void pme_gpu_launch_complex_transforms(gmx_pme_t *pme,
gmx_wallcycle *wcycle)
{
PmeGpu *pmeGpu = pme->gpu;
const bool computeEnergyAndVirial = (pmeGpu->settings.currentFlags & GMX_PME_CALC_ENER_VIR) != 0;
const bool performBackFFT = (pmeGpu->settings.currentFlags & (GMX_PME_CALC_F | GMX_PME_CALC_POT)) != 0;
const unsigned int gridIndex = 0;
t_complex *cfftgrid = pme->cfftgrid[gridIndex];
if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
{
if (!pme_gpu_performs_FFT(pmeGpu))
{
wallcycle_start(wcycle, ewcWAIT_GPU_PME_SPREAD);
pme_gpu_sync_spread_grid(pme->gpu);
wallcycle_stop(wcycle, ewcWAIT_GPU_PME_SPREAD);
}
}
try
{
if (pmeGpu->settings.currentFlags & GMX_PME_SOLVE)
{
/* do R2C 3D-FFT */
parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_REAL_TO_COMPLEX, wcycle);
/* solve in k-space for our local cells */
if (pme_gpu_performs_solve(pmeGpu))
{
const auto gridOrdering = pme_gpu_uses_dd(pmeGpu) ? GridOrdering::YZX : GridOrdering::XYZ;
wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
pme_gpu_solve(pmeGpu, cfftgrid, gridOrdering, computeEnergyAndVirial);
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
}
else
{
wallcycle_start(wcycle, ewcPME_SOLVE_MIXED_MODE);
#pragma omp parallel for num_threads(pme->nthread) schedule(static)
for (int thread = 0; thread < pme->nthread; thread++)
{
solve_pme_yzx(pme, cfftgrid, pme->boxVolume,
computeEnergyAndVirial, pme->nthread, thread);
}
wallcycle_stop(wcycle, ewcPME_SOLVE_MIXED_MODE);
}
}
if (performBackFFT)
{
parallel_3dfft_execute_gpu_wrapper(pme, gridIndex, GMX_FFT_COMPLEX_TO_REAL, wcycle);
}
} GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
}
/*! \brief
* A convenience wrapper for launching either the GPU or CPU FFT.
*
* \param[in] pme The PME structure.
* \param[in] gridIndex The grid index - should currently always be 0.
* \param[in] dir The FFT direction enum.
* \param[in] wcycle The wallclock counter.
*/
void inline parallel_3dfft_execute_gpu_wrapper(gmx_pme_t *pme,
const int gridIndex,
enum gmx_fft_direction dir,
gmx_wallcycle_t wcycle)
{
GMX_ASSERT(gridIndex == 0, "Only single grid supported");
if (pme_gpu_performs_FFT(pme->gpu))
{
wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
pme_gpu_3dfft(pme->gpu, dir, gridIndex);
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
}
else
{
wallcycle_start(wcycle, ewcPME_FFT_MIXED_MODE);
#pragma omp parallel for num_threads(pme->nthread) schedule(static)
for (int thread = 0; thread < pme->nthread; thread++)
{
gmx_parallel_3dfft_execute(pme->pfft_setup[gridIndex], dir, thread, wcycle);
}
wallcycle_stop(wcycle, ewcPME_FFT_MIXED_MODE);
}
}
void pme_gpu_launch_spread(gmx_pme_t *pme,
const rvec *x,
gmx_wallcycle *wcycle)
{
GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
PmeGpu *pmeGpu = pme->gpu;
// The only spot of PME GPU where LAUNCH_GPU counter increases call-count
wallcycle_start(wcycle, ewcLAUNCH_GPU);
// The only spot of PME GPU where ewcsLAUNCH_GPU_PME subcounter increases call-count
wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_PME);
pme_gpu_copy_input_coordinates(pmeGpu, x);
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
const unsigned int gridIndex = 0;
real *fftgrid = pme->fftgrid[gridIndex];
if (pmeGpu->settings.currentFlags & GMX_PME_SPREAD)
{
/* Spread the coefficients on a grid */
const bool computeSplines = true;
const bool spreadCharges = true;
wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
pme_gpu_spread(pmeGpu, gridIndex, fftgrid, computeSplines, spreadCharges);
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
}
}
void mdoutf_tng_close(gmx_mdoutf_t of)
{
if (of->tng || of->tng_low_prec)
{
wallcycle_start(of->wcycle, ewcTRAJ);
gmx_tng_close(&of->tng);
gmx_tng_close(&of->tng_low_prec);
wallcycle_stop(of->wcycle, ewcTRAJ);
}
}
void wallcycle_start_nocount(gmx_wallcycle_t wc, int ewc)
{
if (wc == NULL)
{
return;
}
wallcycle_start(wc, ewc);
wc->wcc[ewc].n--;
}
static void reset_pmeonly_counters(gmx_wallcycle_t wcycle,
gmx_walltime_accounting_t walltime_accounting,
t_nrnb *nrnb,
int64_t step,
bool useGpuForPme)
{
/* Reset all the counters related to performance over the run */
wallcycle_stop(wcycle, ewcRUN);
wallcycle_reset_all(wcycle);
init_nrnb(nrnb);
wallcycle_start(wcycle, ewcRUN);
walltime_accounting_reset_time(walltime_accounting, step);
if (useGpuForPme)
{
resetGpuProfiler();
}
}
static void reset_pmeonly_counters(gmx_wallcycle_t wcycle,
gmx_walltime_accounting_t walltime_accounting,
t_nrnb *nrnb, t_inputrec *ir,
gmx_int64_t step)
{
/* Reset all the counters related to performance over the run */
wallcycle_stop(wcycle, ewcRUN);
wallcycle_reset_all(wcycle);
init_nrnb(nrnb);
if (ir->nsteps >= 0)
{
/* ir->nsteps is not used here, but we update it for consistency */
ir->nsteps -= step - ir->init_step;
}
ir->init_step = step;
wallcycle_start(wcycle, ewcRUN);
walltime_accounting_start(walltime_accounting);
}
bool pme_gpu_try_finish_task(const gmx_pme_t *pme,
gmx_wallcycle *wcycle,
gmx::ArrayRef<const gmx::RVec> *forces,
matrix virial,
real *energy,
GpuTaskCompletion completionKind)
{
GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
wallcycle_start_nocount(wcycle, ewcWAIT_GPU_PME_GATHER);
if (completionKind == GpuTaskCompletion::Check)
{
// Query the PME stream for completion of all tasks enqueued and
// if we're not done, stop the timer before early return.
if (!pme_gpu_stream_query(pme->gpu))
{
wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
return false;
}
}
else
{
// Synchronize the whole PME stream at once, including D2H result transfers.
pme_gpu_synchronize(pme->gpu);
}
wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
// Time the final staged data handling separately with a counting call to get
// the call count right.
wallcycle_start(wcycle, ewcWAIT_GPU_PME_GATHER);
pme_gpu_update_timings(pme->gpu);
pme_gpu_get_staged_results(pme, forces, virial, energy);
wallcycle_stop(wcycle, ewcWAIT_GPU_PME_GATHER);
return true;
}
void
do_md_trajectory_writing(FILE *fplog,
t_commrec *cr,
int nfile,
const t_filenm fnm[],
gmx_int64_t step,
gmx_int64_t step_rel,
double t,
t_inputrec *ir,
t_state *state,
t_state *state_global,
gmx_mtop_t *top_global,
t_forcerec *fr,
gmx_mdoutf_t outf,
t_mdebin *mdebin,
gmx_ekindata_t *ekind,
rvec *f,
rvec *f_global,
int *nchkpt,
gmx_bool bCPT,
gmx_bool bRerunMD,
gmx_bool bLastStep,
gmx_bool bDoConfOut,
gmx_bool bSumEkinhOld
)
{
int mdof_flags;
mdof_flags = 0;
if (do_per_step(step, ir->nstxout))
{
mdof_flags |= MDOF_X;
}
if (do_per_step(step, ir->nstvout))
{
mdof_flags |= MDOF_V;
}
if (do_per_step(step, ir->nstfout))
{
mdof_flags |= MDOF_F;
}
if (do_per_step(step, ir->nstxout_compressed))
{
mdof_flags |= MDOF_X_COMPRESSED;
}
if (bCPT)
{
mdof_flags |= MDOF_CPT;
}
;
#if defined(GMX_FAHCORE) || defined(GMX_WRITELASTSTEP)
if (bLastStep)
{
/* Enforce writing positions and velocities at end of run */
mdof_flags |= (MDOF_X | MDOF_V);
}
#endif
#ifdef GMX_FAHCORE
if (MASTER(cr))
{
fcReportProgress( ir->nsteps, step );
}
#if defined(__native_client__)
fcCheckin(MASTER(cr));
#endif
/* sync bCPT and fc record-keeping */
if (bCPT && MASTER(cr))
{
fcRequestCheckPoint();
}
#endif
if (mdof_flags != 0)
{
wallcycle_start(mdoutf_get_wcycle(outf), ewcTRAJ);
if (bCPT)
{
if (MASTER(cr))
{
if (bSumEkinhOld)
{
state_global->ekinstate.bUpToDate = FALSE;
}
else
{
update_ekinstate(&state_global->ekinstate, ekind);
state_global->ekinstate.bUpToDate = TRUE;
}
update_energyhistory(&state_global->enerhist, mdebin);
}
}
mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags, top_global,
step, t, state, state_global, f, f_global);
if (bCPT)
{
(*nchkpt)++;
}
debug_gmx();
if (bLastStep && step_rel == ir->nsteps &&
bDoConfOut && MASTER(cr) &&
!bRerunMD)
{
/* x and v have been collected in mdoutf_write_to_trajectory_files,
* because a checkpoint file will always be written
* at the last step.
*/
fprintf(stderr, "\nWriting final coordinates.\n");
if (fr->bMolPBC)
{
/* Make molecules whole only for confout writing */
do_pbc_mtop(fplog, ir->ePBC, state->box, top_global, state_global->x);
}
write_sto_conf_mtop(ftp2fn(efSTO, nfile, fnm),
*top_global->name, top_global,
state_global->x, state_global->v,
ir->ePBC, state->box);
debug_gmx();
}
wallcycle_stop(mdoutf_get_wcycle(outf), ewcTRAJ);
}
}
void compute_globals(FILE *fplog, gmx_global_stat_t gstat, t_commrec *cr, t_inputrec *ir,
t_forcerec *fr, gmx_ekindata_t *ekind,
t_state *state, t_state *state_global, t_mdatoms *mdatoms,
t_nrnb *nrnb, t_vcm *vcm, gmx_wallcycle_t wcycle,
gmx_enerdata_t *enerd, tensor force_vir, tensor shake_vir, tensor total_vir,
tensor pres, rvec mu_tot, gmx_constr_t constr,
globsig_t *gs, gmx_bool bInterSimGS,
matrix box, gmx_mtop_t *top_global,
gmx_bool *bSumEkinhOld, int flags)
{
int i, gsi;
real gs_buf[eglsNR];
tensor corr_vir, corr_pres;
gmx_bool bEner, bPres, bTemp;
gmx_bool bStopCM, bGStat,
bReadEkin, bEkinAveVel, bScaleEkin, bConstrain;
real prescorr, enercorr, dvdlcorr, dvdl_ekin;
/* translate CGLO flags to gmx_booleans */
bStopCM = flags & CGLO_STOPCM;
bGStat = flags & CGLO_GSTAT;
bReadEkin = (flags & CGLO_READEKIN);
bScaleEkin = (flags & CGLO_SCALEEKIN);
bEner = flags & CGLO_ENERGY;
bTemp = flags & CGLO_TEMPERATURE;
bPres = (flags & CGLO_PRESSURE);
bConstrain = (flags & CGLO_CONSTRAINT);
/* we calculate a full state kinetic energy either with full-step velocity verlet
or half step where we need the pressure */
bEkinAveVel = (ir->eI == eiVV || (ir->eI == eiVVAK && bPres) || bReadEkin);
/* in initalization, it sums the shake virial in vv, and to
sums ekinh_old in leapfrog (or if we are calculating ekinh_old) for other reasons */
/* ########## Kinetic energy ############## */
if (bTemp)
{
/* Non-equilibrium MD: this is parallellized, but only does communication
* when there really is NEMD.
*/
if (PAR(cr) && (ekind->bNEMD))
{
accumulate_u(cr, &(ir->opts), ekind);
}
debug_gmx();
if (bReadEkin)
{
restore_ekinstate_from_state(cr, ekind, &state_global->ekinstate);
}
else
{
calc_ke_part(state, &(ir->opts), mdatoms, ekind, nrnb, bEkinAveVel);
}
debug_gmx();
}
/* Calculate center of mass velocity if necessary, also parallellized */
if (bStopCM)
{
calc_vcm_grp(0, mdatoms->homenr, mdatoms,
state->x, state->v, vcm);
}
if (bTemp || bStopCM || bPres || bEner || bConstrain)
{
if (!bGStat)
{
/* We will not sum ekinh_old,
* so signal that we still have to do it.
*/
*bSumEkinhOld = TRUE;
}
else
{
if (gs != NULL)
{
for (i = 0; i < eglsNR; i++)
{
gs_buf[i] = gs->sig[i];
}
}
if (PAR(cr))
{
wallcycle_start(wcycle, ewcMoveE);
global_stat(fplog, gstat, cr, enerd, force_vir, shake_vir, mu_tot,
ir, ekind, constr, bStopCM ? vcm : NULL,
gs != NULL ? eglsNR : 0, gs_buf,
top_global, state,
*bSumEkinhOld, flags);
wallcycle_stop(wcycle, ewcMoveE);
}
if (gs != NULL)
{
if (MULTISIM(cr) && bInterSimGS)
{
if (MASTER(cr))
{
/* Communicate the signals between the simulations */
gmx_sum_sim(eglsNR, gs_buf, cr->ms);
}
/* Communicate the signals form the master to the others */
gmx_bcast(eglsNR*sizeof(gs_buf[0]), gs_buf, cr);
}
for (i = 0; i < eglsNR; i++)
{
if (bInterSimGS || gs_simlocal[i])
{
/* Set the communicated signal only when it is non-zero,
* since signals might not be processed at each MD step.
*/
gsi = (gs_buf[i] >= 0 ?
(int)(gs_buf[i] + 0.5) :
(int)(gs_buf[i] - 0.5));
if (gsi != 0)
{
gs->set[i] = gsi;
}
/* Turn off the local signal */
gs->sig[i] = 0;
}
}
}
*bSumEkinhOld = FALSE;
}
}
if (!ekind->bNEMD && debug && bTemp && (vcm->nr > 0))
{
correct_ekin(debug,
0, mdatoms->homenr,
state->v, vcm->group_p[0],
mdatoms->massT, mdatoms->tmass, ekind->ekin);
}
/* Do center of mass motion removal */
if (bStopCM)
{
check_cm_grp(fplog, vcm, ir, 1);
do_stopcm_grp(0, mdatoms->homenr, mdatoms->cVCM,
state->x, state->v, vcm);
inc_nrnb(nrnb, eNR_STOPCM, mdatoms->homenr);
}
if (bEner)
{
/* Calculate the amplitude of the cosine velocity profile */
ekind->cosacc.vcos = ekind->cosacc.mvcos/mdatoms->tmass;
}
if (bTemp)
{
/* Sum the kinetic energies of the groups & calc temp */
/* compute full step kinetic energies if vv, or if vv-avek and we are computing the pressure with IR_NPT_TROTTER */
/* three maincase: VV with AveVel (md-vv), vv with AveEkin (md-vv-avek), leap with AveEkin (md).
Leap with AveVel is not supported; it's not clear that it will actually work.
bEkinAveVel: If TRUE, we simply multiply ekin by ekinscale to get a full step kinetic energy.
If FALSE, we average ekinh_old and ekinh*ekinscale_nhc to get an averaged half step kinetic energy.
*/
enerd->term[F_TEMP] = sum_ekin(&(ir->opts), ekind, &dvdl_ekin,
bEkinAveVel, bScaleEkin);
enerd->dvdl_lin[efptMASS] = (double) dvdl_ekin;
enerd->term[F_EKIN] = trace(ekind->ekin);
}
/* ########## Long range energy information ###### */
if (bEner || bPres || bConstrain)
{
calc_dispcorr(ir, fr, top_global->natoms, box, state->lambda[efptVDW],
corr_pres, corr_vir, &prescorr, &enercorr, &dvdlcorr);
}
if (bEner)
{
enerd->term[F_DISPCORR] = enercorr;
enerd->term[F_EPOT] += enercorr;
enerd->term[F_DVDL_VDW] += dvdlcorr;
}
/* ########## Now pressure ############## */
if (bPres || bConstrain)
{
m_add(force_vir, shake_vir, total_vir);
/* Calculate pressure and apply LR correction if PPPM is used.
* Use the box from last timestep since we already called update().
*/
enerd->term[F_PRES] = calc_pres(fr->ePBC, ir->nwall, box, ekind->ekin, total_vir, pres);
/* Calculate long range corrections to pressure and energy */
/* this adds to enerd->term[F_PRES] and enerd->term[F_ETOT],
and computes enerd->term[F_DISPCORR]. Also modifies the
total_vir and pres tesors */
m_add(total_vir, corr_vir, total_vir);
m_add(pres, corr_pres, pres);
enerd->term[F_PDISPCORR] = prescorr;
enerd->term[F_PRES] += prescorr;
}
}
void
do_md_trajectory_writing(FILE *fplog,
t_commrec *cr,
int nfile,
const t_filenm fnm[],
gmx_int64_t step,
gmx_int64_t step_rel,
double t,
t_inputrec *ir,
t_state *state,
t_state *state_global,
gmx_mtop_t *top_global,
t_forcerec *fr,
gmx_mdoutf_t outf,
t_mdebin *mdebin,
gmx_ekindata_t *ekind,
rvec *f,
rvec *f_global,
int *nchkpt,
gmx_bool bCPT,
gmx_bool bRerunMD,
gmx_bool bLastStep,
gmx_bool bDoConfOut,
gmx_bool bSumEkinhOld
)
{
int mdof_flags;
rvec *x_for_confout = NULL;
mdof_flags = 0;
if (do_per_step(step, ir->nstxout))
{
mdof_flags |= MDOF_X;
}
if (do_per_step(step, ir->nstvout))
{
mdof_flags |= MDOF_V;
}
if (do_per_step(step, ir->nstfout))
{
mdof_flags |= MDOF_F;
}
if (do_per_step(step, ir->nstxout_compressed))
{
mdof_flags |= MDOF_X_COMPRESSED;
}
if (bCPT)
{
mdof_flags |= MDOF_CPT;
}
;
#if defined(GMX_FAHCORE)
if (bLastStep)
{
/* Enforce writing positions and velocities at end of run */
mdof_flags |= (MDOF_X | MDOF_V);
}
if (MASTER(cr))
{
fcReportProgress( ir->nsteps, step );
}
#if defined(__native_client__)
fcCheckin(MASTER(cr));
#endif
/* sync bCPT and fc record-keeping */
if (bCPT && MASTER(cr))
{
fcRequestCheckPoint();
}
#endif
if (mdof_flags != 0)
{
wallcycle_start(mdoutf_get_wcycle(outf), ewcTRAJ);
if (bCPT)
{
if (MASTER(cr))
{
if (bSumEkinhOld)
{
state_global->ekinstate.bUpToDate = FALSE;
}
else
{
update_ekinstate(&state_global->ekinstate, ekind);
state_global->ekinstate.bUpToDate = TRUE;
}
update_energyhistory(state_global->enerhist, mdebin);
}
}
mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags, top_global,
step, t, state, state_global, f, f_global);
if (bCPT)
{
(*nchkpt)++;
}
debug_gmx();
if (bLastStep && step_rel == ir->nsteps &&
bDoConfOut && MASTER(cr) &&
!bRerunMD)
{
if (fr->bMolPBC && state->x == state_global->x)
{
/* This (single-rank) run needs to allocate a
temporary array of size natoms so that any
periodicity removal for mdrun -confout does not
perturb the update and thus the final .edr
output. This makes .cpt restarts look binary
identical, and makes .edr restarts binary
identical. */
snew(x_for_confout, state_global->natoms);
copy_rvecn(state_global->x, x_for_confout, 0, state_global->natoms);
}
else
{
/* With DD, or no bMolPBC, it doesn't matter if
we change state_global->x */
x_for_confout = state_global->x;
}
/* x and v have been collected in mdoutf_write_to_trajectory_files,
* because a checkpoint file will always be written
* at the last step.
*/
fprintf(stderr, "\nWriting final coordinates.\n");
if (fr->bMolPBC)
{
/* Make molecules whole only for confout writing */
do_pbc_mtop(fplog, ir->ePBC, state->box, top_global, x_for_confout);
}
write_sto_conf_mtop(ftp2fn(efSTO, nfile, fnm),
*top_global->name, top_global,
x_for_confout, state_global->v,
ir->ePBC, state->box);
if (fr->bMolPBC && state->x == state_global->x)
{
sfree(x_for_confout);
}
debug_gmx();
}
wallcycle_stop(mdoutf_get_wcycle(outf), ewcTRAJ);
}
}
void do_force_lowlevel(FILE *fplog, gmx_large_int_t step,
t_forcerec *fr, t_inputrec *ir,
t_idef *idef, t_commrec *cr,
t_nrnb *nrnb, gmx_wallcycle_t wcycle,
t_mdatoms *md,
t_grpopts *opts,
rvec x[], history_t *hist,
rvec f[],
rvec f_longrange[],
gmx_enerdata_t *enerd,
t_fcdata *fcd,
gmx_mtop_t *mtop,
gmx_localtop_t *top,
gmx_genborn_t *born,
t_atomtypes *atype,
gmx_bool bBornRadii,
matrix box,
t_lambda *fepvals,
real *lambda,
t_graph *graph,
t_blocka *excl,
rvec mu_tot[],
int flags,
float *cycles_pme)
{
int i, j, status;
int donb_flags;
gmx_bool bDoEpot, bSepDVDL, bSB;
int pme_flags;
matrix boxs;
rvec box_size;
real Vsr, Vlr, Vcorr = 0;
t_pbc pbc;
real dvdgb;
char buf[22];
double clam_i, vlam_i;
real dvdl_dum[efptNR], dvdl, dvdl_nb[efptNR], lam_i[efptNR];
real dvdlsum;
#ifdef GMX_MPI
double t0 = 0.0, t1, t2, t3; /* time measurement for coarse load balancing */
#endif
#define PRINT_SEPDVDL(s, v, dvdlambda) if (bSepDVDL) {fprintf(fplog, sepdvdlformat, s, v, dvdlambda); }
GMX_MPE_LOG(ev_force_start);
set_pbc(&pbc, fr->ePBC, box);
/* reset free energy components */
for (i = 0; i < efptNR; i++)
{
dvdl_nb[i] = 0;
dvdl_dum[i] = 0;
}
/* Reset box */
for (i = 0; (i < DIM); i++)
{
box_size[i] = box[i][i];
}
bSepDVDL = (fr->bSepDVDL && do_per_step(step, ir->nstlog));
debug_gmx();
/* do QMMM first if requested */
if (fr->bQMMM)
{
enerd->term[F_EQM] = calculate_QMMM(cr, x, f, fr, md);
}
if (bSepDVDL)
{
fprintf(fplog, "Step %s: non-bonded V and dVdl for node %d:\n",
gmx_step_str(step, buf), cr->nodeid);
}
/* Call the short range functions all in one go. */
GMX_MPE_LOG(ev_do_fnbf_start);
#ifdef GMX_MPI
/*#define TAKETIME ((cr->npmenodes) && (fr->timesteps < 12))*/
#define TAKETIME FALSE
if (TAKETIME)
{
MPI_Barrier(cr->mpi_comm_mygroup);
t0 = MPI_Wtime();
}
#endif
if (ir->nwall)
{
/* foreign lambda component for walls */
dvdl = do_walls(ir, fr, box, md, x, f, lambda[efptVDW],
enerd->grpp.ener[egLJSR], nrnb);
PRINT_SEPDVDL("Walls", 0.0, dvdl);
enerd->dvdl_lin[efptVDW] += dvdl;
}
/* If doing GB, reset dvda and calculate the Born radii */
if (ir->implicit_solvent)
{
wallcycle_sub_start(wcycle, ewcsNONBONDED);
for (i = 0; i < born->nr; i++)
{
fr->dvda[i] = 0;
}
if (bBornRadii)
{
calc_gb_rad(cr, fr, ir, top, atype, x, &(fr->gblist), born, md, nrnb);
}
wallcycle_sub_stop(wcycle, ewcsNONBONDED);
}
where();
/* We only do non-bonded calculation with group scheme here, the verlet
* calls are done from do_force_cutsVERLET(). */
if (fr->cutoff_scheme == ecutsGROUP && (flags & GMX_FORCE_NONBONDED))
{
donb_flags = 0;
/* Add short-range interactions */
donb_flags |= GMX_NONBONDED_DO_SR;
if (flags & GMX_FORCE_FORCES)
{
donb_flags |= GMX_NONBONDED_DO_FORCE;
}
if (flags & GMX_FORCE_ENERGY)
{
donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
}
if (flags & GMX_FORCE_DO_LR)
{
donb_flags |= GMX_NONBONDED_DO_LR;
}
wallcycle_sub_start(wcycle, ewcsNONBONDED);
do_nonbonded(cr, fr, x, f, f_longrange, md, excl,
&enerd->grpp, box_size, nrnb,
lambda, dvdl_nb, -1, -1, donb_flags);
/* If we do foreign lambda and we have soft-core interactions
* we have to recalculate the (non-linear) energies contributions.
*/
if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
{
for (i = 0; i < enerd->n_lambda; i++)
{
for (j = 0; j < efptNR; j++)
{
lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
}
reset_foreign_enerdata(enerd);
do_nonbonded(cr, fr, x, f, f_longrange, md, excl,
&(enerd->foreign_grpp), box_size, nrnb,
lam_i, dvdl_dum, -1, -1,
(donb_flags & ~GMX_NONBONDED_DO_FORCE) | GMX_NONBONDED_DO_FOREIGNLAMBDA);
sum_epot(&ir->opts, &(enerd->foreign_grpp), enerd->foreign_term);
enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
}
}
wallcycle_sub_stop(wcycle, ewcsNONBONDED);
where();
}
/* If we are doing GB, calculate bonded forces and apply corrections
* to the solvation forces */
/* MRS: Eventually, many need to include free energy contribution here! */
if (ir->implicit_solvent)
{
wallcycle_sub_start(wcycle, ewcsBONDED);
calc_gb_forces(cr, md, born, top, atype, x, f, fr, idef,
ir->gb_algorithm, ir->sa_algorithm, nrnb, bBornRadii, &pbc, graph, enerd);
wallcycle_sub_stop(wcycle, ewcsBONDED);
}
#ifdef GMX_MPI
if (TAKETIME)
{
t1 = MPI_Wtime();
fr->t_fnbf += t1-t0;
}
#endif
if (fepvals->sc_alpha != 0)
{
enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
}
else
{
enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
}
if (fepvals->sc_alpha != 0)
/* even though coulomb part is linear, we already added it, beacuse we
need to go through the vdw calculation anyway */
{
enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
}
else
{
enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
}
Vsr = 0;
if (bSepDVDL)
{
for (i = 0; i < enerd->grpp.nener; i++)
{
Vsr +=
(fr->bBHAM ?
enerd->grpp.ener[egBHAMSR][i] :
enerd->grpp.ener[egLJSR][i])
+ enerd->grpp.ener[egCOULSR][i] + enerd->grpp.ener[egGB][i];
}
dvdlsum = dvdl_nb[efptVDW] + dvdl_nb[efptCOUL];
PRINT_SEPDVDL("VdW and Coulomb SR particle-p.", Vsr, dvdlsum);
}
debug_gmx();
GMX_MPE_LOG(ev_do_fnbf_finish);
if (debug)
{
pr_rvecs(debug, 0, "fshift after SR", fr->fshift, SHIFTS);
}
/* Shift the coordinates. Must be done before bonded forces and PPPM,
* but is also necessary for SHAKE and update, therefore it can NOT
* go when no bonded forces have to be evaluated.
*/
/* Here sometimes we would not need to shift with NBFonly,
* but we do so anyhow for consistency of the returned coordinates.
*/
if (graph)
{
shift_self(graph, box, x);
if (TRICLINIC(box))
{
inc_nrnb(nrnb, eNR_SHIFTX, 2*graph->nnodes);
}
else
{
inc_nrnb(nrnb, eNR_SHIFTX, graph->nnodes);
}
}
/* Check whether we need to do bondeds or correct for exclusions */
if (fr->bMolPBC &&
((flags & GMX_FORCE_BONDED)
|| EEL_RF(fr->eeltype) || EEL_FULL(fr->eeltype)))
{
/* Since all atoms are in the rectangular or triclinic unit-cell,
* only single box vector shifts (2 in x) are required.
*/
set_pbc_dd(&pbc, fr->ePBC, cr->dd, TRUE, box);
}
debug_gmx();
if (flags & GMX_FORCE_BONDED)
{
GMX_MPE_LOG(ev_calc_bonds_start);
wallcycle_sub_start(wcycle, ewcsBONDED);
calc_bonds(fplog, cr->ms,
idef, x, hist, f, fr, &pbc, graph, enerd, nrnb, lambda, md, fcd,
DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL, atype, born,
flags,
fr->bSepDVDL && do_per_step(step, ir->nstlog), step);
/* Check if we have to determine energy differences
* at foreign lambda's.
*/
if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) &&
idef->ilsort != ilsortNO_FE)
{
if (idef->ilsort != ilsortFE_SORTED)
{
gmx_incons("The bonded interactions are not sorted for free energy");
}
for (i = 0; i < enerd->n_lambda; i++)
{
reset_foreign_enerdata(enerd);
for (j = 0; j < efptNR; j++)
{
lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
}
calc_bonds_lambda(fplog, idef, x, fr, &pbc, graph, &(enerd->foreign_grpp), enerd->foreign_term, nrnb, lam_i, md,
fcd, DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL);
sum_epot(&ir->opts, &(enerd->foreign_grpp), enerd->foreign_term);
enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
}
}
debug_gmx();
GMX_MPE_LOG(ev_calc_bonds_finish);
wallcycle_sub_stop(wcycle, ewcsBONDED);
}
where();
*cycles_pme = 0;
if (EEL_FULL(fr->eeltype))
{
bSB = (ir->nwall == 2);
if (bSB)
{
copy_mat(box, boxs);
svmul(ir->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
box_size[ZZ] *= ir->wall_ewald_zfac;
}
clear_mat(fr->vir_el_recip);
if (fr->bEwald)
{
Vcorr = 0;
dvdl = 0;
/* With the Verlet scheme exclusion forces are calculated
* in the non-bonded kernel.
*/
/* The TPI molecule does not have exclusions with the rest
* of the system and no intra-molecular PME grid contributions
* will be calculated in gmx_pme_calc_energy.
*/
if ((ir->cutoff_scheme == ecutsGROUP && fr->n_tpi == 0) ||
ir->ewald_geometry != eewg3D ||
ir->epsilon_surface != 0)
{
int nthreads, t;
wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
if (fr->n_tpi > 0)
{
gmx_fatal(FARGS, "TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
}
nthreads = gmx_omp_nthreads_get(emntBonded);
#pragma omp parallel for num_threads(nthreads) schedule(static)
for (t = 0; t < nthreads; t++)
{
int s, e, i;
rvec *fnv;
tensor *vir;
real *Vcorrt, *dvdlt;
if (t == 0)
{
fnv = fr->f_novirsum;
vir = &fr->vir_el_recip;
Vcorrt = &Vcorr;
dvdlt = &dvdl;
}
else
{
fnv = fr->f_t[t].f;
vir = &fr->f_t[t].vir;
Vcorrt = &fr->f_t[t].Vcorr;
dvdlt = &fr->f_t[t].dvdl[efptCOUL];
for (i = 0; i < fr->natoms_force; i++)
{
clear_rvec(fnv[i]);
}
clear_mat(*vir);
}
*dvdlt = 0;
*Vcorrt =
ewald_LRcorrection(fplog,
fr->excl_load[t], fr->excl_load[t+1],
cr, t, fr,
md->chargeA,
md->nChargePerturbed ? md->chargeB : NULL,
ir->cutoff_scheme != ecutsVERLET,
excl, x, bSB ? boxs : box, mu_tot,
ir->ewald_geometry,
ir->epsilon_surface,
fnv, *vir,
lambda[efptCOUL], dvdlt);
}
if (nthreads > 1)
{
reduce_thread_forces(fr->natoms_force, fr->f_novirsum,
fr->vir_el_recip,
&Vcorr, efptCOUL, &dvdl,
nthreads, fr->f_t);
}
wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
}
if (fr->n_tpi == 0)
{
Vcorr += ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
&dvdl, fr->vir_el_recip);
}
PRINT_SEPDVDL("Ewald excl./charge/dip. corr.", Vcorr, dvdl);
enerd->dvdl_lin[efptCOUL] += dvdl;
}
status = 0;
Vlr = 0;
dvdl = 0;
switch (fr->eeltype)
{
case eelPME:
case eelPMESWITCH:
case eelPMEUSER:
case eelPMEUSERSWITCH:
case eelP3M_AD:
if (cr->duty & DUTY_PME)
{
assert(fr->n_tpi >= 0);
if (fr->n_tpi == 0 || (flags & GMX_FORCE_STATECHANGED))
{
pme_flags = GMX_PME_SPREAD_Q | GMX_PME_SOLVE;
if (flags & GMX_FORCE_FORCES)
{
pme_flags |= GMX_PME_CALC_F;
}
if (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY))
{
pme_flags |= GMX_PME_CALC_ENER_VIR;
}
if (fr->n_tpi > 0)
{
/* We don't calculate f, but we do want the potential */
pme_flags |= GMX_PME_CALC_POT;
}
wallcycle_start(wcycle, ewcPMEMESH);
status = gmx_pme_do(fr->pmedata,
md->start, md->homenr - fr->n_tpi,
x, fr->f_novirsum,
md->chargeA, md->chargeB,
bSB ? boxs : box, cr,
DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
nrnb, wcycle,
fr->vir_el_recip, fr->ewaldcoeff,
&Vlr, lambda[efptCOUL], &dvdl,
pme_flags);
*cycles_pme = wallcycle_stop(wcycle, ewcPMEMESH);
/* We should try to do as little computation after
* this as possible, because parallel PME synchronizes
* the nodes, so we want all load imbalance of the rest
* of the force calculation to be before the PME call.
* DD load balancing is done on the whole time of
* the force call (without PME).
*/
}
if (fr->n_tpi > 0)
{
/* Determine the PME grid energy of the test molecule
* with the PME grid potential of the other charges.
*/
gmx_pme_calc_energy(fr->pmedata, fr->n_tpi,
x + md->homenr - fr->n_tpi,
md->chargeA + md->homenr - fr->n_tpi,
&Vlr);
}
PRINT_SEPDVDL("PME mesh", Vlr, dvdl);
}
break;
case eelEWALD:
Vlr = do_ewald(fplog, FALSE, ir, x, fr->f_novirsum,
md->chargeA, md->chargeB,
box_size, cr, md->homenr,
fr->vir_el_recip, fr->ewaldcoeff,
lambda[efptCOUL], &dvdl, fr->ewald_table);
PRINT_SEPDVDL("Ewald long-range", Vlr, dvdl);
break;
default:
gmx_fatal(FARGS, "No such electrostatics method implemented %s",
eel_names[fr->eeltype]);
}
if (status != 0)
{
gmx_fatal(FARGS, "Error %d in long range electrostatics routine %s",
status, EELTYPE(fr->eeltype));
}
/* Note that with separate PME nodes we get the real energies later */
enerd->dvdl_lin[efptCOUL] += dvdl;
enerd->term[F_COUL_RECIP] = Vlr + Vcorr;
if (debug)
{
fprintf(debug, "Vlr = %g, Vcorr = %g, Vlr_corr = %g\n",
Vlr, Vcorr, enerd->term[F_COUL_RECIP]);
pr_rvecs(debug, 0, "vir_el_recip after corr", fr->vir_el_recip, DIM);
pr_rvecs(debug, 0, "fshift after LR Corrections", fr->fshift, SHIFTS);
}
}
else
{
if (EEL_RF(fr->eeltype))
{
/* With the Verlet scheme exclusion forces are calculated
* in the non-bonded kernel.
*/
if (ir->cutoff_scheme != ecutsVERLET && fr->eeltype != eelRF_NEC)
{
dvdl = 0;
enerd->term[F_RF_EXCL] =
RF_excl_correction(fplog, fr, graph, md, excl, x, f,
fr->fshift, &pbc, lambda[efptCOUL], &dvdl);
}
enerd->dvdl_lin[efptCOUL] += dvdl;
PRINT_SEPDVDL("RF exclusion correction",
enerd->term[F_RF_EXCL], dvdl);
}
}
where();
debug_gmx();
if (debug)
{
print_nrnb(debug, nrnb);
}
debug_gmx();
#ifdef GMX_MPI
if (TAKETIME)
{
t2 = MPI_Wtime();
MPI_Barrier(cr->mpi_comm_mygroup);
t3 = MPI_Wtime();
fr->t_wait += t3-t2;
if (fr->timesteps == 11)
{
fprintf(stderr, "* PP load balancing info: node %d, step %s, rel wait time=%3.0f%% , load string value: %7.2f\n",
cr->nodeid, gmx_step_str(fr->timesteps, buf),
100*fr->t_wait/(fr->t_wait+fr->t_fnbf),
(fr->t_fnbf+fr->t_wait)/fr->t_fnbf);
}
fr->timesteps++;
}
#endif
if (debug)
{
pr_rvecs(debug, 0, "fshift after bondeds", fr->fshift, SHIFTS);
}
GMX_MPE_LOG(ev_force_finish);
}
int gmx_pmeonly(struct gmx_pme_t *pme,
t_commrec *cr, t_nrnb *mynrnb,
gmx_wallcycle_t wcycle,
gmx_walltime_accounting_t walltime_accounting,
real ewaldcoeff_q, real ewaldcoeff_lj,
t_inputrec *ir)
{
int npmedata;
struct gmx_pme_t **pmedata;
gmx_pme_pp_t pme_pp;
int ret;
int natoms;
matrix box;
rvec *x_pp = NULL, *f_pp = NULL;
real *chargeA = NULL, *chargeB = NULL;
real *c6A = NULL, *c6B = NULL;
real *sigmaA = NULL, *sigmaB = NULL;
real lambda_q = 0;
real lambda_lj = 0;
int maxshift_x = 0, maxshift_y = 0;
real energy_q, energy_lj, dvdlambda_q, dvdlambda_lj;
matrix vir_q, vir_lj;
float cycles;
int count;
gmx_bool bEnerVir;
int pme_flags;
gmx_int64_t step, step_rel;
ivec grid_switch;
/* This data will only use with PME tuning, i.e. switching PME grids */
npmedata = 1;
snew(pmedata, npmedata);
pmedata[0] = pme;
pme_pp = gmx_pme_pp_init(cr);
init_nrnb(mynrnb);
count = 0;
do /****** this is a quasi-loop over time steps! */
{
/* The reason for having a loop here is PME grid tuning/switching */
do
{
/* Domain decomposition */
ret = gmx_pme_recv_coeffs_coords(pme_pp,
&natoms,
&chargeA, &chargeB,
&c6A, &c6B,
&sigmaA, &sigmaB,
box, &x_pp, &f_pp,
&maxshift_x, &maxshift_y,
&pme->bFEP_q, &pme->bFEP_lj,
&lambda_q, &lambda_lj,
&bEnerVir,
&pme_flags,
&step,
grid_switch, &ewaldcoeff_q, &ewaldcoeff_lj);
if (ret == pmerecvqxSWITCHGRID)
{
/* Switch the PME grid to grid_switch */
gmx_pmeonly_switch(&npmedata, &pmedata, grid_switch, cr, ir, &pme);
}
if (ret == pmerecvqxRESETCOUNTERS)
{
/* Reset the cycle and flop counters */
reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, ir, step);
}
}
while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
if (ret == pmerecvqxFINISH)
{
/* We should stop: break out of the loop */
break;
}
step_rel = step - ir->init_step;
if (count == 0)
{
wallcycle_start(wcycle, ewcRUN);
walltime_accounting_start(walltime_accounting);
}
wallcycle_start(wcycle, ewcPMEMESH);
dvdlambda_q = 0;
dvdlambda_lj = 0;
clear_mat(vir_q);
clear_mat(vir_lj);
gmx_pme_do(pme, 0, natoms, x_pp, f_pp,
chargeA, chargeB, c6A, c6B, sigmaA, sigmaB, box,
cr, maxshift_x, maxshift_y, mynrnb, wcycle,
vir_q, ewaldcoeff_q, vir_lj, ewaldcoeff_lj,
&energy_q, &energy_lj, lambda_q, lambda_lj, &dvdlambda_q, &dvdlambda_lj,
pme_flags | GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0));
cycles = wallcycle_stop(wcycle, ewcPMEMESH);
gmx_pme_send_force_vir_ener(pme_pp,
f_pp, vir_q, energy_q, vir_lj, energy_lj,
dvdlambda_q, dvdlambda_lj, cycles);
count++;
} /***** end of quasi-loop, we stop with the break above */
while (TRUE);
walltime_accounting_end(walltime_accounting);
return 0;
}
int gmx_pmeonly(struct gmx_pme_t *pme,
const t_commrec *cr, t_nrnb *mynrnb,
gmx_wallcycle *wcycle,
gmx_walltime_accounting_t walltime_accounting,
t_inputrec *ir, PmeRunMode runMode)
{
int ret;
int natoms = 0;
matrix box;
real lambda_q = 0;
real lambda_lj = 0;
int maxshift_x = 0, maxshift_y = 0;
real energy_q, energy_lj, dvdlambda_q, dvdlambda_lj;
matrix vir_q, vir_lj;
float cycles;
int count;
gmx_bool bEnerVir = FALSE;
int64_t step;
/* This data will only use with PME tuning, i.e. switching PME grids */
std::vector<gmx_pme_t *> pmedata;
pmedata.push_back(pme);
auto pme_pp = gmx_pme_pp_init(cr);
//TODO the variable below should be queried from the task assignment info
const bool useGpuForPme = (runMode == PmeRunMode::GPU) || (runMode == PmeRunMode::Mixed);
if (useGpuForPme)
{
changePinningPolicy(&pme_pp->chargeA, pme_get_pinning_policy());
changePinningPolicy(&pme_pp->x, pme_get_pinning_policy());
}
init_nrnb(mynrnb);
count = 0;
do /****** this is a quasi-loop over time steps! */
{
/* The reason for having a loop here is PME grid tuning/switching */
do
{
/* Domain decomposition */
ivec newGridSize;
bool atomSetChanged = false;
real ewaldcoeff_q = 0, ewaldcoeff_lj = 0;
ret = gmx_pme_recv_coeffs_coords(pme_pp.get(),
&natoms,
box,
&maxshift_x, &maxshift_y,
&lambda_q, &lambda_lj,
&bEnerVir,
&step,
&newGridSize,
&ewaldcoeff_q,
&ewaldcoeff_lj,
&atomSetChanged);
if (ret == pmerecvqxSWITCHGRID)
{
/* Switch the PME grid to newGridSize */
pme = gmx_pmeonly_switch(&pmedata, newGridSize, ewaldcoeff_q, ewaldcoeff_lj, cr, ir);
}
if (atomSetChanged)
{
gmx_pme_reinit_atoms(pme, natoms, pme_pp->chargeA.data());
}
if (ret == pmerecvqxRESETCOUNTERS)
{
/* Reset the cycle and flop counters */
reset_pmeonly_counters(wcycle, walltime_accounting, mynrnb, step, useGpuForPme);
}
}
while (ret == pmerecvqxSWITCHGRID || ret == pmerecvqxRESETCOUNTERS);
if (ret == pmerecvqxFINISH)
{
/* We should stop: break out of the loop */
break;
}
if (count == 0)
{
wallcycle_start(wcycle, ewcRUN);
walltime_accounting_start_time(walltime_accounting);
}
wallcycle_start(wcycle, ewcPMEMESH);
dvdlambda_q = 0;
dvdlambda_lj = 0;
clear_mat(vir_q);
clear_mat(vir_lj);
energy_q = 0;
energy_lj = 0;
// TODO Make a struct of array refs onto these per-atom fields
// of pme_pp (maybe box, energy and virial, too; and likewise
// from mdatoms for the other call to gmx_pme_do), so we have
// fewer lines of code and less parameter passing.
const int pmeFlags = GMX_PME_DO_ALL_F | (bEnerVir ? GMX_PME_CALC_ENER_VIR : 0);
gmx::ArrayRef<const gmx::RVec> forces;
if (useGpuForPme)
{
const bool boxChanged = false;
//TODO this should be set properly by gmx_pme_recv_coeffs_coords,
// or maybe use inputrecDynamicBox(ir), at the very least - change this when this codepath is tested!
pme_gpu_prepare_computation(pme, boxChanged, box, wcycle, pmeFlags);
pme_gpu_launch_spread(pme, pme_pp->x.rvec_array(), wcycle);
pme_gpu_launch_complex_transforms(pme, wcycle);
pme_gpu_launch_gather(pme, wcycle, PmeForceOutputHandling::Set);
pme_gpu_wait_finish_task(pme, wcycle, &forces, vir_q, &energy_q);
pme_gpu_reinit_computation(pme, wcycle);
}
else
{
gmx_pme_do(pme, 0, natoms, pme_pp->x.rvec_array(), as_rvec_array(pme_pp->f.data()),
pme_pp->chargeA.data(), pme_pp->chargeB.data(),
pme_pp->sqrt_c6A.data(), pme_pp->sqrt_c6B.data(),
pme_pp->sigmaA.data(), pme_pp->sigmaB.data(), box,
cr, maxshift_x, maxshift_y, mynrnb, wcycle,
vir_q, vir_lj,
&energy_q, &energy_lj, lambda_q, lambda_lj, &dvdlambda_q, &dvdlambda_lj,
pmeFlags);
forces = pme_pp->f;
}
cycles = wallcycle_stop(wcycle, ewcPMEMESH);
gmx_pme_send_force_vir_ener(pme_pp.get(), as_rvec_array(forces.data()),
vir_q, energy_q, vir_lj, energy_lj,
dvdlambda_q, dvdlambda_lj, cycles);
count++;
} /***** end of quasi-loop, we stop with the break above */
while (TRUE);
walltime_accounting_end_time(walltime_accounting);
return 0;
}
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;
}
double do_md_openmm(FILE *fplog,t_commrec *cr,int nfile,const t_filenm fnm[],
const output_env_t oenv, gmx_bool bVerbose,gmx_bool bCompact,
int nstglobalcomm,
gmx_vsite_t *vsite,gmx_constr_t constr,
int stepout,t_inputrec *ir,
gmx_mtop_t *top_global,
t_fcdata *fcd,
t_state *state_global,
t_mdatoms *mdatoms,
t_nrnb *nrnb,gmx_wallcycle_t wcycle,
gmx_edsam_t ed,t_forcerec *fr,
int repl_ex_nst,int repl_ex_seed,
real cpt_period,real max_hours,
const char *deviceOptions,
unsigned long Flags,
gmx_runtime_t *runtime)
{
gmx_mdoutf_t *outf;
gmx_large_int_t step,step_rel;
double run_time;
double t,t0,lam0;
gmx_bool bSimAnn,
bFirstStep,bStateFromTPX,bLastStep,bStartingFromCpt;
gmx_bool bInitStep=TRUE;
gmx_bool do_ene,do_log, do_verbose,
bX,bV,bF,bCPT;
tensor force_vir,shake_vir,total_vir,pres;
int i,m;
int mdof_flags;
rvec mu_tot;
t_vcm *vcm;
int nchkpt=1;
gmx_localtop_t *top;
t_mdebin *mdebin=NULL;
t_state *state=NULL;
rvec *f_global=NULL;
int n_xtc=-1;
rvec *x_xtc=NULL;
gmx_enerdata_t *enerd;
rvec *f=NULL;
gmx_global_stat_t gstat;
gmx_update_t upd=NULL;
t_graph *graph=NULL;
globsig_t gs;
gmx_groups_t *groups;
gmx_ekindata_t *ekind, *ekind_save;
gmx_bool bAppend;
int a0,a1;
matrix lastbox;
real reset_counters=0,reset_counters_now=0;
char sbuf[STEPSTRSIZE],sbuf2[STEPSTRSIZE];
int handled_stop_condition=gmx_stop_cond_none;
const char *ommOptions = NULL;
void *openmmData;
bAppend = (Flags & MD_APPENDFILES);
check_ir_old_tpx_versions(cr,fplog,ir,top_global);
groups = &top_global->groups;
/* Initial values */
init_md(fplog,cr,ir,oenv,&t,&t0,&state_global->lambda,&lam0,
nrnb,top_global,&upd,
nfile,fnm,&outf,&mdebin,
force_vir,shake_vir,mu_tot,&bSimAnn,&vcm,state_global,Flags);
clear_mat(total_vir);
clear_mat(pres);
/* Energy terms and groups */
snew(enerd,1);
init_enerdata(top_global->groups.grps[egcENER].nr,ir->n_flambda,enerd);
snew(f,top_global->natoms);
/* Kinetic energy data */
snew(ekind,1);
init_ekindata(fplog,top_global,&(ir->opts),ekind);
/* needed for iteration of constraints */
snew(ekind_save,1);
init_ekindata(fplog,top_global,&(ir->opts),ekind_save);
/* Copy the cos acceleration to the groups struct */
ekind->cosacc.cos_accel = ir->cos_accel;
gstat = global_stat_init(ir);
debug_gmx();
{
double io = compute_io(ir,top_global->natoms,groups,mdebin->ebin->nener,1);
if ((io > 2000) && MASTER(cr))
fprintf(stderr,
"\nWARNING: This run will generate roughly %.0f Mb of data\n\n",
io);
}
top = gmx_mtop_generate_local_top(top_global,ir);
a0 = 0;
a1 = top_global->natoms;
state = partdec_init_local_state(cr,state_global);
f_global = f;
atoms2md(top_global,ir,0,NULL,a0,a1-a0,mdatoms);
if (vsite)
{
set_vsite_top(vsite,top,mdatoms,cr);
}
if (ir->ePBC != epbcNONE && !ir->bPeriodicMols)
{
graph = mk_graph(fplog,&(top->idef),0,top_global->natoms,FALSE,FALSE);
}
update_mdatoms(mdatoms,state->lambda);
if (deviceOptions[0]=='\0')
{
/* empty options, which should default to OpenMM in this build */
ommOptions=deviceOptions;
}
else
{
if (gmx_strncasecmp(deviceOptions,"OpenMM",6)!=0)
{
gmx_fatal(FARGS, "This Gromacs version currently only works with OpenMM. Use -device \"OpenMM:<options>\"");
}
else
{
ommOptions=strchr(deviceOptions,':');
if (NULL!=ommOptions)
{
/* Increase the pointer to skip the colon */
ommOptions++;
}
}
}
openmmData = openmm_init(fplog, ommOptions, ir, top_global, top, mdatoms, fr, state);
please_cite(fplog,"Friedrichs2009");
if (MASTER(cr))
{
/* Update mdebin with energy history if appending to output files */
if ( Flags & MD_APPENDFILES )
{
restore_energyhistory_from_state(mdebin,&state_global->enerhist);
}
/* Set the initial energy history in state to zero by updating once */
update_energyhistory(&state_global->enerhist,mdebin);
}
if (constr)
{
set_constraints(constr,top,ir,mdatoms,cr);
}
if (!ir->bContinuation)
{
if (mdatoms->cFREEZE && (state->flags & (1<<estV)))
{
/* Set the velocities of frozen particles to zero */
for (i=mdatoms->start; i<mdatoms->start+mdatoms->homenr; i++)
{
for (m=0; m<DIM; m++)
{
if (ir->opts.nFreeze[mdatoms->cFREEZE[i]][m])
{
state->v[i][m] = 0;
}
}
}
}
if (constr)
{
/* Constrain the initial coordinates and velocities */
do_constrain_first(fplog,constr,ir,mdatoms,state,f,
graph,cr,nrnb,fr,top,shake_vir);
}
if (vsite)
{
/* Construct the virtual sites for the initial configuration */
construct_vsites(fplog,vsite,state->x,nrnb,ir->delta_t,NULL,
top->idef.iparams,top->idef.il,
fr->ePBC,fr->bMolPBC,graph,cr,state->box);
}
}
debug_gmx();
if (MASTER(cr))
{
char tbuf[20];
fprintf(fplog,"Initial temperature: %g K\n",enerd->term[F_TEMP]);
fprintf(stderr,"starting mdrun '%s'\n",
*(top_global->name));
if (ir->nsteps >= 0)
{
sprintf(tbuf,"%8.1f",(ir->init_step+ir->nsteps)*ir->delta_t);
}
else
{
sprintf(tbuf,"%s","infinite");
}
if (ir->init_step > 0)
{
fprintf(stderr,"%s steps, %s ps (continuing from step %s, %8.1f ps).\n",
gmx_step_str(ir->init_step+ir->nsteps,sbuf),tbuf,
gmx_step_str(ir->init_step,sbuf2),
ir->init_step*ir->delta_t);
}
else
{
fprintf(stderr,"%s steps, %s ps.\n",
gmx_step_str(ir->nsteps,sbuf),tbuf);
}
}
fprintf(fplog,"\n");
/* Set and write start time */
runtime_start(runtime);
print_date_and_time(fplog,cr->nodeid,"Started mdrun",runtime);
wallcycle_start(wcycle,ewcRUN);
if (fplog)
fprintf(fplog,"\n");
/* safest point to do file checkpointing is here. More general point would be immediately before integrator call */
debug_gmx();
/***********************************************************
*
* Loop over MD steps
*
************************************************************/
/* loop over MD steps or if rerunMD to end of input trajectory */
bFirstStep = TRUE;
/* Skip the first Nose-Hoover integration when we get the state from tpx */
bStateFromTPX = !opt2bSet("-cpi",nfile,fnm);
bInitStep = bFirstStep && bStateFromTPX;
bStartingFromCpt = (Flags & MD_STARTFROMCPT) && bInitStep;
bLastStep = FALSE;
init_global_signals(&gs,cr,ir,repl_ex_nst);
step = ir->init_step;
step_rel = 0;
while (!bLastStep)
{
wallcycle_start(wcycle,ewcSTEP);
GMX_MPE_LOG(ev_timestep1);
bLastStep = (step_rel == ir->nsteps);
t = t0 + step*ir->delta_t;
if (gs.set[eglsSTOPCOND] != 0)
{
bLastStep = TRUE;
}
do_log = do_per_step(step,ir->nstlog) || bFirstStep || bLastStep;
do_verbose = bVerbose &&
(step % stepout == 0 || bFirstStep || bLastStep);
if (MASTER(cr) && do_log)
{
print_ebin_header(fplog,step,t,state->lambda);
}
clear_mat(force_vir);
GMX_MPE_LOG(ev_timestep2);
/* We write a checkpoint at this MD step when:
* either when we signalled through gs (in OpenMM NS works different),
* or at the last step (but not when we do not want confout),
* but never at the first step.
*/
bCPT = ((gs.set[eglsCHKPT] ||
(bLastStep && (Flags & MD_CONFOUT))) &&
step > ir->init_step );
if (bCPT)
{
gs.set[eglsCHKPT] = 0;
}
/* Now we have the energies and forces corresponding to the
* coordinates at time t. We must output all of this before
* the update.
* for RerunMD t is read from input trajectory
*/
GMX_MPE_LOG(ev_output_start);
mdof_flags = 0;
if (do_per_step(step,ir->nstxout))
{
mdof_flags |= MDOF_X;
}
if (do_per_step(step,ir->nstvout))
{
mdof_flags |= MDOF_V;
}
if (do_per_step(step,ir->nstfout))
{
mdof_flags |= MDOF_F;
}
if (do_per_step(step,ir->nstxtcout))
{
mdof_flags |= MDOF_XTC;
}
if (bCPT)
{
mdof_flags |= MDOF_CPT;
};
do_ene = (do_per_step(step,ir->nstenergy) || bLastStep);
if (mdof_flags != 0 || do_ene || do_log)
{
wallcycle_start(wcycle,ewcTRAJ);
bF = (mdof_flags & MDOF_F);
bX = (mdof_flags & (MDOF_X | MDOF_XTC | MDOF_CPT));
bV = (mdof_flags & (MDOF_V | MDOF_CPT));
openmm_copy_state(openmmData, state, &t, f, enerd, bX, bV, bF, do_ene);
upd_mdebin(mdebin, FALSE,TRUE,
t,mdatoms->tmass,enerd,state,lastbox,
shake_vir,force_vir,total_vir,pres,
ekind,mu_tot,constr);
print_ebin(outf->fp_ene,do_ene,FALSE,FALSE,do_log?fplog:NULL,
step,t,
eprNORMAL,bCompact,mdebin,fcd,groups,&(ir->opts));
write_traj(fplog,cr,outf,mdof_flags,top_global,
step,t,state,state_global,f,f_global,&n_xtc,&x_xtc);
if (bCPT)
{
nchkpt++;
bCPT = FALSE;
}
debug_gmx();
if (bLastStep && step_rel == ir->nsteps &&
(Flags & MD_CONFOUT) && MASTER(cr))
{
/* x and v have been collected in write_traj,
* because a checkpoint file will always be written
* at the last step.
*/
fprintf(stderr,"\nWriting final coordinates.\n");
if (ir->ePBC != epbcNONE && !ir->bPeriodicMols)
{
/* Make molecules whole only for confout writing */
do_pbc_mtop(fplog,ir->ePBC,state->box,top_global,state_global->x);
}
write_sto_conf_mtop(ftp2fn(efSTO,nfile,fnm),
*top_global->name,top_global,
state_global->x,state_global->v,
ir->ePBC,state->box);
debug_gmx();
}
wallcycle_stop(wcycle,ewcTRAJ);
}
GMX_MPE_LOG(ev_output_finish);
/* Determine the wallclock run time up till now */
run_time = gmx_gettime() - (double)runtime->real;
/* Check whether everything is still allright */
if (((int)gmx_get_stop_condition() > handled_stop_condition)
#ifdef GMX_THREADS
&& MASTER(cr)
#endif
)
{
/* this is just make gs.sig compatible with the hack
of sending signals around by MPI_Reduce with together with
other floats */
/* NOTE: this only works for serial code. For code that allows
MPI nodes to propagate their condition, see kernel/md.c*/
if ( gmx_get_stop_condition() == gmx_stop_cond_next_ns )
gs.set[eglsSTOPCOND]=1;
if ( gmx_get_stop_condition() == gmx_stop_cond_next )
gs.set[eglsSTOPCOND]=1;
/* < 0 means stop at next step, > 0 means stop at next NS step */
if (fplog)
{
fprintf(fplog,
"\n\nReceived the %s signal, stopping at the next %sstep\n\n",
gmx_get_signal_name(),
gs.sig[eglsSTOPCOND]==1 ? "NS " : "");
fflush(fplog);
}
fprintf(stderr,
"\n\nReceived the %s signal, stopping at the next %sstep\n\n",
gmx_get_signal_name(),
gs.sig[eglsSTOPCOND]==1 ? "NS " : "");
fflush(stderr);
handled_stop_condition=(int)gmx_get_stop_condition();
}
else if (MASTER(cr) &&
(max_hours > 0 && run_time > max_hours*60.0*60.0*0.99) &&
gs.set[eglsSTOPCOND] == 0)
{
/* Signal to terminate the run */
gs.set[eglsSTOPCOND] = 1;
if (fplog)
{
fprintf(fplog,"\nStep %s: Run time exceeded %.3f hours, will terminate the run\n",gmx_step_str(step,sbuf),max_hours*0.99);
}
fprintf(stderr, "\nStep %s: Run time exceeded %.3f hours, will terminate the run\n",gmx_step_str(step,sbuf),max_hours*0.99);
}
/* checkpoints */
if (MASTER(cr) && (cpt_period >= 0 &&
(cpt_period == 0 ||
run_time >= nchkpt*cpt_period*60.0)) &&
gs.set[eglsCHKPT] == 0)
{
gs.set[eglsCHKPT] = 1;
}
/* Time for performance */
if (((step % stepout) == 0) || bLastStep)
{
runtime_upd_proc(runtime);
}
if (do_per_step(step,ir->nstlog))
{
if (fflush(fplog) != 0)
{
gmx_fatal(FARGS,"Cannot flush logfile - maybe you are out of quota?");
}
}
/* Remaining runtime */
if (MULTIMASTER(cr) && (do_verbose || gmx_got_usr_signal() ))
{
print_time(stderr,runtime,step,ir,cr);
}
bFirstStep = FALSE;
bInitStep = FALSE;
bStartingFromCpt = FALSE;
step++;
step_rel++;
openmm_take_one_step(openmmData);
}
/* End of main MD loop */
debug_gmx();
/* Stop the time */
runtime_end(runtime);
if (MASTER(cr))
{
if (ir->nstcalcenergy > 0)
{
print_ebin(outf->fp_ene,FALSE,FALSE,FALSE,fplog,step,t,
eprAVER,FALSE,mdebin,fcd,groups,&(ir->opts));
}
}
openmm_cleanup(fplog, openmmData);
done_mdoutf(outf);
debug_gmx();
runtime->nsteps_done = step_rel;
return 0;
}
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 do_force_lowlevel(FILE *fplog, gmx_large_int_t step,
t_forcerec *fr, t_inputrec *ir,
t_idef *idef, t_commrec *cr,
t_nrnb *nrnb, gmx_wallcycle_t wcycle,
t_mdatoms *md,
t_grpopts *opts,
rvec x[], history_t *hist,
rvec f[],
gmx_enerdata_t *enerd,
t_fcdata *fcd,
gmx_mtop_t *mtop,
gmx_localtop_t *top,
gmx_genborn_t *born,
t_atomtypes *atype,
gmx_bool bBornRadii,
matrix box,
real lambda,
t_graph *graph,
t_blocka *excl,
rvec mu_tot[],
int flags,
float *cycles_pme)
{
int i,status;
int donb_flags;
gmx_bool bDoEpot,bSepDVDL,bSB;
int pme_flags;
matrix boxs;
rvec box_size;
real dvdlambda,Vsr,Vlr,Vcorr=0,vdip,vcharge;
t_pbc pbc;
real dvdgb;
char buf[22];
gmx_enerdata_t ed_lam;
double lam_i;
real dvdl_dum;
#ifdef GMX_MPI
double t0=0.0,t1,t2,t3; /* time measurement for coarse load balancing */
#endif
#define PRINT_SEPDVDL(s,v,dvdl) if (bSepDVDL) fprintf(fplog,sepdvdlformat,s,v,dvdl);
GMX_MPE_LOG(ev_force_start);
set_pbc(&pbc,fr->ePBC,box);
/* Reset box */
for(i=0; (i<DIM); i++)
{
box_size[i]=box[i][i];
}
bSepDVDL=(fr->bSepDVDL && do_per_step(step,ir->nstlog));
debug_gmx();
/* do QMMM first if requested */
if(fr->bQMMM)
{
enerd->term[F_EQM] = calculate_QMMM(cr,x,f,fr,md);
}
if (bSepDVDL)
{
fprintf(fplog,"Step %s: non-bonded V and dVdl for node %d:\n",
gmx_step_str(step,buf),cr->nodeid);
}
/* Call the short range functions all in one go. */
GMX_MPE_LOG(ev_do_fnbf_start);
dvdlambda = 0;
#ifdef GMX_MPI
/*#define TAKETIME ((cr->npmenodes) && (fr->timesteps < 12))*/
#define TAKETIME FALSE
if (TAKETIME)
{
MPI_Barrier(cr->mpi_comm_mygroup);
t0=MPI_Wtime();
}
#endif
if (ir->nwall)
{
dvdlambda = do_walls(ir,fr,box,md,x,f,lambda,
enerd->grpp.ener[egLJSR],nrnb);
PRINT_SEPDVDL("Walls",0.0,dvdlambda);
enerd->dvdl_lin += dvdlambda;
}
/* If doing GB, reset dvda and calculate the Born radii */
if (ir->implicit_solvent)
{
/* wallcycle_start(wcycle,ewcGB); */
for(i=0; i<born->nr; i++)
{
fr->dvda[i]=0;
}
if(bBornRadii)
{
calc_gb_rad(cr,fr,ir,top,atype,x,&(fr->gblist),born,md,nrnb);
}
/* wallcycle_stop(wcycle, ewcGB); */
}
where();
donb_flags = 0;
if (flags & GMX_FORCE_FORCES)
{
donb_flags |= GMX_DONB_FORCES;
}
do_nonbonded(cr,fr,x,f,md,excl,
fr->bBHAM ?
enerd->grpp.ener[egBHAMSR] :
enerd->grpp.ener[egLJSR],
enerd->grpp.ener[egCOULSR],
enerd->grpp.ener[egGB],box_size,nrnb,
lambda,&dvdlambda,-1,-1,donb_flags);
/* If we do foreign lambda and we have soft-core interactions
* we have to recalculate the (non-linear) energies contributions.
*/
if (ir->n_flambda > 0 && (flags & GMX_FORCE_DHDL) && ir->sc_alpha != 0)
{
init_enerdata(mtop->groups.grps[egcENER].nr,ir->n_flambda,&ed_lam);
for(i=0; i<enerd->n_lambda; i++)
{
lam_i = (i==0 ? lambda : ir->flambda[i-1]);
dvdl_dum = 0;
reset_enerdata(&ir->opts,fr,TRUE,&ed_lam,FALSE);
do_nonbonded(cr,fr,x,f,md,excl,
fr->bBHAM ?
ed_lam.grpp.ener[egBHAMSR] :
ed_lam.grpp.ener[egLJSR],
ed_lam.grpp.ener[egCOULSR],
enerd->grpp.ener[egGB], box_size,nrnb,
lam_i,&dvdl_dum,-1,-1,
GMX_DONB_FOREIGNLAMBDA);
sum_epot(&ir->opts,&ed_lam);
enerd->enerpart_lambda[i] += ed_lam.term[F_EPOT];
}
destroy_enerdata(&ed_lam);
}
where();
/* If we are doing GB, calculate bonded forces and apply corrections
* to the solvation forces */
if (ir->implicit_solvent) {
calc_gb_forces(cr,md,born,top,atype,x,f,fr,idef,
ir->gb_algorithm,ir->sa_algorithm,nrnb,bBornRadii,&pbc,graph,enerd);
}
#ifdef GMX_MPI
if (TAKETIME)
{
t1=MPI_Wtime();
fr->t_fnbf += t1-t0;
}
#endif
if (ir->sc_alpha != 0)
{
enerd->dvdl_nonlin += dvdlambda;
}
else
{
enerd->dvdl_lin += dvdlambda;
}
Vsr = 0;
if (bSepDVDL)
{
for(i=0; i<enerd->grpp.nener; i++)
{
Vsr +=
(fr->bBHAM ?
enerd->grpp.ener[egBHAMSR][i] :
enerd->grpp.ener[egLJSR][i])
+ enerd->grpp.ener[egCOULSR][i] + enerd->grpp.ener[egGB][i];
}
}
PRINT_SEPDVDL("VdW and Coulomb SR particle-p.",Vsr,dvdlambda);
debug_gmx();
GMX_MPE_LOG(ev_do_fnbf_finish);
if (debug)
{
pr_rvecs(debug,0,"fshift after SR",fr->fshift,SHIFTS);
}
/* Shift the coordinates. Must be done before bonded forces and PPPM,
* but is also necessary for SHAKE and update, therefore it can NOT
* go when no bonded forces have to be evaluated.
*/
/* Here sometimes we would not need to shift with NBFonly,
* but we do so anyhow for consistency of the returned coordinates.
*/
if (graph)
{
shift_self(graph,box,x);
if (TRICLINIC(box))
{
inc_nrnb(nrnb,eNR_SHIFTX,2*graph->nnodes);
}
else
{
inc_nrnb(nrnb,eNR_SHIFTX,graph->nnodes);
}
}
/* Check whether we need to do bondeds or correct for exclusions */
if (fr->bMolPBC &&
((flags & GMX_FORCE_BONDED)
|| EEL_RF(fr->eeltype) || EEL_FULL(fr->eeltype)))
{
/* Since all atoms are in the rectangular or triclinic unit-cell,
* only single box vector shifts (2 in x) are required.
*/
set_pbc_dd(&pbc,fr->ePBC,cr->dd,TRUE,box);
}
debug_gmx();
if (flags & GMX_FORCE_BONDED)
{
GMX_MPE_LOG(ev_calc_bonds_start);
calc_bonds(fplog,cr->ms,
idef,x,hist,f,fr,&pbc,graph,enerd,nrnb,lambda,md,fcd,
DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL, atype, born,
fr->bSepDVDL && do_per_step(step,ir->nstlog),step);
/* Check if we have to determine energy differences
* at foreign lambda's.
*/
if (ir->n_flambda > 0 && (flags & GMX_FORCE_DHDL) &&
idef->ilsort != ilsortNO_FE)
{
if (idef->ilsort != ilsortFE_SORTED)
{
gmx_incons("The bonded interactions are not sorted for free energy");
}
init_enerdata(mtop->groups.grps[egcENER].nr,ir->n_flambda,&ed_lam);
for(i=0; i<enerd->n_lambda; i++)
{
lam_i = (i==0 ? lambda : ir->flambda[i-1]);
dvdl_dum = 0;
reset_enerdata(&ir->opts,fr,TRUE,&ed_lam,FALSE);
calc_bonds_lambda(fplog,
idef,x,fr,&pbc,graph,&ed_lam,nrnb,lam_i,md,
fcd,
DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL);
sum_epot(&ir->opts,&ed_lam);
enerd->enerpart_lambda[i] += ed_lam.term[F_EPOT];
}
destroy_enerdata(&ed_lam);
}
debug_gmx();
GMX_MPE_LOG(ev_calc_bonds_finish);
}
where();
*cycles_pme = 0;
if (EEL_FULL(fr->eeltype))
{
bSB = (ir->nwall == 2);
if (bSB)
{
copy_mat(box,boxs);
svmul(ir->wall_ewald_zfac,boxs[ZZ],boxs[ZZ]);
box_size[ZZ] *= ir->wall_ewald_zfac;
}
clear_mat(fr->vir_el_recip);
if (fr->bEwald)
{
if (fr->n_tpi == 0)
{
dvdlambda = 0;
Vcorr = ewald_LRcorrection(fplog,md->start,md->start+md->homenr,
cr,fr,
md->chargeA,
md->nChargePerturbed ? md->chargeB : NULL,
excl,x,bSB ? boxs : box,mu_tot,
ir->ewald_geometry,
ir->epsilon_surface,
lambda,&dvdlambda,&vdip,&vcharge);
PRINT_SEPDVDL("Ewald excl./charge/dip. corr.",Vcorr,dvdlambda);
enerd->dvdl_lin += dvdlambda;
}
else
{
if (ir->ewald_geometry != eewg3D || ir->epsilon_surface != 0)
{
gmx_fatal(FARGS,"TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
}
/* The TPI molecule does not have exclusions with the rest
* of the system and no intra-molecular PME grid contributions
* will be calculated in gmx_pme_calc_energy.
*/
Vcorr = 0;
}
}
else
{
Vcorr = shift_LRcorrection(fplog,md->start,md->homenr,cr,fr,
md->chargeA,excl,x,TRUE,box,
fr->vir_el_recip);
}
dvdlambda = 0;
status = 0;
switch (fr->eeltype)
{
case eelPPPM:
status = gmx_pppm_do(fplog,fr->pmedata,FALSE,x,fr->f_novirsum,
md->chargeA,
box_size,fr->phi,cr,md->start,md->homenr,
nrnb,ir->pme_order,&Vlr);
break;
case eelPME:
case eelPMESWITCH:
case eelPMEUSER:
case eelPMEUSERSWITCH:
if (cr->duty & DUTY_PME)
{
if (fr->n_tpi == 0 || (flags & GMX_FORCE_STATECHANGED))
{
pme_flags = GMX_PME_SPREAD_Q | GMX_PME_SOLVE;
if (flags & GMX_FORCE_FORCES)
{
pme_flags |= GMX_PME_CALC_F;
}
if (flags & GMX_FORCE_VIRIAL)
{
pme_flags |= GMX_PME_CALC_ENER_VIR;
}
if (fr->n_tpi > 0)
{
/* We don't calculate f, but we do want the potential */
pme_flags |= GMX_PME_CALC_POT;
}
wallcycle_start(wcycle,ewcPMEMESH);
status = gmx_pme_do(fr->pmedata,
md->start,md->homenr - fr->n_tpi,
x,fr->f_novirsum,
md->chargeA,md->chargeB,
bSB ? boxs : box,cr,
DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
nrnb,wcycle,
fr->vir_el_recip,fr->ewaldcoeff,
&Vlr,lambda,&dvdlambda,
pme_flags);
*cycles_pme = wallcycle_stop(wcycle,ewcPMEMESH);
/* We should try to do as little computation after
* this as possible, because parallel PME synchronizes
* the nodes, so we want all load imbalance of the rest
* of the force calculation to be before the PME call.
* DD load balancing is done on the whole time of
* the force call (without PME).
*/
}
if (fr->n_tpi > 0)
{
/* Determine the PME grid energy of the test molecule
* with the PME grid potential of the other charges.
*/
gmx_pme_calc_energy(fr->pmedata,fr->n_tpi,
x + md->homenr - fr->n_tpi,
md->chargeA + md->homenr - fr->n_tpi,
&Vlr);
}
PRINT_SEPDVDL("PME mesh",Vlr,dvdlambda);
}
else
{
/* Energies and virial are obtained later from the PME nodes */
/* but values have to be zeroed out here */
Vlr=0.0;
}
break;
case eelEWALD:
Vlr = do_ewald(fplog,FALSE,ir,x,fr->f_novirsum,
md->chargeA,md->chargeB,
box_size,cr,md->homenr,
fr->vir_el_recip,fr->ewaldcoeff,
lambda,&dvdlambda,fr->ewald_table);
PRINT_SEPDVDL("Ewald long-range",Vlr,dvdlambda);
break;
default:
Vlr = 0;
gmx_fatal(FARGS,"No such electrostatics method implemented %s",
eel_names[fr->eeltype]);
}
if (status != 0)
{
gmx_fatal(FARGS,"Error %d in long range electrostatics routine %s",
status,EELTYPE(fr->eeltype));
}
enerd->dvdl_lin += dvdlambda;
enerd->term[F_COUL_RECIP] = Vlr + Vcorr;
if (debug)
{
fprintf(debug,"Vlr = %g, Vcorr = %g, Vlr_corr = %g\n",
Vlr,Vcorr,enerd->term[F_COUL_RECIP]);
pr_rvecs(debug,0,"vir_el_recip after corr",fr->vir_el_recip,DIM);
pr_rvecs(debug,0,"fshift after LR Corrections",fr->fshift,SHIFTS);
}
}
else
{
if (EEL_RF(fr->eeltype))
{
dvdlambda = 0;
if (fr->eeltype != eelRF_NEC)
{
enerd->term[F_RF_EXCL] =
RF_excl_correction(fplog,fr,graph,md,excl,x,f,
fr->fshift,&pbc,lambda,&dvdlambda);
}
enerd->dvdl_lin += dvdlambda;
PRINT_SEPDVDL("RF exclusion correction",
enerd->term[F_RF_EXCL],dvdlambda);
}
}
where();
debug_gmx();
if (debug)
{
print_nrnb(debug,nrnb);
}
debug_gmx();
#ifdef GMX_MPI
if (TAKETIME)
{
t2=MPI_Wtime();
MPI_Barrier(cr->mpi_comm_mygroup);
t3=MPI_Wtime();
fr->t_wait += t3-t2;
if (fr->timesteps == 11)
{
fprintf(stderr,"* PP load balancing info: node %d, step %s, rel wait time=%3.0f%% , load string value: %7.2f\n",
cr->nodeid, gmx_step_str(fr->timesteps,buf),
100*fr->t_wait/(fr->t_wait+fr->t_fnbf),
(fr->t_fnbf+fr->t_wait)/fr->t_fnbf);
}
fr->timesteps++;
}
#endif
if (debug)
{
pr_rvecs(debug,0,"fshift after bondeds",fr->fshift,SHIFTS);
}
GMX_MPE_LOG(ev_force_finish);
}
/* TODO Specialize this routine into init-time and loop-time versions?
e.g. bReadEkin is only true when restoring from checkpoint */
void compute_globals(FILE *fplog, gmx_global_stat *gstat, t_commrec *cr, t_inputrec *ir,
t_forcerec *fr, gmx_ekindata_t *ekind,
t_state *state, t_mdatoms *mdatoms,
t_nrnb *nrnb, t_vcm *vcm, gmx_wallcycle_t wcycle,
gmx_enerdata_t *enerd, tensor force_vir, tensor shake_vir, tensor total_vir,
tensor pres, rvec mu_tot, gmx_constr_t constr,
gmx::SimulationSignaller *signalCoordinator,
matrix box, int *totalNumberOfBondedInteractions,
gmx_bool *bSumEkinhOld, int flags)
{
tensor corr_vir, corr_pres;
gmx_bool bEner, bPres, bTemp;
gmx_bool bStopCM, bGStat,
bReadEkin, bEkinAveVel, bScaleEkin, bConstrain;
real prescorr, enercorr, dvdlcorr, dvdl_ekin;
/* translate CGLO flags to gmx_booleans */
bStopCM = flags & CGLO_STOPCM;
bGStat = flags & CGLO_GSTAT;
bReadEkin = (flags & CGLO_READEKIN);
bScaleEkin = (flags & CGLO_SCALEEKIN);
bEner = flags & CGLO_ENERGY;
bTemp = flags & CGLO_TEMPERATURE;
bPres = (flags & CGLO_PRESSURE);
bConstrain = (flags & CGLO_CONSTRAINT);
/* we calculate a full state kinetic energy either with full-step velocity verlet
or half step where we need the pressure */
bEkinAveVel = (ir->eI == eiVV || (ir->eI == eiVVAK && bPres) || bReadEkin);
/* in initalization, it sums the shake virial in vv, and to
sums ekinh_old in leapfrog (or if we are calculating ekinh_old) for other reasons */
/* ########## Kinetic energy ############## */
if (bTemp)
{
/* Non-equilibrium MD: this is parallellized, but only does communication
* when there really is NEMD.
*/
if (PAR(cr) && (ekind->bNEMD))
{
accumulate_u(cr, &(ir->opts), ekind);
}
if (!bReadEkin)
{
calc_ke_part(state, &(ir->opts), mdatoms, ekind, nrnb, bEkinAveVel);
}
}
/* Calculate center of mass velocity if necessary, also parallellized */
if (bStopCM)
{
calc_vcm_grp(0, mdatoms->homenr, mdatoms,
state->x, state->v, vcm);
}
if (bTemp || bStopCM || bPres || bEner || bConstrain)
{
if (!bGStat)
{
/* We will not sum ekinh_old,
* so signal that we still have to do it.
*/
*bSumEkinhOld = TRUE;
}
else
{
gmx::ArrayRef<real> signalBuffer = signalCoordinator->getCommunicationBuffer();
if (PAR(cr))
{
wallcycle_start(wcycle, ewcMoveE);
global_stat(gstat, cr, enerd, force_vir, shake_vir, mu_tot,
ir, ekind, constr, bStopCM ? vcm : NULL,
signalBuffer.size(), signalBuffer.data(),
totalNumberOfBondedInteractions,
*bSumEkinhOld, flags);
wallcycle_stop(wcycle, ewcMoveE);
}
signalCoordinator->finalizeSignals();
*bSumEkinhOld = FALSE;
}
}
if (!ekind->bNEMD && debug && bTemp && (vcm->nr > 0))
{
correct_ekin(debug,
0, mdatoms->homenr,
state->v, vcm->group_p[0],
mdatoms->massT, mdatoms->tmass, ekind->ekin);
}
/* Do center of mass motion removal */
if (bStopCM)
{
check_cm_grp(fplog, vcm, ir, 1);
do_stopcm_grp(0, mdatoms->homenr, mdatoms->cVCM,
state->x, state->v, vcm);
inc_nrnb(nrnb, eNR_STOPCM, mdatoms->homenr);
}
if (bEner)
{
/* Calculate the amplitude of the cosine velocity profile */
ekind->cosacc.vcos = ekind->cosacc.mvcos/mdatoms->tmass;
}
if (bTemp)
{
/* Sum the kinetic energies of the groups & calc temp */
/* compute full step kinetic energies if vv, or if vv-avek and we are computing the pressure with inputrecNptTrotter */
/* three maincase: VV with AveVel (md-vv), vv with AveEkin (md-vv-avek), leap with AveEkin (md).
Leap with AveVel is not supported; it's not clear that it will actually work.
bEkinAveVel: If TRUE, we simply multiply ekin by ekinscale to get a full step kinetic energy.
If FALSE, we average ekinh_old and ekinh*ekinscale_nhc to get an averaged half step kinetic energy.
*/
enerd->term[F_TEMP] = sum_ekin(&(ir->opts), ekind, &dvdl_ekin,
bEkinAveVel, bScaleEkin);
enerd->dvdl_lin[efptMASS] = (double) dvdl_ekin;
enerd->term[F_EKIN] = trace(ekind->ekin);
}
/* ########## Long range energy information ###### */
if (bEner || bPres || bConstrain)
{
calc_dispcorr(ir, fr, box, state->lambda[efptVDW],
corr_pres, corr_vir, &prescorr, &enercorr, &dvdlcorr);
}
if (bEner)
{
enerd->term[F_DISPCORR] = enercorr;
enerd->term[F_EPOT] += enercorr;
enerd->term[F_DVDL_VDW] += dvdlcorr;
}
/* ########## Now pressure ############## */
if (bPres || bConstrain)
{
m_add(force_vir, shake_vir, total_vir);
/* Calculate pressure and apply LR correction if PPPM is used.
* Use the box from last timestep since we already called update().
*/
enerd->term[F_PRES] = calc_pres(fr->ePBC, ir->nwall, box, ekind->ekin, total_vir, pres);
/* Calculate long range corrections to pressure and energy */
/* this adds to enerd->term[F_PRES] and enerd->term[F_ETOT],
and computes enerd->term[F_DISPCORR]. Also modifies the
total_vir and pres tesors */
m_add(total_vir, corr_vir, total_vir);
m_add(pres, corr_pres, pres);
enerd->term[F_PDISPCORR] = prescorr;
enerd->term[F_PRES] += prescorr;
}
}
void do_force_lowlevel(t_forcerec *fr, t_inputrec *ir,
t_idef *idef, t_commrec *cr,
t_nrnb *nrnb, gmx_wallcycle_t wcycle,
t_mdatoms *md,
rvec x[], history_t *hist,
rvec f[],
rvec f_longrange[],
gmx_enerdata_t *enerd,
t_fcdata *fcd,
gmx_localtop_t *top,
gmx_genborn_t *born,
gmx_bool bBornRadii,
matrix box,
t_lambda *fepvals,
real *lambda,
t_graph *graph,
t_blocka *excl,
rvec mu_tot[],
int flags,
float *cycles_pme)
{
int i, j;
int donb_flags;
gmx_bool bSB;
int pme_flags;
matrix boxs;
rvec box_size;
t_pbc pbc;
real dvdl_dum[efptNR], dvdl_nb[efptNR];
#ifdef GMX_MPI
double t0 = 0.0, t1, t2, t3; /* time measurement for coarse load balancing */
#endif
set_pbc(&pbc, fr->ePBC, box);
/* reset free energy components */
for (i = 0; i < efptNR; i++)
{
dvdl_nb[i] = 0;
dvdl_dum[i] = 0;
}
/* Reset box */
for (i = 0; (i < DIM); i++)
{
box_size[i] = box[i][i];
}
debug_gmx();
/* do QMMM first if requested */
if (fr->bQMMM)
{
enerd->term[F_EQM] = calculate_QMMM(cr, x, f, fr);
}
/* Call the short range functions all in one go. */
#ifdef GMX_MPI
/*#define TAKETIME ((cr->npmenodes) && (fr->timesteps < 12))*/
#define TAKETIME FALSE
if (TAKETIME)
{
MPI_Barrier(cr->mpi_comm_mygroup);
t0 = MPI_Wtime();
}
#endif
if (ir->nwall)
{
/* foreign lambda component for walls */
real dvdl_walls = do_walls(ir, fr, box, md, x, f, lambda[efptVDW],
enerd->grpp.ener[egLJSR], nrnb);
enerd->dvdl_lin[efptVDW] += dvdl_walls;
}
/* If doing GB, reset dvda and calculate the Born radii */
if (ir->implicit_solvent)
{
wallcycle_sub_start(wcycle, ewcsNONBONDED);
for (i = 0; i < born->nr; i++)
{
fr->dvda[i] = 0;
}
if (bBornRadii)
{
calc_gb_rad(cr, fr, ir, top, x, &(fr->gblist), born, md, nrnb);
}
wallcycle_sub_stop(wcycle, ewcsNONBONDED);
}
where();
/* We only do non-bonded calculation with group scheme here, the verlet
* calls are done from do_force_cutsVERLET(). */
if (fr->cutoff_scheme == ecutsGROUP && (flags & GMX_FORCE_NONBONDED))
{
donb_flags = 0;
/* Add short-range interactions */
donb_flags |= GMX_NONBONDED_DO_SR;
/* Currently all group scheme kernels always calculate (shift-)forces */
if (flags & GMX_FORCE_FORCES)
{
donb_flags |= GMX_NONBONDED_DO_FORCE;
}
if (flags & GMX_FORCE_VIRIAL)
{
donb_flags |= GMX_NONBONDED_DO_SHIFTFORCE;
}
if (flags & GMX_FORCE_ENERGY)
{
donb_flags |= GMX_NONBONDED_DO_POTENTIAL;
}
if (flags & GMX_FORCE_DO_LR)
{
donb_flags |= GMX_NONBONDED_DO_LR;
}
wallcycle_sub_start(wcycle, ewcsNONBONDED);
do_nonbonded(fr, x, f, f_longrange, md, excl,
&enerd->grpp, nrnb,
lambda, dvdl_nb, -1, -1, donb_flags);
/* If we do foreign lambda and we have soft-core interactions
* we have to recalculate the (non-linear) energies contributions.
*/
if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL) && fepvals->sc_alpha != 0)
{
for (i = 0; i < enerd->n_lambda; i++)
{
real lam_i[efptNR];
for (j = 0; j < efptNR; j++)
{
lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
}
reset_foreign_enerdata(enerd);
do_nonbonded(fr, x, f, f_longrange, md, excl,
&(enerd->foreign_grpp), nrnb,
lam_i, dvdl_dum, -1, -1,
(donb_flags & ~GMX_NONBONDED_DO_FORCE) | GMX_NONBONDED_DO_FOREIGNLAMBDA);
sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
}
}
wallcycle_sub_stop(wcycle, ewcsNONBONDED);
where();
}
/* If we are doing GB, calculate bonded forces and apply corrections
* to the solvation forces */
/* MRS: Eventually, many need to include free energy contribution here! */
if (ir->implicit_solvent)
{
wallcycle_sub_start(wcycle, ewcsLISTED);
calc_gb_forces(cr, md, born, top, x, f, fr, idef,
ir->gb_algorithm, ir->sa_algorithm, nrnb, &pbc, graph, enerd);
wallcycle_sub_stop(wcycle, ewcsLISTED);
}
#ifdef GMX_MPI
if (TAKETIME)
{
t1 = MPI_Wtime();
fr->t_fnbf += t1-t0;
}
#endif
if (fepvals->sc_alpha != 0)
{
enerd->dvdl_nonlin[efptVDW] += dvdl_nb[efptVDW];
}
else
{
enerd->dvdl_lin[efptVDW] += dvdl_nb[efptVDW];
}
if (fepvals->sc_alpha != 0)
/* even though coulomb part is linear, we already added it, beacuse we
need to go through the vdw calculation anyway */
{
enerd->dvdl_nonlin[efptCOUL] += dvdl_nb[efptCOUL];
}
else
{
enerd->dvdl_lin[efptCOUL] += dvdl_nb[efptCOUL];
}
debug_gmx();
if (debug)
{
pr_rvecs(debug, 0, "fshift after SR", fr->fshift, SHIFTS);
}
/* Shift the coordinates. Must be done before listed forces and PPPM,
* but is also necessary for SHAKE and update, therefore it can NOT
* go when no listed forces have to be evaluated.
*
* The shifting and PBC code is deliberately not timed, since with
* the Verlet scheme it only takes non-zero time with triclinic
* boxes, and even then the time is around a factor of 100 less
* than the next smallest counter.
*/
/* Here sometimes we would not need to shift with NBFonly,
* but we do so anyhow for consistency of the returned coordinates.
*/
if (graph)
{
shift_self(graph, box, x);
if (TRICLINIC(box))
{
inc_nrnb(nrnb, eNR_SHIFTX, 2*graph->nnodes);
}
else
{
inc_nrnb(nrnb, eNR_SHIFTX, graph->nnodes);
}
}
/* Check whether we need to do listed interactions or correct for exclusions */
if (fr->bMolPBC &&
((flags & GMX_FORCE_LISTED)
|| EEL_RF(fr->eeltype) || EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype)))
{
/* TODO There are no electrostatics methods that require this
transformation, when using the Verlet scheme, so update the
above conditional. */
/* Since all atoms are in the rectangular or triclinic unit-cell,
* only single box vector shifts (2 in x) are required.
*/
set_pbc_dd(&pbc, fr->ePBC, cr->dd, TRUE, box);
}
debug_gmx();
do_force_listed(wcycle, box, ir->fepvals, cr->ms,
idef, (const rvec *) x, hist, f, fr,
&pbc, graph, enerd, nrnb, lambda, md, fcd,
DOMAINDECOMP(cr) ? cr->dd->gatindex : NULL,
flags);
where();
*cycles_pme = 0;
clear_mat(fr->vir_el_recip);
clear_mat(fr->vir_lj_recip);
/* Do long-range electrostatics and/or LJ-PME, including related short-range
* corrections.
*/
if (EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype))
{
int status = 0;
real Vlr_q = 0, Vlr_lj = 0, Vcorr_q = 0, Vcorr_lj = 0;
real dvdl_long_range_q = 0, dvdl_long_range_lj = 0;
bSB = (ir->nwall == 2);
if (bSB)
{
copy_mat(box, boxs);
svmul(ir->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
box_size[ZZ] *= ir->wall_ewald_zfac;
}
if (EEL_PME_EWALD(fr->eeltype) || EVDW_PME(fr->vdwtype))
{
real dvdl_long_range_correction_q = 0;
real dvdl_long_range_correction_lj = 0;
/* With the Verlet scheme exclusion forces are calculated
* in the non-bonded kernel.
*/
/* The TPI molecule does not have exclusions with the rest
* of the system and no intra-molecular PME grid
* contributions will be calculated in
* gmx_pme_calc_energy.
*/
if ((ir->cutoff_scheme == ecutsGROUP && fr->n_tpi == 0) ||
ir->ewald_geometry != eewg3D ||
ir->epsilon_surface != 0)
{
int nthreads, t;
wallcycle_sub_start(wcycle, ewcsEWALD_CORRECTION);
if (fr->n_tpi > 0)
{
gmx_fatal(FARGS, "TPI with PME currently only works in a 3D geometry with tin-foil boundary conditions");
}
nthreads = gmx_omp_nthreads_get(emntBonded);
#pragma omp parallel for num_threads(nthreads) schedule(static)
for (t = 0; t < nthreads; t++)
{
int i;
rvec *fnv;
tensor *vir_q, *vir_lj;
real *Vcorrt_q, *Vcorrt_lj, *dvdlt_q, *dvdlt_lj;
if (t == 0)
{
fnv = fr->f_novirsum;
vir_q = &fr->vir_el_recip;
vir_lj = &fr->vir_lj_recip;
Vcorrt_q = &Vcorr_q;
Vcorrt_lj = &Vcorr_lj;
dvdlt_q = &dvdl_long_range_correction_q;
dvdlt_lj = &dvdl_long_range_correction_lj;
}
else
{
fnv = fr->f_t[t].f;
vir_q = &fr->f_t[t].vir_q;
vir_lj = &fr->f_t[t].vir_lj;
Vcorrt_q = &fr->f_t[t].Vcorr_q;
Vcorrt_lj = &fr->f_t[t].Vcorr_lj;
dvdlt_q = &fr->f_t[t].dvdl[efptCOUL];
dvdlt_lj = &fr->f_t[t].dvdl[efptVDW];
for (i = 0; i < fr->natoms_force; i++)
{
clear_rvec(fnv[i]);
}
clear_mat(*vir_q);
clear_mat(*vir_lj);
}
*dvdlt_q = 0;
*dvdlt_lj = 0;
ewald_LRcorrection(fr->excl_load[t], fr->excl_load[t+1],
cr, t, fr,
md->chargeA, md->chargeB,
md->sqrt_c6A, md->sqrt_c6B,
md->sigmaA, md->sigmaB,
md->sigma3A, md->sigma3B,
md->nChargePerturbed || md->nTypePerturbed,
ir->cutoff_scheme != ecutsVERLET,
excl, x, bSB ? boxs : box, mu_tot,
ir->ewald_geometry,
ir->epsilon_surface,
fnv, *vir_q, *vir_lj,
Vcorrt_q, Vcorrt_lj,
lambda[efptCOUL], lambda[efptVDW],
dvdlt_q, dvdlt_lj);
}
if (nthreads > 1)
{
reduce_thread_forces(fr->natoms_force, fr->f_novirsum,
fr->vir_el_recip, fr->vir_lj_recip,
&Vcorr_q, &Vcorr_lj,
&dvdl_long_range_correction_q,
&dvdl_long_range_correction_lj,
nthreads, fr->f_t);
}
wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
}
if (EEL_PME_EWALD(fr->eeltype) && fr->n_tpi == 0)
{
/* This is not in a subcounter because it takes a
negligible and constant-sized amount of time */
Vcorr_q += ewald_charge_correction(cr, fr, lambda[efptCOUL], box,
&dvdl_long_range_correction_q,
fr->vir_el_recip);
}
enerd->dvdl_lin[efptCOUL] += dvdl_long_range_correction_q;
enerd->dvdl_lin[efptVDW] += dvdl_long_range_correction_lj;
if ((EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype)) && (cr->duty & DUTY_PME))
{
/* Do reciprocal PME for Coulomb and/or LJ. */
assert(fr->n_tpi >= 0);
if (fr->n_tpi == 0 || (flags & GMX_FORCE_STATECHANGED))
{
pme_flags = GMX_PME_SPREAD | GMX_PME_SOLVE;
if (EEL_PME(fr->eeltype))
{
pme_flags |= GMX_PME_DO_COULOMB;
}
if (EVDW_PME(fr->vdwtype))
{
pme_flags |= GMX_PME_DO_LJ;
}
if (flags & GMX_FORCE_FORCES)
{
pme_flags |= GMX_PME_CALC_F;
}
if (flags & GMX_FORCE_VIRIAL)
{
pme_flags |= GMX_PME_CALC_ENER_VIR;
}
if (fr->n_tpi > 0)
{
/* We don't calculate f, but we do want the potential */
pme_flags |= GMX_PME_CALC_POT;
}
wallcycle_start(wcycle, ewcPMEMESH);
status = gmx_pme_do(fr->pmedata,
0, md->homenr - fr->n_tpi,
x, fr->f_novirsum,
md->chargeA, md->chargeB,
md->sqrt_c6A, md->sqrt_c6B,
md->sigmaA, md->sigmaB,
bSB ? boxs : box, cr,
DOMAINDECOMP(cr) ? dd_pme_maxshift_x(cr->dd) : 0,
DOMAINDECOMP(cr) ? dd_pme_maxshift_y(cr->dd) : 0,
nrnb, wcycle,
fr->vir_el_recip, fr->ewaldcoeff_q,
fr->vir_lj_recip, fr->ewaldcoeff_lj,
&Vlr_q, &Vlr_lj,
lambda[efptCOUL], lambda[efptVDW],
&dvdl_long_range_q, &dvdl_long_range_lj, pme_flags);
*cycles_pme = wallcycle_stop(wcycle, ewcPMEMESH);
if (status != 0)
{
gmx_fatal(FARGS, "Error %d in reciprocal PME routine", status);
}
/* We should try to do as little computation after
* this as possible, because parallel PME synchronizes
* the nodes, so we want all load imbalance of the
* rest of the force calculation to be before the PME
* call. DD load balancing is done on the whole time
* of the force call (without PME).
*/
}
if (fr->n_tpi > 0)
{
if (EVDW_PME(ir->vdwtype))
{
gmx_fatal(FARGS, "Test particle insertion not implemented with LJ-PME");
}
/* Determine the PME grid energy of the test molecule
* with the PME grid potential of the other charges.
*/
gmx_pme_calc_energy(fr->pmedata, fr->n_tpi,
x + md->homenr - fr->n_tpi,
md->chargeA + md->homenr - fr->n_tpi,
&Vlr_q);
}
}
}
if (!EEL_PME(fr->eeltype) && EEL_PME_EWALD(fr->eeltype))
{
Vlr_q = do_ewald(ir, x, fr->f_novirsum,
md->chargeA, md->chargeB,
box_size, cr, md->homenr,
fr->vir_el_recip, fr->ewaldcoeff_q,
lambda[efptCOUL], &dvdl_long_range_q, fr->ewald_table);
}
/* Note that with separate PME nodes we get the real energies later */
enerd->dvdl_lin[efptCOUL] += dvdl_long_range_q;
enerd->dvdl_lin[efptVDW] += dvdl_long_range_lj;
enerd->term[F_COUL_RECIP] = Vlr_q + Vcorr_q;
enerd->term[F_LJ_RECIP] = Vlr_lj + Vcorr_lj;
if (debug)
{
fprintf(debug, "Vlr_q = %g, Vcorr_q = %g, Vlr_corr_q = %g\n",
Vlr_q, Vcorr_q, enerd->term[F_COUL_RECIP]);
pr_rvecs(debug, 0, "vir_el_recip after corr", fr->vir_el_recip, DIM);
pr_rvecs(debug, 0, "fshift after LR Corrections", fr->fshift, SHIFTS);
fprintf(debug, "Vlr_lj: %g, Vcorr_lj = %g, Vlr_corr_lj = %g\n",
Vlr_lj, Vcorr_lj, enerd->term[F_LJ_RECIP]);
pr_rvecs(debug, 0, "vir_lj_recip after corr", fr->vir_lj_recip, DIM);
}
}
else
{
/* Is there a reaction-field exclusion correction needed? */
if (EEL_RF(fr->eeltype) && eelRF_NEC != fr->eeltype)
{
/* With the Verlet scheme, exclusion forces are calculated
* in the non-bonded kernel.
*/
if (ir->cutoff_scheme != ecutsVERLET)
{
real dvdl_rf_excl = 0;
enerd->term[F_RF_EXCL] =
RF_excl_correction(fr, graph, md, excl, x, f,
fr->fshift, &pbc, lambda[efptCOUL], &dvdl_rf_excl);
enerd->dvdl_lin[efptCOUL] += dvdl_rf_excl;
}
}
}
where();
debug_gmx();
if (debug)
{
print_nrnb(debug, nrnb);
}
debug_gmx();
#ifdef GMX_MPI
if (TAKETIME)
{
t2 = MPI_Wtime();
MPI_Barrier(cr->mpi_comm_mygroup);
t3 = MPI_Wtime();
fr->t_wait += t3-t2;
if (fr->timesteps == 11)
{
char buf[22];
fprintf(stderr, "* PP load balancing info: rank %d, step %s, rel wait time=%3.0f%% , load string value: %7.2f\n",
cr->nodeid, gmx_step_str(fr->timesteps, buf),
100*fr->t_wait/(fr->t_wait+fr->t_fnbf),
(fr->t_fnbf+fr->t_wait)/fr->t_fnbf);
}
fr->timesteps++;
}
#endif
if (debug)
{
pr_rvecs(debug, 0, "fshift after bondeds", fr->fshift, SHIFTS);
}
}
