SocketSyncCPU<Dtype>::SocketSyncCPU(shared_ptr<Solver<Dtype> > root_solver,
const vector<shared_ptr<SocketChannel> > &
peers, int rank)
: P2PSyncCPU<Dtype>(root_solver, NULL, root_solver->param()),
peers_(peers),
rank_(rank),
data_send_(peers.size()),
data_recv_(peers.size()),
diff_send_(peers.size()),
diff_recv_(peers.size()) {
chunk(rank_, &own_offs_, &own_size_);
for (int peer = 0; peer < peers_.size(); ++peer) {
if (peer == rank_) {
// Chunk for which we are master, connected to all peers.
// Loops must be imbricated to have buffers created in
// the same order on all boxes.
for (int i = 0; i < peers_.size(); ++i) {
if (i != rank_) {
CreateMasterBuffers(i);
}
}
} else {
// Other chunks are connected to their respective masters
CreateWorkerBuffers(peer);
}
}
}
SocketSync<Dtype>::SocketSync(shared_ptr<Solver<Dtype> > root_solver,
const vector<shared_ptr<SocketChannel> > &
peers, int rank)
: P2PSync<Dtype>(root_solver, NULL, root_solver->param()),
peers_(peers),
rank_(rank),
data_send_(peers.size()),
data_recv_(peers.size()),
diff_send_(peers.size()),
diff_recv_(peers.size()),
ctrl_send_(peers.size()),
ctrl_recv_(peers.size()) {
#ifndef CPU_ONLY
int initial_device;
CUDA_CHECK(cudaGetDevice(&initial_device));
CUDA_CHECK(cudaSetDevice(root_solver->param().device_id()));
chunk(rank_, &own_offs_, &own_size_);
for (int peer = 0; peer < peers_.size(); ++peer) {
if (peer == rank_) {
// Chunk for which we are master, connected to all peers.
// Loops must be imbricated to have buffers created in
// the same order on all boxes.
for (int i = 0; i < peers_.size(); ++i) {
if (i != rank_) {
CreateMasterBuffers(i);
}
}
} else {
// Other chunks are connected to their respective masters
CreateWorkerBuffers(peer);
}
}
CUDA_CHECK(cudaSetDevice(initial_device));
#else
NO_GPU;
#endif
}
