-/* Copyright (c) 2007-2017. The SimGrid Team.
- * All rights reserved. */
+/* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
-#include "private.h"
-#include <vector>
+#include "src/smpi/private.h"
+#include "src/smpi/smpi_coll.hpp"
+#include "src/smpi/smpi_comm.hpp"
+#include "src/smpi/smpi_datatype.hpp"
+#include "src/smpi/smpi_info.hpp"
+#include "src/smpi/smpi_keyvals.hpp"
+#include "src/smpi/smpi_process.hpp"
+#include "src/smpi/smpi_request.hpp"
+#include "src/smpi/smpi_win.hpp"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)");
requests_ = new std::vector<MPI_Request>();
mut_=xbt_mutex_init();
lock_mut_=xbt_mutex_init();
+ atomic_mut_=xbt_mutex_init();
connected_wins_ = new MPI_Win[comm_size];
connected_wins_[rank_] = this;
count_ = 0;
MSG_barrier_destroy(bar_);
xbt_mutex_destroy(mut_);
xbt_mutex_destroy(lock_mut_);
+ xbt_mutex_destroy(atomic_mut_);
if(allocated_ !=0)
xbt_free(base_);
}
int Win::attach (void *base, MPI_Aint size){
- if (!(base_ == MPI_BOTTOM || base_ == 0))
+ if (not(base_ == MPI_BOTTOM || base_ == 0))
return MPI_ERR_ARG;
base_=0;//actually the address will be given in the RMA calls, as being the disp.
size_+=size;
}
int Win::put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank,
- MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype)
+ MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Request* request)
{
//get receiver pointer
MPI_Win recv_win = connected_wins_[target_rank];
//start send
sreq->start();
+
+ if(request!=nullptr){
+ *request=sreq;
+ }else{
+ xbt_mutex_acquire(mut_);
+ requests_->push_back(sreq);
+ xbt_mutex_release(mut_);
+ }
+
//push request to receiver's win
xbt_mutex_acquire(recv_win->mut_);
recv_win->requests_->push_back(rreq);
rreq->start();
xbt_mutex_release(recv_win->mut_);
- //push request to sender's win
- xbt_mutex_acquire(mut_);
- requests_->push_back(sreq);
- xbt_mutex_release(mut_);
+
}else{
Datatype::copy(origin_addr, origin_count, origin_datatype, recv_addr, target_count, target_datatype);
+ if(request!=nullptr)
+ *request = MPI_REQUEST_NULL;
}
return MPI_SUCCESS;
}
int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank,
- MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype)
+ MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Request* request)
{
//get sender pointer
MPI_Win send_win = connected_wins_[target_rank];
comm_->group()->index(target_rank), smpi_process()->index(), SMPI_RMA_TAG+2, comm_,
MPI_OP_NULL);
- //start the send, with another process than us as sender.
+ //start the send, with another process than us as sender.
sreq->start();
//push request to receiver's win
xbt_mutex_acquire(send_win->mut_);
//start recv
rreq->start();
- //push request to sender's win
- xbt_mutex_acquire(mut_);
- requests_->push_back(rreq);
- xbt_mutex_release(mut_);
+
+ if(request!=nullptr){
+ *request=rreq;
+ }else{
+ xbt_mutex_acquire(mut_);
+ requests_->push_back(rreq);
+ xbt_mutex_release(mut_);
+ }
+
}else{
Datatype::copy(send_addr, target_count, target_datatype, origin_addr, origin_count, origin_datatype);
+ if(request!=nullptr)
+ *request=MPI_REQUEST_NULL;
}
return MPI_SUCCESS;
int Win::accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank,
- MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Op op)
+ MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Op op, MPI_Request* request)
{
//get receiver pointer
if(locked != 1)
return MPI_ERR_WIN;
}
- //FIXME: local version
+ //FIXME: local version
if(target_count*target_datatype->get_extent()>recv_win->size_)
return MPI_ERR_ARG;
recv_win->requests_->push_back(rreq);
rreq->start();
xbt_mutex_release(recv_win->mut_);
- //push request to sender's win
- xbt_mutex_acquire(mut_);
- requests_->push_back(sreq);
- xbt_mutex_release(mut_);
+
+ if(request!=nullptr){
+ *request=sreq;
+ }else{
+ xbt_mutex_acquire(mut_);
+ requests_->push_back(sreq);
+ xbt_mutex_release(mut_);
+ }
return MPI_SUCCESS;
}
-int Win::get_accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, void *result_addr,
- int result_count, MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp, int target_count,
- MPI_Datatype target_datatype, MPI_Op op){
+int Win::get_accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, void *result_addr,
+ int result_count, MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp, int target_count,
+ MPI_Datatype target_datatype, MPI_Op op, MPI_Request* request){
//get sender pointer
MPI_Win send_win = connected_wins_[target_rank];
return MPI_ERR_ARG;
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;
+ xbt_mutex_acquire(send_win->atomic_mut_);
get(result_addr, result_count, result_datatype, target_rank,
- target_disp, target_count, target_datatype);
+ target_disp, target_count, target_datatype, &req);
+ if (req != MPI_REQUEST_NULL)
+ Request::wait(&req, MPI_STATUS_IGNORE);
if(op!=MPI_NO_OP)
accumulate(origin_addr, origin_count, origin_datatype, target_rank,
- target_disp, target_count, target_datatype, op);
-
+ target_disp, target_count, target_datatype, op, &req);
+ if (req != MPI_REQUEST_NULL)
+ Request::wait(&req, MPI_STATUS_IGNORE);
+ xbt_mutex_release(send_win->atomic_mut_);
return MPI_SUCCESS;
}
+int Win::compare_and_swap(void *origin_addr, void *compare_addr,
+ void *result_addr, MPI_Datatype datatype, int target_rank,
+ MPI_Aint target_disp){
+ //get sender pointer
+ MPI_Win send_win = connected_wins_[target_rank];
+
+ if(opened_==0){//check that post/start has been done
+ // no fence or start .. lock ok ?
+ int locked=0;
+ for(auto it : send_win->lockers_)
+ if (it == comm_->rank())
+ locked = 1;
+ if(locked != 1)
+ return MPI_ERR_WIN;
+ }
+
+ XBT_DEBUG("Entering MPI_Compare_and_swap with %d", target_rank);
+ MPI_Request req;
+ xbt_mutex_acquire(send_win->atomic_mut_);
+ get(result_addr, 1, datatype, target_rank,
+ target_disp, 1, datatype, &req);
+ if (req != MPI_REQUEST_NULL)
+ Request::wait(&req, MPI_STATUS_IGNORE);
+ if (not memcmp(result_addr, compare_addr, datatype->get_extent())) {
+ put(origin_addr, 1, datatype, target_rank,
+ target_disp, 1, datatype);
+ }
+ xbt_mutex_release(send_win->atomic_mut_);
+ return MPI_SUCCESS;
+}
+
int Win::start(MPI_Group group, int assert){
/* From MPI forum advices
The call to MPI_WIN_COMPLETE does not return until the put call has completed at the origin; and the target window
}
int Win::lock(int lock_type, int rank, int assert){
- if(opened_!=0)
- return MPI_ERR_WIN;
-
MPI_Win target_win = connected_wins_[rank];
if ((lock_type == MPI_LOCK_EXCLUSIVE && target_win->mode_ != MPI_LOCK_SHARED)|| target_win->mode_ == MPI_LOCK_EXCLUSIVE){
if(lock_type == MPI_LOCK_SHARED){//the window used to be exclusive, it's now shared.
xbt_mutex_release(target_win->lock_mut_);
}
- } else if(!(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
+ } 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());
- int finished = finish_comms();
+ 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);
return MPI_SUCCESS;
}
}
int Win::unlock(int rank){
- if(opened_!=0)
- return MPI_ERR_WIN;
-
MPI_Win target_win = connected_wins_[rank];
int target_mode = target_win->mode_;
target_win->mode_= 0;
xbt_mutex_release(target_win->lock_mut_);
}
- int finished = finish_comms();
+ 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);
return MPI_SUCCESS;
}
int size = static_cast<int>(reqqs->size());
if (size > 0) {
size = 0;
- std::vector<MPI_Request>* myreqqs = new std::vector<MPI_Request>();
+ std::vector<MPI_Request> myreqqs;
std::vector<MPI_Request>::iterator iter = reqqs->begin();
while (iter != reqqs->end()){
- if(((*iter)->src() == rank) || ((*iter)->dst() == rank)){
- myreqqs->push_back(*iter);
- iter = reqqs->erase(iter);
- size++;
+ if(((*iter)!=MPI_REQUEST_NULL) && (((*iter)->src() == rank) || ((*iter)->dst() == rank))){
+ myreqqs.push_back(*iter);
+ iter = reqqs->erase(iter);
+ size++;
} else {
++iter;
}
}
if(size >0){
- MPI_Request* treqs = &(*myreqqs)[0];
+ MPI_Request* treqs = &myreqqs[0];
Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
- myreqqs->clear();
- delete myreqqs;
+ myreqqs.clear();
}
}
xbt_mutex_release(mut_);
