static void pull_set_pbcatoms(t_commrec *cr, t_pull *pull,
t_mdatoms *md, rvec *x,
rvec *x_pbc)
{
int g, n, m;
n = 0;
for (g = 0; g < pull->ngroup; g++)
{
if ((g == 0 && PULL_CYL(pull)) || pull->group[g].pbcatom == -1)
{
clear_rvec(x_pbc[g]);
}
else
{
pull_set_pbcatom(cr, &pull->group[g], md, x, x_pbc[g]);
for (m = 0; m < DIM; m++)
{
if (pull->dim[m] == 0)
{
x_pbc[g][m] = 0.0;
}
}
n++;
}
}
if (cr && PAR(cr) && n > 0)
{
/* Sum over the nodes to get x_pbc from the home node of pbcatom */
gmx_sum(pull->ngroup*DIM, x_pbc[0], cr);
}
}
real Umbrella_Communicate(real Q_local,real *k_Q,real *Q_0)
{
int i;
int i_omp=gmx_omp_get_thread_num();
// int n_omp=gmx_omp_get_num_procs(); // Wrong number
int n_omp=udata.n_omp;
// int i_mpi;
static real Q_semilocal[UMB_MAX_OMP];
real Q_global;
if (n_omp>UMB_MAX_OMP) {
fprintf(stderr,"Seg fault is probably about to happen because Q_semilocal is not big enough to accommodate %d omp threads. See %d in %s.\n",n_omp,__LINE__,__FILE__);
}
Q_semilocal[i_omp]=Q_local;
#pragma omp barrier
// #pragma omp master
// {
if (i_omp==0) {
Q_local=0;
// fprintf(stderr,"%f %f %f %f %f\n",Q_semilocal[0],Q_semilocal[1],Q_semilocal[2],Q_semilocal[3],Q_semilocal[4]);
for (i=0; i<n_omp; i++) {
Q_local+=Q_semilocal[i];
}
// fprintf(stderr,"%f %f %f %f %f\n",Q_semilocal[0],Q_semilocal[1],Q_semilocal[2],Q_semilocal[3],Q_semilocal[4]);
// #ifdef GMX_DOUBLE
// MPI_Allreduce(&Q_local,&Q_global,1,MPI_DOUBLE,MPI_SUM,MPI_COMM_WORLD);
// #else
// MPI_Allreduce(&Q_local,&Q_global,1,MPI_FLOAT, MPI_SUM,MPI_COMM_WORLD);
// #endif
// for (i=0; i<n_omp; i++) {
// Q_semilocal[i]=Q_global;
// }
// gmx_sum declared in src/gromacs/legacyheaders/network.h (in main.h)
#ifdef GMX_MPI
gmx_sum(1,&Q_local,udata.cr);
#endif
for (i=0; i<n_omp; i++) {
Q_semilocal[i]=Q_local;
}
// fprintf(stderr,"%f %f %f %f %f\n",Q_semilocal[0],Q_semilocal[1],Q_semilocal[2],Q_semilocal[3],Q_semilocal[4]);
}
#pragma omp barrier
Q_global=Q_semilocal[i_omp];
*k_Q=udata.k_Q;
*Q_0=udata.Q_0+(udata.Q_init-udata.Q_0)*exp(-udata.step/udata.Q_steps);
// #pragma omp master
// {
if (i_omp==0) {
// MPI_Comm_rank(MPI_COMM_WORLD,&i_mpi);
// if (i_mpi==0) { // udata.fp=NULL on other mpi processes.
if ((udata.step % udata.freq)==0 && udata.fp!=NULL) {
fprintf(udata.fp,"%d %g %g\n",udata.step,Q_global,0.5*(*k_Q)*(Q_global-(*Q_0))*(Q_global-(*Q_0)));
}
// }
}
return Q_global;
}
static void pull_set_pbcatoms(t_commrec *cr, t_pull *pull,
rvec *x,
rvec *x_pbc)
{
int g, n, m;
n = 0;
for (g = 0; g < pull->ngroup; g++)
{
if (!pull->group[g].bCalcCOM || pull->group[g].pbcatom == -1)
{
clear_rvec(x_pbc[g]);
}
else
{
pull_set_pbcatom(cr, &pull->group[g], x, x_pbc[g]);
n++;
}
}
if (cr && PAR(cr) && n > 0)
{
/* Sum over the nodes to get x_pbc from the home node of pbcatom */
gmx_sum(pull->ngroup*DIM, x_pbc[0], cr);
}
}
/* Assemble the positions of the group such that every node has all of them.
* The atom indices are retrieved from anrs_loc[0..nr_loc]
* Note that coll_ind[i] = i is needed in the serial case */
extern void communicate_group_positions(
t_commrec *cr,
rvec *xcoll, /* OUT: Collective array of positions */
ivec *shifts, /* IN+OUT: Collective array of shifts for xcoll */
ivec *extra_shifts, /* BUF: Extra shifts since last time step */
const gmx_bool bNS, /* IN: NS step, the shifts have changed */
rvec *x_loc, /* IN: Local positions on this node */
const int nr, /* IN: Total number of atoms in the group */
const int nr_loc, /* IN: Local number of atoms in the group */
int *anrs_loc, /* IN: Local atom numbers */
int *coll_ind, /* IN: Collective index */
rvec *xcoll_old, /* IN+OUT: Positions from the last time step, used to make group whole */
matrix box)
{
int i;
/* Zero out the groups' global position array */
clear_rvecs(nr, xcoll);
/* Put the local positions that this node has into the right place of
* the collective array. Note that in the serial case, coll_ind[i] = i */
for (i=0; i<nr_loc; i++)
copy_rvec(x_loc[anrs_loc[i]], xcoll[coll_ind[i]]);
if (PAR(cr))
{
/* Add the arrays from all nodes together */
gmx_sum(nr*3, xcoll[0], cr);
}
/* To make the group whole, start with a whole group and each
* step move the assembled positions at closest distance to the positions
* from the last step. First shift the positions with the saved shift
* vectors (these are 0 when this routine is called for the first time!) */
shift_positions_group(box, xcoll, shifts, nr);
/* Now check if some shifts changed since the last step.
* This only needs to be done when the shifts are expected to have changed,
* i.e. after neighboursearching */
if (bNS)
{
get_shifts_group(3, box, xcoll, nr, xcoll_old, extra_shifts);
/* Shift with the additional shifts such that we get a whole group now */
shift_positions_group(box, xcoll, extra_shifts, nr);
/* Add the shift vectors together for the next time step */
for (i=0; i<nr; i++)
{
shifts[i][XX] += extra_shifts[i][XX];
shifts[i][YY] += extra_shifts[i][YY];
shifts[i][ZZ] += extra_shifts[i][ZZ];
}
/* Store current correctly-shifted positions for comparison in the next NS time step */
for (i=0; i<nr; i++)
copy_rvec(xcoll[i],xcoll_old[i]);
}
}
static void accumulate_ekin(t_commrec *cr,t_grpopts *opts,
gmx_ekindata_t *ekind)
{
int g;
if(PAR(cr))
for(g=0; (g<opts->ngtc); g++)
gmx_sum(DIM*DIM,ekind->tcstat[g].ekin[0],cr);
}
static void pull_reduce_real(t_commrec *cr,
pull_comm_t *comm,
int n,
real *data)
{
if (cr != NULL && PAR(cr))
{
if (comm->bParticipateAll)
{
/* Sum the contributions over all DD ranks */
gmx_sum(n, data, cr);
}
else
{
#if GMX_MPI
#if MPI_IN_PLACE_EXISTS
MPI_Allreduce(MPI_IN_PLACE, data, n, GMX_MPI_REAL, MPI_SUM,
comm->mpi_comm_com);
#else
real *buf;
snew(buf, n);
MPI_Allreduce(data, buf, n, GMX_MPI_REAL, MPI_SUM,
comm->mpi_comm_com);
/* Copy the result from the buffer to the input/output data */
for (int i = 0; i < n; i++)
{
data[i] = buf[i];
}
sfree(buf);
#endif
#else
gmx_incons("comm->bParticipateAll=FALSE without GMX_MPI");
#endif
}
}
}
/* Estimate the reciprocal space part error of the SPME Ewald sum. */
static real estimate_reciprocal(
t_inputinfo *info,
rvec x[], /* array of particles */
real q[], /* array of charges */
int nr, /* number of charges = size of the charge array */
FILE *fp_out,
gmx_bool bVerbose,
unsigned int seed, /* The seed for the random number generator */
int *nsamples, /* Return the number of samples used if Monte Carlo
* algorithm is used for self energy error estimate */
t_commrec *cr)
{
real e_rec=0; /* reciprocal error estimate */
real e_rec1=0; /* Error estimate term 1*/
real e_rec2=0; /* Error estimate term 2*/
real e_rec3=0; /* Error estimate term 3 */
real e_rec3x=0; /* part of Error estimate term 3 in x */
real e_rec3y=0; /* part of Error estimate term 3 in y */
real e_rec3z=0; /* part of Error estimate term 3 in z */
int i,ci;
int nx,ny,nz; /* grid coordinates */
real q2_all=0; /* sum of squared charges */
rvec gridpx; /* reciprocal grid point in x direction*/
rvec gridpxy; /* reciprocal grid point in x and y direction*/
rvec gridp; /* complete reciprocal grid point in 3 directions*/
rvec tmpvec; /* template to create points from basis vectors */
rvec tmpvec2; /* template to create points from basis vectors */
real coeff=0; /* variable to compute coefficients of the error estimate */
real coeff2=0; /* variable to compute coefficients of the error estimate */
real tmp=0; /* variables to compute different factors from vectors */
real tmp1=0;
real tmp2=0;
gmx_bool bFraction;
/* Random number generator */
gmx_rng_t rng=NULL;
int *numbers=NULL;
/* Index variables for parallel work distribution */
int startglobal,stopglobal;
int startlocal, stoplocal;
int x_per_core;
int xtot;
#ifdef TAKETIME
double t0=0.0;
double t1=0.0;
#endif
rng=gmx_rng_init(seed);
clear_rvec(gridpx);
clear_rvec(gridpxy);
clear_rvec(gridp);
clear_rvec(tmpvec);
clear_rvec(tmpvec2);
for(i=0;i<nr;i++)
{
q2_all += q[i]*q[i];
}
/* Calculate indices for work distribution */
startglobal=-info->nkx[0]/2;
stopglobal = info->nkx[0]/2;
xtot = stopglobal*2+1;
if (PAR(cr))
{
x_per_core = ceil((real)xtot / (real)cr->nnodes);
startlocal = startglobal + x_per_core*cr->nodeid;
stoplocal = startlocal + x_per_core -1;
if (stoplocal > stopglobal)
stoplocal = stopglobal;
}
else
{
startlocal = startglobal;
stoplocal = stopglobal;
x_per_core = xtot;
}
/*
#ifdef GMX_LIB_MPI
MPI_Barrier(MPI_COMM_WORLD);
#endif
*/
#ifdef GMX_LIB_MPI
#ifdef TAKETIME
if (MASTER(cr))
t0 = MPI_Wtime();
#endif
#endif
if (MASTER(cr)){
fprintf(stderr, "Calculating reciprocal error part 1 ...");
}
for(nx=startlocal; nx<=stoplocal; nx++)
{
svmul(nx,info->recipbox[XX],gridpx);
for(ny=-info->nky[0]/2; ny<info->nky[0]/2+1; ny++)
{
svmul(ny,info->recipbox[YY],tmpvec);
rvec_add(gridpx,tmpvec,gridpxy);
for(nz=-info->nkz[0]/2; nz<info->nkz[0]/2+1; nz++)
{
if ( 0 == nx && 0 == ny && 0 == nz )
continue;
svmul(nz,info->recipbox[ZZ],tmpvec);
rvec_add(gridpxy,tmpvec,gridp);
tmp=norm2(gridp);
coeff=exp(-1.0 * M_PI * M_PI * tmp / info->ewald_beta[0] / info->ewald_beta[0] ) ;
coeff/= 2.0 * M_PI * info->volume * tmp;
coeff2=tmp ;
tmp=eps_poly2(nx,info->nkx[0],info->pme_order[0]);
tmp+=eps_poly2(ny,info->nkx[0],info->pme_order[0]);
tmp+=eps_poly2(nz,info->nkx[0],info->pme_order[0]);
tmp1=eps_poly1(nx,info->nkx[0],info->pme_order[0]);
tmp2=eps_poly1(ny,info->nky[0],info->pme_order[0]);
tmp+=2.0 * tmp1 * tmp2;
tmp1=eps_poly1(nz,info->nkz[0],info->pme_order[0]);
tmp2=eps_poly1(ny,info->nky[0],info->pme_order[0]);
tmp+=2.0 * tmp1 * tmp2;
tmp1=eps_poly1(nz,info->nkz[0],info->pme_order[0]);
tmp2=eps_poly1(nx,info->nkx[0],info->pme_order[0]);
tmp+=2.0 * tmp1 * tmp2;
tmp1=eps_poly1(nx,info->nkx[0],info->pme_order[0]);
tmp1+=eps_poly1(ny,info->nky[0],info->pme_order[0]);
tmp1+=eps_poly1(nz,info->nkz[0],info->pme_order[0]);
tmp+= tmp1 * tmp1;
e_rec1+= 32.0 * M_PI * M_PI * coeff * coeff * coeff2 * tmp * q2_all * q2_all / nr ;
tmp1=eps_poly3(nx,info->nkx[0],info->pme_order[0]);
tmp1*=info->nkx[0];
tmp2=iprod(gridp,info->recipbox[XX]);
tmp=tmp1*tmp2;
tmp1=eps_poly3(ny,info->nky[0],info->pme_order[0]);
tmp1*=info->nky[0];
tmp2=iprod(gridp,info->recipbox[YY]);
tmp+=tmp1*tmp2;
tmp1=eps_poly3(nz,info->nkz[0],info->pme_order[0]);
tmp1*=info->nkz[0];
tmp2=iprod(gridp,info->recipbox[ZZ]);
tmp+=tmp1*tmp2;
tmp*=4.0 * M_PI;
tmp1=eps_poly4(nx,info->nkx[0],info->pme_order[0]);
tmp1*=norm2(info->recipbox[XX]);
tmp1*=info->nkx[0] * info->nkx[0];
tmp+=tmp1;
tmp1=eps_poly4(ny,info->nky[0],info->pme_order[0]);
tmp1*=norm2(info->recipbox[YY]);
tmp1*=info->nky[0] * info->nky[0];
tmp+=tmp1;
tmp1=eps_poly4(nz,info->nkz[0],info->pme_order[0]);
tmp1*=norm2(info->recipbox[ZZ]);
tmp1*=info->nkz[0] * info->nkz[0];
tmp+=tmp1;
e_rec2+= 4.0 * coeff * coeff * tmp * q2_all * q2_all / nr ;
}
}
if (MASTER(cr))
fprintf(stderr, "\rCalculating reciprocal error part 1 ... %3.0f%%", 100.0*(nx-startlocal+1)/(x_per_core));
}
if (MASTER(cr))
fprintf(stderr, "\n");
/* Use just a fraction of all charges to estimate the self energy error term? */
bFraction = (info->fracself > 0.0) && (info->fracself < 1.0);
if (bFraction)
{
/* Here xtot is the number of samples taken for the Monte Carlo calculation
* of the average of term IV of equation 35 in Wang2010. Round up to a
* number of samples that is divisible by the number of nodes */
x_per_core = ceil(info->fracself * nr / (real)cr->nnodes);
xtot = x_per_core * cr->nnodes;
}
else
{
/* In this case we use all nr particle positions */
xtot = nr;
x_per_core = ceil( (real)xtot / (real)cr->nnodes );
}
startlocal = x_per_core * cr->nodeid;
stoplocal = min(startlocal + x_per_core, xtot); /* min needed if xtot == nr */
if (bFraction)
{
/* Make shure we get identical results in serial and parallel. Therefore,
* take the sample indices from a single, global random number array that
* is constructed on the master node and that only depends on the seed */
snew(numbers, xtot);
if (MASTER(cr))
{
for (i=0; i<xtot; i++)
{
numbers[i] = floor(gmx_rng_uniform_real(rng) * nr );
}
}
/* Broadcast the random number array to the other nodes */
if (PAR(cr))
{
nblock_bc(cr,xtot,numbers);
}
if (bVerbose && MASTER(cr))
{
fprintf(stdout, "Using %d sample%s to approximate the self interaction error term",
xtot, xtot==1?"":"s");
if (PAR(cr))
fprintf(stdout, " (%d sample%s per node)", x_per_core, x_per_core==1?"":"s");
fprintf(stdout, ".\n");
}
}
/* Return the number of positions used for the Monte Carlo algorithm */
*nsamples = xtot;
for(i=startlocal;i<stoplocal;i++)
{
e_rec3x=0;
e_rec3y=0;
e_rec3z=0;
if (bFraction)
{
/* Randomly pick a charge */
ci = numbers[i];
}
else
{
/* Use all charges */
ci = i;
}
/* for(nx=startlocal; nx<=stoplocal; nx++)*/
for(nx=-info->nkx[0]/2; nx<info->nkx[0]/2+1; nx++)
{
svmul(nx,info->recipbox[XX],gridpx);
for(ny=-info->nky[0]/2; ny<info->nky[0]/2+1; ny++)
{
svmul(ny,info->recipbox[YY],tmpvec);
rvec_add(gridpx,tmpvec,gridpxy);
for(nz=-info->nkz[0]/2; nz<info->nkz[0]/2+1; nz++)
{
if ( 0 == nx && 0 == ny && 0 == nz)
continue;
svmul(nz,info->recipbox[ZZ],tmpvec);
rvec_add(gridpxy,tmpvec,gridp);
tmp=norm2(gridp);
coeff=exp(-1.0 * M_PI * M_PI * tmp / info->ewald_beta[0] / info->ewald_beta[0] );
coeff/= tmp ;
e_rec3x+=coeff*eps_self(nx,info->nkx[0],info->recipbox[XX],info->pme_order[0],x[ci]);
e_rec3y+=coeff*eps_self(ny,info->nky[0],info->recipbox[YY],info->pme_order[0],x[ci]);
e_rec3z+=coeff*eps_self(nz,info->nkz[0],info->recipbox[ZZ],info->pme_order[0],x[ci]);
}
}
}
clear_rvec(tmpvec2);
svmul(e_rec3x,info->recipbox[XX],tmpvec);
rvec_inc(tmpvec2,tmpvec);
svmul(e_rec3y,info->recipbox[YY],tmpvec);
rvec_inc(tmpvec2,tmpvec);
svmul(e_rec3z,info->recipbox[ZZ],tmpvec);
rvec_inc(tmpvec2,tmpvec);
e_rec3 += q[ci]*q[ci]*q[ci]*q[ci]*norm2(tmpvec2) / ( xtot * M_PI * info->volume * M_PI * info->volume);
if (MASTER(cr)){
fprintf(stderr, "\rCalculating reciprocal error part 2 ... %3.0f%%",
100.0*(i+1)/stoplocal);
}
}
if (MASTER(cr))
fprintf(stderr, "\n");
#ifdef GMX_LIB_MPI
#ifdef TAKETIME
if (MASTER(cr))
{
t1= MPI_Wtime() - t0;
fprintf(fp_out, "Recip. err. est. took : %lf s\n", t1);
}
#endif
#endif
#ifdef DEBUG
if (PAR(cr))
{
fprintf(stderr, "Node %3d: nx=[%3d...%3d] e_rec3=%e\n",
cr->nodeid, startlocal, stoplocal, e_rec3);
}
#endif
if (PAR(cr))
{
gmx_sum(1,&e_rec1,cr);
gmx_sum(1,&e_rec2,cr);
gmx_sum(1,&e_rec3,cr);
}
/* e_rec1*=8.0 * q2_all / info->volume / info->volume / nr ;
e_rec2*= q2_all / M_PI / M_PI / info->volume / info->volume / nr ;
e_rec3/= M_PI * M_PI * info->volume * info->volume * nr ;
*/
e_rec=sqrt(e_rec1+e_rec2+e_rec3);
return ONE_4PI_EPS0 * e_rec;
}
static void do_my_pme(FILE *fp,real tm,gmx_bool bVerbose,t_inputrec *ir,
rvec x[],rvec xbuf[],rvec f[],
real charge[],real qbuf[],real qqbuf[],
matrix box,gmx_bool bSort,
t_commrec *cr,t_nsborder *nsb,t_nrnb *nrnb,
t_block *excl,real qtot,
t_forcerec *fr,int index[],FILE *fp_xvg,
int ngroups,unsigned short cENER[])
{
real ener,vcorr,q,xx,dvdl=0,vdip,vcharge;
tensor vir,vir_corr,vir_tot;
rvec mu_tot[2];
int i,m,ii,ig,jg;
real **epme,*qptr;
/* Initiate local variables */
fr->f_el_recip = f;
clear_mat(vir);
clear_mat(vir_corr);
if (ngroups > 1) {
fprintf(fp,"There are %d energy groups\n",ngroups);
snew(epme,ngroups);
for(i=0; (i<ngroups); i++)
snew(epme[i],ngroups);
}
/* Put x is in the box, this part needs to be parallellized properly */
/*put_atoms_in_box(box,nsb->natoms,x);*/
/* Here sorting of X (and q) is done.
* Alternatively, one could just put the atoms in one of the
* cr->nnodes slabs. That is much cheaper than sorting.
*/
for(i=0; (i<nsb->natoms); i++)
index[i] = i;
if (bSort) {
xptr = x;
qsort(index,nsb->natoms,sizeof(index[0]),comp_xptr);
xptr = NULL; /* To trap unintentional use of the ptr */
}
/* After sorting we only need the part that is to be computed on
* this processor. We also compute the mu_tot here (system dipole)
*/
clear_rvec(mu_tot[0]);
for(i=START(nsb); (i<START(nsb)+HOMENR(nsb)); i++) {
ii = index[i];
q = charge[ii];
qbuf[i] = q;
for(m=0; (m<DIM); m++) {
xx = x[ii][m];
xbuf[i][m] = xx;
mu_tot[0][m] += q*xx;
}
clear_rvec(f[ii]);
}
copy_rvec(mu_tot[0],mu_tot[1]);
if (debug) {
pr_rvec(debug,0,"qbuf",qbuf,nsb->natoms,TRUE);
pr_rvecs(debug,0,"xbuf",xbuf,nsb->natoms);
pr_rvecs(debug,0,"box",box,DIM);
}
for(ig=0; (ig<ngroups); ig++) {
for(jg=ig; (jg<ngroups); jg++) {
if (ngroups > 1) {
for(i=START(nsb); (i<START(nsb)+HOMENR(nsb)); i++) {
if ((cENER[i] == ig) || (cENER[i] == jg))
qqbuf[i] = qbuf[i];
else
qqbuf[i] = 0;
}
qptr = qqbuf;
}
else
qptr = qbuf;
ener = do_pme(fp,bVerbose,ir,xbuf,f,qptr,qptr,box,cr,
nsb,nrnb,vir,fr->ewaldcoeff,FALSE,0,&dvdl,FALSE);
vcorr = ewald_LRcorrection(fp,nsb,cr,fr,qptr,qptr,excl,xbuf,box,mu_tot,
ir->ewald_geometry,ir->epsilon_surface,
0,&dvdl,&vdip,&vcharge);
gmx_sum(1,&ener,cr);
gmx_sum(1,&vcorr,cr);
if (ngroups > 1)
epme[ig][jg] = ener+vcorr;
}
}
if (ngroups > 1) {
if (fp_xvg)
fprintf(fp_xvg,"%10.3f",tm);
for(ig=0; (ig<ngroups); ig++) {
for(jg=ig; (jg<ngroups); jg++) {
if (ig != jg)
epme[ig][jg] -= epme[ig][ig]+epme[jg][jg];
if (fp_xvg)
fprintf(fp_xvg," %12.5e",epme[ig][jg]);
}
}
if (fp_xvg)
fprintf(fp_xvg,"\n");
}
else {
fprintf(fp,"Time: %10.3f Energy: %12.5e Correction: %12.5e Total: %12.5e\n",
tm,ener,vcorr,ener+vcorr);
if (fp_xvg)
fprintf(fp_xvg,"%10.3f %12.5e %12.5e %12.5e\n",tm,ener+vcorr,vdip,vcharge);
if (bVerbose) {
m_add(vir,vir_corr,vir_tot);
gmx_sum(9,vir_tot[0],cr);
pr_rvecs(fp,0,"virial",vir_tot,DIM);
}
fflush(fp);
}
}
void do_force(FILE *log,t_commrec *cr,t_commrec *mcr,
t_parm *parm,t_nsborder *nsb,tensor vir_part,tensor pme_vir,
int step,t_nrnb *nrnb,t_topology *top,t_groups *grps,
rvec x[],rvec v[],rvec f[],rvec buf[],
t_mdatoms *mdatoms,real ener[],t_fcdata *fcd,bool bVerbose,
real lambda,t_graph *graph,
bool bNS,bool bNBFonly,t_forcerec *fr, rvec mu_tot,
bool bGatherOnly)
{
static rvec box_size;
static real dvdl_lr = 0;
int cg0,cg1,i,j;
int start,homenr;
static real mu_and_q[DIM+1];
real qsum;
start = START(nsb);
homenr = HOMENR(nsb);
cg0 = CG0(nsb);
cg1 = CG1(nsb);
update_forcerec(log,fr,parm->box);
/* Calculate total (local) dipole moment in a temporary common array.
* This makes it possible to sum them over nodes faster.
*/
calc_mu_and_q(nsb,x,mdatoms->chargeT,mu_and_q,mu_and_q+DIM);
if (fr->ePBC != epbcNONE) {
/* Compute shift vectors every step, because of pressure coupling! */
if (parm->ir.epc != epcNO)
calc_shifts(parm->box,box_size,fr->shift_vec);
if (bNS) {
put_charge_groups_in_box(log,cg0,cg1,parm->box,box_size,
&(top->blocks[ebCGS]),x,fr->cg_cm);
inc_nrnb(nrnb,eNR_RESETX,homenr);
} else if (parm->ir.eI==eiSteep || parm->ir.eI==eiCG)
unshift_self(graph,parm->box,x);
}
else if (bNS)
calc_cgcm(log,cg0,cg1,&(top->blocks[ebCGS]),x,fr->cg_cm);
if (bNS) {
inc_nrnb(nrnb,eNR_CGCM,cg1-cg0);
if (PAR(cr))
move_cgcm(log,cr,fr->cg_cm,nsb->workload);
if (debug)
pr_rvecs(debug,0,"cgcm",fr->cg_cm,nsb->cgtotal);
}
/* Communicate coordinates and sum dipole and net charge if necessary */
if (PAR(cr)) {
move_x(log,cr->left,cr->right,x,nsb,nrnb);
gmx_sum(DIM+1,mu_and_q,cr);
}
for(i=0;i<DIM;i++)
mu_tot[i]=mu_and_q[i];
qsum=mu_and_q[DIM];
/* Reset energies */
reset_energies(&(parm->ir.opts),grps,fr,bNS,ener);
if (bNS) {
if (fr->ePBC != epbcNONE)
/* Calculate intramolecular shift vectors to make molecules whole */
mk_mshift(log,graph,parm->box,x);
/* Reset long range forces if necessary */
if (fr->bTwinRange) {
clear_rvecs(nsb->natoms,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_lr = 0;
ns(log,fr,x,f,parm->box,grps,&(parm->ir.opts),top,mdatoms,
cr,nrnb,nsb,step,lambda,&dvdl_lr);
}
/* Reset PME/Ewald forces if necessary */
if (EEL_LR(fr->eeltype))
clear_rvecs(homenr,fr->f_pme+start);
/* Copy long range forces into normal buffers */
if (fr->bTwinRange) {
for(i=0; i<nsb->natoms; 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 {
clear_rvecs(nsb->natoms,f);
clear_rvecs(SHIFTS,fr->fshift);
}
/* Compute the forces */
force(log,step,fr,&(parm->ir),&(top->idef),nsb,cr,mcr,nrnb,grps,mdatoms,
top->atoms.grps[egcENER].nr,&(parm->ir.opts),
x,f,ener,fcd,bVerbose,parm->box,lambda,graph,&(top->atoms.excl),
bNBFonly,pme_vir,mu_tot,qsum,bGatherOnly);
/* Take long range contribution to free energy into account */
ener[F_DVDL] += dvdl_lr;
#ifdef DEBUG
if (bNS)
print_nrnb(log,nrnb);
#endif
/* The short-range virial from surrounding boxes */
clear_mat(vir_part);
calc_vir(log,SHIFTS,fr->shift_vec,fr->fshift,vir_part);
inc_nrnb(nrnb,eNR_VIRIAL,SHIFTS);
if (debug)
pr_rvecs(debug,0,"vir_shifts",vir_part,DIM);
/* Compute forces due to electric field */
calc_f_el(start,homenr,mdatoms->chargeT,f,parm->ir.ex);
/* When using PME/Ewald we compute the long range virial (pme_vir) 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))
move_f(log,cr->left,cr->right,f,buf,nsb,nrnb);
}
/* Estimate the reciprocal space part error of the SPME Ewald sum. */
static real estimate_reciprocal(
t_inputinfo *info,
rvec x[], /* array of particles */
real q[], /* array of charges */
int nr, /* number of charges = size of the charge array */
FILE *fp_out,
t_commrec *cr)
{
real e_rec=0; /* reciprocal error estimate */
real e_rec1=0; /* Error estimate term 1*/
real e_rec2=0; /* Error estimate term 2*/
real e_rec3=0; /* Error estimate term 3 */
real e_rec3x=0; /* part of Error estimate term 3 in x */
real e_rec3y=0; /* part of Error estimate term 3 in y */
real e_rec3z=0; /* part of Error estimate term 3 in z */
int i,ci;
int nx,ny,nz; /* grid coordinates */
real q2_all=0; /* sum of squared charges */
rvec gridpx; /* reciprocal grid point in x direction*/
rvec gridpxy; /* reciprocal grid point in x and y direction*/
rvec gridp; /* complete reciprocal grid point in 3 directions*/
rvec tmpvec; /* template to create points from basis vectors */
rvec tmpvec2; /* template to create points from basis vectors */
real coeff=0; /* variable to compute coefficients of the error estimate */
real coeff2=0; /* variable to compute coefficients of the error estimate */
real tmp=0; /* variables to compute different factors from vectors */
real tmp1=0;
real tmp2=0;
real xtmp=0;
real ytmp=0;
real ztmp=0;
double ewald_error;
/* Random number generator */
gmx_rng_t rng=NULL;
/*rng=gmx_rng_init(gmx_rng_make_seed()); */
/* Index variables for parallel work distribution */
int startglobal,stopglobal;
int startlocal, stoplocal;
int x_per_core;
int nrsamples;
real xtot;
/* #define TAKETIME */
#ifdef TAKETIME
double t0=0.0;
double t1=0.0;
double t2=0.0;
#endif
rng=gmx_rng_init(cr->nodeid);
clear_rvec(gridpx);
clear_rvec(gridpxy);
clear_rvec(gridp);
clear_rvec(tmpvec);
clear_rvec(tmpvec2);
for(i=0;i<nr;i++)
{
q2_all += q[i]*q[i];
}
/* Calculate indices for work distribution */
startglobal=-info->nkx[0]/2;
stopglobal = info->nkx[0]/2;
xtot = stopglobal*2+1;
if (PAR(cr))
{
x_per_core = ceil(xtot / cr->nnodes);
startlocal = startglobal + x_per_core*cr->nodeid;
stoplocal = startlocal + x_per_core -1;
if (stoplocal > stopglobal)
stoplocal = stopglobal;
}
else
{
startlocal = startglobal;
stoplocal = stopglobal;
x_per_core = xtot;
}
/*
#ifdef GMX_MPI
MPI_Barrier(MPI_COMM_WORLD);
#endif
*/
#ifdef TAKETIME
if (MASTER(cr))
t0 = MPI_Wtime();
#endif
if (MASTER(cr)){
fprintf(stderr, "Calculating reciprocal error part 1 ...");
}
for(nx=startlocal; nx<=stoplocal; nx++)
{
svmul(nx,info->recipbox[XX],gridpx);
for(ny=-info->nky[0]/2; ny<info->nky[0]/2+1; ny++)
{
svmul(ny,info->recipbox[YY],tmpvec);
rvec_add(gridpx,tmpvec,gridpxy);
for(nz=-info->nkz[0]/2; nz<info->nkz[0]/2+1; nz++)
{
if ( 0 == nx && 0 == ny && 0 == nz )
continue;
svmul(nz,info->recipbox[ZZ],tmpvec);
rvec_add(gridpxy,tmpvec,gridp);
tmp=norm2(gridp);
coeff=exp(-1.0 * M_PI * M_PI * tmp / info->ewald_beta[0] / info->ewald_beta[0] ) ;
coeff/= 2.0 * M_PI * info->volume * tmp;
coeff2=tmp ;
tmp=eps_poly2(nx,info->nkx[0],info->pme_order[0]);
tmp+=eps_poly2(ny,info->nkx[0],info->pme_order[0]);
tmp+=eps_poly2(nz,info->nkx[0],info->pme_order[0]);
tmp1=eps_poly1(nx,info->nkx[0],info->pme_order[0]);
tmp2=eps_poly1(ny,info->nky[0],info->pme_order[0]);
tmp+=2.0 * tmp1 * tmp2;
tmp1=eps_poly1(nz,info->nkz[0],info->pme_order[0]);
tmp2=eps_poly1(ny,info->nky[0],info->pme_order[0]);
tmp+=2.0 * tmp1 * tmp2;
tmp1=eps_poly1(nz,info->nkz[0],info->pme_order[0]);
tmp2=eps_poly1(nx,info->nkx[0],info->pme_order[0]);
tmp+=2.0 * tmp1 * tmp2;
tmp1=eps_poly1(nx,info->nkx[0],info->pme_order[0]);
tmp1+=eps_poly1(ny,info->nky[0],info->pme_order[0]);
tmp1+=eps_poly1(nz,info->nkz[0],info->pme_order[0]);
tmp+= tmp1 * tmp1;
e_rec1+= 32.0 * M_PI * M_PI * coeff * coeff * coeff2 * tmp * q2_all * q2_all / nr ;
tmp1=eps_poly3(nx,info->nkx[0],info->pme_order[0]);
tmp1*=info->nkx[0];
tmp2=iprod(gridp,info->recipbox[XX]);
tmp=tmp1*tmp2;
tmp1=eps_poly3(ny,info->nky[0],info->pme_order[0]);
tmp1*=info->nky[0];
tmp2=iprod(gridp,info->recipbox[YY]);
tmp+=tmp1*tmp2;
tmp1=eps_poly3(nz,info->nkz[0],info->pme_order[0]);
tmp1*=info->nkz[0];
tmp2=iprod(gridp,info->recipbox[ZZ]);
tmp+=tmp1*tmp2;
tmp*=4.0 * M_PI;
tmp1=eps_poly4(nx,info->nkx[0],info->pme_order[0]);
tmp1*=norm2(info->recipbox[XX]);
tmp1*=info->nkx[0] * info->nkx[0];
tmp+=tmp1;
tmp1=eps_poly4(ny,info->nky[0],info->pme_order[0]);
tmp1*=norm2(info->recipbox[YY]);
tmp1*=info->nky[0] * info->nky[0];
tmp+=tmp1;
tmp1=eps_poly4(nz,info->nkz[0],info->pme_order[0]);
tmp1*=norm2(info->recipbox[ZZ]);
tmp1*=info->nkz[0] * info->nkz[0];
tmp+=tmp1;
e_rec2+= 4.0 * coeff * coeff * tmp * q2_all * q2_all / nr ;
}
}
if (MASTER(cr))
fprintf(stderr, "\rCalculating reciprocal error part 1 ... %3.0f%%", 100.0*(nx-startlocal+1)/(x_per_core));
}
/*
#ifdef GMX_MPI
MPI_Barrier(MPI_COMM_WORLD);
#endif
*/
if (MASTER(cr))
fprintf(stderr, "\n");
if (info->fracself>0)
{
nrsamples=ceil(info->fracself*nr);
}
else
{
nrsamples=nr;
}
xtot=nrsamples;
startglobal=0;
stopglobal=nr;
if(PAR(cr))
{
x_per_core=ceil(xtot/cr->nnodes);
startlocal=startglobal+x_per_core*cr->nodeid;
stoplocal=startglobal+x_per_core*(cr->nodeid+1);
if (stoplocal>stopglobal)
stoplocal=stopglobal;
}
else
{
startlocal=startglobal;
stoplocal=stopglobal;
x_per_core=xtot;
}
for(i=startlocal;i<stoplocal;i++)
{
e_rec3x=0;
e_rec3y=0;
e_rec3z=0;
if (info->fracself<0) {
ci=i;
}else {
ci=floor(gmx_rng_uniform_real(rng) * nr );
if (ci==nr)
{
ci=nr-1;
}
}
/* for(nx=startlocal; nx<=stoplocal; nx++)*/
for(nx=-info->nkx[0]/2; nx<info->nkx[0]/2+1; nx++)
{
svmul(nx,info->recipbox[XX],gridpx);
for(ny=-info->nky[0]/2; ny<info->nky[0]/2+1; ny++)
{
svmul(ny,info->recipbox[YY],tmpvec);
rvec_add(gridpx,tmpvec,gridpxy);
for(nz=-info->nkz[0]/2; nz<info->nkz[0]/2+1; nz++)
{
if ( 0 == nx && 0 == ny && 0 == nz)
continue;
svmul(nz,info->recipbox[ZZ],tmpvec);
rvec_add(gridpxy,tmpvec,gridp);
tmp=norm2(gridp);
coeff=exp(-1.0 * M_PI * M_PI * tmp / info->ewald_beta[0] / info->ewald_beta[0] );
coeff/= tmp ;
e_rec3x+=coeff*eps_self(nx,info->nkx[0],info->recipbox[XX],info->pme_order[0],x[ci]);
e_rec3y+=coeff*eps_self(ny,info->nky[0],info->recipbox[YY],info->pme_order[0],x[ci]);
e_rec3z+=coeff*eps_self(nz,info->nkz[0],info->recipbox[ZZ],info->pme_order[0],x[ci]);
}
}
}
clear_rvec(tmpvec2);
svmul(e_rec3x,info->recipbox[XX],tmpvec);
rvec_inc(tmpvec2,tmpvec);
svmul(e_rec3y,info->recipbox[YY],tmpvec);
rvec_inc(tmpvec2,tmpvec);
svmul(e_rec3z,info->recipbox[ZZ],tmpvec);
rvec_inc(tmpvec2,tmpvec);
e_rec3 += q[ci]*q[ci]*q[ci]*q[ci]*norm2(tmpvec2) / ( nrsamples * M_PI * info->volume * M_PI * info->volume);
if (MASTER(cr)){
fprintf(stderr, "\rCalculating reciprocal error part 2 ... %3.0f%%",
100.0*(i+1)/stoplocal);
}
}
if (MASTER(cr))
fprintf(stderr, "\n");
#ifdef TAKETIME
if (MASTER(cr))
{
t1= MPI_Wtime() - t0;
fprintf(fp_out, "Recip. err. est. took : %lf s\n", t1);
}
#endif
#ifdef DEBUG
if (PAR(cr))
{
fprintf(stderr, "Node %3d: nx=[%3d...%3d] e_rec3=%e\n",
cr->nodeid, startlocal, stoplocal, e_rec3);
}
#endif
/*
#ifdef GMX_MPI
MPI_Barrier(MPI_COMM_WORLD);
#endif
*/
#ifdef TAKETIME
if (MASTER(cr))
{
t2= MPI_Wtime() - t0;
fprintf(fp_out, "barrier : %lf s\n", t2-t1);
}
#endif
if (PAR(cr))
{
gmx_sum(1,&e_rec1,cr);
gmx_sum(1,&e_rec2,cr);
gmx_sum(1,&e_rec3,cr);
}
#ifdef TAKETIME
if (MASTER(cr))
fprintf(fp_out, "final reduce : %lf s\n", MPI_Wtime() - t0-t2);
#endif
/* e_rec1*=8.0 * q2_all / info->volume / info->volume / nr ;
e_rec2*= q2_all / M_PI / M_PI / info->volume / info->volume / nr ;
e_rec3/= M_PI * M_PI * info->volume * info->volume * nr ;
*/
e_rec=sqrt(e_rec1+e_rec2+e_rec3);
return ONE_4PI_EPS0 * e_rec;
}
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;
}
/* Assemble the positions of the group such that every node has all of them.
* The atom indices are retrieved from anrs_loc[0..nr_loc]
* Note that coll_ind[i] = i is needed in the serial case */
extern void communicate_group_positions(
const t_commrec *cr, /* Pointer to MPI communication data */
rvec *xcoll, /* Collective array of positions */
ivec *shifts, /* Collective array of shifts for xcoll (can be NULL) */
ivec *extra_shifts, /* (optional) Extra shifts since last time step */
const gmx_bool bNS, /* (optional) NS step, the shifts have changed */
const rvec *x_loc, /* Local positions on this node */
const int nr, /* Total number of atoms in the group */
const int nr_loc, /* Local number of atoms in the group */
const int *anrs_loc, /* Local atom numbers */
const int *coll_ind, /* Collective index */
rvec *xcoll_old, /* (optional) Positions from the last time step,
used to make group whole */
const matrix box) /* (optional) The box */
{
int i;
/* Zero out the groups' global position array */
clear_rvecs(nr, xcoll);
/* Put the local positions that this node has into the right place of
* the collective array. Note that in the serial case, coll_ind[i] = i */
for (i = 0; i < nr_loc; i++)
{
copy_rvec(x_loc[anrs_loc[i]], xcoll[coll_ind[i]]);
}
if (PAR(cr))
{
/* Add the arrays from all nodes together */
gmx_sum(nr*3, xcoll[0], cr);
}
/* Now we have all the positions of the group in the xcoll array present on all
* nodes.
*
* The rest of the code is for making the group whole again in case atoms changed
* their PBC representation / crossed a box boundary. We only do that if the
* shifts array is allocated. */
if (nullptr != shifts)
{
/* To make the group whole, start with a whole group and each
* step move the assembled positions at closest distance to the positions
* from the last step. First shift the positions with the saved shift
* vectors (these are 0 when this routine is called for the first time!) */
shift_positions_group(box, xcoll, shifts, nr);
/* Now check if some shifts changed since the last step.
* This only needs to be done when the shifts are expected to have changed,
* i.e. after neighbor searching */
if (bNS)
{
get_shifts_group(3, box, xcoll, nr, xcoll_old, extra_shifts);
/* Shift with the additional shifts such that we get a whole group now */
shift_positions_group(box, xcoll, extra_shifts, nr);
/* Add the shift vectors together for the next time step */
for (i = 0; i < nr; i++)
{
shifts[i][XX] += extra_shifts[i][XX];
shifts[i][YY] += extra_shifts[i][YY];
shifts[i][ZZ] += extra_shifts[i][ZZ];
}
/* Store current correctly-shifted positions for comparison in the next NS time step */
for (i = 0; i < nr; i++)
{
copy_rvec(xcoll[i], xcoll_old[i]);
}
}
}
}
void init_orires(FILE *log,int nfa,t_iatom forceatoms[],t_iparams ip[],
rvec *xref,t_mdatoms *md,t_inputrec *ir,
t_commrec *mcr,t_fcdata *fcd)
{
int i,j,d,ex,nr,*nr_ex;
real mtot;
rvec com;
t_oriresdata *od;
od = &(fcd->orires);
od->fc = ir->orires_fc;
od->nex = 0;
od->S = NULL;
if (ir->orires_tau > 0)
od->edt = exp(-ir->delta_t/ir->orires_tau);
else
od->edt = 0;
od->edt1 = 1 - od->edt;
od->exp_min_t_tau = 1.0;
od->nr = nfa/3;
if (od->nr == 0)
return;
nr_ex = NULL;
for(i=0; i<nfa; i+=3) {
ex = ip[forceatoms[i]].orires.ex;
if (ex >= od->nex) {
srenew(nr_ex,ex+1);
for(j=od->nex; j<ex+1; j++)
nr_ex[j] = 0;
od->nex = ex+1;
}
nr_ex[ex]++;
}
snew(od->S,od->nex);
/* When not doing time averaging, the instaneous and time averaged data
* are indentical and the pointers can point to the same memory.
*/
snew(od->Dinsl,od->nr);
if (mcr)
snew(od->Dins,od->nr);
else
od->Dins = od->Dinsl;
if (fabs(ir->orires_tau) < GMX_REAL_MIN)
od->Dtav = od->Dins;
else
snew(od->Dtav,od->nr);
snew(od->oinsl,od->nr);
if (mcr)
snew(od->oins,od->nr);
else
od->oins = od->oinsl;
if ( fabs(ir->orires_tau) < GMX_REAL_MIN)
od->otav = od->oins;
else
snew(od->otav,od->nr);
snew(od->tmp,od->nex);
snew(od->TMP,od->nex);
for(ex=0; ex<od->nex; ex++) {
snew(od->TMP[ex],5);
for(i=0; i<5; i++)
snew(od->TMP[ex][i],5);
}
od->nref = 0;
for(i=0; i<md->nr; i++)
if (md->cORF[i] == 0)
od->nref++;
snew(od->mref,od->nref);
snew(od->xref,od->nref);
snew(od->xtmp,od->nref);
/* Determine the reference structure on the master node.
* Copy it to the other nodes after checking multi compatibility,
* so we are sure the subsystems match before copying.
*/
clear_rvec(com);
mtot = 0.0;
j = 0;
for(i=0; i<md->nr; i++) {
if (md->cORF[i] == 0) {
od->mref[j] = md->massT[i];
if (mcr==NULL || MASTER(mcr)) {
copy_rvec(xref[i],od->xref[j]);
for(d=0; d<DIM; d++)
com[d] += od->mref[j]*xref[i][d];
}
mtot += od->mref[j];
j++;
}
}
od->invmref = 1.0/mtot;
svmul(od->invmref,com,com);
if (mcr==NULL || MASTER(mcr))
for(j=0; j<od->nref; j++)
rvec_dec(od->xref[j],com);
fprintf(log,"Found %d orientation experiments\n",od->nex);
for(i=0; i<od->nex; i++)
fprintf(log," experiment %d has %d restraints\n",i+1,nr_ex[i]);
sfree(nr_ex);
fprintf(log," the fit group consists of %d atoms and has total mass %g\n",
od->nref,mtot);
if (mcr) {
fprintf(log," the orientation restraints are ensemble averaged over %d systems\n",mcr->nnodes);
check_multi_int(log,mcr,fcd->orires.nr,
"the number of orientation restraints");
check_multi_int(log,mcr,fcd->orires.nref,
"the number of fit atoms for orientation restraining");
/* Copy the reference coordinates from the master to the other nodes */
gmx_sum(DIM*fcd->orires.nref,fcd->orires.xref[0],mcr);
}
}
real calc_orires_dev(t_commrec *mcr,
int nfa,t_iatom forceatoms[],t_iparams ip[],
t_mdatoms *md,rvec x[],t_fcdata *fcd)
{
int fa,d,i,j,type,ex,nref;
real edt,edt1,invn,pfac,r2,invr,corrfac,weight,wsv2,sw,dev;
tensor *S,R,TMP;
rvec5 *Dinsl,*Dins,*Dtav,*rhs;
real *mref,***T;
rvec *xref,*xtmp,com,r_unrot,r;
t_oriresdata *od;
bool bTAV;
static real two_thr=2.0/3.0;
od = &(fcd->orires);
bTAV = (fabs(od->edt)>GMX_REAL_MIN);
edt = od->edt;
edt1 = od->edt1;
S = od->S;
Dinsl= od->Dinsl;
Dins = od->Dins;
Dtav = od->Dtav;
T = od->TMP;
rhs = od->tmp;
nref = od->nref;
mref = od->mref;
xref = od->xref;
xtmp = od->xtmp;
od->exp_min_t_tau *= edt;
if (mcr)
invn = 1.0/mcr->nnodes;
else
invn = 1.0;
j=0;
for(i=0; i<md->nr; i++)
if (md->cORF[i] == 0) {
copy_rvec(x[i],xtmp[j]);
for(d=0; d<DIM; d++)
com[d] += mref[j]*xref[j][d];
j++;
}
svmul(od->invmref,com,com);
for(j=0; j<nref; j++)
rvec_dec(xtmp[j],com);
/* Calculate the rotation matrix to rotate x to the reference orientation */
calc_fit_R(nref,mref,xref,xtmp,R);
copy_mat(R,od->R);
d = 0;
for(fa=0; fa<nfa; fa+=3) {
type = forceatoms[fa];
rvec_sub(x[forceatoms[fa+1]],x[forceatoms[fa+2]],r_unrot);
mvmul(R,r_unrot,r);
r2 = norm2(r);
invr = invsqrt(r2);
/* Calculate the prefactor for the D tensor, this includes the factor 3! */
pfac = ip[type].orires.c*invr*invr*3;
for(i=0; i<ip[type].orires.pow; i++)
pfac *= invr;
Dinsl[d][0] = pfac*(2*r[0]*r[0] + r[1]*r[1] - r2);
Dinsl[d][1] = pfac*(2*r[0]*r[1]);
Dinsl[d][2] = pfac*(2*r[0]*r[2]);
Dinsl[d][3] = pfac*(2*r[1]*r[1] + r[0]*r[0] - r2);
Dinsl[d][4] = pfac*(2*r[1]*r[2]);
if (mcr)
for(i=0; i<5; i++)
Dins[d][i] = Dinsl[d][i]*invn;
d++;
}
if (mcr)
gmx_sum(5*od->nr,Dins[0],mcr);
/* Correction factor to correct for the lack of history for short times */
corrfac = 1.0/(1.0-od->exp_min_t_tau);
/* Calculate the order tensor S for each experiment via optimization */
for(ex=0; ex<od->nex; ex++)
for(i=0; i<5; i++) {
rhs[ex][i] = 0;
for(j=0; j<=i; j++)
T[ex][i][j] = 0;
}
d = 0;
for(fa=0; fa<nfa; fa+=3) {
if (bTAV)
for(i=0; i<5; i++)
Dtav[d][i] = edt*Dtav[d][i] + edt1*Dins[d][i];
type = forceatoms[fa];
ex = ip[type].orires.ex;
weight = ip[type].orires.kfac;
/* Calculate the vector rhs and half the matrix T for the 5 equations */
for(i=0; i<5; i++) {
rhs[ex][i] += Dtav[d][i]*ip[type].orires.obs*weight;
for(j=0; j<=i; j++)
T[ex][i][j] += Dtav[d][i]*Dtav[d][j]*weight;
}
d++;
}
/* Now we have all the data we can calculate S */
for(ex=0; ex<od->nex; ex++) {
/* Correct corrfac and copy one half of T to the other half */
for(i=0; i<5; i++) {
rhs[ex][i] *= corrfac;
T[ex][i][i] *= sqr(corrfac);
for(j=0; j<i; j++) {
T[ex][i][j] *= sqr(corrfac);
T[ex][j][i] = T[ex][i][j];
}
}
m_inv_gen(T[ex],5,T[ex]);
/* Calculate the orientation tensor S for this experiment */
S[ex][0][0] = 0;
S[ex][0][1] = 0;
S[ex][0][2] = 0;
S[ex][1][1] = 0;
S[ex][1][2] = 0;
for(i=0; i<5; i++) {
S[ex][0][0] += 1.5*T[ex][0][i]*rhs[ex][i];
S[ex][0][1] += 1.5*T[ex][1][i]*rhs[ex][i];
S[ex][0][2] += 1.5*T[ex][2][i]*rhs[ex][i];
S[ex][1][1] += 1.5*T[ex][3][i]*rhs[ex][i];
S[ex][1][2] += 1.5*T[ex][4][i]*rhs[ex][i];
}
S[ex][1][0] = S[ex][0][1];
S[ex][2][0] = S[ex][0][2];
S[ex][2][1] = S[ex][1][2];
S[ex][2][2] = -S[ex][0][0] - S[ex][1][1];
}
wsv2 = 0;
sw = 0;
d = 0;
for(fa=0; fa<nfa; fa+=3) {
type = forceatoms[fa];
ex = ip[type].orires.ex;
od->otav[d] = two_thr*
corrfac*(S[ex][0][0]*Dtav[d][0] + S[ex][0][1]*Dtav[d][1] +
S[ex][0][2]*Dtav[d][2] + S[ex][1][1]*Dtav[d][3] +
S[ex][1][2]*Dtav[d][4]);
if (bTAV)
od->oins[d] = two_thr*(S[ex][0][0]*Dins[d][0] + S[ex][0][1]*Dins[d][1] +
S[ex][0][2]*Dins[d][2] + S[ex][1][1]*Dins[d][3] +
S[ex][1][2]*Dins[d][4]);
if (mcr)
/* When ensemble averaging is used recalculate the local orientation
* for output to the energy file.
*/
od->oinsl[d] = two_thr*
(S[ex][0][0]*Dinsl[d][0] + S[ex][0][1]*Dinsl[d][1] +
S[ex][0][2]*Dinsl[d][2] + S[ex][1][1]*Dinsl[d][3] +
S[ex][1][2]*Dinsl[d][4]);
dev = od->otav[d] - ip[type].orires.obs;
wsv2 += ip[type].orires.kfac*sqr(dev);
sw += ip[type].orires.kfac;
d++;
}
od->rmsdev = sqrt(wsv2/sw);
/* Rotate the S matrices back, so we get the correct grad(tr(S D)) */
for(ex=0; ex<od->nex; ex++) {
tmmul(R,S[ex],TMP);
mmul(TMP,R,S[ex]);
}
return od->rmsdev;
/* Approx. 120*nfa/3 flops */
}