//As per the standard, perform a barrier to ensure every async comm is finished
bar_->wait();
- int finished = finish_comms();
- XBT_DEBUG("Win destructor - Finished %d RMA calls", finished);
+ flush_local_all();
if (info_ != MPI_INFO_NULL)
simgrid::smpi::Info::unref(info_);
if (not (assert & MPI_MODE_NOPRECEDE)) {
// This is not the first fence => finalize what came before
bar_->wait();
- mut_->lock();
- // This (simulated) mutex ensures that no process pushes to the vector of requests during the waitall.
- // Without this, the vector could get redimensioned when another process pushes.
- // This would result in the array used by Request::waitall() to be invalidated.
- // Another solution would be to copy the data and cleanup the vector *before* Request::waitall
-
- // start all requests that have been prepared by another process
- if (not requests_.empty()) {
- int size = static_cast<int>(requests_.size());
- MPI_Request* treqs = requests_.data();
- Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
- }
+ flush_local_all();
count_=0;
- mut_->unlock();
}
if (assert & MPI_MODE_NOSUCCEED) // there should be no ops after this one, tell we are closed.
void* recv_addr = static_cast<char*>(recv_win->base_) + target_disp * recv_win->disp_unit_;
- if (target_rank != comm_->rank()) { // This is not for myself, so we need to send messages
+ if (target_rank != rank_) { // This is not for myself, so we need to send messages
XBT_DEBUG("Entering MPI_Put to remote rank %d", target_rank);
// prepare send_request
MPI_Request sreq =
- // TODO cheinrich Check for rank / pid conversion
- Request::rma_send_init(origin_addr, origin_count, origin_datatype, comm_->rank(), target_rank, SMPI_RMA_TAG + 1,
- comm_, MPI_OP_NULL);
+ Request::rma_send_init(origin_addr, origin_count, origin_datatype, rank_, target_rank, SMPI_RMA_TAG + 1, comm_,
+ MPI_OP_NULL);
//prepare receiver request
- // TODO cheinrich Check for rank / pid conversion
- MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, recv_win->comm_->rank(),
- target_rank, SMPI_RMA_TAG + 1, recv_win->comm_, MPI_OP_NULL);
+ MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, rank_, target_rank,
+ SMPI_RMA_TAG + 1, recv_win->comm_, MPI_OP_NULL);
//start send
sreq->start();
const void* send_addr = static_cast<void*>(static_cast<char*>(send_win->base_) + target_disp * send_win->disp_unit_);
XBT_DEBUG("Entering MPI_Get from %d", target_rank);
- if(target_rank != comm_->rank()){
+ if (target_rank != rank_) {
//prepare send_request
- MPI_Request sreq = Request::rma_send_init(send_addr, target_count, target_datatype, target_rank,
- send_win->comm_->rank(), SMPI_RMA_TAG + 2, send_win->comm_, MPI_OP_NULL);
+ MPI_Request sreq = Request::rma_send_init(send_addr, target_count, target_datatype, target_rank, rank_,
+ SMPI_RMA_TAG + 2, send_win->comm_, MPI_OP_NULL);
//prepare receiver request
- MPI_Request rreq = Request::rma_recv_init(
- origin_addr, origin_count, origin_datatype, target_rank,
- comm_->rank(), // TODO cheinrich Check here if comm_->rank() and above send_win->comm_->rank() are correct
- SMPI_RMA_TAG + 2, comm_, MPI_OP_NULL);
+ MPI_Request rreq = Request::rma_recv_init(origin_addr, origin_count, origin_datatype, target_rank, rank_,
+ SMPI_RMA_TAG + 2, comm_, MPI_OP_NULL);
//start the send, with another process than us as sender.
sreq->start();
- //push request to receiver's win
+ // push request to sender's win
send_win->mut_->lock();
send_win->requests_.push_back(sreq);
send_win->mut_->unlock();
// SMPI tags, SMPI_RMA_TAG is set below all the other ones we use)
// prepare send_request
- MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, comm_->rank(), target_rank,
+ MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, rank_, target_rank,
SMPI_RMA_TAG - 3 - count_, comm_, op);
// prepare receiver request
- MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, recv_win->comm_->rank(),
- recv_win->comm_->group()->rank(comm_->group()->actor(target_rank)),
+ MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, rank_, target_rank,
SMPI_RMA_TAG - 3 - count_, recv_win->comm_, op);
count_++;
XBT_DEBUG("Entering MPI_Get_accumulate from %d", target_rank);
//need to be sure ops are correctly ordered, so finish request here ? slow.
- MPI_Request req;
+ MPI_Request req = MPI_REQUEST_NULL;
send_win->atomic_mut_->lock();
get(result_addr, result_count, result_datatype, target_rank,
target_disp, target_count, target_datatype, &req);
for (auto& req : reqs)
Request::unref(&req);
- int finished = finish_comms();
- XBT_DEBUG("Win_complete - Finished %d RMA calls", finished);
+ flush_local_all();
opened_--; //we're closed for business !
Group::unref(dst_group_);
for (auto& req : reqs)
Request::unref(&req);
- int finished = finish_comms();
- XBT_DEBUG("Win_wait - Finished %d RMA calls", finished);
+ flush_local_all();
opened_--; //we're closed for business !
Group::unref(src_group_);
} else if (not(target_win->mode_ == MPI_LOCK_SHARED && lock_type == MPI_LOCK_EXCLUSIVE))
target_win->mode_ += lock_type; // don't set to exclusive if it's already shared
- target_win->lockers_.push_back(comm_->rank());
+ target_win->lockers_.push_back(rank_);
- int finished = finish_comms(rank);
- XBT_DEBUG("Win_lock %d - Finished %d RMA calls", rank, finished);
- finished = target_win->finish_comms(rank_);
- XBT_DEBUG("Win_lock target %d - Finished %d RMA calls", rank, finished);
+ flush(rank);
return MPI_SUCCESS;
}
MPI_Win target_win = connected_wins_[rank];
int target_mode = target_win->mode_;
target_win->mode_= 0;
- target_win->lockers_.remove(comm_->rank());
+ target_win->lockers_.remove(rank_);
if (target_mode==MPI_LOCK_EXCLUSIVE){
target_win->lock_mut_->unlock();
}
- int finished = finish_comms(rank);
- XBT_DEBUG("Win_unlock %d - Finished %d RMA calls", rank, finished);
- finished = target_win->finish_comms(rank_);
- XBT_DEBUG("Win_unlock target %d - Finished %d RMA calls", rank, finished);
+ flush(rank);
return MPI_SUCCESS;
}
}
int Win::flush(int rank){
- MPI_Win target_win = connected_wins_[rank];
- int finished = finish_comms(rank_);
- XBT_DEBUG("Win_flush on local %d - Finished %d RMA calls", rank_, finished);
- finished = target_win->finish_comms(rank);
- XBT_DEBUG("Win_flush on remote %d - Finished %d RMA calls", rank, finished);
+ int finished = finish_comms(rank);
+ XBT_DEBUG("Win_flush on local %d for remote %d - Finished %d RMA calls", rank_, rank, finished);
+ if (rank != rank_) {
+ finished = connected_wins_[rank]->finish_comms(rank_);
+ XBT_DEBUG("Win_flush on remote %d for local %d - Finished %d RMA calls", rank, rank_, finished);
+ }
return MPI_SUCCESS;
}
int Win::flush_local(int rank){
int finished = finish_comms(rank);
- XBT_DEBUG("Win_flush_local for rank %d - Finished %d RMA calls", rank, finished);
+ XBT_DEBUG("Win_flush_local on local %d for remote %d - Finished %d RMA calls", rank_, rank, finished);
return MPI_SUCCESS;
}
int Win::flush_all(){
int finished = finish_comms();
- XBT_DEBUG("Win_flush_all on local - Finished %d RMA calls", finished);
+ XBT_DEBUG("Win_flush_all on local %d - Finished %d RMA calls", rank_, finished);
for (int i = 0; i < comm_->size(); i++) {
- finished = connected_wins_[i]->finish_comms(rank_);
- XBT_DEBUG("Win_flush_all on %d - Finished %d RMA calls", i, finished);
+ if (i != rank_) {
+ finished = connected_wins_[i]->finish_comms(rank_);
+ XBT_DEBUG("Win_flush_all on remote %d for local %d - Finished %d RMA calls", i, rank_, finished);
+ }
}
return MPI_SUCCESS;
}
int Win::flush_local_all(){
int finished = finish_comms();
- XBT_DEBUG("Win_flush_local_all - Finished %d RMA calls", finished);
+ XBT_DEBUG("Win_flush_local_all on local %d - Finished %d RMA calls", rank_, finished);
return MPI_SUCCESS;
}
}
int Win::finish_comms(){
+ // This (simulated) mutex ensures that no process pushes to the vector of requests during the waitall.
+ // Without this, the vector could get redimensioned when another process pushes.
+ // This would result in the array used by Request::waitall() to be invalidated.
+ // Another solution would be to copy the data and cleanup the vector *before* Request::waitall
mut_->lock();
//Finish own requests
int size = static_cast<int>(requests_.size());
}
int Win::finish_comms(int rank){
+ // See comment about the mutex in finish_comms() above
mut_->lock();
// Finish own requests
// Let's see if we're either the destination or the sender of this request