have smp-aware algorithms use number of cores on the node as basis for their computat...

[simgrid.git] / src / smpi / colls / allreduce-smp-rsag.c

diff --git a/src/smpi/colls/allreduce-smp-rsag.c b/src/smpi/colls/allreduce-smp-rsag.c
index 53d1a7f..9de2bcf 100644 (file)
--- a/src/smpi/colls/allreduce-smp-rsag.c
+++ b/src/smpi/colls/allreduce-smp-rsag.c
@@ -19,10 +19,13 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
 {
   int comm_size, rank;
   void *tmp_buf;
-  int tag = 50;
+  int tag = COLL_TAG_ALLREDUCE;
   int mask, src, dst;
   MPI_Status status;
-  int num_core = NUM_CORE;
+  int num_core = simcall_host_get_core(SIMIX_host_self());
+  // do we use the default one or the number of cores in the platform ?
+  // if the number of cores is one, the platform may be simulated with 1 node = 1 core
+  if (num_core == 1) num_core = NUM_CORE;
   /*
      #ifdef MPICH2_REDUCTION
      MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
@@ -65,7 +68,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
       //      if (src < ((inter_rank + 1) * num_core)) {
       if (src < comm_size) {
         smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
-        star_reduction(op, tmp_buf, recv_buf, &count, &dtype);
+        smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
         //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
       }
     } else {
@@ -105,7 +108,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
                    &status);
 
       // result is in rbuf
-      star_reduction(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
+      smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
                      &dtype);
     }