From ac09bc3724aa5d92be061f16f206598b32a85a1a Mon Sep 17 00:00:00 2001
From: "Rahman, Md" <md.rahman@intel.com>
Date: Tue, 13 Feb 2024 13:51:38 -0800
Subject: [PATCH] Using put_nb for iput when OFI inject is disabled

---
 src/data_c.c4 | 57 +++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 55 insertions(+), 2 deletions(-)

diff --git a/src/data_c.c4 b/src/data_c.c4
index 2083fb0b..be007144 100644
--- a/src/data_c.c4
+++ b/src/data_c.c4
@@ -416,6 +416,7 @@ SHMEM_PROF_DEF_CTX_PUT_N_SIGNAL_NBI(`mem')
     shmem_internal_get(ctx, target, source, (SIZE)*nelems, pe);\
   }
 
+#ifndef DISABLE_OFI_INJECT
 #define SHMEM_DEF_IPUT(STYPE,TYPE)                            \
   void SHMEM_FUNCTION_ATTRIBUTES                              \
   SHMEM_FUNC_PROTOTYPE(STYPE##_iput, TYPE *target,            \
@@ -438,8 +439,34 @@ SHMEM_PROF_DEF_CTX_PUT_N_SIGNAL_NBI(`mem')
       source += sst;                                          \
     }                                                         \
   }
+#else
+#define SHMEM_DEF_IPUT(STYPE,TYPE)                            \
+  void SHMEM_FUNCTION_ATTRIBUTES                              \
+  SHMEM_FUNC_PROTOTYPE(STYPE##_iput, TYPE *target,            \
+                       const TYPE *source, ptrdiff_t tst,     \
+                       ptrdiff_t sst, size_t nelems, int pe)  \
+    SHMEM_ERR_CHECK_INITIALIZED();                            \
+    SHMEM_ERR_CHECK_PE(pe);                                   \
+    SHMEM_ERR_CHECK_CTX(ctx);                                 \
+    SHMEM_ERR_CHECK_POSITIVE(tst);                            \
+    SHMEM_ERR_CHECK_POSITIVE(sst);                            \
+    SHMEM_ERR_CHECK_SYMMETRIC(target, sizeof(TYPE) * ((nelems-1) * tst + 1)); \
+    SHMEM_ERR_CHECK_NULL(source, nelems);                     \
+    SHMEM_ERR_CHECK_OVERLAP(target, source,                   \
+                   sizeof(TYPE) * ((nelems-1) * tst + 1),     \
+                   sizeof(TYPE) * ((nelems-1) * sst + 1), 0); \
+    for ( ; nelems > 0 ; --nelems) {                          \
+      long completion = 0;                                    \
+      shmem_internal_put_nb(ctx, target, source,              \
+                            sizeof(TYPE), pe, &completion);   \
+      shmem_internal_put_wait(ctx, &completion);              \
+      target += tst;                                          \
+      source += sst;                                          \
+    }                                                         \
+  }
+#endif
 
-
+#ifndef DISABLE_OFI_INJECT
 #define SHMEM_DEF_IPUT_N(NAME,SIZE)                          \
   void SHMEM_FUNCTION_ATTRIBUTES                             \
   SHMEM_FUNC_PROTOTYPE(iput##NAME, void *target,             \
@@ -463,7 +490,33 @@ SHMEM_PROF_DEF_CTX_PUT_N_SIGNAL_NBI(`mem')
       source = (uint8_t*)source + sst*(SIZE);                \
     }                                                        \
   }
-
+#else
+#define SHMEM_DEF_IPUT_N(NAME,SIZE)                          \
+  void SHMEM_FUNCTION_ATTRIBUTES                             \
+  SHMEM_FUNC_PROTOTYPE(iput##NAME, void *target,             \
+                       const void *source, ptrdiff_t tst,    \
+                       ptrdiff_t sst, size_t nelems, int pe) \
+    SHMEM_ERR_CHECK_INITIALIZED();                           \
+    SHMEM_ERR_CHECK_PE(pe);                                  \
+    SHMEM_ERR_CHECK_CTX(ctx);                                \
+    SHMEM_ERR_CHECK_POSITIVE(tst);                           \
+    SHMEM_ERR_CHECK_POSITIVE(sst);                           \
+    SHMEM_ERR_CHECK_SYMMETRIC(target,                        \
+                           (SIZE) * ((nelems-1) * tst + 1)); \
+    SHMEM_ERR_CHECK_NULL(source, nelems);                    \
+    SHMEM_ERR_CHECK_OVERLAP(target, source,                  \
+                        (SIZE) * ((nelems-1) * tst + 1),     \
+                        (SIZE) * ((nelems-1) * sst + 1), 0); \
+    for ( ; nelems > 0 ; --nelems) {                         \
+      long completion = 0;                                   \
+      shmem_internal_put_nb(ctx, target, source, (SIZE), pe, \
+                            &completion);                    \
+      shmem_internal_put_wait(ctx, &completion);             \
+      target = (uint8_t*)target + tst*(SIZE);                \
+      source = (uint8_t*)source + sst*(SIZE);                \
+    }                                                        \
+  }
+#endif
 
 #define SHMEM_DEF_IGET(STYPE,TYPE)                            \
   void SHMEM_FUNCTION_ATTRIBUTES                              \