double do_tpi(FILE *fplog, t_commrec *cr,
int nfile, const t_filenm fnm[],
const output_env_t oenv, gmx_bool bVerbose, gmx_bool gmx_unused bCompact,
int gmx_unused nstglobalcomm,
gmx_vsite_t gmx_unused *vsite, gmx_constr_t gmx_unused constr,
int gmx_unused stepout,
t_inputrec *inputrec,
gmx_mtop_t *top_global, t_fcdata *fcd,
t_state *state,
t_mdatoms *mdatoms,
t_nrnb *nrnb, gmx_wallcycle_t wcycle,
gmx_edsam_t gmx_unused ed,
t_forcerec *fr,
int gmx_unused repl_ex_nst, int gmx_unused repl_ex_nex, int gmx_unused repl_ex_seed,
gmx_membed_t gmx_unused membed,
real gmx_unused cpt_period, real gmx_unused max_hours,
const char gmx_unused *deviceOptions,
int gmx_unused imdport,
unsigned long gmx_unused Flags,
gmx_walltime_accounting_t walltime_accounting)
{
const char *TPI = "Test Particle Insertion";
gmx_localtop_t *top;
gmx_groups_t *groups;
gmx_enerdata_t *enerd;
rvec *f;
real lambda, t, temp, beta, drmax, epot;
double embU, sum_embU, *sum_UgembU, V, V_all, VembU_all;
t_trxstatus *status;
t_trxframe rerun_fr;
gmx_bool bDispCorr, bCharge, bRFExcl, bNotLastFrame, bStateChanged, bNS;
tensor force_vir, shake_vir, vir, pres;
int cg_tp, a_tp0, a_tp1, ngid, gid_tp, nener, e;
rvec *x_mol;
rvec mu_tot, x_init, dx, x_tp;
int nnodes, frame;
gmx_int64_t frame_step_prev, frame_step;
gmx_int64_t nsteps, stepblocksize = 0, step;
gmx_int64_t rnd_count_stride, rnd_count;
gmx_int64_t seed;
double rnd[4];
int i, start, end;
FILE *fp_tpi = NULL;
char *ptr, *dump_pdb, **leg, str[STRLEN], str2[STRLEN];
double dbl, dump_ener;
gmx_bool bCavity;
int nat_cavity = 0, d;
real *mass_cavity = NULL, mass_tot;
int nbin;
double invbinw, *bin, refvolshift, logV, bUlogV;
real dvdl, prescorr, enercorr, dvdlcorr;
gmx_bool bEnergyOutOfBounds;
const char *tpid_leg[2] = {"direct", "reweighted"};
/* Since there is no upper limit to the insertion energies,
* we need to set an upper limit for the distribution output.
*/
real bU_bin_limit = 50;
real bU_logV_bin_limit = bU_bin_limit + 10;
nnodes = cr->nnodes;
top = gmx_mtop_generate_local_top(top_global, inputrec);
groups = &top_global->groups;
bCavity = (inputrec->eI == eiTPIC);
if (bCavity)
{
ptr = getenv("GMX_TPIC_MASSES");
if (ptr == NULL)
{
nat_cavity = 1;
}
else
{
/* Read (multiple) masses from env var GMX_TPIC_MASSES,
* The center of mass of the last atoms is then used for TPIC.
*/
nat_cavity = 0;
while (sscanf(ptr, "%lf%n", &dbl, &i) > 0)
{
srenew(mass_cavity, nat_cavity+1);
mass_cavity[nat_cavity] = dbl;
fprintf(fplog, "mass[%d] = %f\n",
nat_cavity+1, mass_cavity[nat_cavity]);
nat_cavity++;
ptr += i;
}
if (nat_cavity == 0)
{
gmx_fatal(FARGS, "Found %d masses in GMX_TPIC_MASSES", nat_cavity);
}
}
}
/*
init_em(fplog,TPI,inputrec,&lambda,nrnb,mu_tot,
state->box,fr,mdatoms,top,cr,nfile,fnm,NULL,NULL);*/
/* We never need full pbc for TPI */
fr->ePBC = epbcXYZ;
/* Determine the temperature for the Boltzmann weighting */
temp = inputrec->opts.ref_t[0];
if (fplog)
{
for (i = 1; (i < inputrec->opts.ngtc); i++)
{
if (inputrec->opts.ref_t[i] != temp)
{
fprintf(fplog, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
fprintf(stderr, "\nWARNING: The temperatures of the different temperature coupling groups are not identical\n\n");
}
}
fprintf(fplog,
"\n The temperature for test particle insertion is %.3f K\n\n",
temp);
}
beta = 1.0/(BOLTZ*temp);
/* Number of insertions per frame */
nsteps = inputrec->nsteps;
/* Use the same neighborlist with more insertions points
* in a sphere of radius drmax around the initial point
*/
/* This should be a proper mdp parameter */
drmax = inputrec->rtpi;
/* An environment variable can be set to dump all configurations
* to pdb with an insertion energy <= this value.
*/
dump_pdb = getenv("GMX_TPI_DUMP");
dump_ener = 0;
if (dump_pdb)
{
sscanf(dump_pdb, "%lf", &dump_ener);
}
atoms2md(top_global, inputrec, 0, NULL, top_global->natoms, mdatoms);
update_mdatoms(mdatoms, inputrec->fepvals->init_lambda);
snew(enerd, 1);
init_enerdata(groups->grps[egcENER].nr, inputrec->fepvals->n_lambda, enerd);
snew(f, top_global->natoms);
/* Print to log file */
walltime_accounting_start(walltime_accounting);
wallcycle_start(wcycle, ewcRUN);
print_start(fplog, cr, walltime_accounting, "Test Particle Insertion");
/* The last charge group is the group to be inserted */
cg_tp = top->cgs.nr - 1;
a_tp0 = top->cgs.index[cg_tp];
a_tp1 = top->cgs.index[cg_tp+1];
if (debug)
{
fprintf(debug, "TPI cg %d, atoms %d-%d\n", cg_tp, a_tp0, a_tp1);
}
if (a_tp1 - a_tp0 > 1 &&
(inputrec->rlist < inputrec->rcoulomb ||
inputrec->rlist < inputrec->rvdw))
{
gmx_fatal(FARGS, "Can not do TPI for multi-atom molecule with a twin-range cut-off");
}
snew(x_mol, a_tp1-a_tp0);
bDispCorr = (inputrec->eDispCorr != edispcNO);
bCharge = FALSE;
for (i = a_tp0; i < a_tp1; i++)
{
/* Copy the coordinates of the molecule to be insterted */
copy_rvec(state->x[i], x_mol[i-a_tp0]);
/* Check if we need to print electrostatic energies */
bCharge |= (mdatoms->chargeA[i] != 0 ||
(mdatoms->chargeB && mdatoms->chargeB[i] != 0));
}
bRFExcl = (bCharge && EEL_RF(fr->eeltype) && fr->eeltype != eelRF_NEC);
calc_cgcm(fplog, cg_tp, cg_tp+1, &(top->cgs), state->x, fr->cg_cm);
if (bCavity)
{
if (norm(fr->cg_cm[cg_tp]) > 0.5*inputrec->rlist && fplog)
{
fprintf(fplog, "WARNING: Your TPI molecule is not centered at 0,0,0\n");
fprintf(stderr, "WARNING: Your TPI molecule is not centered at 0,0,0\n");
}
}
else
{
/* Center the molecule to be inserted at zero */
for (i = 0; i < a_tp1-a_tp0; i++)
{
rvec_dec(x_mol[i], fr->cg_cm[cg_tp]);
}
}
if (fplog)
{
fprintf(fplog, "\nWill insert %d atoms %s partial charges\n",
a_tp1-a_tp0, bCharge ? "with" : "without");
fprintf(fplog, "\nWill insert %d times in each frame of %s\n",
(int)nsteps, opt2fn("-rerun", nfile, fnm));
}
if (!bCavity)
{
if (inputrec->nstlist > 1)
{
if (drmax == 0 && a_tp1-a_tp0 == 1)
{
gmx_fatal(FARGS, "Re-using the neighborlist %d times for insertions of a single atom in a sphere of radius %f does not make sense", inputrec->nstlist, drmax);
}
if (fplog)
{
fprintf(fplog, "Will use the same neighborlist for %d insertions in a sphere of radius %f\n", inputrec->nstlist, drmax);
}
}
}
else
{
if (fplog)
{
fprintf(fplog, "Will insert randomly in a sphere of radius %f around the center of the cavity\n", drmax);
}
}
ngid = groups->grps[egcENER].nr;
gid_tp = GET_CGINFO_GID(fr->cginfo[cg_tp]);
nener = 1 + ngid;
if (bDispCorr)
{
nener += 1;
}
if (bCharge)
{
nener += ngid;
if (bRFExcl)
{
nener += 1;
}
if (EEL_FULL(fr->eeltype))
{
nener += 1;
}
}
snew(sum_UgembU, nener);
/* Copy the random seed set by the user */
seed = inputrec->ld_seed;
/* We use the frame step number as one random counter.
* The second counter use the insertion (step) count. But we
* need multiple random numbers per insertion. This number is
* not fixed, since we generate random locations in a sphere
* by putting locations in a cube and some of these fail.
* A count of 20 is already extremely unlikely, so 10000 is
* a safe margin for random numbers per insertion.
*/
rnd_count_stride = 10000;
if (MASTER(cr))
{
fp_tpi = xvgropen(opt2fn("-tpi", nfile, fnm),
"TPI energies", "Time (ps)",
"(kJ mol\\S-1\\N) / (nm\\S3\\N)", oenv);
xvgr_subtitle(fp_tpi, "f. are averages over one frame", oenv);
snew(leg, 4+nener);
e = 0;
sprintf(str, "-kT log(<Ve\\S-\\betaU\\N>/<V>)");
leg[e++] = strdup(str);
sprintf(str, "f. -kT log<e\\S-\\betaU\\N>");
leg[e++] = strdup(str);
sprintf(str, "f. <e\\S-\\betaU\\N>");
leg[e++] = strdup(str);
sprintf(str, "f. V");
leg[e++] = strdup(str);
sprintf(str, "f. <Ue\\S-\\betaU\\N>");
leg[e++] = strdup(str);
for (i = 0; i < ngid; i++)
{
sprintf(str, "f. <U\\sVdW %s\\Ne\\S-\\betaU\\N>",
*(groups->grpname[groups->grps[egcENER].nm_ind[i]]));
leg[e++] = strdup(str);
}
if (bDispCorr)
{
sprintf(str, "f. <U\\sdisp c\\Ne\\S-\\betaU\\N>");
leg[e++] = strdup(str);
}
if (bCharge)
{
for (i = 0; i < ngid; i++)
{
sprintf(str, "f. <U\\sCoul %s\\Ne\\S-\\betaU\\N>",
*(groups->grpname[groups->grps[egcENER].nm_ind[i]]));
leg[e++] = strdup(str);
}
if (bRFExcl)
{
sprintf(str, "f. <U\\sRF excl\\Ne\\S-\\betaU\\N>");
leg[e++] = strdup(str);
}
if (EEL_FULL(fr->eeltype))
{
sprintf(str, "f. <U\\sCoul recip\\Ne\\S-\\betaU\\N>");
leg[e++] = strdup(str);
}
}
xvgr_legend(fp_tpi, 4+nener, (const char**)leg, oenv);
for (i = 0; i < 4+nener; i++)
{
sfree(leg[i]);
}
sfree(leg);
}
clear_rvec(x_init);
V_all = 0;
VembU_all = 0;
invbinw = 10;
nbin = 10;
snew(bin, nbin);
/* Avoid frame step numbers <= -1 */
frame_step_prev = -1;
bNotLastFrame = read_first_frame(oenv, &status, opt2fn("-rerun", nfile, fnm),
&rerun_fr, TRX_NEED_X);
frame = 0;
if (rerun_fr.natoms - (bCavity ? nat_cavity : 0) !=
mdatoms->nr - (a_tp1 - a_tp0))
{
gmx_fatal(FARGS, "Number of atoms in trajectory (%d)%s "
"is not equal the number in the run input file (%d) "
"minus the number of atoms to insert (%d)\n",
rerun_fr.natoms, bCavity ? " minus one" : "",
mdatoms->nr, a_tp1-a_tp0);
}
refvolshift = log(det(rerun_fr.box));
switch (inputrec->eI)
{
case eiTPI:
stepblocksize = inputrec->nstlist;
break;
case eiTPIC:
stepblocksize = 1;
break;
default:
gmx_fatal(FARGS, "Unknown integrator %s", ei_names[inputrec->eI]);
}
#ifdef GMX_SIMD
/* Make sure we don't detect SIMD overflow generated before this point */
gmx_simd_check_and_reset_overflow();
#endif
while (bNotLastFrame)
{
frame_step = rerun_fr.step;
if (frame_step <= frame_step_prev)
{
/* We don't have step number in the trajectory file,
* or we have constant or decreasing step numbers.
* Ensure we have increasing step numbers, since we use
* the step numbers as a counter for random numbers.
*/
frame_step = frame_step_prev + 1;
}
frame_step_prev = frame_step;
lambda = rerun_fr.lambda;
t = rerun_fr.time;
sum_embU = 0;
for (e = 0; e < nener; e++)
{
sum_UgembU[e] = 0;
}
/* Copy the coordinates from the input trajectory */
for (i = 0; i < rerun_fr.natoms; i++)
{
copy_rvec(rerun_fr.x[i], state->x[i]);
}
copy_mat(rerun_fr.box, state->box);
V = det(state->box);
logV = log(V);
bStateChanged = TRUE;
bNS = TRUE;
step = cr->nodeid*stepblocksize;
while (step < nsteps)
{
/* Initialize the second counter for random numbers using
* the insertion step index. This ensures that we get
* the same random numbers independently of how many
* MPI ranks we use. Also for the same seed, we get
* the same initial random sequence for different nsteps.
*/
rnd_count = step*rnd_count_stride;
if (!bCavity)
{
/* Random insertion in the whole volume */
bNS = (step % inputrec->nstlist == 0);
if (bNS)
{
/* Generate a random position in the box */
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd);
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd+2);
for (d = 0; d < DIM; d++)
{
x_init[d] = rnd[d]*state->box[d][d];
}
}
if (inputrec->nstlist == 1)
{
copy_rvec(x_init, x_tp);
}
else
{
/* Generate coordinates within |dx|=drmax of x_init */
do
{
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd);
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd+2);
for (d = 0; d < DIM; d++)
{
dx[d] = (2*rnd[d] - 1)*drmax;
}
}
while (norm2(dx) > drmax*drmax);
rvec_add(x_init, dx, x_tp);
}
}
else
{
/* Random insertion around a cavity location
* given by the last coordinate of the trajectory.
*/
if (step == 0)
{
if (nat_cavity == 1)
{
/* Copy the location of the cavity */
copy_rvec(rerun_fr.x[rerun_fr.natoms-1], x_init);
}
else
{
/* Determine the center of mass of the last molecule */
clear_rvec(x_init);
mass_tot = 0;
for (i = 0; i < nat_cavity; i++)
{
for (d = 0; d < DIM; d++)
{
x_init[d] +=
mass_cavity[i]*rerun_fr.x[rerun_fr.natoms-nat_cavity+i][d];
}
mass_tot += mass_cavity[i];
}
for (d = 0; d < DIM; d++)
{
x_init[d] /= mass_tot;
}
}
}
/* Generate coordinates within |dx|=drmax of x_init */
do
{
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd);
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd+2);
for (d = 0; d < DIM; d++)
{
dx[d] = (2*rnd[d] - 1)*drmax;
}
}
while (norm2(dx) > drmax*drmax);
rvec_add(x_init, dx, x_tp);
}
if (a_tp1 - a_tp0 == 1)
{
/* Insert a single atom, just copy the insertion location */
copy_rvec(x_tp, state->x[a_tp0]);
}
else
{
/* Copy the coordinates from the top file */
for (i = a_tp0; i < a_tp1; i++)
{
copy_rvec(x_mol[i-a_tp0], state->x[i]);
}
/* Rotate the molecule randomly */
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd);
gmx_rng_cycle_2uniform(frame_step, rnd_count++, seed, RND_SEED_TPI, rnd+2);
rotate_conf(a_tp1-a_tp0, state->x+a_tp0, NULL,
2*M_PI*rnd[0],
2*M_PI*rnd[1],
2*M_PI*rnd[2]);
/* Shift to the insertion location */
for (i = a_tp0; i < a_tp1; i++)
{
rvec_inc(state->x[i], x_tp);
}
}
/* Clear some matrix variables */
clear_mat(force_vir);
clear_mat(shake_vir);
clear_mat(vir);
clear_mat(pres);
/* Set the charge group center of mass of the test particle */
copy_rvec(x_init, fr->cg_cm[top->cgs.nr-1]);
/* Calc energy (no forces) on new positions.
* Since we only need the intermolecular energy
* and the RF exclusion terms of the inserted molecule occur
* within a single charge group we can pass NULL for the graph.
* This also avoids shifts that would move charge groups
* out of the box.
*
* Some checks above ensure than we can not have
* twin-range interactions together with nstlist > 1,
* therefore we do not need to remember the LR energies.
*/
/* Make do_force do a single node force calculation */
cr->nnodes = 1;
do_force(fplog, cr, inputrec,
step, nrnb, wcycle, top, &top_global->groups,
state->box, state->x, &state->hist,
f, force_vir, mdatoms, enerd, fcd,
state->lambda,
NULL, fr, NULL, mu_tot, t, NULL, NULL, FALSE,
GMX_FORCE_NONBONDED | GMX_FORCE_ENERGY |
(bNS ? GMX_FORCE_DYNAMICBOX | GMX_FORCE_NS | GMX_FORCE_DO_LR : 0) |
(bStateChanged ? GMX_FORCE_STATECHANGED : 0));
cr->nnodes = nnodes;
bStateChanged = FALSE;
bNS = FALSE;
/* Calculate long range corrections to pressure and energy */
calc_dispcorr(fplog, inputrec, fr, step, top_global->natoms, state->box,
lambda, pres, vir, &prescorr, &enercorr, &dvdlcorr);
/* figure out how to rearrange the next 4 lines MRS 8/4/2009 */
enerd->term[F_DISPCORR] = enercorr;
enerd->term[F_EPOT] += enercorr;
enerd->term[F_PRES] += prescorr;
enerd->term[F_DVDL_VDW] += dvdlcorr;
epot = enerd->term[F_EPOT];
bEnergyOutOfBounds = FALSE;
#ifdef GMX_SIMD_X86_SSE2_OR_HIGHER
/* With SSE the energy can overflow, check for this */
if (gmx_mm_check_and_reset_overflow())
{
if (debug)
{
fprintf(debug, "Found an SSE overflow, assuming the energy is out of bounds\n");
}
bEnergyOutOfBounds = TRUE;
}
#endif
/* If the compiler doesn't optimize this check away
* we catch the NAN energies.
* The epot>GMX_REAL_MAX check catches inf values,
* which should nicely result in embU=0 through the exp below,
* but it does not hurt to check anyhow.
*/
/* Non-bonded Interaction usually diverge at r=0.
* With tabulated interaction functions the first few entries
* should be capped in a consistent fashion between
* repulsion, dispersion and Coulomb to avoid accidental
* negative values in the total energy.
* The table generation code in tables.c does this.
* With user tbales the user should take care of this.
*/
if (epot != epot || epot > GMX_REAL_MAX)
{
bEnergyOutOfBounds = TRUE;
}
if (bEnergyOutOfBounds)
{
if (debug)
{
fprintf(debug, "\n time %.3f, step %d: non-finite energy %f, using exp(-bU)=0\n", t, (int)step, epot);
}
embU = 0;
}
else
{
embU = exp(-beta*epot);
sum_embU += embU;
/* Determine the weighted energy contributions of each energy group */
e = 0;
sum_UgembU[e++] += epot*embU;
if (fr->bBHAM)
{
for (i = 0; i < ngid; i++)
{
sum_UgembU[e++] +=
(enerd->grpp.ener[egBHAMSR][GID(i, gid_tp, ngid)] +
enerd->grpp.ener[egBHAMLR][GID(i, gid_tp, ngid)])*embU;
}
}
else
{
for (i = 0; i < ngid; i++)
{
sum_UgembU[e++] +=
(enerd->grpp.ener[egLJSR][GID(i, gid_tp, ngid)] +
enerd->grpp.ener[egLJLR][GID(i, gid_tp, ngid)])*embU;
}
}
if (bDispCorr)
{
sum_UgembU[e++] += enerd->term[F_DISPCORR]*embU;
}
if (bCharge)
{
for (i = 0; i < ngid; i++)
{
sum_UgembU[e++] +=
(enerd->grpp.ener[egCOULSR][GID(i, gid_tp, ngid)] +
enerd->grpp.ener[egCOULLR][GID(i, gid_tp, ngid)])*embU;
}
if (bRFExcl)
{
sum_UgembU[e++] += enerd->term[F_RF_EXCL]*embU;
}
if (EEL_FULL(fr->eeltype))
{
sum_UgembU[e++] += enerd->term[F_COUL_RECIP]*embU;
}
}
}
if (embU == 0 || beta*epot > bU_bin_limit)
{
bin[0]++;
}
else
{
i = (int)((bU_logV_bin_limit
- (beta*epot - logV + refvolshift))*invbinw
+ 0.5);
if (i < 0)
{
i = 0;
}
if (i >= nbin)
{
realloc_bins(&bin, &nbin, i+10);
}
bin[i]++;
}
if (debug)
{
fprintf(debug, "TPI %7d %12.5e %12.5f %12.5f %12.5f\n",
(int)step, epot, x_tp[XX], x_tp[YY], x_tp[ZZ]);
}
if (dump_pdb && epot <= dump_ener)
{
sprintf(str, "t%g_step%d.pdb", t, (int)step);
sprintf(str2, "t: %f step %d ener: %f", t, (int)step, epot);
write_sto_conf_mtop(str, str2, top_global, state->x, state->v,
inputrec->ePBC, state->box);
}
step++;
if ((step/stepblocksize) % cr->nnodes != cr->nodeid)
{
/* Skip all steps assigned to the other MPI ranks */
step += (cr->nnodes - 1)*stepblocksize;
}
}
if (PAR(cr))
{
/* When running in parallel sum the energies over the processes */
gmx_sumd(1, &sum_embU, cr);
gmx_sumd(nener, sum_UgembU, cr);
}
frame++;
V_all += V;
VembU_all += V*sum_embU/nsteps;
if (fp_tpi)
{
if (bVerbose || frame%10 == 0 || frame < 10)
{
fprintf(stderr, "mu %10.3e <mu> %10.3e\n",
-log(sum_embU/nsteps)/beta, -log(VembU_all/V_all)/beta);
}
fprintf(fp_tpi, "%10.3f %12.5e %12.5e %12.5e %12.5e",
t,
VembU_all == 0 ? 20/beta : -log(VembU_all/V_all)/beta,
sum_embU == 0 ? 20/beta : -log(sum_embU/nsteps)/beta,
sum_embU/nsteps, V);
for (e = 0; e < nener; e++)
{
fprintf(fp_tpi, " %12.5e", sum_UgembU[e]/nsteps);
}
fprintf(fp_tpi, "\n");
fflush(fp_tpi);
}
bNotLastFrame = read_next_frame(oenv, status, &rerun_fr);
} /* End of the loop */
walltime_accounting_end(walltime_accounting);
close_trj(status);
if (fp_tpi != NULL)
{
gmx_fio_fclose(fp_tpi);
}
if (fplog != NULL)
{
fprintf(fplog, "\n");
fprintf(fplog, " <V> = %12.5e nm^3\n", V_all/frame);
fprintf(fplog, " <mu> = %12.5e kJ/mol\n", -log(VembU_all/V_all)/beta);
}
/* Write the Boltzmann factor histogram */
if (PAR(cr))
{
/* When running in parallel sum the bins over the processes */
i = nbin;
global_max(cr, &i);
realloc_bins(&bin, &nbin, i);
gmx_sumd(nbin, bin, cr);
}
if (MASTER(cr))
{
fp_tpi = xvgropen(opt2fn("-tpid", nfile, fnm),
"TPI energy distribution",
"\\betaU - log(V/<V>)", "count", oenv);
sprintf(str, "number \\betaU > %g: %9.3e", bU_bin_limit, bin[0]);
xvgr_subtitle(fp_tpi, str, oenv);
xvgr_legend(fp_tpi, 2, (const char **)tpid_leg, oenv);
for (i = nbin-1; i > 0; i--)
{
bUlogV = -i/invbinw + bU_logV_bin_limit - refvolshift + log(V_all/frame);
fprintf(fp_tpi, "%6.2f %10d %12.5e\n",
bUlogV,
(int)(bin[i]+0.5),
bin[i]*exp(-bUlogV)*V_all/VembU_all);
}
gmx_fio_fclose(fp_tpi);
}
sfree(bin);
sfree(sum_UgembU);
walltime_accounting_set_nsteps_done(walltime_accounting, frame*inputrec->nsteps);
return 0;
}
static void
test_for_replica_exchange(FILE *fplog,
const gmx_multisim_t *ms,
struct gmx_repl_ex *re,
gmx_enerdata_t *enerd,
real vol,
gmx_int64_t step,
real time)
{
int m, i, j, a, b, ap, bp, i0, i1, tmp;
real ediff = 0, delta = 0, dpV = 0;
gmx_bool bPrint, bMultiEx;
gmx_bool *bEx = re->bEx;
real *prob = re->prob;
int *pind = re->destinations; /* permuted index */
gmx_bool bEpot = FALSE;
gmx_bool bDLambda = FALSE;
gmx_bool bVol = FALSE;
gmx_rng_t rng;
bMultiEx = (re->nex > 1); /* multiple exchanges at each state */
fprintf(fplog, "Replica exchange at step " "%"GMX_PRId64 " time %g\n", step, time);
if (re->bNPT)
{
for (i = 0; i < re->nrepl; i++)
{
re->Vol[i] = 0;
}
bVol = TRUE;
re->Vol[re->repl] = vol;
}
if ((re->type == ereTEMP || re->type == ereTL))
{
for (i = 0; i < re->nrepl; i++)
{
re->Epot[i] = 0;
}
bEpot = TRUE;
re->Epot[re->repl] = enerd->term[F_EPOT];
/* temperatures of different states*/
for (i = 0; i < re->nrepl; i++)
{
re->beta[i] = 1.0/(re->q[ereTEMP][i]*BOLTZ);
}
}
else
{
for (i = 0; i < re->nrepl; i++)
{
re->beta[i] = 1.0/(re->temp*BOLTZ); /* we have a single temperature */
}
}
if (re->type == ereLAMBDA || re->type == ereTL)
{
bDLambda = TRUE;
/* lambda differences. */
/* de[i][j] is the energy of the jth simulation in the ith Hamiltonian
minus the energy of the jth simulation in the jth Hamiltonian */
for (i = 0; i < re->nrepl; i++)
{
for (j = 0; j < re->nrepl; j++)
{
re->de[i][j] = 0;
}
}
for (i = 0; i < re->nrepl; i++)
{
re->de[i][re->repl] = (enerd->enerpart_lambda[(int)re->q[ereLAMBDA][i]+1]-enerd->enerpart_lambda[0]);
}
}
/* now actually do the communication */
if (bVol)
{
gmx_sum_sim(re->nrepl, re->Vol, ms);
}
if (bEpot)
{
gmx_sum_sim(re->nrepl, re->Epot, ms);
}
if (bDLambda)
{
for (i = 0; i < re->nrepl; i++)
{
gmx_sum_sim(re->nrepl, re->de[i], ms);
}
}
/* make a duplicate set of indices for shuffling */
for (i = 0; i < re->nrepl; i++)
{
pind[i] = re->ind[i];
}
if (bMultiEx)
{
/* multiple random switch exchange */
int nself = 0;
for (i = 0; i < re->nex + nself; i++)
{
double rnd[2];
gmx_rng_cycle_2uniform(step, i*2, re->seed, RND_SEED_REPLEX, rnd);
/* randomly select a pair */
/* in theory, could reduce this by identifying only which switches had a nonneglibible
probability of occurring (log p > -100) and only operate on those switches */
/* find out which state it is from, and what label that state currently has. Likely
more work that useful. */
i0 = (int)(re->nrepl*rnd[0]);
i1 = (int)(re->nrepl*rnd[1]);
if (i0 == i1)
{
nself++;
continue; /* self-exchange, back up and do it again */
}
a = re->ind[i0]; /* what are the indices of these states? */
b = re->ind[i1];
ap = pind[i0];
bp = pind[i1];
bPrint = FALSE; /* too noisy */
/* calculate the energy difference */
/* if the code changes to flip the STATES, rather than the configurations,
use the commented version of the code */
/* delta = calc_delta(fplog,bPrint,re,a,b,ap,bp); */
delta = calc_delta(fplog, bPrint, re, ap, bp, a, b);
/* we actually only use the first space in the prob and bEx array,
since there are actually many switches between pairs. */
if (delta <= 0)
{
/* accepted */
prob[0] = 1;
bEx[0] = TRUE;
}
else
{
if (delta > PROBABILITYCUTOFF)
{
prob[0] = 0;
}
else
{
prob[0] = exp(-delta);
}
/* roll a number to determine if accepted */
gmx_rng_cycle_2uniform(step, i*2+1, re->seed, RND_SEED_REPLEX, rnd);
bEx[0] = rnd[0] < prob[0];
}
re->prob_sum[0] += prob[0];
if (bEx[0])
{
/* swap the states */
tmp = pind[i0];
pind[i0] = pind[i1];
pind[i1] = tmp;
}
}
re->nattempt[0]++; /* keep track of total permutation trials here */
print_allswitchind(fplog, re->nrepl, pind, re->allswaps, re->tmpswap);
}
else
{
/* standard nearest neighbor replica exchange */
m = (step / re->nst) % 2;
for (i = 1; i < re->nrepl; i++)
{
a = re->ind[i-1];
b = re->ind[i];
bPrint = (re->repl == a || re->repl == b);
if (i % 2 == m)
{
delta = calc_delta(fplog, bPrint, re, a, b, a, b);
if (delta <= 0)
{
/* accepted */
prob[i] = 1;
bEx[i] = TRUE;
}
else
{
double rnd[2];
if (delta > PROBABILITYCUTOFF)
{
prob[i] = 0;
}
else
{
prob[i] = exp(-delta);
}
/* roll a number to determine if accepted */
gmx_rng_cycle_2uniform(step, i, re->seed, RND_SEED_REPLEX, rnd);
bEx[i] = rnd[0] < prob[i];
}
re->prob_sum[i] += prob[i];
if (bEx[i])
{
/* swap these two */
tmp = pind[i-1];
pind[i-1] = pind[i];
pind[i] = tmp;
re->nexchange[i]++; /* statistics for back compatibility */
}
}
else
{
prob[i] = -1;
bEx[i] = FALSE;
}
}
/* print some statistics */
print_ind(fplog, "ex", re->nrepl, re->ind, bEx);
print_prob(fplog, "pr", re->nrepl, prob);
fprintf(fplog, "\n");
re->nattempt[m]++;
}
/* record which moves were made and accepted */
for (i = 0; i < re->nrepl; i++)
{
re->nmoves[re->ind[i]][pind[i]] += 1;
re->nmoves[pind[i]][re->ind[i]] += 1;
}
fflush(fplog); /* make sure we can see what the last exchange was */
}
static int ChooseNewLambda(int nlim, t_expanded *expand, df_history_t *dfhist, int fep_state, real *weighted_lamee, double *p_k,
gmx_int64_t seed, gmx_int64_t step)
{
/* Choose new lambda value, and update transition matrix */
int i, ifep, minfep, maxfep, lamnew, lamtrial, starting_fep_state;
real r1, r2, de, trialprob, tprob = 0;
double *propose, *accept, *remainder;
double pks;
real pnorm;
starting_fep_state = fep_state;
lamnew = fep_state; /* so that there is a default setting -- stays the same */
if (!EWL(expand->elamstats)) /* ignore equilibrating the weights if using WL */
{
if ((expand->lmc_forced_nstart > 0) && (dfhist->n_at_lam[nlim-1] <= expand->lmc_forced_nstart))
{
/* Use a marching method to run through the lambdas and get preliminary free energy data,
before starting 'free' sampling. We start free sampling when we have enough at each lambda */
/* if we have enough at this lambda, move on to the next one */
if (dfhist->n_at_lam[fep_state] == expand->lmc_forced_nstart)
{
lamnew = fep_state+1;
if (lamnew == nlim) /* whoops, stepped too far! */
{
lamnew -= 1;
}
}
else
{
lamnew = fep_state;
}
return lamnew;
}
}
snew(propose, nlim);
snew(accept, nlim);
snew(remainder, nlim);
for (i = 0; i < expand->lmc_repeats; i++)
{
double rnd[2];
gmx_rng_cycle_2uniform(step, i, seed, RND_SEED_EXPANDED, rnd);
for (ifep = 0; ifep < nlim; ifep++)
{
propose[ifep] = 0;
accept[ifep] = 0;
}
if ((expand->elmcmove == elmcmoveGIBBS) || (expand->elmcmove == elmcmoveMETGIBBS))
{
/* use the Gibbs sampler, with restricted range */
if (expand->gibbsdeltalam < 0)
{
minfep = 0;
maxfep = nlim-1;
}
else
{
minfep = fep_state - expand->gibbsdeltalam;
maxfep = fep_state + expand->gibbsdeltalam;
if (minfep < 0)
{
minfep = 0;
}
if (maxfep > nlim-1)
{
maxfep = nlim-1;
}
}
GenerateGibbsProbabilities(weighted_lamee, p_k, &pks, minfep, maxfep);
if (expand->elmcmove == elmcmoveGIBBS)
{
for (ifep = minfep; ifep <= maxfep; ifep++)
{
propose[ifep] = p_k[ifep];
accept[ifep] = 1.0;
}
/* Gibbs sampling */
r1 = rnd[0];
for (lamnew = minfep; lamnew <= maxfep; lamnew++)
{
if (r1 <= p_k[lamnew])
{
break;
}
r1 -= p_k[lamnew];
}
}
else if (expand->elmcmove == elmcmoveMETGIBBS)
{
/* Metropolized Gibbs sampling */
for (ifep = minfep; ifep <= maxfep; ifep++)
{
remainder[ifep] = 1 - p_k[ifep];
}
/* find the proposal probabilities */
if (remainder[fep_state] == 0)
{
/* only the current state has any probability */
/* we have to stay at the current state */
lamnew = fep_state;
}
else
{
for (ifep = minfep; ifep <= maxfep; ifep++)
{
if (ifep != fep_state)
{
propose[ifep] = p_k[ifep]/remainder[fep_state];
}
else
{
propose[ifep] = 0;
}
}
r1 = rnd[0];
for (lamtrial = minfep; lamtrial <= maxfep; lamtrial++)
{
pnorm = p_k[lamtrial]/remainder[fep_state];
if (lamtrial != fep_state)
{
if (r1 <= pnorm)
{
break;
}
r1 -= pnorm;
}
}
/* we have now selected lamtrial according to p(lamtrial)/1-p(fep_state) */
tprob = 1.0;
/* trial probability is min{1,\frac{1 - p(old)}{1-p(new)} MRS 1/8/2008 */
trialprob = (remainder[fep_state])/(remainder[lamtrial]);
if (trialprob < tprob)
{
tprob = trialprob;
}
r2 = rnd[1];
if (r2 < tprob)
{
lamnew = lamtrial;
}
else
{
lamnew = fep_state;
}
}
/* now figure out the acceptance probability for each */
for (ifep = minfep; ifep <= maxfep; ifep++)
{
tprob = 1.0;
if (remainder[ifep] != 0)
{
trialprob = (remainder[fep_state])/(remainder[ifep]);
}
else
{
trialprob = 1.0; /* this state is the only choice! */
}
if (trialprob < tprob)
{
tprob = trialprob;
}
/* probability for fep_state=0, but that's fine, it's never proposed! */
accept[ifep] = tprob;
}
}
if (lamnew > maxfep)
{
/* it's possible some rounding is failing */
if (gmx_within_tol(remainder[fep_state], 0, 50*GMX_DOUBLE_EPS))
{
/* numerical rounding error -- no state other than the original has weight */
lamnew = fep_state;
}
else
{
/* probably not a numerical issue */
int loc = 0;
int nerror = 200+(maxfep-minfep+1)*60;
char *errorstr;
snew(errorstr, nerror);
/* if its greater than maxfep, then something went wrong -- probably underflow in the calculation
of sum weights. Generated detailed info for failure */
loc += sprintf(errorstr, "Something wrong in choosing new lambda state with a Gibbs move -- probably underflow in weight determination.\nDenominator is: %3d%17.10e\n i dE numerator weights\n", 0, pks);
for (ifep = minfep; ifep <= maxfep; ifep++)
{
loc += sprintf(&errorstr[loc], "%3d %17.10e%17.10e%17.10e\n", ifep, weighted_lamee[ifep], p_k[ifep], dfhist->sum_weights[ifep]);
}
gmx_fatal(FARGS, errorstr);
}
}
}
else if ((expand->elmcmove == elmcmoveMETROPOLIS) || (expand->elmcmove == elmcmoveBARKER))
{
/* use the metropolis sampler with trial +/- 1 */
r1 = rnd[0];
if (r1 < 0.5)
{
if (fep_state == 0)
{
lamtrial = fep_state;
}
else
{
lamtrial = fep_state-1;
}
}
else
{
if (fep_state == nlim-1)
{
lamtrial = fep_state;
}
else
{
lamtrial = fep_state+1;
}
}
de = weighted_lamee[lamtrial] - weighted_lamee[fep_state];
if (expand->elmcmove == elmcmoveMETROPOLIS)
{
tprob = 1.0;
trialprob = std::exp(de);
if (trialprob < tprob)
{
tprob = trialprob;
}
propose[fep_state] = 0;
propose[lamtrial] = 1.0; /* note that this overwrites the above line if fep_state = ntrial, which only occurs at the ends */
accept[fep_state] = 1.0; /* doesn't actually matter, never proposed unless fep_state = ntrial, in which case it's 1.0 anyway */
accept[lamtrial] = tprob;
}
else if (expand->elmcmove == elmcmoveBARKER)
{
tprob = 1.0/(1.0+std::exp(-de));
propose[fep_state] = (1-tprob);
propose[lamtrial] += tprob; /* we add, to account for the fact that at the end, they might be the same point */
accept[fep_state] = 1.0;
accept[lamtrial] = 1.0;
}
r2 = rnd[1];
if (r2 < tprob)
{
lamnew = lamtrial;
}
else
{
lamnew = fep_state;
}
}
for (ifep = 0; ifep < nlim; ifep++)
{
dfhist->Tij[fep_state][ifep] += propose[ifep]*accept[ifep];
dfhist->Tij[fep_state][fep_state] += propose[ifep]*(1.0-accept[ifep]);
}
fep_state = lamnew;
}
dfhist->Tij_empirical[starting_fep_state][lamnew] += 1.0;
sfree(propose);
sfree(accept);
sfree(remainder);
return lamnew;
}
