X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/20180cf998fe77ea59e5dbffd7ba0515124c775d..7013392b1a0a3aed2219fa352e81826f8657fbe2:/src/smpi/mpi/smpi_win.cpp diff --git a/src/smpi/mpi/smpi_win.cpp b/src/smpi/mpi/smpi_win.cpp index 5a2544d586..86059ca339 100644 --- a/src/smpi/mpi/smpi_win.cpp +++ b/src/smpi/mpi/smpi_win.cpp @@ -1,17 +1,18 @@ -/* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2019. 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 "smpi_win.hpp" + #include "private.hpp" #include "smpi_coll.hpp" #include "smpi_comm.hpp" #include "smpi_datatype.hpp" #include "smpi_info.hpp" #include "smpi_keyvals.hpp" -#include "smpi_process.hpp" #include "smpi_request.hpp" +#include "src/smpi/include/smpi_actor.hpp" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)"); @@ -32,30 +33,31 @@ Win::Win(void *base, MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm, opened_ = 0; group_ = MPI_GROUP_NULL; requests_ = new std::vector(); - mut_ = xbt_mutex_init(); - lock_mut_ = xbt_mutex_init(); - atomic_mut_ = xbt_mutex_init(); + mut_ = s4u::Mutex::create(); + lock_mut_ = s4u::Mutex::create(); + atomic_mut_ = s4u::Mutex::create(); connected_wins_ = new MPI_Win[comm_size]; connected_wins_[rank_] = this; count_ = 0; if(rank_==0){ - bar_ = MSG_barrier_init(comm_size); + bar_ = new simgrid::s4u::Barrier(comm_size); } mode_=0; comm->add_rma_win(this); + comm->ref(); 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); + Colls::bcast(&(bar_), sizeof(simgrid::s4u::Barrier*), MPI_BYTE, 0, comm); Colls::barrier(comm); } Win::~Win(){ //As per the standard, perform a barrier to ensure every async comm is finished - MSG_barrier_wait(bar_); + bar_->wait(); int finished = finish_comms(); XBT_DEBUG("Win destructor - Finished %d RMA calls", finished); @@ -72,11 +74,10 @@ Win::~Win(){ comm_->remove_rma_win(this); Colls::barrier(comm_); + Comm::unref(comm_); + if (rank_ == 0) - MSG_barrier_destroy(bar_); - xbt_mutex_destroy(mut_); - xbt_mutex_destroy(lock_mut_); - xbt_mutex_destroy(atomic_mut_); + delete bar_; if(allocated_ !=0) xbt_free(base_); @@ -160,8 +161,8 @@ int Win::fence(int assert) opened_=1; if (assert != MPI_MODE_NOPRECEDE) { // This is not the first fence => finalize what came before - MSG_barrier_wait(bar_); - xbt_mutex_acquire(mut_); + 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 redimensionned when another process pushes. // This would result in the array used by Request::waitall() to be invalidated. @@ -174,14 +175,14 @@ int Win::fence(int assert) Request::waitall(size, treqs, MPI_STATUSES_IGNORE); } count_=0; - xbt_mutex_release(mut_); + mut_->unlock(); } 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_); + bar_->wait(); XBT_DEBUG("Leaving fence"); return MPI_SUCCESS; @@ -207,9 +208,9 @@ int Win::put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, return MPI_ERR_ARG; void* recv_addr = static_cast ( static_cast(recv_win->base_) + target_disp * recv_win->disp_unit_); - XBT_DEBUG("Entering MPI_Put to %d", target_rank); if (target_rank != comm_->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 @@ -227,18 +228,19 @@ int Win::put( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, if(request!=nullptr){ *request=sreq; }else{ - xbt_mutex_acquire(mut_); + mut_->lock(); requests_->push_back(sreq); - xbt_mutex_release(mut_); + mut_->unlock(); } //push request to receiver's win - xbt_mutex_acquire(recv_win->mut_); + recv_win->mut_->lock(); recv_win->requests_->push_back(rreq); rreq->start(); - xbt_mutex_release(recv_win->mut_); + recv_win->mut_->unlock(); }else{ + XBT_DEBUG("Entering MPI_Put from myself to myself, rank %d", target_rank); Datatype::copy(origin_addr, origin_count, origin_datatype, recv_addr, target_count, target_datatype); if(request!=nullptr) *request = MPI_REQUEST_NULL; @@ -283,9 +285,9 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, //start the send, with another process than us as sender. sreq->start(); //push request to receiver's win - xbt_mutex_acquire(send_win->mut_); + send_win->mut_->lock(); send_win->requests_->push_back(sreq); - xbt_mutex_release(send_win->mut_); + send_win->mut_->unlock(); //start recv rreq->start(); @@ -293,9 +295,9 @@ int Win::get( void *origin_addr, int origin_count, MPI_Datatype origin_datatype, if(request!=nullptr){ *request=rreq; }else{ - xbt_mutex_acquire(mut_); + mut_->lock(); requests_->push_back(rreq); - xbt_mutex_release(mut_); + mut_->unlock(); } }else{ @@ -346,17 +348,17 @@ int Win::accumulate( void *origin_addr, int origin_count, MPI_Datatype origin_da // start send sreq->start(); // push request to receiver's win - xbt_mutex_acquire(recv_win->mut_); + recv_win->mut_->lock(); recv_win->requests_->push_back(rreq); rreq->start(); - xbt_mutex_release(recv_win->mut_); + recv_win->mut_->unlock(); if (request != nullptr) { *request = sreq; } else { - xbt_mutex_acquire(mut_); + mut_->lock(); requests_->push_back(sreq); - xbt_mutex_release(mut_); + mut_->unlock(); } XBT_DEBUG("Leaving MPI_Win_Accumulate"); @@ -386,7 +388,7 @@ int Win::get_accumulate( void *origin_addr, int origin_count, MPI_Datatype origi 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_); + send_win->atomic_mut_->lock(); get(result_addr, result_count, result_datatype, target_rank, target_disp, target_count, target_datatype, &req); if (req != MPI_REQUEST_NULL) @@ -396,7 +398,7 @@ int Win::get_accumulate( void *origin_addr, int origin_count, MPI_Datatype origi 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_); + send_win->atomic_mut_->unlock(); return MPI_SUCCESS; } @@ -419,7 +421,7 @@ int Win::compare_and_swap(void *origin_addr, void *compare_addr, XBT_DEBUG("Entering MPI_Compare_and_swap with %d", target_rank); MPI_Request req = MPI_REQUEST_NULL; - xbt_mutex_acquire(send_win->atomic_mut_); + send_win->atomic_mut_->lock(); get(result_addr, 1, datatype, target_rank, target_disp, 1, datatype, &req); if (req != MPI_REQUEST_NULL) @@ -428,7 +430,7 @@ int Win::compare_and_swap(void *origin_addr, void *compare_addr, put(origin_addr, 1, datatype, target_rank, target_disp, 1, datatype); } - xbt_mutex_release(send_win->atomic_mut_); + send_win->atomic_mut_->unlock(); return MPI_SUCCESS; } @@ -577,10 +579,10 @@ int Win::lock(int lock_type, int rank, int assert){ 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->lock_mut_->lock(); 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_); + target_win->lock_mut_->unlock(); } } 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 @@ -611,7 +613,7 @@ int Win::unlock(int rank){ target_win->mode_= 0; target_win->lockers_.remove(comm_->rank()); if (target_mode==MPI_LOCK_EXCLUSIVE){ - xbt_mutex_release(target_win->lock_mut_); + target_win->lock_mut_->unlock(); } int finished = finish_comms(rank); @@ -625,18 +627,18 @@ 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; + int ret = this->unlock(i); + if (ret != MPI_SUCCESS) + retval = ret; } return retval; } int Win::flush(int rank){ MPI_Win target_win = connected_wins_[rank]; - int finished = finish_comms(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_); + finished = target_win->finish_comms(rank); XBT_DEBUG("Win_flush on remote %d - Finished %d RMA calls", rank, finished); return MPI_SUCCESS; } @@ -648,11 +650,9 @@ int Win::flush_local(int rank){ } int Win::flush_all(){ - int i=0; - int finished = 0; - finished = finish_comms(); + int finished = finish_comms(); XBT_DEBUG("Win_flush_all on local - Finished %d RMA calls", finished); - for (i=0; isize();i++){ + 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); } @@ -669,9 +669,8 @@ Win* Win::f2c(int id){ return static_cast(F2C::f2c(id)); } - int Win::finish_comms(){ - xbt_mutex_acquire(mut_); + mut_->lock(); //Finish own requests std::vector *reqqs = requests_; int size = static_cast(reqqs->size()); @@ -680,12 +679,12 @@ int Win::finish_comms(){ Request::waitall(size, treqs, MPI_STATUSES_IGNORE); reqqs->clear(); } - xbt_mutex_release(mut_); + mut_->unlock(); return size; } int Win::finish_comms(int rank){ - xbt_mutex_acquire(mut_); + mut_->lock(); //Finish own requests std::vector *reqqs = requests_; int size = static_cast(reqqs->size()); @@ -693,8 +692,13 @@ int Win::finish_comms(int rank){ size = 0; std::vector myreqqs; std::vector::iterator iter = reqqs->begin(); + int proc_id = comm_->group()->actor(rank)->get_pid(); while (iter != reqqs->end()){ - if(((*iter)!=MPI_REQUEST_NULL) && (((*iter)->src() == rank) || ((*iter)->dst() == rank))){ + // Let's see if we're either the destination or the sender of this request + // because we only wait for requests that we are responsible for. + // Also use the process id here since the request itself returns from src() + // and dst() the process id, NOT the rank (which only exists in the context of a communicator). + if (((*iter) != MPI_REQUEST_NULL) && (((*iter)->src() == proc_id) || ((*iter)->dst() == proc_id))) { myreqqs.push_back(*iter); iter = reqqs->erase(iter); size++; @@ -708,10 +712,26 @@ int Win::finish_comms(int rank){ myreqqs.clear(); } } - xbt_mutex_release(mut_); + mut_->unlock(); return size; } - +int Win::shared_query(int rank, MPI_Aint* size, int* disp_unit, void* baseptr) +{ + MPI_Win target_win = rank != MPI_PROC_NULL ? connected_wins_[rank] : nullptr; + for (int i = 0; not target_win && i < comm_->size(); i++) { + if (connected_wins_[i]->size_ > 0) + target_win = connected_wins_[i]; + } + if (target_win) { + *size = target_win->size_; + *disp_unit = target_win->disp_unit_; + *static_cast(baseptr) = target_win->base_; + } else { + *size = 0; + *static_cast(baseptr) = xbt_malloc(0); + } + return MPI_SUCCESS; +} } }