X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9c9f17b4ac25fe6fcf026724a7d3cc4978b775ed..19b3962253112b19308537bc2400de141c119d99:/src/smpi/smpi_win.cpp diff --git a/src/smpi/smpi_win.cpp b/src/smpi/smpi_win.cpp index 5f819980b1..e2d6b5f3a8 100644 --- a/src/smpi/smpi_win.cpp +++ b/src/smpi/smpi_win.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2015. The SimGrid Team. +/* Copyright (c) 2007-2017. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it @@ -14,7 +14,7 @@ namespace smpi{ std::unordered_map 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){ +Win::Win(void *base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm, int allocated, int dynamic): base_(base), size_(size), disp_unit_(disp_unit), assert_(0), info_(info), comm_(comm), allocated_(allocated), dynamic_(dynamic){ int comm_size = comm->size(); rank_ = comm->rank(); XBT_DEBUG("Creating window"); @@ -26,12 +26,14 @@ Win::Win(void *base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm) requests_ = new std::vector(); 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; if(rank_==0){ bar_ = MSG_barrier_init(comm_size); } + mode_=0; comm->add_rma_win(this); @@ -67,10 +69,28 @@ Win::~Win(){ MSG_barrier_destroy(bar_); xbt_mutex_destroy(mut_); xbt_mutex_destroy(lock_mut_); + xbt_mutex_destroy(atomic_mut_); + + if(allocated_ !=0) + xbt_free(base_); cleanup_attr(); } +int Win::attach (void *base, MPI_Aint size){ + if (!(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; + return MPI_SUCCESS; +} + +int Win::detach (void *base){ + base_=MPI_BOTTOM; + size_=-1; + return MPI_SUCCESS; +} + void Win::get_name(char* name, int* length){ if(name_==nullptr){ *length=0; @@ -89,6 +109,13 @@ void Win::get_group(MPI_Group* group){ } } +MPI_Info Win::info(){ + if(info_== MPI_INFO_NULL) + info_ = new Info(); + info_->ref(); + return info_; +} + int Win::rank(){ return rank_; } @@ -105,6 +132,15 @@ int Win::disp_unit(){ return disp_unit_; } +int Win::dynamic(){ + return dynamic_; +} + +void Win::set_info(MPI_Info info){ + if(info_!= MPI_INFO_NULL) + info->ref(); + info_=info; +} void Win::set_name(char* name){ name_ = xbt_strdup(name); @@ -127,18 +163,15 @@ int Win::fence(int assert) int size = static_cast(reqs->size()); // start all requests that have been prepared by another process if (size > 0) { - for (const auto& req : *reqs) { - if (req && (req->flags() & PREPARED)) - req->start(); - } - MPI_Request* treqs = &(*reqs)[0]; - Request::waitall(size, treqs, MPI_STATUSES_IGNORE); } count_=0; xbt_mutex_release(mut_); } + + if(assert==MPI_MODE_NOSUCCEED)//there should be no ops after this one, tell we are closed. + opened_=0; assert_ = assert; MSG_barrier_wait(bar_); @@ -148,7 +181,7 @@ int Win::fence(int assert) } 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]; @@ -178,26 +211,34 @@ int Win::put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, 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); + //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_); - //start send - sreq->start(); - //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]; @@ -238,12 +279,19 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, //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; @@ -251,7 +299,7 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, 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 @@ -284,18 +332,92 @@ int Win::accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_da smpi_process()->index(), comm_->group()->index(target_rank), SMPI_RMA_TAG-3-count_, recv_win->comm_, op); count_++; + + //start send + sreq->start(); //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_); - //start send - sreq->start(); - //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, MPI_Request* request){ + + //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); + //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, &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, &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(! 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; } @@ -407,8 +529,9 @@ int Win::complete(){ 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){ @@ -438,39 +561,92 @@ int Win::wait(){ int Win::lock(int lock_type, int rank, int assert){ MPI_Win target_win = connected_wins_[rank]; - //window already locked, we have to wait - if (lock_type == MPI_LOCK_EXCLUSIVE){ - XBT_DEBUG("Win_lock - Entering lock %d", 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_); - XBT_DEBUG("Win_lock - Released from lock %d", rank); -} + 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 - xbt_mutex_acquire(target_win->mut_); target_win->lockers_.push_back(comm_->rank()); - xbt_mutex_release(target_win->mut_); - - int finished = finish_comms(); - XBT_DEBUG("Win_lock - Finished %d RMA calls", finished); + 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::lock_all(int assert){ + int i=0; + int retval = MPI_SUCCESS; + for (i=0; isize();i++){ + int ret = this->lock(MPI_LOCK_SHARED, i, assert); + if(ret != MPI_SUCCESS) + retval = ret; + } + return retval; +} + int Win::unlock(int rank){ MPI_Win target_win = connected_wins_[rank]; - - xbt_mutex_acquire(target_win->mut_); - int size=target_win->lockers_.size(); + 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(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; +} - if (size<=1){//0 or 1 lockers -> exclusive assumed - xbt_mutex_try_acquire(target_win->lock_mut_); - xbt_mutex_release(target_win->lock_mut_); +int Win::unlock_all(){ + int i=0; + int retval = MPI_SUCCESS; + for (i=0; isize();i++){ + int ret = this->unlock(i); + if(ret != MPI_SUCCESS) + retval = ret; } - xbt_mutex_release(target_win->mut_); - int finished = finish_comms(); - XBT_DEBUG("Win_unlock - Finished %d RMA calls", finished); + return retval; +} + +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); + 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); + return MPI_SUCCESS; +} + +int Win::flush_all(){ + int i=0; + int finished = 0; + finished = finish_comms(); + XBT_DEBUG("Win_flush_all on local - Finished %d RMA calls", finished); + for (i=0; isize();i++){ + finished = connected_wins_[i]->finish_comms(rank_); + XBT_DEBUG("Win_flush_all on %d - Finished %d RMA calls", i, finished); + } + return MPI_SUCCESS; +} +int Win::flush_local_all(){ + int finished = finish_comms(); + XBT_DEBUG("Win_flush_local_all - Finished %d RMA calls", finished); return MPI_SUCCESS; } @@ -485,12 +661,6 @@ int Win::finish_comms(){ std::vector *reqqs = requests_; int size = static_cast(reqqs->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(); @@ -499,6 +669,35 @@ int Win::finish_comms(){ return size; } +int Win::finish_comms(int rank){ + xbt_mutex_acquire(mut_); + //Finish own requests + std::vector *reqqs = requests_; + int size = static_cast(reqqs->size()); + if (size > 0) { + size = 0; + std::vector* myreqqs = new std::vector(); + std::vector::iterator iter = reqqs->begin(); + while (iter != reqqs->end()){ + 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]; + Request::waitall(size, treqs, MPI_STATUSES_IGNORE); + myreqqs->clear(); + delete myreqqs; + } + } + xbt_mutex_release(mut_); + return size; +} + } }