-/* Copyright (c) 2007-2020. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2021. 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_request.hpp"
#include "src/smpi/include/smpi_actor.hpp"
+#include <algorithm>
+
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_rma, smpi, "Logging specific to SMPI (RMA operations)");
+#define CHECK_RMA_REMOTE_WIN(fun, win)\
+ if(target_count*target_datatype->get_extent()>win->size_){\
+ XBT_WARN("%s: Trying to move %zd, which exceeds the window size on target process %d : %zd - Bailing out.",\
+ fun, target_count*target_datatype->get_extent(), target_rank, win->size_);\
+ simgrid::smpi::utils::set_current_buffer(1,"win_base",win->base_);\
+ return MPI_ERR_RMA_RANGE;\
+ }
+
+#define CHECK_WIN_LOCKED(win)\
+ if(opened_==0){ /*check that post/start has been done*/\
+ int locked=0;\
+ for (auto const& it : win->lockers_)\
+ if (it == comm_->rank())\
+ locked = 1;\
+ if(locked != 1)\
+ return MPI_ERR_WIN;\
+ }
namespace simgrid{
namespace smpi{
, disp_unit_(disp_unit)
, info_(info)
, comm_(comm)
+ , connected_wins_(comm->size())
, rank_(comm->rank())
, allocated_(allocated)
, dynamic_(dynamic)
XBT_DEBUG("Creating window");
if(info!=MPI_INFO_NULL)
info->ref();
- int comm_size = comm->size();
- name_ = nullptr;
- opened_ = 0;
- group_ = MPI_GROUP_NULL;
- requests_ = new std::vector<MPI_Request>();
- 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_ = new s4u::Barrier(comm_size);
+ bar_ = new s4u::Barrier(comm->size());
}
- mode_=0;
- errhandler_=MPI_ERRORS_ARE_FATAL;
errhandler_->ref();
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::allgather(&connected_wins_[rank_], sizeof(MPI_Win), MPI_BYTE, connected_wins_.data(), sizeof(MPI_Win),
+ MPI_BYTE, comm);
- colls::bcast(&(bar_), sizeof(s4u::Barrier*), MPI_BYTE, 0, comm);
+ colls::bcast(&bar_, sizeof(s4u::Barrier*), MPI_BYTE, 0, comm);
colls::barrier(comm);
+ this->add_f();
}
Win::~Win(){
int finished = finish_comms();
XBT_DEBUG("Win destructor - Finished %d RMA calls", finished);
- delete requests_;
- delete[] connected_wins_;
- if (name_ != nullptr){
- xbt_free(name_);
- }
if (info_ != MPI_INFO_NULL)
simgrid::smpi::Info::unref(info_);
if (errhandler_ != MPI_ERRHANDLER_NULL)
if(allocated_ !=0)
xbt_free(base_);
+ F2C::free_f(this->f2c_id());
cleanup_attr<Win>();
}
void Win::get_name(char* name, int* length) const
{
- if(name_==nullptr){
- *length=0;
- name=nullptr;
- return;
+ *length = static_cast<int>(name_.length());
+ if (not name_.empty()) {
+ name_.copy(name, *length);
+ name[*length] = '\0';
}
- *length = strlen(name_);
- strncpy(name, name_, *length+1);
}
void Win::get_group(MPI_Group* group){
return rank_;
}
+MPI_Comm Win::comm() const
+{
+ return comm_;
+}
+
MPI_Aint Win::size() const
{
return size_;
}
void Win::set_name(const char* name){
- name_ = xbt_strdup(name);
+ name_ = name;
}
int Win::fence(int assert)
// Without this, the vector could get redimensioned when another process pushes.
// This would result in the array used by Request::waitall() to be invalidated.
// Another solution would be to copy the data and cleanup the vector *before* Request::waitall
- std::vector<MPI_Request> *reqs = requests_;
- int size = static_cast<int>(reqs->size());
+
// start all requests that have been prepared by another process
- if (size > 0) {
- MPI_Request* treqs = &(*reqs)[0];
+ if (not requests_.empty()) {
+ int size = static_cast<int>(requests_.size());
+ MPI_Request* treqs = requests_.data();
Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
}
count_=0;
MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Request* request)
{
//get receiver pointer
- const Win* recv_win = connected_wins_[target_rank];
+ Win* recv_win = connected_wins_[target_rank];
if(opened_==0){//check that post/start has been done
// no fence or start .. lock ok ?
return MPI_ERR_WIN;
}
- if(target_count*target_datatype->get_extent()>recv_win->size_)
- return MPI_ERR_ARG;
+ CHECK_RMA_REMOTE_WIN("MPI_Put", recv_win)
void* recv_addr = static_cast<char*>(recv_win->base_) + target_disp * recv_win->disp_unit_;
*request=sreq;
}else{
mut_->lock();
- requests_->push_back(sreq);
+ requests_.push_back(sreq);
mut_->unlock();
}
//push request to receiver's win
recv_win->mut_->lock();
- recv_win->requests_->push_back(rreq);
+ recv_win->requests_.push_back(rreq);
rreq->start();
recv_win->mut_->unlock();
} else {
MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Request* request)
{
//get sender pointer
- const 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 const& it : send_win->lockers_)
- if (it == comm_->rank())
- locked = 1;
- if(locked != 1)
- return MPI_ERR_WIN;
- }
+ Win* send_win = connected_wins_[target_rank];
- if(target_count*target_datatype->get_extent()>send_win->size_)
- return MPI_ERR_ARG;
+ CHECK_WIN_LOCKED(send_win)
+ CHECK_RMA_REMOTE_WIN("MPI_Get", send_win)
const void* send_addr = static_cast<void*>(static_cast<char*>(send_win->base_) + target_disp * send_win->disp_unit_);
XBT_DEBUG("Entering MPI_Get from %d", target_rank);
sreq->start();
//push request to receiver's win
send_win->mut_->lock();
- send_win->requests_->push_back(sreq);
+ send_win->requests_.push_back(sreq);
send_win->mut_->unlock();
//start recv
*request=rreq;
}else{
mut_->lock();
- requests_->push_back(rreq);
+ requests_.push_back(rreq);
mut_->unlock();
}
} else {
{
XBT_DEBUG("Entering MPI_Win_Accumulate");
//get receiver pointer
- const Win* recv_win = connected_wins_[target_rank];
+ 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 const& 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;
+ CHECK_WIN_LOCKED(recv_win)
+ CHECK_RMA_REMOTE_WIN("MPI_Accumulate", recv_win)
void* recv_addr = static_cast<char*>(recv_win->base_) + target_disp * recv_win->disp_unit_;
XBT_DEBUG("Entering MPI_Accumulate to %d", target_rank);
// prepare receiver request
MPI_Request rreq = Request::rma_recv_init(recv_addr, target_count, target_datatype, recv_win->comm_->rank(),
- recv_win->comm_->group()->rank(comm_->group()->actor(target_rank)), SMPI_RMA_TAG - 3 - count_, recv_win->comm_, op);
+ recv_win->comm_->group()->rank(comm_->group()->actor_pid(target_rank)),
+ SMPI_RMA_TAG - 3 - count_, recv_win->comm_, op);
count_++;
sreq->start();
// push request to receiver's win
recv_win->mut_->lock();
- recv_win->requests_->push_back(rreq);
+ recv_win->requests_.push_back(rreq);
rreq->start();
recv_win->mut_->unlock();
*request = sreq;
} else {
mut_->lock();
- requests_->push_back(sreq);
+ requests_.push_back(sreq);
mut_->unlock();
}
//get sender pointer
const 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 const& 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;
+ CHECK_WIN_LOCKED(send_win)
+ CHECK_RMA_REMOTE_WIN("MPI_Get_Accumulate", send_win)
XBT_DEBUG("Entering MPI_Get_accumulate from %d", target_rank);
//need to be sure ops are correctly ordered, so finish request here ? slow.
//get sender pointer
const 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 const& it : send_win->lockers_)
- if (it == comm_->rank())
- locked = 1;
- if(locked != 1)
- return MPI_ERR_WIN;
- }
+ CHECK_WIN_LOCKED(send_win)
XBT_DEBUG("Entering MPI_Compare_and_swap with %d", target_rank);
MPI_Request req = MPI_REQUEST_NULL;
int i = 0;
int j = 0;
int size = group->size();
- MPI_Request* reqs = xbt_new0(MPI_Request, size);
+ std::vector<MPI_Request> reqs(size);
XBT_DEBUG("Entering MPI_Win_Start");
while (j != size) {
- int src = comm_->group()->rank(group->actor(j));
+ int src = comm_->group()->rank(group->actor_pid(j));
if (src != rank_ && src != MPI_UNDEFINED) { // TODO cheinrich: The check of MPI_UNDEFINED should be useless here
reqs[i] = Request::irecv_init(nullptr, 0, MPI_CHAR, src, SMPI_RMA_TAG + 4, comm_);
i++;
j++;
}
size = i;
- Request::startall(size, reqs);
- Request::waitall(size, reqs, MPI_STATUSES_IGNORE);
+ Request::startall(size, reqs.data());
+ Request::waitall(size, reqs.data(), MPI_STATUSES_IGNORE);
for (i = 0; i < size; i++) {
Request::unref(&reqs[i]);
}
- xbt_free(reqs);
opened_++; //we're open for business !
group_=group;
group->ref();
int i = 0;
int j = 0;
int size = group->size();
- MPI_Request* reqs = xbt_new0(MPI_Request, size);
+ std::vector<MPI_Request> reqs(size);
XBT_DEBUG("Entering MPI_Win_Post");
while(j!=size){
- int dst = comm_->group()->rank(group->actor(j));
+ int dst = comm_->group()->rank(group->actor_pid(j));
if (dst != rank_ && dst != MPI_UNDEFINED) {
reqs[i] = Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG + 4, comm_);
i++;
}
size=i;
- Request::startall(size, reqs);
- Request::waitall(size, reqs, MPI_STATUSES_IGNORE);
+ Request::startall(size, reqs.data());
+ Request::waitall(size, reqs.data(), MPI_STATUSES_IGNORE);
for(i=0;i<size;i++){
Request::unref(&reqs[i]);
}
- xbt_free(reqs);
opened_++; //we're open for business !
group_=group;
group->ref();
}
int Win::complete(){
- if(opened_==0)
- xbt_die("Complete called on already opened MPI_Win");
+ xbt_assert(opened_ != 0, "Complete called on already opened MPI_Win");
XBT_DEBUG("Entering MPI_Win_Complete");
int i = 0;
int j = 0;
int size = group_->size();
- MPI_Request* reqs = xbt_new0(MPI_Request, size);
+ std::vector<MPI_Request> reqs(size);
while(j!=size){
- int dst = comm_->group()->rank(group_->actor(j));
+ int dst = comm_->group()->rank(group_->actor_pid(j));
if (dst != rank_ && dst != MPI_UNDEFINED) {
reqs[i] = Request::send_init(nullptr, 0, MPI_CHAR, dst, SMPI_RMA_TAG + 5, comm_);
i++;
}
size=i;
XBT_DEBUG("Win_complete - Sending sync messages to %d processes", size);
- Request::startall(size, reqs);
- Request::waitall(size, reqs, MPI_STATUSES_IGNORE);
+ Request::startall(size, reqs.data());
+ Request::waitall(size, reqs.data(), MPI_STATUSES_IGNORE);
for(i=0;i<size;i++){
Request::unref(&reqs[i]);
}
- xbt_free(reqs);
int finished = finish_comms();
XBT_DEBUG("Win_complete - Finished %d RMA calls", finished);
int i = 0;
int j = 0;
int size = group_->size();
- MPI_Request* reqs = xbt_new0(MPI_Request, size);
+ std::vector<MPI_Request> reqs(size);
while(j!=size){
- int src = comm_->group()->rank(group_->actor(j));
+ int src = comm_->group()->rank(group_->actor_pid(j));
if (src != rank_ && src != MPI_UNDEFINED) {
reqs[i] = Request::irecv_init(nullptr, 0, MPI_CHAR, src, SMPI_RMA_TAG + 5, comm_);
i++;
}
size=i;
XBT_DEBUG("Win_wait - Receiving sync messages from %d processes", size);
- Request::startall(size, reqs);
- Request::waitall(size, reqs, MPI_STATUSES_IGNORE);
+ Request::startall(size, reqs.data());
+ Request::waitall(size, reqs.data(), MPI_STATUSES_IGNORE);
for(i=0;i<size;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 !
+ opened_--; //we're closed for business !
return MPI_SUCCESS;
}
int Win::finish_comms(){
mut_->lock();
//Finish own requests
- std::vector<MPI_Request> *reqqs = requests_;
- int size = static_cast<int>(reqqs->size());
+ int size = static_cast<int>(requests_.size());
if (size > 0) {
- MPI_Request* treqs = &(*reqqs)[0];
+ MPI_Request* treqs = requests_.data();
Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
- reqqs->clear();
+ requests_.clear();
}
mut_->unlock();
return size;
int Win::finish_comms(int rank){
mut_->lock();
- //Finish own requests
- std::vector<MPI_Request> *reqqs = requests_;
- int size = static_cast<int>(reqqs->size());
+ // Finish own requests
+ // 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).
+ int proc_id = comm_->group()->actor_pid(rank);
+ auto it = std::stable_partition(begin(requests_), end(requests_), [proc_id](const MPI_Request& req) {
+ return (req == MPI_REQUEST_NULL || (req->src() != proc_id && req->dst() != proc_id));
+ });
+ std::vector<MPI_Request> myreqqs(it, end(requests_));
+ requests_.erase(it, end(requests_));
+ int size = static_cast<int>(myreqqs.size());
if (size > 0) {
- size = 0;
- std::vector<MPI_Request> myreqqs;
- auto iter = reqqs->begin();
- int proc_id = comm_->group()->actor(rank)->get_pid();
- while (iter != reqqs->end()){
- // 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++;
- } else {
- ++iter;
- }
- }
- if(size >0){
- MPI_Request* treqs = &myreqqs[0];
- Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
- myreqqs.clear();
- }
+ MPI_Request* treqs = myreqqs.data();
+ Request::waitall(size, treqs, MPI_STATUSES_IGNORE);
+ myreqqs.clear();
}
mut_->unlock();
return size;