namespace simgrid{
namespace smpi{
+std::unordered_map<int, smpi_key_elem> Win::keyvals_;
+int Win::keyval_id_=0;
Win::Win(void *base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm): base_(base), size_(size), disp_unit_(disp_unit), assert_(0), info_(info), comm_(comm){
int comm_size = comm->size();
- int rank = comm->rank();
+ rank_ = comm->rank();
XBT_DEBUG("Creating window");
if(info!=MPI_INFO_NULL)
info->ref();
group_ = MPI_GROUP_NULL;
requests_ = new std::vector<MPI_Request>();
mut_=xbt_mutex_init();
+ lock_mut_=xbt_mutex_init();
connected_wins_ = new MPI_Win[comm_size];
- connected_wins_[rank] = this;
+ connected_wins_[rank_] = this;
count_ = 0;
- if(rank==0){
+ if(rank_==0){
bar_ = MSG_barrier_init(comm_size);
}
- Colls::allgather(&(connected_wins_[rank]), sizeof(MPI_Win), MPI_BYTE, connected_wins_, sizeof(MPI_Win),
+ mode_=0;
+
+ comm->add_rma_win(this);
+
+ Colls::allgather(&(connected_wins_[rank_]), sizeof(MPI_Win), MPI_BYTE, connected_wins_, sizeof(MPI_Win),
MPI_BYTE, comm);
Colls::bcast(&(bar_), sizeof(msg_bar_t), MPI_BYTE, 0, comm);
Win::~Win(){
//As per the standard, perform a barrier to ensure every async comm is finished
MSG_barrier_wait(bar_);
- xbt_mutex_acquire(mut_);
+
+ int finished = finish_comms();
+ XBT_DEBUG("Win destructor - Finished %d RMA calls", finished);
+
delete requests_;
- xbt_mutex_release(mut_);
delete[] connected_wins_;
if (name_ != nullptr){
xbt_free(name_);
MPI_Info_free(&info_);
}
+ comm_->remove_rma_win(this);
+
Colls::barrier(comm_);
int rank=comm_->rank();
if(rank == 0)
MSG_barrier_destroy(bar_);
xbt_mutex_destroy(mut_);
+ xbt_mutex_destroy(lock_mut_);
+
+ cleanup_attr<Win>();
}
void Win::get_name(char* name, int* length){
}
}
+int Win::rank(){
+ return rank_;
+}
+
+MPI_Aint Win::size(){
+ return size_;
+}
+
+void* Win::base(){
+ return base_;
+}
+
+int Win::disp_unit(){
+ return disp_unit_;
+}
+
+
void Win::set_name(char* name){
name_ = xbt_strdup(name);
}
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)
{
- if(opened_==0)//check that post/start has been done
- return MPI_ERR_WIN;
//get receiver pointer
MPI_Win recv_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 : recv_win->lockers_)
+ if (it == comm_->rank())
+ locked = 1;
+ if(locked != 1)
+ return MPI_ERR_WIN;
+ }
+
if(target_count*target_datatype->get_extent()>recv_win->size_)
return MPI_ERR_ARG;
if(target_rank != comm_->rank()){
//prepare send_request
- MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, smpi_process_index(),
+ MPI_Request sreq = Request::rma_send_init(origin_addr, origin_count, origin_datatype, smpi_process()->index(),
comm_->group()->index(target_rank), SMPI_RMA_TAG+1, comm_, MPI_OP_NULL);
//prepare receiver request
- MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, smpi_process_index(),
+ MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, smpi_process()->index(),
comm_->group()->index(target_rank), SMPI_RMA_TAG+1, recv_win->comm_, MPI_OP_NULL);
//push request to receiver's win
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)
{
- if(opened_==0)//check that post/start has been done
- return MPI_ERR_WIN;
//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;
+ }
+
if(target_count*target_datatype->get_extent()>send_win->size_)
return MPI_ERR_ARG;
if(target_rank != comm_->rank()){
//prepare send_request
MPI_Request sreq = Request::rma_send_init(send_addr, target_count, target_datatype,
- comm_->group()->index(target_rank), smpi_process_index(), SMPI_RMA_TAG+2, send_win->comm_,
+ comm_->group()->index(target_rank), smpi_process()->index(), 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,
- comm_->group()->index(target_rank), smpi_process_index(), SMPI_RMA_TAG+2, comm_,
+ 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.
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)
{
- if(opened_==0)//check that post/start has been done
- return MPI_ERR_WIN;
- //FIXME: local version
+
//get receiver pointer
MPI_Win recv_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 : recv_win->lockers_)
+ if (it == comm_->rank())
+ locked = 1;
+ if(locked != 1)
+ return MPI_ERR_WIN;
+ }
+ //FIXME: local version
+
if(target_count*target_datatype->get_extent()>recv_win->size_)
return MPI_ERR_ARG;
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
+ //As the tag will be used for ordering of the operations, substract count from it (to avoid collisions with other 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,
- smpi_process_index(), comm_->group()->index(target_rank), SMPI_RMA_TAG+3+count_, comm_, op);
+ smpi_process()->index(), comm_->group()->index(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,
- smpi_process_index(), comm_->group()->index(target_rank), SMPI_RMA_TAG+3+count_, recv_win->comm_, op);
+ smpi_process()->index(), comm_->group()->index(target_rank), SMPI_RMA_TAG-3-count_, recv_win->comm_, op);
count_++;
//push request to receiver's win
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){
+
+ //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;
+ }
+
+ if(target_count*target_datatype->get_extent()>send_win->size_)
+ return MPI_ERR_ARG;
+
+ XBT_DEBUG("Entering MPI_Get_accumulate from %d", target_rank);
+
+ get(result_addr, result_count, result_datatype, target_rank,
+ target_disp, target_count, target_datatype);
+ accumulate(origin_addr, origin_count, origin_datatype, target_rank,
+ target_disp, target_count, target_datatype, op);
+
+ 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
while (j != size) {
int src = group->index(j);
- if (src != smpi_process_index() && src != MPI_UNDEFINED) {
+ if (src != smpi_process()->index() && src != MPI_UNDEFINED) {
reqs[i] = Request::irecv_init(nullptr, 0, MPI_CHAR, src, SMPI_RMA_TAG + 4, MPI_COMM_WORLD);
i++;
}
while(j!=size){
int dst=group->index(j);
- if(dst!=smpi_process_index() && dst!=MPI_UNDEFINED){
+ if(dst!=smpi_process()->index() && dst!=MPI_UNDEFINED){
reqs[i]=Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+4, MPI_COMM_WORLD);
i++;
}
while(j!=size){
int dst=group_->index(j);
- if(dst!=smpi_process_index() && dst!=MPI_UNDEFINED){
+ if(dst!=smpi_process()->index() && dst!=MPI_UNDEFINED){
reqs[i]=Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG+5, MPI_COMM_WORLD);
i++;
}
}
xbt_free(reqs);
- //now we can finish RMA calls
- xbt_mutex_acquire(mut_);
- std::vector<MPI_Request> *reqqs = requests_;
- size = static_cast<int>(reqqs->size());
-
- XBT_DEBUG("Win_complete - 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))
- req->start();
- }
-
- MPI_Request* treqs = &(*reqqs)[0];
- Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
- reqqs->clear();
- }
- xbt_mutex_release(mut_);
+ int finished = finish_comms();
+ XBT_DEBUG("Win_complete - Finished %d RMA calls", finished);
Group::unref(group_);
opened_--; //we're closed for business !
int Win::wait(){
//naive, blocking implementation.
XBT_DEBUG("Entering MPI_Win_Wait");
- int i=0,j=0;
- int size = group_->size();
+ int i = 0;
+ int j = 0;
+ int size = group_->size();
MPI_Request* reqs = xbt_new0(MPI_Request, size);
while(j!=size){
int src=group_->index(j);
- if(src!=smpi_process_index() && src!=MPI_UNDEFINED){
+ if(src!=smpi_process()->index() && src!=MPI_UNDEFINED){
reqs[i]=Request::irecv_init(nullptr, 0, MPI_CHAR, src,SMPI_RMA_TAG+5, MPI_COMM_WORLD);
i++;
}
Request::unref(&reqs[i]);
}
xbt_free(reqs);
+ int finished = finish_comms();
+ XBT_DEBUG("Win_wait - Finished %d RMA calls", finished);
+
+ Group::unref(group_);
+ opened_--; //we're opened for business !
+ return MPI_SUCCESS;
+}
+
+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){
+ xbt_mutex_acquire(target_win->lock_mut_);
+ target_win->mode_+= lock_type;//add the lock_type to differentiate case when we are switching from EXCLUSIVE to SHARED (no release needed in the unlock)
+ 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
+
+ target_win->lockers_.push_back(comm_->rank());
+
+ int finished = finish_comms();
+ XBT_DEBUG("Win_lock %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;
+ target_win->lockers_.remove(comm_->rank());
+ if (target_mode==MPI_LOCK_EXCLUSIVE){
+ xbt_mutex_release(target_win->lock_mut_);
+ }
+
+ int finished = finish_comms();
+ XBT_DEBUG("Win_unlock %d - Finished %d RMA calls", rank, finished);
+
+ return MPI_SUCCESS;
+}
+
+Win* Win::f2c(int id){
+ return static_cast<Win*>(F2C::f2c(id));
+}
+
+
+int Win::finish_comms(){
xbt_mutex_acquire(mut_);
+ //Finish own requests
std::vector<MPI_Request> *reqqs = requests_;
- size = static_cast<int>(reqqs->size());
-
- XBT_DEBUG("Win_wait - Finishing %d RMA calls", size);
+ int size = static_cast<int>(reqqs->size());
if (size > 0) {
// start all requests that have been prepared by another process
for (const auto& req : *reqqs) {
reqqs->clear();
}
xbt_mutex_release(mut_);
-
- Group::unref(group_);
- opened_--; //we're opened for business !
- return MPI_SUCCESS;
+ return size;
}
-Win* Win::f2c(int id){
- return static_cast<Win*>(F2C::f2c(id));
-}
}
}
