/*! This function computes the gravitational forces for all active
* particles. If needed, a new tree is constructed, otherwise the
* dynamically updated tree is used. Particles are only exported to other
* processors when really needed, thereby allowing a good use of the
* communication buffer.
*/
void gravity_tree(void)
{
long long ntot;
int numnodes, nexportsum = 0;
int i, j, iter = 0;
int *numnodeslist, maxnumnodes, nexport, *numlist, *nrecv, *ndonelist;
double tstart, tend, timetree = 0, timecommsumm = 0, timeimbalance = 0, sumimbalance;
double ewaldcount;
double costtotal, ewaldtot, *costtreelist, *ewaldlist;
double maxt, sumt, *timetreelist, *timecommlist;
double fac, plb, plb_max, sumcomm;
#ifndef NOGRAVITY
int *noffset, *nbuffer, *nsend, *nsend_local;
long long ntotleft;
int ndone, maxfill, ngrp;
int k, place;
int level, sendTask, recvTask;
double ax, ay, az;
MPI_Status status;
#endif
#ifdef ADD_CENTRAL_GRAVITY
int numsinks,root,globalroot,liveStar,liveStarGlobal;
double starData[4],r,h,h_inv,h3_inv,u,starGrav[3],starGravGlobal[3];
#endif
/* set new softening lengths */
if(All.ComovingIntegrationOn)
set_softenings();
/* contruct tree if needed */
tstart = second();
if(TreeReconstructFlag)
{
if(ThisTask == 0)
printf("Tree construction.\n");
force_treebuild(NumPart);
TreeReconstructFlag = 0;
if(ThisTask == 0)
printf("Tree construction done.\n");
}
tend = second();
All.CPU_TreeConstruction += timediff(tstart, tend);
costtotal = ewaldcount = 0;
/* Note: 'NumForceUpdate' has already been determined in find_next_sync_point_and_drift() */
numlist = malloc(NTask * sizeof(int) * NTask);
MPI_Allgather(&NumForceUpdate, 1, MPI_INT, numlist, 1, MPI_INT, MPI_COMM_WORLD);
for(i = 0, ntot = 0; i < NTask; i++)
ntot += numlist[i];
free(numlist);
#ifndef NOGRAVITY
if(ThisTask == 0)
printf("Begin tree force.\n");
#ifdef SELECTIVE_NO_GRAVITY
for(i = 0; i < NumPart; i++)
if(((1 << P[i].Type) & (SELECTIVE_NO_GRAVITY)))
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
#endif
noffset = malloc(sizeof(int) * NTask); /* offsets of bunches in common list */
nbuffer = malloc(sizeof(int) * NTask);
nsend_local = malloc(sizeof(int) * NTask);
nsend = malloc(sizeof(int) * NTask * NTask);
ndonelist = malloc(sizeof(int) * NTask);
i = 0; /* beginn with this index */
ntotleft = ntot; /* particles left for all tasks together */
while(ntotleft > 0)
{
iter++;
for(j = 0; j < NTask; j++)
nsend_local[j] = 0;
/* do local particles and prepare export list */
tstart = second();
for(nexport = 0, ndone = 0; i < NumPart && nexport < All.BunchSizeForce - NTask; i++)
{
if(P[i].Ti_endstep == All.Ti_Current)
{
ndone++;
for(j = 0; j < NTask; j++)
Exportflag[j] = 0;
#ifndef PMGRID
costtotal += force_treeevaluate(i, 0, &ewaldcount);
#else
costtotal += force_treeevaluate_shortrange(i, 0);
#endif
for(j = 0; j < NTask; j++)
{
if(Exportflag[j])
{
for(k = 0; k < 3; k++)
GravDataGet[nexport].u.Pos[k] = P[i].Pos[k];
#ifdef UNEQUALSOFTENINGS
GravDataGet[nexport].Type = P[i].Type;
#ifdef ADAPTIVE_GRAVSOFT_FORGAS
if(P[i].Type == 0)
GravDataGet[nexport].Soft = SphP[i].Hsml;
#endif
#endif
GravDataGet[nexport].w.OldAcc = P[i].OldAcc;
GravDataIndexTable[nexport].Task = j;
GravDataIndexTable[nexport].Index = i;
GravDataIndexTable[nexport].SortIndex = nexport;
nexport++;
nexportsum++;
nsend_local[j]++;
}
}
}
}
tend = second();
timetree += timediff(tstart, tend);
qsort(GravDataIndexTable, nexport, sizeof(struct gravdata_index), grav_tree_compare_key);
for(j = 0; j < nexport; j++)
GravDataIn[j] = GravDataGet[GravDataIndexTable[j].SortIndex];
for(j = 1, noffset[0] = 0; j < NTask; j++)
noffset[j] = noffset[j - 1] + nsend_local[j - 1];
tstart = second();
MPI_Allgather(nsend_local, NTask, MPI_INT, nsend, NTask, MPI_INT, MPI_COMM_WORLD);
tend = second();
timeimbalance += timediff(tstart, tend);
/* now do the particles that need to be exported */
for(level = 1; level < (1 << PTask); level++)
{
tstart = second();
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* get the particles */
MPI_Sendrecv(&GravDataIn[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_GRAV_A,
&GravDataGet[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_GRAV_A, MPI_COMM_WORLD, &status);
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tend = second();
timecommsumm += timediff(tstart, tend);
tstart = second();
for(j = 0; j < nbuffer[ThisTask]; j++)
{
#ifndef PMGRID
costtotal += force_treeevaluate(j, 1, &ewaldcount);
#else
costtotal += force_treeevaluate_shortrange(j, 1);
#endif
}
tend = second();
timetree += timediff(tstart, tend);
tstart = second();
MPI_Barrier(MPI_COMM_WORLD);
tend = second();
timeimbalance += timediff(tstart, tend);
/* get the result */
tstart = second();
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* send the results */
MPI_Sendrecv(&GravDataResult[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_GRAV_B,
&GravDataOut[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_GRAV_B, MPI_COMM_WORLD, &status);
/* add the result to the particles */
for(j = 0; j < nsend_local[recvTask]; j++)
{
place = GravDataIndexTable[noffset[recvTask] + j].Index;
for(k = 0; k < 3; k++)
P[place].GravAccel[k] += GravDataOut[j + noffset[recvTask]].u.Acc[k];
P[place].GravCost += GravDataOut[j + noffset[recvTask]].w.Ninteractions;
}
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tend = second();
timecommsumm += timediff(tstart, tend);
level = ngrp - 1;
}
MPI_Allgather(&ndone, 1, MPI_INT, ndonelist, 1, MPI_INT, MPI_COMM_WORLD);
for(j = 0; j < NTask; j++)
ntotleft -= ndonelist[j];
}
free(ndonelist);
free(nsend);
free(nsend_local);
free(nbuffer);
free(noffset);
/* now add things for comoving integration */
#ifndef PERIODIC
#ifndef PMGRID
if(All.ComovingIntegrationOn)
{
fac = 0.5 * All.Hubble * All.Hubble * All.Omega0 / All.G;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] += fac * P[i].Pos[j];
}
#endif
#endif
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
{
#ifdef PMGRID
ax = P[i].GravAccel[0] + P[i].GravPM[0] / All.G;
ay = P[i].GravAccel[1] + P[i].GravPM[1] / All.G;
az = P[i].GravAccel[2] + P[i].GravPM[2] / All.G;
#else
ax = P[i].GravAccel[0];
ay = P[i].GravAccel[1];
az = P[i].GravAccel[2];
#endif
P[i].OldAcc = sqrt(ax * ax + ay * ay + az * az);
}
if(All.TypeOfOpeningCriterion == 1)
All.ErrTolTheta = 0; /* This will switch to the relative opening criterion for the following force computations */
/* muliply by G */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
{
P[i].GravAccel[j] *= All.G;
//printf("Gravity! %g\n",P[i].GravAccel[j]);
}
/* Finally, the following factor allows a computation of a cosmological simulation
with vacuum energy in physical coordinates */
#ifndef PERIODIC
#ifndef PMGRID
if(All.ComovingIntegrationOn == 0)
{
fac = All.OmegaLambda * All.Hubble * All.Hubble;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] += fac * P[i].Pos[j];
}
#endif
#endif
#ifdef SELECTIVE_NO_GRAVITY
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
#endif
if(ThisTask == 0)
printf("tree is done.\n");
#else /* gravity is switched off */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] = 0;
#endif
#ifdef SINK_GRAV_ONLY
sink_grav();
#endif
#ifdef ADD_CENTRAL_GRAVITY
/* Get the position and mass of the central object and send it to everyone */
numsinks=NumPart - N_gas;
starData[0]=starData[1]=starData[2]=starData[3]= -1.0;
root=-1;
liveStar=0;
for(i=0; i<3; i++)
{
starGrav[i]=0.0;
starGravGlobal[i]=0.0;
}
for(i=0; i<numsinks;i++)
{
if(P[i+N_gas].ID==All.StarID)
{
starData[0] = P[i+N_gas].Pos[0];
starData[1] = P[i+N_gas].Pos[1];
starData[2] = P[i+N_gas].Pos[2];
starData[3] = P[i+N_gas].Mass;
root = ThisTask;
//Do we need to update the star's gravity?
if(P[i+N_gas].Ti_endstep == All.Ti_Current)
{
liveStar=1;
}
}
}
/* Get the node that has the data */
MPI_Barrier(MPI_COMM_WORLD);
MPI_Allreduce(&root,&globalroot,1,MPI_INT,MPI_MAX,MPI_COMM_WORLD);
MPI_Allreduce(&liveStar,&liveStarGlobal,1,MPI_INT,MPI_MAX,MPI_COMM_WORLD);
/* Broadcast it. */
MPI_Bcast(&starData,4,MPI_DOUBLE,globalroot,MPI_COMM_WORLD);
//We have the central object mass and position, add its gravity, it is softened by the type 1 softening...
h = All.ForceSoftening[1];
h_inv = 1.0 / h;
h3_inv = h_inv * h_inv * h_inv;
for(i = 0; i < NumPart; i++)
{
if(P[i].ID != All.StarID)
{
//If we need to update the star's gravity we need to calculate this for all particles...
if(liveStarGlobal)
{
r=sqrt((starData[0]-P[i].Pos[0])*(starData[0]-P[i].Pos[0])+(starData[1]-P[i].Pos[1])*(starData[1]-P[i].Pos[1])+(starData[2]-P[i].Pos[2])*(starData[2]-P[i].Pos[2]));
if(r >= h)
{
fac = 1 / (r*r*r);
}
else
{
u = r * h_inv;
if(u < 0.5)
{
fac = h3_inv * (10.666666666667 + u * u * (32.0 * u - 38.4));
}
else
{
fac = h3_inv * (21.333333333333 - 48.0 * u +
38.4 * u * u - 10.666666666667 * u * u * u - 0.066666666667 / (u * u * u));
}
}
for(j=0;j<3;j++)
{
starGrav[j]+=(P[i].Pos[j]-starData[j])*All.G*P[i].Mass*fac;
}
}
//Otherwise, just give the star's gravity to those that need it
if(P[i].Ti_endstep == All.Ti_Current)
{
r=sqrt((starData[0]-P[i].Pos[0])*(starData[0]-P[i].Pos[0])+(starData[1]-P[i].Pos[1])*(starData[1]-P[i].Pos[1])+(starData[2]-P[i].Pos[2])*(starData[2]-P[i].Pos[2]));
if(r >= h)
{
fac = 1 / (r*r*r);
}
else
{
h_inv = 1.0 / h;
h3_inv = h_inv * h_inv * h_inv;
u = r * h_inv;
if(u < 0.5)
{
fac = h3_inv * (10.666666666667 + u * u * (32.0 * u - 38.4));
}
else
{
fac = h3_inv * (21.333333333333 - 48.0 * u +
38.4 * u * u - 10.666666666667 * u * u * u - 0.066666666667 / (u * u * u));
}
}
for(j=0;j<3;j++)
{
P[i].GravAccel[j]+=(starData[j]-P[i].Pos[j])*All.G*starData[3]*fac;
}
}
}
}
//Gather the forces of the pcles on the star together and add them to the star
if(liveStarGlobal)
{
//Finally we need to combine all the starGrav values for the star...
MPI_Barrier(MPI_COMM_WORLD);
MPI_Allreduce(&starGrav[0],&starGravGlobal[0],1,MPI_DOUBLE,MPI_SUM,MPI_COMM_WORLD);
MPI_Allreduce(&starGrav[1],&starGravGlobal[1],1,MPI_DOUBLE,MPI_SUM,MPI_COMM_WORLD);
MPI_Allreduce(&starGrav[2],&starGravGlobal[2],1,MPI_DOUBLE,MPI_SUM,MPI_COMM_WORLD);
//Finally, find the actual star and add it
if(globalroot==ThisTask)
{
for(i=0; i<numsinks;i++)
{
if(P[i+N_gas].ID==All.StarID)
{
for(j=0;j<3;j++)
{
P[i+N_gas].GravAccel[j]+=starGravGlobal[j];
}
}
}
}
}
MPI_Barrier(MPI_COMM_WORLD);
#endif
/* Now the force computation is finished */
/* gather some diagnostic information */
timetreelist = malloc(sizeof(double) * NTask);
timecommlist = malloc(sizeof(double) * NTask);
costtreelist = malloc(sizeof(double) * NTask);
numnodeslist = malloc(sizeof(int) * NTask);
ewaldlist = malloc(sizeof(double) * NTask);
nrecv = malloc(sizeof(int) * NTask);
numnodes = Numnodestree;
MPI_Gather(&costtotal, 1, MPI_DOUBLE, costtreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&numnodes, 1, MPI_INT, numnodeslist, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&timetree, 1, MPI_DOUBLE, timetreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&timecommsumm, 1, MPI_DOUBLE, timecommlist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&NumPart, 1, MPI_INT, nrecv, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&ewaldcount, 1, MPI_DOUBLE, ewaldlist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Reduce(&nexportsum, &nexport, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(&timeimbalance, &sumimbalance, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if(ThisTask == 0)
{
All.TotNumOfForces += ntot;
fprintf(FdTimings, "Step= %d t= %g dt= %g \n", All.NumCurrentTiStep, All.Time, All.TimeStep);
fprintf(FdTimings, "Nf= %d%09d total-Nf= %d%09d ex-frac= %g iter= %d\n",
(int) (ntot / 1000000000), (int) (ntot % 1000000000),
(int) (All.TotNumOfForces / 1000000000), (int) (All.TotNumOfForces % 1000000000),
nexport / ((double) ntot), iter);
/* note: on Linux, the 8-byte integer could be printed with the format identifier "%qd", but doesn't work on AIX */
fac = NTask / ((double) All.TotNumPart);
for(i = 0, maxt = timetreelist[0], sumt = 0, plb_max = 0,
maxnumnodes = 0, costtotal = 0, sumcomm = 0, ewaldtot = 0; i < NTask; i++)
{
costtotal += costtreelist[i];
sumcomm += timecommlist[i];
if(maxt < timetreelist[i])
maxt = timetreelist[i];
sumt += timetreelist[i];
plb = nrecv[i] * fac;
if(plb > plb_max)
plb_max = plb;
if(numnodeslist[i] > maxnumnodes)
maxnumnodes = numnodeslist[i];
ewaldtot += ewaldlist[i];
}
fprintf(FdTimings, "work-load balance: %g max=%g avg=%g PE0=%g\n",
maxt / (sumt / NTask), maxt, sumt / NTask, timetreelist[0]);
fprintf(FdTimings, "particle-load balance: %g\n", plb_max);
fprintf(FdTimings, "max. nodes: %d, filled: %g\n", maxnumnodes,
maxnumnodes / (All.TreeAllocFactor * All.MaxPart));
fprintf(FdTimings, "part/sec=%g | %g ia/part=%g (%g)\n", ntot / (sumt + 1.0e-20),
ntot / (maxt * NTask), ((double) (costtotal)) / ntot, ((double) ewaldtot) / ntot);
fprintf(FdTimings, "\n");
fflush(FdTimings);
All.CPU_TreeWalk += sumt / NTask;
All.CPU_Imbalance += sumimbalance / NTask;
All.CPU_CommSum += sumcomm / NTask;
}
free(nrecv);
free(ewaldlist);
free(numnodeslist);
free(costtreelist);
free(timecommlist);
free(timetreelist);
}
/*! This function computes the gravitational forces for all active
* particles. If needed, a new tree is constructed, otherwise the
* dynamically updated tree is used. Particles are only exported to other
* processors when really needed, thereby allowing a good use of the
* communication buffer.
*/
void gravity_tree(void)
{
long long ntot;
int numnodes, nexportsum = 0;
int i, j, iter = 0;
int *numnodeslist, maxnumnodes, nexport, *numlist, *nrecv, *ndonelist;
double tstart, tend, timetree = 0, timecommsumm = 0, timeimbalance = 0, sumimbalance;
double ewaldcount;
double costtotal, ewaldtot, *costtreelist, *ewaldlist;
double maxt, sumt, *timetreelist, *timecommlist;
double fac, plb, plb_max, sumcomm;
#ifndef NOGRAVITY
int *noffset, *nbuffer, *nsend, *nsend_local;
long long ntotleft;
int ndone, maxfill, ngrp;
int k, place;
int level, sendTask, recvTask;
double ax, ay, az;
MPI_Status status;
#endif
/* set new softening lengths */
if(All.ComovingIntegrationOn)
set_softenings();
/* contruct tree if needed */
tstart = second();
if(TreeReconstructFlag)
{
if(ThisTask == 0)
printf("Tree construction.\n");
force_treebuild(NumPart);
TreeReconstructFlag = 0;
if(ThisTask == 0)
printf("Tree construction done.\n");
}
tend = second();
All.CPU_TreeConstruction += timediff(tstart, tend);
costtotal = ewaldcount = 0;
/* Note: 'NumForceUpdate' has already been determined in find_next_sync_point_and_drift() */
numlist = malloc(NTask * sizeof(int) * NTask);
MPI_Allgather(&NumForceUpdate, 1, MPI_INT, numlist, 1, MPI_INT, MPI_COMM_WORLD);
for(i = 0, ntot = 0; i < NTask; i++)
ntot += numlist[i];
free(numlist);
#ifndef NOGRAVITY
if(ThisTask == 0)
printf("Begin tree force.\n");
#ifdef SELECTIVE_NO_GRAVITY
for(i = 0; i < NumPart; i++)
if(((1 << P[i].Type) & (SELECTIVE_NO_GRAVITY)))
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
#endif
noffset = malloc(sizeof(int) * NTask); /* offsets of bunches in common list */
nbuffer = malloc(sizeof(int) * NTask);
nsend_local = malloc(sizeof(int) * NTask);
nsend = malloc(sizeof(int) * NTask * NTask);
ndonelist = malloc(sizeof(int) * NTask);
i = 0; /* beginn with this index */
ntotleft = ntot; /* particles left for all tasks together */
while(ntotleft > 0)
{
//printf("nontotleft %d, iter %d\n", ntotleft, iter);
iter++;
for(j = 0; j < NTask; j++)
nsend_local[j] = 0;
/* do local particles and prepare export list */
tstart = second();
for(nexport = 0, ndone = 0; i < NumPart && nexport < All.BunchSizeForce - NTask; i++)
if(P[i].Ti_endstep == All.Ti_Current)
{
ndone++;
for(j = 0; j < NTask; j++)
Exportflag[j] = 0;
#ifndef PMGRID
costtotal += force_treeevaluate(i, 0, &ewaldcount);
#else
costtotal += force_treeevaluate_shortrange(i, 0);
#endif
for(j = 0; j < NTask; j++)
{
if(Exportflag[j])
{
for(k = 0; k < 3; k++)
GravDataGet[nexport].u.Pos[k] = P[i].Pos[k];
// KC 8/11/14 Need to export single particle masses now
GravDataGet[nexport].Mass = P[i].Mass;
#ifdef UNEQUALSOFTENINGS
GravDataGet[nexport].Type = P[i].Type;
#ifdef ADAPTIVE_GRAVSOFT_FORGAS
if(P[i].Type == 0)
GravDataGet[nexport].Soft = SphP[i].Hsml;
#endif
#endif
GravDataGet[nexport].w.OldAcc = P[i].OldAcc;
GravDataIndexTable[nexport].Task = j;
GravDataIndexTable[nexport].Index = i;
GravDataIndexTable[nexport].SortIndex = nexport;
nexport++;
nexportsum++;
nsend_local[j]++;
}
}
}
tend = second();
timetree += timediff(tstart, tend);
qsort(GravDataIndexTable, nexport, sizeof(struct gravdata_index), grav_tree_compare_key);
for(j = 0; j < nexport; j++)
GravDataIn[j] = GravDataGet[GravDataIndexTable[j].SortIndex];
for(j = 1, noffset[0] = 0; j < NTask; j++)
noffset[j] = noffset[j - 1] + nsend_local[j - 1];
tstart = second();
MPI_Allgather(nsend_local, NTask, MPI_INT, nsend, NTask, MPI_INT, MPI_COMM_WORLD);
tend = second();
timeimbalance += timediff(tstart, tend);
/* now do the particles that need to be exported */
for(level = 1; level < (1 << PTask); level++)
{
tstart = second();
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* get the particles */
MPI_Sendrecv(&GravDataIn[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_GRAV_A,
&GravDataGet[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_GRAV_A, MPI_COMM_WORLD, &status);
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tend = second();
timecommsumm += timediff(tstart, tend);
tstart = second();
for(j = 0; j < nbuffer[ThisTask]; j++)
{
#ifndef PMGRID
costtotal += force_treeevaluate(j, 1, &ewaldcount);
#else
costtotal += force_treeevaluate_shortrange(j, 1);
#endif
}
tend = second();
timetree += timediff(tstart, tend);
tstart = second();
MPI_Barrier(MPI_COMM_WORLD);
tend = second();
timeimbalance += timediff(tstart, tend);
/* get the result */
tstart = second();
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* send the results */
MPI_Sendrecv(&GravDataResult[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_GRAV_B,
&GravDataOut[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_GRAV_B, MPI_COMM_WORLD, &status);
/* add the result to the particles */
for(j = 0; j < nsend_local[recvTask]; j++)
{
place = GravDataIndexTable[noffset[recvTask] + j].Index;
for(k = 0; k < 3; k++)
P[place].GravAccel[k] += GravDataOut[j + noffset[recvTask]].u.Acc[k];
P[place].GravCost += GravDataOut[j + noffset[recvTask]].w.Ninteractions;
}
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tend = second();
timecommsumm += timediff(tstart, tend);
level = ngrp - 1;
}
MPI_Allgather(&ndone, 1, MPI_INT, ndonelist, 1, MPI_INT, MPI_COMM_WORLD);
for(j = 0; j < NTask; j++)
ntotleft -= ndonelist[j];
}
free(ndonelist);
free(nsend);
free(nsend_local);
free(nbuffer);
free(noffset);
// KC 10/22/14
// At this point, GravAcce[j] will contain the tree-walked force. If PMGRID is on, then this will
// be the shortrange stuff
#if defined PMGRID && defined DEBUG_NGRAVS_SHORTTREE
for(i = 0; i < NumPart; ++i)
fprintf(stderr, "%d\t%e\t%e\t%e\t%e\t%e\t%e\t%d\n", P[i].ID, P[i].Pos[0], P[i].Pos[1], P[i].Pos[2], P[i].GravAccel[0], P[i].GravAccel[1], P[i].GravAccel[2], P[i].Type);
endrun(5555);
#endif
/* now add things for comoving integration */
#ifndef PERIODIC
#ifndef PMGRID
if(All.ComovingIntegrationOn)
{
fac = 0.5 * All.Hubble * All.Hubble * All.Omega0 / All.G;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] += fac * P[i].Pos[j];
}
#endif
#endif
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
{
#ifdef PMGRID
ax = P[i].GravAccel[0] + P[i].GravPM[0] / All.G;
ay = P[i].GravAccel[1] + P[i].GravPM[1] / All.G;
az = P[i].GravAccel[2] + P[i].GravPM[2] / All.G;
#else
ax = P[i].GravAccel[0];
ay = P[i].GravAccel[1];
az = P[i].GravAccel[2];
#endif
P[i].OldAcc = sqrt(ax * ax + ay * ay + az * az);
}
if(All.TypeOfOpeningCriterion == 1)
All.ErrTolTheta = 0; /* This will switch to the relative opening criterion for the following force computations */
/* muliply by G */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] *= All.G;
/* Finally, the following factor allows a computation of a cosmological simulation
with vacuum energy in physical coordinates */
#ifndef PERIODIC
#ifndef PMGRID
if(All.ComovingIntegrationOn == 0)
{
fac = All.OmegaLambda * All.Hubble * All.Hubble;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] += fac * P[i].Pos[j];
}
#endif
#endif
#ifdef SELECTIVE_NO_GRAVITY
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
#endif
if(ThisTask == 0)
printf("tree is done.\n");
#else /* gravity is switched off */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] = 0;
#endif
/* Now the force computation is finished */
//printf("Tree force computation done.");
/* gather some diagnostic information */
timetreelist = malloc(sizeof(double) * NTask);
timecommlist = malloc(sizeof(double) * NTask);
costtreelist = malloc(sizeof(double) * NTask);
numnodeslist = malloc(sizeof(int) * NTask);
ewaldlist = malloc(sizeof(double) * NTask);
nrecv = malloc(sizeof(int) * NTask);
numnodes = Numnodestree;
MPI_Gather(&costtotal, 1, MPI_DOUBLE, costtreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&numnodes, 1, MPI_INT, numnodeslist, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&timetree, 1, MPI_DOUBLE, timetreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&timecommsumm, 1, MPI_DOUBLE, timecommlist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&NumPart, 1, MPI_INT, nrecv, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&ewaldcount, 1, MPI_DOUBLE, ewaldlist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Reduce(&nexportsum, &nexport, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(&timeimbalance, &sumimbalance, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if(ThisTask == 0)
{
All.TotNumOfForces += ntot;
fprintf(FdTimings, "Step= %d t= %g dt= %g \n", All.NumCurrentTiStep, All.Time, All.TimeStep);
fprintf(FdTimings, "Nf= %d%09d total-Nf= %d%09d ex-frac= %g iter= %d\n",
(int) (ntot / 1000000000), (int) (ntot % 1000000000),
(int) (All.TotNumOfForces / 1000000000), (int) (All.TotNumOfForces % 1000000000),
nexport / ((double) ntot), iter);
/* note: on Linux, the 8-byte integer could be printed with the format identifier "%qd", but doesn't work on AIX */
fac = NTask / ((double) All.TotNumPart);
for(i = 0, maxt = timetreelist[0], sumt = 0, plb_max = 0,
maxnumnodes = 0, costtotal = 0, sumcomm = 0, ewaldtot = 0; i < NTask; i++)
{
costtotal += costtreelist[i];
sumcomm += timecommlist[i];
if(maxt < timetreelist[i])
maxt = timetreelist[i];
sumt += timetreelist[i];
plb = nrecv[i] * fac;
if(plb > plb_max)
plb_max = plb;
if(numnodeslist[i] > maxnumnodes)
maxnumnodes = numnodeslist[i];
ewaldtot += ewaldlist[i];
}
fprintf(FdTimings, "work-load balance: %g max=%g avg=%g PE0=%g\n",
maxt / (sumt / NTask), maxt, sumt / NTask, timetreelist[0]);
fprintf(FdTimings, "particle-load balance: %g\n", plb_max);
fprintf(FdTimings, "max. nodes: %d, filled: %g\n", maxnumnodes,
maxnumnodes / (All.TreeAllocFactor * All.MaxPart));
fprintf(FdTimings, "part/sec=%g | %g ia/part=%g (%g)\n", ntot / (sumt + 1.0e-20),
ntot / (maxt * NTask), ((double) (costtotal)) / ntot, ((double) ewaldtot) / ntot);
fprintf(FdTimings, "\n");
fflush(FdTimings);
All.CPU_TreeWalk += sumt / NTask;
All.CPU_Imbalance += sumimbalance / NTask;
All.CPU_CommSum += sumcomm / NTask;
}
free(nrecv);
free(ewaldlist);
free(numnodeslist);
free(costtreelist);
free(timecommlist);
free(timetreelist);
}
/*! This function computes the gravitational potential for ALL the particles.
* First, the (short-range) tree potential is computed, and then, if needed,
* the long range PM potential is added.
*/
void compute_potential(void)
{
int i;
#ifndef NOGRAVITY
int j, k, ret, sendTask, recvTask;
int ndone, ndone_flag, dummy;
int ngrp, place, nexport, nimport;
double fac;
MPI_Status status;
double r2;
if(All.ComovingIntegrationOn)
set_softenings();
if(ThisTask == 0)
{
printf("Start computation of potential for all particles...\n");
fflush(stdout);
}
CPU_Step[CPU_MISC] += measure_time();
if(TreeReconstructFlag)
{
if(ThisTask == 0)
printf("Tree construction.\n");
CPU_Step[CPU_MISC] += measure_time();
#if defined(SFR) || defined(BLACK_HOLES)
rearrange_particle_sequence();
#endif
force_treebuild(NumPart, NULL);
CPU_Step[CPU_TREEBUILD] += measure_time();
TreeReconstructFlag = 0;
if(ThisTask == 0)
printf("Tree construction done.\n");
}
/* allocate buffers to arrange communication */
All.BunchSize =
(int) ((All.BufferSize * 1024 * 1024) / (sizeof(struct data_index) + sizeof(struct data_nodelist) +
sizeof(struct gravdata_in) + sizeof(struct potdata_out) +
sizemax(sizeof(struct gravdata_in),
sizeof(struct potdata_out))));
DataIndexTable = (struct data_index *) mymalloc(All.BunchSize * sizeof(struct data_index));
DataNodeList = (struct data_nodelist *) mymalloc(All.BunchSize * sizeof(struct data_nodelist));
for(i = 0; i < NumPart; i++)
if(P[i].Ti_current != All.Ti_Current)
drift_particle(i, All.Ti_Current);
i = 0; /* beginn with this index */
do
{
for(j = 0; j < NTask; j++)
{
Send_count[j] = 0;
Exportflag[j] = -1;
}
/* do local particles and prepare export list */
for(nexport = 0; i < NumPart; i++)
{
#ifndef PMGRID
ret = force_treeevaluate_potential(i, 0, &nexport, Send_count);
#else
ret = force_treeevaluate_potential_shortrange(i, 0, &nexport, Send_count);
#endif
if(ret < 0)
break; /* export buffer has filled up */
}
#ifdef MYSORT
mysort_dataindex(DataIndexTable, nexport, sizeof(struct data_index), data_index_compare);
#else
qsort(DataIndexTable, nexport, sizeof(struct data_index), data_index_compare);
#endif
MPI_Allgather(Send_count, NTask, MPI_INT, Sendcount_matrix, NTask, MPI_INT, MPI_COMM_WORLD);
for(j = 0, nimport = 0, Recv_offset[0] = 0, Send_offset[0] = 0; j < NTask; j++)
{
Recv_count[j] = Sendcount_matrix[j * NTask + ThisTask];
nimport += Recv_count[j];
if(j > 0)
{
Send_offset[j] = Send_offset[j - 1] + Send_count[j - 1];
Recv_offset[j] = Recv_offset[j - 1] + Recv_count[j - 1];
}
}
GravDataGet = (struct gravdata_in *) mymalloc(nimport * sizeof(struct gravdata_in));
GravDataIn = (struct gravdata_in *) mymalloc(nexport * sizeof(struct gravdata_in));
/* prepare particle data for export */
for(j = 0; j < nexport; j++)
{
place = DataIndexTable[j].Index;
for(k = 0; k < 3; k++)
GravDataIn[j].Pos[k] = P[place].Pos[k];
#ifdef UNEQUALSOFTENINGS
GravDataIn[j].Type = P[place].Type;
#ifdef ADAPTIVE_GRAVSOFT_FORGAS
if(P[place].Type == 0)
GravDataIn[j].Soft = SphP[place].Hsml;
#endif
#endif
GravDataIn[j].OldAcc = P[place].OldAcc;
for(k = 0; k < NODELISTLENGTH; k++)
GravDataIn[j].NodeList[k] = DataNodeList[DataIndexTable[j].IndexGet].NodeList[k];
}
/* exchange particle data */
for(ngrp = 1; ngrp < (1 << PTask); ngrp++)
{
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(Send_count[recvTask] > 0 || Recv_count[recvTask] > 0)
{
/* get the particles */
MPI_Sendrecv(&GravDataIn[Send_offset[recvTask]],
Send_count[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_POTENTIAL_A,
&GravDataGet[Recv_offset[recvTask]],
Recv_count[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_POTENTIAL_A, MPI_COMM_WORLD, &status);
}
}
}
myfree(GravDataIn);
PotDataResult = (struct potdata_out *) mymalloc(nimport * sizeof(struct potdata_out));
PotDataOut = (struct potdata_out *) mymalloc(nexport * sizeof(struct potdata_out));
/* now do the particles that were sent to us */
for(j = 0; j < nimport; j++)
{
#ifndef PMGRID
force_treeevaluate_potential(j, 1, &dummy, &dummy);
#else
force_treeevaluate_potential_shortrange(j, 1, &dummy, &dummy);
#endif
}
if(i >= NumPart)
ndone_flag = 1;
else
ndone_flag = 0;
MPI_Allreduce(&ndone_flag, &ndone, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
/* get the result */
for(ngrp = 1; ngrp < (1 << PTask); ngrp++)
{
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(Send_count[recvTask] > 0 || Recv_count[recvTask] > 0)
{
/* send the results */
MPI_Sendrecv(&PotDataResult[Recv_offset[recvTask]],
Recv_count[recvTask] * sizeof(struct potdata_out),
MPI_BYTE, recvTask, TAG_POTENTIAL_B,
&PotDataOut[Send_offset[recvTask]],
Send_count[recvTask] * sizeof(struct potdata_out),
MPI_BYTE, recvTask, TAG_POTENTIAL_B, MPI_COMM_WORLD, &status);
}
}
}
/* add the results to the local particles */
for(j = 0; j < nexport; j++)
{
place = DataIndexTable[j].Index;
P[place].p.dPotential += PotDataOut[j].Potential;
}
myfree(PotDataOut);
myfree(PotDataResult);
myfree(GravDataGet);
}
while(ndone < NTask);
myfree(DataNodeList);
myfree(DataIndexTable);
/* add correction to exclude self-potential */
for(i = 0; i < NumPart; i++)
{
#ifdef FLTROUNDOFFREDUCTION
P[i].p.Potential = FLT(P[i].p.dPotential);
#endif
/* remove self-potential */
P[i].p.Potential += P[i].Mass / All.SofteningTable[P[i].Type];
if(All.ComovingIntegrationOn)
if(All.PeriodicBoundariesOn)
P[i].p.Potential -= 2.8372975 * pow(P[i].Mass, 2.0 / 3) *
pow(All.Omega0 * 3 * All.Hubble * All.Hubble / (8 * M_PI * All.G), 1.0 / 3);
}
/* multiply with the gravitational constant */
for(i = 0; i < NumPart; i++)
P[i].p.Potential *= All.G;
#ifdef PMGRID
#ifdef PERIODIC
pmpotential_periodic();
#ifdef PLACEHIGHRESREGION
i = pmpotential_nonperiodic(1);
if(i == 1) /* this is returned if a particle lied outside allowed range */
{
pm_init_regionsize();
pm_setup_nonperiodic_kernel();
i = pmpotential_nonperiodic(1); /* try again */
}
if(i == 1)
endrun(88686);
#endif
#else
i = pmpotential_nonperiodic(0);
if(i == 1) /* this is returned if a particle lied outside allowed range */
{
pm_init_regionsize();
pm_setup_nonperiodic_kernel();
i = pmpotential_nonperiodic(0); /* try again */
}
if(i == 1)
endrun(88687);
#ifdef PLACEHIGHRESREGION
i = pmpotential_nonperiodic(1);
if(i == 1) /* this is returned if a particle lied outside allowed range */
{
pm_init_regionsize();
i = pmpotential_nonperiodic(1);
}
if(i != 0)
endrun(88688);
#endif
#endif
#endif
if(All.ComovingIntegrationOn)
{
#ifndef PERIODIC
fac = -0.5 * All.Omega0 * All.Hubble * All.Hubble;
for(i = 0; i < NumPart; i++)
{
for(k = 0, r2 = 0; k < 3; k++)
r2 += P[i].Pos[k] * P[i].Pos[k];
P[i].p.Potential += fac * r2;
}
#endif
}
else
{
fac = -0.5 * All.OmegaLambda * All.Hubble * All.Hubble;
if(fac != 0)
{
for(i = 0; i < NumPart; i++)
{
for(k = 0, r2 = 0; k < 3; k++)
r2 += P[i].Pos[k] * P[i].Pos[k];
P[i].p.Potential += fac * r2;
}
}
}
if(ThisTask == 0)
{
printf("potential done.\n");
fflush(stdout);
}
#else
for(i = 0; i < NumPart; i++)
P[i].Potential = 0;
#endif
CPU_Step[CPU_POTENTIAL] += measure_time();
}
/*! This function computes the gravitational potential for ALL the particles.
* First, the (short-range) tree potential is computed, and then, if needed,
* the long range PM potential is added.
*/
void compute_potential(void)
{
int i;
#ifndef NOGRAVITY
long long ntot, ntotleft;
int j, k, level, sendTask, recvTask;
int ndone;
int maxfill, ngrp, place, nexport;
int *nsend, *noffset, *nsend_local, *nbuffer, *ndonelist, *numlist;
double fac;
double t0, t1, tstart, tend;
#ifndef NOMPI
MPI_Status status;
#endif
double r2;
t0 = second();
if(All.ComovingIntegrationOn)
set_softenings();
if(ThisTask == 0)
{
printf("Start computation of potential for all particles...\n");
fflush(stdout);
}
tstart = second();
if(TreeReconstructFlag)
{
if(ThisTask == 0)
printf("Tree construction.\n");
force_treebuild(NumPart);
TreeReconstructFlag = 0;
if(ThisTask == 0)
printf("Tree construction done.\n");
}
tend = second();
All.CPU_TreeConstruction += timediff(tstart, tend);
numlist = malloc(NTask * sizeof(int) * NTask);
#ifndef NOMPI
MPI_Allgather(&NumPart, 1, MPI_INT, numlist, 1, MPI_INT, GADGET_WORLD);
#else
numlist[0] = NumPart;
#endif
for(i = 0, ntot = 0; i < NTask; i++)
ntot += numlist[i];
free(numlist);
noffset = malloc(sizeof(int) * NTask); /* offsets of bunches in common list */
nbuffer = malloc(sizeof(int) * NTask);
nsend_local = malloc(sizeof(int) * NTask);
nsend = malloc(sizeof(int) * NTask * NTask);
ndonelist = malloc(sizeof(int) * NTask);
i = 0; /* beginn with this index */
ntotleft = ntot; /* particles left for all tasks together */
while(ntotleft > 0)
{
for(j = 0; j < NTask; j++)
nsend_local[j] = 0;
/* do local particles and prepare export list */
for(nexport = 0, ndone = 0; i < NumPart && nexport < All.BunchSizeForce - NTask; i++)
{
ndone++;
for(j = 0; j < NTask; j++)
Exportflag[j] = 0;
#ifndef PMGRID
force_treeevaluate_potential(i, 0);
#else
force_treeevaluate_potential_shortrange(i, 0);
#endif
for(j = 0; j < NTask; j++)
{
if(Exportflag[j])
{
for(k = 0; k < 3; k++)
GravDataGet[nexport].u.Pos[k] = P[i].Pos[k];
#ifdef UNEQUALSOFTENINGS
GravDataGet[nexport].Type = P[i].Type;
#ifdef ADAPTIVE_GRAVSOFT_FORGAS
if(P[i].Type == 0)
GravDataGet[nexport].Soft = SphP[i].Hsml;
#endif
#endif
GravDataGet[nexport].w.OldAcc = P[i].OldAcc;
GravDataIndexTable[nexport].Task = j;
GravDataIndexTable[nexport].Index = i;
GravDataIndexTable[nexport].SortIndex = nexport;
nexport++;
nsend_local[j]++;
}
}
}
qsort(GravDataIndexTable, nexport, sizeof(struct gravdata_index), grav_tree_compare_key);
for(j = 0; j < nexport; j++)
GravDataIn[j] = GravDataGet[GravDataIndexTable[j].SortIndex];
for(j = 1, noffset[0] = 0; j < NTask; j++)
noffset[j] = noffset[j - 1] + nsend_local[j - 1];
#ifndef NOMPI
MPI_Allgather(nsend_local, NTask, MPI_INT, nsend, NTask, MPI_INT, GADGET_WORLD);
#else
nsend[0] = nsend_local[0];
#endif
/* now do the particles that need to be exported */
for(level = 1; level < (1 << PTask); level++)
{
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
#ifndef NOMPI
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* get the particles */
MPI_Sendrecv(&GravDataIn[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_POTENTIAL_A,
&GravDataGet[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_POTENTIAL_A, GADGET_WORLD, &status);
}
}
#endif
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
for(j = 0; j < nbuffer[ThisTask]; j++)
{
#ifndef PMGRID
force_treeevaluate_potential(j, 1);
#else
force_treeevaluate_potential_shortrange(j, 1);
#endif
}
/* get the result */
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
#ifndef NOMPI
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* send the results */
MPI_Sendrecv(&GravDataResult[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_POTENTIAL_B,
&GravDataOut[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_POTENTIAL_B, GADGET_WORLD, &status);
/* add the result to the particles */
for(j = 0; j < nsend_local[recvTask]; j++)
{
place = GravDataIndexTable[noffset[recvTask] + j].Index;
P[place].Potential += GravDataOut[j + noffset[recvTask]].u.Potential;
}
}
}
#endif
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
level = ngrp - 1;
}
#ifndef NOMPI
MPI_Allgather(&ndone, 1, MPI_INT, ndonelist, 1, MPI_INT, GADGET_WORLD);
#else
ndonelist[0] = ndone;
#endif // NOMPI
for(j = 0; j < NTask; j++)
ntotleft -= ndonelist[j];
}
free(ndonelist);
free(nsend);
free(nsend_local);
free(nbuffer);
free(noffset);
/* add correction to exclude self-potential */
for(i = 0; i < NumPart; i++)
{
/* remove self-potential */
P[i].Potential += P[i].Mass / All.SofteningTable[P[i].Type];
if(All.ComovingIntegrationOn)
if(All.PeriodicBoundariesOn)
P[i].Potential -= 2.8372975 * pow(P[i].Mass, 2.0 / 3) *
pow(All.Omega0 * 3 * All.Hubble * All.Hubble / (8 * M_PI * All.G), 1.0 / 3);
}
/* multiply with the gravitational constant */
for(i = 0; i < NumPart; i++)
P[i].Potential *= All.G;
#ifdef PMGRID
#ifdef PERIODIC
pmpotential_periodic();
#ifdef PLACEHIGHRESREGION
i = pmpotential_nonperiodic(1);
if(i == 1) /* this is returned if a particle lied outside allowed range */
{
pm_init_regionsize();
pm_setup_nonperiodic_kernel();
i = pmpotential_nonperiodic(1); /* try again */
}
if(i == 1)
endrun(88686);
#endif
#else
i = pmpotential_nonperiodic(0);
if(i == 1) /* this is returned if a particle lied outside allowed range */
{
pm_init_regionsize();
pm_setup_nonperiodic_kernel();
i = pmpotential_nonperiodic(0); /* try again */
}
if(i == 1)
endrun(88687);
#ifdef PLACEHIGHRESREGION
i = pmpotential_nonperiodic(1);
if(i == 1) /* this is returned if a particle lied outside allowed range */
{
pm_init_regionsize();
i = pmpotential_nonperiodic(1);
}
if(i != 0)
endrun(88688);
#endif
#endif
#endif
if(All.ComovingIntegrationOn)
{
#ifndef PERIODIC
fac = -0.5 * All.Omega0 * All.Hubble * All.Hubble;
for(i = 0; i < NumPart; i++)
{
for(k = 0, r2 = 0; k < 3; k++)
r2 += P[i].Pos[k] * P[i].Pos[k];
P[i].Potential += fac * r2;
}
#endif
}
else
{
fac = -0.5 * All.OmegaLambda * All.Hubble * All.Hubble;
if(fac != 0)
{
for(i = 0; i < NumPart; i++)
{
for(k = 0, r2 = 0; k < 3; k++)
r2 += P[i].Pos[k] * P[i].Pos[k];
P[i].Potential += fac * r2;
}
}
}
if(ThisTask == 0)
{
printf("potential done.\n");
fflush(stdout);
}
t1 = second();
All.CPU_Potential += timediff(t0, t1);
#else
for(i = 0; i < NumPart; i++)
P[i].Potential = 0;
#endif
}
/* This function computes the gravitational potential for ALL the particles.
* It expects that the particles are predicted to the current time.
* The routine constructs a new force-tree.
*/
void compute_potential()
{
#ifdef GRAPE
compute_potential_grape();
return;
#else
double t0,t1;
int i,j;
double r2,fac;
t0=second();
if(All.ComovingIntegrationOn)
set_softenings();
printf("Start potential computation..."); fflush(stdout);
force_treebuild();
All.NumForcesSinceLastTreeConstruction = All.TreeUpdateFrequency*All.TotNumPart ; /* ensures that new tree will be constructed next time*/
for(i=1;i<=NumPart;i++)
{
force_treeevaluate_potential(i-1);
P[i].Potential += P[i].Mass/All.SofteningTable[P[i].Type]; /* removes self energy */
}
if(All.ComovingIntegrationOn)
{
fac=0.5*All.Omega0*All.Hubble*All.Hubble;
for(i=1;i<=NumPart;i++)
{
#ifdef PERIODIC
P[i].Potential = All.G*P[i].Potential;
#else
for(j=0, r2=0; j<3; j++)
r2 += P[i].PosPred[j]*P[i].PosPred[j];
P[i].Potential = All.G*P[i].Potential - fac*r2;
#endif
}
}
else
{
fac= -0.5*All.OmegaLambda*All.Hubble*All.Hubble;
for(i=1;i<=NumPart;i++)
{
P[i].Potential *= All.G;
if(fac!=0)
{
for(j=0,r2=0;j<3;j++)
r2 += P[i].PosPred[j]*P[i].PosPred[j];
P[i].Potential += fac*r2;
}
}
}
printf("done.\n"); fflush(stdout);
t1=second();
All.CPU_Potential+= timediff(t0,t1);
#endif
}
/*! This function computes the gravitational forces for all active
* particles. If needed, a new tree is constructed, otherwise the
* dynamically updated tree is used. Particles are only exported to other
* processors when really needed, thereby allowing a good use of the
* communication buffer.
*/
void gravity_tree(void)
{
int tim=20; // GX mod, timer to profile calls
TimerBeg(29);
TimerBeg(tim);
long long ntot;
int numnodes, nexportsum = 0;
int i, j, iter = 0;
int *numnodeslist, maxnumnodes, nexport, *numlist, *nrecv, *ndonelist;
double tstart, tend, timetree = 0, timecommsumm = 0, timeimbalance = 0, sumimbalance;
double ewaldcount;
double costtotal, ewaldtot, *costtreelist, *ewaldlist;
double maxt, sumt, *timetreelist, *timecommlist;
double fac, plb, plb_max, sumcomm;
#ifndef NOGRAVITY
int *noffset, *nbuffer, *nsend, *nsend_local;
long long ntotleft;
int ndone,maxfill, ngrp;
int k, place;
int level, sendTask, recvTask;
double ax, ay, az;
MPI_Status status;
#endif
///////////////// GX //////////////////////
int totdone=0;
#if CUDA_DEBUG_GX>0
int not_timestepped_gx=0;
int exporthash_gx=0;
int count_exported_gx=0;
#endif
///////////////// GX //////////////////////
/* set new softening lengths */
if(All.ComovingIntegrationOn)
set_softenings();
/* contruct tree if needed */
tstart = second();
if(TreeReconstructFlag)
{
if(ThisTask == 0)
printf("Tree construction.\n");
force_treebuild(NumPart);
TreeReconstructFlag = 0;
if(ThisTask == 0)
printf("Tree construction done.\n");
}
tend = second();
All.CPU_TreeConstruction += timediff(tstart, tend);
costtotal = ewaldcount = 0;
/* Note: 'NumForceUpdate' has already been determined in find_next_sync_point_and_drift() */
numlist = malloc(NTask * sizeof(int) * NTask);
MPI_Allgather(&NumForceUpdate, 1, MPI_INT, numlist, 1, MPI_INT, MPI_COMM_WORLD);
for(i = 0, ntot = 0; i < NTask; i++)
ntot += numlist[i];
free(numlist);
#ifndef NOGRAVITY
if(ThisTask == 0)
printf("Begin tree force.\n");
#ifdef SELECTIVE_NO_GRAVITY
for(i = 0; i < NumPart; i++)
if(((1 << P[i].Type) & (SELECTIVE_NO_GRAVITY)))
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
#endif
noffset = malloc(sizeof(int) * NTask); /* offsets of bunches in common list */
nbuffer = malloc(sizeof(int) * NTask);
nsend_local = malloc(sizeof(int) * NTask);
nsend = malloc(sizeof(int) * NTask * NTask);
ndonelist = malloc(sizeof(int) * NTask);
i = 0; /* begin with this index */
ntotleft = ntot; /* particles left for all tasks together */
TimerEnd(tim++);
///////////////// GX //////////////////////
// if (s_gx.cudamode>0 && All.MaxPart>1400000) TimersSleep(10); // GPU card runs hot on large sims, this is around N_p=1404928
// if (s_gx.cudamode>0) TimersSleep(10);
TimerBeg(tim);
double starttime,subtime=-1,cpytime=-1;
int Np=-1;
int buffered=0;
if(s_gx.cudamode>0)
{
FUN_MESSAGE(2,"gravity_tree()");
TimerBeg(50);
cpytime=GetTime();
Np=InitializeProlog_gx(NumPart);
TimerEnd(50);
cpytime=GetTime()-cpytime;
}
///////////////// GX //////////////////////
while(ntotleft > 0)
{
TimerBeg(31);
starttime=GetTime();
iter++;
for(j = 0; j < NTask; j++)
nsend_local[j] = 0;
/* do local particles and prepare export list */
tstart = second();
if (s_gx.cudamode==0 || Np<MIN_FORCE_PARTICLES_FOR_GPU_GX) {
ASSERT_GX( !buffered );
ReLaunchChunkManager();
for(nexport = 0, ndone = 0; i < NumPart && nexport < All.BunchSizeForce - NTask; i++) {
if(P[i].Ti_endstep == All.Ti_Current)
{
ndone++;
for(j = 0; j < NTask; j++)
Exportflag[j] = 0;
TimerUpdateCounter(31,1);
#ifndef PMGRID
costtotal += force_treeevaluate(i, 0, &ewaldcount);
#else
costtotal += force_treeevaluate_shortrange(i, 0 );
#endif
#if CUDA_DEBUG_GX>0
int flagexported_gx=0;
#endif
for(j = 0; j < NTask; j++)
{
if(Exportflag[j])
{
ASSERT_GX( NTask>1 );
#if CUDA_DEBUG_GX>0
flagexported_gx=1;
exporthash_gx += (i-j)*(j+ThisTask+1);
#endif
for(k = 0; k < 3; k++)
GravDataGet[nexport].u.Pos[k] = P[i].Pos[k];
#ifdef UNEQUALSOFTENINGS
GravDataGet[nexport].Type = P[i].Type;
#ifdef ADAPTIVE_GRAVSOFT_FORGAS
if(P[i].Type == 0)
GravDataGet[nexport].Soft = SphP[i].Hsml;
#endif
#endif
GravDataGet[nexport].w.OldAcc = P[i].OldAcc;
GravDataIndexTable[nexport].Task = j;
GravDataIndexTable[nexport].Index = i;
GravDataIndexTable[nexport].SortIndex = nexport;
nexport++;
nexportsum++;
nsend_local[j]++;
}
}
#if CUDA_DEBUG_GX>0
if (flagexported_gx) ++count_exported_gx;
#endif
}
#if CUDA_DEBUG_GX>0
else ++not_timestepped_gx;
#endif
}
ManageChuncks(0);
} else {
///////////////// GX //////////////////////
// cudamode>0
///////////////// GX //////////////////////
#ifndef PMGRID
// WARNING Attemping to run in tree-only mode, examine results carefully
// ERROR cannot run in non PMGRID mode
#endif
if (iter==1){
const double tx=GetTime();
TimerBeg(51);
ASSERT_GX(NumPart>=i);
ASSERT_GX(!buffered);
if (iter!=1) ERROR("cuda mode does not support iterations in gravtree calc, try to increasing the 'BufferSize' in the parameter file to surcomevent this problem");
const int Np2=InitializeCalculation_gx(NumPart,P,0);
ASSERT_GX( Np2==Np );
if (Np2==0) WARNING("no particles participate in this timestep");
TimerEnd(51);
cpytime += GetTime() - tx;
subtime=GetTime();
TimerBeg(52);
force_treeevaluate_shortrange_range_gx(0, Np);
buffered=1;
TimerUpdateCounter(31,NumPart-i);
TimerEnd(52);
subtime = GetTime() - subtime;
} else {
cpytime=-1;
subtime=-1;
ASSERT_GX(buffered);
}
for(nexport = 0, ndone = 0; i < NumPart && nexport < All.BunchSizeForce - NTask; i++) {
if(P[i].Ti_endstep == All.Ti_Current)
{
ndone++;
ASSERT_GX( i<NumPart );
ASSERT_GX( buffered );
const struct result_gx r=GetTarget(totdone++,i); // s_gx.result[target];
P[i].GravAccel[0] = r.acc_x;
P[i].GravAccel[1] = r.acc_y;
P[i].GravAccel[2] = r.acc_z;
P[i].GravCost = r.ninteractions;
costtotal += r.ninteractions;
if (s_gx.NTask>1) {
#if CUDA_DEBUG_GX>0
int flagexported_gx=0;
#endif
for(j = 0; j < NTask; j++) {
if (GetExportflag_gx(&s_gx,i,NTask,j)){
ASSERT_GX( NTask>1 );
#if CUDA_DEBUG_GX>0
flagexported_gx=1;
exporthash_gx += (i-j)*(j+ThisTask+1);
#endif
for(k = 0; k < 3; k++) GravDataGet[nexport].u.Pos[k] = P[i].Pos[k];
#ifdef UNEQUALSOFTENINGS
GravDataGet[nexport].Type = P[i].Type;
#ifdef ADAPTIVE_GRAVSOFT_FORGAS
if(P[i].Type == 0) GravDataGet[nexport].Soft = SphP[i].Hsml;
#endif
#endif
GravDataGet[nexport].w.OldAcc = P[i].OldAcc;
GravDataIndexTable[nexport].Task = j;
GravDataIndexTable[nexport].Index = i;
GravDataIndexTable[nexport].SortIndex = nexport;
nexport++;
nexportsum++;
nsend_local[j]++;
}
}
#if CUDA_DEBUG_GX>0
if (flagexported_gx) ++count_exported_gx;
#endif
}
}
#if CUDA_DEBUG_GX>0
else ++not_timestepped_gx;
#endif
}
AssertsOnhasGadgetDataBeenModified_gx(0,1,0);
}
TimerEnd(31);
///////////////// GX //////////////////////
if (iter==1 || !buffered){
PrintInfoFinalize(s_gx,ndone,Np,starttime,cpytime,subtime,0,iter,-1
#if CUDA_DEBUG_GX>0
,not_timestepped_gx,count_exported_gx,nexport,nexportsum,exporthash_gx,costtotal
#else
,0,0,0,0,0,0
#endif
);
subtime=-1;
}
TimerBeg(39);
///////////////// GX //////////////////////
tend = second();
timetree += timediff(tstart, tend);
qsort(GravDataIndexTable, nexport, sizeof(struct gravdata_index), grav_tree_compare_key);
for(j = 0; j < nexport; j++)
GravDataIn[j] = GravDataGet[GravDataIndexTable[j].SortIndex];
for(j = 1, noffset[0] = 0; j < NTask; j++)
noffset[j] = noffset[j - 1] + nsend_local[j - 1];
tstart = second();
MPI_Allgather(nsend_local, NTask, MPI_INT, nsend, NTask, MPI_INT, MPI_COMM_WORLD);
tend = second();
timeimbalance += timediff(tstart, tend);
/* now do the particles that need to be exported */
for(level = 1; level < (1 << PTask); level++)
{
tstart = second();
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* get the particles */
MPI_Sendrecv(&GravDataIn[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_GRAV_A,
&GravDataGet[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_GRAV_A, MPI_COMM_WORLD, &status);
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tend = second();
timecommsumm += timediff(tstart, tend);
TimerBeg(30);
TimerUpdateCounter(30,nbuffer[ThisTask]);
tstart = second();
///////////////// GX //////////////////////
// Do exported particles on the CPU/GPU
{
AssertsOnhasGadgetDataBeenModified_gx(1,1,0);
#if CUDA_DEBUG_GX>1
MESSAGE("INFO: DistRMSGrav=%g",DistRMSGravdata(nbuffer[ThisTask],GravDataGet));
#endif
starttime=GetTime();
const int N=nbuffer[ThisTask];
if (N>0){
if (s_gx.cudamode==0 || N<MIN_FORCE_PARTICLES_FOR_GPU_GX || Np<MIN_FORCE_PARTICLES_FOR_GPU_GX) {
ReLaunchChunkManager();
for(j = 0; j<N ; j++)
{
#ifndef PMGRID
costtotal += force_treeevaluate(j, 1, &ewaldcount);
#else
costtotal += force_treeevaluate_shortrange(j, 1);
#endif
}
ManageChuncks(0);
} else {
ASSERT_GX( buffered );
cpytime=GetTime();
InitializeExportCalculation_gx(N,P[0].Type);
ASSERT_GX( N==s_gx.Np );
subtime=GetTime();
force_treeevaluate_shortrange_range_gx(1, N);
subtime=GetTime()-subtime;
costtotal += FinalizeExportCalculation_gx(N);
cpytime=GetTime()-cpytime-subtime;
ASSERT_GX( N==s_gx.Np );
}
PrintInfoFinalize(s_gx,0,N,starttime,cpytime,subtime,2,iter,level,0,0,nexport,0,0,0);
subtime=-1;
} else {
ReLaunchChunkManager();
ManageChuncks(0);
}
}
///////////////// GX //////////////////////
if (nbuffer[ThisTask]>0) TimerUpdateCounter(30,-1);
TimerEnd(30);
tend = second();
timetree += timediff(tstart, tend);
TimerBeg(33);
tstart = second();
MPI_Barrier(MPI_COMM_WORLD);
tend = second();
timeimbalance += timediff(tstart, tend);
TimerEnd(33);
/* get the result */
tstart = second();
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* send the results */
MPI_Sendrecv(&GravDataResult[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_GRAV_B,
&GravDataOut[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_GRAV_B, MPI_COMM_WORLD, &status);
/* add the result to the particles */
for(j = 0; j < nsend_local[recvTask]; j++)
{
place = GravDataIndexTable[noffset[recvTask] + j].Index;
// comment out in order to disable export forces for debugging
for(k = 0; k < 3; k++)
P[place].GravAccel[k] += GravDataOut[j + noffset[recvTask]].u.Acc[k];
P[place].GravCost += GravDataOut[j + noffset[recvTask]].w.Ninteractions;
}
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tend = second();
timecommsumm += timediff(tstart, tend);
level = ngrp - 1;
}
MPI_Allgather(&ndone, 1, MPI_INT, ndonelist, 1, MPI_INT, MPI_COMM_WORLD);
for(j = 0; j < NTask; j++)
ntotleft -= ndonelist[j];
TimerEnd(39);
}
TimerEnd(tim++);
TimerBeg(tim);
free(ndonelist);
free(nsend);
free(nsend_local);
free(nbuffer);
free(noffset);
/* now add things for comoving integration */
#ifndef PERIODIC
#ifndef PMGRID
if(All.ComovingIntegrationOn)
{
fac = 0.5 * All.Hubble * All.Hubble * All.Omega0 / All.G;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] += fac * P[i].Pos[j];
}
#endif
#endif
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
{
#ifdef PMGRID
ax = P[i].GravAccel[0] + P[i].GravPM[0] / All.G;
ay = P[i].GravAccel[1] + P[i].GravPM[1] / All.G;
az = P[i].GravAccel[2] + P[i].GravPM[2] / All.G;
#else
ax = P[i].GravAccel[0];
ay = P[i].GravAccel[1];
az = P[i].GravAccel[2];
#endif
P[i].OldAcc = sqrt(ax * ax + ay * ay + az * az);
}
if(All.TypeOfOpeningCriterion == 1)
All.ErrTolTheta = 0; /* This will switch to the relative opening criterion for the following force computations */
/* muliply by G */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] *= All.G;
/* Finally, the following factor allows a computation of a cosmological simulation
with vacuum energy in physical coordinates */
#ifndef PERIODIC
#ifndef PMGRID
if(All.ComovingIntegrationOn == 0)
{
fac = All.OmegaLambda * All.Hubble * All.Hubble;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] += fac * P[i].Pos[j];
}
#endif
#endif
#ifdef SELECTIVE_NO_GRAVITY
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
#endif
if(ThisTask == 0)
printf("tree is done.\n");
#else /* gravity is switched off */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep == All.Ti_Current)
for(j = 0; j < 3; j++)
P[i].GravAccel[j] = 0;
#endif
/* Now the force computation is finished */
/* gather some diagnostic information */
timetreelist = malloc(sizeof(double) * NTask);
timecommlist = malloc(sizeof(double) * NTask);
costtreelist = malloc(sizeof(double) * NTask);
numnodeslist = malloc(sizeof(int) * NTask);
ewaldlist = malloc(sizeof(double) * NTask);
nrecv = malloc(sizeof(int) * NTask);
numnodes = Numnodestree;
MPI_Gather(&costtotal, 1, MPI_DOUBLE, costtreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&numnodes, 1, MPI_INT, numnodeslist, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&timetree, 1, MPI_DOUBLE, timetreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&timecommsumm, 1, MPI_DOUBLE, timecommlist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&NumPart, 1, MPI_INT, nrecv, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&ewaldcount, 1, MPI_DOUBLE, ewaldlist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Reduce(&nexportsum, &nexport, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(&timeimbalance, &sumimbalance, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if(ThisTask == 0)
{
All.TotNumOfForces += ntot;
fprintf(FdTimings, "Step= %d t= %g dt= %g \n", All.NumCurrentTiStep, All.Time, All.TimeStep);
fprintf(FdTimings, "Nf= %d%09d total-Nf= %d%09d ex-frac= %g iter= %d\n",
(int) (ntot / 1000000000), (int) (ntot % 1000000000),
(int) (All.TotNumOfForces / 1000000000), (int) (All.TotNumOfForces % 1000000000),
nexport / ((double) ntot), iter);
/* note: on Linux, the 8-byte integer could be printed with the format identifier "%qd", but doesn't work on AIX */
fac = NTask / ((double) All.TotNumPart);
for(i = 0, maxt = timetreelist[0], sumt = 0, plb_max = 0,
maxnumnodes = 0, costtotal = 0, sumcomm = 0, ewaldtot = 0; i < NTask; i++)
{
costtotal += costtreelist[i];
sumcomm += timecommlist[i];
if(maxt < timetreelist[i])
maxt = timetreelist[i];
sumt += timetreelist[i];
plb = nrecv[i] * fac;
if(plb > plb_max)
plb_max = plb;
if(numnodeslist[i] > maxnumnodes)
maxnumnodes = numnodeslist[i];
ewaldtot += ewaldlist[i];
}
fprintf(FdTimings, "work-load balance: %g max=%g avg=%g PE0=%g\n",
maxt / (sumt / NTask), maxt, sumt / NTask, timetreelist[0]);
fprintf(FdTimings, "particle-load balance: %g\n", plb_max);
fprintf(FdTimings, "max. nodes: %d, filled: %g\n", maxnumnodes,
maxnumnodes / (All.TreeAllocFactor * All.MaxPart));
fprintf(FdTimings, "part/sec=%g | %g ia/part=%g (%g)\n", ntot / (sumt + 1.0e-20),
ntot / (maxt * NTask), ((double) (costtotal)) / ntot, ((double) ewaldtot) / ntot);
fprintf(FdTimings, "\n");
fflush(FdTimings);
All.CPU_TreeWalk += sumt / NTask;
All.CPU_Imbalance += sumimbalance / NTask;
All.CPU_CommSum += sumcomm / NTask;
}
free(nrecv);
free(ewaldlist);
free(numnodeslist);
free(costtreelist);
free(timecommlist);
free(timetreelist);
ASSERT_GX( tim==22 );
TimerEnd(tim++);
TimerEnd(29);
//MESSAGE("%6.2f, %6.2f, %6.2f, %6.2f, %6.2f - %5.1f, %5.1f, %5.1f, %5.1f %c force timers d 29,31,30,33,net",TimerGet(29),TimerGet(31),TimerGet(30),TimerGet(33),TimerGet(29)-TimerGet(31)-TimerGet(30),100.0*TimerGet(31)/TimerGet(29),100.0*TimerGet(30)/TimerGet(29),100.0*TimerGet(33)/TimerGet(29),100.0*(TimerGet(29)-TimerGet(31)-TimerGet(30))/TimerGet(29),'%');
//MESSAGE("%6.2f, %6.2f, %6.2f, %6.2f, %6.2f - %5.1f, %5.1f, %5.1f, %5.1f %c force timers a 29,31,30,33,net",TimerGetAccumulated(29),TimerGetAccumulated(31),TimerGetAccumulated(30),TimerGetAccumulated(33),TimerGetAccumulated(29)-TimerGetAccumulated(31)-TimerGetAccumulated(30),100.0*TimerGetAccumulated(31)/TimerGetAccumulated(29),100.0*TimerGetAccumulated(30)/TimerGetAccumulated(29),100.0*TimerGetAccumulated(33)/TimerGetAccumulated(29),100.0*(TimerGetAccumulated(29)-TimerGetAccumulated(31)-TimerGetAccumulated(30))/TimerGetAccumulated(29),'%');
}
/*! This routine does the test of the gravitational tree force by computing
* the force for a random subset of particles with direct summation.
*/
void gravity_forcetest(void)
{
int ntot, iter = 0, ntotleft, nthis;
double tstart, tend, timetree = 0;
int i, j, ndone, ngrp, maxfill, place, ndonetot;
#ifndef NOGRAVITY
int *noffset, *nbuffer, *nsend, *nsend_local;
int k, nexport;
int level, sendTask, recvTask;
double fac1;
#ifndef NOMPI
MPI_Status status;
#endif
#endif
double costtotal, *costtreelist;
double maxt, sumt, *timetreelist;
double fac;
char buf[200];
#ifdef PMGRID
if(All.PM_Ti_endstep != All.Ti_Current)
return;
#endif
if(All.ComovingIntegrationOn)
set_softenings(); /* set new softening lengths */
for(i = 0, NumForceUpdate = 0; i < NumPart; i++)
{
if(P[i].Ti_endstep == All.Ti_Current)
{
if(get_random_number(P[i].ID) < FORCETEST)
{
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
NumForceUpdate++;
}
}
}
/* NumForceUpdate is the number of particles on this processor that want a force update */
#ifndef NOMPI
MPI_Allreduce(&NumForceUpdate, &ntot, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
#else
ntot = NumForceUpdate;
#endif
costtotal = 0;
noffset = malloc(sizeof(int) * NTask); /* offsets of bunches in common list */
nbuffer = malloc(sizeof(int) * NTask);
nsend_local = malloc(sizeof(int) * NTask);
nsend = malloc(sizeof(int) * NTask * NTask);
i = 0; /* beginn with this index */
ntotleft = ntot; /* particles left for all tasks together */
while(ntotleft > 0)
{
iter++;
for(j = 0; j < NTask; j++)
nsend_local[j] = 0;
/* do local particles and prepare export list */
tstart = second();
for(nexport = 0, ndone = 0; i < NumPart && nexport < All.BunchSizeForce - NTask; i++)
if(P[i].Ti_endstep < 0)
{
ndone++;
for(j = 0; j < NTask; j++)
Exportflag[j] = 1;
Exportflag[ThisTask] = 0;
costtotal += force_treeevaluate_direct(i, 0);
for(j = 0; j < NTask; j++)
{
if(Exportflag[j])
{
for(k = 0; k < 3; k++)
GravDataGet[nexport].u.Pos[k] = P[i].Pos[k];
#ifdef UNEQUALSOFTENINGS
GravDataGet[nexport].Type = P[i].Type;
#endif
GravDataGet[nexport].w.OldAcc = P[i].OldAcc;
GravDataIndexTable[nexport].Task = j;
GravDataIndexTable[nexport].Index = i;
GravDataIndexTable[nexport].SortIndex = nexport;
nexport++;
nsend_local[j]++;
}
}
}
tend = second();
timetree += timediff(tstart, tend);
qsort(GravDataIndexTable, nexport, sizeof(struct gravdata_index), grav_tree_compare_key);
for(j = 0; j < nexport; j++)
GravDataIn[j] = GravDataGet[GravDataIndexTable[j].SortIndex];
for(j = 1, noffset[0] = 0; j < NTask; j++)
noffset[j] = noffset[j - 1] + nsend_local[j - 1];
#ifndef NOMPI
MPI_Allgather(nsend_local, NTask, MPI_INT, nsend, NTask, MPI_INT, MPI_COMM_WORLD);
#else
nsend[0] = nsend_local[i];
#endif
/* now do the particles that need to be exported */
for(level = 1; level < (1 << PTask); level++)
{
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* get the particles */
MPI_Sendrecv(&GravDataIn[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_DIRECT_A,
&GravDataGet[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in), MPI_BYTE,
recvTask, TAG_DIRECT_A, MPI_COMM_WORLD, &status);
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
tstart = second();
for(j = 0; j < nbuffer[ThisTask]; j++)
{
costtotal += force_treeevaluate_direct(j, 1);
}
tend = second();
timetree += timediff(tstart, tend);
/* get the result */
for(j = 0; j < NTask; j++)
nbuffer[j] = 0;
for(ngrp = level; ngrp < (1 << PTask); ngrp++)
{
maxfill = 0;
for(j = 0; j < NTask; j++)
{
if((j ^ ngrp) < NTask)
if(maxfill < nbuffer[j] + nsend[(j ^ ngrp) * NTask + j])
maxfill = nbuffer[j] + nsend[(j ^ ngrp) * NTask + j];
}
if(maxfill >= All.BunchSizeForce)
break;
sendTask = ThisTask;
recvTask = ThisTask ^ ngrp;
if(recvTask < NTask)
{
if(nsend[ThisTask * NTask + recvTask] > 0 || nsend[recvTask * NTask + ThisTask] > 0)
{
/* send the results */
MPI_Sendrecv(&GravDataResult[nbuffer[ThisTask]],
nsend[recvTask * NTask + ThisTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_DIRECT_B,
&GravDataOut[noffset[recvTask]],
nsend_local[recvTask] * sizeof(struct gravdata_in),
MPI_BYTE, recvTask, TAG_DIRECT_B, MPI_COMM_WORLD, &status);
/* add the result to the particles */
for(j = 0; j < nsend_local[recvTask]; j++)
{
place = GravDataIndexTable[noffset[recvTask] + j].Index;
for(k = 0; k < 3; k++)
P[place].GravAccelDirect[k] += GravDataOut[j + noffset[recvTask]].u.Acc[k];
}
}
}
for(j = 0; j < NTask; j++)
if((j ^ ngrp) < NTask)
nbuffer[j] += nsend[(j ^ ngrp) * NTask + j];
}
level = ngrp - 1;
}
MPI_Allreduce(&ndone, &ndonetot, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
ntotleft -= ndonetot;
}
free(nsend);
free(nsend_local);
free(nbuffer);
free(noffset);
/* now add things for comoving integration */
if(All.ComovingIntegrationOn)
{
#ifndef PERIODIC
fac1 = 0.5 * All.Hubble * All.Hubble * All.Omega0 / All.G;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
for(j = 0; j < 3; j++)
P[i].GravAccelDirect[j] += fac1 * P[i].Pos[j];
#endif
}
/* muliply by G */
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
for(j = 0; j < 3; j++)
P[i].GravAccelDirect[j] *= All.G;
/* Finally, the following factor allows a computation of cosmological simulation
with vacuum energy in physical coordinates */
if(All.ComovingIntegrationOn == 0)
{
fac1 = All.OmegaLambda * All.Hubble * All.Hubble;
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
for(j = 0; j < 3; j++)
P[i].GravAccelDirect[j] += fac1 * P[i].Pos[j];
}
/* now output the forces to a file */
for(nthis = 0; nthis < NTask; nthis++)
{
if(nthis == ThisTask)
{
sprintf(buf, "%s%s", All.OutputDir, "forcetest.txt");
if(!(FdForceTest = fopen(buf, "a")))
{
printf("error in opening file '%s'\n", buf);
endrun(17);
}
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
{
#ifndef PMGRID
fprintf(FdForceTest, "%d %g %g %g %g %g %g %g %g %g %g %g\n",
P[i].Type, All.Time, All.Time - TimeOfLastTreeConstruction,
P[i].Pos[0], P[i].Pos[1], P[i].Pos[2],
P[i].GravAccelDirect[0], P[i].GravAccelDirect[1], P[i].GravAccelDirect[2],
P[i].GravAccel[0], P[i].GravAccel[1], P[i].GravAccel[2]);
#else
fprintf(FdForceTest, "%d %g %g %g %g %g %g %g %g %g %g %g %g %g %g\n",
P[i].Type, All.Time, All.Time - TimeOfLastTreeConstruction,
P[i].Pos[0], P[i].Pos[1], P[i].Pos[2],
P[i].GravAccelDirect[0], P[i].GravAccelDirect[1], P[i].GravAccelDirect[2],
P[i].GravAccel[0], P[i].GravAccel[1], P[i].GravAccel[2],
P[i].GravPM[0] + P[i].GravAccel[0],
P[i].GravPM[1] + P[i].GravAccel[1], P[i].GravPM[2] + P[i].GravAccel[2]);
#endif
}
fclose(FdForceTest);
}
#ifndef NOMPI
MPI_Barrier(MPI_COMM_WORLD);
#endif
}
for(i = 0; i < NumPart; i++)
if(P[i].Ti_endstep < 0)
P[i].Ti_endstep = -P[i].Ti_endstep - 1;
/* Now the force computation is finished */
timetreelist = malloc(sizeof(double) * NTask);
costtreelist = malloc(sizeof(double) * NTask);
MPI_Gather(&costtotal, 1, MPI_DOUBLE, costtreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Gather(&timetree, 1, MPI_DOUBLE, timetreelist, 1, MPI_DOUBLE, 0, MPI_COMM_WORLD);
if(ThisTask == 0)
{
fac = NTask / ((double) All.TotNumPart);
for(i = 0, maxt = timetreelist[0], sumt = 0, costtotal = 0; i < NTask; i++)
{
costtotal += costtreelist[i];
if(maxt < timetreelist[i])
maxt = timetreelist[i];
sumt += timetreelist[i];
}
fprintf(FdTimings, "DIRECT Nf= %d part/sec=%g | %g ia/part=%g \n", ntot, ntot / (sumt + 1.0e-20),
ntot / (maxt * NTask), ((double) (costtotal)) / ntot);
fprintf(FdTimings, "\n");
fflush(FdTimings);
}
free(costtreelist);
free(timetreelist);
}
