Teams API topic branch

README

@@ -259,6 +259,15 @@ options.
         compute node is determined by its unique hostname, and the number of
         STXs available on a compute node is provided by the libfabric library.
 
+  Team Environment variables:
+
+    SHMEM_TEAMS_MAX (default: 10)
+        Sets the maximum number of available teams per PE, including the
+        predefined teams.  The maximum supported value is 64.
+
+    SHMEM_TEAM_SHARED_ONLY_SELF (default: off)
+        If defined, the predefined team, SHMEM_TEAM_SHARED, will only include
+        the self PE.
 
   Debugging Environment variables:
 

bindings/shmem_bind_c.m4

@@ -96,6 +96,45 @@ $1(int,      int,           `SHM_INTERNAL_INT',           `$2', `$3')
 $1(long,     long,          `SHM_INTERNAL_LONG',          `$2', `$3')
 $1(longlong, long long,     `SHM_INTERNAL_LONG_LONG',     `$2', `$3')')dnl
 dnl
+define(`SHMEM_BIND_C_COLL_AND_OR_XOR', dnl args: macro_name, op_name, op_const
+`$1(uchar,    unsigned char,      `SHM_INTERNAL_UCHAR',      `$2', `$3')
+$1(short,     short,              `SHM_INTERNAL_SHORT',      `$2', `$3')
+$1(ushort,    unsigned short,     `SHM_INTERNAL_USHORT',     `$2', `$3')
+$1(int,       int,                `SHM_INTERNAL_INT',        `$2', `$3')
+$1(uint,      unsigned int,       `SHM_INTERNAL_UINT',       `$2', `$3')
+$1(long,      long,               `SHM_INTERNAL_LONG',       `$2', `$3')
+$1(ulong,     unsigned long,      `SHM_INTERNAL_ULONG',      `$2', `$3')
+$1(longlong,  long long,          `SHM_INTERNAL_LONG_LONG',  `$2', `$3')
+$1(ulonglong, unsigned long long, `SHM_INTERNAL_ULONG_LONG', `$2', `$3')')dnl
+dnl
+define(`SHMEM_BIND_C_COLL_MIN_MAX', dnl args: macro_name, op_name, op_const
+`$1(short,     short,              `SHM_INTERNAL_SHORT',       `$2', `$3')
+$1(ushort,     unsigned short,     `SHM_INTERNAL_USHORT',      `$2', `$3')
+$1(int,        int,                `SHM_INTERNAL_INT',         `$2', `$3')
+$1(uint,       unsigned int,       `SHM_INTERNAL_UINT',        `$2', `$3')
+$1(long,       long,               `SHM_INTERNAL_LONG',        `$2', `$3')
+$1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG',       `$2', `$3')
+$1(longlong,   long long,          `SHM_INTERNAL_LONG_LONG',   `$2', `$3')
+$1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG',  `$2', `$3')
+$1(float,      float,              `SHM_INTERNAL_FLOAT',       `$2', `$3')
+$1(double,     double,             `SHM_INTERNAL_DOUBLE',      `$2', `$3')
+$1(longdouble, long double,        `SHM_INTERNAL_LONG_DOUBLE', `$2', `$3')')dnl
+dnl
+define(`SHMEM_BIND_C_COLL_SUM_PROD', dnl args: macro_name, op_name, op_const
+`$1(short,     short,              `SHM_INTERNAL_SHORT',          `$2', `$3')
+$1(ushort,     unsigned short,     `SHM_INTERNAL_USHORT',         `$2', `$3')
+$1(int,        int,                `SHM_INTERNAL_INT',            `$2', `$3')
+$1(uint,       unsigned int,       `SHM_INTERNAL_UINT',           `$2', `$3')
+$1(long,       long,               `SHM_INTERNAL_LONG',           `$2', `$3')
+$1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG',          `$2', `$3')
+$1(longlong,   long long,          `SHM_INTERNAL_LONG_LONG',      `$2', `$3')
+$1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG',     `$2', `$3')
+$1(float,      float,              `SHM_INTERNAL_FLOAT',          `$2', `$3')
+$1(double,     double,             `SHM_INTERNAL_DOUBLE',         `$2', `$3')
+$1(longdouble, long double,        `SHM_INTERNAL_LONG_DOUBLE',    `$2', `$3')
+$1(complexd,   double _Complex,    `SHM_INTERNAL_DOUBLE_COMPLEX', `$2', `$3')
+$1(complexf,   float _Complex,     `SHM_INTERNAL_FLOAT_COMPLEX',  `$2', `$3')')dnl
+dnl
 define(`SHMEM_BIND_C_COLL_FLOATS', dnl args: macro_name, op_name, op_const
 `$1(float,    float,         `SHM_INTERNAL_FLOAT',         `$2', `$3')
 $1(double,   double,        `SHM_INTERNAL_DOUBLE',        `$2', `$3')

bindings/shmem_bind_c11.m4

@@ -65,3 +65,43 @@ $1(ushort,     unsigned short)$2
 $1(uint,       unsigned int)$2
 $1(ulong,      unsigned long)$2
 $1(ulonglong,  unsigned long long)')dnl
+dnl
+define(`SHMEM_BIND_C11_COLL_AND_OR_XOR', dnl args: macro_name, op_name, op_const
+`$1(uchar,    unsigned char,      `SHM_INTERNAL_UCHAR')$2
+$1(short,     short,              `SHM_INTERNAL_SHORT')$2
+$1(ushort,    unsigned short,     `SHM_INTERNAL_USHORT')$2
+$1(int,       int,                `SHM_INTERNAL_INT')$2
+$1(uint,      unsigned int,       `SHM_INTERNAL_UINT')$2
+$1(long,      long,               `SHM_INTERNAL_LONG')$2
+$1(ulong,     unsigned long,      `SHM_INTERNAL_ULONG')$2
+$1(longlong,  long long,          `SHM_INTERNAL_LONG_LONG')$2
+$1(ulonglong, unsigned long long, `SHM_INTERNAL_ULONG_LONG')')dnl
+dnl
+define(`SHMEM_BIND_C11_COLL_MIN_MAX', dnl args: macro_name, op_name, op_const
+`$1(short,     short,              `SHM_INTERNAL_SHORT')$2
+$1(ushort,     unsigned short,     `SHM_INTERNAL_USHORT')$2
+$1(int,        int,                `SHM_INTERNAL_INT')$2
+$1(uint,       unsigned int,       `SHM_INTERNAL_UINT')$2
+$1(long,       long,               `SHM_INTERNAL_LONG')$2
+$1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG')$2
+$1(longlong,   long long,          `SHM_INTERNAL_LONG_LONG')$2
+$1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG')$2
+$1(float,      float,              `SHM_INTERNAL_FLOAT')$2
+$1(double,     double,             `SHM_INTERNAL_DOUBLE')$2
+$1(longdouble, long double,        `SHM_INTERNAL_LONG_DOUBLE')')dnl
+dnl
+define(`SHMEM_BIND_C11_COLL_SUM_PROD', dnl args: macro_name, op_name, op_const
+`$1(short,     short,              `SHM_INTERNAL_SHORT')$2
+$1(ushort,     unsigned short,     `SHM_INTERNAL_USHORT')$2
+$1(int,        int,                `SHM_INTERNAL_INT')$2
+$1(uint,       unsigned int,       `SHM_INTERNAL_UINT')$2
+$1(long,       long,               `SHM_INTERNAL_LONG')$2
+$1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG')$2
+$1(longlong,   long long,          `SHM_INTERNAL_LONG_LONG')$2
+$1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG')$2
+$1(float,      float,              `SHM_INTERNAL_FLOAT')$2
+$1(double,     double,             `SHM_INTERNAL_DOUBLE')$2
+$1(longdouble, long double,        `SHM_INTERNAL_LONG_DOUBLE')$2
+$1(complexd,   double _Complex,    `SHM_INTERNAL_DOUBLE_COMPLEX')$2
+$1(complexf,   float _Complex,     `SHM_INTERNAL_FLOAT_COMPLEX')')dnl
+dnl

bindings/shmem_bind_cxx.m4

@@ -55,7 +55,7 @@ $1(uint,       unsigned int,       `SHM_INTERNAL_UINT')$2
 $1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG')$2
 $1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG')')dnl
 dnl
-define(`SHMEM_BIND_CXX_SYNC', dnl args: macro_name
+define(`SHMEM_BIND_CXX_SYNC', dnl args: macro_name, end
 SHMEM_BIND_CXX_SYNC_EXTRAS($1,$2)dnl
 `$1(short,      short)$2
 $1(int,        int)$2
@@ -66,8 +66,50 @@ $1(uint,       unsigned int)$2
 $1(ulong,      unsigned long)$2
 $1(ulonglong,  unsigned long long)')dnl
 dnl
+define(`SHMEM_BIND_CXX_COLL_AND_OR_XOR', dnl args: macro_name, end
+`$1(uchar,    unsigned char,      `SHM_INTERNAL_UCHAR')$2
+$1(short,     short,              `SHM_INTERNAL_SHORT')$2
+$1(ushort,    unsigned short,     `SHM_INTERNAL_USHORT')$2
+$1(int,       int,                `SHM_INTERNAL_INT')$2
+$1(uint,      unsigned int,       `SHM_INTERNAL_UINT')$2
+$1(long,      long,               `SHM_INTERNAL_LONG')$2
+$1(ulong,     unsigned long,      `SHM_INTERNAL_ULONG')$2
+$1(longlong,  long long,          `SHM_INTERNAL_LONG_LONG')$2
+$1(ulonglong, unsigned long long, `SHM_INTERNAL_ULONG_LONG')')dnl
+dnl
+define(`SHMEM_BIND_CXX_COLL_MIN_MAX', dnl args: macro_name, op_name, op_const
+`$1(short,     short,              `SHM_INTERNAL_SHORT')$2
+$1(ushort,     unsigned short,     `SHM_INTERNAL_USHORT')$2
+$1(int,        int,                `SHM_INTERNAL_INT')$2
+$1(uint,       unsigned int,       `SHM_INTERNAL_UINT')$2
+$1(long,       long,               `SHM_INTERNAL_LONG')$2
+$1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG')$2
+$1(longlong,   long long,          `SHM_INTERNAL_LONG_LONG')$2
+$1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG')$2
+$1(float,      float,              `SHM_INTERNAL_FLOAT')$2
+$1(double,     double,             `SHM_INTERNAL_DOUBLE')$2
+$1(longdouble, long double,        `SHM_INTERNAL_LONG_DOUBLE')')dnl
+dnl
+define(`SHMEM_BIND_CXX_COLL_SUM_PROD', dnl args: macro_name, op_name, op_const
+`$1(short,     short,              `SHM_INTERNAL_SHORT')$2
+$1(ushort,     unsigned short,     `SHM_INTERNAL_USHORT')$2
+$1(int,        int,                `SHM_INTERNAL_INT')$2
+$1(uint,       unsigned int,       `SHM_INTERNAL_UINT')$2
+$1(long,       long,               `SHM_INTERNAL_LONG')$2
+$1(ulong,      unsigned long,      `SHM_INTERNAL_ULONG')$2
+$1(longlong,   long long,          `SHM_INTERNAL_LONG_LONG')$2
+$1(ulonglong,  unsigned long long, `SHM_INTERNAL_ULONG_LONG')$2
+$1(float,      float,              `SHM_INTERNAL_FLOAT')$2
+$1(double,     double,             `SHM_INTERNAL_DOUBLE')$2
+$1(longdouble, long double,        `SHM_INTERNAL_LONG_DOUBLE')$2
+$1(complexd,   double _Complex,    `SHM_INTERNAL_DOUBLE_COMPLEX')$2
+$1(complexf,   float _Complex,     `SHM_INTERNAL_FLOAT_COMPLEX')')dnl
+dnl
 define(`SHMEM_CXX_DEFINE_FOR_RMA', `SHMEM_BIND_CXX_RMA(`$1')')dnl
 define(`SHMEM_CXX_DEFINE_FOR_AMO', `SHMEM_BIND_CXX_AMO(`$1')')dnl
 define(`SHMEM_CXX_DEFINE_FOR_EXTENDED_AMO', `SHMEM_BIND_CXX_EXTENDED_AMO(`$1')')dnl
 define(`SHMEM_CXX_DEFINE_FOR_BITWISE_AMO', `SHMEM_BIND_CXX_BITWISE_AMO(`$1')')dnl
 define(`SHMEM_CXX_DEFINE_FOR_SYNC', `SHMEM_BIND_CXX_SYNC(`$1')')dnl
+define(`SHMEM_CXX_DEFINE_FOR_COLL_AND_OR_XOR', `SHMEM_BIND_CXX_COLL_AND_OR_XOR(`$1')')dnl
+define(`SHMEM_CXX_DEFINE_FOR_COLL_MIN_MAX', `SHMEM_BIND_CXX_COLL_MIN_MAX(`$1')')dnl
+define(`SHMEM_CXX_DEFINE_FOR_COLL_SUM_PROD', `SHMEM_BIND_CXX_COLL_SUM_PROD(`$1')')dnl

configure.ac

@@ -482,6 +482,8 @@ SHMEM_FIND_INT_TYPE([long], [$transport])
 SHMEM_FIND_INT_TYPE([long long], [$transport])
 SHMEM_FIND_INT_TYPE([ptrdiff_t], [$transport])
 
+SHMEM_FIND_UINT_TYPE([unsigned char], [$transport])
+SHMEM_FIND_UINT_TYPE([unsigned short], [$transport])
 SHMEM_FIND_UINT_TYPE([unsigned int], [$transport])
 SHMEM_FIND_UINT_TYPE([unsigned long], [$transport])
 SHMEM_FIND_UINT_TYPE([unsigned long long], [$transport])

mpp/shmemx-def.h

@@ -22,8 +22,29 @@ typedef struct {
     uint64_t target;
 } shmemx_pcntr_t;
 
+/* Teams */
+typedef struct shmem_impl_team_t {
+    int dummy;
+} * shmemx_team_t;
+
+typedef struct {
+    int num_contexts;
+} shmemx_team_config_t;
+
+#if SHMEM_HAVE_ATTRIBUTE_VISIBILITY == 1
+    __attribute__((visibility("default"))) extern shmemx_team_t SHMEMX_TEAM_WORLD;
+    __attribute__((visibility("default"))) extern shmemx_team_t SHMEMX_TEAM_SHARED;
+#else
+    extern shmemx_team_t SHMEMX_TEAM_WORLD;
+    extern shmemx_team_t SHMEMX_TEAM_SHARED;
+#endif
+
+#define SHMEMX_TEAM_INVALID NULL
+
 #define SHMEMX_CTX_INVALID NULL
 
+#define SHMEMX_TEAM_NUM_CONTEXTS       (1l<<0)
+
 #ifdef __cplusplus
 }
 #endif

mpp/shmemx.h4

@@ -175,6 +175,77 @@ static inline void shmemx_put_signal_nbi(shmem_ctx_t ctx, $2* dest, const $2* so
 }')dnl
 SHMEM_CXX_DEFINE_FOR_RMA(`SHMEM_CXX_PUT_SIGNAL_NBI')
 
+/* Team Collective Routines */
+define(`SHMEM_CXX_BCAST',
+`static inline int shmemx_broadcast(shmemx_team_t team, $2* dest, const $2* source,
+                                   size_t nelems, int PE_root) {
+    return shmemx_$1_broadcast(team, dest, source, nelems, PE_root);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_RMA(`SHMEM_CXX_BCAST')
+
+define(`SHMEM_CXX_COLLECT',
+`static inline int shmemx_collect(shmemx_team_t team, $2* dest, const $2* source,
+                                 size_t nelems) {
+    return shmemx_$1_collect(team, dest, source, nelems);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_RMA(`SHMEM_CXX_COLLECT')
+
+define(`SHMEM_CXX_FCOLLECT',
+`static inline int shmemx_fcollect(shmemx_team_t team, $2* dest, const $2* source,
+                                  size_t nelems) {
+    return shmemx_$1_fcollect(team, dest, source, nelems);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_RMA(`SHMEM_CXX_FCOLLECT')
+
+define(`SHMEM_CXX_AND_REDUCE',
+`static inline int shmemx_and_reduce(shmemx_team_t team, $2* dest, const $2* source,
+                                    size_t nreduce) {
+    return shmemx_$1_and_reduce(team, dest, source, nreduce);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_COLL_AND_OR_XOR(`SHMEM_CXX_AND_REDUCE')
+
+define(`SHMEM_CXX_MIN_REDUCE',
+`static inline int shmemx_min_reduce(shmemx_team_t team, $2* dest, const $2* source,
+                                    size_t nreduce) {
+    return shmemx_$1_min_reduce(team, dest, source, nreduce);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_COLL_MIN_MAX(`SHMEM_CXX_MIN_REDUCE')
+
+define(`SHMEM_CXX_MAX_REDUCE',
+`static inline int shmemx_max_reduce(shmemx_team_t team, $2* dest, const $2* source,
+                                    size_t nreduce) {
+    return shmemx_$1_max_reduce(team, dest, source, nreduce);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_COLL_MIN_MAX(`SHMEM_CXX_MAX_REDUCE')
+
+define(`SHMEM_CXX_SUM_REDUCE',
+`static inline int shmemx_sum_reduce(shmemx_team_t team, $2* dest, const $2* source,
+                                    size_t nreduce) {
+    return shmemx_$1_sum_reduce(team, dest, source, nreduce);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_COLL_SUM_PROD(`SHMEM_CXX_SUM_REDUCE')
+
+define(`SHMEM_CXX_PROD_REDUCE',
+`static inline int shmemx_prod_reduce(shmemx_team_t team, $2* dest, const $2* source,
+                                     size_t nreduce) {
+    return shmemx_$1_prod_reduce(team, dest, source, nreduce);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_COLL_SUM_PROD(`SHMEM_CXX_PROD_REDUCE')
+
+define(`SHMEM_CXX_ALLTOALL',
+`static inline int shmemx_alltoall(shmemx_team_t team, $2* dest, const $2* source,
+                                  size_t nelems) {
+    return shmemx_$1_alltoall(team, dest, source, nelems);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_RMA(`SHMEM_CXX_ALLTOALL')
+
+define(`SHMEM_CXX_ALLTOALLS',
+`static inline int shmemx_alltoalls(shmemx_team_t team, $2* dest, const $2* source,
+                                   ptrdiff_t dst, ptrdiff_t sst, size_t nelems) {
+    return shmemx_$1_alltoalls(team, dest, source, dst, sst, nelems);
+}')dnl
+SHMEM_CXX_DEFINE_FOR_RMA(`SHMEM_CXX_ALLTOALLS')
+
 /* C11 Generic Macros */
 #elif (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(SHMEM_INTERNAL_INCLUDE))
 
@@ -348,6 +419,79 @@ SHMEM_BIND_C11_RMA(`SHMEM_CTX_C11_GEN_PUT_SIGNAL_NBI', `, \') \
 SHMEM_BIND_C11_RMA(`SHMEM_C11_GEN_PUT_SIGNAL_NBI', `, \') \
     )(__VA_ARGS__)
 
+/* Team Collective Routines */
+define(`SHMEM_C11_GEN_BCAST', `        $2*: shmemx_$1_broadcast')dnl
+#define shmemx_broadcast(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_RMA(`SHMEM_C11_GEN_BCAST', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_COLLECT', `        $2*: shmemx_$1_collect')dnl
+#define shmemx_collect(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_RMA(`SHMEM_C11_GEN_COLLECT', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_FCOLLECT', `        $2*: shmemx_$1_fcollect')dnl
+#define shmemx_fcollect(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_RMA(`SHMEM_C11_GEN_FCOLLECT', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_AND_REDUCE', `        $2*: shmemx_$1_and_reduce')dnl
+#define shmemx_and_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_AND_OR_XOR(`SHMEM_C11_GEN_AND_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_OR_REDUCE', `        $2*: shmemx_$1_or_reduce')dnl
+#define shmemx_or_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_AND_OR_XOR(`SHMEM_C11_GEN_OR_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_XOR_REDUCE', `        $2*: shmemx_$1_xor_reduce')dnl
+#define shmemx_xor_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_AND_OR_XOR(`SHMEM_C11_GEN_XOR_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_MIN_REDUCE', `        $2*: shmemx_$1_min_reduce')dnl
+#define shmemx_min_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_MIN_MAX(`SHMEM_C11_GEN_MIN_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_MAX_REDUCE', `        $2*: shmemx_$1_max_reduce')dnl
+#define shmemx_max_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_MIN_MAX(`SHMEM_C11_GEN_MAX_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_SUM_REDUCE', `        $2*: shmemx_$1_sum_reduce')dnl
+#define shmemx_sum_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_SUM_PROD(`SHMEM_C11_GEN_SUM_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_PROD_REDUCE', `        $2*: shmemx_$1_prod_reduce')dnl
+#define shmemx_prod_reduce(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_COLL_SUM_PROD(`SHMEM_C11_GEN_PROD_REDUCE', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_ALLTOALL', `        $2*: shmemx_$1_alltoall')dnl
+#define shmemx_alltoall(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_RMA(`SHMEM_C11_GEN_ALLTOALL', `, \') \
+    )(__VA_ARGS__)
+
+define(`SHMEM_C11_GEN_ALLTOALLS', `        $2*: shmemx_$1_alltoalls')dnl
+#define shmemx_alltoalls(...) \
+    _Generic(SHMEM_C11_TYPE_EVAL_PTR(SHMEM_C11_ARG1(__VA_ARGS__)), \
+SHMEM_BIND_C11_RMA(`SHMEM_C11_GEN_ALLTOALLS', `, \') \
+    )(__VA_ARGS__)
+
 #endif /* C11 */
 
 #endif /* SHMEMX_H */