The code could be refactored to use a series of explicit OpenMP
barriers so that the wallcycle_start/stop are more clearly matched,
but that's too hard for now. The comments make it harder for anyone
to misunderstand or mistakenly change the code.
Change-Id: Iec00e50e26a60efe09afb6f5744d401a9d9c06a3
inc_nrnb(nrnb, eNR_FFT, 2*npme);
}
inc_nrnb(nrnb, eNR_FFT, 2*npme);
}
+ /* Note: this wallcycle region is closed below
+ outside an OpenMP region, so take care if
+ refactoring code here. */
wallcycle_start(wcycle, ewcPME_SPREADGATHER);
}
wallcycle_start(wcycle, ewcPME_SPREADGATHER);
}
inc_nrnb(nrnb, eNR_GATHERFBSP,
pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
inc_nrnb(nrnb, eNR_GATHERFBSP,
pme->pme_order*pme->pme_order*pme->pme_order*pme->atc[0].n);
+ /* Note: this wallcycle region is opened above inside an OpenMP
+ region, so take care if refactoring code here. */
wallcycle_stop(wcycle, ewcPME_SPREADGATHER);
}
wallcycle_stop(wcycle, ewcPME_SPREADGATHER);
}