Correctly compute the number of elements in a segment.

For broadcast send the correct size for all intermediary nodes.

This commit was SVN r12552.

ompi/mca/coll/tuned/coll_tuned.h

@@ -405,7 +405,8 @@ do {
  * expected by the collective call.
  */
 #define COLL_TUNED_COMPUTED_SEGCOUNT(SEGSIZE, TYPELNG, SEGCOUNT)        \
-    if( (SEGSIZE) >= (TYPELNG) ) {                                      \
+    if( ((SEGSIZE) >= (TYPELNG)) &&                                     \
+        ((SEGSIZE) < ((TYPELNG) * (SEGCOUNT))) ) {                      \
         size_t residual;                                                \
         (SEGCOUNT) = (int)((SEGSIZE) / (TYPELNG));                      \
         residual = (SEGSIZE) - (SEGCOUNT) * (TYPELNG);                  \

ompi/mca/coll/tuned/coll_tuned_bcast.c

@@ -140,11 +140,11 @@ ompi_coll_tuned_bcast_intra_generic( void* buffer,
             for( i = 0; i < tree->tree_nextsize; i++ ) {  /* send data to children */
                 /* send data */
 #if defined(COLL_TUNED_BCAST_USE_BLOCKING)
-                err = MCA_PML_CALL(send(tmpbuf, sendcount, datatype,
+                err = MCA_PML_CALL(send(tmpbuf, count_by_segment, datatype,
                                         tree->tree_next[i], MCA_COLL_BASE_TAG_BCAST,
                                         MCA_PML_BASE_SEND_STANDARD, comm));
 #else
-                err = MCA_PML_CALL(isend(tmpbuf, sendcount, datatype,
+                err = MCA_PML_CALL(isend(tmpbuf, count_by_segment, datatype,
                                          tree->tree_next[i], MCA_COLL_BASE_TAG_BCAST,
                                          MCA_PML_BASE_SEND_STANDARD, comm, &send_reqs[i]));
 #endif  /* COLL_TUNED_BCAST_USE_BLOCKING */