Merge pull request #9856 from EmmanuelBRELLE/fix_roundup_libnbc-4.1.x

v4.1.x: Patched libnbc to not round fractions based on float precision

Author: jsquyres

ompi/mca/coll/libnbc/nbc_iallreduce.c

@@ -366,7 +366,7 @@ static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype dat
 
   root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */
   RANK2VRANK(rank, vrank, root);
-  maxr = (int)ceil((log((double)p)/LOG2));
+  maxr = ceil_of_log2(p);
   /* ensure the result ends up in recvbuf on vrank 0 */
   if (0 == (maxr%2)) {
     rbuf = (void *)(-gap);

ompi/mca/coll/libnbc/nbc_ibarrier.c

@@ -45,7 +45,7 @@ static int nbc_barrier_init(struct ompi_communicator_t *comm, ompi_request_t **
       return OMPI_ERR_OUT_OF_RESOURCE;
     }
 
-    maxround = (int)ceil((log((double)p)/LOG2)-1);
+    maxround = ceil_of_log2(p) -1;
 
     for (int round = 0 ; round <= maxround ; ++round) {
       sendpeer = (rank + (1 << round)) % p;

ompi/mca/coll/libnbc/nbc_ibcast.c

@@ -238,7 +238,7 @@ int ompi_coll_libnbc_ibcast(void *buffer, int count, MPI_Datatype datatype, int
 static inline int bcast_sched_binomial(int rank, int p, int root, NBC_Schedule *schedule, void *buffer, int count, MPI_Datatype datatype) {
   int maxr, vrank, peer, res;
 
-  maxr = (int)ceil((log((double)p)/LOG2));
+  maxr = ceil_of_log2(p);
 
   RANK2VRANK(rank, vrank, root);
 

ompi/mca/coll/libnbc/nbc_internal.h

@@ -51,8 +51,15 @@
 extern "C" {
 #endif
 
-/* log(2) */
-#define LOG2 0.69314718055994530941
+/* Dividing very close floats may lead to unexpected roundings */
+static inline int
+ceil_of_log2 (int val) {
+    int ret = 0;
+    while (1 << ret < val) {
+        ret ++;
+    }
+    return ret;
+}
 
 /* true/false */
 #define true 1

ompi/mca/coll/libnbc/nbc_ireduce.c

@@ -367,7 +367,7 @@ static inline int red_sched_binomial (int rank, int p, int root, const void *sen
     vroot = 0;
   }
   RANK2VRANK(rank, vrank, vroot);
-  maxr = (int)ceil((log((double)p)/LOG2));
+  maxr = ceil_of_log2(p);
 
   if (rank != root) {
     inplace = 0;

ompi/mca/coll/libnbc/nbc_ireduce_scatter.c

@@ -83,7 +83,7 @@ static int nbc_reduce_scatter_init(const void* sendbuf, void* recvbuf, const int
     return nbc_get_noop_request(persistent, request);
   }
 
-  maxr = (int) ceil ((log((double) p) / LOG2));
+  maxr = ceil_of_log2(p);
 
   span = opal_datatype_span(&datatype->super, count, &gap);
   span_align = OPAL_ALIGN(span, datatype->super.align, ptrdiff_t);

ompi/mca/coll/libnbc/nbc_ireduce_scatter_block.c

@@ -68,7 +68,7 @@ static int nbc_reduce_scatter_block_init(const void* sendbuf, void* recvbuf, int
     return OMPI_ERR_OUT_OF_RESOURCE;
   }
 
-  maxr = (int)ceil((log((double)p)/LOG2));
+  maxr = ceil_of_log2(p);
 
   count = p * recvcount;