XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)");
-#define RMA_TAG -1234
-
-xbt_bar_t creation_bar = nullptr;
-
typedef struct s_smpi_mpi_win{
void* base;
MPI_Aint size;
MPI_Info info;
int assert;
std::vector<MPI_Request> *requests;
- xbt_bar_t bar;
+ xbt_mutex_t mut;
+ msg_bar_t bar;
MPI_Win* connected_wins;
char* name;
int opened;
MPI_Group group;
+ int count; //for ordering the accs
} s_smpi_mpi_win_t;
win->opened = 0;
win->group = MPI_GROUP_NULL;
win->requests = new std::vector<MPI_Request>();
+ win->mut=xbt_mutex_init();
win->connected_wins = xbt_new0(MPI_Win, comm_size);
win->connected_wins[rank] = win;
-
+ win->count = 0;
if(rank==0){
- win->bar=xbt_barrier_init(comm_size);
+ win->bar = MSG_barrier_init(comm_size);
}
mpi_coll_allgather_fun(&(win->connected_wins[rank]), sizeof(MPI_Win), MPI_BYTE, win->connected_wins, sizeof(MPI_Win),
MPI_BYTE, comm);
- mpi_coll_bcast_fun( &(win->bar), sizeof(xbt_bar_t), MPI_BYTE, 0, comm);
+ mpi_coll_bcast_fun(&(win->bar), sizeof(msg_bar_t), MPI_BYTE, 0, comm);
mpi_coll_barrier_fun(comm);
int smpi_mpi_win_free( MPI_Win* win){
//As per the standard, perform a barrier to ensure every async comm is finished
- xbt_barrier_wait((*win)->bar);
+ MSG_barrier_wait((*win)->bar);
+ xbt_mutex_acquire((*win)->mut);
delete (*win)->requests;
+ xbt_mutex_release((*win)->mut);
xbt_free((*win)->connected_wins);
if ((*win)->name != nullptr){
xbt_free((*win)->name);
mpi_coll_barrier_fun((*win)->comm);
int rank=smpi_comm_rank((*win)->comm);
if(rank == 0)
- xbt_barrier_destroy((*win)->bar);
+ MSG_barrier_destroy((*win)->bar);
+ xbt_mutex_destroy((*win)->mut);
xbt_free(*win);
*win = MPI_WIN_NULL;
return MPI_SUCCESS;
void smpi_mpi_win_get_group(MPI_Win win, MPI_Group* group){
if(win->comm != MPI_COMM_NULL){
*group = smpi_comm_group(win->comm);
+ } else {
+ *group = MPI_GROUP_NULL;
}
}
win->name = xbt_strdup(name);
}
-int smpi_mpi_win_fence( int assert, MPI_Win win){
+int smpi_mpi_win_fence(int assert, MPI_Win win)
+{
XBT_DEBUG("Entering fence");
- if(win->opened==0)
+ if (win->opened == 0)
win->opened=1;
- if(assert != MPI_MODE_NOPRECEDE){
- xbt_barrier_wait(win->bar);
-
+ if (assert != MPI_MODE_NOPRECEDE) {
+ // This is not the first fence => finalize what came before
+ MSG_barrier_wait(win->bar);
+ xbt_mutex_acquire(win->mut);
+ // This (simulated) mutex ensures that no process pushes to the vector of requests during the waitall.
+ // Without this, the vector could get redimensionned when another process pushes.
+ // This would result in the array used by smpi_mpi_waitall() to be invalidated.
+ // Another solution would be to copy the data and cleanup the vector *before* smpi_mpi_waitall
std::vector<MPI_Request> *reqs = win->requests;
int size = static_cast<int>(reqs->size());
// start all requests that have been prepared by another process
- for(auto req: *reqs){
- if (req && (req->flags & PREPARED))
- smpi_mpi_start(req);
- }
+ if (size > 0) {
+ for (const auto& req : *reqs) {
+ if (req && (req->flags & PREPARED))
+ smpi_mpi_start(req);
+ }
- MPI_Request* treqs = &(*reqs)[0];
- smpi_mpi_waitall(size,treqs,MPI_STATUSES_IGNORE);
+ MPI_Request* treqs = &(*reqs)[0];
+
+ smpi_mpi_waitall(size, treqs, MPI_STATUSES_IGNORE);
+ }
+ win->count=0;
+ xbt_mutex_release(win->mut);
}
win->assert = assert;
- xbt_barrier_wait(win->bar);
- XBT_DEBUG("Leaving fence ");
+ MSG_barrier_wait(win->bar);
+ XBT_DEBUG("Leaving fence");
return MPI_SUCCESS;
}
if(target_rank != smpi_comm_rank(win->comm)){
//prepare send_request
MPI_Request sreq = smpi_rma_send_init(origin_addr, origin_count, origin_datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(win->comm),target_rank), RMA_TAG+1, win->comm, MPI_OP_NULL);
+ smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+1, win->comm, MPI_OP_NULL);
//prepare receiver request
MPI_Request rreq = smpi_rma_recv_init(recv_addr, target_count, target_datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(win->comm),target_rank), RMA_TAG+1, recv_win->comm, MPI_OP_NULL);
+ smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+1, recv_win->comm, MPI_OP_NULL);
//push request to receiver's win
+ xbt_mutex_acquire(recv_win->mut);
recv_win->requests->push_back(rreq);
-
+ xbt_mutex_release(recv_win->mut);
//start send
smpi_mpi_start(sreq);
//push request to sender's win
+ xbt_mutex_acquire(win->mut);
win->requests->push_back(sreq);
+ xbt_mutex_release(win->mut);
}else{
smpi_datatype_copy(origin_addr, origin_count, origin_datatype, recv_addr, target_count, target_datatype);
}
if(target_rank != smpi_comm_rank(win->comm)){
//prepare send_request
MPI_Request sreq = smpi_rma_send_init(send_addr, target_count, target_datatype,
- smpi_group_index(smpi_comm_group(win->comm),target_rank), smpi_process_index(), RMA_TAG+2, send_win->comm,
+ smpi_group_index(smpi_comm_group(win->comm),target_rank), smpi_process_index(), SMPI_RMA_TAG+2, send_win->comm,
MPI_OP_NULL);
//prepare receiver request
MPI_Request rreq = smpi_rma_recv_init(origin_addr, origin_count, origin_datatype,
- smpi_group_index(smpi_comm_group(win->comm),target_rank), smpi_process_index(), RMA_TAG+2, win->comm,
+ smpi_group_index(smpi_comm_group(win->comm),target_rank), smpi_process_index(), SMPI_RMA_TAG+2, win->comm,
MPI_OP_NULL);
//start the send, with another process than us as sender.
smpi_mpi_start(sreq);
-
//push request to receiver's win
+ xbt_mutex_acquire(send_win->mut);
send_win->requests->push_back(sreq);
+ xbt_mutex_release(send_win->mut);
//start recv
smpi_mpi_start(rreq);
-
//push request to sender's win
+ xbt_mutex_acquire(win->mut);
win->requests->push_back(rreq);
+ xbt_mutex_release(win->mut);
}else{
smpi_datatype_copy(send_addr, target_count, target_datatype, origin_addr, origin_count, origin_datatype);
}
void* recv_addr = static_cast<void*>(static_cast<char*>(recv_win->base) + target_disp * recv_win->disp_unit);
XBT_DEBUG("Entering MPI_Accumulate to %d", target_rank);
-
+ //As the tag will be used for ordering of the operations, add count to it
//prepare send_request
MPI_Request sreq = smpi_rma_send_init(origin_addr, origin_count, origin_datatype,
- smpi_process_index(), smpi_group_index(smpi_comm_group(win->comm),target_rank), RMA_TAG+3, win->comm, op);
+ smpi_process_index(), smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+3+win->count, win->comm, op);
//prepare receiver request
MPI_Request rreq = smpi_rma_recv_init(recv_addr, target_count, target_datatype,
- smpi_process_index(), smpi_group_index(smpi_comm_group(win->comm),target_rank), RMA_TAG+3, recv_win->comm, op);
+ smpi_process_index(), smpi_group_index(smpi_comm_group(win->comm),target_rank), SMPI_RMA_TAG+3+win->count, recv_win->comm, op);
+
+ win->count++;
//push request to receiver's win
+ xbt_mutex_acquire(recv_win->mut);
recv_win->requests->push_back(rreq);
+ xbt_mutex_release(recv_win->mut);
//start send
smpi_mpi_start(sreq);
//push request to sender's win
+ xbt_mutex_acquire(win->mut);
win->requests->push_back(sreq);
+ xbt_mutex_release(win->mut);
return MPI_SUCCESS;
}
while(j!=size){
int src=smpi_group_index(group,j);
- if(src!=smpi_process_index()){
- reqs[i]=smpi_irecv_init(nullptr, 0, MPI_CHAR, src,RMA_TAG+4, MPI_COMM_WORLD);
+ if(src!=smpi_process_index()&& src!=MPI_UNDEFINED){
+ reqs[i]=smpi_irecv_init(nullptr, 0, MPI_CHAR, src,SMPI_RMA_TAG+4, MPI_COMM_WORLD);
i++;
}
j++;
while(j!=size){
int dst=smpi_group_index(group,j);
- if(dst!=smpi_process_index()){
- reqs[i]=smpi_mpi_send_init(nullptr, 0, MPI_CHAR, dst, RMA_TAG+4, MPI_COMM_WORLD);
+ if(dst!=smpi_process_index() && dst!=MPI_UNDEFINED){
+ reqs[i]=smpi_mpi_send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+4, MPI_COMM_WORLD);
i++;
}
j++;
while(j!=size){
int dst=smpi_group_index(win->group,j);
- if(dst!=smpi_process_index()){
- reqs[i]=smpi_mpi_send_init(nullptr, 0, MPI_CHAR, dst, RMA_TAG+5, MPI_COMM_WORLD);
+ if(dst!=smpi_process_index() && dst!=MPI_UNDEFINED){
+ reqs[i]=smpi_mpi_send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+5, MPI_COMM_WORLD);
i++;
}
j++;
xbt_free(reqs);
//now we can finish RMA calls
-
+ xbt_mutex_acquire(win->mut);
std::vector<MPI_Request> *reqqs = win->requests;
size = static_cast<int>(reqqs->size());
XBT_DEBUG("Win_complete - Finishing %d RMA calls", size);
- // start all requests that have been prepared by another process
- for (auto req: *reqqs){
- if (req && (req->flags & PREPARED))
- smpi_mpi_start(req);
+ if (size > 0) {
+ // start all requests that have been prepared by another process
+ for (const auto& req : *reqqs) {
+ if (req && (req->flags & PREPARED))
+ smpi_mpi_start(req);
+ }
+
+ MPI_Request* treqs = &(*reqqs)[0];
+ smpi_mpi_waitall(size, treqs, MPI_STATUSES_IGNORE);
+ reqqs->clear();
}
+ xbt_mutex_release(win->mut);
- MPI_Request* treqs = &(*reqqs)[0];
- smpi_mpi_waitall(size,treqs,MPI_STATUSES_IGNORE);
- reqqs->clear();
smpi_group_unuse(win->group);
win->opened--; //we're closed for business !
return MPI_SUCCESS;
while(j!=size){
int src=smpi_group_index(win->group,j);
- if(src!=smpi_process_index()){
- reqs[i]=smpi_irecv_init(nullptr, 0, MPI_CHAR, src,RMA_TAG+5, MPI_COMM_WORLD);
+ if(src!=smpi_process_index() && src!=MPI_UNDEFINED){
+ reqs[i]=smpi_irecv_init(nullptr, 0, MPI_CHAR, src,SMPI_RMA_TAG+5, MPI_COMM_WORLD);
i++;
}
j++;
smpi_mpi_request_free(&reqs[i]);
}
xbt_free(reqs);
-
+ xbt_mutex_acquire(win->mut);
std::vector<MPI_Request> *reqqs = win->requests;
size = static_cast<int>(reqqs->size());
- XBT_DEBUG("Win_complete - Finishing %d RMA calls", size);
+ XBT_DEBUG("Win_wait - Finishing %d RMA calls", size);
+ if (size > 0) {
+ // start all requests that have been prepared by another process
+ for (const auto& req : *reqqs) {
+ if (req && (req->flags & PREPARED))
+ smpi_mpi_start(req);
+ }
- // start all requests that have been prepared by another process
- for(auto req: *reqqs){
- if (req && (req->flags & PREPARED))
- smpi_mpi_start(req);
+ MPI_Request* treqs = &(*reqqs)[0];
+ smpi_mpi_waitall(size, treqs, MPI_STATUSES_IGNORE);
+ reqqs->clear();
}
+ xbt_mutex_release(win->mut);
- MPI_Request* treqs = &(*reqqs)[0];
- smpi_mpi_waitall(size,treqs,MPI_STATUSES_IGNORE);
- reqqs->clear();
smpi_group_unuse(win->group);
win->opened--; //we're opened for business !
return MPI_SUCCESS;
