void writefile(AtomVec& atoms, OriginBox& obox){
// The .xyz format is simple:
// Line 1: [Number of atoms]
// Line 2: [Comment line, whatever you want, left blank here]
// Line 3: [element type, here C for carbon]\t[x]\t[y]\t[z]
// Line 4: [element type, here C for carbon]\t[x]\t[y]\t[z]
// ...
// Line N+1: [element type, here C for carbon]\t[x]\t[y]\t[z]
// Line N+2: [element type, here C for carbon]\t[x]\t[y]\t[z]
// Line N+3: [Number of atoms]
// Line N+4: [Comment line, whatever you want, left blank here]
// Line N+5: [element type, here C for carbon]\t[x]\t[y]\t[z]
// ...
// And so on. each set of N atoms/coordinates corresponds to a "frame",
// which then make a movie. There must be the same N in each frame for VMD.
// Note that if this is compiled as a 2D simulation, it will leave out
// the z coordinate, and VMD can't handle that.
ofstream outf;
outf.open("packing.xyz", ios::out);
outf.precision(24);
outf << atoms.size() << endl;
outf << "L=" << obox.L() << endl; // blank line for comment
for(uint i=0; i<atoms.size(); i++){
if(i < Ns){
outf << "C";
} else {
outf << "O";
}
Vec normloc = obox.diff(Vec::Zero(), atoms[i].x);
for(uint j=0; j<NDIM; j++){
outf << "\t" << normloc[j];
}
outf << endl;
};
// Unnecessary extra:
// Write a "tcl" file with the box boundaries
// the "tcl" file is made specifically for VMD
ofstream pbcfile;
pbcfile.open("packing.tcl", ios::out);
pbcfile << "set cell [pbc set {";
for(uint j=0; j<NDIM; j++){
pbcfile << obox.box_shape()[j] << " ";
}
pbcfile << "} -all];\n";
pbcfile << "pbc box -toggle -center origin -color red;\n";
pbcfile << "set natoms [atomselect 0 \"name C\";];\n"
<< "$natoms set radius " << (sigma/2.0) << ";\n"
<< "set natoms [atomselect 0 \"name O\";];\n"
<< "$natoms set radius " << (sigmal/2.0) << ";\n";
// Now you should be able to run "vmd -e LJatoms-pbc.tcl LJatoms.xyz"
// and it will show you the movie and also the bounding box
// if you have .vmdrc in that same directory, you should also be able
// to toggle the box with the "b" button
};
void writefile(ofstream& outf, AtomVec& atoms, Box& bx){
// The .xyz format is simple:
// Line 1: [Number of atoms]
// Line 2: [Comment line, whatever you want, left blank here]
// Line 3: [element type, here C for carbon]\t[x]\t[y]\t[z]
// Line 4: [element type, here C for carbon]\t[x]\t[y]\t[z]
// ...
// Line N+1: [element type, here C for carbon]\t[x]\t[y]\t[z]
// Line N+2: [element type, here C for carbon]\t[x]\t[y]\t[z]
// Line N+3: [Number of atoms]
// Line N+4: [Comment line, whatever you want, left blank here]
// Line N+5: [element type, here C for carbon]\t[x]\t[y]\t[z]
// ...
// And so on. each set of N atoms/coordinates corresponds to a "frame",
// which then make a movie. There must be the same N in each frame for VMD.
// Note that if this is compiled as a 2D simulation, it will leave out
// the z coordinate, and VMD can't handle that.
outf << atoms.size() << endl;
outf << endl; // blank line for comment
for(uint i=0; i<atoms.size(); i++){
outf << "C";
Vec normloc = bx.diff(Vec::Zero(), atoms[i].x);
for(uint j=0; j<NDIM; j++){
outf << "\t" << normloc[j];
}
outf << endl;
};
};
void CommBrick::reverse_comm()
{
int n;
MPI_Request request;
AtomVec *avec = atom->avec;
double **f = atom->f;
double *buf;
// exchange data with another proc
// if other proc is self, just copy
// if comm_f_only set, exchange or copy directly from f, don't pack
for (int iswap = nswap-1; iswap >= 0; iswap--) {
if (sendproc[iswap] != me) {
if (comm_f_only) {
if (size_reverse_recv[iswap])
MPI_Irecv(buf_recv,size_reverse_recv[iswap],MPI_DOUBLE,
sendproc[iswap],0,world,&request);
if (size_reverse_send[iswap]) {
if (size_reverse_send[iswap]) buf = f[firstrecv[iswap]];
else buf = NULL;
MPI_Send(buf,size_reverse_send[iswap],MPI_DOUBLE,
recvproc[iswap],0,world);
}
if (size_reverse_recv[iswap]) MPI_Wait(&request,MPI_STATUS_IGNORE);
} else {
if (size_reverse_recv[iswap])
MPI_Irecv(buf_recv,size_reverse_recv[iswap],MPI_DOUBLE,
sendproc[iswap],0,world,&request);
n = avec->pack_reverse(recvnum[iswap],firstrecv[iswap],buf_send);
if (n) MPI_Send(buf_send,n,MPI_DOUBLE,recvproc[iswap],0,world);
if (size_reverse_recv[iswap]) MPI_Wait(&request,MPI_STATUS_IGNORE);
}
avec->unpack_reverse(sendnum[iswap],sendlist[iswap],buf_recv);
} else {
if (comm_f_only) {
if (sendnum[iswap])
avec->unpack_reverse(sendnum[iswap],sendlist[iswap],
f[firstrecv[iswap]]);
} else {
avec->pack_reverse(recvnum[iswap],firstrecv[iswap],buf_send);
avec->unpack_reverse(sendnum[iswap],sendlist[iswap],buf_send);
}
}
}
}
void FixPour::pre_exchange()
{
int i;
// just return if should not be called on this timestep
if (next_reneighbor != update->ntimestep || lastexec == update->ntimestep) return;
lastexec = update->ntimestep;
// nnew = # to insert this timestep
int nnew = nper;
if (ninserted + nnew > ninsert) nnew = ninsert - ninserted;
// lo/hi current = z (or y) bounds of insertion region this timestep
if (domain->dimension == 3) {
lo_current = zlo + (update->ntimestep - nfirst) * update->dt * rate;
hi_current = zhi + (update->ntimestep - nfirst) * update->dt * rate;
} else {
lo_current = ylo + (update->ntimestep - nfirst) * update->dt * rate;
hi_current = yhi + (update->ntimestep - nfirst) * update->dt * rate;
}
// ncount = # of my atoms that overlap the insertion region
// nprevious = total of ncount across all procs
int ncount = 0;
for (i = 0; i < atom->nlocal; i++)
if (overlap(i)) ncount++;
int nprevious;
MPI_Allreduce(&ncount,&nprevious,1,MPI_INT,MPI_SUM,world);
// xmine is for my atoms
// xnear is for atoms from all procs + atoms to be inserted
double **xmine =
memory->create_2d_double_array(ncount,5,"fix_pour:xmine");
double **xnear =
memory->create_2d_double_array(nprevious+nnew*particles_per_insertion(),5,"fix_pour:xnear");
int nnear = nprevious;
// setup for allgatherv
int n = 5*ncount;
MPI_Allgather(&n,1,MPI_INT,recvcounts,1,MPI_INT,world);
displs[0] = 0;
for (int iproc = 1; iproc < nprocs; iproc++)
displs[iproc] = displs[iproc-1] + recvcounts[iproc-1];
// load up xmine array
double **x = atom->x;
double *radius = atom->radius;
ncount = 0;
for (i = 0; i < atom->nlocal; i++)
if (overlap(i)) {
xmine[ncount][0] = x[i][0];
xmine[ncount][1] = x[i][1];
xmine[ncount][2] = x[i][2];
xmine[ncount][3] = radius[i];
ncount++;
}
// perform allgatherv to acquire list of nearby particles on all procs
double *ptr = NULL;
if (ncount) ptr = xmine[0];
MPI_Allgatherv(ptr,5*ncount,MPI_DOUBLE,
xnear[0],recvcounts,displs,MPI_DOUBLE,world);
// insert new atoms into xnear list, one by one
// check against all nearby atoms and previously inserted ones
// if there is an overlap then try again at other z (3d) or y (2d) coord
// else insert by adding to xnear list
// max = maximum # of insertion attempts for all particles
// h = height, biased to give uniform distribution in time of insertion
int success;
double coord[3],radtmp,rn,h;
int attempt = 0;
int max = nnew * maxattempt;
int ntotal = nprevious+nnew;
while (nnear < ntotal) {
radtmp = rand_pour(radius_lo,radius_hi,radius_ran_style);
success = 0;
while (attempt < max) {
rn = random->uniform();
h = (hi_current-shift_randompos(radtmp)) - rn * (hi_current-lo_current-2.*shift_randompos(radtmp));
attempt++;
xyz_random(h,coord,radtmp);
for (i = 0; i < nnear; i++) {
if(overlaps_xnear_i(coord,radtmp,xnear,i)) break;
}
if (i == nnear) {
success = 1;
break;
}
}
if (success) {
nnear=insert_in_xnear(xnear,nnear,coord,radtmp);
} else break;
}
// warn if not all insertions were performed
ninserted += nnear-nprevious;
if (nnear - nprevious < nnew && me == 0)
error->warning("Less insertions than requested");
// check if new atom is in my sub-box or above it if I'm highest proc
// if so, add to my list via create_atom()
// initialize info about the atom
// type, diameter, density set from fix parameters
// group mask set to "all" plus fix group
// z velocity set to what velocity would be if particle
// had fallen from top of insertion region
// this gives continuous stream of atoms
// set npartner for new atom to 0 (assume not touching any others)
AtomVec *avec = atom->avec;
int j,m,flag;
double denstmp,vxtmp,vytmp,vztmp;
double g = grav;
//double g = 1.; //originally
double *sublo = domain->sublo;
double *subhi = domain->subhi;
int b_id;
int nfix = modify->nfix;
Fix **fix = modify->fix;
for (i = nprevious; i < nnear; i++) {
coord[0] = xnear[i][0];
coord[1] = xnear[i][1];
coord[2] = xnear[i][2];
radtmp = xnear[i][3];
b_id = static_cast<int>(xnear[i][4]);
denstmp = rand_pour(density_lo,density_hi,density_ran_style)*density_scaling();
calc_insert_velocities(i,g,xnear,vxtmp,vytmp,vztmp);
flag = 0;
if (coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1] &&
coord[2] >= sublo[2] && coord[2] < subhi[2]) flag = 1;
else if (domain->dimension == 3 && coord[2] >= domain->boxhi[2] &&
comm->myloc[2] == comm->procgrid[2]-1 &&
coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1]) flag = 1;
else if (domain->dimension == 2 && coord[1] >= domain->boxhi[1] &&
comm->myloc[1] == comm->procgrid[1]-1 &&
coord[0] >= sublo[0] && coord[0] < subhi[0]) flag = 1;
if (flag) {
avec->create_atom(ntype,coord);
m = atom->nlocal - 1;
atom->type[m] = ntype;
atom->radius[m] = radtmp;
atom->density[m] = denstmp;
atom->rmass[m] = 4.0*PI/3.0 * radtmp*radtmp*radtmp * denstmp;
atom->mask[m] = 1 | groupbit;
atom->v[m][0] = vxtmp;
atom->v[m][1] = vytmp;
atom->v[m][2] = vztmp;
for (j = 0; j < nfix; j++)
if (fix[j]->create_attribute) fix[j]->set_arrays(m);
set_body_props(i,m,coord,denstmp,radtmp,vxtmp,vytmp,vztmp,b_id);
}
}
// set tag # of new particles beyond all previous atoms
// reset global natoms
// if global map exists, reset it now instead of waiting for comm
// since deleting atoms messes up ghosts
if (atom->tag_enable) {
atom->tag_extend();
atom->natoms += nnear - nprevious;
if (atom->map_style) {
atom->nghost = 0;
atom->map_init();
atom->map_set();
}
}
finalize_insertion();
// free local memory
memory->destroy_2d_double_array(xmine);
memory->destroy_2d_double_array(xnear);
// next timestep to insert
if (ninserted < ninsert) next_reneighbor += nfreq;
else next_reneighbor = 0;
}
void WriteRestart::write(char *file)
{
// special case where reneighboring is not done in integrator
// on timestep restart file is written (due to build_once being set)
// if box is changing, must be reset, else restart file will have
// wrong box size and atoms will be lost when restart file is read
// other calls to pbc and domain and comm are not made,
// b/c they only make sense if reneighboring is actually performed
if (neighbor->build_once) domain->reset_box();
// natoms = sum of nlocal = value to write into restart file
// if unequal and thermo lostflag is "error", don't write restart file
bigint nblocal = atom->nlocal;
if(region)
{
nblocal = 0.;
for (int i = 0; i < atom->nlocal; i++)
if(region->match(atom->x[i][0],atom->x[i][1],atom->x[i][2]))
nblocal += 1;
}
MPI_Allreduce(&nblocal,&natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (natoms != atom->natoms && output->thermo->lostflag == ERROR && !region)
error->all(FLERR,"Atom count is inconsistent, cannot write restart file");
// check if filename contains "%"
int multiproc;
if (strchr(file,'%')) multiproc = 1;
else multiproc = 0;
// open single restart file or base file for multiproc case
if (me == 0) {
char *hfile;
if (multiproc) {
hfile = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(hfile,"%s%s%s",file,"base",ptr+1);
*ptr = '%';
} else hfile = file;
fp = fopen(hfile,"wb");
if (fp == NULL) {
char str[512];
sprintf(str,"Cannot open restart file %s",hfile);
error->one(FLERR,str);
}
if (multiproc) delete [] hfile;
}
// proc 0 writes header, groups, ntype-length arrays, force field
// all procs write fix info
if (me == 0) {
header();
group->write_restart(fp);
type_arrays();
force_fields();
}
modify->write_restart(fp);
// communication buffer for all my atom's info
// max_size = largest buffer needed by any proc
int max_size;
int send_size = atom->avec->size_restart();
MPI_Allreduce(&send_size,&max_size,1,MPI_INT,MPI_MAX,world);
double *buf;
if (me == 0) memory->create(buf,max_size,"write_restart:buf");
else memory->create(buf,send_size,"write_restart:buf");
// pack my atom data into buf
AtomVec *avec = atom->avec;
int n = 0;
if(!region)
{
for (int i = 0; i < atom->nlocal; i++) n += avec->pack_restart(i,&buf[n]);
}
else
{
for (int i = 0; i < atom->nlocal; i++)
if(region->match(atom->x[i][0],atom->x[i][1],atom->x[i][2]))
n += avec->pack_restart(i,&buf[n]);
send_size = n;
}
// if any fix requires it, remap each atom's coords via PBC
// is because fix changes atom coords (excepting an integrate fix)
// just remap in buffer, not actual atoms
if (modify->restart_pbc_any) {
int triclinic = domain->triclinic;
double *lo,*hi,*period;
if (triclinic == 0) {
lo = domain->boxlo;
hi = domain->boxhi;
period = domain->prd;
} else {
lo = domain->boxlo_lamda;
hi = domain->boxhi_lamda;
period = domain->prd_lamda;
}
int xperiodic = domain->xperiodic;
int yperiodic = domain->yperiodic;
int zperiodic = domain->zperiodic;
double *x;
if(region) error->all(FLERR,"have to implement more here");
int m = 0;
for (int i = 0; i < atom->nlocal; i++) {
x = &buf[m+1];
if (triclinic) domain->x2lamda(x,x);
if (xperiodic) {
if (x[0] < lo[0]) x[0] += period[0];
if (x[0] >= hi[0]) x[0] -= period[0];
x[0] = MAX(x[0],lo[0]);
}
if (yperiodic) {
if (x[1] < lo[1]) x[1] += period[1];
if (x[1] >= hi[1]) x[1] -= period[1];
x[1] = MAX(x[1],lo[1]);
}
if (zperiodic) {
if (x[2] < lo[2]) x[2] += period[2];
if (x[2] >= hi[2]) x[2] -= period[2];
x[2] = MAX(x[2],lo[2]);
}
if (triclinic) domain->lamda2x(x,x);
m += static_cast<int> (buf[m]);
}
}
// if single file:
// write one chunk of atoms per proc to file
// proc 0 pings each proc, receives its chunk, writes to file
// all other procs wait for ping, send their chunk to proc 0
// else if one file per proc:
// each proc opens its own file and writes its chunk directly
if (multiproc == 0) {
int tmp,recv_size;
MPI_Status status;
MPI_Request request;
if (me == 0) {
for (int iproc = 0; iproc < nprocs; iproc++) {
if (iproc) {
MPI_Irecv(buf,max_size,MPI_DOUBLE,iproc,0,world,&request);
MPI_Send(&tmp,0,MPI_INT,iproc,0,world);
MPI_Wait(&request,&status);
MPI_Get_count(&status,MPI_DOUBLE,&recv_size);
} else recv_size = send_size;
fwrite(&recv_size,sizeof(int),1,fp);
fwrite(buf,sizeof(double),recv_size,fp);
}
fclose(fp);
} else {
MPI_Recv(&tmp,0,MPI_INT,0,0,world,&status);
MPI_Rsend(buf,send_size,MPI_DOUBLE,0,0,world);
}
} else {
if (me == 0) fclose(fp);
char *perproc = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(perproc,"%s%d%s",file,me,ptr+1);
*ptr = '%';
fp = fopen(perproc,"wb");
if (fp == NULL) {
char str[512];
sprintf(str,"Cannot open restart file %s",perproc);
error->one(FLERR,str);
}
delete [] perproc;
fwrite(&send_size,sizeof(int),1,fp);
fwrite(buf,sizeof(double),send_size,fp);
fclose(fp);
}
memory->destroy(buf);
// invoke any fixes that write their own restart file
for (int ifix = 0; ifix < modify->nfix; ifix++)
if (modify->fix[ifix]->restart_file)
modify->fix[ifix]->write_restart_file(file);
}
void DeleteAtoms::command(int narg, char **arg)
{
if (domain->box_exist == 0)
error->all(FLERR,"Delete_atoms command before simulation box is defined");
if (narg < 1) error->all(FLERR,"Illegal delete_atoms command");
if (atom->tag_enable == 0)
error->all(FLERR,"Cannot use delete_atoms unless atoms have IDs");
// store state before delete
bigint natoms_previous = atom->natoms;
if(modify->n_fixes_style_strict("contacthistory") > 0)
modify->find_fix_style_strict("contacthistory",0)->pre_exchange();
if(modify->n_fixes_style_strict("bond/propagate/gran") > 0)
modify->find_fix_style_strict("bond/propagate/gran",0)->pre_exchange();
// delete the atoms
if (strcmp(arg[0],"group") == 0) delete_group(narg,arg);
else if (strcmp(arg[0],"region") == 0) delete_region(narg,arg);
else if (strcmp(arg[0],"overlap") == 0) delete_overlap(narg,arg);
else if (strcmp(arg[0],"porosity") == 0) delete_porosity(narg,arg);
else error->all(FLERR,"Illegal delete_atoms command");
// delete local atoms flagged in dlist
// reset nlocal
AtomVec *avec = atom->avec;
int nlocal = atom->nlocal;
int i = 0;
while (i < nlocal) {
if (dlist[i]) {
avec->copy(nlocal-1,i,1);
dlist[i] = dlist[nlocal-1];
nlocal--;
} else i++;
}
atom->nlocal = nlocal;
memory->destroy(dlist);
// if non-molecular system and compress flag set,
// reset atom tags to be contiguous
// set all atom IDs to 0, call tag_extend()
if (atom->molecular == 0 && compress_flag) {
int *tag = atom->tag;
for (i = 0; i < nlocal; i++) tag[i] = 0;
atom->tag_extend();
}
// reset atom->natoms
// reset atom->map if it exists
// set nghost to 0 so old ghosts of deleted atoms won't be mapped
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&atom->natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (atom->map_style) {
atom->nghost = 0;
atom->map_init();
atom->map_set();
}
// print before and after atom count
bigint ndelete = natoms_previous - atom->natoms;
if (comm->me == 0) {
if (screen) fprintf(screen,"Deleted " BIGINT_FORMAT
" atoms, new total = " BIGINT_FORMAT "\n",
ndelete,atom->natoms);
if (logfile) fprintf(logfile,"Deleted " BIGINT_FORMAT
" atoms, new total = " BIGINT_FORMAT "\n",
ndelete,atom->natoms);
}
}
void DeleteAtoms::command(int narg, char **arg)
{
if (domain->box_exist == 0)
error->all(FLERR,"Delete_atoms command before simulation box is defined");
if (narg < 1) error->all(FLERR,"Illegal delete_atoms command");
if (atom->tag_enable == 0)
error->all(FLERR,"Cannot use delete_atoms unless atoms have IDs");
// store state before delete
bigint natoms_previous = atom->natoms;
bigint nbonds_previous = atom->nbonds;
bigint nangles_previous = atom->nangles;
bigint ndihedrals_previous = atom->ndihedrals;
bigint nimpropers_previous = atom->nimpropers;
// delete the atoms
if (strcmp(arg[0],"group") == 0) delete_group(narg,arg);
else if (strcmp(arg[0],"region") == 0) delete_region(narg,arg);
else if (strcmp(arg[0],"overlap") == 0) delete_overlap(narg,arg);
else if (strcmp(arg[0],"porosity") == 0) delete_porosity(narg,arg);
else error->all(FLERR,"Illegal delete_atoms command");
// optionally delete additional bonds or atoms in molecules
if (bond_flag) delete_bond();
if (mol_flag) delete_molecule();
// delete local atoms flagged in dlist
// reset nlocal
AtomVec *avec = atom->avec;
int nlocal = atom->nlocal;
int i = 0;
while (i < nlocal) {
if (dlist[i]) {
avec->copy(nlocal-1,i,1);
dlist[i] = dlist[nlocal-1];
nlocal--;
} else i++;
}
atom->nlocal = nlocal;
memory->destroy(dlist);
// if non-molecular system and compress flag set,
// reset atom tags to be contiguous
// set all atom IDs to 0, call tag_extend()
if (atom->molecular == 0 && compress_flag) {
tagint *tag = atom->tag;
for (i = 0; i < nlocal; i++) tag[i] = 0;
atom->tag_extend();
}
// reset atom->natoms and also topology counts
// reset atom->map if it exists
// set nghost to 0 so old ghosts of deleted atoms won't be mapped
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&atom->natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (atom->map_style) {
atom->nghost = 0;
atom->map_init();
atom->map_set();
}
recount_topology();
// print before and after atom and topology counts
bigint ndelete = natoms_previous - atom->natoms;
bigint ndelete_bonds = nbonds_previous - atom->nbonds;
bigint ndelete_angles = nangles_previous - atom->nangles;
bigint ndelete_dihedrals = ndihedrals_previous - atom->ndihedrals;
bigint ndelete_impropers = nimpropers_previous - atom->nimpropers;
if (comm->me == 0) {
if (screen) {
fprintf(screen,"Deleted " BIGINT_FORMAT
" atoms, new total = " BIGINT_FORMAT "\n",
ndelete,atom->natoms);
if (bond_flag || mol_flag) {
if (nbonds_previous)
fprintf(screen,"Deleted " BIGINT_FORMAT
" bonds, new total = " BIGINT_FORMAT "\n",
ndelete_bonds,atom->nbonds);
if (nangles_previous)
fprintf(screen,"Deleted " BIGINT_FORMAT
" angles, new total = " BIGINT_FORMAT "\n",
ndelete_angles,atom->nangles);
if (ndihedrals_previous)
fprintf(screen,"Deleted " BIGINT_FORMAT
" dihedrals, new total = " BIGINT_FORMAT "\n",
ndelete_dihedrals,atom->ndihedrals);
if (nimpropers_previous)
fprintf(screen,"Deleted " BIGINT_FORMAT
" impropers, new total = " BIGINT_FORMAT "\n",
ndelete_impropers,atom->nimpropers);
}
}
if (logfile) {
fprintf(logfile,"Deleted " BIGINT_FORMAT
" atoms, new total = " BIGINT_FORMAT "\n",
ndelete,atom->natoms);
if (bond_flag || mol_flag) {
if (nbonds_previous)
fprintf(logfile,"Deleted " BIGINT_FORMAT
" bonds, new total = " BIGINT_FORMAT "\n",
ndelete_bonds,atom->nbonds);
if (nangles_previous)
fprintf(logfile,"Deleted " BIGINT_FORMAT
" angles, new total = " BIGINT_FORMAT "\n",
ndelete_angles,atom->nangles);
if (ndihedrals_previous)
fprintf(logfile,"Deleted " BIGINT_FORMAT
" dihedrals, new total = " BIGINT_FORMAT "\n",
ndelete_dihedrals,atom->ndihedrals);
if (nimpropers_previous)
fprintf(logfile,"Deleted " BIGINT_FORMAT
" impropers, new total = " BIGINT_FORMAT "\n",
ndelete_impropers,atom->nimpropers);
}
}
}
}
void FixSRP::post_run()
{
// all bond particles are removed after each run
// useful for write_data and write_restart commands
// since those commands occur between runs
bigint natoms_previous = atom->natoms;
int nlocal = atom->nlocal;
int* dlist;
memory->create(dlist,nlocal,"fix_srp:dlist");
for (int i = 0; i < nlocal; i++){
if(atom->type[i] == bptype)
dlist[i] = 1;
else
dlist[i] = 0;
}
// delete local atoms flagged in dlist
// reset nlocal
AtomVec *avec = atom->avec;
int i = 0;
while (i < nlocal) {
if (dlist[i]) {
avec->copy(nlocal-1,i,1);
dlist[i] = dlist[nlocal-1];
nlocal--;
} else i++;
}
atom->nlocal = nlocal;
memory->destroy(dlist);
// reset atom->natoms
// reset atom->map if it exists
// set nghost to 0 so old ghosts won't be mapped
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&atom->natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (atom->map_style) {
atom->nghost = 0;
atom->map_init();
atom->map_set();
}
// print before and after atom count
bigint ndelete = natoms_previous - atom->natoms;
if (comm->me == 0) {
if (screen) fprintf(screen,"Deleted " BIGINT_FORMAT
" atoms, new total = " BIGINT_FORMAT "\n",
ndelete,atom->natoms);
if (logfile) fprintf(logfile,"Deleted " BIGINT_FORMAT
" atoms, new total = " BIGINT_FORMAT "\n",
ndelete,atom->natoms);
}
// verlet calls box_too_small_check() in post_run
// this check maps all bond partners
// therefore need ghosts
// need to convert to lambda coords before apply pbc
if (domain->triclinic) domain->x2lamda(atom->nlocal);
domain->pbc();
comm->setup();
comm->exchange();
if (atom->sortfreq > 0) atom->sort();
comm->borders();
// change back to box coordinates
if (domain->triclinic) domain->lamda2x(atom->nlocal+atom->nghost);
}
void CommCuda::forward_comm_pack_cuda()
{
static int count=0;
static double kerneltime=0.0;
static double copytime=0.0;
timespec time1,time2,time3;
int n; // initialize comm buffers & exchange memory
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
double **x = atom->x;
cuda->shared_data.domain.xy=domain->xy;
cuda->shared_data.domain.xz=domain->xz;
cuda->shared_data.domain.yz=domain->yz;
cuda->shared_data.domain.prd[0]=domain->prd[0];
cuda->shared_data.domain.prd[1]=domain->prd[1];
cuda->shared_data.domain.prd[2]=domain->prd[2];
cuda->shared_data.domain.triclinic=domain->triclinic;
if(not comm_x_only && not avec->cudable) cuda->downloadAll(); //if not comm_x_only the communication routine of the atom_vec style class is used
// exchange data with another proc
// if other proc is self, just copy
// if comm_x_only set, exchange or copy directly to x, don't unpack
for (int iswap = 0; iswap < nswap; iswap++) {
if (sendproc[iswap] != me)
{
if (comm_x_only)
{
clock_gettime(CLOCK_REALTIME,&time1);
// n = Cuda_CommCuda_PackComm(&cuda->shared_data,sendnum[iswap],iswap,(void*) cuda->shared_data.comm.buf_send[iswap],pbc[iswap],pbc_flag[iswap]);
n = Cuda_CommCuda_PackComm(&cuda->shared_data,sendnum[iswap],iswap,(void*)buf_send,pbc[iswap],pbc_flag[iswap]);
clock_gettime(CLOCK_REALTIME,&time2);
if((sizeof(X_FLOAT)!=sizeof(double)) && n) //some complicated way to safe some transfer size if single precision is used
n=(n+1)*sizeof(X_FLOAT)/sizeof(double);
cuda->shared_data.comm.send_size[iswap]=n;
}
else if (ghost_velocity)
{
clock_gettime(CLOCK_REALTIME,&time1);
// n = Cuda_CommCuda_PackComm_Vel(&cuda->shared_data,sendnum[iswap],iswap,(void*) &buf_send[iswap*maxsend],pbc[iswap],pbc_flag[iswap]);
clock_gettime(CLOCK_REALTIME,&time2);
if((sizeof(X_FLOAT)!=sizeof(double)) && n) //some complicated way to safe some transfer size if single precision is used
n=(n+1)*sizeof(X_FLOAT)/sizeof(double);
cuda->shared_data.comm.send_size[iswap]=n;
}
else
{
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
if(avec->cudable)
n = avec->pack_comm(sendnum[iswap],&iswap,
cuda->shared_data.comm.buf_send[iswap],pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
cuda->shared_data.comm.buf_send[iswap],pbc_flag[iswap],pbc[iswap]);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_recv);
}
}
else //sendproc == me
{
if (comm_x_only)
{
if (sendnum[iswap])
{
n = Cuda_CommCuda_PackComm_Self(&cuda->shared_data,sendnum[iswap],iswap,firstrecv[iswap],pbc[iswap],pbc_flag[iswap]);
if(n<0) error->all(FLERR," # CUDA ERRROR on PackComm_Self");
if((sizeof(X_FLOAT)!=sizeof(double)) && n)
n=(n+1)*sizeof(X_FLOAT)/sizeof(double);
}
}
else if (ghost_velocity)
{
n = avec->pack_comm_vel(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
avec->unpack_comm_vel(recvnum[iswap],firstrecv[iswap],buf_send);
}
else
{
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_send);
}
}
}
if(not comm_x_only && not avec->cudable) cuda->uploadAll();
}
void FixPour::pre_exchange()
{
int i;
// just return if should not be called on this timestep
if (next_reneighbor != update->ntimestep) return;
// nnew = # to insert this timestep
int nnew = nper;
if (ninserted + nnew > ninsert) nnew = ninsert - ninserted;
// lo/hi current = z (or y) bounds of insertion region this timestep
if (domain->dimension == 3) {
lo_current = zlo + (update->ntimestep - nfirst) * update->dt * rate;
hi_current = zhi + (update->ntimestep - nfirst) * update->dt * rate;
} else {
lo_current = ylo + (update->ntimestep - nfirst) * update->dt * rate;
hi_current = yhi + (update->ntimestep - nfirst) * update->dt * rate;
}
// ncount = # of my atoms that overlap the insertion region
// nprevious = total of ncount across all procs
int ncount = 0;
for (i = 0; i < atom->nlocal; i++)
if (overlap(i)) ncount++;
int nprevious;
MPI_Allreduce(&ncount,&nprevious,1,MPI_INT,MPI_SUM,world);
// xmine is for my atoms
// xnear is for atoms from all procs + atoms to be inserted
double **xmine =
memory->create_2d_double_array(ncount,4,"fix_pour:xmine");
double **xnear =
memory->create_2d_double_array(nprevious+nnew,4,"fix_pour:xnear");
int nnear = nprevious;
// setup for allgatherv
int n = 4*ncount;
MPI_Allgather(&n,1,MPI_INT,recvcounts,1,MPI_INT,world);
displs[0] = 0;
for (int iproc = 1; iproc < nprocs; iproc++)
displs[iproc] = displs[iproc-1] + recvcounts[iproc-1];
// load up xmine array
double **x = atom->x;
double *radius = atom->radius;
ncount = 0;
for (i = 0; i < atom->nlocal; i++)
if (overlap(i)) {
xmine[ncount][0] = x[i][0];
xmine[ncount][1] = x[i][1];
xmine[ncount][2] = x[i][2];
xmine[ncount][3] = radius[i];
ncount++;
}
// perform allgatherv to acquire list of nearby particles on all procs
double *ptr = NULL;
if (ncount) ptr = xmine[0];
MPI_Allgatherv(ptr,4*ncount,MPI_DOUBLE,
xnear[0],recvcounts,displs,MPI_DOUBLE,world);
// insert new atoms into xnear list, one by one
// check against all nearby atoms and previously inserted ones
// if there is an overlap then try again at same z (3d) or y (2d) coord
// else insert by adding to xnear list
// max = maximum # of insertion attempts for all particles
// h = height, biased to give uniform distribution in time of insertion
int success;
double coord[3],radtmp,delx,dely,delz,rsq,radsum,rn,h;
int attempt = 0;
int max = nnew * maxattempt;
int ntotal = nprevious+nnew;
while (nnear < ntotal) {
rn = random->uniform();
h = hi_current - rn*rn * (hi_current-lo_current);
radtmp = radius_lo + random->uniform() * (radius_hi-radius_lo);
success = 0;
while (attempt < max) {
attempt++;
xyz_random(h,coord);
for (i = 0; i < nnear; i++) {
delx = coord[0] - xnear[i][0];
dely = coord[1] - xnear[i][1];
delz = coord[2] - xnear[i][2];
rsq = delx*delx + dely*dely + delz*delz;
radsum = radtmp + xnear[i][3];
if (rsq <= radsum*radsum) break;
}
if (i == nnear) {
success = 1;
break;
}
}
if (success) {
xnear[nnear][0] = coord[0];
xnear[nnear][1] = coord[1];
xnear[nnear][2] = coord[2];
xnear[nnear][3] = radtmp;
nnear++;
} else break;
}
// warn if not all insertions were performed
ninserted += nnear-nprevious;
if (nnear - nprevious < nnew && me == 0)
error->warning("Less insertions than requested");
// check if new atom is in my sub-box or above it if I'm highest proc
// if so, add to my list via create_atom()
// initialize info about the atom
// type, diameter, density set from fix parameters
// group mask set to "all" plus fix group
// z velocity set to what velocity would be if particle
// had fallen from top of insertion region
// this gives continuous stream of atoms
// set npartner for new atom to 0 (assume not touching any others)
AtomVec *avec = atom->avec;
int m,flag;
double denstmp,vxtmp,vytmp,vztmp;
double g = 1.0;
double *sublo = domain->sublo;
double *subhi = domain->subhi;
for (i = nprevious; i < nnear; i++) {
coord[0] = xnear[i][0];
coord[1] = xnear[i][1];
coord[2] = xnear[i][2];
radtmp = xnear[i][3];
denstmp = density_lo + random->uniform() * (density_hi-density_lo);
if (domain->dimension == 3) {
vxtmp = vxlo + random->uniform() * (vxhi-vxlo);
vytmp = vylo + random->uniform() * (vyhi-vylo);
vztmp = vz - sqrt(2.0*g*(hi_current-coord[2]));
} else {
vxtmp = vxlo + random->uniform() * (vxhi-vxlo);
vytmp = vy - sqrt(2.0*g*(hi_current-coord[1]));
vztmp = 0.0;
}
flag = 0;
if (coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1] &&
coord[2] >= sublo[2] && coord[2] < subhi[2]) flag = 1;
else if (domain->dimension == 3 && coord[2] >= domain->boxhi[2] &&
comm->myloc[2] == comm->procgrid[2]-1 &&
coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1]) flag = 1;
else if (domain->dimension == 2 && coord[1] >= domain->boxhi[1] &&
comm->myloc[1] == comm->procgrid[1]-1 &&
coord[0] >= sublo[0] && coord[0] < subhi[0]) flag = 1;
if (flag) {
avec->create_atom(ntype,coord);
m = atom->nlocal - 1;
atom->type[m] = ntype;
atom->radius[m] = radtmp;
atom->density[m] = denstmp;
if (domain->dimension == 3)
atom->rmass[m] = 4.0*PI/3.0 * radtmp*radtmp*radtmp * denstmp;
else
atom->rmass[m] = PI * radtmp*radtmp * denstmp;
atom->mask[m] = 1 | groupbit;
atom->v[m][0] = vxtmp;
atom->v[m][1] = vytmp;
atom->v[m][2] = vztmp;
if (fix_history) fix_history->npartner[m] = 0;
}
}
// set tag # of new particles beyond all previous atoms
// reset global natoms
// if global map exists, reset it
if (atom->tag_enable) {
atom->tag_extend();
atom->natoms += nnear - nprevious;
if (atom->map_style) {
atom->map_init();
atom->map_set();
}
}
// free local memory
memory->destroy_2d_double_array(xmine);
memory->destroy_2d_double_array(xnear);
// next timestep to insert
if (ninserted < ninsert) next_reneighbor += nfreq;
else next_reneighbor = 0;
}
void Replicate::command(int narg, char **arg)
{
int i,j,m,n;
if (domain->box_exist == 0)
error->all("Replicate command before simulation box is defined");
if (narg != 3) error->all("Illegal replicate command");
int me = comm->me;
int nprocs = comm->nprocs;
if (me == 0 && screen) fprintf(screen,"Replicating atoms ...\n");
// nrep = total # of replications
int nx = atoi(arg[0]);
int ny = atoi(arg[1]);
int nz = atoi(arg[2]);
int nrep = nx*ny*nz;
// error and warning checks
if (nx <= 0 || ny <= 0 || nz <= 0) error->all("Illegal replicate command");
if (domain->dimension == 2 && nz != 1)
error->all("Cannot replicate 2d simulation in z dimension");
if ((nx > 1 && domain->xperiodic == 0) ||
(ny > 1 && domain->yperiodic == 0) ||
(nz > 1 && domain->zperiodic == 0))
error->warning("Replicating in a non-periodic dimension");
if (atom->nextra_grow || atom->nextra_restart || atom->nextra_store)
error->all("Cannot replicate with fixes that store atom quantities");
// maxtag = largest atom tag across all existing atoms
int maxtag = 0;
for (i = 0; i < atom->nlocal; i++) maxtag = MAX(atom->tag[i],maxtag);
int maxtag_all;
MPI_Allreduce(&maxtag,&maxtag_all,1,MPI_INT,MPI_MAX,world);
maxtag = maxtag_all;
// maxmol = largest molecule tag across all existing atoms
int maxmol = 0;
if (atom->molecular) {
for (i = 0; i < atom->nlocal; i++) maxmol = MAX(atom->molecule[i],maxmol);
int maxmol_all;
MPI_Allreduce(&maxmol,&maxmol_all,1,MPI_INT,MPI_MAX,world);
maxmol = maxmol_all;
}
// unmap existing atoms via image flags
for (i = 0; i < atom->nlocal; i++)
domain->unmap(atom->x[i],atom->image[i]);
// communication buffer for all my atom's info
// max_size = largest buffer needed by any proc
// must do before new Atom class created,
// since size_restart() uses atom->nlocal
int max_size;
int send_size = atom->avec->size_restart();
MPI_Allreduce(&send_size,&max_size,1,MPI_INT,MPI_MAX,world);
double *buf =
(double *) memory->smalloc(max_size*sizeof(double),"replicate:buf");
// old = original atom class
// atom = new replicated atom class
// if old atom style was hybrid, pass sub-style names to create_avec
Atom *old = atom;
atom = new Atom(lmp);
int nstyles = 0;
char **keywords = NULL;
if (strcmp(old->atom_style,"hybrid") == 0) {
AtomVecHybrid *avec_hybrid = (AtomVecHybrid *) old->avec;
nstyles = avec_hybrid->nstyles;
keywords = avec_hybrid->keywords;
}
atom->create_avec(old->atom_style,nstyles,keywords);
// check that new problem size will not be too large
// if N > 2^31, turn off tags
// if molecular, N/Nbonds/etc cannot be > 2^31 else tags/counts invalid
double rep = nrep;
if (rep*old->natoms > MAXATOMS) atom->tag_enable = 0;
if (atom->molecular) {
if (rep*old->natoms > MAXATOMS || rep*old->nbonds > MAXATOMS ||
rep*old->nangles > MAXATOMS || rep*old->ndihedrals > MAXATOMS ||
rep*old->nimpropers > MAXATOMS)
error->all("Too big a problem to replicate with molecular atom style");
}
// assign atom and topology counts in new class from old one
atom->natoms = old->natoms * nrep;
atom->nbonds = old->nbonds * nrep;
atom->nangles = old->nangles * nrep;
atom->ndihedrals = old->ndihedrals * nrep;
atom->nimpropers = old->nimpropers * nrep;
atom->ntypes = old->ntypes;
atom->nbondtypes = old->nbondtypes;
atom->nangletypes = old->nangletypes;
atom->ndihedraltypes = old->ndihedraltypes;
atom->nimpropertypes = old->nimpropertypes;
atom->bond_per_atom = old->bond_per_atom;
atom->angle_per_atom = old->angle_per_atom;
atom->dihedral_per_atom = old->dihedral_per_atom;
atom->improper_per_atom = old->improper_per_atom;
// store old simulation box
int triclinic = domain->triclinic;
double old_xprd = domain->xprd;
double old_yprd = domain->yprd;
double old_zprd = domain->zprd;
double old_xy = domain->xy;
double old_xz = domain->xz;
double old_yz = domain->yz;
// setup new simulation box
domain->boxhi[0] = domain->boxlo[0] + nx*old_xprd;
domain->boxhi[1] = domain->boxlo[1] + ny*old_yprd;
domain->boxhi[2] = domain->boxlo[2] + nz*old_zprd;
if (triclinic) {
domain->xy *= ny;
domain->xz *= nz;
domain->yz *= nz;
}
// new problem setup using new box boundaries
if (nprocs == 1) n = static_cast<int> (atom->natoms);
else n = static_cast<int> (LB_FACTOR * atom->natoms / nprocs);
atom->allocate_type_arrays();
atom->avec->grow(n);
n = atom->nmax;
domain->print_box(" ");
domain->set_initial_box();
domain->set_global_box();
comm->set_procs();
domain->set_local_box();
// copy type arrays to new atom class
if (atom->mass) {
for (int itype = 1; itype <= atom->ntypes; itype++) {
atom->mass_setflag[itype] = old->mass_setflag[itype];
if (atom->mass_setflag[itype]) atom->mass[itype] = old->mass[itype];
}
}
if (atom->dipole) {
for (int itype = 1; itype <= atom->ntypes; itype++) {
atom->dipole_setflag[itype] = old->dipole_setflag[itype];
if (atom->dipole_setflag[itype])
atom->dipole[itype] = old->dipole[itype];
}
}
// set bounds for my proc
// if periodic and I am lo/hi proc, adjust bounds by EPSILON
// insures all replicated atoms will be owned even with round-off
double sublo[3],subhi[3];
if (triclinic == 0) {
sublo[0] = domain->sublo[0]; subhi[0] = domain->subhi[0];
sublo[1] = domain->sublo[1]; subhi[1] = domain->subhi[1];
sublo[2] = domain->sublo[2]; subhi[2] = domain->subhi[2];
} else {
sublo[0] = domain->sublo_lamda[0]; subhi[0] = domain->subhi_lamda[0];
sublo[1] = domain->sublo_lamda[1]; subhi[1] = domain->subhi_lamda[1];
sublo[2] = domain->sublo_lamda[2]; subhi[2] = domain->subhi_lamda[2];
}
if (domain->xperiodic) {
if (comm->myloc[0] == 0) sublo[0] -= EPSILON;
if (comm->myloc[0] == comm->procgrid[0]-1) subhi[0] += EPSILON;
}
if (domain->yperiodic) {
if (comm->myloc[1] == 0) sublo[1] -= EPSILON;
if (comm->myloc[1] == comm->procgrid[1]-1) subhi[1] += EPSILON;
}
if (domain->zperiodic) {
if (comm->myloc[2] == 0) sublo[2] -= EPSILON;
if (comm->myloc[2] == comm->procgrid[2]-1) subhi[2] += EPSILON;
}
// loop over all procs
// if this iteration of loop is me:
// pack my unmapped atom data into buf
// bcast it to all other procs
// performs 3d replicate loop with while loop over atoms in buf
// x = new replicated position, remapped into simulation box
// unpack atom into new atom class from buf if I own it
// adjust tag, mol #, coord, topology info as needed
AtomVec *old_avec = old->avec;
AtomVec *avec = atom->avec;
int ix,iy,iz,image,atom_offset,mol_offset;
double x[3],lamda[3];
double *coord;
int tag_enable = atom->tag_enable;
for (int iproc = 0; iproc < nprocs; iproc++) {
if (me == iproc) {
n = 0;
for (i = 0; i < old->nlocal; i++) n += old_avec->pack_restart(i,&buf[n]);
}
MPI_Bcast(&n,1,MPI_INT,iproc,world);
MPI_Bcast(buf,n,MPI_DOUBLE,iproc,world);
for (ix = 0; ix < nx; ix++) {
for (iy = 0; iy < ny; iy++) {
for (iz = 0; iz < nz; iz++) {
// while loop over one proc's atom list
m = 0;
while (m < n) {
image = (512 << 20) | (512 << 10) | 512;
if (triclinic == 0) {
x[0] = buf[m+1] + ix*old_xprd;
x[1] = buf[m+2] + iy*old_yprd;
x[2] = buf[m+3] + iz*old_zprd;
} else {
x[0] = buf[m+1] + ix*old_xprd + iy*old_xy + iz*old_xz;
x[1] = buf[m+2] + iy*old_yprd + iz*old_yz;
x[2] = buf[m+3] + iz*old_zprd;
}
domain->remap(x,image);
if (triclinic) {
domain->x2lamda(x,lamda);
coord = lamda;
} else coord = x;
if (coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1] &&
coord[2] >= sublo[2] && coord[2] < subhi[2]) {
m += avec->unpack_restart(&buf[m]);
i = atom->nlocal - 1;
if (tag_enable)
atom_offset = iz*ny*nx*maxtag + iy*nx*maxtag + ix*maxtag;
else atom_offset = 0;
mol_offset = iz*ny*nx*maxmol + iy*nx*maxmol + ix*maxmol;
atom->x[i][0] = x[0];
atom->x[i][1] = x[1];
atom->x[i][2] = x[2];
atom->tag[i] += atom_offset;
atom->image[i] = image;
if (atom->molecular) {
if (atom->molecule[i] > 0)
atom->molecule[i] += mol_offset;
if (atom->avec->bonds_allow)
for (j = 0; j < atom->num_bond[i]; j++)
atom->bond_atom[i][j] += atom_offset;
if (atom->avec->angles_allow)
for (j = 0; j < atom->num_angle[i]; j++) {
atom->angle_atom1[i][j] += atom_offset;
atom->angle_atom2[i][j] += atom_offset;
atom->angle_atom3[i][j] += atom_offset;
}
if (atom->avec->dihedrals_allow)
for (j = 0; j < atom->num_dihedral[i]; j++) {
atom->dihedral_atom1[i][j] += atom_offset;
atom->dihedral_atom2[i][j] += atom_offset;
atom->dihedral_atom3[i][j] += atom_offset;
atom->dihedral_atom4[i][j] += atom_offset;
}
if (atom->avec->impropers_allow)
for (j = 0; j < atom->num_improper[i]; j++) {
atom->improper_atom1[i][j] += atom_offset;
atom->improper_atom2[i][j] += atom_offset;
atom->improper_atom3[i][j] += atom_offset;
atom->improper_atom4[i][j] += atom_offset;
}
}
} else m += static_cast<int> (buf[m]);
}
}
}
}
} // end of proc loop
// free communication buffer and old atom class
memory->sfree(buf);
delete old;
// check that all atoms were assigned to procs
double natoms;
double rlocal = atom->nlocal;
MPI_Allreduce(&rlocal,&natoms,1,MPI_DOUBLE,MPI_SUM,world);
if (me == 0) {
if (screen) fprintf(screen," %.15g atoms\n",natoms);
if (logfile) fprintf(logfile," %.15g atoms\n",natoms);
}
if (natoms != atom->natoms)
error->all("Replicate did not assign all atoms correctly");
if (me == 0) {
if (atom->nbonds) {
if (screen) fprintf(screen," %d bonds\n",atom->nbonds);
if (logfile) fprintf(logfile," %d bonds\n",atom->nbonds);
}
if (atom->nangles) {
if (screen) fprintf(screen," %d angles\n",atom->nangles);
if (logfile) fprintf(logfile," %d angles\n",atom->nangles);
}
if (atom->ndihedrals) {
if (screen) fprintf(screen," %d dihedrals\n",atom->ndihedrals);
if (logfile) fprintf(logfile," %d dihedrals\n",atom->ndihedrals);
}
if (atom->nimpropers) {
if (screen) fprintf(screen," %d impropers\n",atom->nimpropers);
if (logfile) fprintf(logfile," %d impropers\n",atom->nimpropers);
}
}
// create global mapping and bond topology now that system is defined
if (atom->map_style) {
atom->map_init();
atom->map_set();
}
if (atom->molecular) {
Special special(lmp);
special.build();
}
}
void CommBrick::borders()
{
int i,n,itype,iswap,dim,ineed,twoneed;
int nsend,nrecv,sendflag,nfirst,nlast,ngroup;
double lo,hi;
int *type;
double **x;
double *buf,*mlo,*mhi;
MPI_Request request;
AtomVec *avec = atom->avec;
// do swaps over all 3 dimensions
iswap = 0;
smax = rmax = 0;
for (dim = 0; dim < 3; dim++) {
nlast = 0;
twoneed = 2*maxneed[dim];
for (ineed = 0; ineed < twoneed; ineed++) {
// find atoms within slab boundaries lo/hi using <= and >=
// check atoms between nfirst and nlast
// for first swaps in a dim, check owned and ghost
// for later swaps in a dim, only check newly arrived ghosts
// store sent atom indices in sendlist for use in future timesteps
x = atom->x;
if (mode == SINGLE) {
lo = slablo[iswap];
hi = slabhi[iswap];
} else {
type = atom->type;
mlo = multilo[iswap];
mhi = multihi[iswap];
}
if (ineed % 2 == 0) {
nfirst = nlast;
nlast = atom->nlocal + atom->nghost;
}
nsend = 0;
// sendflag = 0 if I do not send on this swap
// sendneed test indicates receiver no longer requires data
// e.g. due to non-PBC or non-uniform sub-domains
if (ineed/2 >= sendneed[dim][ineed % 2]) sendflag = 0;
else sendflag = 1;
// find send atoms according to SINGLE vs MULTI
// all atoms eligible versus only atoms in bordergroup
// can only limit loop to bordergroup for first sends (ineed < 2)
// on these sends, break loop in two: owned (in group) and ghost
if (sendflag) {
if (!bordergroup || ineed >= 2) {
if (mode == SINGLE) {
for (i = nfirst; i < nlast; i++)
if (x[i][dim] >= lo && x[i][dim] <= hi) {
if (nsend == maxsendlist[iswap]) grow_list(iswap,nsend);
sendlist[iswap][nsend++] = i;
}
} else {
for (i = nfirst; i < nlast; i++) {
itype = type[i];
if (x[i][dim] >= mlo[itype] && x[i][dim] <= mhi[itype]) {
if (nsend == maxsendlist[iswap]) grow_list(iswap,nsend);
sendlist[iswap][nsend++] = i;
}
}
}
} else {
if (mode == SINGLE) {
ngroup = atom->nfirst;
for (i = 0; i < ngroup; i++)
if (x[i][dim] >= lo && x[i][dim] <= hi) {
if (nsend == maxsendlist[iswap]) grow_list(iswap,nsend);
sendlist[iswap][nsend++] = i;
}
for (i = atom->nlocal; i < nlast; i++)
if (x[i][dim] >= lo && x[i][dim] <= hi) {
if (nsend == maxsendlist[iswap]) grow_list(iswap,nsend);
sendlist[iswap][nsend++] = i;
}
} else {
ngroup = atom->nfirst;
for (i = 0; i < ngroup; i++) {
itype = type[i];
if (x[i][dim] >= mlo[itype] && x[i][dim] <= mhi[itype]) {
if (nsend == maxsendlist[iswap]) grow_list(iswap,nsend);
sendlist[iswap][nsend++] = i;
}
}
for (i = atom->nlocal; i < nlast; i++) {
itype = type[i];
if (x[i][dim] >= mlo[itype] && x[i][dim] <= mhi[itype]) {
if (nsend == maxsendlist[iswap]) grow_list(iswap,nsend);
sendlist[iswap][nsend++] = i;
}
}
}
}
}
// pack up list of border atoms
if (nsend*size_border > maxsend) grow_send(nsend*size_border,0);
if (ghost_velocity)
n = avec->pack_border_vel(nsend,sendlist[iswap],buf_send,
pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_border(nsend,sendlist[iswap],buf_send,
pbc_flag[iswap],pbc[iswap]);
// swap atoms with other proc
// no MPI calls except SendRecv if nsend/nrecv = 0
// put incoming ghosts at end of my atom arrays
// if swapping with self, simply copy, no messages
if (sendproc[iswap] != me) {
MPI_Sendrecv(&nsend,1,MPI_INT,sendproc[iswap],0,
&nrecv,1,MPI_INT,recvproc[iswap],0,world,MPI_STATUS_IGNORE);
if (nrecv*size_border > maxrecv) grow_recv(nrecv*size_border);
if (nrecv) MPI_Irecv(buf_recv,nrecv*size_border,MPI_DOUBLE,
recvproc[iswap],0,world,&request);
if (n) MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
if (nrecv) MPI_Wait(&request,MPI_STATUS_IGNORE);
buf = buf_recv;
} else {
nrecv = nsend;
buf = buf_send;
}
// unpack buffer
if (ghost_velocity)
avec->unpack_border_vel(nrecv,atom->nlocal+atom->nghost,buf);
else
avec->unpack_border(nrecv,atom->nlocal+atom->nghost,buf);
// set all pointers & counters
smax = MAX(smax,nsend);
rmax = MAX(rmax,nrecv);
sendnum[iswap] = nsend;
recvnum[iswap] = nrecv;
size_forward_recv[iswap] = nrecv*size_forward;
size_reverse_send[iswap] = nrecv*size_reverse;
size_reverse_recv[iswap] = nsend*size_reverse;
firstrecv[iswap] = atom->nlocal + atom->nghost;
atom->nghost += nrecv;
iswap++;
}
}
// insure send/recv buffers are long enough for all forward & reverse comm
int max = MAX(maxforward*smax,maxreverse*rmax);
if (max > maxsend) grow_send(max,0);
max = MAX(maxforward*rmax,maxreverse*smax);
if (max > maxrecv) grow_recv(max);
// reset global->local map
if (map_style) atom->map_set();
}
void CommBrick::exchange()
{
int i,m,nsend,nrecv,nrecv1,nrecv2,nlocal;
double lo,hi,value;
double **x;
double *sublo,*subhi;
MPI_Request request;
AtomVec *avec = atom->avec;
// clear global->local map for owned and ghost atoms
// b/c atoms migrate to new procs in exchange() and
// new ghosts are created in borders()
// map_set() is done at end of borders()
// clear ghost count and any ghost bonus data internal to AtomVec
if (map_style) atom->map_clear();
atom->nghost = 0;
atom->avec->clear_bonus();
// insure send buf is large enough for single atom
// bufextra = max size of one atom = allowed overflow of sendbuf
// fixes can change per-atom size requirement on-the-fly
int bufextra_old = bufextra;
maxexchange = maxexchange_atom + maxexchange_fix;
bufextra = maxexchange + BUFEXTRA;
if (bufextra > bufextra_old)
memory->grow(buf_send,maxsend+bufextra,"comm:buf_send");
// subbox bounds for orthogonal or triclinic
if (triclinic == 0) {
sublo = domain->sublo;
subhi = domain->subhi;
} else {
sublo = domain->sublo_lamda;
subhi = domain->subhi_lamda;
}
// loop over dimensions
int dimension = domain->dimension;
for (int dim = 0; dim < dimension; dim++) {
// fill buffer with atoms leaving my box, using < and >=
// when atom is deleted, fill it in with last atom
x = atom->x;
lo = sublo[dim];
hi = subhi[dim];
nlocal = atom->nlocal;
i = nsend = 0;
while (i < nlocal) {
if (x[i][dim] < lo || x[i][dim] >= hi) {
if (nsend > maxsend) grow_send(nsend,1);
nsend += avec->pack_exchange(i,&buf_send[nsend]);
avec->copy(nlocal-1,i,1);
nlocal--;
} else i++;
}
atom->nlocal = nlocal;
// send/recv atoms in both directions
// send size of message first so receiver can realloc buf_recv if needed
// if 1 proc in dimension, no send/recv
// set nrecv = 0 so buf_send atoms will be lost
// if 2 procs in dimension, single send/recv
// if more than 2 procs in dimension, send/recv to both neighbors
if (procgrid[dim] == 1) nrecv = 0;
else {
MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][0],0,
&nrecv1,1,MPI_INT,procneigh[dim][1],0,world,MPI_STATUS_IGNORE);
nrecv = nrecv1;
if (procgrid[dim] > 2) {
MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][1],0,
&nrecv2,1,MPI_INT,procneigh[dim][0],0,world,MPI_STATUS_IGNORE);
nrecv += nrecv2;
}
if (nrecv > maxrecv) grow_recv(nrecv);
MPI_Irecv(buf_recv,nrecv1,MPI_DOUBLE,procneigh[dim][1],0,
world,&request);
MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][0],0,world);
MPI_Wait(&request,MPI_STATUS_IGNORE);
if (procgrid[dim] > 2) {
MPI_Irecv(&buf_recv[nrecv1],nrecv2,MPI_DOUBLE,procneigh[dim][0],0,
world,&request);
MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][1],0,world);
MPI_Wait(&request,MPI_STATUS_IGNORE);
}
}
// check incoming atoms to see if they are in my box
// if so, add to my list
// box check is only for this dimension,
// atom may be passed to another proc in later dims
m = 0;
while (m < nrecv) {
value = buf_recv[m+dim+1];
if (value >= lo && value < hi) m += avec->unpack_exchange(&buf_recv[m]);
else m += static_cast<int> (buf_recv[m]);
}
}
if (atom->firstgroupname) atom->first_reorder();
}
void CommBrick::forward_comm(int dummy)
{
int n;
MPI_Request request;
AtomVec *avec = atom->avec;
double **x = atom->x;
double *buf;
// exchange data with another proc
// if other proc is self, just copy
// if comm_x_only set, exchange or copy directly to x, don't unpack
for (int iswap = 0; iswap < nswap; iswap++) {
if (sendproc[iswap] != me) {
if (comm_x_only) {
if (size_forward_recv[iswap]) {
if (size_forward_recv[iswap]) buf = x[firstrecv[iswap]];
else buf = NULL;
MPI_Irecv(buf,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
}
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
if (n) MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
if (size_forward_recv[iswap]) MPI_Wait(&request,MPI_STATUS_IGNORE);
} else if (ghost_velocity) {
if (size_forward_recv[iswap])
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
n = avec->pack_comm_vel(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
if (n) MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
if (size_forward_recv[iswap]) MPI_Wait(&request,MPI_STATUS_IGNORE);
avec->unpack_comm_vel(recvnum[iswap],firstrecv[iswap],buf_recv);
} else {
if (size_forward_recv[iswap])
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
if (n) MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
if (size_forward_recv[iswap]) MPI_Wait(&request,MPI_STATUS_IGNORE);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_recv);
}
} else {
if (comm_x_only) {
if (sendnum[iswap])
avec->pack_comm(sendnum[iswap],sendlist[iswap],
x[firstrecv[iswap]],pbc_flag[iswap],pbc[iswap]);
} else if (ghost_velocity) {
avec->pack_comm_vel(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
avec->unpack_comm_vel(recvnum[iswap],firstrecv[iswap],buf_send);
} else {
avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_send);
}
}
}
}
void Irregular::migrate_atoms()
{
// clear global->local map since atoms move to new procs
// clear old ghosts so map_set() at end will operate only on local atoms
// exchange() doesn't need to clear ghosts b/c borders()
// is called right after and it clears ghosts and calls map_set()
if (map_style) atom->map_clear();
atom->nghost = 0;
atom->avec->clear_bonus();
// subbox bounds for orthogonal or triclinic box
// other comm/domain data used by coord2proc()
double *sublo,*subhi;
if (triclinic == 0) {
sublo = domain->sublo;
subhi = domain->subhi;
} else {
sublo = domain->sublo_lamda;
subhi = domain->subhi_lamda;
}
uniform = comm->uniform;
xsplit = comm->xsplit;
ysplit = comm->ysplit;
zsplit = comm->zsplit;
boxlo = domain->boxlo;
prd = domain->prd;
// loop over atoms, flag any that are not in my sub-box
// fill buffer with atoms leaving my box, using < and >=
// assign which proc it belongs to via coord2proc()
// if coord2proc() returns me, due to round-off
// in triclinic x2lamda(), then keep atom and don't send
// when atom is deleted, fill it in with last atom
AtomVec *avec = atom->avec;
double **x = atom->x;
int nlocal = atom->nlocal;
int nsend = 0;
int nsendatom = 0;
int *sizes = new int[nlocal];
int *proclist = new int[nlocal];
int igx,igy,igz;
int i = 0;
while (i < nlocal) {
if (x[i][0] < sublo[0] || x[i][0] >= subhi[0] ||
x[i][1] < sublo[1] || x[i][1] >= subhi[1] ||
x[i][2] < sublo[2] || x[i][2] >= subhi[2]) {
proclist[nsendatom] = coord2proc(x[i],igx,igy,igz);
if (proclist[nsendatom] != me) {
if (nsend > maxsend) grow_send(nsend,1);
sizes[nsendatom] = avec->pack_exchange(i,&buf_send[nsend]);
nsend += sizes[nsendatom];
nsendatom++;
avec->copy(nlocal-1,i,1);
nlocal--;
} else i++;
} else i++;
}
atom->nlocal = nlocal;
// create irregular communication plan, perform comm, destroy plan
// returned nrecv = size of buffer needed for incoming atoms
int nrecv = create_atom(nsendatom,sizes,proclist);
if (nrecv > maxrecv) grow_recv(nrecv);
exchange_atom(buf_send,sizes,buf_recv);
destroy_atom();
delete [] sizes;
delete [] proclist;
// add received atoms to my list
int m = 0;
while (m < nrecv) m += avec->unpack_exchange(&buf_recv[m]);
// reset global->local map
if (map_style) atom->map_set();
}
void ReadRestart::command(int narg, char **arg)
{
if (narg != 1) error->all(FLERR,"Illegal read_restart command");
if (domain->box_exist)
error->all(FLERR,"Cannot read_restart after simulation box is defined");
MPI_Comm_rank(world,&me);
MPI_Comm_size(world,&nprocs);
// if filename contains "*", search dir for latest restart file
char *file = new char[strlen(arg[0]) + 16];
if (strchr(arg[0],'*')) {
int n;
if (me == 0) {
file_search(arg[0],file);
n = strlen(file) + 1;
}
MPI_Bcast(&n,1,MPI_INT,0,world);
MPI_Bcast(file,n,MPI_CHAR,0,world);
} else strcpy(file,arg[0]);
// check if filename contains "%"
int multiproc;
if (strchr(file,'%')) multiproc = 1;
else multiproc = 0;
// open single restart file or base file for multiproc case
// auto-detect whether byte swapping needs to be done as file is read
if (me == 0) {
if (screen) fprintf(screen,"Reading restart file ...\n");
char *hfile;
if (multiproc) {
hfile = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(hfile,"%s%s%s",file,"base",ptr+1);
*ptr = '%';
} else hfile = file;
fp = fopen(hfile,"rb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",hfile);
error->one(FLERR,str);
}
swapflag = autodetect(&fp,hfile);
if (multiproc) delete [] hfile;
}
MPI_Bcast(&swapflag,1,MPI_INT,0,world);
// read header info and create atom style and simulation box
header();
domain->box_exist = 1;
// problem setup using info from header
int n;
if (nprocs == 1) n = static_cast<int> (atom->natoms);
else n = static_cast<int> (LB_FACTOR * atom->natoms / nprocs);
atom->allocate_type_arrays();
atom->avec->grow(n);
n = atom->nmax;
domain->print_box(" ");
domain->set_initial_box();
domain->set_global_box();
comm->set_proc_grid();
domain->set_local_box();
// read groups, ntype-length arrays, force field, fix info from file
// nextra = max # of extra quantities stored with each atom
group->read_restart(fp);
type_arrays();
force_fields();
int nextra = modify->read_restart(fp);
atom->nextra_store = nextra;
memory->create(atom->extra,n,nextra,"atom:extra");
// single file:
// nprocs_file = # of chunks in file
// proc 0 reads chunks one at a time and bcasts it to other procs
// each proc unpacks the atoms, saving ones in it's sub-domain
// check for atom in sub-domain differs for orthogonal vs triclinic box
// close restart file when done
AtomVec *avec = atom->avec;
int maxbuf = 0;
double *buf = NULL;
int m;
if (multiproc == 0) {
int triclinic = domain->triclinic;
double *x,lamda[3];
double *coord,*sublo,*subhi;
if (triclinic == 0) {
sublo = domain->sublo;
subhi = domain->subhi;
} else {
sublo = domain->sublo_lamda;
subhi = domain->subhi_lamda;
}
for (int iproc = 0; iproc < nprocs_file; iproc++) {
n = read_int();
if (n > maxbuf) {
maxbuf = n;
memory->destroy(buf);
memory->create(buf,maxbuf,"read_restart:buf");
}
if (n > 0) {
if (me == 0) nread_double(buf,n,fp);
MPI_Bcast(buf,n,MPI_DOUBLE,0,world);
}
m = 0;
while (m < n) {
x = &buf[m+1];
if (triclinic) {
domain->x2lamda(x,lamda);
coord = lamda;
} else coord = x;
if (coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1] &&
coord[2] >= sublo[2] && coord[2] < subhi[2]) {
m += avec->unpack_restart(&buf[m]);
}
else m += static_cast<int> (buf[m]);
}
}
if (me == 0) fclose(fp);
// one file per proc:
// nprocs_file = # of files
// each proc reads 1/P fraction of files, keeping all atoms in the files
// perform irregular comm to migrate atoms to correct procs
// close restart file when done
} else {
if (me == 0) fclose(fp);
char *perproc = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
for (int iproc = me; iproc < nprocs_file; iproc += nprocs) {
*ptr = '\0';
sprintf(perproc,"%s%d%s",file,iproc,ptr+1);
*ptr = '%';
fp = fopen(perproc,"rb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",perproc);
error->one(FLERR,str);
}
nread_int(&n,1,fp);
if (n > maxbuf) {
maxbuf = n;
memory->destroy(buf);
memory->create(buf,maxbuf,"read_restart:buf");
}
if (n > 0) nread_double(buf,n,fp);
m = 0;
while (m < n) m += avec->unpack_restart(&buf[m]);
fclose(fp);
}
delete [] perproc;
// create a temporary fix to hold and migrate extra atom info
// necessary b/c irregular will migrate atoms
if (nextra) {
char cextra[8],fixextra[8];
sprintf(cextra,"%d",nextra);
sprintf(fixextra,"%d",modify->nfix_restart_peratom);
char **newarg = new char*[5];
newarg[0] = (char *) "_read_restart";
newarg[1] = (char *) "all";
newarg[2] = (char *) "READ_RESTART";
newarg[3] = cextra;
newarg[4] = fixextra;
modify->add_fix(5,newarg);
delete [] newarg;
}
// move atoms to new processors via irregular()
// in case read by different proc than wrote restart file
// first do map_init() since irregular->migrate_atoms() will do map_clear()
if (atom->map_style) atom->map_init();
if (domain->triclinic) domain->x2lamda(atom->nlocal);
Irregular *irregular = new Irregular(lmp);
irregular->migrate_atoms();
delete irregular;
if (domain->triclinic) domain->lamda2x(atom->nlocal);
// put extra atom info held by fix back into atom->extra
// destroy temporary fix
if (nextra) {
memory->destroy(atom->extra);
memory->create(atom->extra,atom->nmax,nextra,"atom:extra");
int ifix = modify->find_fix("_read_restart");
FixReadRestart *fix = (FixReadRestart *) modify->fix[ifix];
int *count = fix->count;
double **extra = fix->extra;
double **atom_extra = atom->extra;
int nlocal = atom->nlocal;
for (int i = 0; i < nlocal; i++)
for (int j = 0; j < count[i]; j++)
atom_extra[i][j] = extra[i][j];
modify->delete_fix("_read_restart");
}
}
// clean-up memory
delete [] file;
memory->destroy(buf);
// check that all atoms were assigned to procs
bigint natoms;
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (me == 0) {
if (screen) fprintf(screen," " BIGINT_FORMAT " atoms\n",natoms);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " atoms\n",natoms);
}
if (natoms != atom->natoms)
error->all(FLERR,"Did not assign all atoms correctly");
if (me == 0) {
if (atom->nbonds) {
if (screen) fprintf(screen," " BIGINT_FORMAT " bonds\n",atom->nbonds);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " bonds\n",atom->nbonds);
}
if (atom->nangles) {
if (screen) fprintf(screen," " BIGINT_FORMAT " angles\n",
atom->nangles);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " angles\n",
atom->nangles);
}
if (atom->ndihedrals) {
if (screen) fprintf(screen," " BIGINT_FORMAT " dihedrals\n",
atom->ndihedrals);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " dihedrals\n",
atom->ndihedrals);
}
if (atom->nimpropers) {
if (screen) fprintf(screen," " BIGINT_FORMAT " impropers\n",
atom->nimpropers);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " impropers\n",
atom->nimpropers);
}
}
// check if tags are being used
// create global mapping and bond topology now that system is defined
int flag = 0;
for (int i = 0; i < atom->nlocal; i++)
if (atom->tag[i] > 0) flag = 1;
int flag_all;
MPI_Allreduce(&flag,&flag_all,1,MPI_INT,MPI_MAX,world);
if (atom->natoms > 0 && flag_all == 0) atom->tag_enable = 0;
if (atom->map_style) {
atom->map_init();
atom->map_set();
}
if (atom->molecular) {
Special special(lmp);
special.build();
}
}
void FixEvaporate::pre_exchange()
{
int i,j,m,iwhichglobal,iwhichlocal;
int ndel,ndeltopo[4];
if (update->ntimestep != next_reneighbor) return;
// grow list and mark arrays if necessary
if (atom->nmax > nmax) {
memory->destroy(list);
memory->destroy(mark);
nmax = atom->nmax;
memory->create(list,nmax,"evaporate:list");
memory->create(mark,nmax,"evaporate:mark");
}
// ncount = # of deletable atoms in region that I own
// nall = # on all procs
// nbefore = # on procs before me
// list[ncount] = list of local indices of atoms I can delete
Region *region = domain->regions[iregion];
region->prematch();
double **x = atom->x;
int *mask = atom->mask;
tagint *tag = atom->tag;
int nlocal = atom->nlocal;
int ncount = 0;
for (i = 0; i < nlocal; i++)
if (mask[i] & groupbit)
if (region->match(x[i][0],x[i][1],x[i][2])) list[ncount++] = i;
int nall,nbefore;
MPI_Allreduce(&ncount,&nall,1,MPI_INT,MPI_SUM,world);
MPI_Scan(&ncount,&nbefore,1,MPI_INT,MPI_SUM,world);
nbefore -= ncount;
// ndel = total # of atom deletions, in or out of region
// ndeltopo[1,2,3,4] = ditto for bonds, angles, dihedrals, impropers
// mark[] = 1 if deleted
ndel = 0;
for (i = 0; i < nlocal; i++) mark[i] = 0;
// atomic deletions
// choose atoms randomly across all procs and mark them for deletion
// shrink eligible list as my atoms get marked
// keep ndel,ncount,nall,nbefore current after each atom deletion
if (molflag == 0) {
while (nall && ndel < nflux) {
iwhichglobal = static_cast<int> (nall*random->uniform());
if (iwhichglobal < nbefore) nbefore--;
else if (iwhichglobal < nbefore + ncount) {
iwhichlocal = iwhichglobal - nbefore;
mark[list[iwhichlocal]] = 1;
list[iwhichlocal] = list[ncount-1];
ncount--;
}
ndel++;
nall--;
}
// molecule deletions
// choose one atom in one molecule randomly across all procs
// bcast mol ID and delete all atoms in that molecule on any proc
// update deletion count by total # of atoms in molecule
// shrink list of eligible candidates as any of my atoms get marked
// keep ndel,ndeltopo,ncount,nall,nbefore current after each mol deletion
} else {
int me,proc,iatom,ndelone,ndelall,index;
tagint imolecule;
tagint *molecule = atom->molecule;
int *molindex = atom->molindex;
int *molatom = atom->molatom;
int molecular = atom->molecular;
Molecule **onemols = atom->avec->onemols;
ndeltopo[0] = ndeltopo[1] = ndeltopo[2] = ndeltopo[3] = 0;
while (nall && ndel < nflux) {
// pick an iatom,imolecule on proc me to delete
iwhichglobal = static_cast<int> (nall*random->uniform());
if (iwhichglobal >= nbefore && iwhichglobal < nbefore + ncount) {
iwhichlocal = iwhichglobal - nbefore;
iatom = list[iwhichlocal];
imolecule = molecule[iatom];
me = comm->me;
} else me = -1;
// bcast mol ID to delete all atoms from
// if mol ID > 0, delete any atom in molecule and decrement counters
// if mol ID == 0, delete single iatom
// logic with ndeltopo is to count # of deleted bonds,angles,etc
// for atom->molecular = 1, do this for each deleted atom in molecule
// for atom->molecular = 2, use Molecule counts for just 1st atom in mol
MPI_Allreduce(&me,&proc,1,MPI_INT,MPI_MAX,world);
MPI_Bcast(&imolecule,1,MPI_LMP_TAGINT,proc,world);
ndelone = 0;
for (i = 0; i < nlocal; i++) {
if (imolecule && molecule[i] == imolecule) {
mark[i] = 1;
ndelone++;
if (molecular == 1) {
if (atom->avec->bonds_allow) {
if (force->newton_bond) ndeltopo[0] += atom->num_bond[i];
else {
for (j = 0; j < atom->num_bond[i]; j++) {
if (tag[i] < atom->bond_atom[i][j]) ndeltopo[0]++;
}
}
}
if (atom->avec->angles_allow) {
if (force->newton_bond) ndeltopo[1] += atom->num_angle[i];
else {
for (j = 0; j < atom->num_angle[i]; j++) {
m = atom->map(atom->angle_atom2[i][j]);
if (m >= 0 && m < nlocal) ndeltopo[1]++;
}
}
}
if (atom->avec->dihedrals_allow) {
if (force->newton_bond) ndeltopo[2] += atom->num_dihedral[i];
else {
for (j = 0; j < atom->num_dihedral[i]; j++) {
m = atom->map(atom->dihedral_atom2[i][j]);
if (m >= 0 && m < nlocal) ndeltopo[2]++;
}
}
}
if (atom->avec->impropers_allow) {
if (force->newton_bond) ndeltopo[3] += atom->num_improper[i];
else {
for (j = 0; j < atom->num_improper[i]; j++) {
m = atom->map(atom->improper_atom2[i][j]);
if (m >= 0 && m < nlocal) ndeltopo[3]++;
}
}
}
} else if (molecular == 2) {
if (molatom[i] == 0) {
index = molindex[i];
ndeltopo[0] += onemols[index]->nbonds;
ndeltopo[1] += onemols[index]->nangles;
ndeltopo[2] += onemols[index]->ndihedrals;
ndeltopo[3] += onemols[index]->nimpropers;
}
}
} else if (me == proc && i == iatom) {
mark[i] = 1;
ndelone++;
}
}
// remove any atoms marked for deletion from my eligible list
i = 0;
while (i < ncount) {
if (mark[list[i]]) {
list[i] = list[ncount-1];
ncount--;
} else i++;
}
// update ndel,ncount,nall,nbefore
// ndelall is total atoms deleted on this iteration
// ncount is already correct, so resum to get nall and nbefore
MPI_Allreduce(&ndelone,&ndelall,1,MPI_INT,MPI_SUM,world);
ndel += ndelall;
MPI_Allreduce(&ncount,&nall,1,MPI_INT,MPI_SUM,world);
MPI_Scan(&ncount,&nbefore,1,MPI_INT,MPI_SUM,world);
nbefore -= ncount;
}
}
// delete my marked atoms
// loop in reverse order to avoid copying marked atoms
AtomVec *avec = atom->avec;
for (i = nlocal-1; i >= 0; i--) {
if (mark[i]) {
avec->copy(atom->nlocal-1,i,1);
atom->nlocal--;
}
}
// reset global natoms and bonds, angles, etc
// if global map exists, reset it now instead of waiting for comm
// since deleting atoms messes up ghosts
atom->natoms -= ndel;
if (molflag) {
int all[4];
MPI_Allreduce(ndeltopo,all,4,MPI_INT,MPI_SUM,world);
atom->nbonds -= all[0];
atom->nangles -= all[1];
atom->ndihedrals -= all[2];
atom->nimpropers -= all[3];
}
if (ndel && atom->map_style) {
atom->nghost = 0;
atom->map_init();
atom->map_set();
}
// statistics
ndeleted += ndel;
next_reneighbor = update->ntimestep + nevery;
}
void ReadRestart::command(int narg, char **arg)
{
if (narg != 1) error->all(FLERR,"Illegal read_restart command");
if (domain->box_exist)
error->all(FLERR,"Cannot read_restart after simulation box is defined");
MPI_Comm_rank(world,&me);
MPI_Comm_size(world,&nprocs);
// if filename contains "*", search dir for latest restart file
char *file = new char[strlen(arg[0]) + 16];
if (strchr(arg[0],'*')) {
int n;
if (me == 0) {
file_search(arg[0],file);
n = strlen(file) + 1;
}
MPI_Bcast(&n,1,MPI_INT,0,world);
MPI_Bcast(file,n,MPI_CHAR,0,world);
} else strcpy(file,arg[0]);
// check for multiproc files and an MPI-IO filename
if (strchr(arg[0],'%')) multiproc = 1;
else multiproc = 0;
if (strstr(arg[0],".mpi")) mpiioflag = 1;
else mpiioflag = 0;
if (multiproc && mpiioflag)
error->all(FLERR,
"Read restart MPI-IO output not allowed with '%' in filename");
if (mpiioflag) {
mpiio = new RestartMPIIO(lmp);
if (!mpiio->mpiio_exists)
error->all(FLERR,"Reading from MPI-IO filename when "
"MPIIO package is not installed");
}
// open single restart file or base file for multiproc case
if (me == 0) {
if (screen) fprintf(screen,"Reading restart file ...\n");
char *hfile;
if (multiproc) {
hfile = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(hfile,"%s%s%s",file,"base",ptr+1);
*ptr = '%';
} else hfile = file;
fp = fopen(hfile,"rb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",hfile);
error->one(FLERR,str);
}
if (multiproc) delete [] hfile;
}
// read magic string, endian flag, numeric version
magic_string();
endian();
int incompatible = version_numeric();
// read header info which creates simulation box
header(incompatible);
domain->box_exist = 1;
// problem setup using info from header
int n;
if (nprocs == 1) n = static_cast<int> (atom->natoms);
else n = static_cast<int> (LB_FACTOR * atom->natoms / nprocs);
atom->allocate_type_arrays();
atom->avec->grow(n);
n = atom->nmax;
domain->print_box(" ");
domain->set_initial_box();
domain->set_global_box();
comm->set_proc_grid();
domain->set_local_box();
// read groups, ntype-length arrays, force field, fix info from file
// nextra = max # of extra quantities stored with each atom
group->read_restart(fp);
type_arrays();
force_fields();
int nextra = modify->read_restart(fp);
atom->nextra_store = nextra;
memory->create(atom->extra,n,nextra,"atom:extra");
// read file layout info
file_layout();
// close header file if in multiproc mode
if (multiproc && me == 0) fclose(fp);
// read per-proc info
AtomVec *avec = atom->avec;
int maxbuf = 0;
double *buf = NULL;
int m,flag;
// MPI-IO input from single file
if (mpiioflag) {
// add calls to RestartMPIIO class
// reopen header file
// perform reads
// allow for different # of procs reading than wrote the file
// mpiio->open(file);
// mpiio->read();
// mpiio->close();
// then process atom info as
//m = 0;
//while (m < n) m += avec->unpack_restart(&buf[m]);
}
// input of single native file
// nprocs_file = # of chunks in file
// proc 0 reads a chunk and bcasts it to other procs
// each proc unpacks the atoms, saving ones in it's sub-domain
// check for atom in sub-domain differs for orthogonal vs triclinic box
else if (multiproc == 0) {
int triclinic = domain->triclinic;
double *x,lamda[3];
double *coord,*sublo,*subhi;
if (triclinic == 0) {
sublo = domain->sublo;
subhi = domain->subhi;
} else {
sublo = domain->sublo_lamda;
subhi = domain->subhi_lamda;
}
for (int iproc = 0; iproc < nprocs_file; iproc++) {
if (read_int() != PERPROC)
error->all(FLERR,"Invalid flag in peratom section of restart file");
n = read_int();
if (n > maxbuf) {
maxbuf = n;
memory->destroy(buf);
memory->create(buf,maxbuf,"read_restart:buf");
}
read_double_vec(n,buf);
m = 0;
while (m < n) {
x = &buf[m+1];
if (triclinic) {
domain->x2lamda(x,lamda);
coord = lamda;
} else coord = x;
if (coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1] &&
coord[2] >= sublo[2] && coord[2] < subhi[2]) {
m += avec->unpack_restart(&buf[m]);
}
else m += static_cast<int> (buf[m]);
}
}
if (me == 0) fclose(fp);
}
// input of multiple native files with procs <= files
// # of files = multiproc_file
// each proc reads a subset of files, striding by nprocs
// each proc keeps all atoms in all perproc chunks in its files
else if (nprocs <= multiproc_file) {
char *procfile = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
for (int iproc = me; iproc < multiproc_file; iproc += nprocs) {
*ptr = '\0';
sprintf(procfile,"%s%d%s",file,iproc,ptr+1);
*ptr = '%';
fp = fopen(procfile,"rb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",procfile);
error->one(FLERR,str);
}
fread(&flag,sizeof(int),1,fp);
if (flag != PROCSPERFILE)
error->one(FLERR,"Invalid flag in peratom section of restart file");
int procsperfile;
fread(&procsperfile,sizeof(int),1,fp);
for (int i = 0; i < procsperfile; i++) {
fread(&flag,sizeof(int),1,fp);
if (flag != PERPROC)
error->one(FLERR,"Invalid flag in peratom section of restart file");
fread(&n,sizeof(int),1,fp);
if (n > maxbuf) {
maxbuf = n;
memory->destroy(buf);
memory->create(buf,maxbuf,"read_restart:buf");
}
fread(buf,sizeof(double),n,fp);
m = 0;
while (m < n) m += avec->unpack_restart(&buf[m]);
}
fclose(fp);
}
delete [] procfile;
}
// input of multiple native files with procs > files
// # of files = multiproc_file
// cluster procs based on # of files
// 1st proc in each cluster reads per-proc chunks from file
// sends chunks round-robin to other procs in its cluster
// each proc keeps all atoms in its perproc chunks in file
else {
// nclusterprocs = # of procs in my cluster that read from one file
// filewriter = 1 if this proc reads file, else 0
// fileproc = ID of proc in my cluster who reads from file
// clustercomm = MPI communicator within my cluster of procs
int nfile = multiproc_file;
int icluster = static_cast<int> ((bigint) me * nfile/nprocs);
int fileproc = static_cast<int> ((bigint) icluster * nprocs/nfile);
int fcluster = static_cast<int> ((bigint) fileproc * nfile/nprocs);
if (fcluster < icluster) fileproc++;
int fileprocnext =
static_cast<int> ((bigint) (icluster+1) * nprocs/nfile);
fcluster = static_cast<int> ((bigint) fileprocnext * nfile/nprocs);
if (fcluster < icluster+1) fileprocnext++;
int nclusterprocs = fileprocnext - fileproc;
int filereader = 0;
if (me == fileproc) filereader = 1;
MPI_Comm clustercomm;
MPI_Comm_split(world,icluster,0,&clustercomm);
if (filereader) {
char *procfile = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(procfile,"%s%d%s",file,icluster,ptr+1);
*ptr = '%';
fp = fopen(procfile,"rb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",procfile);
error->one(FLERR,str);
}
delete [] procfile;
}
int flag,procsperfile;
if (filereader) {
fread(&flag,sizeof(int),1,fp);
if (flag != PROCSPERFILE)
error->one(FLERR,"Invalid flag in peratom section of restart file");
fread(&procsperfile,sizeof(int),1,fp);
}
MPI_Bcast(&procsperfile,1,MPI_INT,0,clustercomm);
int tmp,iproc;
MPI_Status status;
MPI_Request request;
for (int i = 0; i < procsperfile; i++) {
if (filereader) {
fread(&flag,sizeof(int),1,fp);
if (flag != PERPROC)
error->one(FLERR,"Invalid flag in peratom section of restart file");
fread(&n,sizeof(int),1,fp);
if (n > maxbuf) {
maxbuf = n;
memory->destroy(buf);
memory->create(buf,maxbuf,"read_restart:buf");
}
fread(buf,sizeof(double),n,fp);
if (i % nclusterprocs) {
iproc = me + (i % nclusterprocs);
MPI_Send(&n,1,MPI_INT,iproc,0,world);
MPI_Recv(&tmp,0,MPI_INT,iproc,0,world,&status);
MPI_Rsend(buf,n,MPI_DOUBLE,iproc,0,world);
}
} else if (i % nclusterprocs == me - fileproc) {
MPI_Recv(&n,1,MPI_INT,fileproc,0,world,&status);
if (n > maxbuf) {
maxbuf = n;
memory->destroy(buf);
memory->create(buf,maxbuf,"read_restart:buf");
}
MPI_Irecv(buf,n,MPI_DOUBLE,fileproc,0,world,&request);
MPI_Send(&tmp,0,MPI_INT,fileproc,0,world);
MPI_Wait(&request,&status);
}
if (i % nclusterprocs == me - fileproc) {
m = 0;
while (m < n) m += avec->unpack_restart(&buf[m]);
}
}
if (filereader) fclose(fp);
MPI_Comm_free(&clustercomm);
}
// clean-up memory
delete [] file;
memory->destroy(buf);
// for multiproc or MPI-IO files:
// perform irregular comm to migrate atoms to correct procs
if (multiproc || mpiioflag) {
// create a temporary fix to hold and migrate extra atom info
// necessary b/c irregular will migrate atoms
if (nextra) {
char cextra[8],fixextra[8];
sprintf(cextra,"%d",nextra);
sprintf(fixextra,"%d",modify->nfix_restart_peratom);
char **newarg = new char*[5];
newarg[0] = (char *) "_read_restart";
newarg[1] = (char *) "all";
newarg[2] = (char *) "READ_RESTART";
newarg[3] = cextra;
newarg[4] = fixextra;
modify->add_fix(5,newarg);
delete [] newarg;
}
// move atoms to new processors via irregular()
// in case read by different proc than wrote restart file
// first do map_init() since irregular->migrate_atoms() will do map_clear()
if (atom->map_style) atom->map_init();
if (domain->triclinic) domain->x2lamda(atom->nlocal);
Irregular *irregular = new Irregular(lmp);
irregular->migrate_atoms();
delete irregular;
if (domain->triclinic) domain->lamda2x(atom->nlocal);
// put extra atom info held by fix back into atom->extra
// destroy temporary fix
if (nextra) {
memory->destroy(atom->extra);
memory->create(atom->extra,atom->nmax,nextra,"atom:extra");
int ifix = modify->find_fix("_read_restart");
FixReadRestart *fix = (FixReadRestart *) modify->fix[ifix];
int *count = fix->count;
double **extra = fix->extra;
double **atom_extra = atom->extra;
int nlocal = atom->nlocal;
for (int i = 0; i < nlocal; i++)
for (int j = 0; j < count[i]; j++)
atom_extra[i][j] = extra[i][j];
modify->delete_fix("_read_restart");
}
}
// check that all atoms were assigned to procs
bigint natoms;
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (me == 0) {
if (screen) fprintf(screen," " BIGINT_FORMAT " atoms\n",natoms);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " atoms\n",natoms);
}
if (natoms != atom->natoms)
error->all(FLERR,"Did not assign all atoms correctly");
if (me == 0) {
if (atom->nbonds) {
if (screen) fprintf(screen," " BIGINT_FORMAT " bonds\n",atom->nbonds);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " bonds\n",atom->nbonds);
}
if (atom->nangles) {
if (screen) fprintf(screen," " BIGINT_FORMAT " angles\n",
atom->nangles);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " angles\n",
atom->nangles);
}
if (atom->ndihedrals) {
if (screen) fprintf(screen," " BIGINT_FORMAT " dihedrals\n",
atom->ndihedrals);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " dihedrals\n",
atom->ndihedrals);
}
if (atom->nimpropers) {
if (screen) fprintf(screen," " BIGINT_FORMAT " impropers\n",
atom->nimpropers);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " impropers\n",
atom->nimpropers);
}
}
// check if tags are being used
// create global mapping and bond topology now that system is defined
flag = 0;
for (int i = 0; i < atom->nlocal; i++)
if (atom->tag[i] > 0) flag = 1;
int flag_all;
MPI_Allreduce(&flag,&flag_all,1,MPI_INT,MPI_MAX,world);
if (flag_all == 0) atom->tag_enable = 0;
if (atom->map_style) {
atom->map_init();
atom->map_set();
}
if (atom->molecular) {
Special special(lmp);
special.build();
}
}
void CommCuda::forward_comm_unpack_cuda()
{
static int count=0;
static double kerneltime=0.0;
static double copytime=0.0;
timespec time1,time2,time3;
int n;
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
double **x = atom->x;
cuda->shared_data.domain.xy=domain->xy;
cuda->shared_data.domain.xz=domain->xz;
cuda->shared_data.domain.yz=domain->yz;
cuda->shared_data.domain.prd[0]=domain->prd[0];
cuda->shared_data.domain.prd[1]=domain->prd[1];
cuda->shared_data.domain.prd[2]=domain->prd[2];
cuda->shared_data.domain.triclinic=domain->triclinic;
if(not comm_x_only && not avec->cudable) cuda->downloadAll(); //if not comm_x_only the communication routine of the atom_vec style class is used
// exchange data with another proc
// if other proc is self, just copy
// if comm_x_only set, exchange or copy directly to x, don't unpack
for (int iswap = 0; iswap < nswap; iswap++) {
if (sendproc[iswap] != me)
{
if (comm_x_only)
{
//Cuda_CommCuda_UnpackComm(&cuda->shared_data,recvnum[iswap],firstrecv[iswap],cuda->shared_data.comm.buf_recv[iswap],iswap); //Unpack for cpu exchange happens implicitely since buf==x[firstrecv]
Cuda_CommCuda_UnpackComm(&cuda->shared_data,recvnum[iswap],firstrecv[iswap],buf_recv,iswap); //Unpack for cpu exchange happens implicitely since buf==x[firstrecv]
}
else if (ghost_velocity)
{
//Cuda_CommCuda_UnpackComm_Vel(&cuda->shared_data,recvnum[iswap],firstrecv[iswap],(void*)&buf_recv[iswap*maxrecv]); //Unpack for cpu exchange happens implicitely since buf==x[firstrecv]
}
else
{
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
if(avec->cudable)
n = avec->pack_comm(sendnum[iswap],&iswap,
buf_send,pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_recv);
}
}
else //sendproc == me
{
if (comm_x_only)
{
if (sendnum[iswap])
{
}
}
else if (ghost_velocity)
{
}
else
{
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_send);
}
}
}
if(not comm_x_only && not avec->cudable) cuda->uploadAll();
}
void CommCuda::forward_comm_cuda()
{
static int count=0;
static double kerneltime=0.0;
static double copytime=0.0;
timespec time1,time2,time3;
int n;
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
double **x = atom->x;
cuda->shared_data.domain.xy=domain->xy;
cuda->shared_data.domain.xz=domain->xz;
cuda->shared_data.domain.yz=domain->yz;
cuda->shared_data.domain.prd[0]=domain->prd[0];
cuda->shared_data.domain.prd[1]=domain->prd[1];
cuda->shared_data.domain.prd[2]=domain->prd[2];
cuda->shared_data.domain.triclinic=domain->triclinic;
if(not comm_x_only && not avec->cudable)
{
cuda->downloadAll();
Comm::forward_comm();
cuda->uploadAll();
return;
}
// exchange data with another proc
// if other proc is self, just copy
// if comm_x_only set, exchange or copy directly to x, don't unpack
for (int iswap = 0; iswap < nswap; iswap++) {
if (sendproc[iswap] != me)
{
if (comm_x_only)
{
int size_forward_recv_now=0;
if((sizeof(X_FLOAT)!=sizeof(double)) && size_forward_recv[iswap]) //some complicated way to safe some transfer size if single precision is used
size_forward_recv_now=(size_forward_recv[iswap]+1)*sizeof(X_FLOAT)/sizeof(double);
else
size_forward_recv_now=size_forward_recv[iswap];
clock_gettime(CLOCK_REALTIME,&time1);
MPI_Irecv(buf_recv,size_forward_recv_now,MPI_DOUBLE,
recvproc[iswap],0,world,&request);
n = Cuda_CommCuda_PackComm(&cuda->shared_data,sendnum[iswap],iswap,(void*) buf_send,pbc[iswap],pbc_flag[iswap]);
clock_gettime(CLOCK_REALTIME,&time2);
if((sizeof(X_FLOAT)!=sizeof(double)) && n) //some complicated way to safe some transfer size if single precision is used
n=(n+1)*sizeof(X_FLOAT)/sizeof(double);
//printf("RecvSize: %i SendSize: %i\n",size_forward_recv_now,n);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
clock_gettime(CLOCK_REALTIME,&time3);
cuda->shared_data.cuda_timings.comm_forward_mpi_upper+=
time3.tv_sec-time1.tv_sec+1.0*(time3.tv_nsec-time1.tv_nsec)/1000000000;
cuda->shared_data.cuda_timings.comm_forward_mpi_lower+=
time3.tv_sec-time2.tv_sec+1.0*(time3.tv_nsec-time2.tv_nsec)/1000000000;
Cuda_CommCuda_UnpackComm(&cuda->shared_data,recvnum[iswap],firstrecv[iswap],(void*)buf_recv,iswap); //Unpack for cpu exchange happens implicitely since buf==x[firstrecv]
}
else if (ghost_velocity)
{
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
if(avec->cudable)
n = avec->pack_comm_vel(sendnum[iswap],&iswap,
buf_send,pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_comm_vel(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
avec->unpack_comm_vel(recvnum[iswap],firstrecv[iswap],buf_recv);
}
else
{
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
if(avec->cudable)
n = avec->pack_comm(sendnum[iswap],&iswap,
buf_send,pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_recv);
}
}
else //sendproc == me
{
cuda->self_comm=1;
if (comm_x_only)
{
if (sendnum[iswap])
{
n = Cuda_CommCuda_PackComm_Self(&cuda->shared_data,sendnum[iswap],iswap,firstrecv[iswap],pbc[iswap],pbc_flag[iswap]);
if(n<0) error->all(FLERR," # CUDA ERRROR on PackComm_Self");
if((sizeof(X_FLOAT)!=sizeof(double)) && n)
n=(n+1)*sizeof(X_FLOAT)/sizeof(double);
}
}
else if (ghost_velocity)
{
n = avec->pack_comm_vel(sendnum[iswap],&iswap,
(double*) firstrecv,pbc_flag[iswap],pbc[iswap]);
//avec->unpack_comm_vel(recvnum[iswap],firstrecv[iswap],(double*) firstrecv);
}
else
{
n = avec->pack_comm(sendnum[iswap],&iswap,
(double*) firstrecv,pbc_flag[iswap],pbc[iswap]);
//avec->unpack_comm(recvnum[iswap],firstrecv[iswap],(double*) firstrecv);
}
cuda->self_comm=0;
}
}
}
void CommCuda::exchange_cuda()
{
int i,m,nsend,nrecv,nrecv1,nrecv2,nlocal;
double lo,hi,value;
double **x;
double *sublo,*subhi,*buf;
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
timespec time1,time2,time3;
// clear global->local map for owned and ghost atoms
// b/c atoms migrate to new procs in exchange() and
// new ghosts are created in borders()
// map_set() is done at end of borders()
if(map_style) cuda->cu_tag->download();
if (map_style) atom->map_clear();
// subbox bounds for orthogonal or triclinic
if (triclinic == 0) {
sublo = domain->sublo;
subhi = domain->subhi;
} else {
sublo = domain->sublo_lamda;
subhi = domain->subhi_lamda;
}
// loop over dimensions
for (int dim = 0; dim < 3; dim++) {
// fill buffer with atoms leaving my box, using < and >=
// when atom is deleted, fill it in with last atom
cuda->shared_data.exchange_dim=dim;
nlocal = atom->nlocal;
avec->maxsend=&maxsend;
nsend=avec->pack_exchange(dim,(double*) &buf_send);
nlocal = atom->nlocal;
atom->nlocal = nlocal;
// send/recv atoms in both directions
// if 1 proc in dimension, no send/recv, set recv buf to send buf
// if 2 procs in dimension, single send/recv
// if more than 2 procs in dimension, send/recv to both neighbors
clock_gettime(CLOCK_REALTIME,&time1);
if (procgrid[dim] == 1) {
nrecv = nsend;
buf = buf_send;
} else {
MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][0],0,
&nrecv1,1,MPI_INT,procneigh[dim][1],0,world,&status);
nrecv = nrecv1;
if (procgrid[dim] > 2) {
MPI_Sendrecv(&nsend,1,MPI_INT,procneigh[dim][1],0,
&nrecv2,1,MPI_INT,procneigh[dim][0],0,world,&status);
nrecv += nrecv2;
}
if (nrecv+1 > maxrecv) grow_recv(nrecv+1);
MPI_Irecv(buf_recv,nrecv1,MPI_DOUBLE,procneigh[dim][1],0,
world,&request);
MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][0],0,world);
MPI_Wait(&request,&status);
if (procgrid[dim] > 2) {
MPI_Irecv(&buf_recv[nrecv1],nrecv2,MPI_DOUBLE,procneigh[dim][0],0,
world,&request);
MPI_Send(buf_send,nsend,MPI_DOUBLE,procneigh[dim][1],0,world);
MPI_Wait(&request,&status);
if((nrecv1==0)||(nrecv2==0)) buf_recv[nrecv]=0;
}
buf = buf_recv;
}
//printf("nsend: %i nrecv: %i\n",nsend,nrecv);
// check incoming atoms to see if they are in my box
// if so, add to my list
clock_gettime(CLOCK_REALTIME,&time2);
cuda->shared_data.cuda_timings.comm_exchange_mpi+=
time2.tv_sec-time1.tv_sec+1.0*(time2.tv_nsec-time1.tv_nsec)/1000000000;
if(nrecv)
{
avec->maxsend=&maxsend;
avec->unpack_exchange(buf);
}
}
if(atom->firstgroupname) cuda->downloadAll();
if(atom->firstgroupname) atom->first_reorder();
if(atom->firstgroupname) cuda->uploadAll();
}
void CommCuda::forward_comm_transfer_cuda()
{
static int count=0;
static double kerneltime=0.0;
static double copytime=0.0;
timespec time1,time2,time3;
int n;
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
double **x = atom->x;
cuda->shared_data.domain.xy=domain->xy;
cuda->shared_data.domain.xz=domain->xz;
cuda->shared_data.domain.yz=domain->yz;
cuda->shared_data.domain.prd[0]=domain->prd[0];
cuda->shared_data.domain.prd[1]=domain->prd[1];
cuda->shared_data.domain.prd[2]=domain->prd[2];
cuda->shared_data.domain.triclinic=domain->triclinic;
if(not comm_x_only && not avec->cudable) cuda->downloadAll(); //if not comm_x_only the communication routine of the atom_vec style class is used
//printf("A\n");
// exchange data with another proc
// if other proc is self, just copy
// if comm_x_only set, exchange or copy directly to x, don't unpack
for (int iswap = 0; iswap < nswap; iswap++) {
if (sendproc[iswap] != me)
{
if (comm_x_only)
{
int size_forward_recv_now=0;
if((sizeof(X_FLOAT)!=sizeof(double)) && size_forward_recv[iswap]) //some complicated way to safe some transfer size if single precision is used
size_forward_recv_now=(size_forward_recv[iswap]+1)*sizeof(X_FLOAT)/sizeof(double);
else
size_forward_recv_now=size_forward_recv[iswap];
//printf("A: %i \n",size_forward_recv_now/1024*4);
//MPI_Irecv(cuda->shared_data.comm.buf_recv[iswap],size_forward_recv_now,MPI_DOUBLE,
// recvproc[iswap],0,world,&request);
MPI_Irecv(buf_recv,size_forward_recv_now,MPI_DOUBLE,
recvproc[iswap],0,world,&request);
//printf("%p %p %i\n",buf_send, cuda->shared_data.comm.buf_send_dev[iswap], cuda->shared_data.comm.send_size[iswap]*sizeof(double));
//memcpy(buf_send,cuda->shared_data.comm.buf_send[iswap],cuda->shared_data.comm.send_size[iswap]*sizeof(double));
// CudaWrapper_SyncStream(1);
//printf("B: %i \n",cuda->shared_data.comm.send_size[iswap]/1024*4);
CudaWrapper_DownloadCudaDataAsync((void*) buf_send, cuda->shared_data.comm.buf_send_dev[iswap], cuda->shared_data.comm.send_size[iswap]*sizeof(double),2);
//MPI_Send(cuda->shared_data.comm.buf_send[iswap],cuda->shared_data.comm.send_size[iswap],MPI_DOUBLE,sendproc[iswap],0,world);
clock_gettime(CLOCK_REALTIME,&time1);
CudaWrapper_SyncStream(2);
//printf("C: %i \n",cuda->shared_data.comm.send_size[iswap]/1024*4);
clock_gettime(CLOCK_REALTIME,&time2);
cuda->shared_data.cuda_timings.comm_forward_download+=
time2.tv_sec-time1.tv_sec+1.0*(time2.tv_nsec-time1.tv_nsec)/1000000000;
MPI_Send(buf_send,cuda->shared_data.comm.send_size[iswap],MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
//printf("D: %i \n",cuda->shared_data.comm.send_size[iswap]/1024*4);
CudaWrapper_UploadCudaDataAsync((void*) buf_recv,cuda->shared_data.comm.buf_recv_dev[iswap], size_forward_recv_now*sizeof(double),2);
clock_gettime(CLOCK_REALTIME,&time1);
CudaWrapper_SyncStream(2);
//printf("E: %i \n",cuda->shared_data.comm.send_size[iswap]/1024*4);
//memcpy(cuda->shared_data.comm.buf_recv[iswap],buf_recv,size_forward_recv_now*sizeof(double));
//printf("RecvSize: %i SendSize: %i\n",size_forward_recv_now*sizeof(double),cuda->shared_data.comm.send_size[iswap]*sizeof(double));
clock_gettime(CLOCK_REALTIME,&time3);
cuda->shared_data.cuda_timings.comm_forward_upload+=
time3.tv_sec-time1.tv_sec+1.0*(time3.tv_nsec-time1.tv_nsec)/1000000000;
cuda->shared_data.cuda_timings.comm_forward_mpi_lower+=
time3.tv_sec-time2.tv_sec+1.0*(time3.tv_nsec-time2.tv_nsec)/1000000000;
clock_gettime(CLOCK_REALTIME,&time3);
cuda->shared_data.cuda_timings.comm_forward_mpi_upper+=
time3.tv_sec-time1.tv_sec+1.0*(time3.tv_nsec-time1.tv_nsec)/1000000000;
}
else if (ghost_velocity)
{
/* int size_forward_recv_now=0;
if((sizeof(X_FLOAT)!=sizeof(double)) && size_forward_recv[iswap]) //some complicated way to safe some transfer size if single precision is used
size_forward_recv_now=(size_forward_recv[iswap]+1)*sizeof(X_FLOAT)/sizeof(double);
else
size_forward_recv_now=size_forward_recv[iswap];
clock_gettime(CLOCK_REALTIME,&time1);
MPI_Irecv(cuda->shared_data.comm.buf_recv[iswap],size_forward_recv_now,MPI_DOUBLE,
recvproc[iswap],0,world,&request);
clock_gettime(CLOCK_REALTIME,&time2);
MPI_Send(cuda->shared_data.comm.buf_send[iswap],cuda->shared_data.comm.send_size[iswap],MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
clock_gettime(CLOCK_REALTIME,&time3);
cuda->shared_data.cuda_timings.comm_forward_mpi_upper+=
time3.tv_sec-time1.tv_sec+1.0*(time3.tv_nsec-time1.tv_nsec)/1000000000;
cuda->shared_data.cuda_timings.comm_forward_mpi_lower+=
time3.tv_sec-time2.tv_sec+1.0*(time3.tv_nsec-time2.tv_nsec)/1000000000;*/
}
else
{
MPI_Irecv(buf_recv,size_forward_recv[iswap],MPI_DOUBLE,
recvproc[iswap],0,world,&request);
if(avec->cudable)
n = avec->pack_comm(sendnum[iswap],&iswap,
buf_send,pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_recv);
}
}
else //sendproc == me
{
if (comm_x_only)
{
if (sendnum[iswap])
{
}
}
else if (ghost_velocity)
{
}
else
{
n = avec->pack_comm(sendnum[iswap],sendlist[iswap],
buf_send,pbc_flag[iswap],pbc[iswap]);
avec->unpack_comm(recvnum[iswap],firstrecv[iswap],buf_send);
}
}
}
if(not comm_x_only && not avec->cudable) cuda->uploadAll();
}
void FixSRP::setup_pre_force(int /*zz*/)
{
double **x = atom->x;
double **xold;
tagint *tag = atom->tag;
tagint *tagold;
int *type = atom->type;
int* dlist;
AtomVec *avec = atom->avec;
int **bondlist = neighbor->bondlist;
int nlocal, nlocal_old;
nlocal = nlocal_old = atom->nlocal;
bigint nall = atom->nlocal + atom->nghost;
int nbondlist = neighbor->nbondlist;
int i,j,n;
// make a copy of all coordinates and tags
// that is consistent with the bond list as
// atom->x will be affected by creating/deleting atoms.
// also compile list of local atoms to be deleted.
memory->create(xold,nall,3,"fix_srp:xold");
memory->create(tagold,nall,"fix_srp:tagold");
memory->create(dlist,nall,"fix_srp:dlist");
for (i = 0; i < nall; i++){
xold[i][0] = x[i][0];
xold[i][1] = x[i][1];
xold[i][2] = x[i][2];
tagold[i]=tag[i];
dlist[i] = (type[i] == bptype) ? 1 : 0;
for (n = 0; n < 2; n++)
array[i][n] = 0.0;
}
// delete local atoms flagged in dlist
i = 0;
int ndel = 0;
while (i < nlocal) {
if (dlist[i]) {
avec->copy(nlocal-1,i,1);
dlist[i] = dlist[nlocal-1];
nlocal--;
ndel++;
} else i++;
}
atom->nlocal = nlocal;
memory->destroy(dlist);
int nadd = 0;
double rsqold = 0.0;
double delx, dely, delz, rmax, rsq, rsqmax;
double xone[3];
for (n = 0; n < nbondlist; n++) {
// consider only the user defined bond type
// btype of zero considers all bonds
if(btype > 0 && bondlist[n][2] != btype)
continue;
i = bondlist[n][0];
j = bondlist[n][1];
// position of bond i
xone[0] = (xold[i][0] + xold[j][0])*0.5;
xone[1] = (xold[i][1] + xold[j][1])*0.5;
xone[2] = (xold[i][2] + xold[j][2])*0.5;
// record longest bond
// this used to set ghost cutoff
delx = xold[j][0] - xold[i][0];
dely = xold[j][1] - xold[i][1];
delz = xold[j][2] - xold[i][2];
rsq = delx*delx + dely*dely + delz*delz;
if(rsq > rsqold) rsqold = rsq;
// make one particle for each bond
// i is local
// if newton bond, always make particle
// if j is local, always make particle
// if j is ghost, decide from tag
if ((force->newton_bond) || (j < nlocal_old) || (tagold[i] > tagold[j])) {
atom->natoms++;
avec->create_atom(bptype,xone);
// pack tag i/j into buffer for comm
array[atom->nlocal-1][0] = static_cast<double>(tagold[i]);
array[atom->nlocal-1][1] = static_cast<double>(tagold[j]);
nadd++;
}
}
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&atom->natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
// free temporary storage
memory->destroy(xold);
memory->destroy(tagold);
char str[128];
int nadd_all = 0, ndel_all = 0;
MPI_Allreduce(&ndel,&ndel_all,1,MPI_INT,MPI_SUM,world);
MPI_Allreduce(&nadd,&nadd_all,1,MPI_INT,MPI_SUM,world);
if(comm->me == 0){
sprintf(str, "Removed/inserted %d/%d bond particles.", ndel_all,nadd_all);
error->message(FLERR,str);
}
// check ghost comm distances
// warn and change if shorter from estimate
// ghost atoms must be present for bonds on edge of neighbor cutoff
// extend cutghost slightly more than half of the longest bond
MPI_Allreduce(&rsqold,&rsqmax,1,MPI_DOUBLE,MPI_MAX,world);
rmax = sqrt(rsqmax);
double cutneighmax_srp = neighbor->cutneighmax + 0.51*rmax;
double length0,length1,length2;
if (domain->triclinic) {
double *h_inv = domain->h_inv;
length0 = sqrt(h_inv[0]*h_inv[0] + h_inv[5]*h_inv[5] + h_inv[4]*h_inv[4]);
length1 = sqrt(h_inv[1]*h_inv[1] + h_inv[3]*h_inv[3]);
length2 = h_inv[2];
} else length0 = length1 = length2 = 1.0;
// find smallest cutghost.
// comm->cutghost is stored in fractional coordinates for triclinic
double cutghostmin = comm->cutghost[0]/length0;
if (cutghostmin > comm->cutghost[1]/length1)
cutghostmin = comm->cutghost[1]/length1;
if (cutghostmin > comm->cutghost[2]/length2)
cutghostmin = comm->cutghost[2]/length2;
// stop if cutghost is insufficient
if (cutneighmax_srp > cutghostmin){
sprintf(str, "Communication cutoff too small for fix srp. "
"Need %f, current %f.", cutneighmax_srp, cutghostmin);
error->all(FLERR,str);
}
// assign tags for new atoms, update map
atom->tag_extend();
if (atom->map_style) {
atom->nghost = 0;
atom->map_init();
atom->map_set();
}
// put new particles in the box before exchange
// move owned to new procs
// get ghosts
// build neigh lists again
// if triclinic, lambda coords needed for pbc, exchange, borders
if (domain->triclinic) domain->x2lamda(atom->nlocal);
domain->pbc();
comm->setup();
if (neighbor->style) neighbor->setup_bins();
comm->exchange();
if (atom->sortfreq > 0) atom->sort();
comm->borders();
// back to box coords
if (domain->triclinic) domain->lamda2x(atom->nlocal+atom->nghost);
domain->image_check();
domain->box_too_small_check();
modify->setup_pre_neighbor();
neighbor->build(1);
neighbor->ncalls = 0;
// new atom counts
nlocal = atom->nlocal;
nall = atom->nlocal + atom->nghost;
// zero all forces
for(i = 0; i < nall; i++)
atom->f[i][0] = atom->f[i][1] = atom->f[i][2] = 0.0;
// do not include bond particles in thermo output
// remove them from all groups. set their velocity to zero.
for(i=0; i< nlocal; i++)
if(atom->type[i] == bptype) {
atom->mask[i] = 0;
atom->v[i][0] = atom->v[i][1] = atom->v[i][2] = 0.0;
}
}
void CommCuda::reverse_comm()
{
int n;
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
double **f = atom->f;
double *buf;
if(not comm_f_only && not avec->cudable) cuda->downloadAll(); //not yet implemented in CUDA but only needed for non standard atom styles
// exchange data with another proc
// if other proc is self, just copy
// if comm_f_only set, exchange or copy directly from f, don't pack
for (int iswap = nswap-1; iswap >= 0; iswap--) {
if (sendproc[iswap] != me) {
if (comm_f_only) {
int size_recv_now=size_reverse_recv[iswap];
if((sizeof(F_FLOAT)!=sizeof(double))&& size_reverse_recv[iswap])
size_recv_now=(size_recv_now+1)*sizeof(F_FLOAT)/sizeof(double);
MPI_Irecv(buf_recv,size_recv_now,MPI_DOUBLE,
sendproc[iswap],0,world,&request);
buf=buf_send;
if (size_reverse_send[iswap])
{
Cuda_CommCuda_PackReverse(&cuda->shared_data,size_reverse_send[iswap]/3,firstrecv[iswap],buf);
}
else buf=NULL;
int size_reverse_send_now=size_reverse_send[iswap];
if((sizeof(F_FLOAT)!=sizeof(double))&& size_reverse_send[iswap])
size_reverse_send_now=(size_reverse_send_now+1)*sizeof(F_FLOAT)/sizeof(double);
MPI_Send(buf,size_reverse_send_now,MPI_DOUBLE,
recvproc[iswap],0,world);
MPI_Wait(&request,&status);
Cuda_CommCuda_UnpackReverse(&cuda->shared_data,sendnum[iswap],iswap,buf_recv);
} else {
MPI_Irecv(buf_recv,size_reverse_recv[iswap],MPI_DOUBLE,
sendproc[iswap],0,world,&request);
n = avec->pack_reverse(recvnum[iswap],firstrecv[iswap],buf_send);
MPI_Send(buf_send,n,MPI_DOUBLE,recvproc[iswap],0,world);
MPI_Wait(&request,&status);
avec->unpack_reverse(sendnum[iswap],sendlist[iswap],buf_recv);
}
} else {
if (comm_f_only) {
if (sendnum[iswap])
Cuda_CommCuda_UnpackReverse_Self(&cuda->shared_data,sendnum[iswap],iswap,firstrecv[iswap]);
} else {
n = avec->pack_reverse(recvnum[iswap],firstrecv[iswap],buf_send);
avec->unpack_reverse(sendnum[iswap],sendlist[iswap],buf_send);
}
}
}
if(not comm_f_only && not avec->cudable) cuda->uploadAll(); //not yet implemented in CUDA but only needed for non standard atom styles
}
void Replicate::command(int narg, char **arg)
{
int i,j,m,n;
if (domain->box_exist == 0)
error->all(FLERR,"Replicate command before simulation box is defined");
if (narg != 3) error->all(FLERR,"Illegal replicate command");
int me = comm->me;
int nprocs = comm->nprocs;
if (me == 0 && screen) fprintf(screen,"Replicating atoms ...\n");
// nrep = total # of replications
int nx = force->inumeric(FLERR,arg[0]);
int ny = force->inumeric(FLERR,arg[1]);
int nz = force->inumeric(FLERR,arg[2]);
int nrep = nx*ny*nz;
// error and warning checks
if (nx <= 0 || ny <= 0 || nz <= 0)
error->all(FLERR,"Illegal replicate command");
if (domain->dimension == 2 && nz != 1)
error->all(FLERR,"Cannot replicate 2d simulation in z dimension");
if ((nx > 1 && domain->xperiodic == 0) ||
(ny > 1 && domain->yperiodic == 0) ||
(nz > 1 && domain->zperiodic == 0)) {
if (comm->me == 0)
error->warning(FLERR,"Replicating in a non-periodic dimension");
}
if (atom->nextra_grow || atom->nextra_restart || atom->nextra_store)
error->all(FLERR,"Cannot replicate with fixes that store atom quantities");
// maxtag = largest atom tag across all existing atoms
tagint maxtag = 0;
if (atom->tag_enable) {
for (i = 0; i < atom->nlocal; i++) maxtag = MAX(atom->tag[i],maxtag);
tagint maxtag_all;
MPI_Allreduce(&maxtag,&maxtag_all,1,MPI_LMP_TAGINT,MPI_MAX,world);
maxtag = maxtag_all;
}
// maxmol = largest molecule tag across all existing atoms
tagint maxmol = 0;
if (atom->molecule_flag) {
for (i = 0; i < atom->nlocal; i++) maxmol = MAX(atom->molecule[i],maxmol);
tagint maxmol_all;
MPI_Allreduce(&maxmol,&maxmol_all,1,MPI_LMP_TAGINT,MPI_MAX,world);
maxmol = maxmol_all;
}
// unmap existing atoms via image flags
for (i = 0; i < atom->nlocal; i++)
domain->unmap(atom->x[i],atom->image[i]);
// communication buffer for all my atom's info
// max_size = largest buffer needed by any proc
// must do before new Atom class created,
// since size_restart() uses atom->nlocal
int max_size;
int send_size = atom->avec->size_restart();
MPI_Allreduce(&send_size,&max_size,1,MPI_INT,MPI_MAX,world);
double *buf;
memory->create(buf,max_size,"replicate:buf");
// old = original atom class
// atom = new replicated atom class
// also set atomKK for Kokkos version of Atom class
Atom *old = atom;
if (lmp->kokkos) atom = new AtomKokkos(lmp);
else atom = new Atom(lmp);
atomKK = (AtomKokkos*) atom;
atom->settings(old);
atom->create_avec(old->atom_style,old->avec->nargcopy,old->avec->argcopy,0);
// check that new system will not be too large
// new tags cannot exceed MAXTAGINT
// new system sizes cannot exceed MAXBIGINT
if (atom->tag_enable) {
bigint maxnewtag = maxtag + (nrep-1)*old->natoms;
if (maxnewtag < 0 || maxnewtag >= MAXTAGINT)
error->all(FLERR,"Replicated system atom IDs are too big");
}
if (nrep*old->natoms < 0 || nrep*old->natoms >= MAXBIGINT ||
nrep*old->nbonds < 0 || nrep*old->nbonds >= MAXBIGINT ||
nrep*old->nangles < 0 || nrep*old->nangles >= MAXBIGINT ||
nrep*old->ndihedrals < 0 || nrep*old->ndihedrals >= MAXBIGINT ||
nrep*old->nimpropers < 0 || nrep*old->nimpropers >= MAXBIGINT)
error->all(FLERR,"Replicated system is too big");
// assign atom and topology counts in new class from old one
atom->natoms = old->natoms * nrep;
atom->nbonds = old->nbonds * nrep;
atom->nangles = old->nangles * nrep;
atom->ndihedrals = old->ndihedrals * nrep;
atom->nimpropers = old->nimpropers * nrep;
atom->ntypes = old->ntypes;
atom->nbondtypes = old->nbondtypes;
atom->nangletypes = old->nangletypes;
atom->ndihedraltypes = old->ndihedraltypes;
atom->nimpropertypes = old->nimpropertypes;
atom->bond_per_atom = old->bond_per_atom;
atom->angle_per_atom = old->angle_per_atom;
atom->dihedral_per_atom = old->dihedral_per_atom;
atom->improper_per_atom = old->improper_per_atom;
// store old simulation box
int triclinic = domain->triclinic;
double old_xprd = domain->xprd;
double old_yprd = domain->yprd;
double old_zprd = domain->zprd;
double old_xy = domain->xy;
double old_xz = domain->xz;
double old_yz = domain->yz;
// setup new simulation box
domain->boxhi[0] = domain->boxlo[0] + nx*old_xprd;
domain->boxhi[1] = domain->boxlo[1] + ny*old_yprd;
domain->boxhi[2] = domain->boxlo[2] + nz*old_zprd;
if (triclinic) {
domain->xy *= ny;
domain->xz *= nz;
domain->yz *= nz;
}
// new problem setup using new box boundaries
if (nprocs == 1) n = static_cast<int> (atom->natoms);
else n = static_cast<int> (LB_FACTOR * atom->natoms / nprocs);
atom->allocate_type_arrays();
atom->avec->grow(n);
n = atom->nmax;
domain->print_box(" ");
domain->set_initial_box();
domain->set_global_box();
comm->set_proc_grid();
domain->set_local_box();
// copy type arrays to new atom class
if (atom->mass) {
for (int itype = 1; itype <= atom->ntypes; itype++) {
atom->mass_setflag[itype] = old->mass_setflag[itype];
if (atom->mass_setflag[itype]) atom->mass[itype] = old->mass[itype];
}
}
// set bounds for my proc
// if periodic and I am lo/hi proc, adjust bounds by EPSILON
// insures all replicated atoms will be owned even with round-off
double epsilon[3];
if (triclinic) epsilon[0] = epsilon[1] = epsilon[2] = EPSILON;
else {
epsilon[0] = domain->prd[0] * EPSILON;
epsilon[1] = domain->prd[1] * EPSILON;
epsilon[2] = domain->prd[2] * EPSILON;
}
double sublo[3],subhi[3];
if (triclinic == 0) {
sublo[0] = domain->sublo[0]; subhi[0] = domain->subhi[0];
sublo[1] = domain->sublo[1]; subhi[1] = domain->subhi[1];
sublo[2] = domain->sublo[2]; subhi[2] = domain->subhi[2];
} else {
sublo[0] = domain->sublo_lamda[0]; subhi[0] = domain->subhi_lamda[0];
sublo[1] = domain->sublo_lamda[1]; subhi[1] = domain->subhi_lamda[1];
sublo[2] = domain->sublo_lamda[2]; subhi[2] = domain->subhi_lamda[2];
}
if (comm->layout != LAYOUT_TILED) {
if (domain->xperiodic) {
if (comm->myloc[0] == 0) sublo[0] -= epsilon[0];
if (comm->myloc[0] == comm->procgrid[0]-1) subhi[0] += epsilon[0];
}
if (domain->yperiodic) {
if (comm->myloc[1] == 0) sublo[1] -= epsilon[1];
if (comm->myloc[1] == comm->procgrid[1]-1) subhi[1] += epsilon[1];
}
if (domain->zperiodic) {
if (comm->myloc[2] == 0) sublo[2] -= epsilon[2];
if (comm->myloc[2] == comm->procgrid[2]-1) subhi[2] += epsilon[2];
}
} else {
if (domain->xperiodic) {
if (comm->mysplit[0][0] == 0.0) sublo[0] -= epsilon[0];
if (comm->mysplit[0][1] == 1.0) subhi[0] += epsilon[0];
}
if (domain->yperiodic) {
if (comm->mysplit[1][0] == 0.0) sublo[1] -= epsilon[1];
if (comm->mysplit[1][1] == 1.0) subhi[1] += epsilon[1];
}
if (domain->zperiodic) {
if (comm->mysplit[2][0] == 0.0) sublo[2] -= epsilon[2];
if (comm->mysplit[2][1] == 1.0) subhi[2] += epsilon[2];
}
}
// loop over all procs
// if this iteration of loop is me:
// pack my unmapped atom data into buf
// bcast it to all other procs
// performs 3d replicate loop with while loop over atoms in buf
// x = new replicated position, remapped into simulation box
// unpack atom into new atom class from buf if I own it
// adjust tag, mol #, coord, topology info as needed
AtomVec *old_avec = old->avec;
AtomVec *avec = atom->avec;
int ix,iy,iz;
tagint atom_offset,mol_offset;
imageint image;
double x[3],lamda[3];
double *coord;
int tag_enable = atom->tag_enable;
for (int iproc = 0; iproc < nprocs; iproc++) {
if (me == iproc) {
n = 0;
for (i = 0; i < old->nlocal; i++) n += old_avec->pack_restart(i,&buf[n]);
}
MPI_Bcast(&n,1,MPI_INT,iproc,world);
MPI_Bcast(buf,n,MPI_DOUBLE,iproc,world);
for (ix = 0; ix < nx; ix++) {
for (iy = 0; iy < ny; iy++) {
for (iz = 0; iz < nz; iz++) {
// while loop over one proc's atom list
m = 0;
while (m < n) {
image = ((imageint) IMGMAX << IMG2BITS) |
((imageint) IMGMAX << IMGBITS) | IMGMAX;
if (triclinic == 0) {
x[0] = buf[m+1] + ix*old_xprd;
x[1] = buf[m+2] + iy*old_yprd;
x[2] = buf[m+3] + iz*old_zprd;
} else {
x[0] = buf[m+1] + ix*old_xprd + iy*old_xy + iz*old_xz;
x[1] = buf[m+2] + iy*old_yprd + iz*old_yz;
x[2] = buf[m+3] + iz*old_zprd;
}
domain->remap(x,image);
if (triclinic) {
domain->x2lamda(x,lamda);
coord = lamda;
} else coord = x;
if (coord[0] >= sublo[0] && coord[0] < subhi[0] &&
coord[1] >= sublo[1] && coord[1] < subhi[1] &&
coord[2] >= sublo[2] && coord[2] < subhi[2]) {
m += avec->unpack_restart(&buf[m]);
i = atom->nlocal - 1;
if (tag_enable)
atom_offset = iz*ny*nx*maxtag + iy*nx*maxtag + ix*maxtag;
else atom_offset = 0;
mol_offset = iz*ny*nx*maxmol + iy*nx*maxmol + ix*maxmol;
atom->x[i][0] = x[0];
atom->x[i][1] = x[1];
atom->x[i][2] = x[2];
atom->tag[i] += atom_offset;
atom->image[i] = image;
if (atom->molecular) {
if (atom->molecule[i] > 0)
atom->molecule[i] += mol_offset;
if (atom->molecular == 1) {
if (atom->avec->bonds_allow)
for (j = 0; j < atom->num_bond[i]; j++)
atom->bond_atom[i][j] += atom_offset;
if (atom->avec->angles_allow)
for (j = 0; j < atom->num_angle[i]; j++) {
atom->angle_atom1[i][j] += atom_offset;
atom->angle_atom2[i][j] += atom_offset;
atom->angle_atom3[i][j] += atom_offset;
}
if (atom->avec->dihedrals_allow)
for (j = 0; j < atom->num_dihedral[i]; j++) {
atom->dihedral_atom1[i][j] += atom_offset;
atom->dihedral_atom2[i][j] += atom_offset;
atom->dihedral_atom3[i][j] += atom_offset;
atom->dihedral_atom4[i][j] += atom_offset;
}
if (atom->avec->impropers_allow)
for (j = 0; j < atom->num_improper[i]; j++) {
atom->improper_atom1[i][j] += atom_offset;
atom->improper_atom2[i][j] += atom_offset;
atom->improper_atom3[i][j] += atom_offset;
atom->improper_atom4[i][j] += atom_offset;
}
}
}
} else m += static_cast<int> (buf[m]);
}
}
}
}
}
// free communication buffer and old atom class
memory->destroy(buf);
delete old;
// check that all atoms were assigned to procs
bigint natoms;
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (me == 0) {
if (screen) fprintf(screen," " BIGINT_FORMAT " atoms\n",natoms);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " atoms\n",natoms);
}
if (natoms != atom->natoms)
error->all(FLERR,"Replicate did not assign all atoms correctly");
if (me == 0) {
if (atom->nbonds) {
if (screen) fprintf(screen," " BIGINT_FORMAT " bonds\n",atom->nbonds);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " bonds\n",atom->nbonds);
}
if (atom->nangles) {
if (screen) fprintf(screen," " BIGINT_FORMAT " angles\n",
atom->nangles);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " angles\n",
atom->nangles);
}
if (atom->ndihedrals) {
if (screen) fprintf(screen," " BIGINT_FORMAT " dihedrals\n",
atom->ndihedrals);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " dihedrals\n",
atom->ndihedrals);
}
if (atom->nimpropers) {
if (screen) fprintf(screen," " BIGINT_FORMAT " impropers\n",
atom->nimpropers);
if (logfile) fprintf(logfile," " BIGINT_FORMAT " impropers\n",
atom->nimpropers);
}
}
// check that atom IDs are valid
atom->tag_check();
// create global mapping of atoms
if (atom->map_style) {
atom->map_init();
atom->map_set();
}
// create special bond lists for molecular systems
if (atom->molecular == 1) {
Special special(lmp);
special.build();
}
}
void CommCuda::borders_cuda()
{
int i,n,itype,iswap,dim,ineed,maxneed,smax,rmax;
int nsend,nrecv,nfirst,nlast,ngroup;
double lo,hi;
int *type;
double **x;
double *buf,*mlo,*mhi;
MPI_Request request;
MPI_Status status;
AtomVec *avec = atom->avec;
timespec time1,time2,time3;
// clear old ghosts
atom->nghost = 0;
// do swaps over all 3 dimensions
iswap = 0;
smax = rmax = 0;
cuda->shared_data.comm.nsend=0;
for (dim = 0; dim < 3; dim++) {
nlast = 0;
maxneed = 2*need[dim];
for (ineed = 0; ineed < maxneed; ineed++) {
// find atoms within slab boundaries lo/hi using <= and >=
// check atoms between nfirst and nlast
// for first swaps in a dim, check owned and ghost
// for later swaps in a dim, only check newly arrived ghosts
// store sent atom indices in list for use in future timesteps
x = atom->x;
if (style == SINGLE) {
lo = slablo[iswap];
hi = slabhi[iswap];
} else {
type = atom->type;
mlo = multilo[iswap];
mhi = multihi[iswap];
}
if (ineed % 2 == 0) {
nfirst = nlast;
nlast = atom->nlocal + atom->nghost;
}
nsend = 0;
// find send atoms according to SINGLE vs MULTI
// all atoms eligible versus atoms in bordergroup
// only need to limit loop to bordergroup for first sends (ineed < 2)
// on these sends, break loop in two: owned (in group) and ghost
do
{
if(nsend>=maxsendlist[iswap]) grow_list(iswap,static_cast <int> (nsend*1.05));
nsend=Cuda_CommCuda_BuildSendlist(&cuda->shared_data,bordergroup,ineed,style==SINGLE?1:0,atom->nfirst,nfirst,nlast,dim,iswap);
}while(nsend>=maxsendlist[iswap]);
// pack up list of border atoms
if (nsend*size_border > maxsend)
grow_send(nsend*size_border,0);
if (ghost_velocity)
n = avec->pack_border_vel(nsend,&iswap,buf_send,
pbc_flag[iswap],pbc[iswap]);
else
n = avec->pack_border(nsend,&iswap,buf_send,
pbc_flag[iswap],pbc[iswap]);
// swap atoms with other proc
// put incoming ghosts at end of my atom arrays
// if swapping with self, simply copy, no messages
clock_gettime(CLOCK_REALTIME,&time1);
if (sendproc[iswap] != me) {
MPI_Sendrecv(&nsend,1,MPI_INT,sendproc[iswap],0,
&nrecv,1,MPI_INT,recvproc[iswap],0,world,&status);
if (nrecv*size_border > maxrecv)
grow_recv(nrecv*size_border);
MPI_Irecv(buf_recv,nrecv*size_border,MPI_DOUBLE,
recvproc[iswap],0,world,&request);
MPI_Send(buf_send,n,MPI_DOUBLE,sendproc[iswap],0,world);
MPI_Wait(&request,&status);
buf = buf_recv;
} else {
nrecv = nsend;
buf = buf_send;
}
clock_gettime(CLOCK_REALTIME,&time2);
cuda->shared_data.cuda_timings.comm_border_mpi+=
time2.tv_sec-time1.tv_sec+1.0*(time2.tv_nsec-time1.tv_nsec)/1000000000;
// unpack buffer
if (ghost_velocity)
avec->unpack_border_vel(nrecv,atom->nlocal+atom->nghost,buf);
else
avec->unpack_border(nrecv,atom->nlocal+atom->nghost,buf);
// set all pointers & counters
smax = MAX(smax,nsend);
rmax = MAX(rmax,nrecv);
sendnum[iswap] = nsend;
recvnum[iswap] = nrecv;
size_forward_recv[iswap] = nrecv*size_forward;
size_reverse_send[iswap] = nrecv*size_reverse;
size_reverse_recv[iswap] = nsend*size_reverse;
firstrecv[iswap] = atom->nlocal + atom->nghost;
atom->nghost += nrecv;
iswap++;
}
}
// insure send/recv buffers are long enough for all forward & reverse comm
int max = MAX(maxforward*smax,maxreverse*rmax);
if (max > maxsend) grow_send(max,0);
max = MAX(maxforward*rmax,maxreverse*smax);
if (max > maxrecv) grow_recv(max);
// reset global->local map
if(map_style)
{
cuda->cu_tag->download();
atom->map_set();
}
cuda->setSystemParams();
cuda->shared_data.atom.nghost+=n;
}
void WriteRestart::write(char *file)
{
// special case where reneighboring is not done in integrator
// on timestep restart file is written (due to build_once being set)
// if box is changing, must be reset, else restart file will have
// wrong box size and atoms will be lost when restart file is read
// other calls to pbc and domain and comm are not made,
// b/c they only make sense if reneighboring is actually performed
if (neighbor->build_once) domain->reset_box();
// natoms = sum of nlocal = value to write into restart file
// if unequal and thermo lostflag is "error", don't write restart file
bigint nblocal = atom->nlocal;
MPI_Allreduce(&nblocal,&natoms,1,MPI_LMP_BIGINT,MPI_SUM,world);
if (natoms != atom->natoms && output->thermo->lostflag == ERROR)
error->all(FLERR,"Atom count is inconsistent, cannot write restart file");
// open single restart file or base file for multiproc case
if (me == 0) {
char *hfile;
if (multiproc) {
hfile = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(hfile,"%s%s%s",file,"base",ptr+1);
*ptr = '%';
} else hfile = file;
fp = fopen(hfile,"wb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",hfile);
error->one(FLERR,str);
}
if (multiproc) delete [] hfile;
}
// proc 0 writes magic string, endian flag, numeric version
if (me == 0) {
magic_string();
endian();
version_numeric();
}
// proc 0 writes header, groups, pertype info, force field info
if (me == 0) {
header();
group->write_restart(fp);
type_arrays();
force_fields();
}
// all procs write fix info
modify->write_restart(fp);
// communication buffer for my atom info
// max_size = largest buffer needed by any proc
int max_size;
int send_size = atom->avec->size_restart();
MPI_Allreduce(&send_size,&max_size,1,MPI_INT,MPI_MAX,world);
double *buf;
memory->create(buf,max_size,"write_restart:buf");
// all procs write file layout info which may include per-proc sizes
file_layout(send_size);
// header info is complete
// if multiproc output:
// close header file, open multiname file on each writing proc,
// write PROCSPERFILE into new file
if (multiproc) {
if (me == 0) fclose(fp);
char *multiname = new char[strlen(file) + 16];
char *ptr = strchr(file,'%');
*ptr = '\0';
sprintf(multiname,"%s%d%s",file,icluster,ptr+1);
*ptr = '%';
if (filewriter) {
fp = fopen(multiname,"wb");
if (fp == NULL) {
char str[128];
sprintf(str,"Cannot open restart file %s",multiname);
error->one(FLERR,str);
}
write_int(PROCSPERFILE,nclusterprocs);
}
delete [] multiname;
}
// pack my atom data into buf
AtomVec *avec = atom->avec;
int n = 0;
for (int i = 0; i < atom->nlocal; i++) n += avec->pack_restart(i,&buf[n]);
// if any fix requires it, remap each atom's coords via PBC
// is because fix changes atom coords (excepting an integrate fix)
// just remap in buffer, not actual atoms
if (modify->restart_pbc_any) {
int triclinic = domain->triclinic;
double *lo,*hi,*period;
if (triclinic == 0) {
lo = domain->boxlo;
hi = domain->boxhi;
period = domain->prd;
} else {
lo = domain->boxlo_lamda;
hi = domain->boxhi_lamda;
period = domain->prd_lamda;
}
int xperiodic = domain->xperiodic;
int yperiodic = domain->yperiodic;
int zperiodic = domain->zperiodic;
double *x;
int m = 0;
for (int i = 0; i < atom->nlocal; i++) {
x = &buf[m+1];
if (triclinic) domain->x2lamda(x,x);
if (xperiodic) {
if (x[0] < lo[0]) x[0] += period[0];
if (x[0] >= hi[0]) x[0] -= period[0];
x[0] = MAX(x[0],lo[0]);
}
if (yperiodic) {
if (x[1] < lo[1]) x[1] += period[1];
if (x[1] >= hi[1]) x[1] -= period[1];
x[1] = MAX(x[1],lo[1]);
}
if (zperiodic) {
if (x[2] < lo[2]) x[2] += period[2];
if (x[2] >= hi[2]) x[2] -= period[2];
x[2] = MAX(x[2],lo[2]);
}
if (triclinic) domain->lamda2x(x,x);
m += static_cast<int> (buf[m]);
}
}
// MPI-IO output to single file
if (mpiioflag) {
if (me == 0) fclose(fp);
mpiio->openForWrite(file);
mpiio->write(headerOffset,send_size,buf);
mpiio->close();
}
// output of one or more native files
// filewriter = 1 = this proc writes to file
// ping each proc in my cluster, receive its data, write data to file
// else wait for ping from fileproc, send my data to fileproc
else {
int tmp,recv_size;
MPI_Status status;
MPI_Request request;
if (filewriter) {
for (int iproc = 0; iproc < nclusterprocs; iproc++) {
if (iproc) {
MPI_Irecv(buf,max_size,MPI_DOUBLE,me+iproc,0,world,&request);
MPI_Send(&tmp,0,MPI_INT,me+iproc,0,world);
MPI_Wait(&request,&status);
MPI_Get_count(&status,MPI_DOUBLE,&recv_size);
} else recv_size = send_size;
write_double_vec(PERPROC,recv_size,buf);
}
fclose(fp);
} else {
MPI_Recv(&tmp,0,MPI_INT,fileproc,0,world,&status);
MPI_Rsend(buf,send_size,MPI_DOUBLE,fileproc,0,world);
}
}
// clean up
memory->destroy(buf);
// invoke any fixes that write their own restart file
for (int ifix = 0; ifix < modify->nfix; ifix++)
if (modify->fix[ifix]->restart_file)
modify->fix[ifix]->write_restart_file(file);
}
