coll/libnbc/ialltoallv: skip send/recv 0-byte data

As per MPI specification the amount of data sent must be equal to the
amount of data received for each communication pair, and therefore
both count and datatype size should be accounted for to determine if
the data is 0-byte and therefore skippable.

Signed-off-by: Wenduo Wang <wenduwan@amazon.com>

ompi/mca/coll/libnbc/nbc_ialltoallv.c

@@ -107,7 +107,7 @@ static int nbc_alltoallv_init(const void* sendbuf, const int *sendcounts, const
   }
 
 
-  if (!inplace && sendcounts[rank] != 0) {
+  if (!inplace && 0 < sendcounts[rank] && 0 < sndext) {
     rbuf = (char *) recvbuf + rdispls[rank] * rcvext;
     sbuf = (char *) sendbuf + sdispls[rank] * sndext;
     res = NBC_Sched_copy (sbuf, false, sendcounts[rank], sendtype,
@@ -272,7 +272,7 @@ static inline int a2av_sched_linear(int rank, int p, NBC_Schedule *schedule,
     }
 
     /* post send */
-    if (sendcounts[i] != 0) {
+    if (0 < sendcounts[i] && 0 < sndext) {
       char *sbuf = ((char *) sendbuf) + (sdispls[i] * sndext);
       res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtype, i, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -281,7 +281,7 @@ static inline int a2av_sched_linear(int rank, int p, NBC_Schedule *schedule,
     }
 
     /* post receive */
-    if (recvcounts[i] != 0) {
+    if (0 < recvcounts[i] && 0 < rcvext) {
       char *rbuf = ((char *) recvbuf) + (rdispls[i] * rcvext);
       res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtype, i, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -306,7 +306,7 @@ static inline int a2av_sched_pairwise(int rank, int p, NBC_Schedule *schedule,
     int rcvpeer = (rank + p - i) %p;
 
     /* post send */
-    if (sendcounts[sndpeer] != 0) {
+    if (0 < sendcounts[sndpeer] && 0 < sndext) {
       char *sbuf = ((char *) sendbuf) + (sdispls[sndpeer] * sndext);
       res = NBC_Sched_send(sbuf, false, sendcounts[sndpeer], sendtype, sndpeer, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -315,7 +315,7 @@ static inline int a2av_sched_pairwise(int rank, int p, NBC_Schedule *schedule,
     }
 
     /* post receive */
-    if (recvcounts[rcvpeer] != 0) {
+    if (0 < recvcounts[rcvpeer] && 0 < rcvext) {
       char *rbuf = ((char *) recvbuf) + (rdispls[rcvpeer] * rcvext);
       res = NBC_Sched_recv(rbuf, false, recvcounts[rcvpeer], recvtype, rcvpeer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
@@ -338,34 +338,34 @@ static inline int a2av_sched_inplace(int rank, int p, NBC_Schedule *schedule,
     char *sbuf = (char *) buf + displs[speer] * ext;
     char *rbuf = (char *) buf + displs[rpeer] * ext;
 
-    if (0 != counts[rpeer]) {
+    if (0 < counts[rpeer] && 0 < ext) {
       res = NBC_Sched_copy (rbuf, false, counts[rpeer], type,
                             (void *)(-gap), true, counts[rpeer], type,
                             schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
     }
-    if (0 != counts[speer]) {
+    if (0 < counts[speer] && 0 < ext) {
       res = NBC_Sched_send (sbuf, false , counts[speer], type, speer, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
     }
-    if (0 != counts[rpeer]) {
+    if (0 < counts[rpeer] && 0 < ext) {
       res = NBC_Sched_recv (rbuf, false , counts[rpeer], type, rpeer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
     }
 
-    if (0 != counts[rpeer]) {
+    if (0 < counts[rpeer] && 0 < ext) {
       res = NBC_Sched_send ((void *)(-gap), true, counts[rpeer], type, rpeer, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
       }
     }
-    if (0 != counts[speer]) {
+    if (0 < counts[speer] && 0 < ext) {
       res = NBC_Sched_recv (sbuf, false, counts[speer], type, speer, schedule, true);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;
@@ -374,15 +374,17 @@ static inline int a2av_sched_inplace(int rank, int p, NBC_Schedule *schedule,
   }
   if (0 == (p%2)) {
     int peer = (rank + p/2) % p;
-
     char *tbuf = (char *) buf + displs[peer] * ext;
-    res = NBC_Sched_copy (tbuf, false, counts[peer], type,
-                          (void *)(-gap), true, counts[peer], type,
-                          schedule, true);
-    if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
-      return res;
+
+    if (0 < counts[peer] && 0 < ext) {
+        res = NBC_Sched_copy(tbuf, false, counts[peer], type, (void *) (-gap), true, counts[peer],
+                             type, schedule, true);
+        if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
+            return res;
+        }
     }
-    if (0 != counts[peer]) {
+
+    if (0 < counts[peer] && 0 < ext) {
       res = NBC_Sched_send ((void *)(-gap), true , counts[peer], type, peer, schedule, false);
       if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
         return res;