/*! \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 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;
}
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 pme_gpu_reinit_computation(const gmx_pme_t *pme,
gmx_wallcycle *wcycle)
{
GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
pme_gpu_clear_grids(pme->gpu);
pme_gpu_clear_energy_virial(pme->gpu);
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
}
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;
}
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);
}
void pme_gpu_launch_gather(const gmx_pme_t *pme,
gmx_wallcycle gmx_unused *wcycle,
PmeForceOutputHandling forceTreatment)
{
GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
if (!pme_gpu_performs_gather(pme->gpu))
{
return;
}
wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
const unsigned int gridIndex = 0;
real *fftgrid = pme->fftgrid[gridIndex];
pme_gpu_gather(pme->gpu, forceTreatment, reinterpret_cast<float *>(fftgrid));
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
}
void pme_gpu_prepare_computation(gmx_pme_t *pme,
bool needToUpdateBox,
const matrix box,
gmx_wallcycle *wcycle,
int flags)
{
GMX_ASSERT(pme_gpu_active(pme), "This should be a GPU run of PME but it is not enabled.");
GMX_ASSERT(pme->nnodes > 0, "");
GMX_ASSERT(pme->nnodes == 1 || pme->ndecompdim > 0, "");
PmeGpu *pmeGpu = pme->gpu;
pmeGpu->settings.currentFlags = flags;
// TODO these flags are only here to honor the CPU PME code, and probably should be removed
bool shouldUpdateBox = false;
for (int i = 0; i < DIM; ++i)
{
for (int j = 0; j <= i; ++j)
{
shouldUpdateBox |= (pmeGpu->common->previousBox[i][j] != box[i][j]);
pmeGpu->common->previousBox[i][j] = box[i][j];
}
}
if (needToUpdateBox || shouldUpdateBox) // || is to make the first computation always update
{
wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_PME);
pme_gpu_update_input_box(pmeGpu, box);
wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_PME);
wallcycle_stop(wcycle, ewcLAUNCH_GPU);
if (!pme_gpu_performs_solve(pmeGpu))
{
// TODO remove code duplication and add test coverage
matrix scaledBox;
pmeGpu->common->boxScaler->scaleBox(box, scaledBox);
gmx::invertBoxMatrix(scaledBox, pme->recipbox);
pme->boxVolume = scaledBox[XX][XX] * scaledBox[YY][YY] * scaledBox[ZZ][ZZ];
}
}
}
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;
}
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);
}
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_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);
}
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(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);
}
}
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);
}
}
/* 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_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);
}
}
int mdrunner(gmx_hw_opt_t *hw_opt,
FILE *fplog, t_commrec *cr, int nfile,
const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
gmx_bool bCompact, int nstglobalcomm,
ivec ddxyz, int dd_node_order, real rdd, real rconstr,
const char *dddlb_opt, real dlb_scale,
const char *ddcsx, const char *ddcsy, const char *ddcsz,
const char *nbpu_opt, int nstlist_cmdline,
gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
int repl_ex_seed, real pforce, real cpt_period, real max_hours,
int imdport, unsigned long Flags)
{
gmx_bool bForceUseGPU, bTryUseGPU, bRerunMD;
t_inputrec *inputrec;
t_state *state = NULL;
matrix box;
gmx_ddbox_t ddbox = {0};
int npme_major, npme_minor;
t_nrnb *nrnb;
gmx_mtop_t *mtop = NULL;
t_mdatoms *mdatoms = NULL;
t_forcerec *fr = NULL;
t_fcdata *fcd = NULL;
real ewaldcoeff_q = 0;
real ewaldcoeff_lj = 0;
struct gmx_pme_t **pmedata = NULL;
gmx_vsite_t *vsite = NULL;
gmx_constr_t constr;
int nChargePerturbed = -1, nTypePerturbed = 0, status;
gmx_wallcycle_t wcycle;
gmx_bool bReadEkin;
gmx_walltime_accounting_t walltime_accounting = NULL;
int rc;
gmx_int64_t reset_counters;
gmx_edsam_t ed = NULL;
int nthreads_pme = 1;
int nthreads_pp = 1;
gmx_membed_t membed = NULL;
gmx_hw_info_t *hwinfo = NULL;
/* The master rank decides early on bUseGPU and broadcasts this later */
gmx_bool bUseGPU = FALSE;
/* CAUTION: threads may be started later on in this function, so
cr doesn't reflect the final parallel state right now */
snew(inputrec, 1);
snew(mtop, 1);
if (Flags & MD_APPENDFILES)
{
fplog = NULL;
}
bRerunMD = (Flags & MD_RERUN);
bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
/* Detect hardware, gather information. This is an operation that is
* global for this process (MPI rank). */
hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
gmx_print_detected_hardware(fplog, cr, hwinfo);
if (fplog != NULL)
{
/* Print references after all software/hardware printing */
please_cite(fplog, "Abraham2015");
please_cite(fplog, "Pall2015");
please_cite(fplog, "Pronk2013");
please_cite(fplog, "Hess2008b");
please_cite(fplog, "Spoel2005a");
please_cite(fplog, "Lindahl2001a");
please_cite(fplog, "Berendsen95a");
}
snew(state, 1);
if (SIMMASTER(cr))
{
/* Read (nearly) all data required for the simulation */
read_tpx_state(ftp2fn(efTPR, nfile, fnm), inputrec, state, NULL, mtop);
if (inputrec->cutoff_scheme == ecutsVERLET)
{
/* Here the master rank decides if all ranks will use GPUs */
bUseGPU = (hwinfo->gpu_info.n_dev_compatible > 0 ||
getenv("GMX_EMULATE_GPU") != NULL);
/* TODO add GPU kernels for this and replace this check by:
* (bUseGPU && (ir->vdwtype == evdwPME &&
* ir->ljpme_combination_rule == eljpmeLB))
* update the message text and the content of nbnxn_acceleration_supported.
*/
if (bUseGPU &&
!nbnxn_gpu_acceleration_supported(fplog, cr, inputrec, bRerunMD))
{
/* Fallback message printed by nbnxn_acceleration_supported */
if (bForceUseGPU)
{
gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
}
bUseGPU = FALSE;
}
prepare_verlet_scheme(fplog, cr,
inputrec, nstlist_cmdline, mtop, state->box,
bUseGPU);
}
else
{
if (nstlist_cmdline > 0)
{
gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
}
if (hwinfo->gpu_info.n_dev_compatible > 0)
{
md_print_warn(cr, fplog,
"NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
" To use a GPU, set the mdp option: cutoff-scheme = Verlet\n");
}
if (bForceUseGPU)
{
gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
}
#ifdef GMX_TARGET_BGQ
md_print_warn(cr, fplog,
"NOTE: There is no SIMD implementation of the group scheme kernels on\n"
" BlueGene/Q. You will observe better performance from using the\n"
" Verlet cut-off scheme.\n");
#endif
}
if (inputrec->eI == eiSD2)
{
md_print_warn(cr, fplog, "The stochastic dynamics integrator %s is deprecated, since\n"
"it is slower than integrator %s and is slightly less accurate\n"
"with constraints. Use the %s integrator.",
ei_names[inputrec->eI], ei_names[eiSD1], ei_names[eiSD1]);
}
}
/* Check and update the hardware options for internal consistency */
check_and_update_hw_opt_1(hw_opt, cr);
/* Early check for externally set process affinity. */
gmx_check_thread_affinity_set(fplog, cr,
hw_opt, hwinfo->nthreads_hw_avail, FALSE);
#ifdef GMX_THREAD_MPI
if (SIMMASTER(cr))
{
if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
{
gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
}
/* Since the master knows the cut-off scheme, update hw_opt for this.
* This is done later for normal MPI and also once more with tMPI
* for all tMPI ranks.
*/
check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
/* NOW the threads will be started: */
hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
hw_opt,
inputrec, mtop,
cr, fplog, bUseGPU);
if (hw_opt->nthreads_tmpi > 1)
{
t_commrec *cr_old = cr;
/* now start the threads. */
cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
oenv, bVerbose, bCompact, nstglobalcomm,
ddxyz, dd_node_order, rdd, rconstr,
dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
nbpu_opt, nstlist_cmdline,
nsteps_cmdline, nstepout, resetstep, nmultisim,
repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
cpt_period, max_hours,
Flags);
/* the main thread continues here with a new cr. We don't deallocate
the old cr because other threads may still be reading it. */
if (cr == NULL)
{
gmx_comm("Failed to spawn threads");
}
}
}
#endif
/* END OF CAUTION: cr is now reliable */
/* g_membed initialisation *
* Because we change the mtop, init_membed is called before the init_parallel *
* (in case we ever want to make it run in parallel) */
if (opt2bSet("-membed", nfile, fnm))
{
if (MASTER(cr))
{
fprintf(stderr, "Initializing membed");
}
membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
}
if (PAR(cr))
{
/* now broadcast everything to the non-master nodes/threads: */
init_parallel(cr, inputrec, mtop);
/* The master rank decided on the use of GPUs,
* broadcast this information to all ranks.
*/
gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
}
if (fplog != NULL)
{
pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
fprintf(fplog, "\n");
}
/* now make sure the state is initialized and propagated */
set_state_entries(state, inputrec);
/* A parallel command line option consistency check that we can
only do after any threads have started. */
if (!PAR(cr) &&
(ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
{
gmx_fatal(FARGS,
"The -dd or -npme option request a parallel simulation, "
#ifndef GMX_MPI
"but %s was compiled without threads or MPI enabled"
#else
#ifdef GMX_THREAD_MPI
"but the number of threads (option -nt) is 1"
#else
"but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
#endif
#endif
, output_env_get_program_display_name(oenv)
);
}
if (bRerunMD &&
(EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
{
gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
}
if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
{
gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
}
if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
{
if (cr->npmenodes > 0)
{
gmx_fatal_collective(FARGS, cr, NULL,
"PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
}
cr->npmenodes = 0;
}
if (bUseGPU && cr->npmenodes < 0)
{
/* With GPUs we don't automatically use PME-only ranks. PME ranks can
* improve performance with many threads per GPU, since our OpenMP
* scaling is bad, but it's difficult to automate the setup.
*/
cr->npmenodes = 0;
}
#ifdef GMX_FAHCORE
if (MASTER(cr))
{
fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
}
#endif
/* NMR restraints must be initialized before load_checkpoint,
* since with time averaging the history is added to t_state.
* For proper consistency check we therefore need to extend
* t_state here.
* So the PME-only nodes (if present) will also initialize
* the distance restraints.
*/
snew(fcd, 1);
/* This needs to be called before read_checkpoint to extend the state */
init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
state);
if (DEFORM(*inputrec))
{
/* Store the deform reference box before reading the checkpoint */
if (SIMMASTER(cr))
{
copy_mat(state->box, box);
}
if (PAR(cr))
{
gmx_bcast(sizeof(box), box, cr);
}
/* Because we do not have the update struct available yet
* in which the reference values should be stored,
* we store them temporarily in static variables.
* This should be thread safe, since they are only written once
* and with identical values.
*/
tMPI_Thread_mutex_lock(&deform_init_box_mutex);
deform_init_init_step_tpx = inputrec->init_step;
copy_mat(box, deform_init_box_tpx);
tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
}
if (opt2bSet("-cpi", nfile, fnm))
{
/* Check if checkpoint file exists before doing continuation.
* This way we can use identical input options for the first and subsequent runs...
*/
if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
{
load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
cr, ddxyz,
inputrec, state, &bReadEkin,
(Flags & MD_APPENDFILES),
(Flags & MD_APPENDFILESSET));
if (bReadEkin)
{
Flags |= MD_READ_EKIN;
}
}
}
if (MASTER(cr) && (Flags & MD_APPENDFILES))
{
gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr,
Flags, &fplog);
}
/* override nsteps with value from cmdline */
override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
if (SIMMASTER(cr))
{
copy_mat(state->box, box);
}
if (PAR(cr))
{
gmx_bcast(sizeof(box), box, cr);
}
/* Essential dynamics */
if (opt2bSet("-ei", nfile, fnm))
{
/* Open input and output files, allocate space for ED data structure */
ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
}
if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
inputrec->eI == eiNM))
{
cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
dddlb_opt, dlb_scale,
ddcsx, ddcsy, ddcsz,
mtop, inputrec,
box, state->x,
&ddbox, &npme_major, &npme_minor);
make_dd_communicators(fplog, cr, dd_node_order);
/* Set overallocation to avoid frequent reallocation of arrays */
set_over_alloc_dd(TRUE);
}
else
{
/* PME, if used, is done on all nodes with 1D decomposition */
cr->npmenodes = 0;
cr->duty = (DUTY_PP | DUTY_PME);
npme_major = 1;
npme_minor = 1;
if (inputrec->ePBC == epbcSCREW)
{
gmx_fatal(FARGS,
"pbc=%s is only implemented with domain decomposition",
epbc_names[inputrec->ePBC]);
}
}
if (PAR(cr))
{
/* After possible communicator splitting in make_dd_communicators.
* we can set up the intra/inter node communication.
*/
gmx_setup_nodecomm(fplog, cr);
}
/* Initialize per-physical-node MPI process/thread ID and counters. */
gmx_init_intranode_counters(cr);
#ifdef GMX_MPI
if (MULTISIM(cr))
{
md_print_info(cr, fplog,
"This is simulation %d out of %d running as a composite GROMACS\n"
"multi-simulation job. Setup for this simulation:\n\n",
cr->ms->sim, cr->ms->nsim);
}
md_print_info(cr, fplog, "Using %d MPI %s\n",
cr->nnodes,
#ifdef GMX_THREAD_MPI
cr->nnodes == 1 ? "thread" : "threads"
#else
cr->nnodes == 1 ? "process" : "processes"
#endif
);
fflush(stderr);
#endif
/* Check and update hw_opt for the cut-off scheme */
check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
/* Check and update hw_opt for the number of MPI ranks */
check_and_update_hw_opt_3(hw_opt);
gmx_omp_nthreads_init(fplog, cr,
hwinfo->nthreads_hw_avail,
hw_opt->nthreads_omp,
hw_opt->nthreads_omp_pme,
(cr->duty & DUTY_PP) == 0,
inputrec->cutoff_scheme == ecutsVERLET);
#ifndef NDEBUG
if (integrator[inputrec->eI].func != do_tpi &&
inputrec->cutoff_scheme == ecutsVERLET)
{
gmx_feenableexcept();
}
#endif
if (bUseGPU)
{
/* Select GPU id's to use */
gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
&hw_opt->gpu_opt);
}
else
{
/* Ignore (potentially) manually selected GPUs */
hw_opt->gpu_opt.n_dev_use = 0;
}
/* check consistency across ranks of things like SIMD
* support and number of GPUs selected */
gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
/* Now that we know the setup is consistent, check for efficiency */
check_resource_division_efficiency(hwinfo, hw_opt, Flags & MD_NTOMPSET,
cr, fplog);
if (DOMAINDECOMP(cr))
{
/* When we share GPUs over ranks, we need to know this for the DLB */
dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
}
/* getting number of PP/PME threads
PME: env variable should be read only on one node to make sure it is
identical everywhere;
*/
/* TODO nthreads_pp is only used for pinning threads.
* This is a temporary solution until we have a hw topology library.
*/
nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
nthreads_pme = gmx_omp_nthreads_get(emntPME);
wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
if (PAR(cr))
{
/* Master synchronizes its value of reset_counters with all nodes
* including PME only nodes */
reset_counters = wcycle_get_reset_counters(wcycle);
gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
wcycle_set_reset_counters(wcycle, reset_counters);
}
snew(nrnb, 1);
if (cr->duty & DUTY_PP)
{
bcast_state(cr, state);
/* Initiate forcerecord */
fr = mk_forcerec();
fr->hwinfo = hwinfo;
fr->gpu_opt = &hw_opt->gpu_opt;
init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box,
opt2fn("-table", nfile, fnm),
opt2fn("-tabletf", nfile, fnm),
opt2fn("-tablep", nfile, fnm),
opt2fn("-tableb", nfile, fnm),
nbpu_opt,
FALSE,
pforce);
/* version for PCA_NOT_READ_NODE (see md.c) */
/*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
"nofile","nofile","nofile","nofile",FALSE,pforce);
*/
/* Initialize QM-MM */
if (fr->bQMMM)
{
init_QMMMrec(cr, mtop, inputrec, fr);
}
/* Initialize the mdatoms structure.
* mdatoms is not filled with atom data,
* as this can not be done now with domain decomposition.
*/
mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
/* Initialize the virtual site communication */
vsite = init_vsite(mtop, cr, FALSE);
calc_shifts(box, fr->shift_vec);
/* With periodic molecules the charge groups should be whole at start up
* and the virtual sites should not be far from their proper positions.
*/
if (!inputrec->bContinuation && MASTER(cr) &&
!(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
{
/* Make molecules whole at start of run */
if (fr->ePBC != epbcNONE)
{
do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
}
if (vsite)
{
/* Correct initial vsite positions are required
* for the initial distribution in the domain decomposition
* and for the initial shell prediction.
*/
construct_vsites_mtop(vsite, mtop, state->x);
}
}
if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
{
ewaldcoeff_q = fr->ewaldcoeff_q;
ewaldcoeff_lj = fr->ewaldcoeff_lj;
pmedata = &fr->pmedata;
}
else
{
pmedata = NULL;
}
}
else
{
/* This is a PME only node */
/* We don't need the state */
done_state(state);
ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
snew(pmedata, 1);
}
if (hw_opt->thread_affinity != threadaffOFF)
{
/* Before setting affinity, check whether the affinity has changed
* - which indicates that probably the OpenMP library has changed it
* since we first checked).
*/
gmx_check_thread_affinity_set(fplog, cr,
hw_opt, hwinfo->nthreads_hw_avail, TRUE);
/* Set the CPU affinity */
gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
}
/* Initiate PME if necessary,
* either on all nodes or on dedicated PME nodes only. */
if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
{
if (mdatoms)
{
nChargePerturbed = mdatoms->nChargePerturbed;
if (EVDW_PME(inputrec->vdwtype))
{
nTypePerturbed = mdatoms->nTypePerturbed;
}
}
if (cr->npmenodes > 0)
{
/* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
}
if (cr->duty & DUTY_PME)
{
status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
mtop ? mtop->natoms : 0, nChargePerturbed, nTypePerturbed,
(Flags & MD_REPRODUCIBLE), nthreads_pme);
if (status != 0)
{
gmx_fatal(FARGS, "Error %d initializing PME", status);
}
}
}
if (integrator[inputrec->eI].func == do_md)
{
/* Turn on signal handling on all nodes */
/*
* (A user signal from the PME nodes (if any)
* is communicated to the PP nodes.
*/
signal_handler_install();
}
if (cr->duty & DUTY_PP)
{
/* Assumes uniform use of the number of OpenMP threads */
walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
if (inputrec->bPull)
{
/* Initialize pull code */
inputrec->pull_work =
init_pull(fplog, inputrec->pull, inputrec, nfile, fnm,
mtop, cr, oenv, inputrec->fepvals->init_lambda,
EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
}
if (inputrec->bRot)
{
/* Initialize enforced rotation code */
init_rot(fplog, inputrec, nfile, fnm, cr, state->x, box, mtop, oenv,
bVerbose, Flags);
}
if (inputrec->eSwapCoords != eswapNO)
{
/* Initialize ion swapping code */
init_swapcoords(fplog, bVerbose, inputrec, opt2fn_master("-swap", nfile, fnm, cr),
mtop, state->x, state->box, &state->swapstate, cr, oenv, Flags);
}
constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
if (DOMAINDECOMP(cr))
{
GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
Flags & MD_DDBONDCHECK, fr->cginfo_mb);
set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, &ddbox);
setup_dd_grid(fplog, cr->dd);
}
/* Now do whatever the user wants us to do (how flexible...) */
integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
oenv, bVerbose, bCompact,
nstglobalcomm,
vsite, constr,
nstepout, inputrec, mtop,
fcd, state,
mdatoms, nrnb, wcycle, ed, fr,
repl_ex_nst, repl_ex_nex, repl_ex_seed,
membed,
cpt_period, max_hours,
imdport,
Flags,
walltime_accounting);
if (inputrec->bPull)
{
finish_pull(inputrec->pull_work);
}
if (inputrec->bRot)
{
finish_rot(inputrec->rot);
}
}
else
{
GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
/* do PME only */
walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
}
wallcycle_stop(wcycle, ewcRUN);
/* Finish up, write some stuff
* if rerunMD, don't write last frame again
*/
finish_run(fplog, cr,
inputrec, nrnb, wcycle, walltime_accounting,
fr ? fr->nbv : NULL,
EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
/* Free GPU memory and context */
free_gpu_resources(fr, cr, &hwinfo->gpu_info, fr ? fr->gpu_opt : NULL);
if (opt2bSet("-membed", nfile, fnm))
{
sfree(membed);
}
gmx_hardware_info_free(hwinfo);
/* Does what it says */
print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
walltime_accounting_destroy(walltime_accounting);
/* PLUMED */
if(plumedswitch){
plumed_finalize(plumedmain);
}
/* END PLUMED */
/* Close logfile already here if we were appending to it */
if (MASTER(cr) && (Flags & MD_APPENDFILES))
{
gmx_log_close(fplog);
}
rc = (int)gmx_get_stop_condition();
done_ed(&ed);
#ifdef GMX_THREAD_MPI
/* we need to join all threads. The sub-threads join when they
exit this function, but the master thread needs to be told to
wait for that. */
if (PAR(cr) && MASTER(cr))
{
tMPI_Finalize();
}
#endif
return rc;
}