oshmem: Align OSHMEM API with spec v1.3 (Add scoll/alltoall interface)

oshmem/include/shmem.h.in

@@ -101,13 +101,15 @@ enum shmem_wait_ops {
 #define _SHMEM_BCAST_SYNC_SIZE          (1 + _SHMEM_BARRIER_SYNC_SIZE)
 #define _SHMEM_COLLECT_SYNC_SIZE        (1 + _SHMEM_BCAST_SYNC_SIZE)
 #define _SHMEM_REDUCE_SYNC_SIZE         (1 + _SHMEM_BCAST_SYNC_SIZE)
+#define _SHMEM_ALLTOALL_SYNC_SIZE       (1)
 #define _SHMEM_REDUCE_MIN_WRKDATA_SIZE  (1)
 #define _SHMEM_SYNC_VALUE               (-1)
 
 #define SHMEM_BARRIER_SYNC_SIZE        _SHMEM_BARRIER_SYNC_SIZE
 #define SHMEM_BCAST_SYNC_SIZE          _SHMEM_BCAST_SYNC_SIZE
 #define SHMEM_COLLECT_SYNC_SIZE        _SHMEM_COLLECT_SYNC_SIZE
 #define SHMEM_REDUCE_SYNC_SIZE         _SHMEM_REDUCE_SYNC_SIZE
+#define SHMEM_ALLTOALL_SYNC_SIZE       _SHMEM_ALLTOALL_SYNC_SIZE
 #define SHMEM_REDUCE_MIN_WRKDATA_SIZE  _SHMEM_REDUCE_MIN_WRKDATA_SIZE
 #define SHMEM_SYNC_VALUE               _SHMEM_SYNC_VALUE
 
@@ -191,6 +193,24 @@ OSHMEM_DECLSPEC void shmem_iput32(void* target, const void* source, ptrdiff_t ts
 OSHMEM_DECLSPEC void shmem_iput64(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 OSHMEM_DECLSPEC void shmem_iput128(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 
+/*
+ * Nonblocking put routines
+ */
+OSHMEM_DECLSPEC  void shmem_putmem_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_char_put_nbi(char *target, const char *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_short_put_nbi(short *target, const short *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_int_put_nbi(int* target, const int* source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_long_put_nbi(long *target, const long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_longlong_put_nbi(long long *target, const long long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_float_put_nbi(float *target, const float *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_double_put_nbi(double *target, const double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_longdouble_put_nbi(long double *target, const long double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_put8_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_put16_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_put32_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_put64_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_put128_nbi(void *target, const void *source, size_t len, int pe);
+
 /*
  * Elemental get routines
  */
@@ -232,6 +252,24 @@ OSHMEM_DECLSPEC void shmem_iget32(void* target, const void* source, ptrdiff_t ts
 OSHMEM_DECLSPEC void shmem_iget64(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 OSHMEM_DECLSPEC void shmem_iget128(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 
+/*
+ * Nonblocking data get routines
+ */
+OSHMEM_DECLSPEC  void shmem_getmem_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_char_get_nbi(char *target, const char *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_short_get_nbi(short *target, const short *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_int_get_nbi(int *target, const int *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_long_get_nbi(long *target, const long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_longlong_get_nbi(long long *target, const long long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_float_get_nbi(float *target, const float *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_double_get_nbi(double *target, const double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_longdouble_get_nbi(long double *target, const long double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_get8_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_get16_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_get32_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_get64_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void shmem_get128_nbi(void *target, const void *source, size_t len, int pe);
+
 /*
  * Atomic operations
  */

oshmem/mca/scoll/base/scoll_base_frame.c

@@ -46,6 +46,7 @@ static void scoll_base_module_construct(mca_scoll_base_module_t *m)
     m->scoll_broadcast = NULL;
     m->scoll_collect = NULL;
     m->scoll_reduce = NULL;
+    m->scoll_alltoall = NULL;
     m->scoll_module_enable = NULL;
 }
 

oshmem/mca/scoll/base/scoll_base_select.c

@@ -118,6 +118,22 @@ static int scoll_null_reduce(struct oshmem_group_t *group,
     return OSHMEM_SUCCESS;
 }
 
+static int scoll_null_alltoall(struct oshmem_group_t *group,
+                              void *target,
+                              const void *source,
+                              ptrdiff_t dst, ptrdiff_t sst,
+                              size_t nlong,
+                              long *pSync,
+                              int alg)
+{
+    if (oshmem_proc_group_is_member(group)) {
+        SCOLL_ERROR("internal error");
+        oshmem_shmem_abort(-1);
+        return OSHMEM_ERROR;
+    }
+    return OSHMEM_SUCCESS;
+}
+
 /*
  * Stuff for the OBJ interface
  */
@@ -160,6 +176,7 @@ int mca_scoll_base_group_unselect(struct oshmem_group_t * group)
     CLOSE(group, broadcast);
     CLOSE(group, collect);
     CLOSE(group, reduce);
+    CLOSE(group, alltoall);
 
     /* All done */
     return OSHMEM_SUCCESS;
@@ -184,6 +201,7 @@ int mca_scoll_base_select(struct oshmem_group_t *group)
         group->g_scoll.scoll_broadcast = scoll_null_broadcast;
         group->g_scoll.scoll_collect = scoll_null_collect;
         group->g_scoll.scoll_reduce = scoll_null_reduce;
+        group->g_scoll.scoll_alltoall = scoll_null_alltoall;
         return OSHMEM_SUCCESS;
     }
     SCOLL_VERBOSE(10,
@@ -206,10 +224,11 @@ int mca_scoll_base_select(struct oshmem_group_t *group)
         if (OSHMEM_SUCCESS != ret) {
             mca_scoll_base_group_unselect(group);
         } else {
+            COPY(avail->ac_module, group, barrier);
             COPY(avail->ac_module, group, broadcast);
             COPY(avail->ac_module, group, collect);
             COPY(avail->ac_module, group, reduce);
-            COPY(avail->ac_module, group, barrier);
+            COPY(avail->ac_module, group, alltoall);
         }
         OBJ_RELEASE(avail->ac_module);
         OBJ_RELEASE(avail);
@@ -220,16 +239,17 @@ int mca_scoll_base_select(struct oshmem_group_t *group)
     if ((NULL == group->g_scoll.scoll_barrier)
             || (NULL == group->g_scoll.scoll_broadcast)
             || (NULL == group->g_scoll.scoll_collect)
-            || (NULL == group->g_scoll.scoll_reduce)) {
+            || (NULL == group->g_scoll.scoll_reduce)
+            || (NULL == group->g_scoll.scoll_alltoall)) {
         mca_scoll_base_group_unselect(group);
         return OSHMEM_ERR_NOT_FOUND;
     }
 
     return OSHMEM_SUCCESS;
 }
 
-static int avail_coll_compare (opal_list_item_t **a,
-                               opal_list_item_t **b)
+static int avail_coll_compare(opal_list_item_t **a,
+                              opal_list_item_t **b)
 {
     avail_com_t *acom = (avail_com_t *) *a;
     avail_com_t *bcom = (avail_com_t *) *b;

oshmem/mca/scoll/basic/Makefile.am

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2013      Mellanox Technologies, Inc.
+# Copyright (c) 2013-2016 Mellanox Technologies, Inc.
 #                         All rights reserved.
 # $COPYRIGHT$
 #
@@ -15,7 +15,8 @@ sources = \
 	scoll_basic_barrier.c \
 	scoll_basic_broadcast.c \
 	scoll_basic_collect.c \
-	scoll_basic_reduce.c
+	scoll_basic_reduce.c \
+	scoll_basic_alltoall.c
 
 
 # Make the output library in this directory, and name it either

oshmem/mca/scoll/basic/scoll_basic.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2013      Mellanox Technologies, Inc.
+ * Copyright (c) 2013-2016 Mellanox Technologies, Inc.
  *                         All rights reserved.
  * $COPYRIGHT$
  *
@@ -77,6 +77,13 @@ int mca_scoll_basic_reduce(struct oshmem_group_t *group,
                            long *pSync,
                            void *pWrk,
                            int alg);
+int mca_scoll_basic_alltoall(struct oshmem_group_t *group,
+                             void *target,
+                             const void *source,
+                             ptrdiff_t dst, ptrdiff_t sst,
+                             size_t nlong,
+                             long *pSync,
+                             int alg);
 
 static inline unsigned int scoll_log2(unsigned long val)
 {

oshmem/mca/scoll/basic/scoll_basic_module.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2013      Mellanox Technologies, Inc.
+ * Copyright (c) 2013-2016 Mellanox Technologies, Inc.
  *                         All rights reserved.
  * $COPYRIGHT$
  *
@@ -52,9 +52,10 @@ mca_scoll_basic_query(struct oshmem_group_t *group, int *priority)
         module->super.scoll_broadcast = mca_scoll_basic_broadcast;
         module->super.scoll_collect = mca_scoll_basic_collect;
         module->super.scoll_reduce = mca_scoll_basic_reduce;
+        module->super.scoll_alltoall = mca_scoll_basic_alltoall;
         module->super.scoll_module_enable = mca_scoll_basic_enable;
         return &(module->super);
     }
 
-    return NULL ;
+    return NULL;
 }

oshmem/mca/scoll/fca/scoll_fca.h

@@ -93,6 +93,8 @@ struct mca_scoll_fca_module_t {
     mca_scoll_base_module_t *previous_collect_module;
     mca_scoll_base_module_reduce_fn_t previous_reduce;
     mca_scoll_base_module_t *previous_reduce_module;
+    mca_scoll_base_module_alltoall_fn_t previous_alltoall;
+    mca_scoll_base_module_t *previous_alltoall_module;
 };
 typedef struct mca_scoll_fca_module_t mca_scoll_fca_module_t;
 OBJ_CLASS_DECLARATION(mca_scoll_fca_module_t);

oshmem/mca/scoll/fca/scoll_fca_module.c

@@ -391,6 +391,7 @@ static int _save_coll_handlers(mca_scoll_fca_module_t *fca_module)
     FCA_SAVE_PREV_SCOLL_API(broadcast);
     FCA_SAVE_PREV_SCOLL_API(collect);
     FCA_SAVE_PREV_SCOLL_API(reduce);
+    FCA_SAVE_PREV_SCOLL_API(alltoall);
 
     return OSHMEM_SUCCESS;
 }
@@ -450,6 +451,7 @@ static void mca_scoll_fca_module_clear(mca_scoll_fca_module_t *fca_module)
     fca_module->previous_broadcast = NULL;
     fca_module->previous_collect = NULL;
     fca_module->previous_reduce = NULL;
+    fca_module->previous_alltoall = NULL;
 }
 
 static void mca_scoll_fca_module_construct(mca_scoll_fca_module_t *fca_module)
@@ -465,6 +467,7 @@ static void mca_scoll_fca_module_destruct(mca_scoll_fca_module_t *fca_module)
     OBJ_RELEASE(fca_module->previous_broadcast_module);
     OBJ_RELEASE(fca_module->previous_collect_module);
     OBJ_RELEASE(fca_module->previous_reduce_module);
+    OBJ_RELEASE(fca_module->previous_alltoall_module);
     if (fca_module->fca_comm)
         _destroy_fca_comm(fca_module);
     free(fca_module->local_ranks);
@@ -541,6 +544,7 @@ mca_scoll_fca_comm_query(struct oshmem_group_t *comm, int *priority)
     fca_module->super.scoll_broadcast =
             mca_scoll_fca_component.fca_enable_bcast ? mca_scoll_fca_broadcast :
                                                        NULL;
+    fca_module->super.scoll_alltoall = NULL;
 
     *priority = mca_scoll_fca_component.fca_priority;
     module = &fca_module->super;

oshmem/mca/scoll/mpi/scoll_mpi.h

@@ -69,6 +69,8 @@ struct mca_scoll_mpi_module_t {
     mca_scoll_base_module_t *previous_barrier_module;
     mca_scoll_base_module_collect_fn_t previous_collect;
     mca_scoll_base_module_t *previous_collect_module;
+    mca_scoll_base_module_alltoall_fn_t previous_alltoall;
+    mca_scoll_base_module_t *previous_alltoall_module;
 };
 typedef struct mca_scoll_mpi_module_t mca_scoll_mpi_module_t;
 

oshmem/mca/scoll/mpi/scoll_mpi_module.c

@@ -29,6 +29,7 @@ static void mca_scoll_mpi_module_clear(mca_scoll_mpi_module_t *mpi_module)
     mpi_module->previous_broadcast    = NULL;
     mpi_module->previous_reduce       = NULL;
     mpi_module->previous_collect      = NULL;
+    mpi_module->previous_alltoall     = NULL;
 }
 
 static void mca_scoll_mpi_module_construct(mca_scoll_mpi_module_t *mpi_module)
@@ -43,6 +44,7 @@ static void mca_scoll_mpi_module_destruct(mca_scoll_mpi_module_t *mpi_module)
     OBJ_RELEASE(mpi_module->previous_broadcast_module);
     OBJ_RELEASE(mpi_module->previous_reduce_module);
     OBJ_RELEASE(mpi_module->previous_collect_module);
+    OBJ_RELEASE(mpi_module->previous_alltoall_module);
 
     mca_scoll_mpi_module_clear(mpi_module);
     /* Free ompi_comm */
@@ -68,6 +70,7 @@ static int mca_scoll_mpi_save_coll_handlers(mca_scoll_base_module_t *module, osh
     MPI_SAVE_PREV_SCOLL_API(broadcast);
     MPI_SAVE_PREV_SCOLL_API(reduce);
     MPI_SAVE_PREV_SCOLL_API(collect);
+    MPI_SAVE_PREV_SCOLL_API(alltoall);
     return OSHMEM_SUCCESS;
 }
 
@@ -173,6 +176,7 @@ mca_scoll_mpi_comm_query(oshmem_group_t *osh_group, int *priority)
     mpi_module->super.scoll_broadcast = mca_scoll_mpi_broadcast;
     mpi_module->super.scoll_reduce = mca_scoll_mpi_reduce;
     mpi_module->super.scoll_collect = mca_scoll_mpi_collect;
+    mpi_module->super.scoll_alltoall = NULL;
 
     *priority = cm->mpi_priority;
     module = &mpi_module->super;

oshmem/mca/scoll/scoll.h

@@ -89,7 +89,6 @@ typedef struct mca_scoll_base_component_1_0_0_t mca_scoll_base_component_t;
 typedef int
 (*mca_scoll_base_module_enable_1_0_0_fn_t)(struct mca_scoll_base_module_1_0_0_t* module,
                                            struct oshmem_group_t *comm);
-typedef int (*mca_scoll_base_module_ft_event_fn_t)(int state);
 
 #define SCOLL_DEFAULT_ALG   (-1)
 
@@ -139,6 +138,13 @@ typedef int (*mca_scoll_base_module_reduce_fn_t)(struct oshmem_group_t *group,
                                                  long *pSync,
                                                  void *pWrk,
                                                  int alg);
+typedef int (*mca_scoll_base_module_alltoall_fn_t)(struct oshmem_group_t *group,
+                                                  void *target,
+                                                  const void *source,
+                                                  ptrdiff_t dst, ptrdiff_t sst,
+                                                  size_t nlong,
+                                                  long *pSync,
+                                                  int alg);
 
 struct mca_scoll_base_module_1_0_0_t {
     /** Collective modules all inherit from opal_object */
@@ -149,13 +155,14 @@ struct mca_scoll_base_module_1_0_0_t {
     mca_scoll_base_module_broadcast_fn_t scoll_broadcast;
     mca_scoll_base_module_collect_fn_t scoll_collect;
     mca_scoll_base_module_reduce_fn_t scoll_reduce;
+    mca_scoll_base_module_alltoall_fn_t scoll_alltoall;
     mca_scoll_base_module_enable_1_0_0_fn_t scoll_module_enable;
 };
 typedef struct mca_scoll_base_module_1_0_0_t mca_scoll_base_module_1_0_0_t;
 
-/** Per guidence in mca.h, use the unversioned struct name if you just
+/** Per guidance in mca.h, use the unversioned struct name if you just
  want to always keep up with the most recent version of the
- interace. */
+ interface. */
 typedef struct mca_scoll_base_module_1_0_0_t mca_scoll_base_module_t;
 OSHMEM_DECLSPEC OBJ_CLASS_DECLARATION(mca_scoll_base_module_t);
 
@@ -171,7 +178,7 @@ OSHMEM_DECLSPEC OBJ_CLASS_DECLARATION(mca_scoll_base_module_t);
 /*
  * Collectives group cache structure
  *
- * Collectives gorup cache structure, used to find functions to
+ * Collectives group cache structure, used to find functions to
  * implement collective algorithms and their associated modules.
  */
 struct mca_scoll_base_group_scoll_t {
@@ -183,6 +190,8 @@ struct mca_scoll_base_group_scoll_t {
     mca_scoll_base_module_1_0_0_t *scoll_collect_module;
     mca_scoll_base_module_reduce_fn_t scoll_reduce;
     mca_scoll_base_module_1_0_0_t *scoll_reduce_module;
+    mca_scoll_base_module_alltoall_fn_t scoll_alltoall;
+    mca_scoll_base_module_1_0_0_t *scoll_alltoall_module;
 };
 typedef struct mca_scoll_base_group_scoll_t mca_scoll_base_group_scoll_t;