/*
* Broadcast a buffer.
*/
void Buffer::bcast(MPI::Intracomm& comm, int source)
{
int comm_size = comm.Get_size();
int myRank = comm.Get_rank();
if (source > comm_size - 1 || source < 0) {
UTIL_THROW("Source rank out of bounds");
}
int sendBytes;
if (myRank == source) {
sendBytes = sendPtr_ - sendBufferBegin_;
comm.Bcast(&sendBytes, 1, MPI::INT, source);
comm.Bcast(sendBufferBegin_, sendBytes, MPI::CHAR, source);
sendPtr_ = sendBufferBegin_;
sendType_ = NONE;
} else {
comm.Bcast(&sendBytes, 1, MPI::INT, source);
comm.Bcast(recvBufferBegin_, sendBytes, MPI::CHAR, source);
recvPtr_ = recvBufferBegin_;
recvType_ = NONE;
}
if (sendBytes > maxSendLocal_) {
maxSendLocal_ = sendBytes;
}
}
/// Octree情報を他rankにブロードキャスト.
void BCMOctree::broadcast(MPI::Intracomm& comm)
{
assert(comm.Get_rank() == 0);
rootGrid->broadcast(comm);
int numLeafNode = leafNodeArray.size();
int ibuf[2];
ibuf[0] = numLeafNode;
ibuf[1] = ordering;
comm.Bcast(&ibuf, 2, MPI::INT, 0);
size_t size = Pedigree::GetSerializeSize();
unsigned char* buf = new unsigned char[size * numLeafNode];
size_t ip = 0;
for (int id = 0; id < rootGrid->getSize(); id++) {
packPedigrees(rootNodes[id], ip, buf);
}
comm.Bcast(buf, size*numLeafNode, MPI::BYTE, 0);
delete[] buf;
}
ifstream& FullyDistSpVec<IT,NT>::ReadDistribute (ifstream& infile, int master)
{
IT total_nnz;
MPI::Intracomm World = commGrid->GetWorld();
int neighs = World.Get_size(); // number of neighbors (including oneself)
int buffperneigh = MEMORYINBYTES / (neighs * (sizeof(IT) + sizeof(NT)));
int * displs = new int[neighs];
for (int i=0; i<neighs; ++i)
displs[i] = i*buffperneigh;
int * curptrs = NULL;
int recvcount = 0;
IT * inds = NULL;
NT * vals = NULL;
int rank = World.Get_rank();
if(rank == master) // 1 processor only
{
inds = new IT [ buffperneigh * neighs ];
vals = new NT [ buffperneigh * neighs ];
curptrs = new int[neighs];
fill_n(curptrs, neighs, 0); // fill with zero
if (infile.is_open())
{
infile.clear();
infile.seekg(0);
infile >> glen >> total_nnz;
World.Bcast(&glen, 1, MPIType<IT>(), master);
IT tempind;
NT tempval;
double loadval;
IT cnz = 0;
while ( (!infile.eof()) && cnz < total_nnz)
{
infile >> tempind;
//infile >> tempval;
infile >> loadval;
tempval = static_cast<NT>(loadval);
tempind--;
IT locind;
int rec = Owner(tempind, locind); // recipient (owner) processor
inds[ rec * buffperneigh + curptrs[rec] ] = locind;
vals[ rec * buffperneigh + curptrs[rec] ] = tempval;
++ (curptrs[rec]);
if(curptrs[rec] == buffperneigh || (cnz == (total_nnz-1)) ) // one buffer is full, or file is done !
{
// first, send the receive counts ...
World.Scatter(curptrs, 1, MPI::INT, &recvcount, 1, MPI::INT, master);
// generate space for own recv data ... (use arrays because vector<bool> is cripled, if NT=bool)
IT * tempinds = new IT[recvcount];
NT * tempvals = new NT[recvcount];
// then, send all buffers that to their recipients ...
World.Scatterv(inds, curptrs, displs, MPIType<IT>(), tempinds, recvcount, MPIType<IT>(), master);
World.Scatterv(vals, curptrs, displs, MPIType<NT>(), tempvals, recvcount, MPIType<NT>(), master);
// now push what is ours to tuples
for(IT i=0; i< recvcount; ++i)
{
ind.push_back( tempinds[i] ); // already offset'd by the sender
num.push_back( tempvals[i] );
}
// reset current pointers so that we can reuse {inds,vals} buffers
fill_n(curptrs, neighs, 0);
DeleteAll(tempinds, tempvals);
}
++ cnz;
}
assert (cnz == total_nnz);
// Signal the end of file to other processors along the diagonal
fill_n(curptrs, neighs, numeric_limits<int>::max());
World.Scatter(curptrs, 1, MPI::INT, &recvcount, 1, MPI::INT, master);
}