Have MPI collectives display their root and MPI_Op to handle mismatches lazily

[simgrid.git] / src / smpi / bindings / smpi_pmpi_coll.cpp

diff --git a/src/smpi/bindings/smpi_pmpi_coll.cpp b/src/smpi/bindings/smpi_pmpi_coll.cpp
index 2ae8177..7df4731 100644 (file)
--- a/src/smpi/bindings/smpi_pmpi_coll.cpp
+++ b/src/smpi/bindings/smpi_pmpi_coll.cpp
@@ -1,4 +1,4 @@
-/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved.          */
+/* Copyright (c) 2007-2022. 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. */
@@ -37,7 +37,7 @@ int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request)
 {
   CHECK_COMM(1)
   CHECK_REQUEST(2)
-
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier")
   const SmpiBenchGuard suspend_bench;
   aid_t pid = simgrid::s4u::this_actor::get_pid();
   TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier",
@@ -58,8 +58,7 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c
   return PMPI_Ibcast(buf, count, datatype, root, comm, MPI_REQUEST_IGNORED);
 }
 
-int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, 
-                   int root, MPI_Comm comm, MPI_Request* request)
+int PMPI_Ibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request)
 {
   SET_BUF1(buf)
   CHECK_COMM(5)
@@ -68,6 +67,9 @@ int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype,
   CHECK_BUFFER(1, buf, count, datatype)
   CHECK_ROOT(4)
   CHECK_REQUEST(6)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast");
+  name += " with root " + std::to_string(root);
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   aid_t pid = simgrid::s4u::this_actor::get_pid();
@@ -99,7 +101,6 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void
 {
   CHECK_COMM(8)
   SET_BUF1(sendbuf)
-  SET_BUF2(recvbuf)
   int rank = comm->rank();
   if(sendbuf != MPI_IN_PLACE){
     CHECK_COUNT(2, sendcount)
@@ -107,6 +108,7 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void
     CHECK_BUFFER(1,sendbuf, sendcount, sendtype)
   }
   if(rank == root){
+    SET_BUF2(recvbuf)
     CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
     CHECK_TYPE(6, recvtype)
     CHECK_COUNT(5, recvcount)
@@ -116,6 +118,9 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void
   }
   CHECK_ROOT(7)
   CHECK_REQUEST(9)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather");
+  name += " with root " + std::to_string(root);
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const void* real_sendbuf   = sendbuf;
   int real_sendcount         = sendcount;
@@ -159,7 +164,6 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
 {
   CHECK_COMM(9)
   SET_BUF1(sendbuf)
-  SET_BUF2(recvbuf)
   int rank = comm->rank();
   if(sendbuf != MPI_IN_PLACE){
     CHECK_TYPE(3, sendtype)
@@ -167,6 +171,7 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
   }
   CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
   if(rank == root){
+    SET_BUF2(recvbuf)
     CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
     CHECK_TYPE(6, recvtype)
     CHECK_NULL(5, MPI_ERR_COUNT, recvcounts)
@@ -176,6 +181,9 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
   }
   CHECK_ROOT(8)
   CHECK_REQUEST(10)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv");
+  name += " with root " + std::to_string(root);
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   if (rank == root){
     for (int i = 0; i < comm->size(); i++) {
@@ -196,13 +204,16 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
   aid_t pid        = simgrid::s4u::this_actor::get_pid();
 
   auto trace_recvcounts = std::make_shared<std::vector<int>>();
-  trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]);
+  if (rank == root)
+    trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]);
+  else //this is not significant outside of root, put 0 as we don't know if recvcounts is initialized
+    trace_recvcounts->insert(trace_recvcounts->end(), comm->size(), 0);
 
   TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv",
                      new simgrid::instr::VarCollTIData(
                          request == MPI_REQUEST_IGNORED ? "gatherv" : "igatherv", root,
-                         real_sendcount,
-                         nullptr, recvtype->size(), trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype),
+                         sendcount,
+                         nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype),
                          simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::gatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype,
@@ -237,6 +248,7 @@ int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, v
   CHECK_BUFFER(1, sendbuf, sendcount, sendtype)
   CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
   CHECK_REQUEST(8)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather")
 
   if (sendbuf == MPI_IN_PLACE) {
     sendbuf   = static_cast<char*>(recvbuf) + recvtype->get_extent() * recvcount * comm->rank();
@@ -294,6 +306,7 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype,
     CHECK_COUNT(5, recvcounts[i])
     CHECK_BUFFER(4, recvbuf, recvcounts[i], recvtype)
   }
+  CHECK_COLLECTIVE(comm, MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv")
 
   const SmpiBenchGuard suspend_bench;
   if (sendbuf == MPI_IN_PLACE) {
@@ -310,7 +323,7 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype,
       pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv",
       new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "allgatherv" : "iallgatherv", -1,
                                         sendcount, nullptr,
-                                        recvtype->size(), trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype),
+                                        -1, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype),
                                         simgrid::smpi::Datatype::encode(recvtype)));
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm);
@@ -331,10 +344,10 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
                   MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request)
 {
   CHECK_COMM(8)
-  SET_BUF1(sendbuf)
   SET_BUF2(recvbuf)
   int rank = comm->rank();
   if(rank == root){
+    SET_BUF1(sendbuf)
     CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
     CHECK_COUNT(2, sendcount)
     CHECK_TYPE(3, sendtype)
@@ -349,6 +362,9 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi
   }
   CHECK_ROOT(8)
   CHECK_REQUEST(9)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter");
+  name += " with root " + std::to_string(root);
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   if (recvbuf == MPI_IN_PLACE) {
     recvtype  = sendtype;
@@ -386,12 +402,10 @@ int PMPI_Scatterv(const void *sendbuf, const int *sendcounts, const int *displs,
 int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount,
                    MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request)
 {
-  SET_BUF1(sendbuf)
   SET_BUF2(recvbuf)
   CHECK_COMM(9)
   int rank = comm->rank();
   if(recvbuf != MPI_IN_PLACE){
-    CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
     CHECK_COUNT(5, recvcount)
     CHECK_TYPE(7, recvtype)
     CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
@@ -399,6 +413,8 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs
   CHECK_ROOT(9)
   CHECK_REQUEST(10)
   if (rank == root) {
+    SET_BUF1(sendbuf)
+    CHECK_NOT_IN_PLACE_ROOT(1, sendbuf)
     CHECK_NULL(2, MPI_ERR_COUNT, sendcounts)
     CHECK_NULL(3, MPI_ERR_ARG, displs)
     CHECK_TYPE(4, sendtype)
@@ -413,17 +429,24 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs
   } else {
     CHECK_NOT_IN_PLACE_ROOT(4, recvbuf)
   }
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv");
+  name += " with root " + std::to_string(root);
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
 
   aid_t pid        = simgrid::s4u::this_actor::get_pid();
 
   auto trace_sendcounts = std::make_shared<std::vector<int>>();
-  trace_sendcounts->insert(trace_sendcounts->end(), &sendcounts[0], &sendcounts[comm->size()]);
+  if (rank == root)
+    trace_sendcounts->insert(trace_sendcounts->end(), &sendcounts[0], &sendcounts[comm->size()]);
+  else //this is not significant outside of root, put 0 as we don't know if sendcounts is initialized
+    trace_sendcounts->insert(trace_sendcounts->end(), comm->size(), 0);
+
 
   TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv",
                      new simgrid::instr::VarCollTIData(
-                         request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, sendtype->size(),
+                         request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, -1,
                          trace_sendcounts, recvcount,
                          nullptr, simgrid::smpi::Datatype::encode(sendtype),
                          simgrid::smpi::Datatype::encode(recvtype)));
@@ -446,18 +469,22 @@ int PMPI_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat
 {
   CHECK_COMM(7)
   SET_BUF1(sendbuf)
-  SET_BUF2(recvbuf)
   int rank = comm->rank();
   CHECK_TYPE(4, datatype)
   CHECK_COUNT(3, count)
   CHECK_BUFFER(1, sendbuf, count, datatype)
   if(rank == root){
+    SET_BUF2(recvbuf)
     CHECK_NOT_IN_PLACE(2, recvbuf)
     CHECK_BUFFER(5, recvbuf, count, datatype)
   }
   CHECK_OP(5, op, datatype)
   CHECK_ROOT(7)
   CHECK_REQUEST(8)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce");
+  name += " with op " + op->name();
+  name += " and root " + std::to_string(root);
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   aid_t pid = simgrid::s4u::this_actor::get_pid();
@@ -508,6 +535,9 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype
   CHECK_BUFFER(1, sendbuf, count, datatype)
   CHECK_BUFFER(2, recvbuf, count, datatype)
   CHECK_REQUEST(7)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Alleduce" : "PMPI_Iallreduce");
+  name += " with op " + op->name();
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   std::vector<unsigned char> tmp_sendbuf;
@@ -545,6 +575,9 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat
   CHECK_BUFFER(2,recvbuf,count, datatype)
   CHECK_REQUEST(7)
   CHECK_OP(5, op, datatype)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan");
+  name += " with op " + op->name();
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   aid_t pid = simgrid::s4u::this_actor::get_pid();
@@ -552,8 +585,8 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype);
 
   TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan",
-                     new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1,
-                                                     count, simgrid::smpi::Datatype::encode(datatype)));
+                     new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, 0.0,
+                                                    count, 0, simgrid::smpi::Datatype::encode(datatype), ""));
 
   int retval;
   if (request == MPI_REQUEST_IGNORED)
@@ -580,6 +613,9 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat
   CHECK_BUFFER(2, recvbuf, count, datatype)
   CHECK_REQUEST(7)
   CHECK_OP(5, op, datatype)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan");
+  name += " with op " + op->name();
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   aid_t pid = simgrid::s4u::this_actor::get_pid();
@@ -587,8 +623,8 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat
   const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype);
 
   TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan",
-                     new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1,
-                                                     count, simgrid::smpi::Datatype::encode(datatype)));
+                     new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, 0.0,
+                                                    count, 0, simgrid::smpi::Datatype::encode(datatype), ""));
 
   int retval;
   if (request == MPI_REQUEST_IGNORED)
@@ -621,6 +657,9 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun
     CHECK_BUFFER(1, sendbuf, recvcounts[i], datatype)
     CHECK_BUFFER(2, recvbuf, recvcounts[i], datatype)
   }
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter");
+  name += " with op " + op->name();
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   aid_t pid                          = simgrid::s4u::this_actor::get_pid();
@@ -637,8 +676,8 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun
 
   TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter",
                      new simgrid::instr::VarCollTIData(
-                         request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, datatype->size(), nullptr,
-                         0, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), ""));
+                         request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1, nullptr,
+                         -1 , trace_recvcounts, std::to_string(0), simgrid::smpi::Datatype::encode(datatype)));
 
   if (request == MPI_REQUEST_IGNORED)
     simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm);
@@ -667,6 +706,9 @@ int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount
   CHECK_BUFFER(2, recvbuf, recvcount, datatype)
   CHECK_REQUEST(7)
   CHECK_OP(5, op, datatype)
+  std::string name = (request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block");
+  name += " with op " + op->name();
+  CHECK_COLLECTIVE(comm, name.c_str())
 
   const SmpiBenchGuard suspend_bench;
   int count = comm->size();
@@ -679,8 +721,8 @@ int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount
 
   TRACE_smpi_comm_in(
       pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block",
-      new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, 0,
-                                        nullptr, 0, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), ""));
+      new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1,
+                                        nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), ""));
 
   std::vector<int> recvcounts(count);
   for (int i      = 0; i < count; i++)
@@ -715,6 +757,7 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo
   CHECK_COUNT(5, recvcount)
   CHECK_BUFFER(4, recvbuf, recvcount, recvtype)
   CHECK_REQUEST(8)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall")
 
   aid_t pid                  = simgrid::s4u::this_actor::get_pid();
   int real_sendcount         = sendcount;
@@ -773,6 +816,7 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd
   CHECK_NULL(6, MPI_ERR_COUNT, recvcounts)
   CHECK_NULL(7, MPI_ERR_ARG, recvdispls)
   CHECK_REQUEST(10)
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv")
 
   aid_t pid = simgrid::s4u::this_actor::get_pid();
   int size = comm->size();
@@ -801,7 +845,7 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd
   std::vector<unsigned char> tmp_sendbuf;
   std::vector<int> tmp_sendcounts;
   std::vector<int> tmp_senddispls;
-  const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR);
+  const void* real_sendbuf;
 
   if (sendbuf == MPI_IN_PLACE) {
     tmp_sendcounts.assign(recvcounts, recvcounts + size);
@@ -817,6 +861,7 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd
     if (((recvdispls[i] + recvcounts[i]) * dt_size_recv) > maxsize)
       maxsize = (recvdispls[i] + recvcounts[i]) * dt_size_recv;
   }
+  real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR);
 
   if(recvtype->size() * recvcounts[comm->rank()] !=  real_sendtype->size() * real_sendcounts[comm->rank()]){
     XBT_WARN("MPI_(I)Alltoallv : receive size from me differs from sent size to me : %zu vs %zu", recvtype->size() * recvcounts[comm->rank()], real_sendtype->size() * real_sendcounts[comm->rank()]);
@@ -876,6 +921,7 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd
     CHECK_TYPE(8, recvtypes[i])
     CHECK_BUFFER(5, recvbuf, recvcounts[i], recvtypes[i])
   }
+  CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw")
 
   const SmpiBenchGuard suspend_bench;