Skip to content
Snippets Groups Projects
Commit 05122845 authored by Rene Halver's avatar Rene Halver
Browse files

merge fixes

parents e3584bf9 d781c2c4
Branches
Tags
No related merge requests found
Pipeline #17487 passed
Stage: build
Show whitespace changes
Inline Side-by-side
example/ALL_test.cpp
+ 157
− 138
View file @ 05122845
......@@ -140,7 +140,9 @@ void read_points(std::vector<ALL_Point<double>>& points,
MPI_Comm comm)
{
MPI_File file;
MPI_Barrier(comm);
int err = MPI_File_open(comm, filename, MPI_MODE_RDONLY, MPI_INFO_NULL, &file);
MPI_Barrier(comm);
n_points = 0;
......@@ -404,6 +406,7 @@ int main(int argc, char** argv)
// setup of vector of points on each MPI rank
std::vector<double> dummy(sys_dim);
std::vector<ALL_Point<double>> points;
int max_neighbors, loc_neighbors;
// setup of cartesian communicator
int local_rank;
......@@ -525,6 +528,7 @@ int main(int argc, char** argv)
ALL_Point<double> up(u);
std::vector<ALL_Point<double>> vertices(nvertices,lp);
std::vector<ALL_Point<double>> new_vertices(nvertices,lp);
std::vector<ALL_Point<double>> old_vertices(nvertices,lp);
switch(chosen_method)
{
......@@ -595,9 +599,11 @@ int main(int argc, char** argv)
MPI_Allreduce(&n_points,&max_particles,1,MPI_INT,MPI_MAX,cart_comm);
//if (local_rank == 0)
// std::cout << "maximum number of points on any process: "
// << max_particles << std::endl;
max_particles = (int)std::ceil((double)max_particles * 1.5);
if (local_rank == 0)
std::cout << "maximum number of points on any process: "
<< max_particles << std::endl;
int max_neig = 27;
recv = new double[max_neig * (sys_dim+1) * max_particles];
......@@ -1004,13 +1010,15 @@ int main(int argc, char** argv)
MPI_Barrier(cart_comm);
if (d_ratio < limit_efficiency)
{
gamma /= 2.0;
gamma *= 2.0;
limit_efficiency /= 2.0;
}
if (local_rank == 0) std::cout << "loop " << i_loop << ": " << std::endl;
std::vector<double> work;
std::vector<int> n_bins(3,-1);
double histogram_width = ALL_HISTOGRAM_DEFAULT_WIDTH / (double)(i_loop + 1);
//double histogram_width = ALL_HISTOGRAM_DEFAULT_WIDTH / (double)(i_loop + 1);
double histogram_width = ALL_HISTOGRAM_DEFAULT_WIDTH / gamma;
if (!weighted_points)
{
......@@ -1037,58 +1045,46 @@ int main(int argc, char** argv)
work = std::vector<double>(1,(double)n_points);
else
{
std::vector<int> n_bins(3,-1);
std::vector<double> lb(3,-1.0);
std::vector<double> ub(3,-1.0);
std::vector<double> overlap(3,0.0);
int offset = 0;
double lb(-1.0);
double ub(-1.0);
double overlap(0.0);
int d = 2 - i_loop % 3;
// compute number of bins in each direction
for (int d = 0; d < 3; ++d)
{
lb.at(d) = std::ceil(lp.x(d) / histogram_width) * histogram_width;
ub.at(d) = std::ceil(up.x(d) / histogram_width) * histogram_width;
n_bins.at(d) = (ub.at(3) - lb.at(3)) / histogram_width;
overlap.at(d) = 0.0;
}
work = std::vector<double>(n_bins[0]*
n_bins[1]*
n_bins[2], 0.0);
lb = std::ceil(lp.x(d) / histogram_width) * histogram_width;
ub = std::ceil(up.x(d) / histogram_width) * histogram_width;
n_bins.at(d) = (ub - lb) / histogram_width;
work = std::vector<double>(n_bins.at(d), 0.0);
// compute histogram of work load
for (int d = 0; d < 3; ++d)
{
for (auto p : points)
{
int idx = offset + (int)( ( p.x(d) - lb.at(d) / histogram_width ) );
int idx = (int)( ( ( p.x(d) - lb ) / histogram_width ) );
if (idx >= 0)
{
work.at(idx) += 1.0;
else
overlap.at(d) += 1.0;
}
offset += n_bins[d];
else
overlap += 1.0;
}
// exchange overlapping workload (histograms might overlap
// over the domain boundaries
offset = 0;
for (int d = 0; d < 3; ++d)
{
int rank_left, rank_right;
MPI_Cart_shift(cart_comm,0,1,&rank_left,&rank_right);
MPI_Request sreq, rreq;
MPI_Status ssta, rsta;
std::vector<double> recv_work(3);
double recv_work;
MPI_Isend(&overlap.at(d),
MPI_Isend(&overlap,
1,
MPI_DOUBLE,
rank_left,
0,
cart_comm,
&sreq);
MPI_Irecv(&recv_work.at(d),
MPI_Irecv(&recv_work,
1,
MPI_DOUBLE,
rank_right,
......@@ -1099,10 +1095,7 @@ int main(int argc, char** argv)
MPI_Wait(&rreq,&rsta);
if (local_coords[d] != global_dim[d] - 1)
work.at(offset + n_bins[d] - 1) += recv_work.at(d);
offset += n_bins[d];
}
work.at(n_bins.at(d) - 1) += recv_work;
}
}
else
......@@ -1116,60 +1109,46 @@ int main(int argc, char** argv)
}
else
{
std::vector<int> n_bins(3,-1);
std::vector<double> lb(3,-1.0);
std::vector<double> ub(3,-1.0);
std::vector<double> overlap(3,0.0);
int offset = 0;
double lb(-1.0);
double ub(-1.0);
double overlap(0.0);
int d = 2 - i_loop % 3;
// compute number of bins in each direction
for (int d = 0; d < 3; ++d)
{
lb.at(d) = std::ceil(lp.x(d) / histogram_width) * histogram_width;
ub.at(d) = std::ceil(up.x(d) / histogram_width) * histogram_width;
n_bins.at(d) = (ub.at(d) - lb.at(d)) / histogram_width;
overlap.at(d) = 0.0;
}
work = std::vector<double>(n_bins[0]*
n_bins[1]*
n_bins[2], 0.0);
lb = std::ceil(lp.x(d) / histogram_width) * histogram_width;
ub = std::ceil(up.x(d) / histogram_width) * histogram_width;
n_bins.at(d) = (ub - lb) / histogram_width;
work = std::vector<double>(n_bins.at(d), 0.0);
// compute histogram of work load
for (int d = 0; d < 3; ++d)
{
for (auto p : points)
{
int idx = offset + (int)( ( p.x(d) - lb.at(d) / histogram_width ) );
int idx = (int)( ( ( p.x(d) - lb ) / histogram_width ) );
if (idx >= 0)
{
work.at(idx) += p.get_weight();
}
else
overlap.at(d) += p.get_weight();
}
offset += n_bins[d];
overlap += p.get_weight();
}
// exchange overlapping workload (histograms might overlap
// over the domain boundaries
offset = 0;
for (int d = 0; d < 3; ++d)
{
int rank_left, rank_right;
MPI_Cart_shift(cart_comm,0,1,&rank_left,&rank_right);
MPI_Request sreq, rreq;
MPI_Status ssta, rsta;
std::vector<double> recv_work(3);
double recv_work;
MPI_Isend(&overlap.at(d),
MPI_Isend(&overlap,
1,
MPI_DOUBLE,
rank_left,
0,
cart_comm,
&sreq);
MPI_Irecv(&recv_work.at(d),
MPI_Irecv(&recv_work,
1,
MPI_DOUBLE,
rank_right,
......@@ -1180,9 +1159,7 @@ int main(int argc, char** argv)
MPI_Wait(&rreq,&rsta);
if (local_coords[d] != global_dim[d] - 1)
work.at(offset + n_bins[d] - 1) += recv_work.at(d);
offset += n_bins[d];
work.at(n_bins.at(d) - 1) += recv_work;
}
#ifdef ALL_VORONOI
if (chosen_method == ALL_LB_t::VORONOI)
......@@ -1204,12 +1181,17 @@ int main(int argc, char** argv)
}
#endif
}
}
MPI_Barrier(cart_comm);
if (local_rank == 0)
std::cout << "finished computation of work" << std::endl;
//lb_obj.set_work((double)n_points);
lb_obj.set_work(work);
lb_obj.set_communicator(cart_comm);
lb_obj.setup(chosen_method);
if (chosen_method == ALL_LB_t::HISTOGRAM)
lb_obj.set_method_data(chosen_method,n_bins.data());
lb_obj.set_sys_size(chosen_method, sys_size);
......@@ -1223,7 +1205,18 @@ int main(int argc, char** argv)
lb_obj.balance(chosen_method);
new_vertices = lb_obj.get_result_vertices();
old_vertices = vertices;
vertices = new_vertices;
std::cout << "vertices " << i_loop << " " << local_rank <<
" " << vertices.at(0).x(0) <<
" " << vertices.at(0).x(1) <<
" " << vertices.at(0).x(2) <<
"| " << vertices.at(1).x(0) <<
" " << vertices.at(1).x(1) <<
" " << vertices.at(1).x(2) <<
std::endl;
if (chosen_method == ALL_LB_t::VORONOI)
{
vertices.resize(2);
......@@ -1385,10 +1378,15 @@ int main(int argc, char** argv)
lb_obj.get_neighbors(chosen_method,neighbors);
lb_obj.get_neighbors(chosen_method,&n_neighbors);
offset_neig[0] = 0;
offset_neig[1] = n_neighbors[0];
loc_neighbors = 0;
for (int n = 0; n < 6; ++n)
loc_neighbors += n_neighbors[n];
MPI_Allreduce(&loc_neighbors,&max_neighbors,1,MPI_INT,MPI_MAX,cart_comm);
for (int i = 0; i < sys_dim; ++i)
{
......@@ -1415,6 +1413,7 @@ int main(int argc, char** argv)
{
n_transfer[j] = 0;
}
for (auto p = points.begin(); p != points.end(); ++p)
{
if (p->x(i) < vertices.at(0).x(i))
......@@ -1430,12 +1429,15 @@ int main(int argc, char** argv)
n_transfer[0]++;
if (n_transfer[0] > max_particles)
{
std::stringstream ss;
ss << "Trying to send more particles than buffer size allows! "
<< " n_transfer: " << n_transfer[0]
<< " max_particles: " << max_particles;
throw ALL_Invalid_Argument_Exception(
__FILE__,
__func__,
__LINE__,
"Trying to send more particles than buffer \
size allows!"
ss.str().c_str()
);
}
}
......@@ -1458,12 +1460,15 @@ int main(int argc, char** argv)
n_transfer[1]++;
if (n_transfer[1] > max_particles)
{
std::stringstream ss;
ss << "Trying to send more particles than buffer size allows! "
<< " n_transfer: " << n_transfer[0]
<< " max_particles: " << max_particles;
throw ALL_Invalid_Argument_Exception(
__FILE__,
__func__,
__LINE__,
"Trying to send more particles than buffer \
size allows!"
ss.str().c_str()
);
}
}
......@@ -1533,7 +1538,7 @@ int main(int argc, char** argv)
if (n_neighbors[2*i] > MAX_NEIG)
{
throw ALL_Invalid_Argument_Exception(
throw ALL_Internal_Error_Exception(
__FILE__,
__func__,
__LINE__,
......@@ -2187,6 +2192,16 @@ int main(int argc, char** argv)
#endif
}
break;
case ALL_LB_t::HISTOGRAM:
{
int curr_dim = 2 - (i_loop % 3);
// dertermine current dimension
// compare old domains with new domains
// transfer points from old domains to new domains
std::cout << "No Comm Yet" << std::endl;
break;
}
default:
break;
}
......@@ -2278,7 +2293,11 @@ int main(int argc, char** argv)
of.open("minmax.dat", std::ios::out | std::ios::app);
else
of.open("minmax_w.dat", std::ios::out | std::ios::app);
of << i_loop+1 << " " << d_min << " " << d_max << " " << d_ratio << std::endl;
of << i_loop+1 << " "
<< d_min << " "
<< d_max << " "
<< d_ratio << " "
<< max_neighbors << std::endl;
of.close();
if (i_loop % OUTPUT_INTV == 0)
{
......
include/ALL_Staggered.hpp
+ 6
− 1
View file @ 05122845
......@@ -143,7 +143,12 @@ template <class T, class W> ALL_Staggered_LB<T,W>::~ALL_Staggered_LB()
if (global_dims) delete global_dims;
if (local_coords) delete local_coords;
if (periodicity) delete periodicity;
if (communicators) delete communicators;
if (communicators)
{
for (int i = 1; i < 3; ++i)
MPI_Comm_free(communicators + i);
delete communicators;
}
if (n_neighbors) delete n_neighbors;
}
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment