fix for reduce and allreduce for an unsymmetric case

This commit was SVN r7802.

ompi/mca/coll/hierarch/coll_hierarch.c

@@ -171,8 +171,8 @@ mca_coll_hierarch_comm_query(struct ompi_communicator_t *comm, int *priority,
     
     size = ompi_comm_size(comm);
     
-    if ( size == 1 ) {
+    if ( size < 3 ) {
-	/* No need for hierarchical collectives . */
+	/* No need for hierarchical collectives for 1 or 2 procs. */
 	return NULL;
     }
 
@@ -550,10 +550,18 @@ struct ompi_communicator_t*  mca_coll_hierarch_get_llcomm (int root,
 	  continue;
 	}
 
-	if (llead->offset >= offset ) {
+	if (llead->offset == offset ) {
 	    found = 1;
 	    break;
 	}
+#if 0
+	else if () {
+	  /* the offset of root = maxoffset of this color and
+	   * the offset on llead is larger then offset of root.
+	   * then we can also use this llead structure 
+	   */
+	}
+#endif
     }
     
     if ( !found ) {

ompi/mca/coll/hierarch/coll_hierarch_allreduce.c

@@ -61,8 +61,6 @@ int mca_coll_hierarch_allreduce_intra(void *sbuf, void *rbuf, int count,
     llcomm = mca_coll_hierarch_get_llcomm ( root, data, &llroot, &lroot);
 
     if ( MPI_COMM_NULL != lcomm ) {
-      lrank = ompi_comm_rank (lcomm);
-	if ( lrank == lroot ) {
       ompi_ddt_get_extent(dtype, &lb, &extent);
       ompi_ddt_get_true_extent(dtype, &true_lb, &true_extent);
       
@@ -71,7 +69,6 @@ int mca_coll_hierarch_allreduce_intra(void *sbuf, void *rbuf, int count,
 	return OMPI_ERR_OUT_OF_RESOURCE;
       }
       tmpbuf = tbuf - lb;
-	}
       
       if ( MPI_IN_PLACE != sbuf ) {
 	ret = lcomm->c_coll.coll_reduce (sbuf, tmpbuf, count, dtype, 
@@ -87,9 +84,15 @@ int mca_coll_hierarch_allreduce_intra(void *sbuf, void *rbuf, int count,
     }
 
     if ( MPI_UNDEFINED != llroot ) {
+      if ( MPI_COMM_NULL != lcomm ) {
 	ret = llcomm->c_coll.coll_allreduce (tmpbuf, rbuf, count, dtype,
 					     op, llcomm);
       }
+      else {
+	ret = llcomm->c_coll.coll_allreduce (sbuf, rbuf, count, dtype,
+					     op, llcomm);
+      }
+    }	
 
     if ( MPI_COMM_NULL != lcomm ) {
 	ret = lcomm->c_coll.coll_bcast(rbuf, count, dtype, lroot, lcomm );

ompi/mca/coll/hierarch/coll_hierarch_reduce.c

@@ -60,8 +60,6 @@ int mca_coll_hierarch_reduce_intra(void *sbuf, void *rbuf, int count,
     llcomm = mca_coll_hierarch_get_llcomm ( root, data, &llroot, &lroot);
 
     if ( MPI_COMM_NULL != lcomm ) {
-      lrank = ompi_comm_rank (lcomm);
-	if ( lrank == lroot ) {
       ompi_ddt_get_extent(dtype, &lb, &extent);
       ompi_ddt_get_true_extent(dtype, &true_lb, &true_extent);
       
@@ -70,7 +68,7 @@ int mca_coll_hierarch_reduce_intra(void *sbuf, void *rbuf, int count,
 	return OMPI_ERR_OUT_OF_RESOURCE;
       }
       tmpbuf = tbuf - lb;
-	}
+    
 
       if ( MPI_IN_PLACE != sbuf ) {
 	ret = lcomm->c_coll.coll_reduce (sbuf, tmpbuf, count, dtype, 
@@ -86,9 +84,15 @@ int mca_coll_hierarch_reduce_intra(void *sbuf, void *rbuf, int count,
     }
 
     if ( MPI_UNDEFINED != llroot ) {
+      if ( MPI_COMM_NULL != lcomm ) { 
 	ret = llcomm->c_coll.coll_reduce (tmpbuf, rbuf, count, dtype,
 					  op, llroot, llcomm);
       }
+      else {
+	ret = llcomm->c_coll.coll_reduce (sbuf, rbuf, count, dtype,
+					  op, llroot, llcomm);
+      }
+    }
 
  exit:
     if ( NULL != tmpbuf ) {