gmx_wallcycle_t wallcycle_init(FILE *fplog, int resetstep, t_commrec gmx_unused *cr,
int nthreads_pp, int nthreads_pme)
{
gmx_wallcycle_t wc;
if (!wallcycle_have_counter())
{
return NULL;
}
snew(wc, 1);
wc->haveInvalidCount = FALSE;
wc->wc_barrier = FALSE;
wc->wcc_all = NULL;
wc->wc_depth = 0;
wc->ewc_prev = -1;
wc->reset_counters = resetstep;
wc->nthreads_pp = nthreads_pp;
wc->nthreads_pme = nthreads_pme;
wc->cycles_sum = NULL;
#ifdef GMX_MPI
if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != NULL)
{
if (fplog)
{
fprintf(fplog, "\nWill call MPI_Barrier before each cycle start/stop call\n\n");
}
wc->wc_barrier = TRUE;
wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
}
#endif
snew(wc->wcc, ewcNR);
if (getenv("GMX_CYCLE_ALL") != NULL)
{
if (fplog)
{
fprintf(fplog, "\nWill time all the code during the run\n\n");
}
snew(wc->wcc_all, ewcNR*ewcNR);
}
#ifdef GMX_CYCLE_SUBCOUNTERS
snew(wc->wcsc, ewcsNR);
#endif
#ifdef DEBUG_WCYCLE
wc->count_depth = 0;
#endif
return wc;
}
gmx_wallcycle_t wallcycle_init(FILE *fplog,t_commrec *cr)
{
gmx_wallcycle_t wc;
char *env_ptr;
if (!wallcycle_have_counter())
{
return NULL;
}
snew(wc,1);
wc->wc_barrier = FALSE;
wc->wcc_all = NULL;
wc->wc_depth = 0;
wc->ewc_prev = -1;
#ifdef GMX_MPI
if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != NULL)
{
if (fplog)
{
fprintf(fplog,"\nWill call MPI_Barrier before each cycle start/stop call\n\n");
}
wc->wc_barrier = TRUE;
wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
}
#endif
snew(wc->wcc,ewcNR);
if (getenv("GMX_CYCLE_ALL") != NULL)
{
if (fplog)
{
fprintf(fplog,"\nWill time all the code during the run\n\n");
}
snew(wc->wcc_all,ewcNR*ewcNR);
}
/* Read variable GMX_RESET_COUNTER from environment */
wc->reset_counters = -1;
if ((env_ptr=getenv("GMX_RESET_COUNTERS")) != NULL)
{
sscanf(env_ptr,gmx_step_pfmt,&wc->reset_counters);
}
return wc;
}
gmx_wallcycle_t wallcycle_init(FILE *fplog,int resetstep,t_commrec *cr)
{
gmx_wallcycle_t wc;
if (!wallcycle_have_counter())
{
return NULL;
}
snew(wc,1);
wc->wc_barrier = FALSE;
wc->wcc_all = NULL;
wc->wc_depth = 0;
wc->ewc_prev = -1;
wc->reset_counters = resetstep;
#ifdef GMX_MPI
if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != NULL)
{
if (fplog)
{
fprintf(fplog,"\nWill call MPI_Barrier before each cycle start/stop call\n\n");
}
wc->wc_barrier = TRUE;
wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
}
#endif
snew(wc->wcc,ewcNR);
if (getenv("GMX_CYCLE_ALL") != NULL)
{
/*#ifndef GMX_THREADS*/
if (fplog)
{
fprintf(fplog,"\nWill time all the code during the run\n\n");
}
snew(wc->wcc_all,ewcNR*ewcNR);
/*#else*/
gmx_fatal(FARGS, "GMX_CYCLE_ALL is incompatible with threaded code");
/*#endif*/
}
return wc;
}
void pme_loadbal_init(pme_load_balancing_t **pme_lb_p,
t_commrec *cr,
FILE *fp_log,
const t_inputrec *ir,
matrix box,
const interaction_const_t *ic,
struct gmx_pme_t *pmedata,
gmx_bool bUseGPU,
gmx_bool *bPrinting)
{
pme_load_balancing_t *pme_lb;
real spm, sp;
int d;
snew(pme_lb, 1);
pme_lb->bSepPMERanks = !(cr->duty & DUTY_PME);
/* Initially we turn on balancing directly on based on PP/PME imbalance */
pme_lb->bTriggerOnDLB = FALSE;
/* Any number of stages >= 2 is supported */
pme_lb->nstage = 2;
pme_lb->cutoff_scheme = ir->cutoff_scheme;
if (pme_lb->cutoff_scheme == ecutsVERLET)
{
pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
pme_lb->rbuf_vdw = pme_lb->rbuf_coulomb;
}
else
{
if (ic->rcoulomb > ic->rlist)
{
pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
}
else
{
pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
}
if (ic->rvdw > ic->rlist)
{
pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
}
else
{
pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
}
}
copy_mat(box, pme_lb->box_start);
if (ir->ePBC == epbcXY && ir->nwall == 2)
{
svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
}
pme_lb->n = 1;
snew(pme_lb->setup, pme_lb->n);
pme_lb->rcut_vdw = ic->rvdw;
pme_lb->rcut_coulomb_start = ir->rcoulomb;
pme_lb->nstcalclr_start = ir->nstcalclr;
pme_lb->cur = 0;
pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
pme_lb->setup[0].rlist = ic->rlist;
pme_lb->setup[0].rlistlong = ic->rlistlong;
pme_lb->setup[0].nstcalclr = ir->nstcalclr;
pme_lb->setup[0].grid[XX] = ir->nkx;
pme_lb->setup[0].grid[YY] = ir->nky;
pme_lb->setup[0].grid[ZZ] = ir->nkz;
pme_lb->setup[0].ewaldcoeff_q = ic->ewaldcoeff_q;
pme_lb->setup[0].ewaldcoeff_lj = ic->ewaldcoeff_lj;
pme_lb->setup[0].pmedata = pmedata;
spm = 0;
for (d = 0; d < DIM; d++)
{
sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
if (sp > spm)
{
spm = sp;
}
}
pme_lb->setup[0].spacing = spm;
if (ir->fourier_spacing > 0)
{
pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
}
else
{
pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
}
pme_lb->stage = 0;
pme_lb->fastest = 0;
pme_lb->lower_limit = 0;
pme_lb->start = 0;
pme_lb->end = 0;
pme_lb->elimited = epmelblimNO;
pme_lb->cycles_n = 0;
pme_lb->cycles_c = 0;
/* Tune with GPUs and/or separate PME ranks.
* When running only on a CPU without PME ranks, PME tuning will only help
* with small numbers of atoms in the cut-off sphere.
*/
pme_lb->bActive = (wallcycle_have_counter() && (bUseGPU ||
pme_lb->bSepPMERanks));
/* With GPUs and no separate PME ranks we can't measure the PP/PME
* imbalance, so we start balancing right away.
* Otherwise we only start balancing after we observe imbalance.
*/
pme_lb->bBalance = (pme_lb->bActive && (bUseGPU && !pme_lb->bSepPMERanks));
pme_lb->step_rel_stop = PMETunePeriod*ir->nstlist;
/* Delay DD load balancing when GPUs are used */
if (pme_lb->bActive && DOMAINDECOMP(cr) && cr->dd->nnodes > 1 && bUseGPU)
{
/* Lock DLB=auto to off (does nothing when DLB=yes/no.
* With GPUs and separate PME nodes, we want to first
* do PME tuning without DLB, since DLB might limit
* the cut-off, which never improves performance.
* We allow for DLB + PME tuning after a first round of tuning.
*/
dd_dlb_lock(cr->dd);
if (dd_dlb_is_locked(cr->dd))
{
md_print_warn(cr, fp_log, "NOTE: DLB will not turn on during the first phase of PME tuning\n");
}
}
*pme_lb_p = pme_lb;
*bPrinting = pme_lb->bBalance;
}
void pme_loadbal_init(pme_load_balancing_t **pme_lb_p,
const t_inputrec *ir, matrix box,
const interaction_const_t *ic,
struct gmx_pme_t *pmedata,
gmx_bool bUseGPU, gmx_bool bSepPMERanks,
gmx_bool *bPrinting)
{
pme_load_balancing_t *pme_lb;
real spm, sp;
int d;
snew(pme_lb, 1);
pme_lb->bSepPMERanks = bSepPMERanks;
/* Any number of stages >= 2 is supported */
pme_lb->nstage = 2;
pme_lb->cutoff_scheme = ir->cutoff_scheme;
if (pme_lb->cutoff_scheme == ecutsVERLET)
{
pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
pme_lb->rbuf_vdw = pme_lb->rbuf_coulomb;
}
else
{
if (ic->rcoulomb > ic->rlist)
{
pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
}
else
{
pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
}
if (ic->rvdw > ic->rlist)
{
pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
}
else
{
pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
}
}
copy_mat(box, pme_lb->box_start);
if (ir->ePBC == epbcXY && ir->nwall == 2)
{
svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
}
pme_lb->n = 1;
snew(pme_lb->setup, pme_lb->n);
pme_lb->rcut_vdw = ic->rvdw;
pme_lb->rcut_coulomb_start = ir->rcoulomb;
pme_lb->nstcalclr_start = ir->nstcalclr;
pme_lb->cur = 0;
pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
pme_lb->setup[0].rlist = ic->rlist;
pme_lb->setup[0].rlistlong = ic->rlistlong;
pme_lb->setup[0].nstcalclr = ir->nstcalclr;
pme_lb->setup[0].grid[XX] = ir->nkx;
pme_lb->setup[0].grid[YY] = ir->nky;
pme_lb->setup[0].grid[ZZ] = ir->nkz;
pme_lb->setup[0].ewaldcoeff_q = ic->ewaldcoeff_q;
pme_lb->setup[0].ewaldcoeff_lj = ic->ewaldcoeff_lj;
pme_lb->setup[0].pmedata = pmedata;
spm = 0;
for (d = 0; d < DIM; d++)
{
sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
if (sp > spm)
{
spm = sp;
}
}
pme_lb->setup[0].spacing = spm;
if (ir->fourier_spacing > 0)
{
pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
}
else
{
pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
}
pme_lb->stage = 0;
pme_lb->fastest = 0;
pme_lb->start = 0;
pme_lb->end = 0;
pme_lb->elimited = epmelblimNO;
pme_lb->cycles_n = 0;
pme_lb->cycles_c = 0;
/* Tune with GPUs and/or separate PME ranks.
* When running only on a CPU without PME ranks, PME tuning will only help
* with small numbers of atoms in the cut-off sphere.
*/
pme_lb->bActive = (wallcycle_have_counter() && (bUseGPU || bSepPMERanks));
/* With GPUs and no separate PME ranks we can't measure the PP/PME
* imbalance, so we start balancing right away.
* Otherwise we only start balancing after we observe imbalance.
*/
pme_lb->bBalance = (pme_lb->bActive && (bUseGPU && !bSepPMERanks));
*pme_lb_p = pme_lb;
*bPrinting = pme_lb->bBalance;
}
