v2.x: oshmem: Align OSHMEM API with spec v1.3

oshmem/include/pshmem.h

@@ -128,6 +128,24 @@ OSHMEM_DECLSPEC void pshmem_iput32(void* target, const void* source, ptrdiff_t t
 OSHMEM_DECLSPEC void pshmem_iput64(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 OSHMEM_DECLSPEC void pshmem_iput128(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 
+/*
+ * Nonblocking put routines
+ */
+OSHMEM_DECLSPEC  void pshmem_putmem_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_char_put_nbi(char *target, const char *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_short_put_nbi(short *target, const short *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_int_put_nbi(int* target, const int* source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_long_put_nbi(long *target, const long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_longlong_put_nbi(long long *target, const long long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_float_put_nbi(float *target, const float *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_double_put_nbi(double *target, const double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_longdouble_put_nbi(long double *target, const long double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_put8_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_put16_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_put32_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_put64_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_put128_nbi(void *target, const void *source, size_t len, int pe);
+
 /*
  * Elemental get routines
  */
@@ -169,6 +187,24 @@ OSHMEM_DECLSPEC void pshmem_iget32(void* target, const void* source, ptrdiff_t t
 OSHMEM_DECLSPEC void pshmem_iget64(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 OSHMEM_DECLSPEC void pshmem_iget128(void* target, const void* source, ptrdiff_t tst, ptrdiff_t sst,size_t len, int pe);
 
+/*
+ * Nonblocking data get routines
+ */
+OSHMEM_DECLSPEC  void pshmem_getmem_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_char_get_nbi(char *target, const char *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_short_get_nbi(short *target, const short *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_int_get_nbi(int *target, const int *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_long_get_nbi(long *target, const long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_longlong_get_nbi(long long *target, const long long *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_float_get_nbi(float *target, const float *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_double_get_nbi(double *target, const double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_longdouble_get_nbi(long double *target, const long double *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_get8_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_get16_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_get32_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_get64_nbi(void *target, const void *source, size_t len, int pe);
+OSHMEM_DECLSPEC  void pshmem_get128_nbi(void *target, const void *source, size_t len, int pe);
+
 /*
  * Atomic operations
  */

oshmem/include/shmem.h.in

@@ -193,6 +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
  */
@@ -234,6 +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/spml/base/base.h

@@ -73,6 +73,16 @@ OSHMEM_DECLSPEC int mca_spml_base_oob_get_mkeys(int pe,
 
 OSHMEM_DECLSPEC void mca_spml_base_rmkey_unpack(sshmem_mkey_t *mkey, int pe);
 OSHMEM_DECLSPEC void mca_spml_base_rmkey_free(sshmem_mkey_t *mkey);
+OSHMEM_DECLSPEC int mca_spml_base_put_nb(void *dst_addr,
+                                         size_t size,
+                                         void *src_addr,
+                                         int dst,
+                                         void **handle);
+OSHMEM_DECLSPEC int mca_spml_base_get_nb(void *dst_addr,
+                                         size_t size,
+                                         void *src_addr,
+                                         int src,
+                                         void **handle);
 
 /*
  * MCA framework

oshmem/mca/spml/base/spml_base.c

@@ -165,3 +165,15 @@ void mca_spml_base_rmkey_unpack(sshmem_mkey_t *mkey, int pe)
 void mca_spml_base_rmkey_free(sshmem_mkey_t *mkey)
 {
 }
+
+int mca_spml_base_put_nb(void *dst_addr, size_t size,
+                         void *src_addr, int dst, void **handle)
+{
+    return OSHMEM_ERROR;
+}
+
+int mca_spml_base_get_nb(void *dst_addr, size_t size,
+                         void *src_addr, int src, void **handle)
+{
+    return OSHMEM_ERROR;
+}

oshmem/mca/spml/ikrit/spml_ikrit.c

@@ -221,6 +221,7 @@ mca_spml_ikrit_t mca_spml_ikrit = {
         mca_spml_ikrit_put,
         mca_spml_ikrit_put_nb,
         mca_spml_ikrit_get,
+        mca_spml_ikrit_get_nb,
         mca_spml_ikrit_recv,
         mca_spml_ikrit_send,
         mca_spml_base_wait,
@@ -871,6 +872,15 @@ static inline int mca_spml_ikrit_get_shm(void *src_addr,
     return OSHMEM_SUCCESS;
 }
 
+int mca_spml_ikrit_get_nb(void* src_addr,
+                          size_t size,
+                          void* dst_addr,
+                          int src,
+                          void **handle)
+{
+    return mca_spml_ikrit_get_async(src_addr, size, dst_addr, src);
+}
+
 int mca_spml_ikrit_get(void *src_addr, size_t size, void *dst_addr, int src)
 {
     mxm_send_req_t sreq;

oshmem/mca/spml/ikrit/spml_ikrit.h

@@ -134,6 +134,11 @@ extern int mca_spml_ikrit_get(void* dst_addr,
                               size_t size,
                               void* src_addr,
                               int src);
+extern int mca_spml_ikrit_get_nb(void* src_addr,
+                                 size_t size,
+                                 void* dst_addr,
+                                 int src,
+                                 void **handle);
 /* extension. used 4 fence implementation b4 fence was added to mxm */
 extern int mca_spml_ikrit_get_async(void *src_addr,
                                     size_t size,

oshmem/mca/spml/spml.h

@@ -210,17 +210,35 @@ typedef int (*mca_spml_base_module_put_nb_fn_t)(void *dst_addr,
  * Blocking data transfer from remote PE.
  * Read data from remote PE.
  *
- * @param dst_addr - The address on the local PE, to write the result of the get operation to.
- * @param size     - The number of bytes to be read.
- * @param src_addr - The address on the remote PE, to read from.
- * @param src      - The ID of the remote PE.
- * @return         - OSHMEM_SUCCESS or failure status.
+ * @param dst_addr The address on the local PE, to write the result of the get operation to.
+ * @param size     The number of bytes to be read.
+ * @param src_addr The address on the remote PE, to read from.
+ * @param src      The ID of the remote PE.
+ * @return         OSHMEM_SUCCESS or failure status.
  */
 typedef int (*mca_spml_base_module_get_fn_t)(void *dst_addr,
                                              size_t size,
                                              void *src_addr,
                                              int src);
 
+/**
+ * Non-blocking data transfer from remote PE.
+ * Read data from remote PE.
+ *
+ * @param dst_addr The address on the local PE, to write the result of the get operation to.
+ * @param size     The number of bytes to be read.
+ * @param src_addr The address on the remote PE, to read from.
+ * @param src      The ID of the remote PE.
+ * @param handle   The address of a handle to be passed to shmem_wait_nb() or
+ *                 shmem_test_nb() to wait or poll for the completion of the transfer.
+ * @return         - OSHMEM_SUCCESS or failure status.
+ */
+typedef int (*mca_spml_base_module_get_nb_fn_t)(void *dst_addr,
+                                               size_t size,
+                                               void *src_addr,
+                                               int src,
+                                               void **handle);
+
 /**
  *  Post a receive and wait for completion.
  *
@@ -257,7 +275,7 @@ typedef int (*mca_spml_base_module_fence_fn_t)(void);
  *
  * @return         - OSHMEM_SUCCESS or failure status.
  */
-typedef int (*mca_spml_base_module_wait_nb_fn_t)(void*);
+typedef int (*mca_spml_base_module_wait_nb_fn_t)(void *);
 
 /**
  *  SPML instance.
@@ -275,6 +293,7 @@ struct mca_spml_base_module_1_0_0_t {
     mca_spml_base_module_put_fn_t spml_put;
     mca_spml_base_module_put_nb_fn_t spml_put_nb;
     mca_spml_base_module_get_fn_t spml_get;
+    mca_spml_base_module_get_nb_fn_t spml_get_nb;
 
     mca_spml_base_module_recv_fn_t spml_recv;
     mca_spml_base_module_send_fn_t spml_send;

oshmem/mca/spml/ucx/spml_ucx.c

@@ -53,8 +53,9 @@ mca_spml_ucx_t mca_spml_ucx = {
         mca_spml_ucx_deregister,
         mca_spml_base_oob_get_mkeys,
         mca_spml_ucx_put,
-        NULL, /* todo: mca_spml_ucx_put_nb, */
+        mca_spml_ucx_put_nb,
         mca_spml_ucx_get,
+        mca_spml_ucx_get_nb,
         mca_spml_ucx_recv,
         mca_spml_ucx_send,
         mca_spml_base_wait,
@@ -390,6 +391,19 @@ int mca_spml_ucx_get(void *src_addr, size_t size, void *dst_addr, int src)
     return ucx_status_to_oshmem(status);
 }
 
+int mca_spml_ucx_get_nb(void *src_addr, size_t size, void *dst_addr, int src, void **handle)
+{
+    void *rva;
+    ucs_status_t status;
+    spml_ucx_mkey_t *ucx_mkey;
+
+    ucx_mkey = mca_spml_ucx_get_mkey(src, src_addr, &rva);
+    status = ucp_get_nbi(mca_spml_ucx.ucp_peers[src].ucp_conn, dst_addr, size,
+                     (uint64_t)rva, ucx_mkey->rkey);
+
+    return ucx_status_to_oshmem(status);
+}
+
 int mca_spml_ucx_put(void* dst_addr, size_t size, void* src_addr, int dst)
 {
     void *rva;
@@ -403,6 +417,19 @@ int mca_spml_ucx_put(void* dst_addr, size_t size, void* src_addr, int dst)
     return ucx_status_to_oshmem(status);
 }
 
+int mca_spml_ucx_put_nb(void* dst_addr, size_t size, void* src_addr, int dst, void **handle)
+{
+    void *rva;
+    ucs_status_t status;
+    spml_ucx_mkey_t *ucx_mkey;
+
+    ucx_mkey = mca_spml_ucx_get_mkey(dst, dst_addr, &rva);
+    status = ucp_put_nbi(mca_spml_ucx.ucp_peers[dst].ucp_conn, src_addr, size,
+                     (uint64_t)rva, ucx_mkey->rkey);
+
+    return ucx_status_to_oshmem(status);
+}
+
 int mca_spml_ucx_fence(void)
 {
     ucs_status_t err;

oshmem/mca/spml/ucx/spml_ucx.h

@@ -72,6 +72,11 @@ extern int mca_spml_ucx_get(void* dst_addr,
                               size_t size,
                               void* src_addr,
                               int src);
+extern int mca_spml_ucx_get_nb(void* dst_addr,
+                              size_t size,
+                              void* src_addr,
+                              int src,
+                              void **handle);
 
 extern int mca_spml_ucx_put(void* dst_addr,
                               size_t size,

oshmem/mca/spml/yoda/spml_yoda.c

@@ -57,6 +57,7 @@ mca_spml_yoda_module_t mca_spml_yoda = {
         mca_spml_yoda_put,
         mca_spml_yoda_put_nb,
         mca_spml_yoda_get,
+        mca_spml_yoda_get_nb,
         mca_spml_yoda_recv,
         mca_spml_yoda_send,
         mca_spml_base_wait,
@@ -906,6 +907,8 @@ int mca_spml_yoda_put_nb(void* dst_addr,
 {
     UNREFERENCED_PARAMETER(handle);
 
+    /* TODO: real nonblocking operation is needed
+     */
     return mca_spml_yoda_put_internal(dst_addr, size, src_addr, dst, 1);
 }
 
@@ -979,6 +982,17 @@ int mca_spml_yoda_enable(bool enable)
     return OSHMEM_SUCCESS;
 }
 
+int mca_spml_yoda_get_nb(void* src_addr,
+                          size_t size,
+                          void* dst_addr,
+                          int src,
+                          void **handle)
+{
+    /* TODO: real nonblocking operation is needed
+     */
+    return mca_spml_yoda_get(src_addr, size, dst_addr, src);
+}
+
 /**
  * shmem_get reads data from a remote address
  * in the symmetric heap via RDMA READ.

oshmem/mca/spml/yoda/spml_yoda.h

@@ -104,6 +104,11 @@ extern int mca_spml_yoda_get(void* dst_addr,
                              size_t size,
                              void* src_addr,
                              int src);
+extern int mca_spml_yoda_get_nb(void* dst_addr,
+                                size_t size,
+                                void* src_addr,
+                                int dst,
+                                void **handle);
 extern int mca_spml_yoda_put(void* dst_addr,
                              size_t size,
                              void* src_addr,

oshmem/shmem/c/Makefile.am

@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2013-2015 Mellanox Technologies, Inc.
+# Copyright (c) 2013-2016 Mellanox Technologies, Inc.
 #                         All rights reserved
 # Copyright (c) 2014 Cisco Systems, Inc.  All rights reserved.
 # $COPYRIGHT$
@@ -41,6 +41,8 @@ OSHMEM_API_SOURCES = \
 	shmem_wait.c \
 	shmem_iget.c \
 	shmem_iput.c \
+	shmem_get_nb.c \
+	shmem_put_nb.c \
 	shmem_udcflush.c \
 	shmem_udcflush_line.c \
 	shmem_set_cache_inv.c \

oshmem/shmem/c/profile/Makefile.am

@@ -53,6 +53,8 @@ OSHMEM_API_SOURCES = \
 	pshmem_wait.c \
 	pshmem_iget.c \
 	pshmem_iput.c \
+	pshmem_get_nb.c \
+	pshmem_put_nb.c \
 	pshmem_udcflush.c \
 	pshmem_udcflush_line.c \
 	pshmem_set_cache_inv.c \