OSHMEM: added processing of zero-length collectives

- according spec 1.4, annex C shmem collectives should process
  calls where number of elements is zero independently from pointer
  value
- added zero-count processing - it just call barrier to
  sync ranks

Signed-off-by: Sergey Oblomov <sergeyo@mellanox.com>

Author: Sergey Oblomov

oshmem/mca/scoll/basic/scoll_basic_alltoall.c

@@ -61,15 +61,17 @@ int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
         return OSHMEM_ERR_BAD_PARAM;
     }
 
-    if ((sst == 1) && (dst == 1)) {
-        rc = a2a_alg_simple(group, target, source, nelems, element_size);
-    } else {
-        rc = a2as_alg_simple(group, target, source, dst, sst, nelems,
-                             element_size);
-    }
+    if (nelems) {
+        if ((sst == 1) && (dst == 1)) {
+            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;
+        if (rc != OSHMEM_SUCCESS) {
+           return rc;
+        }
     }
 
     /* quiet is needed because scoll level barrier does not

oshmem/mca/scoll/basic/scoll_basic_broadcast.c

@@ -131,7 +131,7 @@ 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) {
+    if ((PE_root == group->my_pe) && nlong) {
         int pe_cur = 0;
 
         SCOLL_VERBOSE(14,
@@ -192,6 +192,16 @@ static int _algorithm_binomial_tree(struct oshmem_group_t *group,
                   "[#%d] pSync[0] = %ld root = #%d",
                   group->my_pe, pSync[0], PE_root);
 
+    if (OPAL_UNLIKELY(!nlong)) {
+        SCOLL_VERBOSE(14, "[#%d] Wait for operation completion", group->my_pe);
+        /* wait until root finishes sending data  */
+        rc = BARRIER_FUNC(group,
+                (pSync + 1),
+                SCOLL_DEFAULT_ALG);
+        return rc;
+    }
+
+
     vrank = (my_id + group->proc_count - root_id) % group->proc_count;
     hibit = opal_hibit(vrank, dim);
 

oshmem/mca/scoll/basic/scoll_basic_collect.c

@@ -66,7 +66,7 @@ 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) {
+        if (nlong_type && nlong) {
             alg = (alg == SCOLL_DEFAULT_ALG ?
                     mca_scoll_basic_param_collect_algorithm : alg);
             switch (alg) {
@@ -156,7 +156,7 @@ 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) {
+    if ((PE_root == group->my_pe) && nlong) {
         int pe_cur = 0;
 
         memcpy((void*) ((unsigned char*) target + 0 * nlong),
@@ -543,7 +543,7 @@ 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) {
+    if ((PE_root == group->my_pe) && nlong) {
         long value = 0;
         int pe_cur = 0;
         long wait_pe_count = 0;

oshmem/mca/scoll/basic/scoll_basic_reduce.c

@@ -79,8 +79,9 @@ int mca_scoll_basic_reduce(struct oshmem_group_t *group,
         int i = 0;
 
         if (pSync) {
-            alg = (alg == SCOLL_DEFAULT_ALG ?
-                    mca_scoll_basic_param_reduce_algorithm : alg);
+            alg = (nlong ? (alg == SCOLL_DEFAULT_ALG ?
+                                   mca_scoll_basic_param_reduce_algorithm : alg) :
+                           SCOLL_ALG_REDUCE_CENTRAL_COUNTER );
             switch (alg) {
             case SCOLL_ALG_REDUCE_CENTRAL_COUNTER:
                 {
@@ -185,7 +186,7 @@ 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) {
+    if ((PE_root == group->my_pe) && nlong) {
         int pe_cur = 0;
         void *target_cur = NULL;
 

oshmem/mca/scoll/mpi/scoll_mpi_ops.c

@@ -61,7 +61,7 @@ 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) {
+    if ((INT_MAX < nlong) || !nlong) {
         MPI_COLL_VERBOSE(20,"RUNNING FALLBACK BCAST");
         PREVIOUS_SCOLL_FN(mpi_module, broadcast, group,
                 PE_root,
@@ -104,7 +104,7 @@ 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) {
+    if ((nlong_type == true) && nlong) {
         sbuf = (void *) source;
         rbuf = target;
         stype =  &ompi_mpi_char.dt;
@@ -184,7 +184,7 @@ 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) {
+    if ((INT_MAX < count) || !nlong) {
         MPI_COLL_VERBOSE(20,"RUNNING FALLBACK REDUCE");
         PREVIOUS_SCOLL_FN(mpi_module, reduce, group,
                 op,

oshmem/runtime/runtime.h

@@ -200,6 +200,13 @@ OSHMEM_DECLSPEC int oshmem_shmem_register_params(void);
         RUNTIME_CHECK_ERROR("Required address %p is not in symmetric space\n", ((void*)x));    \
         oshmem_shmem_abort(-1);                                                         \
     }
+/* Check if address is in symmetric space or size is zero */
+#define RUNTIME_CHECK_ADDR_SIZE(x,s)    \
+    if (OPAL_UNLIKELY((s) && !MCA_MEMHEAP_CALL(is_symmetric_addr((x)))))        \
+    {                                                                                   \
+        RUNTIME_CHECK_ERROR("Required address %p is not in symmetric space\n", ((void*)x));    \
+        oshmem_shmem_abort(-1);                                                         \
+    }
 #define RUNTIME_CHECK_WITH_MEMHEAP_SIZE(x)    \
     if (OPAL_UNLIKELY((long)(x) > MCA_MEMHEAP_CALL(size)))        \
     {                                                                                   \
@@ -212,6 +219,7 @@ OSHMEM_DECLSPEC int oshmem_shmem_register_params(void);
 #define RUNTIME_CHECK_INIT()
 #define RUNTIME_CHECK_PE(x)
 #define RUNTIME_CHECK_ADDR(x)
+#define RUNTIME_CHECK_ADDR_SIZE(x,s)
 #define RUNTIME_CHECK_WITH_MEMHEAP_SIZE(x)
 
 #endif  /* OSHMEM_PARAM_CHECK */

oshmem/shmem/c/shmem_alltoall.c

@@ -30,7 +30,7 @@ static void _shmem_alltoall(void *target,
                             int PE_size,
                             long *pSync);
 
-#define SHMEM_TYPE_ALLTOALL(name, element_size)     \
+#define SHMEM_TYPE_ALLTOALL(name, element_size)                      \
     void shmem##name(void *target,                                   \
                      const void *source,                             \
                      size_t nelems,                                  \
@@ -40,15 +40,15 @@ static void _shmem_alltoall(void *target,
                      long *pSync)                                    \
 {                                                                    \
     RUNTIME_CHECK_INIT();                                            \
-    RUNTIME_CHECK_ADDR(target);                                      \
-    RUNTIME_CHECK_ADDR(source);                                      \
+    RUNTIME_CHECK_ADDR_SIZE(target, nelems);                         \
+    RUNTIME_CHECK_ADDR_SIZE(source, nelems);                         \
                                                                      \
     _shmem_alltoall(target, source, 1, 1, nelems, element_size,      \
                        PE_start, logPE_stride, PE_size,              \
                        pSync);                                       \
 }
 
-#define SHMEM_TYPE_ALLTOALLS(name, element_size)     \
+#define SHMEM_TYPE_ALLTOALLS(name, element_size)                     \
     void shmem##name(void *target,                                   \
                      const void *source,                             \
                      ptrdiff_t dst, ptrdiff_t sst,                   \
@@ -59,8 +59,8 @@ static void _shmem_alltoall(void *target,
                      long *pSync)                                    \
 {                                                                    \
     RUNTIME_CHECK_INIT();                                            \
-    RUNTIME_CHECK_ADDR(target);                                      \
-    RUNTIME_CHECK_ADDR(source);                                      \
+    RUNTIME_CHECK_ADDR_SIZE(target, nelems);                         \
+    RUNTIME_CHECK_ADDR_SIZE(source, nelems);                         \
                                                                      \
     _shmem_alltoall(target, source, dst, sst, nelems, element_size,  \
                        PE_start, logPE_stride, PE_size,              \

oshmem/shmem/c/shmem_broadcast.c

@@ -29,7 +29,7 @@ static void _shmem_broadcast(void *target,
                               int PE_size,
                               long *pSync);
 
-#define SHMEM_TYPE_BROADCAST(name, element_size)     \
+#define SHMEM_TYPE_BROADCAST(name, element_size)                    \
     void shmem##name( void *target,                                 \
                       const void *source,                           \
                       size_t nelems,                                \
@@ -40,10 +40,10 @@ static void _shmem_broadcast(void *target,
                       long *pSync)                                  \
 {                                                                   \
     RUNTIME_CHECK_INIT();                                           \
-    RUNTIME_CHECK_ADDR(target);                                     \
-    RUNTIME_CHECK_ADDR(source);                                     \
+    RUNTIME_CHECK_ADDR_SIZE(target, nelems);                        \
+    RUNTIME_CHECK_ADDR_SIZE(source, nelems);                        \
                                                                     \
-    _shmem_broadcast( target, source, nelems * element_size,       \
+    _shmem_broadcast( target, source, nelems * element_size,        \
                        PE_root, PE_start, logPE_stride, PE_size,    \
                        pSync);                                      \
 }

oshmem/shmem/c/shmem_collect.c

@@ -39,10 +39,10 @@ static void _shmem_collect(void *target,
                       long *pSync)                              \
 {                                                               \
     RUNTIME_CHECK_INIT();                                       \
-    RUNTIME_CHECK_ADDR(target);                                 \
-    RUNTIME_CHECK_ADDR(source);                                 \
+    RUNTIME_CHECK_ADDR_SIZE(target, nelems);                    \
+    RUNTIME_CHECK_ADDR_SIZE(source, nelems);                    \
                                                                 \
-    _shmem_collect( target, source, nelems * element_size,     \
+    _shmem_collect( target, source, nelems * element_size,      \
                      PE_start, logPE_stride, PE_size,           \
                      pSync,                                     \
                      nelems_type);                              \

oshmem/shmem/c/shmem_reduce.c

@@ -26,8 +26,8 @@
  * object of every PE in the active set. The active set of PEs is defined by the triple PE_start,
  * logPE_stride and PE_size.
  */
-#define SHMEM_TYPE_REDUCE_OP(name, type_name, type, prefix)    \
-    void prefix##type_name##_##name##_to_all( type *target,                     \
+#define SHMEM_TYPE_REDUCE_OP(name, type_name, type, prefix)                                 \
+    void prefix##type_name##_##name##_to_all( type *target,                                 \
                                         const type *source,                                 \
                                         int nreduce,                                        \
                                         int PE_start,                                       \
@@ -40,8 +40,8 @@
     oshmem_group_t*  group = NULL;                                                          \
                                                                                             \
     RUNTIME_CHECK_INIT();                                                                   \
-    RUNTIME_CHECK_ADDR(target);                                                             \
-    RUNTIME_CHECK_ADDR(source);                                                             \
+    RUNTIME_CHECK_ADDR_SIZE(target, nreduce);                                               \
+    RUNTIME_CHECK_ADDR_SIZE(source, nreduce);                                               \
                                                                                             \
     {                                                                                       \
         group = oshmem_proc_group_create_nofail(PE_start, 1<<logPE_stride, PE_size);        \