,fr->f_novirsum,fr->vir_el_recip,&e,&dvdl,
&cycles_pme);
if (bSepDVDL)
fprintf(fplog,sepdvdlformat,"PME mesh",e,dvdl);
enerd->term[F_COUL_RECIP] += e;
enerd->term[F_DVDL] += dvdl;
if (wcycle)
dd_cycles_add(cr->dd,cycles_pme,ddCyclPME);
wallcycle_stop(wcycle,ewcPP_PMEWAITRECVF);
}
#endif
if (bDoForces && fr->bF_NoVirSum) {
if (vsite) {
/* Spread the mesh force on virtual sites to the other particles...
* This is parallellized. MPI communication is performed
* if the constructing atoms aren't local.
*/
wallcycle_start(wcycle,ewcVSITESPREAD);
spread_vsite_f(fplog,vsite,x,fr->f_novirsum,NULL,nrnb,
&top->idef,fr->ePBC,fr->bMolPBC,graph,box,cr);
wallcycle_stop(wcycle,ewcVSITESPREAD);
}
/* Now add the forces, this is local */
if (fr->bDomDec) {
sum_forces(0,fr->f_novirsum_n,f,fr->f_novirsum);
} else {
sum_forces(start,start+homenr,f,fr->f_novirsum);
}
if (EEL_FULL(fr->eeltype)) {
/* Add the mesh contribution to the virial */
m_add(vir_force,fr->vir_el_recip,vir_force);
}
if (debug)
pr_rvecs(debug,0,"vir_force",vir_force,DIM);
}
/* Sum the potential energy terms from group contributions */
sum_epot(&(inputrec->opts),enerd);
if (fr->print_force >= 0 && bDoForces)
print_large_forces(stderr,mdatoms,cr,step,fr->print_force,x,f);
}
void init_pull(FILE *fplog, t_inputrec *ir, int nfile, const t_filenm fnm[],
gmx_mtop_t *mtop, t_commrec *cr, const output_env_t oenv, real lambda,
gmx_bool bOutFile, unsigned long Flags)
{
t_pull *pull;
t_pull_group *pgrp;
int c, g, start = 0, end = 0, m;
pull = ir->pull;
pull->ePBC = ir->ePBC;
switch (pull->ePBC)
{
case epbcNONE: pull->npbcdim = 0; break;
case epbcXY: pull->npbcdim = 2; break;
default: pull->npbcdim = 3; break;
}
if (fplog)
{
gmx_bool bAbs, bCos;
bAbs = FALSE;
for (c = 0; c < pull->ncoord; c++)
{
if (pull->group[pull->coord[c].group[0]].nat == 0 ||
pull->group[pull->coord[c].group[1]].nat == 0)
{
bAbs = TRUE;
}
}
fprintf(fplog, "\nWill apply %s COM pulling in geometry '%s'\n",
EPULLTYPE(ir->ePull), EPULLGEOM(pull->eGeom));
fprintf(fplog, "with %d pull coordinate%s and %d group%s\n",
pull->ncoord, pull->ncoord == 1 ? "" : "s",
pull->ngroup, pull->ngroup == 1 ? "" : "s");
if (bAbs)
{
fprintf(fplog, "with an absolute reference\n");
}
bCos = FALSE;
for (g = 0; g < pull->ngroup; g++)
{
if (pull->group[g].nat > 1 &&
pull->group[g].pbcatom < 0)
{
/* We are using cosine weighting */
fprintf(fplog, "Cosine weighting is used for group %d\n", g);
bCos = TRUE;
}
}
if (bCos)
{
please_cite(fplog, "Engin2010");
}
}
/* We always add the virial contribution,
* except for geometry = direction_periodic where this is impossible.
*/
pull->bVirial = (pull->eGeom != epullgDIRPBC);
if (getenv("GMX_NO_PULLVIR") != NULL)
{
if (fplog)
{
fprintf(fplog, "Found env. var., will not add the virial contribution of the COM pull forces\n");
}
pull->bVirial = FALSE;
}
if (cr && PARTDECOMP(cr))
{
pd_at_range(cr, &start, &end);
}
pull->rbuf = NULL;
pull->dbuf = NULL;
pull->dbuf_cyl = NULL;
pull->bRefAt = FALSE;
pull->cosdim = -1;
for (g = 0; g < pull->ngroup; g++)
{
pgrp = &pull->group[g];
pgrp->epgrppbc = epgrppbcNONE;
if (pgrp->nat > 0)
{
/* Determine if we need to take PBC into account for calculating
* the COM's of the pull groups.
*/
for (m = 0; m < pull->npbcdim; m++)
{
if (pull->dim[m] && pgrp->nat > 1)
{
if (pgrp->pbcatom >= 0)
{
pgrp->epgrppbc = epgrppbcREFAT;
pull->bRefAt = TRUE;
}
else
{
if (pgrp->weight)
{
gmx_fatal(FARGS, "Pull groups can not have relative weights and cosine weighting at same time");
}
pgrp->epgrppbc = epgrppbcCOS;
if (pull->cosdim >= 0 && pull->cosdim != m)
{
gmx_fatal(FARGS, "Can only use cosine weighting with pulling in one dimension (use mdp option pull_dim)");
}
pull->cosdim = m;
}
}
}
/* Set the indices */
init_pull_group_index(fplog, cr, start, end, g, pgrp, pull->dim, mtop, ir, lambda);
if (PULL_CYL(pull) && pgrp->invtm == 0)
{
gmx_fatal(FARGS, "Can not have frozen atoms in a cylinder pull group");
}
}
else
{
/* Absolute reference, set the inverse mass to zero */
pgrp->invtm = 0;
pgrp->wscale = 1;
}
}
/* if we use dynamic reference groups, do some initialising for them */
if (PULL_CYL(pull))
{
if (ir->ePull == epullCONSTRAINT && pull->ncoord > 1)
{
/* We can't easily update the single reference group with multiple
* constraints. This would require recalculating COMs.
*/
gmx_fatal(FARGS, "Constraint COM pulling supports only one coordinate with geometry=cylinder, you can use umbrella pulling with multiple coordinates");
}
for (c = 0; c < pull->ncoord; c++)
{
if (pull->group[pull->coord[c].group[0]].nat == 0)
{
gmx_fatal(FARGS, "Dynamic reference groups are not supported when using absolute reference!\n");
}
}
snew(pull->dyna, pull->ncoord);
}
/* Only do I/O when we are doing dynamics and if we are the MASTER */
pull->out_x = NULL;
pull->out_f = NULL;
if (bOutFile)
{
if (pull->nstxout > 0)
{
pull->out_x = open_pull_out(opt2fn("-px", nfile, fnm), pull, oenv, TRUE, Flags);
}
if (pull->nstfout > 0)
{
pull->out_f = open_pull_out(opt2fn("-pf", nfile, fnm), pull, oenv,
FALSE, Flags);
}
}
}
gmx_bool constrain(FILE *fplog,gmx_bool bLog,gmx_bool bEner,
struct gmx_constr *constr,
t_idef *idef,t_inputrec *ir,gmx_ekindata_t *ekind,
t_commrec *cr,
gmx_large_int_t step,int delta_step,
t_mdatoms *md,
rvec *x,rvec *xprime,rvec *min_proj,
gmx_bool bMolPBC,matrix box,
real lambda,real *dvdlambda,
rvec *v,tensor *vir,
t_nrnb *nrnb,int econq,gmx_bool bPscal,
real veta, real vetanew)
{
gmx_bool bOK,bDump;
int start,homenr,nrend;
int i,j,d;
int ncons,settle_error;
tensor vir_r_m_dr;
rvec *vstor;
real invdt,vir_fac,t;
t_ilist *settle;
int nsettle;
t_pbc pbc,*pbc_null;
char buf[22];
t_vetavars vetavar;
int nth,th;
if (econq == econqForceDispl && !EI_ENERGY_MINIMIZATION(ir->eI))
{
gmx_incons("constrain called for forces displacements while not doing energy minimization, can not do this while the LINCS and SETTLE constraint connection matrices are mass weighted");
}
bOK = TRUE;
bDump = FALSE;
start = md->start;
homenr = md->homenr;
nrend = start+homenr;
/* set constants for pressure control integration */
init_vetavars(&vetavar,econq!=econqCoord,
veta,vetanew,ir,ekind,bPscal);
if (ir->delta_t == 0)
{
invdt = 0;
}
else
{
invdt = 1/ir->delta_t;
}
if (ir->efep != efepNO && EI_DYNAMICS(ir->eI))
{
/* Set the constraint lengths for the step at which this configuration
* is meant to be. The invmasses should not be changed.
*/
lambda += delta_step*ir->fepvals->delta_lambda;
}
if (vir != NULL)
{
clear_mat(vir_r_m_dr);
}
where();
settle = &idef->il[F_SETTLE];
nsettle = settle->nr/(1+NRAL(F_SETTLE));
if (nsettle > 0)
{
nth = gmx_omp_nthreads_get(emntSETTLE);
}
else
{
nth = 1;
}
if (nth > 1 && constr->vir_r_m_dr_th == NULL)
{
snew(constr->vir_r_m_dr_th,nth);
snew(constr->settle_error,nth);
}
settle_error = -1;
/* We do not need full pbc when constraints do not cross charge groups,
* i.e. when dd->constraint_comm==NULL.
* Note that PBC for constraints is different from PBC for bondeds.
* For constraints there is both forward and backward communication.
*/
if (ir->ePBC != epbcNONE &&
(cr->dd || bMolPBC) && !(cr->dd && cr->dd->constraint_comm==NULL))
{
/* With pbc=screw the screw has been changed to a shift
* by the constraint coordinate communication routine,
* so that here we can use normal pbc.
*/
pbc_null = set_pbc_dd(&pbc,ir->ePBC,cr->dd,FALSE,box);
}
else
{
pbc_null = NULL;
}
/* Communicate the coordinates required for the non-local constraints
* for LINCS and/or SETTLE.
*/
if (cr->dd)
{
dd_move_x_constraints(cr->dd,box,x,xprime);
}
else if (PARTDECOMP(cr))
{
pd_move_x_constraints(cr,x,xprime);
}
if (constr->lincsd != NULL)
{
bOK = constrain_lincs(fplog,bLog,bEner,ir,step,constr->lincsd,md,cr,
x,xprime,min_proj,
box,pbc_null,lambda,dvdlambda,
invdt,v,vir!=NULL,vir_r_m_dr,
econq,nrnb,
constr->maxwarn,&constr->warncount_lincs);
if (!bOK && constr->maxwarn >= 0)
{
if (fplog != NULL)
{
fprintf(fplog,"Constraint error in algorithm %s at step %s\n",
econstr_names[econtLINCS],gmx_step_str(step,buf));
}
bDump = TRUE;
}
}
if (constr->nblocks > 0)
{
switch (econq) {
case (econqCoord):
bOK = bshakef(fplog,constr->shaked,
homenr,md->invmass,constr->nblocks,constr->sblock,
idef,ir,x,xprime,nrnb,
constr->lagr,lambda,dvdlambda,
invdt,v,vir!=NULL,vir_r_m_dr,
constr->maxwarn>=0,econq,&vetavar);
break;
case (econqVeloc):
bOK = bshakef(fplog,constr->shaked,
homenr,md->invmass,constr->nblocks,constr->sblock,
idef,ir,x,min_proj,nrnb,
constr->lagr,lambda,dvdlambda,
invdt,NULL,vir!=NULL,vir_r_m_dr,
constr->maxwarn>=0,econq,&vetavar);
break;
default:
gmx_fatal(FARGS,"Internal error, SHAKE called for constraining something else than coordinates");
break;
}
if (!bOK && constr->maxwarn >= 0)
{
if (fplog != NULL)
{
fprintf(fplog,"Constraint error in algorithm %s at step %s\n",
econstr_names[econtSHAKE],gmx_step_str(step,buf));
}
bDump = TRUE;
}
}
if (nsettle > 0)
{
int calcvir_atom_end;
if (vir == NULL)
{
calcvir_atom_end = 0;
}
else
{
calcvir_atom_end = md->start + md->homenr;
}
switch (econq)
{
case econqCoord:
#pragma omp parallel for num_threads(nth) schedule(static)
for(th=0; th<nth; th++)
{
int start_th,end_th;
if (th > 0)
{
clear_mat(constr->vir_r_m_dr_th[th]);
}
start_th = (nsettle* th )/nth;
end_th = (nsettle*(th+1))/nth;
if (start_th >= 0 && end_th - start_th > 0)
{
csettle(constr->settled,
end_th-start_th,
settle->iatoms+start_th*(1+NRAL(F_SETTLE)),
pbc_null,
x[0],xprime[0],
invdt,v?v[0]:NULL,calcvir_atom_end,
th == 0 ? vir_r_m_dr : constr->vir_r_m_dr_th[th],
th == 0 ? &settle_error : &constr->settle_error[th],
&vetavar);
}
}
inc_nrnb(nrnb,eNR_SETTLE,nsettle);
if (v != NULL)
{
inc_nrnb(nrnb,eNR_CONSTR_V,nsettle*3);
}
if (vir != NULL)
{
inc_nrnb(nrnb,eNR_CONSTR_VIR,nsettle*3);
}
break;
case econqVeloc:
case econqDeriv:
case econqForce:
case econqForceDispl:
#pragma omp parallel for num_threads(nth) schedule(static)
for(th=0; th<nth; th++)
{
int start_th,end_th;
if (th > 0)
{
clear_mat(constr->vir_r_m_dr_th[th]);
}
start_th = (nsettle* th )/nth;
end_th = (nsettle*(th+1))/nth;
if (start_th >= 0 && end_th - start_th > 0)
{
settle_proj(fplog,constr->settled,econq,
end_th-start_th,
settle->iatoms+start_th*(1+NRAL(F_SETTLE)),
pbc_null,
x,
xprime,min_proj,calcvir_atom_end,
th == 0 ? vir_r_m_dr : constr->vir_r_m_dr_th[th],
&vetavar);
}
}
/* This is an overestimate */
inc_nrnb(nrnb,eNR_SETTLE,nsettle);
break;
case econqDeriv_FlexCon:
/* Nothing to do, since the are no flexible constraints in settles */
break;
default:
gmx_incons("Unknown constraint quantity for settle");
}
}
if (settle->nr > 0)
{
/* Combine virial and error info of the other threads */
for(i=1; i<nth; i++)
{
m_add(vir_r_m_dr,constr->vir_r_m_dr_th[i],vir_r_m_dr);
settle_error = constr->settle_error[i];
}
if (econq == econqCoord && settle_error >= 0)
{
bOK = FALSE;
if (constr->maxwarn >= 0)
{
char buf[256];
sprintf(buf,
"\nstep " gmx_large_int_pfmt ": Water molecule starting at atom %d can not be "
"settled.\nCheck for bad contacts and/or reduce the timestep if appropriate.\n",
step,ddglatnr(cr->dd,settle->iatoms[settle_error*(1+NRAL(F_SETTLE))+1]));
if (fplog)
{
fprintf(fplog,"%s",buf);
}
fprintf(stderr,"%s",buf);
constr->warncount_settle++;
if (constr->warncount_settle > constr->maxwarn)
{
too_many_constraint_warnings(-1,constr->warncount_settle);
}
bDump = TRUE;
}
}
}
free_vetavars(&vetavar);
if (vir != NULL)
{
switch (econq)
{
case econqCoord:
vir_fac = 0.5/(ir->delta_t*ir->delta_t);
break;
case econqVeloc:
vir_fac = 0.5/ir->delta_t;
break;
case econqForce:
case econqForceDispl:
vir_fac = 0.5;
break;
default:
vir_fac = 0;
gmx_incons("Unsupported constraint quantity for virial");
}
if (EI_VV(ir->eI))
{
vir_fac *= 2; /* only constraining over half the distance here */
}
for(i=0; i<DIM; i++)
{
for(j=0; j<DIM; j++)
{
(*vir)[i][j] = vir_fac*vir_r_m_dr[i][j];
}
}
}
if (bDump)
{
dump_confs(fplog,step,constr->warn_mtop,start,homenr,cr,x,xprime,box);
}
if (econq == econqCoord)
{
if (ir->ePull == epullCONSTRAINT)
{
if (EI_DYNAMICS(ir->eI))
{
t = ir->init_t + (step + delta_step)*ir->delta_t;
}
else
{
t = ir->init_t;
}
set_pbc(&pbc,ir->ePBC,box);
pull_constraint(ir->pull,md,&pbc,cr,ir->delta_t,t,x,xprime,v,*vir);
}
if (constr->ed && delta_step > 0)
{
/* apply the essential dynamcs constraints here */
do_edsam(ir,step,md,cr,xprime,v,box,constr->ed);
}
}
return bOK;
}
int
calc_gb_rad_still_sse2_double(t_commrec *cr, t_forcerec *fr,
int natoms, gmx_localtop_t *top,
const t_atomtypes *atype, double *x, t_nblist *nl,
gmx_genborn_t *born)
{
int i,k,n,ii,is3,ii3,nj0,nj1,offset;
int jnrA,jnrB,j3A,j3B;
int *mdtype;
double shX,shY,shZ;
int *jjnr;
double *shiftvec;
double gpi_ai,gpi2;
double factor;
double *gb_radius;
double *vsolv;
double *work;
double *dadx;
__m128d ix,iy,iz;
__m128d jx,jy,jz;
__m128d dx,dy,dz;
__m128d tx,ty,tz;
__m128d rsq,rinv,rinv2,rinv4,rinv6;
__m128d ratio,gpi,rai,raj,vai,vaj,rvdw;
__m128d ccf,dccf,theta,cosq,term,sinq,res,prod,prod_ai,tmp;
__m128d mask,icf4,icf6,mask_cmp;
const __m128d half = _mm_set1_pd(0.5);
const __m128d three = _mm_set1_pd(3.0);
const __m128d one = _mm_set1_pd(1.0);
const __m128d two = _mm_set1_pd(2.0);
const __m128d zero = _mm_set1_pd(0.0);
const __m128d four = _mm_set1_pd(4.0);
const __m128d still_p5inv = _mm_set1_pd(STILL_P5INV);
const __m128d still_pip5 = _mm_set1_pd(STILL_PIP5);
const __m128d still_p4 = _mm_set1_pd(STILL_P4);
factor = 0.5 * ONE_4PI_EPS0;
gb_radius = born->gb_radius;
vsolv = born->vsolv;
work = born->gpol_still_work;
jjnr = nl->jjnr;
shiftvec = fr->shift_vec[0];
dadx = fr->dadx;
jnrA = jnrB = 0;
jx = _mm_setzero_pd();
jy = _mm_setzero_pd();
jz = _mm_setzero_pd();
n = 0;
for(i=0;i<natoms;i++)
{
work[i]=0;
}
for(i=0;i<nl->nri;i++)
{
ii = nl->iinr[i];
ii3 = ii*3;
is3 = 3*nl->shift[i];
shX = shiftvec[is3];
shY = shiftvec[is3+1];
shZ = shiftvec[is3+2];
nj0 = nl->jindex[i];
nj1 = nl->jindex[i+1];
ix = _mm_set1_pd(shX+x[ii3+0]);
iy = _mm_set1_pd(shY+x[ii3+1]);
iz = _mm_set1_pd(shZ+x[ii3+2]);
/* Polarization energy for atom ai */
gpi = _mm_setzero_pd();
rai = _mm_load1_pd(gb_radius+ii);
prod_ai = _mm_set1_pd(STILL_P4*vsolv[ii]);
for(k=nj0;k<nj1-1;k+=2)
{
jnrA = jjnr[k];
jnrB = jjnr[k+1];
j3A = 3*jnrA;
j3B = 3*jnrB;
GMX_MM_LOAD_1RVEC_2POINTERS_PD(x+j3A,x+j3B,jx,jy,jz);
GMX_MM_LOAD_2VALUES_PD(gb_radius+jnrA,gb_radius+jnrB,raj);
GMX_MM_LOAD_2VALUES_PD(vsolv+jnrA,vsolv+jnrB,vaj);
dx = _mm_sub_pd(ix,jx);
dy = _mm_sub_pd(iy,jy);
dz = _mm_sub_pd(iz,jz);
rsq = gmx_mm_calc_rsq_pd(dx,dy,dz);
rinv = gmx_mm_invsqrt_pd(rsq);
rinv2 = _mm_mul_pd(rinv,rinv);
rinv4 = _mm_mul_pd(rinv2,rinv2);
rinv6 = _mm_mul_pd(rinv4,rinv2);
rvdw = _mm_add_pd(rai,raj);
ratio = _mm_mul_pd(rsq, gmx_mm_inv_pd( _mm_mul_pd(rvdw,rvdw)));
mask_cmp = _mm_cmple_pd(ratio,still_p5inv);
/* gmx_mm_sincos_pd() is quite expensive, so avoid calculating it if we can! */
if( 0 == _mm_movemask_pd(mask_cmp) )
{
/* if ratio>still_p5inv for ALL elements */
ccf = one;
dccf = _mm_setzero_pd();
}
else
{
ratio = _mm_min_pd(ratio,still_p5inv);
theta = _mm_mul_pd(ratio,still_pip5);
gmx_mm_sincos_pd(theta,&sinq,&cosq);
term = _mm_mul_pd(half,_mm_sub_pd(one,cosq));
ccf = _mm_mul_pd(term,term);
dccf = _mm_mul_pd(_mm_mul_pd(two,term),
_mm_mul_pd(sinq,theta));
}
prod = _mm_mul_pd(still_p4,vaj);
icf4 = _mm_mul_pd(ccf,rinv4);
icf6 = _mm_mul_pd( _mm_sub_pd( _mm_mul_pd(four,ccf),dccf), rinv6);
GMX_MM_INCREMENT_2VALUES_PD(work+jnrA,work+jnrB,_mm_mul_pd(prod_ai,icf4));
gpi = _mm_add_pd(gpi, _mm_mul_pd(prod,icf4) );
_mm_store_pd(dadx,_mm_mul_pd(prod,icf6));
dadx+=2;
_mm_store_pd(dadx,_mm_mul_pd(prod_ai,icf6));
dadx+=2;
}
if(k<nj1)
{
jnrA = jjnr[k];
j3A = 3*jnrA;
GMX_MM_LOAD_1RVEC_1POINTER_PD(x+j3A,jx,jy,jz);
GMX_MM_LOAD_1VALUE_PD(gb_radius+jnrA,raj);
GMX_MM_LOAD_1VALUE_PD(vsolv+jnrA,vaj);
dx = _mm_sub_sd(ix,jx);
dy = _mm_sub_sd(iy,jy);
dz = _mm_sub_sd(iz,jz);
rsq = gmx_mm_calc_rsq_pd(dx,dy,dz);
rinv = gmx_mm_invsqrt_pd(rsq);
rinv2 = _mm_mul_sd(rinv,rinv);
rinv4 = _mm_mul_sd(rinv2,rinv2);
rinv6 = _mm_mul_sd(rinv4,rinv2);
rvdw = _mm_add_sd(rai,raj);
ratio = _mm_mul_sd(rsq, gmx_mm_inv_pd( _mm_mul_pd(rvdw,rvdw)));
mask_cmp = _mm_cmple_sd(ratio,still_p5inv);
/* gmx_mm_sincos_pd() is quite expensive, so avoid calculating it if we can! */
if( 0 == _mm_movemask_pd(mask_cmp) )
{
/* if ratio>still_p5inv for ALL elements */
ccf = one;
dccf = _mm_setzero_pd();
}
else
{
ratio = _mm_min_sd(ratio,still_p5inv);
theta = _mm_mul_sd(ratio,still_pip5);
gmx_mm_sincos_pd(theta,&sinq,&cosq);
term = _mm_mul_sd(half,_mm_sub_sd(one,cosq));
ccf = _mm_mul_sd(term,term);
dccf = _mm_mul_sd(_mm_mul_sd(two,term),
_mm_mul_sd(sinq,theta));
}
prod = _mm_mul_sd(still_p4,vaj);
icf4 = _mm_mul_sd(ccf,rinv4);
icf6 = _mm_mul_sd( _mm_sub_sd( _mm_mul_sd(four,ccf),dccf), rinv6);
GMX_MM_INCREMENT_1VALUE_PD(work+jnrA,_mm_mul_sd(prod_ai,icf4));
gpi = _mm_add_sd(gpi, _mm_mul_sd(prod,icf4) );
_mm_store_pd(dadx,_mm_mul_pd(prod,icf6));
dadx+=2;
_mm_store_pd(dadx,_mm_mul_pd(prod_ai,icf6));
dadx+=2;
}
gmx_mm_update_1pot_pd(gpi,work+ii);
}
/* Sum up the polarization energy from other nodes */
if(PARTDECOMP(cr))
{
gmx_sum(natoms, work, cr);
}
else if(DOMAINDECOMP(cr))
{
dd_atom_sum_real(cr->dd, work);
}
/* Compute the radii */
for(i=0;i<fr->natoms_force;i++) /* PELA born->nr */
{
if(born->use[i] != 0)
{
gpi_ai = born->gpol[i] + work[i]; /* add gpi to the initial pol energy gpi_ai*/
gpi2 = gpi_ai * gpi_ai;
born->bRad[i] = factor*gmx_invsqrt(gpi2);
fr->invsqrta[i] = gmx_invsqrt(born->bRad[i]);
}
}
/* Extra (local) communication required for DD */
if(DOMAINDECOMP(cr))
{
dd_atom_spread_real(cr->dd, born->bRad);
dd_atom_spread_real(cr->dd, fr->invsqrta);
}
return 0;
}
int
calc_gb_rad_hct_obc_sse2_double(t_commrec *cr, t_forcerec * fr, int natoms, gmx_localtop_t *top,
const t_atomtypes *atype, double *x, t_nblist *nl, gmx_genborn_t *born,t_mdatoms *md,int gb_algorithm)
{
int i,ai,k,n,ii,ii3,is3,nj0,nj1,at0,at1,offset;
int jnrA,jnrB;
int j3A,j3B;
double shX,shY,shZ;
double rr,rr_inv,rr_inv2,sum_tmp,sum,sum2,sum3,gbr;
double sum_ai2, sum_ai3,tsum,tchain,doffset;
double *obc_param;
double *gb_radius;
double *work;
int * jjnr;
double *dadx;
double *shiftvec;
double min_rad,rad;
__m128d ix,iy,iz,jx,jy,jz;
__m128d dx,dy,dz,t1,t2,t3,t4;
__m128d rsq,rinv,r;
__m128d rai,rai_inv,raj, raj_inv,rai_inv2,sk,sk2,lij,dlij,duij;
__m128d uij,lij2,uij2,lij3,uij3,diff2;
__m128d lij_inv,sk2_inv,prod,log_term,tmp,tmp_sum;
__m128d sum_ai, tmp_ai,sk_ai,sk_aj,sk2_ai,sk2_aj,sk2_rinv;
__m128d dadx1,dadx2;
__m128d logterm;
__m128d mask;
__m128d obc_mask1,obc_mask2,obc_mask3;
__m128d oneeighth = _mm_set1_pd(0.125);
__m128d onefourth = _mm_set1_pd(0.25);
const __m128d half = _mm_set1_pd(0.5);
const __m128d three = _mm_set1_pd(3.0);
const __m128d one = _mm_set1_pd(1.0);
const __m128d two = _mm_set1_pd(2.0);
const __m128d zero = _mm_set1_pd(0.0);
const __m128d neg = _mm_set1_pd(-1.0);
/* Set the dielectric offset */
doffset = born->gb_doffset;
gb_radius = born->gb_radius;
obc_param = born->param;
work = born->gpol_hct_work;
jjnr = nl->jjnr;
dadx = fr->dadx;
shiftvec = fr->shift_vec[0];
jx = _mm_setzero_pd();
jy = _mm_setzero_pd();
jz = _mm_setzero_pd();
jnrA = jnrB = 0;
for(i=0;i<born->nr;i++)
{
work[i] = 0;
}
for(i=0;i<nl->nri;i++)
{
ii = nl->iinr[i];
ii3 = ii*3;
is3 = 3*nl->shift[i];
shX = shiftvec[is3];
shY = shiftvec[is3+1];
shZ = shiftvec[is3+2];
nj0 = nl->jindex[i];
nj1 = nl->jindex[i+1];
ix = _mm_set1_pd(shX+x[ii3+0]);
iy = _mm_set1_pd(shY+x[ii3+1]);
iz = _mm_set1_pd(shZ+x[ii3+2]);
rai = _mm_load1_pd(gb_radius+ii);
rai_inv= gmx_mm_inv_pd(rai);
sum_ai = _mm_setzero_pd();
sk_ai = _mm_load1_pd(born->param+ii);
sk2_ai = _mm_mul_pd(sk_ai,sk_ai);
for(k=nj0;k<nj1-1;k+=2)
{
jnrA = jjnr[k];
jnrB = jjnr[k+1];
j3A = 3*jnrA;
j3B = 3*jnrB;
GMX_MM_LOAD_1RVEC_2POINTERS_PD(x+j3A,x+j3B,jx,jy,jz);
GMX_MM_LOAD_2VALUES_PD(gb_radius+jnrA,gb_radius+jnrB,raj);
GMX_MM_LOAD_2VALUES_PD(obc_param+jnrA,obc_param+jnrB,sk_aj);
dx = _mm_sub_pd(ix, jx);
dy = _mm_sub_pd(iy, jy);
dz = _mm_sub_pd(iz, jz);
rsq = gmx_mm_calc_rsq_pd(dx,dy,dz);
rinv = gmx_mm_invsqrt_pd(rsq);
r = _mm_mul_pd(rsq,rinv);
/* Compute raj_inv aj1-4 */
raj_inv = gmx_mm_inv_pd(raj);
/* Evaluate influence of atom aj -> ai */
t1 = _mm_add_pd(r,sk_aj);
t2 = _mm_sub_pd(r,sk_aj);
t3 = _mm_sub_pd(sk_aj,r);
obc_mask1 = _mm_cmplt_pd(rai, t1);
obc_mask2 = _mm_cmplt_pd(rai, t2);
obc_mask3 = _mm_cmplt_pd(rai, t3);
uij = gmx_mm_inv_pd(t1);
lij = _mm_or_pd( _mm_and_pd(obc_mask2,gmx_mm_inv_pd(t2)),
_mm_andnot_pd(obc_mask2,rai_inv));
dlij = _mm_and_pd(one,obc_mask2);
uij2 = _mm_mul_pd(uij, uij);
uij3 = _mm_mul_pd(uij2,uij);
lij2 = _mm_mul_pd(lij, lij);
lij3 = _mm_mul_pd(lij2,lij);
diff2 = _mm_sub_pd(uij2,lij2);
lij_inv = gmx_mm_invsqrt_pd(lij2);
sk2_aj = _mm_mul_pd(sk_aj,sk_aj);
sk2_rinv = _mm_mul_pd(sk2_aj,rinv);
prod = _mm_mul_pd(onefourth,sk2_rinv);
logterm = gmx_mm_log_pd(_mm_mul_pd(uij,lij_inv));
t1 = _mm_sub_pd(lij,uij);
t2 = _mm_mul_pd(diff2,
_mm_sub_pd(_mm_mul_pd(onefourth,r),
prod));
t3 = _mm_mul_pd(half,_mm_mul_pd(rinv,logterm));
t1 = _mm_add_pd(t1,_mm_add_pd(t2,t3));
t4 = _mm_mul_pd(two,_mm_sub_pd(rai_inv,lij));
t4 = _mm_and_pd(t4,obc_mask3);
t1 = _mm_mul_pd(half,_mm_add_pd(t1,t4));
sum_ai = _mm_add_pd(sum_ai, _mm_and_pd(t1,obc_mask1) );
t1 = _mm_add_pd(_mm_mul_pd(half,lij2),
_mm_mul_pd(prod,lij3));
t1 = _mm_sub_pd(t1,
_mm_mul_pd(onefourth,
_mm_add_pd(_mm_mul_pd(lij,rinv),
_mm_mul_pd(lij3,r))));
t2 = _mm_mul_pd(onefourth,
_mm_add_pd(_mm_mul_pd(uij,rinv),
_mm_mul_pd(uij3,r)));
t2 = _mm_sub_pd(t2,
_mm_add_pd(_mm_mul_pd(half,uij2),
_mm_mul_pd(prod,uij3)));
t3 = _mm_mul_pd(_mm_mul_pd(onefourth,logterm),
_mm_mul_pd(rinv,rinv));
t3 = _mm_sub_pd(t3,
_mm_mul_pd(_mm_mul_pd(diff2,oneeighth),
_mm_add_pd(one,
_mm_mul_pd(sk2_rinv,rinv))));
t1 = _mm_mul_pd(rinv,
_mm_add_pd(_mm_mul_pd(dlij,t1),
_mm_add_pd(t2,t3)));
dadx1 = _mm_and_pd(t1,obc_mask1);
/* Evaluate influence of atom ai -> aj */
t1 = _mm_add_pd(r,sk_ai);
t2 = _mm_sub_pd(r,sk_ai);
t3 = _mm_sub_pd(sk_ai,r);
obc_mask1 = _mm_cmplt_pd(raj, t1);
obc_mask2 = _mm_cmplt_pd(raj, t2);
obc_mask3 = _mm_cmplt_pd(raj, t3);
uij = gmx_mm_inv_pd(t1);
lij = _mm_or_pd( _mm_and_pd(obc_mask2,gmx_mm_inv_pd(t2)),
_mm_andnot_pd(obc_mask2,raj_inv));
dlij = _mm_and_pd(one,obc_mask2);
uij2 = _mm_mul_pd(uij, uij);
uij3 = _mm_mul_pd(uij2,uij);
lij2 = _mm_mul_pd(lij, lij);
lij3 = _mm_mul_pd(lij2,lij);
diff2 = _mm_sub_pd(uij2,lij2);
lij_inv = gmx_mm_invsqrt_pd(lij2);
sk2_rinv = _mm_mul_pd(sk2_ai,rinv);
prod = _mm_mul_pd(onefourth,sk2_rinv);
logterm = gmx_mm_log_pd(_mm_mul_pd(uij,lij_inv));
t1 = _mm_sub_pd(lij,uij);
t2 = _mm_mul_pd(diff2,
_mm_sub_pd(_mm_mul_pd(onefourth,r),
prod));
t3 = _mm_mul_pd(half,_mm_mul_pd(rinv,logterm));
t1 = _mm_add_pd(t1,_mm_add_pd(t2,t3));
t4 = _mm_mul_pd(two,_mm_sub_pd(raj_inv,lij));
t4 = _mm_and_pd(t4,obc_mask3);
t1 = _mm_mul_pd(half,_mm_add_pd(t1,t4));
GMX_MM_INCREMENT_2VALUES_PD(work+jnrA,work+jnrB,_mm_and_pd(t1,obc_mask1));
t1 = _mm_add_pd(_mm_mul_pd(half,lij2),
_mm_mul_pd(prod,lij3));
t1 = _mm_sub_pd(t1,
_mm_mul_pd(onefourth,
_mm_add_pd(_mm_mul_pd(lij,rinv),
_mm_mul_pd(lij3,r))));
t2 = _mm_mul_pd(onefourth,
_mm_add_pd(_mm_mul_pd(uij,rinv),
_mm_mul_pd(uij3,r)));
t2 = _mm_sub_pd(t2,
_mm_add_pd(_mm_mul_pd(half,uij2),
_mm_mul_pd(prod,uij3)));
t3 = _mm_mul_pd(_mm_mul_pd(onefourth,logterm),
_mm_mul_pd(rinv,rinv));
t3 = _mm_sub_pd(t3,
_mm_mul_pd(_mm_mul_pd(diff2,oneeighth),
_mm_add_pd(one,
_mm_mul_pd(sk2_rinv,rinv))));
t1 = _mm_mul_pd(rinv,
_mm_add_pd(_mm_mul_pd(dlij,t1),
_mm_add_pd(t2,t3)));
dadx2 = _mm_and_pd(t1,obc_mask1);
_mm_store_pd(dadx,dadx1);
dadx += 2;
_mm_store_pd(dadx,dadx2);
dadx += 2;
} /* end normal inner loop */
if(k<nj1)
{
jnrA = jjnr[k];
j3A = 3*jnrA;
GMX_MM_LOAD_1RVEC_1POINTER_PD(x+j3A,jx,jy,jz);
GMX_MM_LOAD_1VALUE_PD(gb_radius+jnrA,raj);
GMX_MM_LOAD_1VALUE_PD(obc_param+jnrA,sk_aj);
dx = _mm_sub_sd(ix, jx);
dy = _mm_sub_sd(iy, jy);
dz = _mm_sub_sd(iz, jz);
rsq = gmx_mm_calc_rsq_pd(dx,dy,dz);
rinv = gmx_mm_invsqrt_pd(rsq);
r = _mm_mul_sd(rsq,rinv);
/* Compute raj_inv aj1-4 */
raj_inv = gmx_mm_inv_pd(raj);
/* Evaluate influence of atom aj -> ai */
t1 = _mm_add_sd(r,sk_aj);
t2 = _mm_sub_sd(r,sk_aj);
t3 = _mm_sub_sd(sk_aj,r);
obc_mask1 = _mm_cmplt_sd(rai, t1);
obc_mask2 = _mm_cmplt_sd(rai, t2);
obc_mask3 = _mm_cmplt_sd(rai, t3);
uij = gmx_mm_inv_pd(t1);
lij = _mm_or_pd(_mm_and_pd(obc_mask2,gmx_mm_inv_pd(t2)),
_mm_andnot_pd(obc_mask2,rai_inv));
dlij = _mm_and_pd(one,obc_mask2);
uij2 = _mm_mul_sd(uij, uij);
uij3 = _mm_mul_sd(uij2,uij);
lij2 = _mm_mul_sd(lij, lij);
lij3 = _mm_mul_sd(lij2,lij);
diff2 = _mm_sub_sd(uij2,lij2);
lij_inv = gmx_mm_invsqrt_pd(lij2);
sk2_aj = _mm_mul_sd(sk_aj,sk_aj);
sk2_rinv = _mm_mul_sd(sk2_aj,rinv);
prod = _mm_mul_sd(onefourth,sk2_rinv);
logterm = gmx_mm_log_pd(_mm_mul_sd(uij,lij_inv));
t1 = _mm_sub_sd(lij,uij);
t2 = _mm_mul_sd(diff2,
_mm_sub_sd(_mm_mul_pd(onefourth,r),
prod));
t3 = _mm_mul_sd(half,_mm_mul_sd(rinv,logterm));
t1 = _mm_add_sd(t1,_mm_add_sd(t2,t3));
t4 = _mm_mul_sd(two,_mm_sub_sd(rai_inv,lij));
t4 = _mm_and_pd(t4,obc_mask3);
t1 = _mm_mul_sd(half,_mm_add_sd(t1,t4));
sum_ai = _mm_add_sd(sum_ai, _mm_and_pd(t1,obc_mask1) );
t1 = _mm_add_sd(_mm_mul_sd(half,lij2),
_mm_mul_sd(prod,lij3));
t1 = _mm_sub_sd(t1,
_mm_mul_sd(onefourth,
_mm_add_sd(_mm_mul_sd(lij,rinv),
_mm_mul_sd(lij3,r))));
t2 = _mm_mul_sd(onefourth,
_mm_add_sd(_mm_mul_sd(uij,rinv),
_mm_mul_sd(uij3,r)));
t2 = _mm_sub_sd(t2,
_mm_add_sd(_mm_mul_sd(half,uij2),
_mm_mul_sd(prod,uij3)));
t3 = _mm_mul_sd(_mm_mul_sd(onefourth,logterm),
_mm_mul_sd(rinv,rinv));
t3 = _mm_sub_sd(t3,
_mm_mul_sd(_mm_mul_sd(diff2,oneeighth),
_mm_add_sd(one,
_mm_mul_sd(sk2_rinv,rinv))));
t1 = _mm_mul_sd(rinv,
_mm_add_sd(_mm_mul_sd(dlij,t1),
_mm_add_pd(t2,t3)));
dadx1 = _mm_and_pd(t1,obc_mask1);
/* Evaluate influence of atom ai -> aj */
t1 = _mm_add_sd(r,sk_ai);
t2 = _mm_sub_sd(r,sk_ai);
t3 = _mm_sub_sd(sk_ai,r);
obc_mask1 = _mm_cmplt_sd(raj, t1);
obc_mask2 = _mm_cmplt_sd(raj, t2);
obc_mask3 = _mm_cmplt_sd(raj, t3);
uij = gmx_mm_inv_pd(t1);
lij = _mm_or_pd( _mm_and_pd(obc_mask2,gmx_mm_inv_pd(t2)),
_mm_andnot_pd(obc_mask2,raj_inv));
dlij = _mm_and_pd(one,obc_mask2);
uij2 = _mm_mul_sd(uij, uij);
uij3 = _mm_mul_sd(uij2,uij);
lij2 = _mm_mul_sd(lij, lij);
lij3 = _mm_mul_sd(lij2,lij);
diff2 = _mm_sub_sd(uij2,lij2);
lij_inv = gmx_mm_invsqrt_pd(lij2);
sk2_rinv = _mm_mul_sd(sk2_ai,rinv);
prod = _mm_mul_sd(onefourth,sk2_rinv);
logterm = gmx_mm_log_pd(_mm_mul_sd(uij,lij_inv));
t1 = _mm_sub_sd(lij,uij);
t2 = _mm_mul_sd(diff2,
_mm_sub_sd(_mm_mul_sd(onefourth,r),
prod));
t3 = _mm_mul_sd(half,_mm_mul_sd(rinv,logterm));
t1 = _mm_add_sd(t1,_mm_add_sd(t2,t3));
t4 = _mm_mul_sd(two,_mm_sub_sd(raj_inv,lij));
t4 = _mm_and_pd(t4,obc_mask3);
t1 = _mm_mul_sd(half,_mm_add_sd(t1,t4));
GMX_MM_INCREMENT_1VALUE_PD(work+jnrA,_mm_and_pd(t1,obc_mask1));
t1 = _mm_add_sd(_mm_mul_sd(half,lij2),
_mm_mul_sd(prod,lij3));
t1 = _mm_sub_sd(t1,
_mm_mul_sd(onefourth,
_mm_add_sd(_mm_mul_sd(lij,rinv),
_mm_mul_sd(lij3,r))));
t2 = _mm_mul_sd(onefourth,
_mm_add_sd(_mm_mul_sd(uij,rinv),
_mm_mul_sd(uij3,r)));
t2 = _mm_sub_sd(t2,
_mm_add_sd(_mm_mul_sd(half,uij2),
_mm_mul_sd(prod,uij3)));
t3 = _mm_mul_sd(_mm_mul_sd(onefourth,logterm),
_mm_mul_sd(rinv,rinv));
t3 = _mm_sub_sd(t3,
_mm_mul_sd(_mm_mul_sd(diff2,oneeighth),
_mm_add_sd(one,
_mm_mul_sd(sk2_rinv,rinv))));
t1 = _mm_mul_sd(rinv,
_mm_add_sd(_mm_mul_sd(dlij,t1),
_mm_add_sd(t2,t3)));
dadx2 = _mm_and_pd(t1,obc_mask1);
_mm_store_pd(dadx,dadx1);
dadx += 2;
_mm_store_pd(dadx,dadx2);
dadx += 2;
}
gmx_mm_update_1pot_pd(sum_ai,work+ii);
}
/* Parallel summations */
if(PARTDECOMP(cr))
{
gmx_sum(natoms, work, cr);
}
else if(DOMAINDECOMP(cr))
{
dd_atom_sum_real(cr->dd, work);
}
if(gb_algorithm==egbHCT)
{
/* HCT */
for(i=0;i<fr->natoms_force;i++) /* PELA born->nr */
{
if(born->use[i] != 0)
{
rr = top->atomtypes.gb_radius[md->typeA[i]]-doffset;
sum = 1.0/rr - work[i];
min_rad = rr + doffset;
rad = 1.0/sum;
born->bRad[i] = rad > min_rad ? rad : min_rad;
fr->invsqrta[i] = gmx_invsqrt(born->bRad[i]);
}
}
/* Extra communication required for DD */
if(DOMAINDECOMP(cr))
{
dd_atom_spread_real(cr->dd, born->bRad);
dd_atom_spread_real(cr->dd, fr->invsqrta);
}
}
else
{
/* OBC */
for(i=0;i<fr->natoms_force;i++) /* PELA born->nr */
{
if(born->use[i] != 0)
{
rr = top->atomtypes.gb_radius[md->typeA[i]];
rr_inv2 = 1.0/rr;
rr = rr-doffset;
rr_inv = 1.0/rr;
sum = rr * work[i];
sum2 = sum * sum;
sum3 = sum2 * sum;
tsum = tanh(born->obc_alpha*sum-born->obc_beta*sum2+born->obc_gamma*sum3);
born->bRad[i] = rr_inv - tsum*rr_inv2;
born->bRad[i] = 1.0 / born->bRad[i];
fr->invsqrta[i]=gmx_invsqrt(born->bRad[i]);
tchain = rr * (born->obc_alpha-2*born->obc_beta*sum+3*born->obc_gamma*sum2);
born->drobc[i] = (1.0-tsum*tsum)*tchain*rr_inv2;
}
}
/* Extra (local) communication required for DD */
if(DOMAINDECOMP(cr))
{
dd_atom_spread_real(cr->dd, born->bRad);
dd_atom_spread_real(cr->dd, fr->invsqrta);
dd_atom_spread_real(cr->dd, born->drobc);
}
}
return 0;
}
void do_mmcg (int natoms, // number of atoms in simulation
t_inputrec *inputrec, // input record and box stuff
t_mdatoms *md, // the atoms
t_state *state, // positions & velocities
gmx_mtop_t *top, // global topology
t_commrec *cr, // communicators
rvec *cg_cm, // centre of mass of charge groups
int *allcgid, // charge groups ids
int allcgnr, // charge groups number
int *allsolid, // solvent groups ids
int allsolnr, // solvent groups number
FILE *log) // logfile
{
int i, j, p, q, qmin;
real dvmod, shell2w2, dmin, d;
rvec vecdist, dv, *v;
shell2w2 = pow (inputrec->mmcg.shell2wt, 2.0); // Threshold
v = state->v; // Velocities
t_block cgs; // charge groups
cgs = gmx_mtop_global_cgs(top);
rvec *all_cg_cm=NULL;
snew(all_cg_cm,allcgnr);
// for DD, we need to get cg_cm of all given charge groups (fr->cg_cm is local),
// the nearest one from a monitored water
// may be more than one DD cell distant.
if (DOMAINDECOMP(cr)) {
if(cr->nnodes!=1) gmx_barrier(cr);
for(i=0; i<allcgnr; i++) {
int sender = 0,senderf,k;
for(j=0; j<cr->dd->ncg_home; j++) { // is the cg a home cg ?
if(cr->dd->index_gl[j]==allcgid[i] ) {
sender = cr->sim_nodeid;
for(k=0;k<3;k++) all_cg_cm[i][k]=cg_cm[j][k];//FIXME improve ! compilation error when done with pointers
}
}
MPI_Allreduce(&sender,&senderf,1,MPI_INT,MPI_SUM,cr->dd->mpi_comm_all);
for(k=0;k<3;k++) MPI_Bcast(&all_cg_cm[i][k],sizeof(all_cg_cm[i][k]),MPI_BYTE,senderf,cr->dd->mpi_comm_all); //FIXME again !
}
}
for(i=0; i<allsolnr; i++) { // Loop over waters - START
p = allsolid[i];
if (PARTDECOMP(cr)) { // water i is in the node
if((cgs.index[p]>=md->start) && (cgs.index[p]<(md->start+md->homenr))) {
for (j=0; j<allcgnr; j++) { // Looking for min dist (dmin)
q = allcgid[j]; // and for the nearest charge group (qmin)
d = distance2(cg_cm[p],cg_cm[q]);
if(!j) { dmin = d; qmin = q; }
else if (d < dmin) { dmin = d; qmin = q; }
}
if (dmin >= shell2w2) { // Modifing velocity
rvec_sub(cg_cm[p], cg_cm[qmin], vecdist);
unitv (vecdist, vecdist);
for (j=cgs.index[p]; j<(3+cgs.index[p]); j++) {
dvmod = iprod (v[j], vecdist);
if (dvmod <= 0) continue;
svmul (2.0*dvmod, vecdist, dv);
rvec_dec (&v[j], dv); // Warning (=>?)
}
}
}
}
if (DOMAINDECOMP(cr)) {
int g_atnr; // global atom ID
int l_atnr; // local atom ID in the DD cell
int l_cgnr; // local charge group number
for(g_atnr=cgs.index[p];g_atnr<=cgs.index[p]+2;g_atnr++) {
if(ga2la_get_home(cr->dd->ga2la,g_atnr,&l_atnr)) {// search in global to local lookup table
// if the atom (not water) is in home atoms
// and get local atom number
l_cgnr = cr->dd->la2lc[l_atnr]; // get local charge group number
for (j=0; j<allcgnr; j++) { // Looking for min dist (dmin)
// and for the nearest charge group (qmin)
d = distance2(cg_cm[l_cgnr],all_cg_cm[j]);
if(!j) { dmin = d; qmin = j; }
else if (d < dmin) { dmin = d; qmin = j; }
}
if (dmin >= shell2w2) { // Modifing velocity
rvec_sub(cg_cm[l_cgnr], all_cg_cm[qmin], vecdist);
unitv (vecdist, vecdist);
dvmod = iprod (v[l_atnr], vecdist);
if (dvmod <= 0) continue;
svmul (2.0*dvmod, vecdist, dv);
rvec_dec (&v[l_atnr], dv); // Warning (=>?)
}
}
}
}
} // Loop over waters - END
return;
}
