paracel::deque_type< paracel::coroutine<paracel::str_type> >
files_partition(paracel::list_type<paracel::str_type> & name_list,
int np, const paracel::str_type & pattern = "",
int blk_sz = BLK_SZ) {
paracel::deque_type< paracel::coroutine<paracel::str_type> > func_loaders;
if(pattern == "linesplit" || pattern == "fvec") blk_sz = 1;
np = np * blk_sz;
paracel::list_type<long> displs(name_list.size() + 1, 0);
for(size_t i = 0; i < displs.size() - 1; ++i) {
std::ifstream f(name_list[i], std::ios::binary | std::ios::ate);
long tmp = f.tellg();
displs[i + 1] = displs[i] + tmp;
}
long sz = displs[displs.size() - 1];
int nbk = np;
long bk_sz = sz / (long)nbk;
long e;
for(int i = 0; i < nbk; ++i) {
long s = (long)i * bk_sz;
if(i == nbk -1) {
e = sz;
} else {
e = (i + 1) * bk_sz;
}
paracel::coroutine<paracel::str_type> files_load_lines(std::bind(paracel::files_load_lines_impl, std::placeholders::_1, name_list, displs, s, e));
func_loaders.push_back(std::move(files_load_lines));
}
return func_loaders;
}
std::vector<int> scatter_vector_block_decomp(std::vector<int>& global_vec, MPI_Comm comm)
{
// get MPI Communicator properties
int rank, p;
MPI_Comm_size(comm, &p);
MPI_Comm_rank(comm, &rank);
// the local vector size (MPI restricts message sizes to `int`)
int local_size;
// init result
std::vector<int> local_elements;
if (rank == 0)
{
/* I am the root process */
// get size of global array and bcast
unsigned long long n = global_vec.size();
MPI_Bcast(&n, 1, MPI_UNSIGNED_LONG_LONG, 0, comm);
local_size = block_decompose(n, p, rank);
// scatter-v the actual data
std::vector<int> counts(p);
std::vector<int> displs(p);
displs[0] = 0;
for (int i = 0; i < p; ++i)
{
counts[i] = block_decompose(n, p, i);
if (i > 0)
displs[i] = displs[i-1] + counts[i-1];
}
local_elements.resize(local_size);
MPI_Scatterv(&global_vec[0], &counts[0], &displs[0], MPI_INT,
&local_elements[0], local_size, MPI_INT, 0, comm);
}
else
{
/* I am NOT the root process */
// receive the size of my local array
unsigned long long n;
MPI_Bcast(&n, 1, MPI_UNSIGNED_LONG_LONG, 0, comm);
local_size = block_decompose(n, p, rank);
// resize result buffer
local_elements.resize(local_size);
// actually receive all the data
MPI_Scatterv(NULL, NULL, NULL, MPI_INT,
&local_elements[0], local_size, MPI_INT, 0, comm);
}
// return local array
return local_elements;
}
std::vector<int> gather_vectors(std::vector<int>& local_vec, MPI_Comm comm)
{
// get MPI parameters
int rank;
int p;
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &p);
// get local size
int local_size = local_vec.size();
// init result
std::vector<int> result;
// master process: receive results
if (rank == 0)
{
// gather local array sizes, sizes are restricted to `int` by MPI anyway
// therefore use int
std::vector<int> local_sizes(p);
MPI_Gather(&local_size, 1, MPI_INT, &local_sizes[0], 1, MPI_INT, 0, comm);
// gather-v to collect all the elements
int total_size = std::accumulate(local_sizes.begin(), local_sizes.end(), 0);
result.resize(total_size);
// get receive displacements
std::vector<int> displs(p, 0);
for (int i = 1; i < p; ++i)
displs[i] = displs[i-1] + local_sizes[i-1];
// gather v the vector data to the root
MPI_Gatherv(&local_vec[0], local_size, MPI_INT,
&result[0], &local_sizes[0], &displs[0], MPI_INT, 0, comm);
}
// else: send results
else {
// gather local array sizes
MPI_Gather(&local_size, 1, MPI_INT, NULL, 1, MPI_INT, 0, comm);
// sent the actual data
MPI_Gatherv(&local_vec[0], local_size, MPI_INT,
NULL, NULL, NULL, MPI_INT, 0, comm);
}
return result;
}
vector<vector<int> > MPIDistribution::MPILayerGatherVariables(vector<vector<int> > data)
{
MPI_Comm* comm = MPIGetCommLayer(); // communicator for nodes in layer
vector<vector<int> > gatheredData(data.size());
for(int i=0;i<data.size();i++)
{
data[i].push_back(0); // will be removed later, allgatherv may not like zero-sized vectors
int sizeLocal = data[i].size();
int totalSize;
MPI_Allreduce(&sizeLocal,&totalSize, 1,MPI_INT,MPI_SUM,*comm);
vector<int> totalIndexes(totalSize);
vector<int> rcounts(m_mpiSizeLocal);
vector<int> displs(m_mpiSizeLocal);
// counts from each node
MPI_Allgather( &sizeLocal, 1, MPI_INT, &rcounts[0], 1, MPI_INT, *comm);
for(int j=1;j<rcounts.size();j++)
{
displs[j] = rcounts[j-1]+displs[j-1];
}
MPI_Allgatherv( &data[i][0],data[i].size(),MPI_INT,&totalIndexes[0],&rcounts[0],&displs[0],MPI_INT,*comm);
// remove all artificial values
for(int j=rcounts.size()-1;j>-1;j--)
{
totalIndexes.erase(totalIndexes.begin() + displs[j] + rcounts[j] - 1);
}
gatheredData[i] = totalIndexes;
}
return gatheredData;
}