oshmem: Align OSHMEM API with spec v1.3 (Add nonblocking put/get Fortran functions)

oshmem/shmem/fortran/Makefile.am

@@ -69,6 +69,7 @@ liboshmem_fortran_la_SOURCES += \
                     shmem_iput8_f.c \
                     shmem_logical_iput_f.c \
                     shmem_real_iput_f.c \
+                    shmem_put_nb_f.c \
                     shmem_character_get_f.c \
                     shmem_complex_get_f.c \
                     shmem_double_get_f.c \
@@ -91,6 +92,7 @@ liboshmem_fortran_la_SOURCES += \
                     shmem_integer_iget_f.c \
                     shmem_logical_iget_f.c \
                     shmem_real_iget_f.c \
+                    shmem_get_nb_f.c \
                     shmem_swap_f.c \
                     shmem_int4_swap_f.c \
                     shmem_int8_swap_f.c \

oshmem/shmem/fortran/profile/Makefile.am

@@ -61,6 +61,7 @@ nodist_liboshmem_fortran_pshmem_la_SOURCES = \
 	pshmem_iput8_f.c \
 	pshmem_logical_iput_f.c \
 	pshmem_real_iput_f.c \
+	pshmem_put_nb_f.c \
 	pshmem_character_get_f.c \
 	pshmem_complex_get_f.c \
 	pshmem_double_get_f.c \
@@ -83,6 +84,7 @@ nodist_liboshmem_fortran_pshmem_la_SOURCES = \
 	pshmem_integer_iget_f.c \
 	pshmem_logical_iget_f.c \
 	pshmem_real_iget_f.c \
+	pshmem_get_nb_f.c \
 	pshmem_swap_f.c \
 	pshmem_int4_swap_f.c \
 	pshmem_int8_swap_f.c \

oshmem/shmem/fortran/profile/defines.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2014      Mellanox Technologies, Inc.
+ * Copyright (c) 2014-2016 Mellanox Technologies, Inc.
  *                         All rights reserved.
  * $COPYRIGHT$
  *
@@ -118,6 +118,10 @@
 #define shmem_character_put_ pshmem_character_put_
 #define shmem_character_put__ pshmem_character_put__
 
+#define SHMEM_CHARACTER_PUT_NBI PSHMEM_CHARACTER_PUT_NBI
+#define shmem_character_put_nbi_ pshmem_character_put_nbi_
+#define shmem_character_put_nbi__ pshmem_character_put_nbi__
+
 #define SHMEM_COLLECT4 PSHMEM_COLLECT4
 #define shmem_collect4_ pshmem_collect4_
 #define shmem_collect4__ pshmem_collect4__
@@ -166,6 +170,10 @@
 #define shmem_complex_put_ pshmem_complex_put_
 #define shmem_complex_put__ pshmem_complex_put__
 
+#define SHMEM_COMPLEX_PUT_NBI PSHMEM_COMPLEX_PUT_NBI
+#define shmem_complex_put_nbi_ pshmem_complex_put_nbi_
+#define shmem_complex_put_nbi__ pshmem_complex_put_nbi__
+
 #define SHMEM_DOUBLE_GET PSHMEM_DOUBLE_GET
 #define shmem_double_get_ pshmem_double_get_
 #define shmem_double_get__ pshmem_double_get__
@@ -182,6 +190,10 @@
 #define shmem_double_put_ pshmem_double_put_
 #define shmem_double_put__ pshmem_double_put__
 
+#define SHMEM_DOUBLE_PUT_NBI PSHMEM_DOUBLE_PUT_NBI
+#define shmem_double_put_nbi_ pshmem_double_put_nbi_
+#define shmem_double_put_nbi__ pshmem_double_put_nbi__
+
 #define SHMEM_FENCE PSHMEM_FENCE
 #define shmem_fence_ pshmem_fence_
 #define shmem_fence__ pshmem_fence__
@@ -230,6 +242,102 @@
 #define shmem_iget8_ pshmem_iget8_
 #define shmem_iget8__ pshmem_iget8__
 
+#define SHMEM_GETMEM_NBI PSHMEM_GETMEM_NBI
+#define shmem_getmem_nbi_ pshmem_getmem_nbi_
+#define shmem_getmem_nbi__ pshmem_getmem_nbi__
+
+#define SHMEM_CHARACTER_GET_NBI PSHMEM_CHARACTER_GET_NBI
+#define shmem_character_get_nbi_ pshmem_character_get_nbi_
+#define shmem_character_get_nbi__ pshmem_character_get_nbi__
+
+#define SHMEM_COMPLEX_GET_NBI PSHMEM_COMPLEX_GET_NBI
+#define shmem_complex_get_nbi_ pshmem_complex_get_nbi_
+#define shmem_complex_get_nbi__ pshmem_complex_get_nbi__
+
+#define SHMEM_DOUBLE_GET_NBI PSHMEM_DOUBLE_GET_NBI
+#define shmem_double_get_nbi_ pshmem_double_get_nbi_
+#define shmem_double_get_nbi__ pshmem_double_get_nbi__
+
+#define SHMEM_INTEGER_GET_NBI PSHMEM_INTEGER_GET_NBI
+#define shmem_integer_get_nbi_ pshmem_integer_get_nbi_
+#define shmem_integer_get_nbi__ pshmem_integer_get_nbi__
+
+#define SHMEM_LOGICAL_GET_NBI PSHMEM_LOGICAL_GET_NBI
+#define shmem_logical_get_nbi_ pshmem_logical_get_nbi_
+#define shmem_logical_get_nbi__ pshmem_logical_get_nbi__
+
+#define SHMEM_REAL_GET_NBI PSHMEM_REAL_GET_NBI
+#define shmem_real_get_nbi_ pshmem_real_get_nbi_
+#define shmem_real_get_nbi__ pshmem_real_get_nbi__
+
+#define SHMEM_GET4_NBI PSHMEM_GET4_NBI
+#define shmem_get4_nbi_ pshmem_get4_nbi_
+#define shmem_get4_nbi__ pshmem_get4_nbi__
+
+#define SHMEM_GET8_NBI PSHMEM_GET8_NBI
+#define shmem_get8_nbi_ pshmem_get8_nbi_
+#define shmem_get8_nbi__ pshmem_get8_nbi__
+
+#define SHMEM_GET32_NBI PSHMEM_GET32_NBI
+#define shmem_get32_nbi_ pshmem_get32_nbi_
+#define shmem_get32_nbi__ pshmem_get32_nbi__
+
+#define SHMEM_GET64_NBI PSHMEM_GET64_NBI
+#define shmem_get64_nbi_ pshmem_get64_nbi_
+#define shmem_get64_nbi__ pshmem_get64_nbi__
+
+#define SHMEM_GET128_NBI PSHMEM_GET128_NBI
+#define shmem_get128_nbi_ pshmem_get128_nbi_
+#define shmem_get128_nbi__ pshmem_get128_nbi__
+
+#define SHMEM_PUTMEM_NBI PSHMEM_PUTMEM_NBI
+#define shmem_putmem_nbi_ pshmem_putmem_nbi_
+#define shmem_putmem_nbi__ pshmem_putmem_nbi__
+
+#define SHMEM_CHARACTER_PUT_NBI PSHMEM_CHARACTER_PUT_NBI
+#define shmem_character_put_nbi_ pshmem_character_put_nbi_
+#define shmem_character_put_nbi__ pshmem_character_put_nbi__
+
+#define SHMEM_COMPLEX_PUT_NBI PSHMEM_COMPLEX_PUT_NBI
+#define shmem_complex_put_nbi_ pshmem_complex_put_nbi_
+#define shmem_complex_put_nbi__ pshmem_complex_put_nbi__
+
+#define SHMEM_DOUBLE_PUT_NBI PSHMEM_DOUBLE_PUT_NBI
+#define shmem_double_put_nbi_ pshmem_double_put_nbi_
+#define shmem_double_put_nbi__ pshmem_double_put_nbi__
+
+#define SHMEM_INTEGER_PUT_NBI PSHMEM_INTEGER_PUT_NBI
+#define shmem_integer_put_nbi_ pshmem_integer_put_nbi_
+#define shmem_integer_put_nbi__ pshmem_integer_put_nbi__
+
+#define SHMEM_LOGICAL_PUT_NBI PSHMEM_LOGICAL_PUT_NBI
+#define shmem_logical_put_nbi_ pshmem_logical_put_nbi_
+#define shmem_logical_put_nbi__ pshmem_logical_put_nbi__
+
+#define SHMEM_REAL_PUT_NBI PSHMEM_REAL_PUT_NBI
+#define shmem_real_put_nbi_ pshmem_real_put_nbi_
+#define shmem_real_put_nbi__ pshmem_real_put_nbi__
+
+#define SHMEM_PUT4_NBI PSHMEM_PUT4_NBI
+#define shmem_put4_nbi_ pshmem_put4_nbi_
+#define shmem_put4_nbi__ pshmem_put4_nbi__
+
+#define SHMEM_PUT8_NBI PSHMEM_PUT8_NBI
+#define shmem_put8_nbi_ pshmem_put8_nbi_
+#define shmem_put8_nbi__ pshmem_put8_nbi__
+
+#define SHMEM_PUT32_NBI PSHMEM_PUT32_NBI
+#define shmem_put32_nbi_ pshmem_put32_nbi_
+#define shmem_put32_nbi__ pshmem_put32_nbi__
+
+#define SHMEM_PUT64_NBI PSHMEM_PUT64_NBI
+#define shmem_put64_nbi_ pshmem_put64_nbi_
+#define shmem_put64_nbi__ pshmem_put64_nbi__
+
+#define SHMEM_PUT128_NBI PSHMEM_PUT128_NBI
+#define shmem_put128_nbi_ pshmem_put128_nbi_
+#define shmem_put128_nbi__ pshmem_put128_nbi__
+
 #define SHMEM_INT4_ADD PSHMEM_INT4_ADD
 #define shmem_int4_add_ pshmem_int4_add_
 #define shmem_int4_add__ pshmem_int4_add__

oshmem/shmem/fortran/profile/prototypes_pshmem.h

@@ -36,6 +36,7 @@ PN (void, pshpclmove, PSHPCLMOVE, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MP
 PN (FORTRAN_POINTER_T*, pshmem_ptr, PSHMEM_PTR, (FORTRAN_POINTER_T target, MPI_Fint *pe));
 PN (ompi_fortran_logical_t, pshmem_pe_accessible, PSHMEM_PE_ACCESSIBLE, (MPI_Fint *pe));
 PN (MPI_Fint, pshmem_addr_accessible, PSHMEM_ADDR_ACCESSIBLE, (FORTRAN_POINTER_T addr, MPI_Fint *pe));
+
 PN (void, pshmem_put, PSHMEM_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, pshmem_character_put, PSHMEM_CHARACTER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, pshmem_complex_put, PSHMEM_COMPLEX_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
@@ -49,6 +50,7 @@ PN (void, pshmem_put32, PSHMEM_PUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER
 PN (void, pshmem_put64, PSHMEM_PUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, pshmem_put128, PSHMEM_PUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, pshmem_putmem, PSHMEM_PUTMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+
 PN (void, pshmem_iput4, PSHMEM_IPUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_iput8, PSHMEM_IPUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_iput32, PSHMEM_IPUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
@@ -59,6 +61,20 @@ PN (void, pshmem_double_iput, PSHMEM_DOUBLE_IPUT, (FORTRAN_POINTER_T target, FOR
 PN (void, pshmem_integer_iput, PSHMEM_INTEGER_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_logical_iput, PSHMEM_LOGICAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_real_iput, PSHMEM_REAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+
+PN (void, pshmem_putmem_nbi, PSHMEM_PUTMEM_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_character_put_nbi, PSHMEM_CHARACTER_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_complex_put_nbi, PSHMEM_COMPLEX_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_double_put_nbi, PSHMEM_DOUBLE_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_integer_put_nbi, PSHMEM_INTEGER_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_logical_put_nbi, PSHMEM_LOGICAL_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_real_put_nbi, PSHMEM_REAL_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put4_nbi, PSHMEM_PUT4_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put8_nbi, PSHMEM_PUT8_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put32_nbi, PSHMEM_PUT32_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put64_nbi, PSHMEM_PUT64_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put128_nbi, PSHMEM_PUT128_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+
 PN (void, pshmem_character_get, PSHMEM_CHARACTER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_complex_get, PSHMEM_COMPLEX_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_double_get, PSHMEM_DOUBLE_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
@@ -71,6 +87,7 @@ PN (void, pshmem_get128, PSHMEM_GET128, (FORTRAN_POINTER_T target, FORTRAN_POINT
 PN (void, pshmem_getmem, PSHMEM_GETMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_logical_get, PSHMEM_LOGICAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_real_get, PSHMEM_REAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+
 PN (void, pshmem_iget4, PSHMEM_IGET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_iget8, PSHMEM_IGET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_iget32, PSHMEM_IGET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
@@ -81,6 +98,20 @@ PN (void, pshmem_double_iget, PSHMEM_DOUBLE_IGET, (FORTRAN_POINTER_T target, FOR
 PN (void, pshmem_integer_iget, PSHMEM_INTEGER_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_logical_iget, PSHMEM_LOGICAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, pshmem_real_iget, PSHMEM_REAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+
+PN (void, pshmem_getmem_nbi, PSHMEM_GETMEM_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_character_get_nbi, PSHMEM_CHARACTER_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_complex_get_nbi, PSHMEM_COMPLEX_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_double_get_nbi, PSHMEM_DOUBLE_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_integer_get_nbi, PSHMEM_INTEGER_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_logical_get_nbi, PSHMEM_LOGICAL_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_real_get_nbi, PSHMEM_REAL_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get4_nbi, PSHMEM_GET4_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get8_nbi, PSHMEM_GET8_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get32_nbi, PSHMEM_GET32_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get64_nbi, PSHMEM_GET64_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get128_nbi, PSHMEM_GET128_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+
 PN (MPI_Fint, pshmem_swap, PSHMEM_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_integer4_t, pshmem_int4_swap, PSHMEM_INT4_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_integer8_t, pshmem_int8_swap, PSHMEM_INT8_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));

oshmem/shmem/fortran/prototypes_shmem.h

@@ -39,6 +39,7 @@ PN (void, shpclmove, SHPCLMOVE, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_
 PN (FORTRAN_POINTER_T*, shmem_ptr, SHMEM_PTR, (FORTRAN_POINTER_T target, MPI_Fint *pe));
 PN (ompi_fortran_logical_t, shmem_pe_accessible, SHMEM_PE_ACCESSIBLE, (MPI_Fint *pe));
 PN (MPI_Fint, shmem_addr_accessible, SHMEM_ADDR_ACCESSIBLE, (FORTRAN_POINTER_T addr, MPI_Fint *pe));
+
 PN (void, shmem_put, SHMEM_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, shmem_character_put, SHMEM_CHARACTER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, shmem_complex_put, SHMEM_COMPLEX_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
@@ -52,6 +53,7 @@ PN (void, shmem_put32, SHMEM_PUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T
 PN (void, shmem_put64, SHMEM_PUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, shmem_put128, SHMEM_PUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
 PN (void, shmem_putmem, SHMEM_PUTMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+
 PN (void, shmem_iput4, SHMEM_IPUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_iput8, SHMEM_IPUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_iput32, SHMEM_IPUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
@@ -62,6 +64,20 @@ PN (void, shmem_double_iput, SHMEM_DOUBLE_IPUT, (FORTRAN_POINTER_T target, FORTR
 PN (void, shmem_integer_iput, SHMEM_INTEGER_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_logical_iput, SHMEM_LOGICAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_real_iput, SHMEM_REAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+
+PN (void, shmem_putmem_nbi, SHMEM_PUTMEM_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_character_put_nbi, SHMEM_CHARACTER_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_complex_put_nbi, SHMEM_COMPLEX_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_double_put_nbi, SHMEM_DOUBLE_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_integer_put_nbi, SHMEM_INTEGER_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_logical_put_nbi, SHMEM_LOGICAL_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_real_put_nbi, SHMEM_REAL_PUT_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put4_nbi, SHMEM_PUT4_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put8_nbi, SHMEM_PUT8_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put32_nbi, SHMEM_PUT32_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put64_nbi, SHMEM_PUT64_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put128_nbi, SHMEM_PUT128_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+
 PN (void, shmem_character_get, SHMEM_CHARACTER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_complex_get, SHMEM_COMPLEX_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_double_get, SHMEM_DOUBLE_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
@@ -74,6 +90,7 @@ PN (void, shmem_get128, SHMEM_GET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER
 PN (void, shmem_getmem, SHMEM_GETMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_logical_get, SHMEM_LOGICAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_real_get, SHMEM_REAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+
 PN (void, shmem_iget4, SHMEM_IGET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_iget8, SHMEM_IGET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_iget32, SHMEM_IGET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
@@ -84,6 +101,21 @@ PN (void, shmem_double_iget, SHMEM_DOUBLE_IGET, (FORTRAN_POINTER_T target, FORTR
 PN (void, shmem_integer_iget, SHMEM_INTEGER_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_logical_iget, SHMEM_LOGICAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
 PN (void, shmem_real_iget, SHMEM_REAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+
+PN (void, shmem_getmem_nbi, SHMEM_GETMEM_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_character_get_nbi, SHMEM_CHARACTER_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_complex_get_nbi, SHMEM_COMPLEX_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_double_get_nbi, SHMEM_DOUBLE_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_integer_get_nbi, SHMEM_INTEGER_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_logical_get_nbi, SHMEM_LOGICAL_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_real_get_nbi, SHMEM_REAL_GET_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get4_nbi, SHMEM_GET4_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get8_nbi, SHMEM_GET8_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get32_nbi, SHMEM_GET32_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get64_nbi, SHMEM_GET64_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get128_nbi, SHMEM_GET128_NBI, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+
+
 PN (MPI_Fint, shmem_swap, SHMEM_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_integer4_t, shmem_int4_swap, SHMEM_INT4_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_integer8_t, shmem_int8_swap, SHMEM_INT8_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));

oshmem/shmem/fortran/shmem_get_nb_f.c

@@ -0,0 +1,250 @@
+   /*
+ * Copyright (c) 2016      Mellanox Technologies, Inc.
+ *                         All rights reserved.
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#include "oshmem_config.h"
+#include "oshmem/shmem/fortran/bindings.h"
+#include "oshmem/include/shmem.h"
+#include "oshmem/shmem/shmem_api_logger.h"
+#include "oshmem/runtime/runtime.h"
+#include "oshmem/mca/spml/spml.h"
+#include "ompi/datatype/ompi_datatype.h"
+#include "stdio.h"
+
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GETMEM_NBI, shmem_getmem_nbi)
+
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CHARACTER_GET_NBI, shmem_character_get_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMPLEX_GET_NBI, shmem_complex_get_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_DOUBLE_GET_NBI, shmem_double_get_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INTEGER_GET_NBI, shmem_integer_get_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_LOGICAL_GET_NBI, shmem_logical_get_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL_GET_NBI, shmem_real_get_nbi)
+
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET4_NBI, shmem_get4_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET8_NBI, shmem_get8_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET32_NBI, shmem_get32_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET64_NBI, shmem_get64_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET128_NBI, shmem_get128_nbi)
+
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_GETMEM_NBI,
+        shmem_getmem_nbi_,
+        shmem_getmem_nbi__,
+        shmem_getmem_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_getmem_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len),
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_CHARACTER_GET_NBI,
+        shmem_character_get_nbi_,
+        shmem_character_get_nbi__,
+        shmem_character_get_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe))
+
+void shmem_character_get_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    size_t character_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_character.dt, &character_type_size);
+
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * character_type_size,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMPLEX_GET_NBI,
+        shmem_complex_get_nbi_,
+        shmem_complex_get_nbi__,
+        shmem_complex_get_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_complex_get_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    size_t complex_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_cplex.dt, &complex_type_size);
+
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * complex_type_size,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_DOUBLE_GET_NBI,
+        shmem_double_get_nbi_,
+        shmem_double_get_nbi__,
+        shmem_double_get_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_double_get_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    size_t double_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_dblprec.dt, &double_type_size);
+
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * double_type_size,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INTEGER_GET_NBI,
+        shmem_integer_get_nbi_,
+        shmem_integer_get_nbi__,
+        shmem_integer_get_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_integer_get_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    size_t integer_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_integer.dt, &integer_type_size);
+
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * integer_type_size,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_LOGICAL_GET_NBI,
+        shmem_logical_get_nbi_,
+        shmem_logical_get_nbi__,
+        shmem_logical_get_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_logical_get_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    size_t logical_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_logical.dt, &logical_type_size);
+
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * logical_type_size,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_REAL_GET_NBI,
+        shmem_real_get_nbi_,
+        shmem_real_get_nbi__,
+        shmem_real_get_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_real_get_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    size_t real_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_real.dt, &real_type_size);
+
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * real_type_size,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_GET4_NBI,
+        shmem_get4_nbi_,
+        shmem_get4_nbi__,
+        shmem_get4_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_get4_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * 4,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_GET8_NBI,
+        shmem_get8_nbi_,
+        shmem_get8_nbi__,
+        shmem_get8_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_get8_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * 8,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_GET32_NBI,
+        shmem_get32_nbi_,
+        shmem_get32_nbi__,
+        shmem_get32_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_get32_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * 4,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_GET64_NBI,
+        shmem_get64_nbi_,
+        shmem_get64_nbi__,
+        shmem_get64_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_get64_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * 8,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_GET128_NBI,
+        shmem_get128_nbi_,
+        shmem_get128_nbi__,
+        shmem_get128_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe),
+        (target,source,len,pe) )
+
+void shmem_get128_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(get_nb(FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*len) * 16,
+        FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*pe), NULL));
+}

oshmem/shmem/fortran/shmem_put_nb_f.c

@@ -0,0 +1,248 @@
+/*
+ * Copyright (c) 2016      Mellanox Technologies, Inc.
+ *                         All rights reserved.
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#include "oshmem_config.h"
+#include "oshmem/shmem/fortran/bindings.h"
+#include "oshmem/include/shmem.h"
+#include "oshmem/shmem/shmem_api_logger.h"
+#include "oshmem/runtime/runtime.h"
+#include "ompi/datatype/ompi_datatype.h"
+#include "stdio.h"
+
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUTMEM_NBI, shmem_putmem_nbi)
+
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CHARACTER_PUT_NBI, shmem_character_put_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMPLEX_PUT_NBI, shmem_complex_put_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_DOUBLE_PUT_NBI, shmem_double_put_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INTEGER_PUT_NBI, shmem_integer_put_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_LOGICAL_PUT_NBI, shmem_logical_put_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL_PUT_NBI, shmem_real_put_nbi)
+
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT4_NBI, shmem_put4_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT8_NBI, shmem_put8_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT32_NBI, shmem_put32_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT64_NBI, shmem_put64_nbi)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT128_NBI, shmem_put128_nbi)
+
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUTMEM_NBI,
+        shmem_putmem_nbi_,
+        shmem_putmem_nbi__,
+        shmem_putmem_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_putmem_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length),
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_CHARACTER_PUT_NBI,
+        shmem_character_put_nbi_,
+        shmem_character_put_nbi__,
+        shmem_character_put_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_character_put_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    size_t character_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_character.dt, &character_type_size);
+
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * character_type_size,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMPLEX_PUT_NBI,
+        shmem_complex_put_nbi_,
+        shmem_complex_put_nbi__,
+        shmem_complex_put_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_complex_put_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    size_t complex_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_cplex.dt, &complex_type_size);
+
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * complex_type_size,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_DOUBLE_PUT_NBI,
+        shmem_double_put_nbi_,
+        shmem_double_put_nbi__,
+        shmem_double_put_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_double_put_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    size_t double_precision_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_dblprec.dt, &double_precision_type_size);
+
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * double_precision_type_size,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INTEGER_PUT_NBI,
+        shmem_integer_put_nbi_,
+        shmem_integer_put_nbi__,
+        shmem_integer_put_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_integer_put_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    size_t integer_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_integer.dt, &integer_type_size);
+
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * integer_type_size,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_LOGICAL_PUT_NBI,
+        shmem_logical_put_nbi_,
+        shmem_logical_put_nbi__,
+        shmem_logical_put_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_logical_put_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    size_t logical_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_logical.dt, &logical_type_size);
+
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * logical_type_size,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_REAL_PUT_NBI,
+        shmem_real_put_nbi_,
+        shmem_real_put_nbi__,
+        shmem_real_put_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_real_put_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    size_t real_type_size = 0;
+    ompi_datatype_type_size(&ompi_mpi_real.dt, &real_type_size);
+
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * real_type_size,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUT4_NBI,
+        shmem_put4_nbi_,
+        shmem_put4_nbi__,
+        shmem_put4_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_put4_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * 4,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUT8_NBI,
+        shmem_put8_nbi_,
+        shmem_put8_nbi__,
+        shmem_put8_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_put8_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * 8,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUT32_NBI,
+        shmem_put32_nbi_,
+        shmem_put32_nbi__,
+        shmem_put32_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_put32_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * 4,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUT64_NBI,
+        shmem_put64_nbi_,
+        shmem_put64_nbi__,
+        shmem_put64_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_put64_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * 8,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUT128_NBI,
+        shmem_put128_nbi_,
+        shmem_put128_nbi__,
+        shmem_put128_nbi_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe),
+        (target,source,length,pe) )
+
+void shmem_put128_nbi_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put_nb(FPTR_2_VOID_PTR(target),
+        OMPI_FINT_2_INT(*length) * 16,
+        FPTR_2_VOID_PTR(source),
+        OMPI_FINT_2_INT(*pe), NULL));
+}