From e40e78a8761909f61c3f70ea61b0476b6eac3bc4 Mon Sep 17 00:00:00 2001
From: Augustin Degomme <degomme@idpann.imag.fr>
Date: Wed, 3 Apr 2013 16:25:18 +0200
Subject: [PATCH] adapt two collectives of starmpi to avoid timing issues, by
 using only smpi calls inside

---
 src/smpi/colls/allreduce-rdb.c      | 24 ++++++++++++------------
 src/smpi/colls/allreduce-redbcast.c |  4 ++--
 src/smpi/colls/colls.h              |  1 +
 3 files changed, 15 insertions(+), 14 deletions(-)

diff --git a/src/smpi/colls/allreduce-rdb.c b/src/smpi/colls/allreduce-rdb.c
index 5e3cf46460..fec95bc5f9 100644
--- a/src/smpi/colls/allreduce-rdb.c
+++ b/src/smpi/colls/allreduce-rdb.c
@@ -6,7 +6,7 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
 {
   int nprocs, rank, type_size, tag = 543;
   int mask, dst, pof2, newrank, rem, newdst;
-  MPI_Aint extent;
+  MPI_Aint extent, lb;
   MPI_Status status;
   void *tmp_buf = NULL;
   /*
@@ -19,20 +19,20 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
      uop  = op_ptr->op;
      #endif
    */
-  MPI_Comm_size(comm, &nprocs);
-  MPI_Comm_rank(comm, &rank);
+  nprocs=smpi_comm_size(comm);
+  rank=smpi_comm_rank(comm);
 
-  MPI_Type_extent(dtype, &extent);
+  smpi_datatype_extent(dtype,&lb, &extent);
   tmp_buf = (void *) malloc(count * extent);
   if (!tmp_buf) {
     printf("Could not allocate memory for tmp_buf\n");
     return 1;
   }
 
-  MPI_Sendrecv(sbuff, count, dtype, rank, 500,
+  smpi_mpi_sendrecv(sbuff, count, dtype, rank, 500,
                rbuff, count, dtype, rank, 500, comm, &status);
 
-  MPI_Type_size(dtype, &type_size);
+  type_size=smpi_datatype_size(dtype);
 
   // find nearest power-of-two less than or equal to comm_size
   pof2 = 1;
@@ -52,7 +52,7 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
     // even       
     if (rank % 2 == 0) {
 
-      MPI_Send(rbuff, count, dtype, rank + 1, tag, comm);
+      smpi_mpi_send(rbuff, count, dtype, rank + 1, tag, comm);
 
       // temporarily set the rank to -1 so that this
       // process does not pariticipate in recursive
@@ -60,7 +60,7 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
       newrank = -1;
     } else                      // odd
     {
-      MPI_Recv(tmp_buf, count, dtype, rank - 1, tag, comm, &status);
+      smpi_mpi_recv(tmp_buf, count, dtype, rank - 1, tag, comm, &status);
       // do the reduction on received data. since the
       // ordering is right, it doesn't matter whether
       // the operation is commutative or not.
@@ -92,7 +92,7 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
 
       // Send the most current data, which is in recvbuf. Recv
       // into tmp_buf 
-      MPI_Sendrecv(rbuff, count, dtype, dst, tag, tmp_buf, count, dtype,
+      smpi_mpi_sendrecv(rbuff, count, dtype, dst, tag, tmp_buf, count, dtype,
                    dst, tag, comm, &status);
 
       // tmp_buf contains data received in this step.
@@ -108,7 +108,7 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
         star_reduction(op, rbuff, tmp_buf, &count, &dtype);
 
         // copy result back into recvbuf
-        MPI_Sendrecv(tmp_buf, count, dtype, rank, tag, rbuff, count,
+        smpi_mpi_sendrecv(tmp_buf, count, dtype, rank, tag, rbuff, count,
                      dtype, rank, tag, comm, &status);
       }
       mask <<= 1;
@@ -120,9 +120,9 @@ int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
 
   if (rank < 2 * rem) {
     if (rank % 2)               // odd 
-      MPI_Send(rbuff, count, dtype, rank - 1, tag, comm);
+      smpi_mpi_send(rbuff, count, dtype, rank - 1, tag, comm);
     else                        // even 
-      MPI_Recv(rbuff, count, dtype, rank + 1, tag, comm, &status);
+      smpi_mpi_recv(rbuff, count, dtype, rank + 1, tag, comm, &status);
   }
 
   free(tmp_buf);
diff --git a/src/smpi/colls/allreduce-redbcast.c b/src/smpi/colls/allreduce-redbcast.c
index c00fc2f025..bc6f286e00 100644
--- a/src/smpi/colls/allreduce-redbcast.c
+++ b/src/smpi/colls/allreduce-redbcast.c
@@ -4,7 +4,7 @@ int smpi_coll_tuned_allreduce_redbcast(void *buf, void *buf2, int count,
                                        MPI_Datatype datatype, MPI_Op op,
                                        MPI_Comm comm)
 {
-  MPI_Reduce(buf, buf2, count, datatype, op, 0, comm);
-  MPI_Bcast(buf2, count, datatype, 0, comm);
+  smpi_mpi_reduce(buf, buf2, count, datatype, op, 0, comm);
+  smpi_mpi_bcast(buf2, count, datatype, 0, comm);
   return 0;
 }
diff --git a/src/smpi/colls/colls.h b/src/smpi/colls/colls.h
index 43f5de93b8..242430d129 100644
--- a/src/smpi/colls/colls.h
+++ b/src/smpi/colls/colls.h
@@ -3,6 +3,7 @@
 
 #include <math.h>
 #include "smpi/mpi.h"
+#include "smpi/private.h"
 #include "xbt.h"
 
 void star_reduction(MPI_Op op, void *src, void *target, int *count, MPI_Datatype *dtype);
-- 
2.20.1