X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/711f572ba93269d5151544df3a6df09018c803a0..6b01c1856641f3a79c1eec8f4f0f02e941302090:/src/smpi/mpi/smpi_comm.cpp

diff --git a/src/smpi/mpi/smpi_comm.cpp b/src/smpi/mpi/smpi_comm.cpp
index 0d7b5c76a8..c1226526b0 100644
--- a/src/smpi/mpi/smpi_comm.cpp
+++ b/src/smpi/mpi/smpi_comm.cpp
@@ -1,4 +1,4 @@
-/* Copyright (c) 2010-2020. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2010-2021. 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. */
@@ -16,8 +16,18 @@
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_comm, smpi, "Logging specific to SMPI (comm)");
 
-simgrid::smpi::Comm mpi_MPI_COMM_UNINITIALIZED;
-MPI_Comm MPI_COMM_UNINITIALIZED=&mpi_MPI_COMM_UNINITIALIZED;
+simgrid::smpi::Comm smpi_MPI_COMM_UNINITIALIZED;
+MPI_Comm MPI_COMM_UNINITIALIZED=&smpi_MPI_COMM_UNINITIALIZED;
+/**
+ * Setting MPI_COMM_WORLD to MPI_COMM_UNINITIALIZED (it's a variable)
+ * is important because the implementation of MPI_Comm checks
+ * "this == MPI_COMM_UNINITIALIZED"? If yes, it uses smpi_process()->comm_world()
+ * instead of "this".
+ * This is basically how we only have one global variable but all processes have
+ * different communicators (the one their SMPI instance uses).
+ *
+ */
+MPI_Comm MPI_COMM_WORLD = MPI_COMM_UNINITIALIZED;
 
 /* Support for cartesian topology was added, but there are 2 other types of topology, graph et dist graph. In order to
  * support them, we have to add a field SMPI_Topo_type, and replace the MPI_Topology field by an union. */
@@ -34,6 +44,8 @@ Comm::Comm(MPI_Group group, MPI_Topology topo, bool smp, int in_id)
   errhandler_->ref();
   //First creation of comm is done before SIMIX_run, so only do comms for others
   if(in_id==MPI_UNDEFINED && smp==0 && this->rank()!=MPI_UNDEFINED ){
+    this->add_f();
+    group->c2f();
     int id;
     if(this->rank()==0){
       static int global_id_ = 0;
@@ -53,6 +65,8 @@ void Comm::destroy(Comm* comm)
     Comm::destroy(smpi_process()->comm_world());
     return;
   }
+  if(comm != MPI_COMM_WORLD)
+    comm->mark_as_deleted();
   Comm::unref(comm);
 }
 
@@ -99,7 +113,10 @@ int Comm::dup(MPI_Comm* newcomm){
   if(info_!=MPI_INFO_NULL)
     (*newcomm)->info_ = new simgrid::smpi::Info(info_);
   //duplicate errhandler
-  (*newcomm)->set_errhandler(errhandler_);
+  if (errhandlers_ != nullptr)//MPI_COMM_WORLD, only grab our own
+    (*newcomm)->set_errhandler(errhandlers_[this->rank()]);
+  else
+    (*newcomm)->set_errhandler(errhandler_);
   return ret;
 }
 
@@ -136,7 +153,7 @@ int Comm::rank() const
 {
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->rank();
-  return group_->rank(s4u::Actor::self());
+  return group_->rank(s4u::this_actor::get_pid());
 }
 
 int Comm::id() const
@@ -153,14 +170,23 @@ void Comm::get_name(char* name, int* len) const
   if(this == MPI_COMM_WORLD && name_.empty()) {
     strncpy(name, "MPI_COMM_WORLD", 15);
     *len = 14;
-  } else if(this == MPI_COMM_SELF && name_.empty()) {
-    strncpy(name, "MPI_COMM_SELF", 14);
-    *len = 13;
   } else {
     *len = snprintf(name, MPI_MAX_NAME_STRING+1, "%s", name_.c_str());
   }
 }
 
+std::string Comm::name() const
+{
+  int size;
+  char name[MPI_MAX_NAME_STRING+1];
+  this->get_name(name, &size);
+  if (name[0]=='\0')
+    return std::string("MPI_Comm");
+  else
+    return std::string(name);
+}
+
+
 void Comm::set_name (const char* name)
 {
   if (this == MPI_COMM_UNINITIALIZED){
@@ -233,27 +259,21 @@ MPI_Comm Comm::split(int color, int key)
   if (this == MPI_COMM_UNINITIALIZED)
     return smpi_process()->comm_world()->split(color, key);
   int system_tag = -123;
-  int* recvbuf;
 
   MPI_Group group_root = nullptr;
   MPI_Group group_out  = nullptr;
-  MPI_Group group      = this->group();
+  const Group* group   = this->group();
   int myrank           = this->rank();
   int size             = this->size();
   /* Gather all colors and keys on rank 0 */
-  int* sendbuf = xbt_new(int, 2);
-  sendbuf[0] = color;
-  sendbuf[1] = key;
-  if (myrank == 0) {
-    recvbuf = xbt_new(int, 2 * size);
-  } else {
-    recvbuf = nullptr;
-  }
-  gather__default(sendbuf, 2, MPI_INT, recvbuf, 2, MPI_INT, 0, this);
-  xbt_free(sendbuf);
+  const std::array<int, 2> sendbuf = {{color, key}};
+  std::vector<int> recvbuf;
+  if (myrank == 0)
+    recvbuf.resize(2 * size);
+  gather__default(sendbuf.data(), 2, MPI_INT, recvbuf.data(), 2, MPI_INT, 0, this);
   /* Do the actual job */
   if (myrank == 0) {
-    MPI_Group* group_snd = xbt_new(MPI_Group, size);
+    std::vector<MPI_Group> group_snd(size);
     std::vector<std::pair<int, int>> rankmap;
     rankmap.reserve(size);
     for (int i = 0; i < size; i++) {
@@ -274,10 +294,10 @@ MPI_Comm Comm::split(int color, int key)
           group_root = group_out; /* Save root's group */
         }
         for (unsigned j = 0; j < rankmap.size(); j++) {
-          s4u::Actor* actor = group->actor(rankmap[j].second);
+          aid_t actor = group->actor(rankmap[j].second);
           group_out->set_mapping(actor, j);
         }
-        MPI_Request* requests = xbt_new(MPI_Request, rankmap.size());
+        std::vector<MPI_Request> requests(rankmap.size());
         int reqs              = 0;
         for (auto const& rank : rankmap) {
           if (rank.second != 0) {
@@ -289,12 +309,9 @@ MPI_Comm Comm::split(int color, int key)
         if(i != 0 && group_out != MPI_COMM_WORLD->group() && group_out != MPI_GROUP_EMPTY)
           Group::unref(group_out);
 
-        Request::waitall(reqs, requests, MPI_STATUS_IGNORE);
-        xbt_free(requests);
+        Request::waitall(reqs, requests.data(), MPI_STATUS_IGNORE);
       }
     }
-    xbt_free(recvbuf);
-    xbt_free(group_snd);
     group_out = group_root; /* exit with root's group */
   } else {
     if(color != MPI_UNDEFINED) {
@@ -328,41 +345,48 @@ void Comm::unref(Comm* comm){
     return;
   }
   comm->refcount_--;
-  Group::unref(comm->group_);
 
   if(comm->refcount_==0){
+    if(simgrid::smpi::F2C::lookup() != nullptr)
+      F2C::free_f(comm->f2c_id());
     comm->cleanup_smp();
     comm->cleanup_attr<Comm>();
     if (comm->info_ != MPI_INFO_NULL)
       simgrid::smpi::Info::unref(comm->info_);
-    if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
+    if(comm->errhandlers_!=nullptr){
+      for (int i=0; i<comm->size(); i++)
+	if (comm->errhandlers_[i]!=MPI_ERRHANDLER_NULL)
+          simgrid::smpi::Errhandler::unref(comm->errhandlers_[i]);
+      delete[] comm->errhandlers_;
+    } else if (comm->errhandler_ != MPI_ERRHANDLER_NULL)
       simgrid::smpi::Errhandler::unref(comm->errhandler_);
-    delete comm;
   }
+  Group::unref(comm->group_);
+  if(comm->refcount_==0)
+    delete comm;
 }
 
 MPI_Comm Comm::find_intra_comm(int * leader){
   //get the indices of all processes sharing the same simix host
-  auto& actor_list        = sg_host_self()->pimpl_->actor_list_;
   int intra_comm_size     = 0;
   int min_index           = INT_MAX; // the minimum index will be the leader
-  for (auto& actor : actor_list) {
-    int index = actor.get_pid();
-    if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) { // Is this process in the current group?
+  sg_host_self()->get_impl()->foreach_actor([this, &intra_comm_size, &min_index](auto& actor) {
+    if (this->group()->rank(actor.get_pid()) != MPI_UNDEFINED) { // Is this process in the current group?
       intra_comm_size++;
+      int index = actor.get_pid();
       if (index < min_index)
         min_index = index;
     }
-  }
+  });
   XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
   auto* group_intra = new Group(intra_comm_size);
   int i = 0;
-  for (auto& actor : actor_list) {
-    if (this->group()->rank(actor.get_ciface()) != MPI_UNDEFINED) {
-      group_intra->set_mapping(actor.get_ciface(), i);
+  sg_host_self()->get_impl()->foreach_actor([this, group_intra, &i](auto& actor) {
+    if (this->group()->rank(actor.get_pid()) != MPI_UNDEFINED) {
+      group_intra->set_mapping(actor.get_pid(), i);
       i++;
     }
-  }
+  });
   *leader=min_index;
   return new Comm(group_intra, nullptr, true);
 }
@@ -429,7 +453,7 @@ void Comm::init_smp(){
   if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && this!=MPI_COMM_WORLD){
     //create leader_communicator
     for (i=0; i< leader_group_size;i++)
-      leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
+      leaders_group->set_mapping(leader_list[i], i);
     leader_comm = new Comm(leaders_group, nullptr, true);
     this->set_leaders_comm(leader_comm);
     this->set_intra_comm(comm_intra);
@@ -437,7 +461,7 @@ void Comm::init_smp(){
     // create intracommunicator
   }else{
     for (i=0; i< leader_group_size;i++)
-      leaders_group->set_mapping(s4u::Actor::by_pid(leader_list[i]).get(), i);
+      leaders_group->set_mapping(leader_list[i], i);
 
     if(this->get_leaders_comm()==MPI_COMM_NULL){
       leader_comm = new Comm(leaders_group, nullptr, true);
@@ -469,7 +493,7 @@ void Comm::init_smp(){
     }
     is_uniform_=is_uniform;
   }
-  bcast__scatter_LR_allgather(&(is_uniform_),1, MPI_INT, 0, comm_intra );
+  bcast__scatter_LR_allgather(&is_uniform_, 1, MPI_INT, 0, comm_intra);
 
   if (smpi_cfg_privatization() == SmpiPrivStrategies::MMAP) {
     // we need to switch as the called function may silently touch global variables
@@ -477,7 +501,7 @@ void Comm::init_smp(){
   }
   // Are the ranks blocked ? = allocated contiguously on the SMP nodes
   int is_blocked=1;
-  int prev=this->group()->rank(comm_intra->group()->actor(0));
+  int prev      = this->group()->rank(comm_intra->group()->actor(0));
   for (i = 1; i < my_local_size; i++) {
     int that = this->group()->rank(comm_intra->group()->actor(i));
     if (that != prev + 1) {
@@ -488,7 +512,7 @@ void Comm::init_smp(){
   }
 
   int global_blocked;
-  allreduce__default(&is_blocked, &(global_blocked), 1, MPI_INT, MPI_LAND, this);
+  allreduce__default(&is_blocked, &global_blocked, 1, MPI_INT, MPI_LAND, this);
 
   if(MPI_COMM_WORLD==MPI_COMM_UNINITIALIZED || this==MPI_COMM_WORLD){
     if(this->rank()==0){
@@ -508,8 +532,8 @@ MPI_Comm Comm::f2c(int id) {
     return MPI_COMM_SELF;
   } else if(id==0){
     return MPI_COMM_WORLD;
-  } else if(F2C::f2c_lookup() != nullptr && id >= 0) {
-    const auto& lookup = F2C::f2c_lookup();
+  } else if (F2C::lookup() != nullptr && id >= 0) {
+    const auto& lookup = F2C::lookup();
     auto comm          = lookup->find(id);
     return comm == lookup->end() ? MPI_COMM_NULL : static_cast<MPI_Comm>(comm->second);
   } else {
@@ -518,7 +542,7 @@ MPI_Comm Comm::f2c(int id) {
 }
 
 void Comm::free_f(int id) {
-  F2C::f2c_lookup()->erase(id);
+  F2C::lookup()->erase(id);
 }
 
 void Comm::add_rma_win(MPI_Win win){
@@ -558,18 +582,36 @@ void Comm::set_info(MPI_Info info)
 
 MPI_Errhandler Comm::errhandler()
 {
-  if (errhandler_ != MPI_ERRHANDLER_NULL)
-    errhandler_->ref();
-  return errhandler_;
+  if (this != MPI_COMM_WORLD){
+    if (errhandler_ != MPI_ERRHANDLER_NULL)
+      errhandler_->ref();
+    return errhandler_;
+  } else {
+    if(errhandlers_==nullptr)
+      return MPI_ERRORS_ARE_FATAL;
+    else {
+      if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
+        errhandlers_[this->rank()]->ref();
+      return errhandlers_[this->rank()];
+    }
+  }
 }
 
 void Comm::set_errhandler(MPI_Errhandler errhandler)
 {
-  if (errhandler_ != MPI_ERRHANDLER_NULL)
-    simgrid::smpi::Errhandler::unref(errhandler_);
-  errhandler_ = errhandler;
-  if (errhandler_ != MPI_ERRHANDLER_NULL)
-    errhandler_->ref();
+  if(this != MPI_COMM_WORLD){
+    if (errhandler_ != MPI_ERRHANDLER_NULL)
+      simgrid::smpi::Errhandler::unref(errhandler_);
+    errhandler_ = errhandler;
+  }else{
+    if(errhandlers_==nullptr)
+      errhandlers_= new MPI_Errhandler[this->size()]{MPI_ERRHANDLER_NULL};
+    if(errhandlers_[this->rank()] != MPI_ERRHANDLER_NULL)
+      simgrid::smpi::Errhandler::unref(errhandlers_[this->rank()]);
+    errhandlers_[this->rank()]=errhandler;
+  }
+  if (errhandler != MPI_ERRHANDLER_NULL)
+    errhandler->ref();
 }
 
 MPI_Comm Comm::split_type(int type, int /*key*/, const Info*)