/*! Calculates the long-range non-periodic potential using the PM method. The
* potential is Gaussian filtered with Asmth, given in mesh-cell units. We
* carry out a CIC charge assignment, and compute the potenial by Fourier
* transform methods. The CIC kernel is deconvolved.
*/
int pmpotential_nonperiodic(int grnr)
{
double dx, dy, dz;
double fac, to_slab_fac;
double re, im, pot;
int i, j, slab, level, sendTask, recvTask, flag, flagsum;
int x, y, z, ip;
int slab_x, slab_y, slab_z;
int slab_xx, slab_yy, slab_zz;
int meshmin[3], meshmax[3], sendmin, sendmax, recvmin, recvmax;
int dimx, dimy, dimz, recv_dimx, recv_dimy, recv_dimz;
MPI_Status status;
if(ThisTask == 0)
printf("Starting non-periodic PM-potential calculation.\n");
fac = All.G / pow(All.TotalMeshSize[grnr], 4) * pow(All.TotalMeshSize[grnr] / GRID, 3); /* to get potential */
to_slab_fac = GRID / All.TotalMeshSize[grnr];
/* first, establish the extension of the local patch in GRID (for binning) */
for(j = 0; j < 3; j++)
{
meshmin[j] = GRID;
meshmax[j] = 0;
}
for(i = 0, flag = 0; i < NumPart; i++)
{
if (P[i].Type == 0 && P[i].ID < 0) /*SINK*/
continue;
#ifdef PLACEHIGHRESREGION
if(grnr == 0 || (grnr == 1 && P[i].Type == 0 && P[i].Mass*1.0e10/All.HubbleParam < All.RefinementMass || grnr == 1 && P[i].Type == 1 && P[i].Mass*1.0e10/All.HubbleParam < (All.Omega0/All.OmegaBaryon-1.0)*All.RefinementMass))
#endif
{
for(j = 0; j < 3; j++)
{
if(P[i].Pos[j] < All.Xmintot[grnr][j] || P[i].Pos[j] > All.Xmaxtot[grnr][j])
{
if(flag == 0)
{
printf
("Particle Id=%d on task=%d with coordinates (%g|%g|%g) lies outside PM mesh.\nStopping\n",
(int)P[i].ID, ThisTask, P[i].Pos[0], P[i].Pos[1], P[i].Pos[2]);
fflush(stdout);
}
flag++;
break;
}
}
}
if(flag > 0)
continue;
if(P[i].Pos[0] >= All.Corner[grnr][0] && P[i].Pos[0] < All.UpperCorner[grnr][0])
if(P[i].Pos[1] >= All.Corner[grnr][1] && P[i].Pos[1] < All.UpperCorner[grnr][1])
if(P[i].Pos[2] >= All.Corner[grnr][2] && P[i].Pos[2] < All.UpperCorner[grnr][2])
{
for(j = 0; j < 3; j++)
{
slab = to_slab_fac * (P[i].Pos[j] - All.Corner[grnr][j]);
if(slab < meshmin[j])
meshmin[j] = slab;
if(slab > meshmax[j])
meshmax[j] = slab;
}
}
}
MPI_Allreduce(&flag, &flagsum, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
if(flagsum > 0)
{
if(ThisTask == 0)
{
printf("In total %d particles were outside allowed range.\n", flagsum);
fflush(stdout);
}
return 1; /* error - need to return because particle were outside allowed range */
}
MPI_Allgather(meshmin, 3, MPI_INT, meshmin_list, 3, MPI_INT, MPI_COMM_WORLD);
MPI_Allgather(meshmax, 3, MPI_INT, meshmax_list, 3, MPI_INT, MPI_COMM_WORLD);
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
force_treefree();
pm_init_nonperiodic_allocate((dimx + 4) * (dimy + 4) * (dimz + 4));
for(i = 0; i < dimx * dimy * dimz; i++)
workspace[i] = 0;
for(i = 0; i < NumPart; i++)
{
if (P[i].Type == 0 && P[i].ID < 0) /*SINK*/
continue;
if(P[i].Pos[0] < All.Corner[grnr][0] || P[i].Pos[0] >= All.UpperCorner[grnr][0])
continue;
if(P[i].Pos[1] < All.Corner[grnr][1] || P[i].Pos[1] >= All.UpperCorner[grnr][1])
continue;
if(P[i].Pos[2] < All.Corner[grnr][2] || P[i].Pos[2] >= All.UpperCorner[grnr][2])
continue;
slab_x = to_slab_fac * (P[i].Pos[0] - All.Corner[grnr][0]);
dx = to_slab_fac * (P[i].Pos[0] - All.Corner[grnr][0]) - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * (P[i].Pos[1] - All.Corner[grnr][1]);
dy = to_slab_fac * (P[i].Pos[1] - All.Corner[grnr][1]) - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * (P[i].Pos[2] - All.Corner[grnr][2]);
dz = to_slab_fac * (P[i].Pos[2] - All.Corner[grnr][2]) - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
workspace[(slab_x * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_x * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * dy * (1.0 - dz);
workspace[(slab_x * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * dz;
workspace[(slab_x * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * dy * dz;
workspace[(slab_xx * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (dx) * dy * (1.0 - dz);
workspace[(slab_xx * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (dx) * (1.0 - dy) * dz;
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (dx) * dy * dz;
}
for(i = 0; i < fftsize; i++) /* clear local density field */
rhogrid[i] = 0;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * GRID;
sendmax = -1;
for(slab_x = meshmin[0]; slab_x < meshmax[0] + 2; slab_x++)
if(slab_to_task[slab_x] == recvTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -1)
sendmin = 0;
/* check how much we have to receive */
recvmin = 2 * GRID;
recvmax = -1;
for(slab_x = meshmin_list[3 * recvTask]; slab_x < meshmax_list[3 * recvTask] + 2; slab_x++)
if(slab_to_task[slab_x] == sendTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -1)
recvmin = 0;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 2;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 2;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 2;
if(level > 0)
{
MPI_Sendrecv(workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE, recvTask,
TAG_NONPERIOD_C, forcegrid,
(recvmax - recvmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_NONPERIOD_C, MPI_COMM_WORLD, &status);
}
else
{
memcpy(forcegrid, workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real));
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
slab_xx = slab_x - first_slab_of_task[ThisTask];
if(slab_xx >= 0 && slab_xx < slabs_per_task[ThisTask])
{
for(slab_y = meshmin_list[3 * recvTask + 1];
slab_y <= meshmax_list[3 * recvTask + 1] + 1; slab_y++)
{
slab_yy = slab_y;
for(slab_z = meshmin_list[3 * recvTask + 2];
slab_z <= meshmax_list[3 * recvTask + 2] + 1; slab_z++)
{
slab_zz = slab_z;
rhogrid[GRID * GRID2 * slab_xx + GRID2 * slab_yy + slab_zz] +=
forcegrid[((slab_x - recvmin) * recv_dimy +
(slab_y - meshmin_list[3 * recvTask + 1])) * recv_dimz +
(slab_z - meshmin_list[3 * recvTask + 2])];
}
}
}
}
}
}
}
/* Do the FFT of the density field */
rfftwnd_mpi(fft_forward_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
/* multiply with the Fourier transform of the Green's function (kernel) */
for(y = 0; y < nslab_y; y++)
for(x = 0; x < GRID; x++)
for(z = 0; z < GRID / 2 + 1; z++)
{
ip = GRID * (GRID / 2 + 1) * y + (GRID / 2 + 1) * x + z;
re =
fft_of_rhogrid[ip].re * fft_of_kernel[grnr][ip].re -
fft_of_rhogrid[ip].im * fft_of_kernel[grnr][ip].im;
im =
fft_of_rhogrid[ip].re * fft_of_kernel[grnr][ip].im +
fft_of_rhogrid[ip].im * fft_of_kernel[grnr][ip].re;
fft_of_rhogrid[ip].re = fac * re;
fft_of_rhogrid[ip].im = fac * im;
}
/* get the potential by inverse FFT */
rfftwnd_mpi(fft_inverse_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
/* Now rhogrid holds the potential */
/* construct the potential for the local patch */
/* if we have a high-res mesh, establish the extension of the local patch in GRID (for reading out the
* forces)
*/
#ifdef PLACEHIGHRESREGION
if(grnr == 1)
{
for(j = 0; j < 3; j++)
{
meshmin[j] = GRID;
meshmax[j] = 0;
}
for(i = 0; i < NumPart; i++)
{
if (P[i].Type == 0 && P[i].ID < 0) /*SINK*/
continue;
if(!(P[i].Type == 0 && P[i].Mass*1.0e10/All.HubbleParam < All.RefinementMass || P[i].Type == 1 && P[i].Mass*1.0e10/All.HubbleParam < (All.Omega0/All.OmegaBaryon-1.0)*All.RefinementMass))
continue;
if(P[i].Pos[0] >= All.Corner[grnr][0] && P[i].Pos[0] < All.UpperCorner[grnr][0])
if(P[i].Pos[1] >= All.Corner[grnr][1] && P[i].Pos[1] < All.UpperCorner[grnr][1])
if(P[i].Pos[2] >= All.Corner[grnr][2] && P[i].Pos[2] < All.UpperCorner[grnr][2])
{
for(j = 0; j < 3; j++)
{
slab = to_slab_fac * (P[i].Pos[j] - All.Corner[grnr][j]);
if(slab < meshmin[j])
meshmin[j] = slab;
if(slab > meshmax[j])
meshmax[j] = slab;
}
}
}
MPI_Allgather(meshmin, 3, MPI_INT, meshmin_list, 3, MPI_INT, MPI_COMM_WORLD);
MPI_Allgather(meshmax, 3, MPI_INT, meshmax_list, 3, MPI_INT, MPI_COMM_WORLD);
}
#endif
dimx = meshmax[0] - meshmin[0] + 6;
dimy = meshmax[1] - meshmin[1] + 6;
dimz = meshmax[2] - meshmin[2] + 6;
for(j = 0; j < 3; j++)
{
if(meshmin[j] < 2)
endrun(131231);
if(meshmax[j] > GRID / 2 - 3)
endrun(131288);
}
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * GRID;
sendmax = -GRID;
for(slab_x = meshmin_list[3 * recvTask] - 2; slab_x < meshmax_list[3 * recvTask] + 4; slab_x++)
if(slab_to_task[slab_x] == sendTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -GRID)
sendmin = sendmax + 1;
/* check how much we have to receive */
recvmin = 2 * GRID;
recvmax = -GRID;
for(slab_x = meshmin[0] - 2; slab_x < meshmax[0] + 4; slab_x++)
if(slab_to_task[slab_x] == recvTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -GRID)
recvmin = recvmax + 1;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 6;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 6;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 6;
/* prepare what we want to send */
if(sendmax - sendmin >= 0)
{
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
slab_xx = slab_x - first_slab_of_task[ThisTask];
for(slab_y = meshmin_list[3 * recvTask + 1] - 2;
slab_y < meshmax_list[3 * recvTask + 1] + 4; slab_y++)
{
slab_yy = slab_y;
for(slab_z = meshmin_list[3 * recvTask + 2] - 2;
slab_z < meshmax_list[3 * recvTask + 2] + 4; slab_z++)
{
slab_zz = slab_z;
forcegrid[((slab_x - sendmin) * recv_dimy +
(slab_y - (meshmin_list[3 * recvTask + 1] - 2))) * recv_dimz +
slab_z - (meshmin_list[3 * recvTask + 2] - 2)] =
rhogrid[GRID * GRID2 * slab_xx + GRID2 * slab_yy + slab_zz];
}
}
}
}
if(level > 0)
{
MPI_Sendrecv(forcegrid,
(sendmax - sendmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real),
MPI_BYTE, recvTask, TAG_NONPERIOD_D,
workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
(recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_NONPERIOD_D, MPI_COMM_WORLD, &status);
}
else
{
memcpy(workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
forcegrid, (recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real));
}
}
}
}
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
recv_dimx = meshmax[0] - meshmin[0] + 6;
recv_dimy = meshmax[1] - meshmin[1] + 6;
recv_dimz = meshmax[2] - meshmin[2] + 6;
for(x = 0; x < meshmax[0] - meshmin[0] + 2; x++)
for(y = 0; y < meshmax[1] - meshmin[1] + 2; y++)
for(z = 0; z < meshmax[2] - meshmin[2] + 2; z++)
{
forcegrid[(x * dimy + y) * dimz + z]
= workspace[((x + 2) * recv_dimy + (y + 2)) * recv_dimz + (z + 2)];
}
/* read out the potential */
for(i = 0; i < NumPart; i++)
{
if (P[i].Type == 0 && P[i].ID < 0) /*SINK*/
continue;
#ifdef PLACEHIGHRESREGION
if(grnr == 1)
if(!(P[i].Type == 0 && P[i].Mass*1.0e10/All.HubbleParam < All.RefinementMass || P[i].Type == 1 && P[i].Mass*1.0e10/All.HubbleParam < (All.Omega0/All.OmegaBaryon-1.0)*All.RefinementMass))
continue;
#endif
slab_x = to_slab_fac * (P[i].Pos[0] - All.Corner[grnr][0]);
dx = to_slab_fac * (P[i].Pos[0] - All.Corner[grnr][0]) - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * (P[i].Pos[1] - All.Corner[grnr][1]);
dy = to_slab_fac * (P[i].Pos[1] - All.Corner[grnr][1]) - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * (P[i].Pos[2] - All.Corner[grnr][2]);
dz = to_slab_fac * (P[i].Pos[2] - All.Corner[grnr][2]) - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
pot = forcegrid[(slab_x * dimy + slab_y) * dimz + slab_z] * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
pot += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_z] * (1.0 - dx) * dy * (1.0 - dz);
pot += forcegrid[(slab_x * dimy + slab_y) * dimz + slab_zz] * (1.0 - dx) * (1.0 - dy) * dz;
pot += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_zz] * (1.0 - dx) * dy * dz;
pot += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_z] * (dx) * (1.0 - dy) * (1.0 - dz);
pot += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_z] * (dx) * dy * (1.0 - dz);
pot += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_zz] * (dx) * (1.0 - dy) * dz;
pot += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_zz] * (dx) * dy * dz;
P[i].Potential += pot;
}
pm_init_nonperiodic_free();
force_treeallocate(All.TreeAllocFactor * All.MaxPart, All.MaxPart);
All.NumForcesSinceLastDomainDecomp = 1 + All.TotNumPart * All.TreeDomainUpdateFrequency;
if(ThisTask == 0)
printf("done PM-potential.\n");
return 0;
}
/*! Calculates the long-range periodic force given the particle positions
* using the PM method. The force is Gaussian filtered with Asmth, given in
* mesh-cell units. We carry out a CIC charge assignment, and compute the
* potenial by Fourier transform methods. The potential is finite differenced
* using a 4-point finite differencing formula, and the forces are
* interpolated tri-linearly to the particle positions. The CIC kernel is
* deconvolved. Note that the particle distribution is not in the slab
* decomposition that is used for the FFT. Instead, overlapping patches
* between local domains and FFT slabs are communicated as needed.
*/
void pmforce_periodic(void)
{
double k2, kx, ky, kz, smth;
double dx, dy, dz;
double fx, fy, fz, ff;
double asmth2, fac, acc_dim;
int i, j, slab, level, sendTask, recvTask;
int x, y, z, xl, yl, zl, xr, yr, zr, xll, yll, zll, xrr, yrr, zrr, ip, dim;
int slab_x, slab_y, slab_z;
int slab_xx, slab_yy, slab_zz;
int meshmin[3], meshmax[3], sendmin, sendmax, recvmin, recvmax;
int rep, ncont, cont_sendmin[2], cont_sendmax[2], cont_recvmin[2], cont_recvmax[2];
int dimx, dimy, dimz, recv_dimx, recv_dimy, recv_dimz;
#ifndef NOMPI
MPI_Status status;
#endif
if(ThisTask == 0)
{
printf("Starting periodic PM calculation.\n");
fflush(stdout);
}
force_treefree();
asmth2 = (2 * M_PI) * All.Asmth[0] / All.BoxSize;
asmth2 *= asmth2;
fac = All.G / (M_PI * All.BoxSize); /* to get potential */
fac *= 1 / (2 * All.BoxSize / PMGRID); /* for finite differencing */
/* first, establish the extension of the local patch in the PMGRID */
for(j = 0; j < 3; j++)
{
meshmin[j] = PMGRID;
meshmax[j] = 0;
}
for(i = 0; i < NumPart; i++)
{
for(j = 0; j < 3; j++)
{
slab = to_slab_fac * P[i].Pos[j];
if(slab >= PMGRID)
slab = PMGRID - 1;
if(slab < meshmin[j])
meshmin[j] = slab;
if(slab > meshmax[j])
meshmax[j] = slab;
}
}
#ifndef NOMPI
MPI_Allgather(meshmin, 3, MPI_INT, meshmin_list, 3, MPI_INT, MPI_COMM_WORLD);
MPI_Allgather(meshmax, 3, MPI_INT, meshmax_list, 3, MPI_INT, MPI_COMM_WORLD);
#else
for( i = 0; i < 3; i++){
mesmmin_list[0+i] = meshmin[i];
meshmax_list[0+i] = meshmax[i];
}
#endif
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
pm_init_periodic_allocate((dimx + 4) * (dimy + 4) * (dimz + 4));
for(i = 0; i < dimx * dimy * dimz; i++)
workspace[i] = 0;
for(i = 0; i < NumPart; i++)
{
slab_x = to_slab_fac * P[i].Pos[0];
if(slab_x >= PMGRID)
slab_x = PMGRID - 1;
dx = to_slab_fac * P[i].Pos[0] - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * P[i].Pos[1];
if(slab_y >= PMGRID)
slab_y = PMGRID - 1;
dy = to_slab_fac * P[i].Pos[1] - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * P[i].Pos[2];
if(slab_z >= PMGRID)
slab_z = PMGRID - 1;
dz = to_slab_fac * P[i].Pos[2] - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
workspace[(slab_x * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_x * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * dy * (1.0 - dz);
workspace[(slab_x * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * dz;
workspace[(slab_x * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * dy * dz;
workspace[(slab_xx * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (dx) * dy * (1.0 - dz);
workspace[(slab_xx * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (dx) * (1.0 - dy) * dz;
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (dx) * dy * dz;
}
for(i = 0; i < fftsize; i++) /* clear local density field */
rhogrid[i] = 0;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * PMGRID;
sendmax = -1;
for(slab_x = meshmin[0]; slab_x < meshmax[0] + 2; slab_x++)
if(slab_to_task[slab_x % PMGRID] == recvTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -1)
sendmin = 0;
/* check how much we have to receive */
recvmin = 2 * PMGRID;
recvmax = -1;
for(slab_x = meshmin_list[3 * recvTask]; slab_x < meshmax_list[3 * recvTask] + 2; slab_x++)
if(slab_to_task[slab_x % PMGRID] == sendTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -1)
recvmin = 0;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 2;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 2;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 2;
if(level > 0)
{
#ifndef NOMPI
MPI_Sendrecv(workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE, recvTask,
TAG_PERIODIC_A, forcegrid,
(recvmax - recvmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_PERIODIC_A, MPI_COMM_WORLD, &status);
#else
memcpy(forcegrid, workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real));
#endif
}
else
{
memcpy(forcegrid, workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real));
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
slab_xx = (slab_x % PMGRID) - first_slab_of_task[ThisTask];
if(slab_xx >= 0 && slab_xx < slabs_per_task[ThisTask])
{
for(slab_y = meshmin_list[3 * recvTask + 1];
slab_y <= meshmax_list[3 * recvTask + 1] + 1; slab_y++)
{
slab_yy = slab_y;
if(slab_yy >= PMGRID)
slab_yy -= PMGRID;
for(slab_z = meshmin_list[3 * recvTask + 2];
slab_z <= meshmax_list[3 * recvTask + 2] + 1; slab_z++)
{
slab_zz = slab_z;
if(slab_zz >= PMGRID)
slab_zz -= PMGRID;
rhogrid[PMGRID * PMGRID2 * slab_xx + PMGRID2 * slab_yy + slab_zz] +=
forcegrid[((slab_x - recvmin) * recv_dimy +
(slab_y - meshmin_list[3 * recvTask + 1])) * recv_dimz +
(slab_z - meshmin_list[3 * recvTask + 2])];
}
}
}
}
}
}
}
/* Do the FFT of the density field */
rfftwnd_mpi(fft_forward_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
/* multiply with Green's function for the potential */
for(y = slabstart_y; y < slabstart_y + nslab_y; y++)
for(x = 0; x < PMGRID; x++)
for(z = 0; z < PMGRID / 2 + 1; z++)
{
if(x > PMGRID / 2)
kx = x - PMGRID;
else
kx = x;
if(y > PMGRID / 2)
ky = y - PMGRID;
else
ky = y;
if(z > PMGRID / 2)
kz = z - PMGRID;
else
kz = z;
k2 = kx * kx + ky * ky + kz * kz;
if(k2 > 0)
{
smth = -exp(-k2 * asmth2) / k2;
/* do deconvolution */
fx = fy = fz = 1;
if(kx != 0)
{
fx = (M_PI * kx) / PMGRID;
fx = sin(fx) / fx;
}
if(ky != 0)
{
fy = (M_PI * ky) / PMGRID;
fy = sin(fy) / fy;
}
if(kz != 0)
{
fz = (M_PI * kz) / PMGRID;
fz = sin(fz) / fz;
}
ff = 1 / (fx * fy * fz);
smth *= ff * ff * ff * ff;
/* end deconvolution */
ip = PMGRID * (PMGRID / 2 + 1) * (y - slabstart_y) + (PMGRID / 2 + 1) * x + z;
fft_of_rhogrid[ip].re *= smth;
fft_of_rhogrid[ip].im *= smth;
}
}
if(slabstart_y == 0)
fft_of_rhogrid[0].re = fft_of_rhogrid[0].im = 0.0;
/* Do the FFT to get the potential */
rfftwnd_mpi(fft_inverse_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
/* Now rhogrid holds the potential */
/* construct the potential for the local patch */
dimx = meshmax[0] - meshmin[0] + 6;
dimy = meshmax[1] - meshmin[1] + 6;
dimz = meshmax[2] - meshmin[2] + 6;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * PMGRID;
sendmax = -PMGRID;
for(slab_x = meshmin_list[3 * recvTask] - 2; slab_x < meshmax_list[3 * recvTask] + 4; slab_x++)
if(slab_to_task[(slab_x + PMGRID) % PMGRID] == sendTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -PMGRID)
sendmin = sendmax + 1;
/* check how much we have to receive */
recvmin = 2 * PMGRID;
recvmax = -PMGRID;
for(slab_x = meshmin[0] - 2; slab_x < meshmax[0] + 4; slab_x++)
if(slab_to_task[(slab_x + PMGRID) % PMGRID] == recvTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -PMGRID)
recvmin = recvmax + 1;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 6;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 6;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 6;
ncont = 1;
cont_sendmin[0] = sendmin;
cont_sendmax[0] = sendmax;
cont_sendmin[1] = sendmax + 1;
cont_sendmax[1] = sendmax;
cont_recvmin[0] = recvmin;
cont_recvmax[0] = recvmax;
cont_recvmin[1] = recvmax + 1;
cont_recvmax[1] = recvmax;
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
if(slab_to_task[(slab_x + PMGRID) % PMGRID] != ThisTask)
{
/* non-contiguous */
cont_sendmax[0] = slab_x - 1;
while(slab_to_task[(slab_x + PMGRID) % PMGRID] != ThisTask)
slab_x++;
cont_sendmin[1] = slab_x;
ncont++;
}
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
if(slab_to_task[(slab_x + PMGRID) % PMGRID] != recvTask)
{
/* non-contiguous */
cont_recvmax[0] = slab_x - 1;
while(slab_to_task[(slab_x + PMGRID) % PMGRID] != recvTask)
slab_x++;
cont_recvmin[1] = slab_x;
if(ncont == 1)
ncont++;
}
}
for(rep = 0; rep < ncont; rep++)
{
sendmin = cont_sendmin[rep];
sendmax = cont_sendmax[rep];
recvmin = cont_recvmin[rep];
recvmax = cont_recvmax[rep];
/* prepare what we want to send */
if(sendmax - sendmin >= 0)
{
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
slab_xx = ((slab_x + PMGRID) % PMGRID) - first_slab_of_task[ThisTask];
for(slab_y = meshmin_list[3 * recvTask + 1] - 2;
slab_y < meshmax_list[3 * recvTask + 1] + 4; slab_y++)
{
slab_yy = (slab_y + PMGRID) % PMGRID;
for(slab_z = meshmin_list[3 * recvTask + 2] - 2;
slab_z < meshmax_list[3 * recvTask + 2] + 4; slab_z++)
{
slab_zz = (slab_z + PMGRID) % PMGRID;
forcegrid[((slab_x - sendmin) * recv_dimy +
(slab_y - (meshmin_list[3 * recvTask + 1] - 2))) * recv_dimz +
slab_z - (meshmin_list[3 * recvTask + 2] - 2)] =
rhogrid[PMGRID * PMGRID2 * slab_xx + PMGRID2 * slab_yy + slab_zz];
}
}
}
}
if(level > 0)
{
#ifndef NOMPI
MPI_Sendrecv(forcegrid,
(sendmax - sendmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real),
MPI_BYTE, recvTask, TAG_PERIODIC_B,
workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
(recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_PERIODIC_B, MPI_COMM_WORLD, &status);
#else
memcpy(workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
forcegrid, (recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real));
#endif
}
else
{
memcpy(workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
forcegrid, (recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real));
}
}
}
}
}
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
recv_dimx = meshmax[0] - meshmin[0] + 6;
recv_dimy = meshmax[1] - meshmin[1] + 6;
recv_dimz = meshmax[2] - meshmin[2] + 6;
for(dim = 0; dim < 3; dim++) /* Calculate each component of the force. */
{
/* get the force component by finite differencing the potential */
/* note: "workspace" now contains the potential for the local patch, plus a suffiently large buffer region */
for(x = 0; x < meshmax[0] - meshmin[0] + 2; x++)
for(y = 0; y < meshmax[1] - meshmin[1] + 2; y++)
for(z = 0; z < meshmax[2] - meshmin[2] + 2; z++)
{
xrr = xll = xr = xl = x;
yrr = yll = yr = yl = y;
zrr = zll = zr = zl = z;
switch (dim)
{
case 0:
xr = x + 1;
xrr = x + 2;
xl = x - 1;
xll = x - 2;
break;
case 1:
yr = y + 1;
yl = y - 1;
yrr = y + 2;
yll = y - 2;
break;
case 2:
zr = z + 1;
zl = z - 1;
zrr = z + 2;
zll = z - 2;
break;
}
forcegrid[(x * dimy + y) * dimz + z]
=
fac * ((4.0 / 3) *
(workspace[((xl + 2) * recv_dimy + (yl + 2)) * recv_dimz + (zl + 2)]
- workspace[((xr + 2) * recv_dimy + (yr + 2)) * recv_dimz + (zr + 2)]) -
(1.0 / 6) *
(workspace[((xll + 2) * recv_dimy + (yll + 2)) * recv_dimz + (zll + 2)] -
workspace[((xrr + 2) * recv_dimy + (yrr + 2)) * recv_dimz + (zrr + 2)]));
}
/* read out the forces */
for(i = 0; i < NumPart; i++)
{
slab_x = to_slab_fac * P[i].Pos[0];
if(slab_x >= PMGRID)
slab_x = PMGRID - 1;
dx = to_slab_fac * P[i].Pos[0] - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * P[i].Pos[1];
if(slab_y >= PMGRID)
slab_y = PMGRID - 1;
dy = to_slab_fac * P[i].Pos[1] - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * P[i].Pos[2];
if(slab_z >= PMGRID)
slab_z = PMGRID - 1;
dz = to_slab_fac * P[i].Pos[2] - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
acc_dim =
forcegrid[(slab_x * dimy + slab_y) * dimz + slab_z] * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
acc_dim += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_z] * (1.0 - dx) * dy * (1.0 - dz);
acc_dim += forcegrid[(slab_x * dimy + slab_y) * dimz + slab_zz] * (1.0 - dx) * (1.0 - dy) * dz;
acc_dim += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_zz] * (1.0 - dx) * dy * dz;
acc_dim += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_z] * (dx) * (1.0 - dy) * (1.0 - dz);
acc_dim += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_z] * (dx) * dy * (1.0 - dz);
acc_dim += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_zz] * (dx) * (1.0 - dy) * dz;
acc_dim += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_zz] * (dx) * dy * dz;
P[i].GravPM[dim] = acc_dim;
}
}
pm_init_periodic_free();
force_treeallocate(All.TreeAllocFactor * All.MaxPart, All.MaxPart);
All.NumForcesSinceLastDomainDecomp = 1 + All.TotNumPart * All.TreeDomainUpdateFrequency;
if(ThisTask == 0)
{
printf("done PM.\n");
fflush(stdout);
}
}
/*! Calculates the long-range periodic force given the particle positions
* using the PM method. The force is Gaussian filtered with Asmth, given in
* mesh-cell units. We carry out a CIC charge assignment, and compute the
* potenial by Fourier transform methods. The potential is finite differenced
* using a 4-point finite differencing formula, and the forces are
* interpolated tri-linearly to the particle positions. The CIC kernel is
* deconvolved. Note that the particle distribution is not in the slab
* decomposition that is used for the FFT. Instead, overlapping patches
* between local domains and FFT slabs are communicated as needed.
*/
void pmforce_periodic(void)
{
double k2, kx, ky, kz, smth;
double dx, dy, dz;
double fx, fy, fz, ff;
double asmth2, fac, acc_dim;
int i, j, slab, level, sendTask, recvTask;
int x, y, z, xl, yl, zl, xr, yr, zr, xll, yll, zll, xrr, yrr, zrr, ip, dim;
int slab_x, slab_y, slab_z;
int slab_xx, slab_yy, slab_zz;
int meshmin[3], meshmax[3], sendmin, sendmax, recvmin, recvmax;
int rep, ncont, cont_sendmin[2], cont_sendmax[2], cont_recvmin[2], cont_recvmax[2];
int dimx, dimy, dimz, recv_dimx, recv_dimy, recv_dimz;
MPI_Status status;
if(ThisTask == 0)
{
printf("Starting periodic PM calculation.\n");
fflush(stdout);
}
#ifdef FFTW3
if(fftw_plan_exists)
{
/* macro defined in callgrind.h */
// CALLGRIND_START_INSTRUMENTATION;
}
#else
// CALLGRIND_START_INSTRUMENTATION;
#endif
force_treefree();
asmth2 = (2 * M_PI) * All.Asmth[0] / All.BoxSize;
asmth2 *= asmth2;
fac = All.G / (M_PI * All.BoxSize); /* to get potential */
fac *= 1 / (2 * All.BoxSize / PMGRID); /* for finite differencing */
/* first, establish the extension of the local patch in the PMGRID */
for(j = 0; j < 3; j++)
{
meshmin[j] = PMGRID;
meshmax[j] = 0;
}
for(i = 0; i < NumPart; i++)
{
for(j = 0; j < 3; j++)
{
slab = to_slab_fac * P[i].Pos[j];
if(slab >= PMGRID)
slab = PMGRID - 1;
if(slab < meshmin[j])
meshmin[j] = slab;
if(slab > meshmax[j])
meshmax[j] = slab;
}
}
MPI_Allgather(meshmin, 3, MPI_INT, meshmin_list, 3, MPI_INT, MPI_COMM_WORLD);
MPI_Allgather(meshmax, 3, MPI_INT, meshmax_list, 3, MPI_INT, MPI_COMM_WORLD);
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
pm_init_periodic_allocate((dimx + 4) * (dimy + 4) * (dimz + 4));
#ifdef FFTW3
if(!fftw_plan_exists)
{
/* Create plan for in-place r2c DFT */
fft_forward_plan = fftw_mpi_plan_dft_r2c_3d(PMGRID, PMGRID, PMGRID, rhogrid, fft_of_rhogrid,
MPI_COMM_WORLD, FFTW_PATIENT | FFTW_MPI_TRANSPOSED_OUT);
fft_inverse_plan = fftw_mpi_plan_dft_c2r_3d(PMGRID, PMGRID, PMGRID, fft_of_rhogrid, rhogrid,
MPI_COMM_WORLD, FFTW_PATIENT | FFTW_MPI_TRANSPOSED_IN);
fftw_plan_exists = true; // use C99 bool type
if(ThisTask == 0)
printf("Created new FFTW3 plan.\n");
} else {
/* do nothing, the plan has already been created by previous call to this function */
}
#endif
/* For FFTW3, there is a different convention for fftsize for real-to-complex transforms, i.e.
fftsize is the size of the complex data (number of complex values), NOT the size of the real data!
We attempt to take care of this by defining fftsize to be fftsize_real when using FFTW3. */
for(i = 0; i < dimx * dimy * dimz; i++)
workspace[i] = 0;
for(i = 0; i < NumPart; i++)
{
slab_x = to_slab_fac * P[i].Pos[0];
if(slab_x >= PMGRID)
slab_x = PMGRID - 1;
dx = to_slab_fac * P[i].Pos[0] - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * P[i].Pos[1];
if(slab_y >= PMGRID)
slab_y = PMGRID - 1;
dy = to_slab_fac * P[i].Pos[1] - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * P[i].Pos[2];
if(slab_z >= PMGRID)
slab_z = PMGRID - 1;
dz = to_slab_fac * P[i].Pos[2] - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
workspace[(slab_x * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_x * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * dy * (1.0 - dz);
workspace[(slab_x * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * dz;
workspace[(slab_x * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * dy * dz;
workspace[(slab_xx * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (dx) * dy * (1.0 - dz);
workspace[(slab_xx * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (dx) * (1.0 - dy) * dz;
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (dx) * dy * dz;
}
for(i = 0; i < fftsize; i++) /* clear local density field */
rhogrid[i] = 0;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * PMGRID;
sendmax = -1;
for(slab_x = meshmin[0]; slab_x < meshmax[0] + 2; slab_x++)
if(slab_to_task[slab_x % PMGRID] == recvTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -1)
sendmin = 0;
/* check how much we have to receive */
recvmin = 2 * PMGRID;
recvmax = -1;
for(slab_x = meshmin_list[3 * recvTask]; slab_x < meshmax_list[3 * recvTask] + 2; slab_x++)
if(slab_to_task[slab_x % PMGRID] == sendTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -1)
recvmin = 0;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 2;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 2;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 2;
if(level > 0)
{
MPI_Sendrecv(workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE, recvTask,
TAG_PERIODIC_A, forcegrid,
(recvmax - recvmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_PERIODIC_A, MPI_COMM_WORLD, &status);
}
else
{
memcpy(forcegrid, workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real));
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
slab_xx = (slab_x % PMGRID) - first_slab_of_task[ThisTask];
if(slab_xx >= 0 && slab_xx < slabs_per_task[ThisTask])
{
for(slab_y = meshmin_list[3 * recvTask + 1];
slab_y <= meshmax_list[3 * recvTask + 1] + 1; slab_y++)
{
slab_yy = slab_y;
if(slab_yy >= PMGRID)
slab_yy -= PMGRID;
for(slab_z = meshmin_list[3 * recvTask + 2];
slab_z <= meshmax_list[3 * recvTask + 2] + 1; slab_z++)
{
slab_zz = slab_z;
if(slab_zz >= PMGRID)
slab_zz -= PMGRID;
rhogrid[PMGRID * PMGRID2 * slab_xx + PMGRID2 * slab_yy + slab_zz] +=
forcegrid[((slab_x - recvmin) * recv_dimy +
(slab_y - meshmin_list[3 * recvTask + 1])) * recv_dimz +
(slab_z - meshmin_list[3 * recvTask + 2])];
}
}
}
}
}
}
}
#ifdef DEBUG_FFT
double norm_density = 0.;
for(i = 0; i < fftsize; i++)
{
norm_density += rhogrid[i]*rhogrid[i];
}
/* Write out rhogrid to a 'fft-snapshot' file */
if (ThisTask == 0) {
FILE *fp;
/* Print the norm of the fft */
printf("L2-norm of density: %f\n", norm_density);
printf("First five values of density: %f, %f, %f, %f, %f\n", rhogrid[0],rhogrid[1],rhogrid[2],rhogrid[3], rhogrid[4]);
}
#endif
/* Do the FFT of the density field */
#ifdef FFTW3
fftw_execute_dft_r2c(fft_forward_plan, rhogrid, fft_of_rhogrid);
#else
rfftwnd_mpi(fft_forward_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
#endif
#ifdef DEBUG_FFT
double norm_complex = 0.;
for(i = 0; i < fftsize; i++)
{
norm_complex += rhogrid[i]*rhogrid[i];
}
/* Write out rhogrid to a 'fft-snapshot' file */
if (ThisTask == 0) {
FILE *fp;
/* Print the norm of the fft */
printf("L2-norm of complex rhogrid: %f\n", norm_complex);
printf("First two values of complex fft: %f + i*%f, %f + i*%f\n", rhogrid[0],rhogrid[1],rhogrid[2],rhogrid[3]);
}
#endif
/* multiply with Green's function for the potential */
for(y = slabstart_y; y < slabstart_y + nslab_y; y++)
for(x = 0; x < PMGRID; x++)
for(z = 0; z < PMGRID / 2 + 1; z++)
{
if(x > PMGRID / 2)
kx = x - PMGRID;
else
kx = x;
if(y > PMGRID / 2)
ky = y - PMGRID;
else
ky = y;
if(z > PMGRID / 2)
kz = z - PMGRID;
else
kz = z;
k2 = kx * kx + ky * ky + kz * kz;
if(k2 > 0)
{
smth = -exp(-k2 * asmth2) / k2;
/* do deconvolution */
fx = fy = fz = 1;
if(kx != 0)
{
fx = (M_PI * kx) / PMGRID;
fx = sin(fx) / fx;
}
if(ky != 0)
{
fy = (M_PI * ky) / PMGRID;
fy = sin(fy) / fy;
}
if(kz != 0)
{
fz = (M_PI * kz) / PMGRID;
fz = sin(fz) / fz;
}
ff = 1 / (fx * fy * fz);
smth *= ff * ff * ff * ff;
/* end deconvolution */
ip = PMGRID * (PMGRID / 2 + 1) * (y - slabstart_y) + (PMGRID / 2 + 1) * x + z;
c_re(fft_of_rhogrid[ip]) *= smth;
c_im(fft_of_rhogrid[ip]) *= smth;
}
}
if(slabstart_y == 0)
c_re(fft_of_rhogrid[0]) = c_im(fft_of_rhogrid[0]) = 0.0;
/* Do the FFT to get the potential */
#ifdef FFTW3
fftw_execute_dft_c2r(fft_inverse_plan, fft_of_rhogrid, rhogrid);
/* Now normalize the output
for(i = 0; i < fftsize; i++)
rhogrid[i] = rhogrid[i] / (PMGRID*PMGRID*PMGRID); */
#else
rfftwnd_mpi(fft_inverse_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
#endif
#ifdef DEBUG_FFT
double norm = 0.;
for(i = 0; i < fftsize; i++)
{
norm += rhogrid[i]*rhogrid[i];
}
/* Write out rhogrid to a 'fft-snapshot' file */
if (ThisTask == 0) {
FILE *fp;
/* Print the norm of the fft */
printf("L2-norm of rhogrid: %f\n", norm);
printf("First five values of fft: %f, %f, %f, %f, %f\n\n", rhogrid[0],rhogrid[1],rhogrid[2],rhogrid[3],rhogrid[4]);
/* fp = fopen("rhogrid.0", "wb"); /* add suffix to indicate which node this is */
/* fwrite(rhogrid, sizeof(rhogrid[0]), fftsize, fp);*/
/* fclose(fp); */
}
#endif
/* Now rhogrid holds the potential */
/* construct the potential for the local patch */
dimx = meshmax[0] - meshmin[0] + 6;
dimy = meshmax[1] - meshmin[1] + 6;
dimz = meshmax[2] - meshmin[2] + 6;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * PMGRID;
sendmax = -PMGRID;
for(slab_x = meshmin_list[3 * recvTask] - 2; slab_x < meshmax_list[3 * recvTask] + 4; slab_x++)
if(slab_to_task[(slab_x + PMGRID) % PMGRID] == sendTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -PMGRID)
sendmin = sendmax + 1;
/* check how much we have to receive */
recvmin = 2 * PMGRID;
recvmax = -PMGRID;
for(slab_x = meshmin[0] - 2; slab_x < meshmax[0] + 4; slab_x++)
if(slab_to_task[(slab_x + PMGRID) % PMGRID] == recvTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -PMGRID)
recvmin = recvmax + 1;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 6;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 6;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 6;
ncont = 1;
cont_sendmin[0] = sendmin;
cont_sendmax[0] = sendmax;
cont_sendmin[1] = sendmax + 1;
cont_sendmax[1] = sendmax;
cont_recvmin[0] = recvmin;
cont_recvmax[0] = recvmax;
cont_recvmin[1] = recvmax + 1;
cont_recvmax[1] = recvmax;
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
if(slab_to_task[(slab_x + PMGRID) % PMGRID] != ThisTask)
{
/* non-contiguous */
cont_sendmax[0] = slab_x - 1;
while(slab_to_task[(slab_x + PMGRID) % PMGRID] != ThisTask)
slab_x++;
cont_sendmin[1] = slab_x;
ncont++;
}
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
if(slab_to_task[(slab_x + PMGRID) % PMGRID] != recvTask)
{
/* non-contiguous */
cont_recvmax[0] = slab_x - 1;
while(slab_to_task[(slab_x + PMGRID) % PMGRID] != recvTask)
slab_x++;
cont_recvmin[1] = slab_x;
if(ncont == 1)
ncont++;
}
}
for(rep = 0; rep < ncont; rep++)
{
sendmin = cont_sendmin[rep];
sendmax = cont_sendmax[rep];
recvmin = cont_recvmin[rep];
recvmax = cont_recvmax[rep];
/* prepare what we want to send */
if(sendmax - sendmin >= 0)
{
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
slab_xx = ((slab_x + PMGRID) % PMGRID) - first_slab_of_task[ThisTask];
for(slab_y = meshmin_list[3 * recvTask + 1] - 2;
slab_y < meshmax_list[3 * recvTask + 1] + 4; slab_y++)
{
slab_yy = (slab_y + PMGRID) % PMGRID;
for(slab_z = meshmin_list[3 * recvTask + 2] - 2;
slab_z < meshmax_list[3 * recvTask + 2] + 4; slab_z++)
{
slab_zz = (slab_z + PMGRID) % PMGRID;
forcegrid[((slab_x - sendmin) * recv_dimy +
(slab_y - (meshmin_list[3 * recvTask + 1] - 2))) * recv_dimz +
slab_z - (meshmin_list[3 * recvTask + 2] - 2)] =
rhogrid[PMGRID * PMGRID2 * slab_xx + PMGRID2 * slab_yy + slab_zz];
}
}
}
}
if(level > 0)
{
MPI_Sendrecv(forcegrid,
(sendmax - sendmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real),
MPI_BYTE, recvTask, TAG_PERIODIC_B,
workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
(recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_PERIODIC_B, MPI_COMM_WORLD, &status);
}
else
{
memcpy(workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
forcegrid, (recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real));
}
}
}
}
}
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
recv_dimx = meshmax[0] - meshmin[0] + 6;
recv_dimy = meshmax[1] - meshmin[1] + 6;
recv_dimz = meshmax[2] - meshmin[2] + 6;
for(dim = 0; dim < 3; dim++) /* Calculate each component of the force. */
{
/* get the force component by finite differencing the potential */
/* note: "workspace" now contains the potential for the local patch, plus a suffiently large buffer region */
for(x = 0; x < meshmax[0] - meshmin[0] + 2; x++)
for(y = 0; y < meshmax[1] - meshmin[1] + 2; y++)
for(z = 0; z < meshmax[2] - meshmin[2] + 2; z++)
{
xrr = xll = xr = xl = x;
yrr = yll = yr = yl = y;
zrr = zll = zr = zl = z;
switch (dim)
{
case 0:
xr = x + 1;
xrr = x + 2;
xl = x - 1;
xll = x - 2;
break;
case 1:
yr = y + 1;
yl = y - 1;
yrr = y + 2;
yll = y - 2;
break;
case 2:
zr = z + 1;
zl = z - 1;
zrr = z + 2;
zll = z - 2;
break;
}
forcegrid[(x * dimy + y) * dimz + z]
=
fac * ((4.0 / 3) *
(workspace[((xl + 2) * recv_dimy + (yl + 2)) * recv_dimz + (zl + 2)]
- workspace[((xr + 2) * recv_dimy + (yr + 2)) * recv_dimz + (zr + 2)]) -
(1.0 / 6) *
(workspace[((xll + 2) * recv_dimy + (yll + 2)) * recv_dimz + (zll + 2)] -
workspace[((xrr + 2) * recv_dimy + (yrr + 2)) * recv_dimz + (zrr + 2)]));
}
/* read out the forces */
for(i = 0; i < NumPart; i++)
{
slab_x = to_slab_fac * P[i].Pos[0];
if(slab_x >= PMGRID)
slab_x = PMGRID - 1;
dx = to_slab_fac * P[i].Pos[0] - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * P[i].Pos[1];
if(slab_y >= PMGRID)
slab_y = PMGRID - 1;
dy = to_slab_fac * P[i].Pos[1] - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * P[i].Pos[2];
if(slab_z >= PMGRID)
slab_z = PMGRID - 1;
dz = to_slab_fac * P[i].Pos[2] - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
acc_dim =
forcegrid[(slab_x * dimy + slab_y) * dimz + slab_z] * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
acc_dim += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_z] * (1.0 - dx) * dy * (1.0 - dz);
acc_dim += forcegrid[(slab_x * dimy + slab_y) * dimz + slab_zz] * (1.0 - dx) * (1.0 - dy) * dz;
acc_dim += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_zz] * (1.0 - dx) * dy * dz;
acc_dim += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_z] * (dx) * (1.0 - dy) * (1.0 - dz);
acc_dim += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_z] * (dx) * dy * (1.0 - dz);
acc_dim += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_zz] * (dx) * (1.0 - dy) * dz;
acc_dim += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_zz] * (dx) * dy * dz;
P[i].GravPM[dim] = acc_dim;
}
}
pm_init_periodic_free();
force_treeallocate(All.TreeAllocFactor * All.MaxPart, All.MaxPart);
All.NumForcesSinceLastDomainDecomp = 1 + All.TotNumPart * All.TreeDomainUpdateFrequency;
if(ThisTask == 0)
{
printf("done PM.\n");
fflush(stdout);
}
#ifdef FFTW3
if(fftw_plan_exists)
{
/* macro defined in callgrind.h */
// CALLGRIND_STOP_INSTRUMENTATION;
}
#else
// CALLGRIND_STOP_INSTRUMENTATION;
#endif
}
/*! Calculates the long-range potential using the PM method. The potential is
* Gaussian filtered with Asmth, given in mesh-cell units. We carry out a CIC
* charge assignment, and compute the potenial by Fourier transform
* methods. The CIC kernel is deconvolved.
*/
void pmpotential_periodic(void)
{
double k2, kx, ky, kz, smth;
double dx, dy, dz;
double fx, fy, fz, ff;
double asmth2, fac;
int i, j, slab, level, sendTask, recvTask;
int x, y, z, ip;
int slab_x, slab_y, slab_z;
int slab_xx, slab_yy, slab_zz;
int meshmin[3], meshmax[3], sendmin, sendmax, recvmin, recvmax;
int rep, ncont, cont_sendmin[2], cont_sendmax[2], cont_recvmin[2], cont_recvmax[2];
int dimx, dimy, dimz, recv_dimx, recv_dimy, recv_dimz;
MPI_Status status;
if(ThisTask == 0)
{
printf("Starting periodic PM calculation.\n");
fflush(stdout);
}
asmth2 = (2 * M_PI) * All.Asmth[0] / All.BoxSize;
asmth2 *= asmth2;
fac = All.G / (M_PI * All.BoxSize); /* to get potential */
force_treefree();
/* first, establish the extension of the local patch in the PMGRID */
for(j = 0; j < 3; j++)
{
meshmin[j] = PMGRID;
meshmax[j] = 0;
}
for(i = 0; i < NumPart; i++)
{
for(j = 0; j < 3; j++)
{
slab = to_slab_fac * P[i].Pos[j];
if(slab >= PMGRID)
slab = PMGRID - 1;
if(slab < meshmin[j])
meshmin[j] = slab;
if(slab > meshmax[j])
meshmax[j] = slab;
}
}
MPI_Allgather(meshmin, 3, MPI_INT, meshmin_list, 3, MPI_INT, MPI_COMM_WORLD);
MPI_Allgather(meshmax, 3, MPI_INT, meshmax_list, 3, MPI_INT, MPI_COMM_WORLD);
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
pm_init_periodic_allocate((dimx + 4) * (dimy + 4) * (dimz + 4));
#ifdef FFTW3
/* Create plan for in-place r2c DFT */
fft_forward_plan = fftw_mpi_plan_dft_r2c_3d(PMGRID, PMGRID, PMGRID, rhogrid, fft_of_rhogrid,
MPI_COMM_WORLD, FFTW_ESTIMATE | FFTW_MPI_TRANSPOSED_OUT);
fft_inverse_plan = fftw_mpi_plan_dft_c2r_3d(PMGRID, PMGRID, PMGRID, fft_of_rhogrid, rhogrid,
MPI_COMM_WORLD, FFTW_ESTIMATE | FFTW_MPI_TRANSPOSED_IN);
#endif
for(i = 0; i < dimx * dimy * dimz; i++)
workspace[i] = 0;
for(i = 0; i < NumPart; i++)
{
slab_x = to_slab_fac * P[i].Pos[0];
if(slab_x >= PMGRID)
slab_x = PMGRID - 1;
dx = to_slab_fac * P[i].Pos[0] - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * P[i].Pos[1];
if(slab_y >= PMGRID)
slab_y = PMGRID - 1;
dy = to_slab_fac * P[i].Pos[1] - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * P[i].Pos[2];
if(slab_z >= PMGRID)
slab_z = PMGRID - 1;
dz = to_slab_fac * P[i].Pos[2] - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
workspace[(slab_x * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_x * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (1.0 - dx) * dy * (1.0 - dz);
workspace[(slab_x * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * (1.0 - dy) * dz;
workspace[(slab_x * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (1.0 - dx) * dy * dz;
workspace[(slab_xx * dimy + slab_y) * dimz + slab_z] += P[i].Mass * (dx) * (1.0 - dy) * (1.0 - dz);
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_z] += P[i].Mass * (dx) * dy * (1.0 - dz);
workspace[(slab_xx * dimy + slab_y) * dimz + slab_zz] += P[i].Mass * (dx) * (1.0 - dy) * dz;
workspace[(slab_xx * dimy + slab_yy) * dimz + slab_zz] += P[i].Mass * (dx) * dy * dz;
}
for(i = 0; i < fftsize; i++) /* clear local density field */
rhogrid[i] = 0;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * PMGRID;
sendmax = -1;
for(slab_x = meshmin[0]; slab_x < meshmax[0] + 2; slab_x++)
if(slab_to_task[slab_x % PMGRID] == recvTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -1)
sendmin = 0;
/* check how much we have to receive */
recvmin = 2 * PMGRID;
recvmax = -1;
for(slab_x = meshmin_list[3 * recvTask]; slab_x < meshmax_list[3 * recvTask] + 2; slab_x++)
if(slab_to_task[slab_x % PMGRID] == sendTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -1)
recvmin = 0;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 2;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 2;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 2;
if(level > 0)
{
MPI_Sendrecv(workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE, recvTask,
TAG_PERIODIC_C, forcegrid,
(recvmax - recvmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_PERIODIC_C, MPI_COMM_WORLD, &status);
}
else
{
memcpy(forcegrid, workspace + (sendmin - meshmin[0]) * dimy * dimz,
(sendmax - sendmin + 1) * dimy * dimz * sizeof(fftw_real));
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
slab_xx = (slab_x % PMGRID) - first_slab_of_task[ThisTask];
if(slab_xx >= 0 && slab_xx < slabs_per_task[ThisTask])
{
for(slab_y = meshmin_list[3 * recvTask + 1];
slab_y <= meshmax_list[3 * recvTask + 1] + 1; slab_y++)
{
slab_yy = slab_y;
if(slab_yy >= PMGRID)
slab_yy -= PMGRID;
for(slab_z = meshmin_list[3 * recvTask + 2];
slab_z <= meshmax_list[3 * recvTask + 2] + 1; slab_z++)
{
slab_zz = slab_z;
if(slab_zz >= PMGRID)
slab_zz -= PMGRID;
rhogrid[PMGRID * PMGRID2 * slab_xx + PMGRID2 * slab_yy + slab_zz] +=
forcegrid[((slab_x - recvmin) * recv_dimy +
(slab_y - meshmin_list[3 * recvTask + 1])) * recv_dimz +
(slab_z - meshmin_list[3 * recvTask + 2])];
}
}
}
}
}
}
}
/* Do the FFT of the density field */
#ifdef FFTW3
fftw_execute(fft_forward_plan);
#else
rfftwnd_mpi(fft_forward_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
#endif
/* multiply with Green's function for the potential */
for(y = slabstart_y; y < slabstart_y + nslab_y; y++)
for(x = 0; x < PMGRID; x++)
for(z = 0; z < PMGRID / 2 + 1; z++)
{
if(x > PMGRID / 2)
kx = x - PMGRID;
else
kx = x;
if(y > PMGRID / 2)
ky = y - PMGRID;
else
ky = y;
if(z > PMGRID / 2)
kz = z - PMGRID;
else
kz = z;
k2 = kx * kx + ky * ky + kz * kz;
if(k2 > 0)
{
smth = -exp(-k2 * asmth2) / k2 * fac;
/* do deconvolution */
fx = fy = fz = 1;
if(kx != 0)
{
fx = (M_PI * kx) / PMGRID;
fx = sin(fx) / fx;
}
if(ky != 0)
{
fy = (M_PI * ky) / PMGRID;
fy = sin(fy) / fy;
}
if(kz != 0)
{
fz = (M_PI * kz) / PMGRID;
fz = sin(fz) / fz;
}
ff = 1 / (fx * fy * fz);
smth *= ff * ff * ff * ff;
/* end deconvolution */
ip = PMGRID * (PMGRID / 2 + 1) * (y - slabstart_y) + (PMGRID / 2 + 1) * x + z;
c_re(fft_of_rhogrid[ip]) *= smth;
c_im(fft_of_rhogrid[ip]) *= smth;
}
}
if(slabstart_y == 0)
c_re(fft_of_rhogrid[0]) = c_im(fft_of_rhogrid[0]) = 0.0;
/* Do the FFT to get the potential */
#ifdef FFTW3
fftw_execute(fft_inverse_plan);
/* Now normalize the output */
for(i = 0; i < fftsize; i++)
rhogrid[i] = rhogrid[i] / (PMGRID*PMGRID*PMGRID);
#else
rfftwnd_mpi(fft_inverse_plan, 1, rhogrid, workspace, FFTW_TRANSPOSED_ORDER);
#endif
/* note: "rhogrid" now contains the potential */
dimx = meshmax[0] - meshmin[0] + 6;
dimy = meshmax[1] - meshmin[1] + 6;
dimz = meshmax[2] - meshmin[2] + 6;
for(level = 0; level < (1 << PTask); level++) /* note: for level=0, target is the same task */
{
sendTask = ThisTask;
recvTask = ThisTask ^ level;
if(recvTask < NTask)
{
/* check how much we have to send */
sendmin = 2 * PMGRID;
sendmax = -PMGRID;
for(slab_x = meshmin_list[3 * recvTask] - 2; slab_x < meshmax_list[3 * recvTask] + 4; slab_x++)
if(slab_to_task[(slab_x + PMGRID) % PMGRID] == sendTask)
{
if(slab_x < sendmin)
sendmin = slab_x;
if(slab_x > sendmax)
sendmax = slab_x;
}
if(sendmax == -PMGRID)
sendmin = sendmax + 1;
/* check how much we have to receive */
recvmin = 2 * PMGRID;
recvmax = -PMGRID;
for(slab_x = meshmin[0] - 2; slab_x < meshmax[0] + 4; slab_x++)
if(slab_to_task[(slab_x + PMGRID) % PMGRID] == recvTask)
{
if(slab_x < recvmin)
recvmin = slab_x;
if(slab_x > recvmax)
recvmax = slab_x;
}
if(recvmax == -PMGRID)
recvmin = recvmax + 1;
if((recvmax - recvmin) >= 0 || (sendmax - sendmin) >= 0) /* ok, we have a contribution to the slab */
{
recv_dimx = meshmax_list[3 * recvTask + 0] - meshmin_list[3 * recvTask + 0] + 6;
recv_dimy = meshmax_list[3 * recvTask + 1] - meshmin_list[3 * recvTask + 1] + 6;
recv_dimz = meshmax_list[3 * recvTask + 2] - meshmin_list[3 * recvTask + 2] + 6;
ncont = 1;
cont_sendmin[0] = sendmin;
cont_sendmax[0] = sendmax;
cont_sendmin[1] = sendmax + 1;
cont_sendmax[1] = sendmax;
cont_recvmin[0] = recvmin;
cont_recvmax[0] = recvmax;
cont_recvmin[1] = recvmax + 1;
cont_recvmax[1] = recvmax;
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
if(slab_to_task[(slab_x + PMGRID) % PMGRID] != ThisTask)
{
/* non-contiguous */
cont_sendmax[0] = slab_x - 1;
while(slab_to_task[(slab_x + PMGRID) % PMGRID] != ThisTask)
slab_x++;
cont_sendmin[1] = slab_x;
ncont++;
}
}
for(slab_x = recvmin; slab_x <= recvmax; slab_x++)
{
if(slab_to_task[(slab_x + PMGRID) % PMGRID] != recvTask)
{
/* non-contiguous */
cont_recvmax[0] = slab_x - 1;
while(slab_to_task[(slab_x + PMGRID) % PMGRID] != recvTask)
slab_x++;
cont_recvmin[1] = slab_x;
if(ncont == 1)
ncont++;
}
}
for(rep = 0; rep < ncont; rep++)
{
sendmin = cont_sendmin[rep];
sendmax = cont_sendmax[rep];
recvmin = cont_recvmin[rep];
recvmax = cont_recvmax[rep];
/* prepare what we want to send */
if(sendmax - sendmin >= 0)
{
for(slab_x = sendmin; slab_x <= sendmax; slab_x++)
{
slab_xx = ((slab_x + PMGRID) % PMGRID) - first_slab_of_task[ThisTask];
for(slab_y = meshmin_list[3 * recvTask + 1] - 2;
slab_y < meshmax_list[3 * recvTask + 1] + 4; slab_y++)
{
slab_yy = (slab_y + PMGRID) % PMGRID;
for(slab_z = meshmin_list[3 * recvTask + 2] - 2;
slab_z < meshmax_list[3 * recvTask + 2] + 4; slab_z++)
{
slab_zz = (slab_z + PMGRID) % PMGRID;
forcegrid[((slab_x - sendmin) * recv_dimy +
(slab_y - (meshmin_list[3 * recvTask + 1] - 2))) * recv_dimz +
slab_z - (meshmin_list[3 * recvTask + 2] - 2)] =
rhogrid[PMGRID * PMGRID2 * slab_xx + PMGRID2 * slab_yy + slab_zz];
}
}
}
}
if(level > 0)
{
MPI_Sendrecv(forcegrid,
(sendmax - sendmin + 1) * recv_dimy * recv_dimz * sizeof(fftw_real),
MPI_BYTE, recvTask, TAG_PERIODIC_D,
workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
(recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real), MPI_BYTE,
recvTask, TAG_PERIODIC_D, MPI_COMM_WORLD, &status);
}
else
{
memcpy(workspace + (recvmin - (meshmin[0] - 2)) * dimy * dimz,
forcegrid, (recvmax - recvmin + 1) * dimy * dimz * sizeof(fftw_real));
}
}
}
}
}
dimx = meshmax[0] - meshmin[0] + 2;
dimy = meshmax[1] - meshmin[1] + 2;
dimz = meshmax[2] - meshmin[2] + 2;
recv_dimx = meshmax[0] - meshmin[0] + 6;
recv_dimy = meshmax[1] - meshmin[1] + 6;
recv_dimz = meshmax[2] - meshmin[2] + 6;
for(x = 0; x < meshmax[0] - meshmin[0] + 2; x++)
for(y = 0; y < meshmax[1] - meshmin[1] + 2; y++)
for(z = 0; z < meshmax[2] - meshmin[2] + 2; z++)
{
forcegrid[(x * dimy + y) * dimz + z] =
workspace[((x + 2) * recv_dimy + (y + 2)) * recv_dimz + (z + 2)];
}
/* read out the potential */
for(i = 0; i < NumPart; i++)
{
slab_x = to_slab_fac * P[i].Pos[0];
if(slab_x >= PMGRID)
slab_x = PMGRID - 1;
dx = to_slab_fac * P[i].Pos[0] - slab_x;
slab_x -= meshmin[0];
slab_xx = slab_x + 1;
slab_y = to_slab_fac * P[i].Pos[1];
if(slab_y >= PMGRID)
slab_y = PMGRID - 1;
dy = to_slab_fac * P[i].Pos[1] - slab_y;
slab_y -= meshmin[1];
slab_yy = slab_y + 1;
slab_z = to_slab_fac * P[i].Pos[2];
if(slab_z >= PMGRID)
slab_z = PMGRID - 1;
dz = to_slab_fac * P[i].Pos[2] - slab_z;
slab_z -= meshmin[2];
slab_zz = slab_z + 1;
P[i].Potential +=
forcegrid[(slab_x * dimy + slab_y) * dimz + slab_z] * (1.0 - dx) * (1.0 - dy) * (1.0 - dz);
P[i].Potential += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_z] * (1.0 - dx) * dy * (1.0 - dz);
P[i].Potential += forcegrid[(slab_x * dimy + slab_y) * dimz + slab_zz] * (1.0 - dx) * (1.0 - dy) * dz;
P[i].Potential += forcegrid[(slab_x * dimy + slab_yy) * dimz + slab_zz] * (1.0 - dx) * dy * dz;
P[i].Potential += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_z] * (dx) * (1.0 - dy) * (1.0 - dz);
P[i].Potential += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_z] * (dx) * dy * (1.0 - dz);
P[i].Potential += forcegrid[(slab_xx * dimy + slab_y) * dimz + slab_zz] * (dx) * (1.0 - dy) * dz;
P[i].Potential += forcegrid[(slab_xx * dimy + slab_yy) * dimz + slab_zz] * (dx) * dy * dz;
}
pm_init_periodic_free();
force_treeallocate(All.TreeAllocFactor * All.MaxPart, All.MaxPart);
All.NumForcesSinceLastDomainDecomp = 1 + All.TotNumPart * All.TreeDomainUpdateFrequency;
if(ThisTask == 0)
{
printf("done PM-Potential.\n");
fflush(stdout);
}
}
/* This function reads or writes the restart files.
* Each processor writes its own restart file, with the
* I/O being done in parallel. To avoid congestion of the disks
* you can tell the program to restrict the number of files
* that are simultaneously written to NumFilesWrittenInParallel.
*
* If modus>0 the restart()-routine reads,
* if modus==0 it writes a restart file.
*/
void restart(int modus)
{
char buf[200], buf_bak[200], buf_mv[500];
double save_PartAllocFactor, save_TreeAllocFactor;
int i, nprocgroup, masterTask, groupTask, old_MaxPart, old_MaxNodes;
struct global_data_all_processes all_task0;
#if defined(SFR) || defined(BLACK_HOLES)
#ifdef NO_TREEDATA_IN_RESTART
if(modus == 0)
{
rearrange_particle_sequence();
All.NumForcesSinceLastDomainDecomp = 1 + All.TreeDomainUpdateFrequency * All.TotNumPart; /* ensures that new tree will be constructed */
}
#endif
#endif
sprintf(buf, "%s%s.%d", All.OutputDir, All.RestartFile, ThisTask);
sprintf(buf_bak, "%s%s.%d.bak", All.OutputDir, All.RestartFile, ThisTask);
sprintf(buf_mv, "mv %s %s", buf, buf_bak);
if((NTask < All.NumFilesWrittenInParallel))
{
printf
("Fatal error.\nNumber of processors must be a smaller or equal than `NumFilesWrittenInParallel'.\n");
endrun(2131);
}
nprocgroup = NTask / All.NumFilesWrittenInParallel;
if((NTask % All.NumFilesWrittenInParallel))
{
nprocgroup++;
}
masterTask = (ThisTask / nprocgroup) * nprocgroup;
for(groupTask = 0; groupTask < nprocgroup; groupTask++)
{
if(ThisTask == (masterTask + groupTask)) /* ok, it's this processor's turn */
{
if(modus)
{
if(!(fd = fopen(buf, "r")))
{
printf("Restart file '%s' not found.\n", buf);
endrun(7870);
}
}
else
{
system(buf_mv); /* move old restart files to .bak files */
if(!(fd = fopen(buf, "w")))
{
printf("Restart file '%s' cannot be opened.\n", buf);
endrun(7878);
}
}
save_PartAllocFactor = All.PartAllocFactor;
save_TreeAllocFactor = All.TreeAllocFactor;
/* common data */
byten(&All, sizeof(struct global_data_all_processes), modus);
if(ThisTask == 0 && modus > 0)
all_task0 = All;
if(modus > 0 && groupTask == 0) /* read */
{
MPI_Bcast(&all_task0, sizeof(struct global_data_all_processes), MPI_BYTE, 0, MPI_COMM_WORLD);
}
old_MaxPart = All.MaxPart;
old_MaxNodes = All.TreeAllocFactor * All.MaxPart;
if(modus) /* read */
{
if(All.PartAllocFactor != save_PartAllocFactor)
{
All.PartAllocFactor = save_PartAllocFactor;
All.MaxPart = All.PartAllocFactor * (All.TotNumPart / NTask);
All.MaxPartSph = All.PartAllocFactor * (All.TotN_gas / NTask);
#ifdef INHOMOG_GASDISTR_HINT
All.MaxPartSph = All.MaxPart;
#endif
save_PartAllocFactor = -1;
}
if(All.TreeAllocFactor != save_TreeAllocFactor)
{
All.TreeAllocFactor = save_TreeAllocFactor;
save_TreeAllocFactor = -1;
}
if(all_task0.Time != All.Time)
{
printf("The restart file on task=%d is not consistent with the one on task=0\n", ThisTask);
fflush(stdout);
endrun(16);
}
allocate_memory();
}
in(&NumPart, modus);
if(NumPart > All.MaxPart)
{
printf
("it seems you have reduced(!) 'PartAllocFactor' below the value of %g needed to load the restart file.\n",
NumPart / (((double) All.TotNumPart) / NTask));
printf("fatal error\n");
endrun(22);
}
/* Particle data */
byten(&P[0], NumPart * sizeof(struct particle_data), modus);
in(&N_gas, modus);
if(N_gas > 0)
{
if(N_gas > All.MaxPartSph)
{
printf
("SPH: it seems you have reduced(!) 'PartAllocFactor' below the value of %g needed to load the restart file.\n",
N_gas / (((double) All.TotN_gas) / NTask));
printf("fatal error\n");
endrun(222);
}
/* Sph-Particle data */
byten(&SphP[0], N_gas * sizeof(struct sph_particle_data), modus);
}
/* write state of random number generator */
byten(gsl_rng_state(random_generator), gsl_rng_size(random_generator), modus);
#ifndef NO_TREEDATA_IN_RESTART
/* now store relevant data for tree */
#ifdef SFR
in(&Stars_converted, modus);
#endif
if(modus) /* read */
{
ngb_treeallocate(MAX_NGB);
force_treeallocate(All.TreeAllocFactor * All.MaxPart, All.MaxPart);
}
in(&Numnodestree, modus);
if(Numnodestree > MaxNodes)
{
printf
("Tree storage: it seems you have reduced(!) 'PartAllocFactor' below the value needed to load the restart file (task=%d). "
"Numnodestree=%d MaxNodes=%d\n", ThisTask, Numnodestree, MaxNodes);
endrun(221);
}
byten(Nodes_base, Numnodestree * sizeof(struct NODE), modus);
byten(Extnodes_base, Numnodestree * sizeof(struct extNODE), modus);
byten(Father, NumPart * sizeof(int), modus);
byten(Nextnode, NumPart * sizeof(int), modus);
byten(Nextnode + All.MaxPart, MAXTOPNODES * sizeof(int), modus);
byten(DomainStartList, NTask * sizeof(int), modus);
byten(DomainEndList, NTask * sizeof(int), modus);
byten(DomainTask, MAXTOPNODES * sizeof(int), modus);
byten(DomainNodeIndex, MAXTOPNODES * sizeof(int), modus);
byten(DomainTreeNodeLen, MAXTOPNODES * sizeof(FLOAT), modus);
byten(DomainHmax, MAXTOPNODES * sizeof(FLOAT), modus);
byten(DomainMoment, MAXTOPNODES * sizeof(struct DomainNODE), modus);
byten(DomainCorner, 3 * sizeof(double), modus);
byten(DomainCenter, 3 * sizeof(double), modus);
byten(&DomainLen, sizeof(double), modus);
byten(&DomainFac, sizeof(double), modus);
byten(&DomainMyStart, sizeof(int), modus);
byten(&DomainMyLast, sizeof(int), modus);
if(modus) /* read */
if(All.PartAllocFactor != save_PartAllocFactor || All.TreeAllocFactor != save_TreeAllocFactor)
{
for(i = 0; i < NumPart; i++)
Father[i] += (All.MaxPart - old_MaxPart);
for(i = 0; i < NumPart; i++)
if(Nextnode[i] >= old_MaxPart)
{
if(Nextnode[i] >= old_MaxPart + old_MaxNodes)
Nextnode[i] += (All.MaxPart - old_MaxPart) + (MaxNodes - old_MaxPart);
else
Nextnode[i] += (All.MaxPart - old_MaxPart);
}
for(i = 0; i < Numnodestree; i++)
{
if(Nodes_base[i].u.d.sibling >= old_MaxPart)
{
if(Nodes_base[i].u.d.sibling >= old_MaxPart + old_MaxNodes)
Nodes_base[i].u.d.sibling +=
(All.MaxPart - old_MaxPart) + (MaxNodes - old_MaxNodes);
else
Nodes_base[i].u.d.sibling += (All.MaxPart - old_MaxPart);
}
if(Nodes_base[i].u.d.father >= old_MaxPart)
{
if(Nodes_base[i].u.d.father >= old_MaxPart + old_MaxNodes)
Nodes_base[i].u.d.father += (All.MaxPart - old_MaxPart) + (MaxNodes - old_MaxNodes);
else
Nodes_base[i].u.d.father += (All.MaxPart - old_MaxPart);
}
if(Nodes_base[i].u.d.nextnode >= old_MaxPart)
{
if(Nodes_base[i].u.d.nextnode >= old_MaxPart + old_MaxNodes)
Nodes_base[i].u.d.nextnode +=
(All.MaxPart - old_MaxPart) + (MaxNodes - old_MaxNodes);
else
Nodes_base[i].u.d.nextnode += (All.MaxPart - old_MaxPart);
}
}
for(i = 0; i < MAXTOPNODES; i++)
if(Nextnode[i + All.MaxPart] >= old_MaxPart)
{
if(Nextnode[i + All.MaxPart] >= old_MaxPart + old_MaxNodes)
Nextnode[i + All.MaxPart] += (All.MaxPart - old_MaxPart) + (MaxNodes - old_MaxNodes);
else
Nextnode[i + All.MaxPart] += (All.MaxPart - old_MaxPart);
}
for(i = 0; i < MAXTOPNODES; i++)
if(DomainNodeIndex[i] >= old_MaxPart)
{
if(DomainNodeIndex[i] >= old_MaxPart + old_MaxNodes)
DomainNodeIndex[i] += (All.MaxPart - old_MaxPart) + (MaxNodes - old_MaxNodes);
else
DomainNodeIndex[i] += (All.MaxPart - old_MaxPart);
}
}
#endif
fclose(fd);
}
else /* wait inside the group */
{
if(modus > 0 && groupTask == 0) /* read */
{
MPI_Bcast(&all_task0, sizeof(struct global_data_all_processes), MPI_BYTE, 0, MPI_COMM_WORLD);
}
}
MPI_Barrier(MPI_COMM_WORLD);
}
}
