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
posres_wrapper_lambda(struct gmx_wallcycle *wcycle,
const t_lambda *fepvals,
const t_idef *idef,
const struct t_pbc *pbc,
const rvec x[],
gmx_enerdata_t *enerd,
real *lambda,
t_forcerec *fr)
{
real v;
int i;
if (0 == idef->il[F_POSRES].nr)
{
return;
}
wallcycle_sub_start_nocount(wcycle, ewcsRESTRAINTS);
for (i = 0; i < enerd->n_lambda; i++)
{
real dvdl_dum = 0, lambda_dum;
lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : fepvals->all_lambda[efptRESTRAINT][i-1]);
v = posres(idef->il[F_POSRES].nr, idef->il[F_POSRES].iatoms,
idef->iparams_posres,
x, NULL, NULL,
fr->ePBC == epbcNONE ? NULL : pbc, lambda_dum, &dvdl_dum,
fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
enerd->enerpart_lambda[i] += v;
}
wallcycle_sub_stop(wcycle, ewcsRESTRAINTS);
}
/*! \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_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 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);
}
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];
}
}
}
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);
}
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 = fr->nthread_ewc;
#pragma omp parallel for num_threads(nthreads) schedule(static)
for (t = 0; t < nthreads; t++)
{
try
{
tensor *vir_q, *vir_lj;
real *Vcorrt_q, *Vcorrt_lj, *dvdlt_q, *dvdlt_lj;
if (t == 0)
{
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
{
vir_q = &fr->ewc_t[t].vir_q;
vir_lj = &fr->ewc_t[t].vir_lj;
Vcorrt_q = &fr->ewc_t[t].Vcorr_q;
Vcorrt_lj = &fr->ewc_t[t].Vcorr_lj;
dvdlt_q = &fr->ewc_t[t].dvdl[efptCOUL];
dvdlt_lj = &fr->ewc_t[t].dvdl[efptVDW];
clear_mat(*vir_q);
clear_mat(*vir_lj);
}
*dvdlt_q = 0;
*dvdlt_lj = 0;
/* Threading is only supported with the Verlet cut-off
* scheme and then only single particle forces (no
* exclusion forces) are calculated, so we can store
* the forces in the normal, single fr->f_novirsum array.
*/
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,
fr->f_novirsum, *vir_q, *vir_lj,
Vcorrt_q, Vcorrt_lj,
lambda[efptCOUL], lambda[efptVDW],
dvdlt_q, dvdlt_lj);
}
GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
}
if (nthreads > 1)
{
reduce_thread_energies(fr->vir_el_recip, fr->vir_lj_recip,
&Vcorr_q, &Vcorr_lj,
&dvdl_long_range_correction_q,
&dvdl_long_range_correction_lj,
nthreads, fr->ewc_t);
}
wallcycle_sub_stop(wcycle, ewcsEWALD_CORRECTION);
}
void
do_force_listed(gmx_wallcycle *wcycle,
matrix box,
const t_lambda *fepvals,
const gmx_multisim_t *ms,
const t_idef *idef,
const rvec x[],
history_t *hist,
rvec f[],
t_forcerec *fr,
const struct t_pbc *pbc,
const struct t_graph *graph,
gmx_enerdata_t *enerd,
t_nrnb *nrnb,
real *lambda,
const t_mdatoms *md,
t_fcdata *fcd,
int *global_atom_index,
int flags)
{
t_pbc pbc_full; /* Full PBC is needed for position restraints */
if (!(flags & GMX_FORCE_LISTED))
{
return;
}
if ((idef->il[F_POSRES].nr > 0) ||
(idef->il[F_FBPOSRES].nr > 0))
{
/* Not enough flops to bother counting */
set_pbc(&pbc_full, fr->ePBC, box);
}
calc_listed(ms, wcycle, idef, x, hist, f, fr, pbc, &pbc_full,
graph, enerd, nrnb, lambda, md, fcd,
global_atom_index, flags);
/* Check if we have to determine energy differences
* at foreign lambda's.
*/
if (fepvals->n_lambda > 0 && (flags & GMX_FORCE_DHDL))
{
posres_wrapper_lambda(wcycle, fepvals, idef, &pbc_full, x, enerd, lambda, fr);
if (idef->ilsort != ilsortNO_FE)
{
wallcycle_sub_start(wcycle, ewcsLISTED_FEP);
if (idef->ilsort != ilsortFE_SORTED)
{
gmx_incons("The bonded interactions are not sorted for free energy");
}
for (int i = 0; i < enerd->n_lambda; i++)
{
real lam_i[efptNR];
reset_foreign_enerdata(enerd);
for (int j = 0; j < efptNR; j++)
{
lam_i[j] = (i == 0 ? lambda[j] : fepvals->all_lambda[j][i-1]);
}
calc_listed_lambda(idef, x, fr, pbc, graph, &(enerd->foreign_grpp), enerd->foreign_term, nrnb, lam_i, md,
fcd, global_atom_index);
sum_epot(&(enerd->foreign_grpp), enerd->foreign_term);
enerd->enerpart_lambda[i] += enerd->foreign_term[F_EPOT];
}
wallcycle_sub_stop(wcycle, ewcsLISTED_FEP);
}
}
debug_gmx();
}
void calc_listed(const gmx_multisim_t *ms,
gmx_wallcycle *wcycle,
const t_idef *idef,
const rvec x[], history_t *hist,
rvec f[], t_forcerec *fr,
const struct t_pbc *pbc,
const struct t_pbc *pbc_full,
const struct t_graph *g,
gmx_enerdata_t *enerd, t_nrnb *nrnb,
real *lambda,
const t_mdatoms *md,
t_fcdata *fcd, int *global_atom_index,
int force_flags)
{
gmx_bool bCalcEnerVir;
int i;
real dvdl[efptNR]; /* The dummy array is to have a place to store the dhdl at other values
of lambda, which will be thrown away in the end*/
const t_pbc *pbc_null;
int thread;
assert(fr->nthreads == idef->nthreads);
bCalcEnerVir = (force_flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY));
for (i = 0; i < efptNR; i++)
{
dvdl[i] = 0.0;
}
if (fr->bMolPBC)
{
pbc_null = pbc;
}
else
{
pbc_null = NULL;
}
#ifdef DEBUG
if (g && debug)
{
p_graph(debug, "Bondage is fun", g);
}
#endif
if ((idef->il[F_POSRES].nr > 0) ||
(idef->il[F_FBPOSRES].nr > 0) ||
(idef->il[F_ORIRES].nr > 0) ||
(idef->il[F_DISRES].nr > 0))
{
/* TODO Use of restraints triggers further function calls
inside the loop over calc_one_bond(), but those are too
awkward to account to this subtimer properly in the present
code. We don't test / care much about performance with
restraints, anyway. */
wallcycle_sub_start(wcycle, ewcsRESTRAINTS);
if (idef->il[F_POSRES].nr > 0)
{
posres_wrapper(nrnb, idef, pbc_full, x, enerd, lambda, fr);
}
if (idef->il[F_FBPOSRES].nr > 0)
{
fbposres_wrapper(nrnb, idef, pbc_full, x, enerd, fr);
}
/* Do pre force calculation stuff which might require communication */
if (idef->il[F_ORIRES].nr > 0)
{
enerd->term[F_ORIRESDEV] =
calc_orires_dev(ms, idef->il[F_ORIRES].nr,
idef->il[F_ORIRES].iatoms,
idef->iparams, md, x,
pbc_null, fcd, hist);
}
if (idef->il[F_DISRES].nr)
{
calc_disres_R_6(idef->il[F_DISRES].nr,
idef->il[F_DISRES].iatoms,
idef->iparams, x, pbc_null,
fcd, hist);
#ifdef GMX_MPI
if (fcd->disres.nsystems > 1)
{
gmx_sum_sim(2*fcd->disres.nres, fcd->disres.Rt_6, ms);
}
#endif
}
wallcycle_sub_stop(wcycle, ewcsRESTRAINTS);
}
wallcycle_sub_start(wcycle, ewcsLISTED);
#pragma omp parallel for num_threads(fr->nthreads) schedule(static)
for (thread = 0; thread < fr->nthreads; thread++)
{
int ftype;
real *epot, v;
/* thread stuff */
rvec *ft, *fshift;
real *dvdlt;
gmx_grppairener_t *grpp;
if (thread == 0)
{
ft = f;
fshift = fr->fshift;
epot = enerd->term;
grpp = &enerd->grpp;
dvdlt = dvdl;
}
else
{
zero_thread_forces(&fr->f_t[thread], fr->natoms_force,
fr->red_nblock, 1<<fr->red_ashift);
ft = fr->f_t[thread].f;
fshift = fr->f_t[thread].fshift;
epot = fr->f_t[thread].ener;
grpp = &fr->f_t[thread].grpp;
dvdlt = fr->f_t[thread].dvdl;
}
/* Loop over all bonded force types to calculate the bonded forces */
for (ftype = 0; (ftype < F_NRE); ftype++)
{
if (idef->il[ftype].nr > 0 && ftype_is_bonded_potential(ftype))
{
v = calc_one_bond(thread, ftype, idef, x,
ft, fshift, fr, pbc_null, g, grpp,
nrnb, lambda, dvdlt,
md, fcd, bCalcEnerVir,
global_atom_index);
epot[ftype] += v;
}
}
}
wallcycle_sub_stop(wcycle, ewcsLISTED);
if (fr->nthreads > 1)
{
wallcycle_sub_start(wcycle, ewcsLISTED_BUF_OPS);
reduce_thread_forces(fr->natoms_force, f, fr->fshift,
enerd->term, &enerd->grpp, dvdl,
fr->nthreads, fr->f_t,
fr->red_nblock, 1<<fr->red_ashift,
bCalcEnerVir,
force_flags & GMX_FORCE_DHDL);
wallcycle_sub_stop(wcycle, ewcsLISTED_BUF_OPS);
}
/* Remaining code does not have enough flops to bother counting */
if (force_flags & GMX_FORCE_DHDL)
{
for (i = 0; i < efptNR; i++)
{
enerd->dvdl_nonlin[i] += dvdl[i];
}
}
/* Copy the sum of violations for the distance restraints from fcd */
if (fcd)
{
enerd->term[F_DISRESVIOL] = fcd->disres.sumviol;
}
}
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 calc_listed(const t_commrec *cr,
const gmx_multisim_t *ms,
struct gmx_wallcycle *wcycle,
const t_idef *idef,
const rvec x[], history_t *hist,
rvec f[],
gmx::ForceWithVirial *forceWithVirial,
const t_forcerec *fr,
const struct t_pbc *pbc,
const struct t_pbc *pbc_full,
const struct t_graph *g,
gmx_enerdata_t *enerd, t_nrnb *nrnb,
const real *lambda,
const t_mdatoms *md,
t_fcdata *fcd, int *global_atom_index,
int force_flags)
{
gmx_bool bCalcEnerVir;
const t_pbc *pbc_null;
bonded_threading_t *bt = fr->bondedThreading;
bCalcEnerVir = ((force_flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0);
if (fr->bMolPBC)
{
pbc_null = pbc;
}
else
{
pbc_null = nullptr;
}
if ((idef->il[F_POSRES].nr > 0) ||
(idef->il[F_FBPOSRES].nr > 0) ||
fcd->orires.nr > 0 ||
fcd->disres.nres > 0)
{
/* TODO Use of restraints triggers further function calls
inside the loop over calc_one_bond(), but those are too
awkward to account to this subtimer properly in the present
code. We don't test / care much about performance with
restraints, anyway. */
wallcycle_sub_start(wcycle, ewcsRESTRAINTS);
if (idef->il[F_POSRES].nr > 0)
{
posres_wrapper(nrnb, idef, pbc_full, x, enerd, lambda, fr,
forceWithVirial);
}
if (idef->il[F_FBPOSRES].nr > 0)
{
fbposres_wrapper(nrnb, idef, pbc_full, x, enerd, fr,
forceWithVirial);
}
/* Do pre force calculation stuff which might require communication */
if (fcd->orires.nr > 0)
{
/* This assertion is to ensure we have whole molecules.
* Unfortunately we do not have an mdrun state variable that tells
* us if molecules in x are not broken over PBC, so we have to make
* do with checking graph!=nullptr, which should tell us if we made
* molecules whole before calling the current function.
*/
GMX_RELEASE_ASSERT(fr->ePBC == epbcNONE || g != nullptr, "With orientation restraints molecules should be whole");
enerd->term[F_ORIRESDEV] =
calc_orires_dev(ms, idef->il[F_ORIRES].nr,
idef->il[F_ORIRES].iatoms,
idef->iparams, md, x,
pbc_null, fcd, hist);
}
if (fcd->disres.nres > 0)
{
calc_disres_R_6(cr, ms,
idef->il[F_DISRES].nr,
idef->il[F_DISRES].iatoms,
x, pbc_null,
fcd, hist);
}
wallcycle_sub_stop(wcycle, ewcsRESTRAINTS);
}
if (bt->haveBondeds)
{
wallcycle_sub_start(wcycle, ewcsLISTED);
/* The dummy array is to have a place to store the dhdl at other values
of lambda, which will be thrown away in the end */
real dvdl[efptNR] = {0};
calcBondedForces(idef, x, fr, pbc_null, g, enerd, nrnb, lambda, dvdl, md,
fcd, bCalcEnerVir, global_atom_index);
wallcycle_sub_stop(wcycle, ewcsLISTED);
wallcycle_sub_start(wcycle, ewcsLISTED_BUF_OPS);
reduce_thread_output(fr->natoms_force, f, fr->fshift,
enerd->term, &enerd->grpp, dvdl,
bt,
bCalcEnerVir,
(force_flags & GMX_FORCE_DHDL) != 0);
if (force_flags & GMX_FORCE_DHDL)
{
for (int i = 0; i < efptNR; i++)
{
enerd->dvdl_nonlin[i] += dvdl[i];
}
}
wallcycle_sub_stop(wcycle, ewcsLISTED_BUF_OPS);
}
/* Copy the sum of violations for the distance restraints from fcd */
if (fcd)
{
enerd->term[F_DISRESVIOL] = fcd->disres.sumviol;
}
}