OSHMEM/COLL: optimization on zero-length ops

- removed barrier call on zero-length operations

Signed-off-by: Sergey Oblomov <sergeyo@mellanox.com>
(cherry picked from commit ff2fd0679eb4b31bfd840395d73746612e2670f4)

oshmem/mca/scoll/basic/scoll_basic_alltoall.c

@@ -61,17 +61,20 @@ int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
         return OSHMEM_ERR_BAD_PARAM;
     }
 
-    if (nelems) {
+    /* Do nothing on zero-length request */
-        if ((sst == 1) && (dst == 1)) {
+    if (OPAL_UNLIKELY(!nelems)) {
-            rc = a2a_alg_simple(group, target, source, nelems, element_size);
+        return OPAL_SUCCESS;
-        } else {
+    }
-            rc = a2as_alg_simple(group, target, source, dst, sst, nelems,
-                                 element_size);
-        }
 
-        if (rc != OSHMEM_SUCCESS) {
+    if ((sst == 1) && (dst == 1)) {
-           return rc;
+        rc = a2a_alg_simple(group, target, source, nelems, element_size);
-        }
+    } else {
+        rc = a2as_alg_simple(group, target, source, dst, sst, nelems,
+                             element_size);
+    }
+
+    if (rc != OSHMEM_SUCCESS) {
+       return rc;
     }
 
     /* quiet is needed because scoll level barrier does not

oshmem/mca/scoll/basic/scoll_basic_broadcast.c

@@ -55,6 +55,11 @@ int mca_scoll_basic_broadcast(struct oshmem_group_t *group,
     if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
         int i = 0;
 
+        /* Do nothing on zero-length request */
+        if (OPAL_UNLIKELY(!nlong)) {
+            return OSHMEM_SUCCESS;
+        }
+
         if (pSync) {
             alg = (alg == SCOLL_DEFAULT_ALG ?
                     mca_scoll_basic_param_broadcast_algorithm : alg);
@@ -131,7 +136,7 @@ static int _algorithm_central_counter(struct oshmem_group_t *group,
                   group->my_pe, pSync[0], PE_root);
 
     /* Check if this PE is the root */
-    if ((PE_root == group->my_pe) && nlong) {
+    if (PE_root == group->my_pe) {
         int pe_cur = 0;
 
         SCOLL_VERBOSE(14,

oshmem/mca/scoll/basic/scoll_basic_collect.c

@@ -66,7 +66,12 @@ int mca_scoll_basic_collect(struct oshmem_group_t *group,
     if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
         int i = 0;
 
-        if (nlong_type && nlong) {
+        /* Do nothing on zero-length request */
+        if (OPAL_UNLIKELY(!nlong)) {
+            return OPAL_SUCCESS;
+        }
+
+        if (nlong_type) {
             alg = (alg == SCOLL_DEFAULT_ALG ?
                     mca_scoll_basic_param_collect_algorithm : alg);
             switch (alg) {
@@ -156,7 +161,7 @@ static int _algorithm_f_central_counter(struct oshmem_group_t *group,
                   group->my_pe);
     SCOLL_VERBOSE(15, "[#%d] pSync[0] = %ld", group->my_pe, pSync[0]);
 
-    if ((PE_root == group->my_pe) && nlong) {
+    if (PE_root == group->my_pe) {
         int pe_cur = 0;
 
         memcpy((void*) ((unsigned char*) target + 0 * nlong),
@@ -543,7 +548,7 @@ static int _algorithm_central_collector(struct oshmem_group_t *group,
     /* Set own data size */
     pSync[0] = (nlong ? (long)nlong : SHMEM_SYNC_READY);
 
-    if ((PE_root == group->my_pe) && nlong) {
+    if (PE_root == group->my_pe) {
         long value = 0;
         int pe_cur = 0;
         long wait_pe_count = 0;

oshmem/mca/scoll/basic/scoll_basic_reduce.c

@@ -78,10 +78,14 @@ int mca_scoll_basic_reduce(struct oshmem_group_t *group,
     if ((rc == OSHMEM_SUCCESS) && oshmem_proc_group_is_member(group)) {
         int i = 0;
 
+        /* Do nothing on zero-length request */
+        if (OPAL_UNLIKELY(!nlong)) {
+            return OSHMEM_SUCCESS;
+        }
+
         if (pSync) {
-            alg = (nlong ? (alg == SCOLL_DEFAULT_ALG ?
+            alg = (alg == SCOLL_DEFAULT_ALG ?
-                                   mca_scoll_basic_param_reduce_algorithm : alg) :
+                    mca_scoll_basic_param_reduce_algorithm : alg);
-                           SCOLL_ALG_REDUCE_CENTRAL_COUNTER );
             switch (alg) {
             case SCOLL_ALG_REDUCE_CENTRAL_COUNTER:
                 {
@@ -186,7 +190,7 @@ static int _algorithm_central_counter(struct oshmem_group_t *group,
 
     SCOLL_VERBOSE(12, "[#%d] Reduce algorithm: Central Counter", group->my_pe);
 
-    if ((PE_root == group->my_pe) && nlong) {
+    if (PE_root == group->my_pe) {
         int pe_cur = 0;
         void *target_cur = NULL;
 

oshmem/mca/scoll/mpi/scoll_mpi_ops.c

@@ -54,6 +54,12 @@ int mca_scoll_mpi_broadcast(struct oshmem_group_t *group,
     }
     dtype = &ompi_mpi_char.dt;
     root = oshmem_proc_group_find_id(group, PE_root);
+
+    /* Do nothing on zero-length request */
+    if (OPAL_UNLIKELY(!nlong)) {
+        return OSHMEM_SUCCESS;
+    }
+
     /* Open SHMEM specification has the following constrains (page 85):
      * "If using C/C++, nelems must be of type integer. If you are using Fortran, it must be a
      *  default integer value". And also fortran signature says "INTEGER".
@@ -61,7 +67,7 @@ int mca_scoll_mpi_broadcast(struct oshmem_group_t *group,
      *  and considering this contradiction, we cast size_t to int here
      *  in case if the value is less than INT_MAX and fallback to previous module otherwise. */
 #ifdef INCOMPATIBLE_SHMEM_OMPI_COLL_APIS
-    if ((INT_MAX < nlong) || !nlong) {
+    if (INT_MAX < nlong) {
         MPI_COLL_VERBOSE(20,"RUNNING FALLBACK BCAST");
         PREVIOUS_SCOLL_FN(mpi_module, broadcast, group,
                 PE_root,
@@ -104,7 +110,13 @@ int mca_scoll_mpi_collect(struct oshmem_group_t *group,
     void *sbuf, *rbuf;
     MPI_COLL_VERBOSE(20,"RUNNING MPI ALLGATHER");
     mpi_module = (mca_scoll_mpi_module_t *) group->g_scoll.scoll_collect_module;
-    if ((nlong_type == true) && nlong) {
+
+    /* Do nothing on zero-length request */
+    if (OPAL_UNLIKELY(!nlong)) {
+        return OSHMEM_SUCCESS;
+    }
+
+    if (nlong_type == true) {
         sbuf = (void *) source;
         rbuf = target;
         stype =  &ompi_mpi_char.dt;
@@ -177,6 +189,12 @@ int mca_scoll_mpi_reduce(struct oshmem_group_t *group,
     dtype = shmem_dtype_to_ompi_dtype(op);
     h_op = shmem_op_to_ompi_op(op->op);
     count = nlong/op->dt_size;
+
+    /* Do nothing on zero-length request */
+    if (OPAL_UNLIKELY(!nlong)) {
+        return OSHMEM_SUCCESS;
+    }
+
     /* Open SHMEM specification has the following constrains (page 85):
      * "If using C/C++, nelems must be of type integer. If you are using Fortran, it must be a
      *  default integer value". And also fortran signature says "INTEGER".
@@ -184,7 +202,7 @@ int mca_scoll_mpi_reduce(struct oshmem_group_t *group,
      *  and considering this contradiction, we cast size_t to int here
      *  in case if the value is less than INT_MAX and fallback to previous module otherwise. */
 #ifdef INCOMPATIBLE_SHMEM_OMPI_COLL_APIS
-    if ((INT_MAX < count) || !nlong) {
+    if (INT_MAX < count) {
         MPI_COLL_VERBOSE(20,"RUNNING FALLBACK REDUCE");
         PREVIOUS_SCOLL_FN(mpi_module, reduce, group,
                 op,