[mpich-commits] [mpich] MPICH primary repository branch, master, updated. v3.2-435-ge8f5d0b

Service Account noreply at mpich.org
Fri Sep 9 14:43:34 CDT 2016


This is an automated email from the git hooks/post-receive script. It was
generated because a ref change was pushed to the repository containing
the project "MPICH primary repository".

The branch, master has been updated
       via  e8f5d0b7f0441f55669024b45ca93870fecde385 (commit)
       via  7217189afc2428ce45417523015e70cae4ddcc82 (commit)
       via  6370ba6ce11942523e944f324298b6fe5209b4f5 (commit)
       via  7adf2a94c124f6b5a5ecbd0f77c9ca70a8ed8451 (commit)
      from  5a604db4c16372bf67d3dabfaa049e4bdd774f0e (commit)

Those revisions listed above that are new to this repository have
not appeared on any other notification email; so we list those
revisions in full, below.

- Log -----------------------------------------------------------------
http://git.mpich.org/mpich.git/commitdiff/e8f5d0b7f0441f55669024b45ca93870fecde385

commit e8f5d0b7f0441f55669024b45ca93870fecde385
Author: Pavan Balaji <balaji at anl.gov>
Date:   Tue Aug 30 13:56:40 2016 -0500

    CH4: Replace __*_INLINE__ with MPL_STATIC_INLINE_PREFIX
    
    Signed-off-by: Ken Raffenetti <raffenet at mcs.anl.gov>

diff --git a/src/mpid/ch4/include/mpid_ticketlock.h b/src/mpid/ch4/include/mpid_ticketlock.h
index 0dc5567..6f7898e 100644
--- a/src/mpid/ch4/include/mpid_ticketlock.h
+++ b/src/mpid/ch4/include/mpid_ticketlock.h
@@ -10,8 +10,6 @@
 #ifndef MPID_TICKETLOCK_H_INCLUDED
 #define MPID_TICKETLOCK_H_INCLUDED
 
-#define __MUTEX_INLINE__ __attribute__((always_inline))static inline
-
 #define MPIDI_CH4_CACHELINE_SIZE 64
 
 typedef union MPIDI_CH4_Ticket_lock {
@@ -23,20 +21,20 @@ typedef union MPIDI_CH4_Ticket_lock {
     } s;
 } MPIDI_CH4_Ticket_lock __attribute__ ((aligned(MPIDI_CH4_CACHELINE_SIZE)));
 
-__MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_acquire(MPIDI_CH4_Ticket_lock * m)
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4I_Thread_mutex_acquire(MPIDI_CH4_Ticket_lock * m)
 {
     uint16_t u = __sync_fetch_and_add(&m->s.clients, 1);
     while (m->s.ticket != u)
         asm volatile ("pause\n":::"memory");
 }
 
-__MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_release(MPIDI_CH4_Ticket_lock * m)
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4I_Thread_mutex_release(MPIDI_CH4_Ticket_lock * m)
 {
     asm volatile ("":::"memory");
     m->s.ticket++;
 }
 
-__MUTEX_INLINE__ int MPIDI_CH4I_Thread_mutex_try_acquire(MPIDI_CH4_Ticket_lock * m)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4I_Thread_mutex_try_acquire(MPIDI_CH4_Ticket_lock * m)
 {
     uint16_t u = m->s.clients;
     uint16_t u2 = u + 1;
@@ -49,25 +47,29 @@ __MUTEX_INLINE__ int MPIDI_CH4I_Thread_mutex_try_acquire(MPIDI_CH4_Ticket_lock *
     return EBUSY;
 }
 
-__MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_lock(MPIDI_CH4_Ticket_lock * m, int *mpi_error)
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4I_Thread_mutex_lock(MPIDI_CH4_Ticket_lock * m,
+                                                           int *mpi_error)
 {
     MPIDI_CH4I_Thread_mutex_acquire(m);
     *mpi_error = 0;
 }
 
-__MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_unlock(MPIDI_CH4_Ticket_lock * m, int *mpi_error)
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4I_Thread_mutex_unlock(MPIDI_CH4_Ticket_lock * m,
+                                                             int *mpi_error)
 {
     MPIDI_CH4I_Thread_mutex_release(m);
     *mpi_error = 0;
 }
 
-__MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_create(MPIDI_CH4_Ticket_lock * m, int *mpi_error)
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4I_Thread_mutex_create(MPIDI_CH4_Ticket_lock * m,
+                                                             int *mpi_error)
 {
     m->u = 0;
     *mpi_error = 0;
 }
 
-__MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_destroy(MPIDI_CH4_Ticket_lock * m, int *mpi_error)
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4I_Thread_mutex_destroy(MPIDI_CH4_Ticket_lock * m,
+                                                              int *mpi_error)
 {
     m->u = 0;
     *mpi_error = 0;
@@ -78,7 +80,7 @@ __MUTEX_INLINE__ void MPIDI_CH4I_Thread_mutex_destroy(MPIDI_CH4_Ticket_lock * m,
 /* 2)  Implement it from scratch                                                      */
 /* Currently only async.c is using condition variables, so we should figure out what  */
 /* we really want from the cv implementations                                         */
-__MUTEX_INLINE__ void
+MPL_STATIC_INLINE_PREFIX void
 MPIDI_CH4I_Thread_cond_wait(MPIDU_Thread_cond_t * cond, MPIDI_CH4_Ticket_lock * m, int *mpi_error)
 {
     MPIR_Assert(0);
diff --git a/src/mpid/ch4/include/mpidpost.h b/src/mpid/ch4/include/mpidpost.h
index 43a9658..770c442 100644
--- a/src/mpid/ch4/include/mpidpost.h
+++ b/src/mpid/ch4/include/mpidpost.h
@@ -14,7 +14,7 @@
 #include "mpidu_datatype.h"
 #include "mpidch4.h"
 
-__ALWAYS_INLINE__ void MPID_Request_create_hook(MPIR_Request * req)
+MPL_STATIC_INLINE_PREFIX void MPID_Request_create_hook(MPIR_Request * req)
 {
     MPIDI_CH4U_REQUEST(req, req) = NULL;
 #ifdef MPIDI_BUILD_CH4_SHM
@@ -22,7 +22,7 @@ __ALWAYS_INLINE__ void MPID_Request_create_hook(MPIR_Request * req)
 #endif
 }
 
-__ALWAYS_INLINE__ void MPID_Request_free_hook(MPIR_Request * req)
+MPL_STATIC_INLINE_PREFIX void MPID_Request_free_hook(MPIR_Request * req)
 {
     return;
 }
diff --git a/src/mpid/ch4/include/mpidpre.h b/src/mpid/ch4/include/mpidpre.h
index 698cd4a..60f379b 100644
--- a/src/mpid/ch4/include/mpidpre.h
+++ b/src/mpid/ch4/include/mpidpre.h
@@ -33,8 +33,6 @@ typedef struct {
 typedef int MPID_Progress_state;
 #define HAVE_GPID_ROUTINES
 
-#define __ALWAYS_INLINE__ __attribute__((always_inline)) static inline
-
 #define CH4_COMPILE_TIME_ASSERT(expr_)                                  \
   do { switch(0) { case 0: case (expr_): default: break; } } while (0)
 
@@ -198,9 +196,9 @@ typedef struct {
 #define MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(req)  NULL
 #endif
 
-__ALWAYS_INLINE__ void MPID_Request_create_hook(struct MPIR_Request *req);
+MPL_STATIC_INLINE_PREFIX void MPID_Request_create_hook(struct MPIR_Request *req);
 
-__ALWAYS_INLINE__ void MPID_Request_free_hook(struct MPIR_Request *req);
+MPL_STATIC_INLINE_PREFIX void MPID_Request_free_hook(struct MPIR_Request *req);
 
 typedef struct MPIDI_CH4U_win_shared_info {
     uint32_t disp_unit;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_events.h b/src/mpid/ch4/netmod/ofi/ofi_events.h
index 191bc51..74e5c9d 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_events.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_events.h
@@ -16,9 +16,10 @@
 #include "ofi_am_events.h"
 #include "ofi_control.h"
 
-__ALWAYS_INLINE__ int MPIDI_OFI_get_huge_event(struct fi_cq_tagged_entry *wc, MPIR_Request * req);
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_get_huge_event(struct fi_cq_tagged_entry *wc,
+                                                      MPIR_Request * req);
 
-__ALWAYS_INLINE__ int MPIDI_OFI_cqe_get_source(struct fi_cq_tagged_entry *wc, int do_data)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_cqe_get_source(struct fi_cq_tagged_entry *wc, int do_data)
 {
     if (do_data)
         return wc->data;
@@ -30,7 +31,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_cqe_get_source(struct fi_cq_tagged_entry *wc, in
 #define FUNCNAME MPIDI_OFI_peek_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_peek_event(struct fi_cq_tagged_entry *wc, MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_peek_event(struct fi_cq_tagged_entry *wc,
+                                                  MPIR_Request * rreq)
 {
     size_t count;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_NETMOD_PEEK_EVENT);
@@ -49,7 +51,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_peek_event(struct fi_cq_tagged_entry *wc, MPIR_R
 #define FUNCNAME MPIDI_OFI_peek_empty_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_peek_empty_event(struct fi_cq_tagged_entry *wc, MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_peek_empty_event(struct fi_cq_tagged_entry *wc,
+                                                        MPIR_Request * rreq)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_NETMOD_PEEK_EMPTY_EVENT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_NETMOD_PEEK_EMPTY_EVENT);
@@ -79,7 +82,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_peek_empty_event(struct fi_cq_tagged_entry *wc,
 #define FUNCNAME MPIDI_OFI_recv_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_recv_event(struct fi_cq_tagged_entry *wc, MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_recv_event(struct fi_cq_tagged_entry *wc,
+                                                  MPIR_Request * rreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPI_Aint last;
@@ -162,7 +166,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_recv_event(struct fi_cq_tagged_entry *wc, MPIR_R
 #define FUNCNAME MPIDI_OFI_recv_huge_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_recv_huge_event(struct fi_cq_tagged_entry *wc, MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_recv_huge_event(struct fi_cq_tagged_entry *wc,
+                                                       MPIR_Request * rreq)
 {
     MPIDI_OFI_huge_recv_t *recv;
     MPIR_Comm *comm_ptr;
@@ -196,7 +201,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_recv_huge_event(struct fi_cq_tagged_entry *wc, M
 #define FUNCNAME MPIDI_OFI_send_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send_event(struct fi_cq_tagged_entry *wc, MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send_event(struct fi_cq_tagged_entry *wc,
+                                                  MPIR_Request * sreq)
 {
     int c;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SEND_EVENT);
@@ -220,7 +226,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_send_event(struct fi_cq_tagged_entry *wc, MPIR_R
 #define FUNCNAME MPIDI_OFI_send_huge_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send_huge_event(struct fi_cq_tagged_entry *wc, MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send_huge_event(struct fi_cq_tagged_entry *wc,
+                                                       MPIR_Request * sreq)
 {
     int mpi_errno = MPI_SUCCESS;
     int c;
@@ -275,7 +282,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_send_huge_event(struct fi_cq_tagged_entry *wc, M
 #define FUNCNAME MPIDI_OFI_ssend_ack_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_ssend_ack_event(struct fi_cq_tagged_entry *wc, MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_ssend_ack_event(struct fi_cq_tagged_entry *wc,
+                                                       MPIR_Request * sreq)
 {
     int mpi_errno;
     MPIDI_OFI_ssendack_request_t *req = (MPIDI_OFI_ssendack_request_t *) sreq;
@@ -287,7 +295,7 @@ __ALWAYS_INLINE__ int MPIDI_OFI_ssend_ack_event(struct fi_cq_tagged_entry *wc, M
     return mpi_errno;
 }
 
-__ALWAYS_INLINE__ uintptr_t MPIDI_OFI_recv_rbase(MPIDI_OFI_huge_recv_t * recv)
+MPL_STATIC_INLINE_PREFIX uintptr_t MPIDI_OFI_recv_rbase(MPIDI_OFI_huge_recv_t * recv)
 {
 #ifdef USE_OFI_MR_SCALABLE
     return 0;
@@ -300,7 +308,8 @@ __ALWAYS_INLINE__ uintptr_t MPIDI_OFI_recv_rbase(MPIDI_OFI_huge_recv_t * recv)
 #define FUNCNAME MPIDI_OFI_get_huge_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_get_huge_event(struct fi_cq_tagged_entry *wc, MPIR_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_get_huge_event(struct fi_cq_tagged_entry *wc,
+                                                      MPIR_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIDI_OFI_huge_recv_t *recv = (MPIDI_OFI_huge_recv_t *) req;
@@ -355,7 +364,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_get_huge_event(struct fi_cq_tagged_entry *wc, MP
 #define FUNCNAME MPIDI_OFI_chunk_done_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_chunk_done_event(struct fi_cq_tagged_entry *wc, MPIR_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_chunk_done_event(struct fi_cq_tagged_entry *wc,
+                                                        MPIR_Request * req)
 {
     int c;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_CHUNK_DONE_EVENT);
@@ -376,7 +386,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_chunk_done_event(struct fi_cq_tagged_entry *wc,
 #define FUNCNAME MPIDI_OFI_inject_emu_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_inject_emu_event(struct fi_cq_tagged_entry *wc, MPIR_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_inject_emu_event(struct fi_cq_tagged_entry *wc,
+                                                        MPIR_Request * req)
 {
     int incomplete;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_INJECT_EMU_EVENT);
@@ -398,7 +409,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_inject_emu_event(struct fi_cq_tagged_entry *wc,
 #define FUNCNAME MPIDI_OFI_rma_done_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_rma_done_event(struct fi_cq_tagged_entry *wc, MPIR_Request * in_req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_rma_done_event(struct fi_cq_tagged_entry *wc,
+                                                      MPIR_Request * in_req)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_OFI_RMA_DONE_EVENT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_OFI_RMA_DONE_EVENT);
@@ -414,8 +426,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_rma_done_event(struct fi_cq_tagged_entry *wc, MP
 #define FUNCNAME MPIDI_OFI_accept_probe_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_accept_probe_event(struct fi_cq_tagged_entry *wc,
-                                                   MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_accept_probe_event(struct fi_cq_tagged_entry *wc,
+                                                          MPIR_Request * rreq)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_OFI_ACCEPT_PROBE_EVENT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_OFI_ACCEPT_PROBE_EVENT);
@@ -432,8 +444,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_accept_probe_event(struct fi_cq_tagged_entry *wc
 #define FUNCNAME MPIDI_OFI_dynproc_done_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_dynproc_done_event(struct fi_cq_tagged_entry *wc,
-                                                   MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_dynproc_done_event(struct fi_cq_tagged_entry *wc,
+                                                          MPIR_Request * rreq)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_OFI_DYNPROC_DONE_EVENT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_OFI_DYNPROC_DONE_EVENT);
@@ -447,7 +459,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_dynproc_done_event(struct fi_cq_tagged_entry *wc
 #define FUNCNAME MPIDI_OFI_am_isend_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_am_isend_event(struct fi_cq_tagged_entry *wc, MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_am_isend_event(struct fi_cq_tagged_entry *wc,
+                                                      MPIR_Request * sreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIDI_OFI_am_header_t *msg_hdr;
@@ -488,7 +501,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_am_isend_event(struct fi_cq_tagged_entry *wc, MP
 #define FUNCNAME MPIDI_OFI_am_recv_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_am_recv_event(struct fi_cq_tagged_entry *wc, MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_am_recv_event(struct fi_cq_tagged_entry *wc,
+                                                     MPIR_Request * rreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIDI_OFI_am_header_t *am_hdr;
@@ -545,8 +559,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_am_recv_event(struct fi_cq_tagged_entry *wc, MPI
 #define FUNCNAME MPIDI_OFI_am_read_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_am_read_event(struct fi_cq_tagged_entry *wc,
-                                              MPIR_Request * dont_use_me)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_am_read_event(struct fi_cq_tagged_entry *wc,
+                                                     MPIR_Request * dont_use_me)
 {
     int mpi_errno = MPI_SUCCESS;
     void *netmod_context = NULL;
@@ -585,7 +599,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_am_read_event(struct fi_cq_tagged_entry *wc,
 #define FUNCNAME MPIDI_OFI_am_repost_event
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_am_repost_event(struct fi_cq_tagged_entry *wc, MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_am_repost_event(struct fi_cq_tagged_entry *wc,
+                                                       MPIR_Request * rreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_REPOST_BUFFER);
@@ -597,8 +612,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_am_repost_event(struct fi_cq_tagged_entry *wc, M
     return mpi_errno;
 }
 
-__ALWAYS_INLINE__ int MPIDI_OFI_dispatch_function(struct fi_cq_tagged_entry *wc,
-                                                  MPIR_Request * req, int buffered)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_dispatch_function(struct fi_cq_tagged_entry *wc,
+                                                         MPIR_Request * req, int buffered)
 {
     int mpi_errno;
 
@@ -688,7 +703,7 @@ __ALWAYS_INLINE__ int MPIDI_OFI_dispatch_function(struct fi_cq_tagged_entry *wc,
 #define FUNCNAME MPIDI_OFI_get_buffered
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_get_buffered(struct fi_cq_tagged_entry *wc, ssize_t num)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_get_buffered(struct fi_cq_tagged_entry *wc, ssize_t num)
 {
     int rc = 0;
 
@@ -716,8 +731,8 @@ __ALWAYS_INLINE__ int MPIDI_OFI_get_buffered(struct fi_cq_tagged_entry *wc, ssiz
 #define FUNCNAME MPIDI_OFI_handle_cq_entries
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_handle_cq_entries(struct fi_cq_tagged_entry *wc,
-                                                  ssize_t num, int buffered)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_handle_cq_entries(struct fi_cq_tagged_entry *wc,
+                                                         ssize_t num, int buffered)
 {
     int i, mpi_errno = MPI_SUCCESS;
     MPIR_Request *req;
@@ -740,7 +755,7 @@ __ALWAYS_INLINE__ int MPIDI_OFI_handle_cq_entries(struct fi_cq_tagged_entry *wc,
 #define FUNCNAME MPIDI_OFI_handle_cq_error
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_handle_cq_error(ssize_t ret)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_handle_cq_error(ssize_t ret)
 {
     int mpi_errno = MPI_SUCCESS;
     struct fi_cq_err_entry e;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_impl.h b/src/mpid/ch4/netmod/ofi/ofi_impl.h
index 79aec52..ddbc9f9 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_impl.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_impl.h
@@ -226,7 +226,7 @@
 
 #define WINFO(w,rank) MPIDI_CH4U_WINFO(w,rank)
 
-__ALWAYS_INLINE__ uintptr_t MPIDI_OFI_winfo_base(MPIR_Win * w, int rank)
+MPL_STATIC_INLINE_PREFIX uintptr_t MPIDI_OFI_winfo_base(MPIR_Win * w, int rank)
 {
 #if MPIDI_OFI_ENABLE_MR_SCALABLE
     return 0;
@@ -235,7 +235,7 @@ __ALWAYS_INLINE__ uintptr_t MPIDI_OFI_winfo_base(MPIR_Win * w, int rank)
 #endif
 }
 
-__ALWAYS_INLINE__ uint64_t MPIDI_OFI_winfo_mr_key(MPIR_Win * w, int rank)
+MPL_STATIC_INLINE_PREFIX uint64_t MPIDI_OFI_winfo_mr_key(MPIR_Win * w, int rank)
 {
 #if MPIDI_OFI_ENABLE_MR_SCALABLE
     return MPIDI_OFI_WIN(w).mr_key;
@@ -245,40 +245,40 @@ __ALWAYS_INLINE__ uint64_t MPIDI_OFI_winfo_mr_key(MPIR_Win * w, int rank)
 }
 
 #ifdef MPIDI_OFI_CONFIG_USE_SCALABLE_ENDPOINTS
-__ALWAYS_INLINE__ void MPIDI_OFI_win_conditional_cntr_incr(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_win_conditional_cntr_incr(MPIR_Win * win)
 {
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_win_cntr_incr(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_win_cntr_incr(MPIR_Win * win)
 {
     (*MPIDI_OFI_WIN(win).issued_cntr)++;
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_conditional_cntr_incr()
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_conditional_cntr_incr()
 {
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_cntr_incr()
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_cntr_incr()
 {
     MPIDI_Global.rma_issued_cntr++;
 }
 #else
-__ALWAYS_INLINE__ void MPIDI_OFI_win_conditional_cntr_incr(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_win_conditional_cntr_incr(MPIR_Win * win)
 {
     (*MPIDI_OFI_WIN(win).issued_cntr)++;
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_win_cntr_incr(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_win_cntr_incr(MPIR_Win * win)
 {
     (*MPIDI_OFI_WIN(win).issued_cntr)++;
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_conditional_cntr_incr()
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_conditional_cntr_incr()
 {
     MPIDI_Global.rma_issued_cntr++;
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_cntr_incr()
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_cntr_incr()
 {
     MPIDI_Global.rma_issued_cntr++;
 }
@@ -306,7 +306,7 @@ extern void MPIDI_OFI_index_allocator_destroy(void *_indexmap);
 /* Common Utility functions used by the
  * C and C++ components
  */
-__ALWAYS_INLINE__ MPIDI_OFI_win_request_t *MPIDI_OFI_win_request_alloc_and_init(int extra)
+MPL_STATIC_INLINE_PREFIX MPIDI_OFI_win_request_t *MPIDI_OFI_win_request_alloc_and_init(int extra)
 {
     MPIDI_OFI_win_request_t *req;
     req = (MPIDI_OFI_win_request_t *) MPIR_Request_create(MPIR_REQUEST_KIND__RMA);
@@ -317,13 +317,13 @@ __ALWAYS_INLINE__ MPIDI_OFI_win_request_t *MPIDI_OFI_win_request_alloc_and_init(
     return req;
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_win_datatype_unmap(MPIDI_OFI_win_datatype_t * dt)
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_win_datatype_unmap(MPIDI_OFI_win_datatype_t * dt)
 {
     if (dt->map != &dt->__map)
         MPL_free(dt->map);
 }
 
-__ALWAYS_INLINE__ void MPIDI_OFI_win_request_complete(MPIDI_OFI_win_request_t * req)
+MPL_STATIC_INLINE_PREFIX void MPIDI_OFI_win_request_complete(MPIDI_OFI_win_request_t * req)
 {
     int count;
     MPIR_Assert(HANDLE_GET_MPI_KIND(req->handle) == MPIR_REQUEST);
@@ -338,7 +338,7 @@ __ALWAYS_INLINE__ void MPIDI_OFI_win_request_complete(MPIDI_OFI_win_request_t *
     }
 }
 
-__ALWAYS_INLINE__ fi_addr_t MPIDI_OFI_comm_to_phys(MPIR_Comm * comm, int rank, int ep_family)
+MPL_STATIC_INLINE_PREFIX fi_addr_t MPIDI_OFI_comm_to_phys(MPIR_Comm * comm, int rank, int ep_family)
 {
 #ifdef MPIDI_OFI_CONFIG_USE_SCALABLE_ENDPOINTS
     int ep_num = MPIDI_OFI_COMM_TO_EP(comm, rank);
@@ -350,7 +350,7 @@ __ALWAYS_INLINE__ fi_addr_t MPIDI_OFI_comm_to_phys(MPIR_Comm * comm, int rank, i
 #endif
 }
 
-__ALWAYS_INLINE__ fi_addr_t MPIDI_OFI_to_phys(int rank, int ep_family)
+MPL_STATIC_INLINE_PREFIX fi_addr_t MPIDI_OFI_to_phys(int rank, int ep_family)
 {
 #ifdef MPIDI_OFI_CONFIG_USE_SCALABLE_ENDPOINTS
     int ep_num = 0;
@@ -362,13 +362,14 @@ __ALWAYS_INLINE__ fi_addr_t MPIDI_OFI_to_phys(int rank, int ep_family)
 #endif
 }
 
-__ALWAYS_INLINE__ bool MPIDI_OFI_is_tag_sync(uint64_t match_bits)
+MPL_STATIC_INLINE_PREFIX bool MPIDI_OFI_is_tag_sync(uint64_t match_bits)
 {
     return (0 != (MPIDI_OFI_SYNC_SEND & match_bits));
 }
 
-__ALWAYS_INLINE__ uint64_t MPIDI_OFI_init_sendtag(MPIR_Context_id_t contextid,
-                                                  int source, int tag, uint64_t type, int do_data)
+MPL_STATIC_INLINE_PREFIX uint64_t MPIDI_OFI_init_sendtag(MPIR_Context_id_t contextid,
+                                                         int source, int tag, uint64_t type,
+                                                         int do_data)
 {
     uint64_t match_bits;
     match_bits = contextid;
@@ -384,9 +385,9 @@ __ALWAYS_INLINE__ uint64_t MPIDI_OFI_init_sendtag(MPIR_Context_id_t contextid,
 }
 
 /* receive posting */
-__ALWAYS_INLINE__ uint64_t MPIDI_OFI_init_recvtag(uint64_t * mask_bits,
-                                                  MPIR_Context_id_t contextid,
-                                                  int source, int tag, int do_data)
+MPL_STATIC_INLINE_PREFIX uint64_t MPIDI_OFI_init_recvtag(uint64_t * mask_bits,
+                                                         MPIR_Context_id_t contextid,
+                                                         int source, int tag, int do_data)
 {
     uint64_t match_bits = 0;
     *mask_bits = MPIDI_OFI_PROTOCOL_MASK;
@@ -416,17 +417,17 @@ __ALWAYS_INLINE__ uint64_t MPIDI_OFI_init_recvtag(uint64_t * mask_bits,
     return match_bits;
 }
 
-__ALWAYS_INLINE__ int MPIDI_OFI_init_get_tag(uint64_t match_bits)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_init_get_tag(uint64_t match_bits)
 {
     return ((int) (match_bits & MPIDI_OFI_TAG_MASK));
 }
 
-__ALWAYS_INLINE__ int MPIDI_OFI_init_get_source(uint64_t match_bits)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_init_get_source(uint64_t match_bits)
 {
     return ((int) ((match_bits & MPIDI_OFI_SOURCE_MASK) >> MPIDI_OFI_TAG_SHIFT));
 }
 
-__ALWAYS_INLINE__ MPIR_Request *MPIDI_OFI_context_to_request(void *context)
+MPL_STATIC_INLINE_PREFIX MPIR_Request *MPIDI_OFI_context_to_request(void *context)
 {
     char *base = (char *) context;
     return (MPIR_Request *) container_of(base, MPIR_Request, dev.ch4.netmod);
@@ -436,10 +437,10 @@ __ALWAYS_INLINE__ MPIR_Request *MPIDI_OFI_context_to_request(void *context)
 #define FUNCNAME MPIDI_OFI_send_handler
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send_handler(struct fid_ep *ep, const void *buf, size_t len,
-                                             void *desc, uint64_t dest, fi_addr_t dest_addr,
-                                             uint64_t tag, void *context, int is_inject,
-                                             int do_data, int do_lock)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send_handler(struct fid_ep *ep, const void *buf, size_t len,
+                                                    void *desc, uint64_t dest, fi_addr_t dest_addr,
+                                                    uint64_t tag, void *context, int is_inject,
+                                                    int do_data, int do_lock)
 {
     int mpi_errno = MPI_SUCCESS;
 
diff --git a/src/mpid/ch4/netmod/ofi/ofi_progress.h b/src/mpid/ch4/netmod/ofi/ofi_progress.h
index e5da34b..1d292b3 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_progress.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_progress.h
@@ -15,7 +15,7 @@
 #include "ofi_events.h"
 #include "ofi_am_events.h"
 
-__ALWAYS_INLINE__
+MPL_STATIC_INLINE_PREFIX
     int MPIDI_OFI_progress_generic(void *netmod_context, int blocking, int do_am, int do_tagged)
 {
     int mpi_errno;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_recv.h b/src/mpid/ch4/netmod/ofi/ofi_recv.h
index 375b73d..da88499 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_recv.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_recv.h
@@ -20,14 +20,14 @@
 #define FUNCNAME MPIDI_OFI_do_irecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_do_irecv(void *buf,
-                                         int count,
-                                         MPI_Datatype datatype,
-                                         int rank,
-                                         int tag,
-                                         MPIR_Comm * comm,
-                                         int context_offset,
-                                         MPIR_Request ** request, int mode, uint64_t flags)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_do_irecv(void *buf,
+                                                int count,
+                                                MPI_Datatype datatype,
+                                                int rank,
+                                                int tag,
+                                                MPIR_Comm * comm,
+                                                int context_offset,
+                                                MPIR_Request ** request, int mode, uint64_t flags)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq = NULL;
@@ -120,14 +120,14 @@ __ALWAYS_INLINE__ int MPIDI_OFI_do_irecv(void *buf,
 #define FUNCNAME MPIDI_NM_mpi_recv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv(void *buf,
-                                        int count,
-                                        MPI_Datatype datatype,
-                                        int rank,
-                                        int tag,
-                                        MPIR_Comm * comm,
-                                        int context_offset,
-                                        MPI_Status * status, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv(void *buf,
+                                               int count,
+                                               MPI_Datatype datatype,
+                                               int rank,
+                                               int tag,
+                                               MPIR_Comm * comm,
+                                               int context_offset,
+                                               MPI_Status * status, MPIR_Request ** request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RECV);
@@ -142,13 +142,13 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_recv(void *buf,
 #define FUNCNAME MPIDI_NM_mpi_recv_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv_init(void *buf,
-                                             int count,
-                                             MPI_Datatype datatype,
-                                             int rank,
-                                             int tag,
-                                             MPIR_Comm * comm,
-                                             int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv_init(void *buf,
+                                                    int count,
+                                                    MPI_Datatype datatype,
+                                                    int rank,
+                                                    int tag,
+                                                    MPIR_Comm * comm,
+                                                    int context_offset, MPIR_Request ** request)
 {
     MPIR_Request *rreq;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RECV_INIT);
@@ -187,10 +187,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_recv_init(void *buf,
 #define FUNCNAME MPIDI_NM_mpi_imrecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_imrecv(void *buf,
-                                          int count,
-                                          MPI_Datatype datatype,
-                                          MPIR_Request * message, MPIR_Request ** rreqp)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_imrecv(void *buf,
+                                                 int count,
+                                                 MPI_Datatype datatype,
+                                                 MPIR_Request * message, MPIR_Request ** rreqp)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq;
@@ -223,13 +223,13 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_imrecv(void *buf,
 #define FUNCNAME MPIDI_NM_mpi_irecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_irecv(void *buf,
-                                         int count,
-                                         MPI_Datatype datatype,
-                                         int rank,
-                                         int tag,
-                                         MPIR_Comm * comm, int context_offset,
-                                         MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_irecv(void *buf,
+                                                int count,
+                                                MPI_Datatype datatype,
+                                                int rank,
+                                                int tag,
+                                                MPIR_Comm * comm, int context_offset,
+                                                MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_IRECV);
@@ -244,7 +244,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_irecv(void *buf,
 #define FUNCNAME MPIDI_NM_mpi_cancel_recv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
 
     int mpi_errno = MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_rma.h b/src/mpid/ch4/netmod/ofi/ofi_rma.h
index 531c96e..7d1b87a 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_rma.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_rma.h
@@ -162,16 +162,16 @@ static inline void MPIDI_OFI_win_datatype_map(MPIDI_OFI_win_datatype_t * dt)
     MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_NETMOD_OFI_WIN_DATATYPE_MAP);
 }
 
-__ALWAYS_INLINE__ int MPIDI_OFI_allocate_win_request_put_get(MPIR_Win * win,
-                                                             int origin_count,
-                                                             int target_count,
-                                                             int target_rank,
-                                                             MPI_Datatype origin_datatype,
-                                                             MPI_Datatype target_datatype,
-                                                             MPIDI_OFI_win_request_t ** winreq,
-                                                             uint64_t * flags,
-                                                             struct fid_ep **ep,
-                                                             MPIR_Request ** sigreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_allocate_win_request_put_get(MPIR_Win * win,
+                                                                    int origin_count,
+                                                                    int target_count,
+                                                                    int target_rank,
+                                                                    MPI_Datatype origin_datatype,
+                                                                    MPI_Datatype target_datatype,
+                                                                    MPIDI_OFI_win_request_t **
+                                                                    winreq, uint64_t * flags,
+                                                                    struct fid_ep **ep,
+                                                                    MPIR_Request ** sigreq)
 {
     int mpi_errno = MPI_SUCCESS;
     size_t o_size, t_size;
@@ -200,16 +200,16 @@ __ALWAYS_INLINE__ int MPIDI_OFI_allocate_win_request_put_get(MPIR_Win * win,
     goto fn_exit;
 }
 
-__ALWAYS_INLINE__ int MPIDI_OFI_allocate_win_request_accumulate(MPIR_Win * win,
-                                                                int origin_count,
-                                                                int target_count,
-                                                                int target_rank,
-                                                                MPI_Datatype origin_datatype,
-                                                                MPI_Datatype target_datatype,
-                                                                MPIDI_OFI_win_request_t ** winreq,
-                                                                uint64_t * flags,
-                                                                struct fid_ep **ep,
-                                                                MPIR_Request ** sigreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_allocate_win_request_accumulate(MPIR_Win * win,
+                                                                       int origin_count,
+                                                                       int target_count,
+                                                                       int target_rank,
+                                                                       MPI_Datatype origin_datatype,
+                                                                       MPI_Datatype target_datatype,
+                                                                       MPIDI_OFI_win_request_t **
+                                                                       winreq, uint64_t * flags,
+                                                                       struct fid_ep **ep,
+                                                                       MPIR_Request ** sigreq)
 {
     int mpi_errno = MPI_SUCCESS;
     size_t o_size, t_size;
@@ -239,19 +239,23 @@ __ALWAYS_INLINE__ int MPIDI_OFI_allocate_win_request_accumulate(MPIR_Win * win,
     goto fn_exit;
 }
 
-__ALWAYS_INLINE__ int MPIDI_OFI_allocate_win_request_get_accumulate(MPIR_Win * win,
-                                                                    int origin_count,
-                                                                    int target_count,
-                                                                    int result_count,
-                                                                    int target_rank,
-                                                                    MPI_Op op,
-                                                                    MPI_Datatype origin_datatype,
-                                                                    MPI_Datatype target_datatype,
-                                                                    MPI_Datatype result_datatype,
-                                                                    MPIDI_OFI_win_request_t **
-                                                                    winreq, uint64_t * flags,
-                                                                    struct fid_ep **ep,
-                                                                    MPIR_Request ** sigreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_allocate_win_request_get_accumulate(MPIR_Win * win,
+                                                                           int origin_count,
+                                                                           int target_count,
+                                                                           int result_count,
+                                                                           int target_rank,
+                                                                           MPI_Op op,
+                                                                           MPI_Datatype
+                                                                           origin_datatype,
+                                                                           MPI_Datatype
+                                                                           target_datatype,
+                                                                           MPI_Datatype
+                                                                           result_datatype,
+                                                                           MPIDI_OFI_win_request_t
+                                                                           ** winreq,
+                                                                           uint64_t * flags,
+                                                                           struct fid_ep **ep,
+                                                                           MPIR_Request ** sigreq)
 {
     int mpi_errno = MPI_SUCCESS;
     size_t o_size, t_size, r_size;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_send.h b/src/mpid/ch4/netmod/ofi/ofi_send.h
index 43a7e4c..3ad10f0 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_send.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_send.h
@@ -25,10 +25,11 @@
 #define FUNCNAME MPIDI_OFI_send_lightweight
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send_lightweight(const void *buf,
-                                                 size_t data_sz,
-                                                 int rank,
-                                                 int tag, MPIR_Comm * comm, int context_offset)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send_lightweight(const void *buf,
+                                                        size_t data_sz,
+                                                        int rank,
+                                                        int tag, MPIR_Comm * comm,
+                                                        int context_offset)
 {
     int mpi_errno = MPI_SUCCESS;
     uint64_t match_bits;
@@ -55,13 +56,13 @@ __ALWAYS_INLINE__ int MPIDI_OFI_send_lightweight(const void *buf,
 #define FUNCNAME MPIDI_OFI_send_lightweight_request
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send_lightweight_request(const void *buf,
-                                                         size_t data_sz,
-                                                         int rank,
-                                                         int tag,
-                                                         MPIR_Comm * comm,
-                                                         int context_offset,
-                                                         MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send_lightweight_request(const void *buf,
+                                                                size_t data_sz,
+                                                                int rank,
+                                                                int tag,
+                                                                MPIR_Comm * comm,
+                                                                int context_offset,
+                                                                MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     uint64_t match_bits;
@@ -91,11 +92,11 @@ __ALWAYS_INLINE__ int MPIDI_OFI_send_lightweight_request(const void *buf,
 #define FUNCNAME MPIDI_OFI_send_normal
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send_normal(MPIDI_OFI_SENDPARAMS,
-                                            int dt_contig,
-                                            size_t data_sz,
-                                            MPIR_Datatype * dt_ptr,
-                                            MPI_Aint dt_true_lb, uint64_t type)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send_normal(MPIDI_OFI_SENDPARAMS,
+                                                   int dt_contig,
+                                                   size_t data_sz,
+                                                   MPIR_Datatype * dt_ptr,
+                                                   MPI_Aint dt_true_lb, uint64_t type)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq = NULL;
@@ -251,7 +252,7 @@ __ALWAYS_INLINE__ int MPIDI_OFI_send_normal(MPIDI_OFI_SENDPARAMS,
 #define FUNCNAME MPIDI_OFI_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_send(MPIDI_OFI_SENDPARAMS, int noreq, uint64_t syncflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_send(MPIDI_OFI_SENDPARAMS, int noreq, uint64_t syncflag)
 {
     int dt_contig, mpi_errno;
     size_t data_sz;
@@ -284,7 +285,7 @@ __ALWAYS_INLINE__ int MPIDI_OFI_send(MPIDI_OFI_SENDPARAMS, int noreq, uint64_t s
 #define FUNCNAME MPIDI_OFI_persistent_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_OFI_persistent_send(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_OFI_persistent_send(MPIDI_OFI_SENDPARAMS)
 {
     MPIR_Request *sreq;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_NM_PSEND);
@@ -334,7 +335,7 @@ __ALWAYS_INLINE__ int MPIDI_OFI_persistent_send(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_send(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_send(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SEND);
@@ -348,7 +349,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_send(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_rsend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_rsend(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rsend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RSEND);
@@ -363,7 +364,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_rsend(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_irsend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_irsend(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_irsend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_IRSEND);
@@ -377,7 +378,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_irsend(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_ssend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_ssend(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ssend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SSEND);
@@ -392,7 +393,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_ssend(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_isend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_isend(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_isend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_ISEND);
@@ -406,7 +407,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_isend(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_issend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_issend(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_issend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_ISSEND);
@@ -421,7 +422,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_issend(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_startall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
     int rc = MPI_SUCCESS, i;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_STARTALL);
@@ -492,7 +493,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[]
 #define FUNCNAME MPIDI_NM_mpi_send_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_send_init(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_send_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SEND_INIT);
@@ -507,7 +508,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_send_init(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_ssend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_ssend_init(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ssend_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SSEND_INIT);
@@ -522,7 +523,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_ssend_init(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_bsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_bsend_init(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_bsend_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_BSEND_INIT);
@@ -537,7 +538,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_bsend_init(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_rsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_rsend_init(MPIDI_OFI_SENDPARAMS)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rsend_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RSEND_INIT);
@@ -552,7 +553,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_rsend_init(MPIDI_OFI_SENDPARAMS)
 #define FUNCNAME MPIDI_NM_mpi_cancel_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_CANCEL_SEND);
diff --git a/src/mpid/ch4/netmod/ucx/ucx_recv.h b/src/mpid/ch4/netmod/ucx/ucx_recv.h
index e166096..4d762ad 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_recv.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_recv.h
@@ -11,12 +11,12 @@
 
 #include "ucx_impl.h"
 
-__ALWAYS_INLINE__ int ucx_irecv_continous(void *buf,
-                                          size_t data_sz,
-                                          int rank,
-                                          int tag,
-                                          MPIR_Comm * comm,
-                                          int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int ucx_irecv_continous(void *buf,
+                                                 size_t data_sz,
+                                                 int rank,
+                                                 int tag,
+                                                 MPIR_Comm * comm,
+                                                 int context_offset, MPIR_Request ** request)
 {
 
     int mpi_errno = MPI_SUCCESS;
@@ -56,13 +56,13 @@ __ALWAYS_INLINE__ int ucx_irecv_continous(void *buf,
     goto fn_exit;
 }
 
-__ALWAYS_INLINE__ int ucx_irecv_non_continous(void *buf,
-                                              size_t count,
-                                              int rank,
-                                              int tag,
-                                              MPIR_Comm * comm,
-                                              int context_offset, MPIR_Request ** request,
-                                              MPIR_Datatype * datatype)
+MPL_STATIC_INLINE_PREFIX int ucx_irecv_non_continous(void *buf,
+                                                     size_t count,
+                                                     int rank,
+                                                     int tag,
+                                                     MPIR_Comm * comm,
+                                                     int context_offset, MPIR_Request ** request,
+                                                     MPIR_Datatype * datatype)
 {
 
     int mpi_errno = MPI_SUCCESS;
@@ -134,34 +134,34 @@ static inline int do_irecv(void *buf,
 
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv(void *buf,
-                                        int count,
-                                        MPI_Datatype datatype,
-                                        int rank,
-                                        int tag,
-                                        MPIR_Comm * comm,
-                                        int context_offset,
-                                        MPI_Status * status, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv(void *buf,
+                                               int count,
+                                               MPI_Datatype datatype,
+                                               int rank,
+                                               int tag,
+                                               MPIR_Comm * comm,
+                                               int context_offset,
+                                               MPI_Status * status, MPIR_Request ** request)
 {
 
     return do_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv_init(void *buf,
-                                             int count,
-                                             MPI_Datatype datatype,
-                                             int rank,
-                                             int tag,
-                                             MPIR_Comm * comm,
-                                             int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv_init(void *buf,
+                                                    int count,
+                                                    MPI_Datatype datatype,
+                                                    int rank,
+                                                    int tag,
+                                                    MPIR_Comm * comm,
+                                                    int context_offset, MPIR_Request ** request)
 {
     return MPIDI_CH4U_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_imrecv(void *buf,
-                                          int count,
-                                          MPI_Datatype datatype,
-                                          MPIR_Request * message, MPIR_Request ** rreqp)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_imrecv(void *buf,
+                                                 int count,
+                                                 MPI_Datatype datatype,
+                                                 MPIR_Request * message, MPIR_Request ** rreqp)
 {
     ucp_tag_message_h message_handler;
     int mpi_errno = MPI_SUCCESS;
@@ -218,13 +218,13 @@ __ALWAYS_INLINE__ int MPIDI_NM_mpi_imrecv(void *buf,
     goto fn_exit;
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_mpi_irecv(void *buf,
-                                         int count,
-                                         MPI_Datatype datatype,
-                                         int rank,
-                                         int tag,
-                                         MPIR_Comm * comm, int context_offset,
-                                         MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_irecv(void *buf,
+                                                int count,
+                                                MPI_Datatype datatype,
+                                                int rank,
+                                                int tag,
+                                                MPIR_Comm * comm, int context_offset,
+                                                MPIR_Request ** request)
 {
 
 
diff --git a/src/mpid/ch4/netmod/ucx/ucx_send.h b/src/mpid/ch4/netmod/ucx/ucx_send.h
index 0217ca4..5400320 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_send.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_send.h
@@ -16,12 +16,12 @@
 #define FUNCNAME ucx_send_continous
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int ucx_send_continous(const void *buf,
-                                         size_t data_sz,
-                                         int rank,
-                                         int tag,
-                                         MPIR_Comm * comm, int context_offset,
-                                         MPIR_Request ** request, int have_request)
+MPL_STATIC_INLINE_PREFIX int ucx_send_continous(const void *buf,
+                                                size_t data_sz,
+                                                int rank,
+                                                int tag,
+                                                MPIR_Comm * comm, int context_offset,
+                                                MPIR_Request ** request, int have_request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *req;
@@ -73,12 +73,12 @@ __ALWAYS_INLINE__ int ucx_send_continous(const void *buf,
 
 }
 
-__ALWAYS_INLINE__ int ucx_sync_send_continous(const void *buf,
-                                              size_t data_sz,
-                                              int rank,
-                                              int tag,
-                                              MPIR_Comm * comm, int context_offset,
-                                              MPIR_Request ** request, int have_request)
+MPL_STATIC_INLINE_PREFIX int ucx_sync_send_continous(const void *buf,
+                                                     size_t data_sz,
+                                                     int rank,
+                                                     int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request ** request, int have_request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *req;
@@ -122,13 +122,13 @@ __ALWAYS_INLINE__ int ucx_sync_send_continous(const void *buf,
 
 }
 
-__ALWAYS_INLINE__ int ucx_sync_send_non_continous(const void *buf,
-                                                  size_t count,
-                                                  int rank,
-                                                  int tag,
-                                                  MPIR_Comm * comm, int context_offset,
-                                                  MPIR_Request ** request, int have_request,
-                                                  MPIR_Datatype * datatype)
+MPL_STATIC_INLINE_PREFIX int ucx_sync_send_non_continous(const void *buf,
+                                                         size_t count,
+                                                         int rank,
+                                                         int tag,
+                                                         MPIR_Comm * comm, int context_offset,
+                                                         MPIR_Request ** request, int have_request,
+                                                         MPIR_Datatype * datatype)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *req;
@@ -181,13 +181,13 @@ __ALWAYS_INLINE__ int ucx_sync_send_non_continous(const void *buf,
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
 
-__ALWAYS_INLINE__ int ucx_send_non_continous(const void *buf,
-                                             size_t count,
-                                             int rank,
-                                             int tag,
-                                             MPIR_Comm * comm, int context_offset,
-                                             MPIR_Request ** request, int have_request,
-                                             MPIR_Datatype * datatype)
+MPL_STATIC_INLINE_PREFIX int ucx_send_non_continous(const void *buf,
+                                                    size_t count,
+                                                    int rank,
+                                                    int tag,
+                                                    MPIR_Comm * comm, int context_offset,
+                                                    MPIR_Request ** request, int have_request,
+                                                    MPIR_Datatype * datatype)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *req;

http://git.mpich.org/mpich.git/commitdiff/7217189afc2428ce45417523015e70cae4ddcc82

commit 7217189afc2428ce45417523015e70cae4ddcc82
Author: Pavan Balaji <balaji at anl.gov>
Date:   Tue Aug 30 10:42:55 2016 -0500

    Namespace certain functions as hooks
    
    Some existing functions are more accurately described as hooks,
    providing functionality beyond what the upper-layer code is already
    doing. In this commit we rename the CH4 init/finalize routines and
    device layer request create/free routines.
    
    Signed-off-by: Ken Raffenetti <raffenet at mcs.anl.gov>

diff --git a/src/include/mpir_request.h b/src/include/mpir_request.h
index af212ed..7085904 100644
--- a/src/include/mpir_request.h
+++ b/src/include/mpir_request.h
@@ -224,7 +224,7 @@ static inline MPIR_Request *MPIR_Request_create(MPIR_Request_kind_t kind)
             break;
         }
 
-        MPID_Request_init(req);
+        MPID_Request_create_hook(req);
     }
     else
     {
@@ -281,7 +281,7 @@ static inline void MPIR_Request_free(MPIR_Request *req)
             MPL_free(req->u.ureq.greq_fns);
         }
 
-        MPID_Request_finalize(req);
+        MPID_Request_free_hook(req);
 
         MPIR_Handle_obj_free(&MPIR_Request_mem, req);
     }
diff --git a/src/mpid/ch3/include/mpidpre.h b/src/mpid/ch3/include/mpidpre.h
index 926612e..edd7ec6 100644
--- a/src/mpid/ch3/include/mpidpre.h
+++ b/src/mpid/ch3/include/mpidpre.h
@@ -693,8 +693,8 @@ int MPID_Get_processor_name( char *name, int namelen, int *resultlen);
 int MPID_Get_universe_size(int  * universe_size);
 int MPID_Comm_get_lpid(MPIR_Comm *comm_ptr, int idx, int * lpid_ptr, MPL_bool is_remote);
 
-void MPID_Request_init(MPIR_Request *);
-void MPID_Request_finalize(MPIR_Request *);
+void MPID_Request_create_hook(MPIR_Request *);
+void MPID_Request_free_hook(MPIR_Request *);
 int MPID_Request_complete(MPIR_Request *);
 
 void *MPID_Alloc_mem( size_t size, MPIR_Info *info );
diff --git a/src/mpid/ch3/src/ch3u_request.c b/src/mpid/ch3/src/ch3u_request.c
index 2fec8f3..5176fc3 100644
--- a/src/mpid/ch3/src/ch3u_request.c
+++ b/src/mpid/ch3/src/ch3u_request.c
@@ -24,10 +24,10 @@
 /* See the comments above about request creation.  Some routines will
    use macros in mpidimpl.h *instead* of this routine */
 #undef FUNCNAME
-#define FUNCNAME MPID_Request_init
+#define FUNCNAME MPID_Request_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-void MPID_Request_init(MPIR_Request *req)
+void MPID_Request_create_hook(MPIR_Request *req)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_REQUEST_INIT);
 
@@ -588,7 +588,7 @@ int MPID_Request_complete(MPIR_Request *req)
     goto fn_exit;
 }
 
-void MPID_Request_finalize(MPIR_Request *req)
+void MPID_Request_free_hook(MPIR_Request *req)
 {
     if (req->dev.datatype_ptr != NULL) {
         MPIDU_Datatype_release(req->dev.datatype_ptr);
diff --git a/src/mpid/ch4/include/mpidpost.h b/src/mpid/ch4/include/mpidpost.h
index 40edeb7..43a9658 100644
--- a/src/mpid/ch4/include/mpidpost.h
+++ b/src/mpid/ch4/include/mpidpost.h
@@ -14,7 +14,7 @@
 #include "mpidu_datatype.h"
 #include "mpidch4.h"
 
-__ALWAYS_INLINE__ void MPID_Request_init(MPIR_Request * req)
+__ALWAYS_INLINE__ void MPID_Request_create_hook(MPIR_Request * req)
 {
     MPIDI_CH4U_REQUEST(req, req) = NULL;
 #ifdef MPIDI_BUILD_CH4_SHM
@@ -22,7 +22,7 @@ __ALWAYS_INLINE__ void MPID_Request_init(MPIR_Request * req)
 #endif
 }
 
-__ALWAYS_INLINE__ void MPID_Request_finalize(MPIR_Request * req)
+__ALWAYS_INLINE__ void MPID_Request_free_hook(MPIR_Request * req)
 {
     return;
 }
diff --git a/src/mpid/ch4/include/mpidpre.h b/src/mpid/ch4/include/mpidpre.h
index 1a045ca..698cd4a 100644
--- a/src/mpid/ch4/include/mpidpre.h
+++ b/src/mpid/ch4/include/mpidpre.h
@@ -198,9 +198,9 @@ typedef struct {
 #define MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(req)  NULL
 #endif
 
-__ALWAYS_INLINE__ void MPID_Request_init(struct MPIR_Request *req);
+__ALWAYS_INLINE__ void MPID_Request_create_hook(struct MPIR_Request *req);
 
-__ALWAYS_INLINE__ void MPID_Request_finalize(struct MPIR_Request *req);
+__ALWAYS_INLINE__ void MPID_Request_free_hook(struct MPIR_Request *req);
 
 typedef struct MPIDI_CH4U_win_shared_info {
     uint32_t disp_unit;
diff --git a/src/mpid/ch4/netmod/include/netmod.h b/src/mpid/ch4/netmod/include/netmod.h
index 2a8f621..12b8ba2 100644
--- a/src/mpid/ch4/netmod/include/netmod.h
+++ b/src/mpid/ch4/netmod/include/netmod.h
@@ -504,11 +504,12 @@ extern MPIDI_NM_native_funcs_t *MPIDI_NM_native_func;
 extern int MPIDI_num_netmods;
 extern char MPIDI_NM_strings[][MPIDI_MAX_NETMOD_STRING_LEN];
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_init(int rank, int size, int appnum, int *tag_ub,
-                                               MPIR_Comm * comm_world, MPIR_Comm * comm_self,
-                                               int spawned, int num_contexts,
-                                               void **netmod_contexts) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_finalize(void) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_init_hook(int rank, int size, int appnum, int *tag_ub,
+                                                    MPIR_Comm * comm_world, MPIR_Comm * comm_self,
+                                                    int spawned, int num_contexts,
+                                                    void **netmod_contexts)
+    MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_finalize_hook(void) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_progress(void *netmod_context,
                                                int blocking) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_reg_handler(int handler_id,
diff --git a/src/mpid/ch4/netmod/include/netmod_impl.h b/src/mpid/ch4/netmod/include/netmod_impl.h
index ad16dbc..e2f77ad 100644
--- a/src/mpid/ch4/netmod/include/netmod_impl.h
+++ b/src/mpid/ch4/netmod/include/netmod_impl.h
@@ -16,16 +16,16 @@
 #ifndef NETMOD_DIRECT
 #ifndef NETMOD_DISABLE_INLINES
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_init(int rank, int size, int appnum, int *tag_ub,
-                                               MPIR_Comm * comm_world, MPIR_Comm * comm_self,
-                                               int spawned, int num_contexts,
-                                               void **netmod_contexts)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_init_hook(int rank, int size, int appnum, int *tag_ub,
+                                                    MPIR_Comm * comm_world, MPIR_Comm * comm_self,
+                                                    int spawned, int num_contexts,
+                                                    void **netmod_contexts)
 {
     return MPIDI_NM_func->mpi_init(rank, size, appnum, tag_ub, comm_world, comm_self, spawned,
                                    num_contexts, netmod_contexts);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_finalize(void)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_finalize_hook(void)
 {
     return MPIDI_NM_func->mpi_finalize();
 };
diff --git a/src/mpid/ch4/netmod/ofi/func_table.c b/src/mpid/ch4/netmod/ofi/func_table.c
index f8789f2..bd524ed 100644
--- a/src/mpid/ch4/netmod/ofi/func_table.c
+++ b/src/mpid/ch4/netmod/ofi/func_table.c
@@ -14,8 +14,8 @@
 #include <mpidimpl.h>
 #include "netmod_direct.h"
 MPIDI_NM_funcs_t MPIDI_NM_ofi_funcs = {
-    MPIDI_NM_mpi_init,
-    MPIDI_NM_mpi_finalize,
+    MPIDI_NM_mpi_init_hook,
+    MPIDI_NM_mpi_finalize_hook,
     MPIDI_NM_progress,
     MPIDI_NM_mpi_comm_connect,
     MPIDI_NM_mpi_comm_disconnect,
diff --git a/src/mpid/ch4/netmod/ofi/ofi_init.h b/src/mpid/ch4/netmod/ofi/ofi_init.h
index 9046bf2..be23ba0 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_init.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_init.h
@@ -482,16 +482,16 @@ static inline int MPIDI_OFI_init_generic(int rank,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_mpi_init
+#define FUNCNAME MPIDI_NM_mpi_init_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_mpi_init(int rank,
-                                    int size,
-                                    int appnum,
-                                    int *tag_ub,
-                                    MPIR_Comm * comm_world,
-                                    MPIR_Comm * comm_self,
-                                    int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init_hook(int rank,
+                                         int size,
+                                         int appnum,
+                                         int *tag_ub,
+                                         MPIR_Comm * comm_world,
+                                         MPIR_Comm * comm_self,
+                                         int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno;
     mpi_errno = MPIDI_OFI_init_generic(rank, size, appnum, tag_ub, comm_world,
@@ -589,7 +589,7 @@ static inline int MPIDI_OFI_finalize_generic(int do_scalable_ep, int do_am, int
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_mpi_finalize(void)
+static inline int MPIDI_NM_mpi_finalize_hook(void)
 {
     return MPIDI_OFI_finalize_generic(MPIDI_OFI_ENABLE_SCALABLE_ENDPOINTS,
                                       MPIDI_OFI_ENABLE_AM, MPIDI_OFI_ENABLE_STX_RMA);
diff --git a/src/mpid/ch4/netmod/portals4/func_table.c b/src/mpid/ch4/netmod/portals4/func_table.c
index e707315..5a4e501 100644
--- a/src/mpid/ch4/netmod/portals4/func_table.c
+++ b/src/mpid/ch4/netmod/portals4/func_table.c
@@ -14,8 +14,8 @@
 #include "netmod_direct.h"
 
 MPIDI_NM_funcs_t MPIDI_NM_portals4_funcs = {
-    MPIDI_NM_mpi_init,
-    MPIDI_NM_mpi_finalize,
+    MPIDI_NM_mpi_init_hook,
+    MPIDI_NM_mpi_finalize_hook,
     MPIDI_NM_progress,
     MPIDI_NM_mpi_comm_connect,
     MPIDI_NM_mpi_comm_disconnect,
diff --git a/src/mpid/ch4/netmod/portals4/ptl_init.h b/src/mpid/ch4/netmod/portals4/ptl_init.h
index 06b1f51..5766515 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_init.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_init.h
@@ -40,16 +40,16 @@ static inline int MPIDI_PTL_append_overflow(size_t i)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_mpi_init
+#define FUNCNAME MPIDI_NM_mpi_init_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_mpi_init(int rank,
-                                    int size,
-                                    int appnum,
-                                    int *tag_ub,
-                                    MPIR_Comm * comm_world,
-                                    MPIR_Comm * comm_self,
-                                    int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init_hook(int rank,
+                                         int size,
+                                         int appnum,
+                                         int *tag_ub,
+                                         MPIR_Comm * comm_world,
+                                         MPIR_Comm * comm_self,
+                                         int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS;
     int ret;
@@ -183,7 +183,7 @@ static inline int MPIDI_NM_mpi_init(int rank,
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_mpi_finalize(void)
+static inline int MPIDI_NM_mpi_finalize_hook(void)
 {
     int mpi_errno = MPI_SUCCESS;
     int ret, i;
diff --git a/src/mpid/ch4/netmod/stubnm/globals.c b/src/mpid/ch4/netmod/stubnm/globals.c
index f87b80b..bfb95c2 100644
--- a/src/mpid/ch4/netmod/stubnm/globals.c
+++ b/src/mpid/ch4/netmod/stubnm/globals.c
@@ -14,8 +14,8 @@
 #include <mpidimpl.h>
 #include "netmod_direct.h"
 MPIDI_NM_funcs_t MPIDI_NM_stubnm_funcs = {
-    MPIDI_NM_mpi_init,
-    MPIDI_NM_mpi_finalize,
+    MPIDI_NM_mpi_init_hook,
+    MPIDI_NM_mpi_finalize_hook,
     MPIDI_NM_progress,
     MPIDI_NM_mpi_comm_connect,
     MPIDI_NM_mpi_comm_disconnect,
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_init.h b/src/mpid/ch4/netmod/stubnm/stubnm_init.h
index fbe2d18..ca58e7e 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_init.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_init.h
@@ -13,13 +13,13 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_mpi_init(int rank,
-                                    int size,
-                                    int appnum,
-                                    int *tag_ub,
-                                    MPIR_Comm * comm_world,
-                                    MPIR_Comm * comm_self,
-                                    int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init_hook(int rank,
+                                         int size,
+                                         int appnum,
+                                         int *tag_ub,
+                                         MPIR_Comm * comm_world,
+                                         MPIR_Comm * comm_self,
+                                         int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -28,7 +28,7 @@ static inline int MPIDI_NM_mpi_init(int rank,
     return mpi_errno;
 }
 
-static inline int MPIDI_NM_mpi_finalize(void)
+static inline int MPIDI_NM_mpi_finalize_hook(void)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
diff --git a/src/mpid/ch4/netmod/ucx/func_table.c b/src/mpid/ch4/netmod/ucx/func_table.c
index aa2a904..404b891 100644
--- a/src/mpid/ch4/netmod/ucx/func_table.c
+++ b/src/mpid/ch4/netmod/ucx/func_table.c
@@ -12,8 +12,8 @@
 #include <mpidimpl.h>
 #include "netmod_direct.h"
 MPIDI_NM_funcs_t MPIDI_NM_ucx_funcs = {
-    MPIDI_NM_mpi_init,
-    MPIDI_NM_mpi_finalize,
+    MPIDI_NM_mpi_init_hook,
+    MPIDI_NM_mpi_finalize_hook,
     MPIDI_NM_progress,
     MPIDI_NM_mpi_comm_connect,
     MPIDI_NM_mpi_comm_disconnect,
diff --git a/src/mpid/ch4/netmod/ucx/ucx_init.h b/src/mpid/ch4/netmod/ucx/ucx_init.h
index acbd31f..23354ef 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_init.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_init.h
@@ -15,16 +15,16 @@
 #include "pmi.h"
 #include <ucp/api/ucp.h>
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_mpi_init
+#define FUNCNAME MPIDI_NM_mpi_init_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_mpi_init(int rank,
-                                    int size,
-                                    int appnum,
-                                    int *tag_ub,
-                                    MPIR_Comm * comm_world,
-                                    MPIR_Comm * comm_self,
-                                    int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init_hook(int rank,
+                                         int size,
+                                         int appnum,
+                                         int *tag_ub,
+                                         MPIR_Comm * comm_world,
+                                         MPIR_Comm * comm_self,
+                                         int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS, thr_err, pmi_errno;
     int str_errno = MPL_STR_SUCCESS;
@@ -127,7 +127,7 @@ static inline int MPIDI_NM_mpi_init(int rank,
 
 }
 
-static inline int MPIDI_NM_mpi_finalize(void)
+static inline int MPIDI_NM_mpi_finalize_hook(void)
 {
     int mpi_errno = MPI_SUCCESS, thr_err, pmi_errno;
     int i, j, max_n_avts;
diff --git a/src/mpid/ch4/shm/include/shm.h b/src/mpid/ch4/shm/include/shm.h
index 597922e..a87940e 100644
--- a/src/mpid/ch4/shm/include/shm.h
+++ b/src/mpid/ch4/shm/include/shm.h
@@ -527,8 +527,8 @@ extern MPIDI_SHM_native_funcs_t *MPIDI_SHM_native_func;
 extern int MPIDI_num_shms;
 extern char MPIDI_SHM_strings[][MPIDI_MAX_SHM_STRING_LEN];
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_init(int rank, int size) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_finalize(void) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_init_hook(int rank, int size) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_finalize_hook(void) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_progress(int blocking) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_reg_handler(int handler_id,
                                                       MPIDI_SHM_am_origin_handler_fn
diff --git a/src/mpid/ch4/shm/include/shm_impl.h b/src/mpid/ch4/shm/include/shm_impl.h
index ad23af0..a15f4d4 100644
--- a/src/mpid/ch4/shm/include/shm_impl.h
+++ b/src/mpid/ch4/shm/include/shm_impl.h
@@ -15,12 +15,12 @@
 #ifndef SHM_DIRECT
 #ifndef SHM_DISABLE_INLINES
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_init(int rank, int size)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_init_hook(int rank, int size)
 {
     return MPIDI_SHM_func->mpi_init(rank, size);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_finalize(void)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_finalize_hook(void)
 {
     return MPIDI_SHM_func->mpi_finalize();
 };
diff --git a/src/mpid/ch4/shm/posix/func_table.c b/src/mpid/ch4/shm/posix/func_table.c
index 24e9709..6e0e1f4 100644
--- a/src/mpid/ch4/shm/posix/func_table.c
+++ b/src/mpid/ch4/shm/posix/func_table.c
@@ -13,8 +13,8 @@
 #include <mpidimpl.h>
 #include "shm_direct.h"
 MPIDI_SHM_funcs_t MPIDI_SHM_posix_funcs = {
-    MPIDI_SHM_mpi_init,
-    MPIDI_SHM_mpi_finalize,
+    MPIDI_SHM_mpi_init_hook,
+    MPIDI_SHM_mpi_finalize_hook,
     MPIDI_SHM_progress,
     MPIDI_SHM_am_reg_handler,
     MPIDI_SHM_mpi_comm_connect,
diff --git a/src/mpid/ch4/shm/posix/posix_init.h b/src/mpid/ch4/shm/posix/posix_init.h
index 1f2754b..eb47bb6 100644
--- a/src/mpid/ch4/shm/posix/posix_init.h
+++ b/src/mpid/ch4/shm/posix/posix_init.h
@@ -22,8 +22,8 @@ extern MPIDI_POSIX_mem_region_t MPIDI_POSIX_mem_region;
 extern char *MPIDI_POSIX_asym_base_addr;
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_init)
-static inline int MPIDI_SHM_mpi_init(int rank, int size)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_init_hook)
+static inline int MPIDI_SHM_mpi_init_hook(int rank, int size)
 {
     int mpi_errno = MPI_SUCCESS;
     int num_local = 0;
@@ -229,8 +229,8 @@ static inline int MPIDI_SHM_mpi_init(int rank, int size)
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_finalize)
-static inline int MPIDI_SHM_mpi_finalize(void)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_finalize_hook)
+static inline int MPIDI_SHM_mpi_finalize_hook(void)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_SHM_FINALIZE);
diff --git a/src/mpid/ch4/shm/stubshm/func_table.c b/src/mpid/ch4/shm/stubshm/func_table.c
index acb9672..7e4e2c1 100644
--- a/src/mpid/ch4/shm/stubshm/func_table.c
+++ b/src/mpid/ch4/shm/stubshm/func_table.c
@@ -13,8 +13,8 @@
 #include <mpidimpl.h>
 #include "shm_direct.h"
 MPIDI_SHM_funcs_t MPIDI_SHM_stubshm_funcs = {
-    MPIDI_SHM_mpi_init,
-    MPIDI_SHM_mpi_finalize,
+    MPIDI_SHM_mpi_init_hook,
+    MPIDI_SHM_mpi_finalize_hook,
     MPIDI_SHM_progress,
     MPIDI_SHM_am_reg_handler,
     MPIDI_SHM_mpi_comm_connect,
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_init.h b/src/mpid/ch4/shm/stubshm/stubshm_init.h
index 6b61d99..f4e8da8 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_init.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_init.h
@@ -13,13 +13,13 @@
 
 #include "stubshm_impl.h"
 
-static inline int MPIDI_SHM_mpi_init(int rank, int size)
+static inline int MPIDI_SHM_mpi_init_hook(int rank, int size)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_mpi_finalize(void)
+static inline int MPIDI_SHM_mpi_finalize_hook(void)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/src/ch4_init.h b/src/mpid/ch4/src/ch4_init.h
index 6613e82..9b40821 100644
--- a/src/mpid/ch4/src/ch4_init.h
+++ b/src/mpid/ch4/src/ch4_init.h
@@ -248,9 +248,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Init(int *argc,
     MPIR_Process.attrs.tag_ub = (1ULL << MPIDI_CH4U_TAG_SHIFT) - 1;
     /* discuss */
 
-    mpi_errno = MPIDI_NM_mpi_init(rank, size, appnum, &MPIR_Process.attrs.tag_ub,
-                                  MPIR_Process.comm_world,
-                                  MPIR_Process.comm_self, has_parent, 1, &netmod_contexts);
+    mpi_errno = MPIDI_NM_mpi_init_hook(rank, size, appnum, &MPIR_Process.attrs.tag_ub,
+                                       MPIR_Process.comm_world,
+                                       MPIR_Process.comm_self, has_parent, 1, &netmod_contexts);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POPFATAL(mpi_errno);
     }
@@ -279,7 +279,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Init(int *argc,
 #endif
 
 #ifdef MPIDI_BUILD_CH4_SHM
-    mpi_errno = MPIDI_SHM_mpi_init(rank, size);
+    mpi_errno = MPIDI_SHM_mpi_init_hook(rank, size);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POPFATAL(mpi_errno);
@@ -346,11 +346,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Finalize(void)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_FINALIZE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_FINALIZE);
 
-    mpi_errno = MPIDI_NM_mpi_finalize();
+    mpi_errno = MPIDI_NM_mpi_finalize_hook();
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 #ifdef MPIDI_BUILD_CH4_SHM
-    mpi_errno = MPIDI_SHM_mpi_finalize();
+    mpi_errno = MPIDI_SHM_mpi_finalize_hook();
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 #endif

http://git.mpich.org/mpich.git/commitdiff/6370ba6ce11942523e944f324298b6fe5209b4f5

commit 6370ba6ce11942523e944f324298b6fe5209b4f5
Author: Pavan Balaji <balaji at anl.gov>
Date:   Tue Aug 30 10:42:55 2016 -0500

    CH4: Namespace MPI-level functionality used by the netmod/shmmod
    
    Signed-off-by: Ken Raffenetti <raffenet at mcs.anl.gov>

diff --git a/src/mpid/ch4/netmod/include/netmod.h b/src/mpid/ch4/netmod/include/netmod.h
index e9d4e58..2a8f621 100644
--- a/src/mpid/ch4/netmod/include/netmod.h
+++ b/src/mpid/ch4/netmod/include/netmod.h
@@ -26,21 +26,21 @@ typedef int (*MPIDI_NM_am_target_handler_fn)
   size_t * data_sz, int *is_contig, MPIDI_NM_am_completion_handler_fn * cmpl_handler_fn,        /* completion handler */
   MPIR_Request ** req);         /* if allocated, need pointer to completion function */
 
-typedef int (*MPIDI_NM_init_t) (int rank, int size, int appnum, int *tag_ub, MPIR_Comm * comm_world,
-                                MPIR_Comm * comm_self, int spawned, int num_contexts,
-                                void **netmod_contexts);
-typedef int (*MPIDI_NM_finalize_t) (void);
+typedef int (*MPIDI_NM_mpi_init_t) (int rank, int size, int appnum, int *tag_ub,
+                                    MPIR_Comm * comm_world, MPIR_Comm * comm_self, int spawned,
+                                    int num_contexts, void **netmod_contexts);
+typedef int (*MPIDI_NM_mpi_finalize_t) (void);
 typedef int (*MPIDI_NM_progress_t) (void *netmod_context, int blocking);
 typedef int (*MPIDI_NM_am_reg_handler_t) (int handler_id,
                                           MPIDI_NM_am_origin_handler_fn origin_handler_fn,
                                           MPIDI_NM_am_target_handler_fn target_handler_fn);
-typedef int (*MPIDI_NM_comm_connect_t) (const char *port_name, MPIR_Info * info, int root,
-                                        MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
-typedef int (*MPIDI_NM_comm_disconnect_t) (MPIR_Comm * comm_ptr);
-typedef int (*MPIDI_NM_open_port_t) (MPIR_Info * info_ptr, char *port_name);
-typedef int (*MPIDI_NM_close_port_t) (const char *port_name);
-typedef int (*MPIDI_NM_comm_accept_t) (const char *port_name, MPIR_Info * info, int root,
-                                       MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
+typedef int (*MPIDI_NM_mpi_comm_connect_t) (const char *port_name, MPIR_Info * info, int root,
+                                            MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
+typedef int (*MPIDI_NM_mpi_comm_disconnect_t) (MPIR_Comm * comm_ptr);
+typedef int (*MPIDI_NM_mpi_open_port_t) (MPIR_Info * info_ptr, char *port_name);
+typedef int (*MPIDI_NM_mpi_close_port_t) (const char *port_name);
+typedef int (*MPIDI_NM_mpi_comm_accept_t) (const char *port_name, MPIR_Info * info, int root,
+                                           MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
 typedef int (*MPIDI_NM_am_send_hdr_t) (int rank, MPIR_Comm * comm, int handler_id,
                                        const void *am_hdr, size_t am_hdr_sz, void *netmod_context);
 typedef int (*MPIDI_NM_am_isend_t) (int rank, MPIR_Comm * comm, int handler_id, const void *am_hdr,
@@ -67,296 +67,313 @@ typedef int (*MPIDI_NM_getallincomm_t) (MPIR_Comm * comm_ptr, int local_size,
 typedef int (*MPIDI_NM_gpid_tolpidarray_t) (int size, MPIR_Gpid gpid[], int lpid[]);
 typedef int (*MPIDI_NM_create_intercomm_from_lpids_t) (MPIR_Comm * newcomm_ptr, int size,
                                                        const int lpids[]);
-typedef int (*MPIDI_NM_comm_create_hook_t) (MPIR_Comm * comm);
-typedef int (*MPIDI_NM_comm_free_hook_t) (MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_comm_create_hook_t) (MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_comm_free_hook_t) (MPIR_Comm * comm);
 typedef void (*MPIDI_NM_am_request_init_t) (MPIR_Request * req);
 typedef void (*MPIDI_NM_am_request_finalize_t) (MPIR_Request * req);
-typedef int (*MPIDI_NM_send_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                int tag, MPIR_Comm * comm, int context_offset,
-                                MPIR_Request ** request);
-typedef int (*MPIDI_NM_ssend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                 int tag, MPIR_Comm * comm, int context_offset,
-                                 MPIR_Request ** request);
-typedef int (*MPIDI_NM_startall_t) (int count, MPIR_Request * requests[]);
-typedef int (*MPIDI_NM_send_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+typedef int (*MPIDI_NM_mpi_send_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+                                    int tag, MPIR_Comm * comm, int context_offset,
+                                    MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_ssend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
                                      int tag, MPIR_Comm * comm, int context_offset,
                                      MPIR_Request ** request);
-typedef int (*MPIDI_NM_ssend_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                      int tag, MPIR_Comm * comm, int context_offset,
-                                      MPIR_Request ** request);
-typedef int (*MPIDI_NM_rsend_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                      int tag, MPIR_Comm * comm, int context_offset,
-                                      MPIR_Request ** request);
-typedef int (*MPIDI_NM_bsend_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+typedef int (*MPIDI_NM_mpi_startall_t) (int count, MPIR_Request * requests[]);
+typedef int (*MPIDI_NM_mpi_send_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                         int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_ssend_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                          int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_rsend_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                          int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_bsend_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                          int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_isend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+                                     int tag, MPIR_Comm * comm, int context_offset,
+                                     MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_issend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
                                       int tag, MPIR_Comm * comm, int context_offset,
                                       MPIR_Request ** request);
-typedef int (*MPIDI_NM_isend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                 int tag, MPIR_Comm * comm, int context_offset,
-                                 MPIR_Request ** request);
-typedef int (*MPIDI_NM_issend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                  int tag, MPIR_Comm * comm, int context_offset,
-                                  MPIR_Request ** request);
-typedef int (*MPIDI_NM_cancel_send_t) (MPIR_Request * sreq);
-typedef int (*MPIDI_NM_recv_init_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
+typedef int (*MPIDI_NM_mpi_cancel_send_t) (MPIR_Request * sreq);
+typedef int (*MPIDI_NM_mpi_recv_init_t) (void *buf, int count, MPI_Datatype datatype, int rank,
+                                         int tag, MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_recv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
+                                    MPIR_Comm * comm, int context_offset, MPI_Status * status,
+                                    MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_irecv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request);
-typedef int (*MPIDI_NM_recv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
-                                MPIR_Comm * comm, int context_offset, MPI_Status * status,
-                                MPIR_Request ** request);
-typedef int (*MPIDI_NM_irecv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request);
-typedef int (*MPIDI_NM_imrecv_t) (void *buf, int count, MPI_Datatype datatype,
-                                  MPIR_Request * message, MPIR_Request ** rreqp);
-typedef int (*MPIDI_NM_cancel_recv_t) (MPIR_Request * rreq);
-typedef void *(*MPIDI_NM_alloc_mem_t) (size_t size, MPIR_Info * info_ptr);
-typedef int (*MPIDI_NM_free_mem_t) (void *ptr);
-typedef int (*MPIDI_NM_improbe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
-                                   int *flag, MPIR_Request ** message, MPI_Status * status);
-typedef int (*MPIDI_NM_iprobe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
-                                  int *flag, MPI_Status * status);
-typedef int (*MPIDI_NM_win_set_info_t) (MPIR_Win * win, MPIR_Info * info);
-typedef int (*MPIDI_NM_win_shared_query_t) (MPIR_Win * win, int rank, MPI_Aint * size,
-                                            int *disp_unit, void *baseptr);
-typedef int (*MPIDI_NM_put_t) (const void *origin_addr, int origin_count,
-                               MPI_Datatype origin_datatype, int target_rank, MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_start_t) (MPIR_Group * group, int assert, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_complete_t) (MPIR_Win * win);
-typedef int (*MPIDI_NM_win_post_t) (MPIR_Group * group, int assert, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_wait_t) (MPIR_Win * win);
-typedef int (*MPIDI_NM_win_test_t) (MPIR_Win * win, int *flag);
-typedef int (*MPIDI_NM_win_lock_t) (int lock_type, int rank, int assert, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_unlock_t) (int rank, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_get_info_t) (MPIR_Win * win, MPIR_Info ** info_p_p);
-typedef int (*MPIDI_NM_get_t) (void *origin_addr, int origin_count, MPI_Datatype origin_datatype,
-                               int target_rank, MPI_Aint target_disp, int target_count,
-                               MPI_Datatype target_datatype, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_free_t) (MPIR_Win ** win_ptr);
-typedef int (*MPIDI_NM_win_fence_t) (int assert, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_create_t) (void *base, MPI_Aint length, int disp_unit, MPIR_Info * info,
-                                      MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr);
-typedef int (*MPIDI_NM_accumulate_t) (const void *origin_addr, int origin_count,
-                                      MPI_Datatype origin_datatype, int target_rank,
-                                      MPI_Aint target_disp, int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_attach_t) (MPIR_Win * win, void *base, MPI_Aint size);
-typedef int (*MPIDI_NM_win_allocate_shared_t) (MPI_Aint size, int disp_unit, MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr, void **base_ptr,
-                                               MPIR_Win ** win_ptr);
-typedef int (*MPIDI_NM_rput_t) (const void *origin_addr, int origin_count,
-                                MPI_Datatype origin_datatype, int target_rank, MPI_Aint target_disp,
-                                int target_count, MPI_Datatype target_datatype, MPIR_Win * win,
-                                MPIR_Request ** request);
-typedef int (*MPIDI_NM_win_flush_local_t) (int rank, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_detach_t) (MPIR_Win * win, const void *base);
-typedef int (*MPIDI_NM_compare_and_swap_t) (const void *origin_addr, const void *compare_addr,
-                                            void *result_addr, MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win);
-typedef int (*MPIDI_NM_raccumulate_t) (const void *origin_addr, int origin_count,
-                                       MPI_Datatype origin_datatype, int target_rank,
-                                       MPI_Aint target_disp, int target_count,
-                                       MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win,
-                                       MPIR_Request ** request);
-typedef int (*MPIDI_NM_rget_accumulate_t) (const void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, void *result_addr,
-                                           int result_count, MPI_Datatype result_datatype,
-                                           int target_rank, MPI_Aint target_disp, int target_count,
+typedef int (*MPIDI_NM_mpi_imrecv_t) (void *buf, int count, MPI_Datatype datatype,
+                                      MPIR_Request * message, MPIR_Request ** rreqp);
+typedef int (*MPIDI_NM_mpi_cancel_recv_t) (MPIR_Request * rreq);
+typedef void *(*MPIDI_NM_mpi_alloc_mem_t) (size_t size, MPIR_Info * info_ptr);
+typedef int (*MPIDI_NM_mpi_free_mem_t) (void *ptr);
+typedef int (*MPIDI_NM_mpi_improbe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
+                                       int *flag, MPIR_Request ** message, MPI_Status * status);
+typedef int (*MPIDI_NM_mpi_iprobe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
+                                      int *flag, MPI_Status * status);
+typedef int (*MPIDI_NM_mpi_win_set_info_t) (MPIR_Win * win, MPIR_Info * info);
+typedef int (*MPIDI_NM_mpi_win_shared_query_t) (MPIR_Win * win, int rank, MPI_Aint * size,
+                                                int *disp_unit, void *baseptr);
+typedef int (*MPIDI_NM_mpi_put_t) (const void *origin_addr, int origin_count,
+                                   MPI_Datatype origin_datatype, int target_rank,
+                                   MPI_Aint target_disp, int target_count,
+                                   MPI_Datatype target_datatype, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_start_t) (MPIR_Group * group, int assert, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_complete_t) (MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_post_t) (MPIR_Group * group, int assert, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_wait_t) (MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_test_t) (MPIR_Win * win, int *flag);
+typedef int (*MPIDI_NM_mpi_win_lock_t) (int lock_type, int rank, int assert, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_unlock_t) (int rank, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_get_info_t) (MPIR_Win * win, MPIR_Info ** info_p_p);
+typedef int (*MPIDI_NM_mpi_get_t) (void *origin_addr, int origin_count,
+                                   MPI_Datatype origin_datatype, int target_rank,
+                                   MPI_Aint target_disp, int target_count,
+                                   MPI_Datatype target_datatype, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_free_t) (MPIR_Win ** win_ptr);
+typedef int (*MPIDI_NM_mpi_win_fence_t) (int assert, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_create_t) (void *base, MPI_Aint length, int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr);
+typedef int (*MPIDI_NM_mpi_accumulate_t) (const void *origin_addr, int origin_count,
+                                          MPI_Datatype origin_datatype, int target_rank,
+                                          MPI_Aint target_disp, int target_count,
+                                          MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_attach_t) (MPIR_Win * win, void *base, MPI_Aint size);
+typedef int (*MPIDI_NM_mpi_win_allocate_shared_t) (MPI_Aint size, int disp_unit,
+                                                   MPIR_Info * info_ptr, MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr);
+typedef int (*MPIDI_NM_mpi_rput_t) (const void *origin_addr, int origin_count,
+                                    MPI_Datatype origin_datatype, int target_rank,
+                                    MPI_Aint target_disp, int target_count,
+                                    MPI_Datatype target_datatype, MPIR_Win * win,
+                                    MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_win_flush_local_t) (int rank, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_detach_t) (MPIR_Win * win, const void *base);
+typedef int (*MPIDI_NM_mpi_compare_and_swap_t) (const void *origin_addr, const void *compare_addr,
+                                                void *result_addr, MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_raccumulate_t) (const void *origin_addr, int origin_count,
+                                           MPI_Datatype origin_datatype, int target_rank,
+                                           MPI_Aint target_disp, int target_count,
                                            MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win,
                                            MPIR_Request ** request);
-typedef int (*MPIDI_NM_fetch_and_op_t) (const void *origin_addr, void *result_addr,
-                                        MPI_Datatype datatype, int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_allocate_t) (MPI_Aint size, int disp_unit, MPIR_Info * info,
-                                        MPIR_Comm * comm, void *baseptr, MPIR_Win ** win);
-typedef int (*MPIDI_NM_win_flush_t) (int rank, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_flush_local_all_t) (MPIR_Win * win);
-typedef int (*MPIDI_NM_win_unlock_all_t) (MPIR_Win * win);
-typedef int (*MPIDI_NM_win_create_dynamic_t) (MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win);
-typedef int (*MPIDI_NM_rget_t) (void *origin_addr, int origin_count, MPI_Datatype origin_datatype,
-                                int target_rank, MPI_Aint target_disp, int target_count,
-                                MPI_Datatype target_datatype, MPIR_Win * win,
-                                MPIR_Request ** request);
-typedef int (*MPIDI_NM_win_sync_t) (MPIR_Win * win);
-typedef int (*MPIDI_NM_win_flush_all_t) (MPIR_Win * win);
-typedef int (*MPIDI_NM_get_accumulate_t) (const void *origin_addr, int origin_count,
-                                          MPI_Datatype origin_datatype, void *result_addr,
-                                          int result_count, MPI_Datatype result_datatype,
-                                          int target_rank, MPI_Aint target_disp, int target_count,
-                                          MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win);
-typedef int (*MPIDI_NM_win_lock_all_t) (int assert, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_rget_accumulate_t) (const void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, void *result_addr,
+                                               int result_count, MPI_Datatype result_datatype,
+                                               int target_rank, MPI_Aint target_disp,
+                                               int target_count, MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_fetch_and_op_t) (const void *origin_addr, void *result_addr,
+                                            MPI_Datatype datatype, int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_allocate_t) (MPI_Aint size, int disp_unit, MPIR_Info * info,
+                                            MPIR_Comm * comm, void *baseptr, MPIR_Win ** win);
+typedef int (*MPIDI_NM_mpi_win_flush_t) (int rank, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_flush_local_all_t) (MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_unlock_all_t) (MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_create_dynamic_t) (MPIR_Info * info, MPIR_Comm * comm,
+                                                  MPIR_Win ** win);
+typedef int (*MPIDI_NM_mpi_rget_t) (void *origin_addr, int origin_count,
+                                    MPI_Datatype origin_datatype, int target_rank,
+                                    MPI_Aint target_disp, int target_count,
+                                    MPI_Datatype target_datatype, MPIR_Win * win,
+                                    MPIR_Request ** request);
+typedef int (*MPIDI_NM_mpi_win_sync_t) (MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_flush_all_t) (MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_get_accumulate_t) (const void *origin_addr, int origin_count,
+                                              MPI_Datatype origin_datatype, void *result_addr,
+                                              int result_count, MPI_Datatype result_datatype,
+                                              int target_rank, MPI_Aint target_disp,
+                                              int target_count, MPI_Datatype target_datatype,
+                                              MPI_Op op, MPIR_Win * win);
+typedef int (*MPIDI_NM_mpi_win_lock_all_t) (int assert, MPIR_Win * win);
 typedef int (*MPIDI_NM_rank_is_local_t) (int target, MPIR_Comm * comm);
-typedef int (*MPIDI_NM_barrier_t) (MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_bcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
-                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_allreduce_t) (const void *sendbuf, void *recvbuf, int count,
-                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                     MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_allgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
+typedef int (*MPIDI_NM_mpi_barrier_t) (MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_bcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
                                      MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_allgatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts, const int *displs,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm,
-                                      MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_scatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_scatterv_t) (const void *sendbuf, const int *sendcounts, const int *displs,
-                                    MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_gather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                  void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_gatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, const int *recvcounts, const int *displs,
-                                   MPI_Datatype recvtype, int root, MPIR_Comm * comm,
-                                   MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_alltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_alltoallv_t) (const void *sendbuf, const int *sendcounts, const int *sdispls,
-                                     MPI_Datatype sendtype, void *recvbuf, const int *recvcounts,
-                                     const int *rdispls, MPI_Datatype recvtype, MPIR_Comm * comm,
-                                     MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_alltoallw_t) (const void *sendbuf, const int *sendcounts, const int *sdispls,
-                                     const MPI_Datatype sendtypes[], void *recvbuf,
-                                     const int *recvcounts, const int *rdispls,
-                                     const MPI_Datatype recvtypes[], MPIR_Comm * comm,
-                                     MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_reduce_t) (const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, int root, MPIR_Comm * comm_ptr,
-                                  MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_reduce_scatter_t) (const void *sendbuf, void *recvbuf, const int *recvcounts,
-                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+typedef int (*MPIDI_NM_mpi_allreduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                         MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                         MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_allgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_allgatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts, const int *displs,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm,
                                           MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_reduce_scatter_block_t) (const void *sendbuf, void *recvbuf, int recvcount,
-                                                MPI_Datatype datatype, MPI_Op op,
-                                                MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_scan_t) (const void *sendbuf, void *recvbuf, int count,
-                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_exscan_t) (const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                  MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_NM_neighbor_allgather_t) (const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm);
-typedef int (*MPIDI_NM_neighbor_allgatherv_t) (const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int *recvcounts, const int *displs,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm);
-typedef int (*MPIDI_NM_neighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                              const int *sdispls, MPI_Datatype sendtype,
-                                              void *recvbuf, const int *recvcounts,
-                                              const int *rdispls, MPI_Datatype recvtype,
-                                              MPIR_Comm * comm);
-typedef int (*MPIDI_NM_neighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                              const MPI_Aint * sdispls,
-                                              const MPI_Datatype * sendtypes, void *recvbuf,
-                                              const int *recvcounts, const MPI_Aint * rdispls,
-                                              const MPI_Datatype * recvtypes, MPIR_Comm * comm);
-typedef int (*MPIDI_NM_neighbor_alltoall_t) (const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                             MPI_Datatype recvtype, MPIR_Comm * comm);
-typedef int (*MPIDI_NM_ineighbor_allgather_t) (const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm,
-                                               MPI_Request * req);
-typedef int (*MPIDI_NM_ineighbor_allgatherv_t) (const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int *recvcounts, const int *displs,
-                                                MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                MPI_Request * req);
-typedef int (*MPIDI_NM_ineighbor_alltoall_t) (const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm,
-                                              MPI_Request * req);
-typedef int (*MPIDI_NM_ineighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                               const int *sdispls, MPI_Datatype sendtype,
-                                               void *recvbuf, const int *recvcounts,
-                                               const int *rdispls, MPI_Datatype recvtype,
-                                               MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_ineighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                               const MPI_Aint * sdispls,
-                                               const MPI_Datatype * sendtypes, void *recvbuf,
-                                               const int *recvcounts, const MPI_Aint * rdispls,
-                                               const MPI_Datatype * recvtypes, MPIR_Comm * comm,
-                                               MPI_Request * req);
-typedef int (*MPIDI_NM_ibarrier_t) (MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_ibcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
-                                  MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_iallgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_iallgatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+typedef int (*MPIDI_NM_mpi_scatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_scatterv_t) (const void *sendbuf, const int *sendcounts,
+                                        const int *displs, MPI_Datatype sendtype, void *recvbuf,
+                                        int recvcount, MPI_Datatype recvtype, int root,
+                                        MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_gather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                      void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
+                                      MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_gatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
                                        void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_iallreduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype recvtype, int root, MPIR_Comm * comm,
+                                       MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_alltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_alltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                         const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
+                                         const int *recvcounts, const int *rdispls,
+                                         MPI_Datatype recvtype, MPIR_Comm * comm,
+                                         MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_alltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                         const int *sdispls, const MPI_Datatype sendtypes[],
+                                         void *recvbuf, const int *recvcounts, const int *rdispls,
+                                         const MPI_Datatype recvtypes[], MPIR_Comm * comm,
+                                         MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_reduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, int root,
+                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_reduce_scatter_t) (const void *sendbuf, void *recvbuf,
+                                              const int *recvcounts, MPI_Datatype datatype,
+                                              MPI_Op op, MPIR_Comm * comm_ptr,
+                                              MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_reduce_scatter_block_t) (const void *sendbuf, void *recvbuf,
+                                                    int recvcount, MPI_Datatype datatype, MPI_Op op,
+                                                    MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_scan_t) (const void *sendbuf, void *recvbuf, int count,
+                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                    MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_exscan_t) (const void *sendbuf, void *recvbuf, int count,
                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                      MPI_Request * req);
-typedef int (*MPIDI_NM_ialltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_ialltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
-                                      const int *recvcounts, const int *rdispls,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_ialltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int *recvcounts, const int *rdispls,
-                                      const MPI_Datatype recvtypes[], MPIR_Comm * comm,
-                                      MPI_Request * req);
-typedef int (*MPIDI_NM_iexscan_t) (const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                   MPI_Request * req);
-typedef int (*MPIDI_NM_igather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                   MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_igatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm,
-                                    MPI_Request * req);
-typedef int (*MPIDI_NM_ireduce_scatter_block_t) (const void *sendbuf, void *recvbuf, int recvcount,
-                                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                                 MPI_Request * req);
-typedef int (*MPIDI_NM_ireduce_scatter_t) (const void *sendbuf, void *recvbuf,
-                                           const int *recvcounts, MPI_Datatype datatype, MPI_Op op,
-                                           MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_ireduce_t) (const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root, MPIR_Comm * comm_ptr,
-                                   MPI_Request * req);
-typedef int (*MPIDI_NM_iscan_t) (const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                 MPI_Request * req);
-typedef int (*MPIDI_NM_iscatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                    MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_NM_iscatterv_t) (const void *sendbuf, const int *sendcounts, const int *displs,
-                                     MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                     MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                      MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_NM_mpi_neighbor_allgather_t) (const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_neighbor_allgatherv_t) (const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int *recvcounts, const int *displs,
+                                                   MPI_Datatype recvtype, MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_neighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                                  const int *sdispls, MPI_Datatype sendtype,
+                                                  void *recvbuf, const int *recvcounts,
+                                                  const int *rdispls, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_neighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                                  const MPI_Aint * sdispls,
+                                                  const MPI_Datatype * sendtypes, void *recvbuf,
+                                                  const int *recvcounts, const MPI_Aint * rdispls,
+                                                  const MPI_Datatype * recvtypes, MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_neighbor_alltoall_t) (const void *sendbuf, int sendcount,
+                                                 MPI_Datatype sendtype, void *recvbuf,
+                                                 int recvcount, MPI_Datatype recvtype,
+                                                 MPIR_Comm * comm);
+typedef int (*MPIDI_NM_mpi_ineighbor_allgather_t) (const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ineighbor_allgatherv_t) (const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int *recvcounts, const int *displs,
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                    MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ineighbor_alltoall_t) (const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ineighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                                   const int *sdispls, MPI_Datatype sendtype,
+                                                   void *recvbuf, const int *recvcounts,
+                                                   const int *rdispls, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ineighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                                   const MPI_Aint * sdispls,
+                                                   const MPI_Datatype * sendtypes, void *recvbuf,
+                                                   const int *recvcounts, const MPI_Aint * rdispls,
+                                                   const MPI_Datatype * recvtypes, MPIR_Comm * comm,
+                                                   MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ibarrier_t) (MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ibcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
+                                      MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_iallgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_iallgatherv_t) (const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm,
+                                           MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_iallreduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                          MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ialltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ialltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
+                                          const int *recvcounts, const int *rdispls,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm,
+                                          MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ialltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int *recvcounts, const int *rdispls,
+                                          const MPI_Datatype recvtypes[], MPIR_Comm * comm,
+                                          MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_iexscan_t) (const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                       MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_igather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_igatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm,
+                                        MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ireduce_scatter_block_t) (const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm,
+                                                     MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ireduce_scatter_t) (const void *sendbuf, void *recvbuf,
+                                               const int *recvcounts, MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_ireduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_iscan_t) (const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                      MPI_Request * req);
-typedef int (*MPIDI_NM_type_create_hook_t) (MPIR_Datatype * datatype_p);
-typedef int (*MPIDI_NM_type_free_hook_t) (MPIR_Datatype * datatype_p);
-typedef int (*MPIDI_NM_op_create_hook_t) (MPIR_Op * op_p);
-typedef int (*MPIDI_NM_op_free_hook_t) (MPIR_Op * op_p);
+typedef int (*MPIDI_NM_mpi_iscatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_iscatterv_t) (const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype, void *recvbuf,
+                                         int recvcount, MPI_Datatype recvtype, int root,
+                                         MPIR_Comm * comm_ptr, MPI_Request * req);
+typedef int (*MPIDI_NM_mpi_type_create_hook_t) (MPIR_Datatype * datatype_p);
+typedef int (*MPIDI_NM_mpi_type_free_hook_t) (MPIR_Datatype * datatype_p);
+typedef int (*MPIDI_NM_mpi_op_create_hook_t) (MPIR_Op * op_p);
+typedef int (*MPIDI_NM_mpi_op_free_hook_t) (MPIR_Op * op_p);
 
 typedef struct MPIDI_NM_funcs {
-    MPIDI_NM_init_t init;
-    MPIDI_NM_finalize_t finalize;
+    MPIDI_NM_mpi_init_t mpi_init;
+    MPIDI_NM_mpi_finalize_t mpi_finalize;
     MPIDI_NM_progress_t progress;
-    MPIDI_NM_comm_connect_t comm_connect;
-    MPIDI_NM_comm_disconnect_t comm_disconnect;
-    MPIDI_NM_open_port_t open_port;
-    MPIDI_NM_close_port_t close_port;
-    MPIDI_NM_comm_accept_t comm_accept;
+    MPIDI_NM_mpi_comm_connect_t mpi_comm_connect;
+    MPIDI_NM_mpi_comm_disconnect_t mpi_comm_disconnect;
+    MPIDI_NM_mpi_open_port_t mpi_open_port;
+    MPIDI_NM_mpi_close_port_t mpi_close_port;
+    MPIDI_NM_mpi_comm_accept_t mpi_comm_accept;
     /* Routines that handle addressing */
     MPIDI_NM_comm_get_lpid_t comm_get_lpid;
     MPIDI_NM_gpid_get_t gpid_get;
     MPIDI_NM_getallincomm_t getallincomm;
     MPIDI_NM_gpid_tolpidarray_t gpid_tolpidarray;
     MPIDI_NM_create_intercomm_from_lpids_t create_intercomm_from_lpids;
-    MPIDI_NM_comm_create_hook_t comm_create_hook;
-    MPIDI_NM_comm_free_hook_t comm_free_hook;
+    MPIDI_NM_mpi_comm_create_hook_t mpi_comm_create_hook;
+    MPIDI_NM_mpi_comm_free_hook_t mpi_comm_free_hook;
     /* Request allocation routines */
     MPIDI_NM_am_request_init_t am_request_init;
     MPIDI_NM_am_request_finalize_t am_request_finalize;
@@ -372,112 +389,112 @@ typedef struct MPIDI_NM_funcs {
 } MPIDI_NM_funcs_t;
 
 typedef struct MPIDI_NM_native_funcs {
-    MPIDI_NM_send_t send;
-    MPIDI_NM_ssend_t ssend;
-    MPIDI_NM_startall_t startall;
-    MPIDI_NM_send_init_t send_init;
-    MPIDI_NM_ssend_init_t ssend_init;
-    MPIDI_NM_rsend_init_t rsend_init;
-    MPIDI_NM_bsend_init_t bsend_init;
-    MPIDI_NM_isend_t isend;
-    MPIDI_NM_issend_t issend;
-    MPIDI_NM_cancel_send_t cancel_send;
-    MPIDI_NM_recv_init_t recv_init;
-    MPIDI_NM_recv_t recv;
-    MPIDI_NM_irecv_t irecv;
-    MPIDI_NM_imrecv_t imrecv;
-    MPIDI_NM_cancel_recv_t cancel_recv;
-    MPIDI_NM_alloc_mem_t alloc_mem;
-    MPIDI_NM_free_mem_t free_mem;
-    MPIDI_NM_improbe_t improbe;
-    MPIDI_NM_iprobe_t iprobe;
-    MPIDI_NM_win_set_info_t win_set_info;
-    MPIDI_NM_win_shared_query_t win_shared_query;
-    MPIDI_NM_put_t put;
-    MPIDI_NM_win_start_t win_start;
-    MPIDI_NM_win_complete_t win_complete;
-    MPIDI_NM_win_post_t win_post;
-    MPIDI_NM_win_wait_t win_wait;
-    MPIDI_NM_win_test_t win_test;
-    MPIDI_NM_win_lock_t win_lock;
-    MPIDI_NM_win_unlock_t win_unlock;
-    MPIDI_NM_win_get_info_t win_get_info;
-    MPIDI_NM_get_t get;
-    MPIDI_NM_win_free_t win_free;
-    MPIDI_NM_win_fence_t win_fence;
-    MPIDI_NM_win_create_t win_create;
-    MPIDI_NM_accumulate_t accumulate;
-    MPIDI_NM_win_attach_t win_attach;
-    MPIDI_NM_win_allocate_shared_t win_allocate_shared;
-    MPIDI_NM_rput_t rput;
-    MPIDI_NM_win_flush_local_t win_flush_local;
-    MPIDI_NM_win_detach_t win_detach;
-    MPIDI_NM_compare_and_swap_t compare_and_swap;
-    MPIDI_NM_raccumulate_t raccumulate;
-    MPIDI_NM_rget_accumulate_t rget_accumulate;
-    MPIDI_NM_fetch_and_op_t fetch_and_op;
-    MPIDI_NM_win_allocate_t win_allocate;
-    MPIDI_NM_win_flush_t win_flush;
-    MPIDI_NM_win_flush_local_all_t win_flush_local_all;
-    MPIDI_NM_win_unlock_all_t win_unlock_all;
-    MPIDI_NM_win_create_dynamic_t win_create_dynamic;
-    MPIDI_NM_rget_t rget;
-    MPIDI_NM_win_sync_t win_sync;
-    MPIDI_NM_win_flush_all_t win_flush_all;
-    MPIDI_NM_get_accumulate_t get_accumulate;
-    MPIDI_NM_win_lock_all_t win_lock_all;
+    MPIDI_NM_mpi_send_t mpi_send;
+    MPIDI_NM_mpi_ssend_t mpi_ssend;
+    MPIDI_NM_mpi_startall_t mpi_startall;
+    MPIDI_NM_mpi_send_init_t mpi_send_init;
+    MPIDI_NM_mpi_ssend_init_t mpi_ssend_init;
+    MPIDI_NM_mpi_rsend_init_t mpi_rsend_init;
+    MPIDI_NM_mpi_bsend_init_t mpi_bsend_init;
+    MPIDI_NM_mpi_isend_t mpi_isend;
+    MPIDI_NM_mpi_issend_t mpi_issend;
+    MPIDI_NM_mpi_cancel_send_t mpi_cancel_send;
+    MPIDI_NM_mpi_recv_init_t mpi_recv_init;
+    MPIDI_NM_mpi_recv_t mpi_recv;
+    MPIDI_NM_mpi_irecv_t mpi_irecv;
+    MPIDI_NM_mpi_imrecv_t mpi_imrecv;
+    MPIDI_NM_mpi_cancel_recv_t mpi_cancel_recv;
+    MPIDI_NM_mpi_alloc_mem_t mpi_alloc_mem;
+    MPIDI_NM_mpi_free_mem_t mpi_free_mem;
+    MPIDI_NM_mpi_improbe_t mpi_improbe;
+    MPIDI_NM_mpi_iprobe_t mpi_iprobe;
+    MPIDI_NM_mpi_win_set_info_t mpi_win_set_info;
+    MPIDI_NM_mpi_win_shared_query_t mpi_win_shared_query;
+    MPIDI_NM_mpi_put_t mpi_put;
+    MPIDI_NM_mpi_win_start_t mpi_win_start;
+    MPIDI_NM_mpi_win_complete_t mpi_win_complete;
+    MPIDI_NM_mpi_win_post_t mpi_win_post;
+    MPIDI_NM_mpi_win_wait_t mpi_win_wait;
+    MPIDI_NM_mpi_win_test_t mpi_win_test;
+    MPIDI_NM_mpi_win_lock_t mpi_win_lock;
+    MPIDI_NM_mpi_win_unlock_t mpi_win_unlock;
+    MPIDI_NM_mpi_win_get_info_t mpi_win_get_info;
+    MPIDI_NM_mpi_get_t mpi_get;
+    MPIDI_NM_mpi_win_free_t mpi_win_free;
+    MPIDI_NM_mpi_win_fence_t mpi_win_fence;
+    MPIDI_NM_mpi_win_create_t mpi_win_create;
+    MPIDI_NM_mpi_accumulate_t mpi_accumulate;
+    MPIDI_NM_mpi_win_attach_t mpi_win_attach;
+    MPIDI_NM_mpi_win_allocate_shared_t mpi_win_allocate_shared;
+    MPIDI_NM_mpi_rput_t mpi_rput;
+    MPIDI_NM_mpi_win_flush_local_t mpi_win_flush_local;
+    MPIDI_NM_mpi_win_detach_t mpi_win_detach;
+    MPIDI_NM_mpi_compare_and_swap_t mpi_compare_and_swap;
+    MPIDI_NM_mpi_raccumulate_t mpi_raccumulate;
+    MPIDI_NM_mpi_rget_accumulate_t mpi_rget_accumulate;
+    MPIDI_NM_mpi_fetch_and_op_t mpi_fetch_and_op;
+    MPIDI_NM_mpi_win_allocate_t mpi_win_allocate;
+    MPIDI_NM_mpi_win_flush_t mpi_win_flush;
+    MPIDI_NM_mpi_win_flush_local_all_t mpi_win_flush_local_all;
+    MPIDI_NM_mpi_win_unlock_all_t mpi_win_unlock_all;
+    MPIDI_NM_mpi_win_create_dynamic_t mpi_win_create_dynamic;
+    MPIDI_NM_mpi_rget_t mpi_rget;
+    MPIDI_NM_mpi_win_sync_t mpi_win_sync;
+    MPIDI_NM_mpi_win_flush_all_t mpi_win_flush_all;
+    MPIDI_NM_mpi_get_accumulate_t mpi_get_accumulate;
+    MPIDI_NM_mpi_win_lock_all_t mpi_win_lock_all;
     MPIDI_NM_rank_is_local_t rank_is_local;
     /* Collectives */
-    MPIDI_NM_barrier_t barrier;
-    MPIDI_NM_bcast_t bcast;
-    MPIDI_NM_allreduce_t allreduce;
-    MPIDI_NM_allgather_t allgather;
-    MPIDI_NM_allgatherv_t allgatherv;
-    MPIDI_NM_scatter_t scatter;
-    MPIDI_NM_scatterv_t scatterv;
-    MPIDI_NM_gather_t gather;
-    MPIDI_NM_gatherv_t gatherv;
-    MPIDI_NM_alltoall_t alltoall;
-    MPIDI_NM_alltoallv_t alltoallv;
-    MPIDI_NM_alltoallw_t alltoallw;
-    MPIDI_NM_reduce_t reduce;
-    MPIDI_NM_reduce_scatter_t reduce_scatter;
-    MPIDI_NM_reduce_scatter_block_t reduce_scatter_block;
-    MPIDI_NM_scan_t scan;
-    MPIDI_NM_exscan_t exscan;
-    MPIDI_NM_neighbor_allgather_t neighbor_allgather;
-    MPIDI_NM_neighbor_allgatherv_t neighbor_allgatherv;
-    MPIDI_NM_neighbor_alltoall_t neighbor_alltoall;
-    MPIDI_NM_neighbor_alltoallv_t neighbor_alltoallv;
-    MPIDI_NM_neighbor_alltoallw_t neighbor_alltoallw;
-    MPIDI_NM_ineighbor_allgather_t ineighbor_allgather;
-    MPIDI_NM_ineighbor_allgatherv_t ineighbor_allgatherv;
-    MPIDI_NM_ineighbor_alltoall_t ineighbor_alltoall;
-    MPIDI_NM_ineighbor_alltoallv_t ineighbor_alltoallv;
-    MPIDI_NM_ineighbor_alltoallw_t ineighbor_alltoallw;
-    MPIDI_NM_ibarrier_t ibarrier;
-    MPIDI_NM_ibcast_t ibcast;
-    MPIDI_NM_iallgather_t iallgather;
-    MPIDI_NM_iallgatherv_t iallgatherv;
-    MPIDI_NM_iallreduce_t iallreduce;
-    MPIDI_NM_ialltoall_t ialltoall;
-    MPIDI_NM_ialltoallv_t ialltoallv;
-    MPIDI_NM_ialltoallw_t ialltoallw;
-    MPIDI_NM_iexscan_t iexscan;
-    MPIDI_NM_igather_t igather;
-    MPIDI_NM_igatherv_t igatherv;
-    MPIDI_NM_ireduce_scatter_block_t ireduce_scatter_block;
-    MPIDI_NM_ireduce_scatter_t ireduce_scatter;
-    MPIDI_NM_ireduce_t ireduce;
-    MPIDI_NM_iscan_t iscan;
-    MPIDI_NM_iscatter_t iscatter;
-    MPIDI_NM_iscatterv_t iscatterv;
+    MPIDI_NM_mpi_barrier_t mpi_barrier;
+    MPIDI_NM_mpi_bcast_t mpi_bcast;
+    MPIDI_NM_mpi_allreduce_t mpi_allreduce;
+    MPIDI_NM_mpi_allgather_t mpi_allgather;
+    MPIDI_NM_mpi_allgatherv_t mpi_allgatherv;
+    MPIDI_NM_mpi_scatter_t mpi_scatter;
+    MPIDI_NM_mpi_scatterv_t mpi_scatterv;
+    MPIDI_NM_mpi_gather_t mpi_gather;
+    MPIDI_NM_mpi_gatherv_t mpi_gatherv;
+    MPIDI_NM_mpi_alltoall_t mpi_alltoall;
+    MPIDI_NM_mpi_alltoallv_t mpi_alltoallv;
+    MPIDI_NM_mpi_alltoallw_t mpi_alltoallw;
+    MPIDI_NM_mpi_reduce_t mpi_reduce;
+    MPIDI_NM_mpi_reduce_scatter_t mpi_reduce_scatter;
+    MPIDI_NM_mpi_reduce_scatter_block_t mpi_reduce_scatter_block;
+    MPIDI_NM_mpi_scan_t mpi_scan;
+    MPIDI_NM_mpi_exscan_t mpi_exscan;
+    MPIDI_NM_mpi_neighbor_allgather_t mpi_neighbor_allgather;
+    MPIDI_NM_mpi_neighbor_allgatherv_t mpi_neighbor_allgatherv;
+    MPIDI_NM_mpi_neighbor_alltoall_t mpi_neighbor_alltoall;
+    MPIDI_NM_mpi_neighbor_alltoallv_t mpi_neighbor_alltoallv;
+    MPIDI_NM_mpi_neighbor_alltoallw_t mpi_neighbor_alltoallw;
+    MPIDI_NM_mpi_ineighbor_allgather_t mpi_ineighbor_allgather;
+    MPIDI_NM_mpi_ineighbor_allgatherv_t mpi_ineighbor_allgatherv;
+    MPIDI_NM_mpi_ineighbor_alltoall_t mpi_ineighbor_alltoall;
+    MPIDI_NM_mpi_ineighbor_alltoallv_t mpi_ineighbor_alltoallv;
+    MPIDI_NM_mpi_ineighbor_alltoallw_t mpi_ineighbor_alltoallw;
+    MPIDI_NM_mpi_ibarrier_t mpi_ibarrier;
+    MPIDI_NM_mpi_ibcast_t mpi_ibcast;
+    MPIDI_NM_mpi_iallgather_t mpi_iallgather;
+    MPIDI_NM_mpi_iallgatherv_t mpi_iallgatherv;
+    MPIDI_NM_mpi_iallreduce_t mpi_iallreduce;
+    MPIDI_NM_mpi_ialltoall_t mpi_ialltoall;
+    MPIDI_NM_mpi_ialltoallv_t mpi_ialltoallv;
+    MPIDI_NM_mpi_ialltoallw_t mpi_ialltoallw;
+    MPIDI_NM_mpi_iexscan_t mpi_iexscan;
+    MPIDI_NM_mpi_igather_t mpi_igather;
+    MPIDI_NM_mpi_igatherv_t mpi_igatherv;
+    MPIDI_NM_mpi_ireduce_scatter_block_t mpi_ireduce_scatter_block;
+    MPIDI_NM_mpi_ireduce_scatter_t mpi_ireduce_scatter;
+    MPIDI_NM_mpi_ireduce_t mpi_ireduce;
+    MPIDI_NM_mpi_iscan_t mpi_iscan;
+    MPIDI_NM_mpi_iscatter_t mpi_iscatter;
+    MPIDI_NM_mpi_iscatterv_t mpi_iscatterv;
     /* Datatype hooks */
-    MPIDI_NM_type_create_hook_t type_create_hook;
-    MPIDI_NM_type_free_hook_t type_free_hook;
+    MPIDI_NM_mpi_type_create_hook_t mpi_type_create_hook;
+    MPIDI_NM_mpi_type_free_hook_t mpi_type_free_hook;
     /* Op hooks */
-    MPIDI_NM_op_create_hook_t op_create_hook;
-    MPIDI_NM_op_free_hook_t op_free_hook;
+    MPIDI_NM_mpi_op_create_hook_t mpi_op_create_hook;
+    MPIDI_NM_mpi_op_free_hook_t mpi_op_free_hook;
 } MPIDI_NM_native_funcs_t;
 
 extern MPIDI_NM_funcs_t *MPIDI_NM_funcs[];
@@ -487,11 +504,11 @@ extern MPIDI_NM_native_funcs_t *MPIDI_NM_native_func;
 extern int MPIDI_num_netmods;
 extern char MPIDI_NM_strings[][MPIDI_MAX_NETMOD_STRING_LEN];
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_init(int rank, int size, int appnum, int *tag_ub,
-                                           MPIR_Comm * comm_world, MPIR_Comm * comm_self,
-                                           int spawned, int num_contexts,
-                                           void **netmod_contexts) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_finalize(void) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_init(int rank, int size, int appnum, int *tag_ub,
+                                               MPIR_Comm * comm_world, MPIR_Comm * comm_self,
+                                               int spawned, int num_contexts,
+                                               void **netmod_contexts) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_finalize(void) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_progress(void *netmod_context,
                                                int blocking) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_reg_handler(int handler_id,
@@ -499,19 +516,20 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_reg_handler(int handler_id,
                                                      origin_handler_fn,
                                                      MPIDI_NM_am_target_handler_fn
                                                      target_handler_fn) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_connect(const char *port_name, MPIR_Info * info,
-                                                   int root, MPIR_Comm * comm,
-                                                   MPIR_Comm **
-                                                   newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_disconnect(MPIR_Comm *
-                                                      comm_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_open_port(MPIR_Info * info_ptr,
-                                                char *port_name) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_close_port(const char *port_name) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_accept(const char *port_name, MPIR_Info * info,
-                                                  int root, MPIR_Comm * comm,
-                                                  MPIR_Comm **
-                                                  newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_connect(const char *port_name, MPIR_Info * info,
+                                                       int root, MPIR_Comm * comm,
+                                                       MPIR_Comm **
+                                                       newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm *
+                                                          comm_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr,
+                                                    char *port_name) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_close_port(const char *port_name)
+    MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_accept(const char *port_name, MPIR_Info * info,
+                                                      int root, MPIR_Comm * comm,
+                                                      MPIR_Comm **
+                                                      newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_send_hdr(int rank, MPIR_Comm * comm, int handler_id,
                                                   const void *am_hdr, size_t am_hdr_sz,
                                                   void *netmod_context) MPL_STATIC_INLINE_SUFFIX;
@@ -525,9 +543,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_isendv(int rank, MPIR_Comm * comm, int
                                                 const void *data, MPI_Count count,
                                                 MPI_Datatype datatype, MPIR_Request * sreq,
                                                 void *netmod_context) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_send_hdr_reply(MPIR_Context_id_t context_id,
-                                                        int src_rank, int handler_id,
-                                                        const void *am_hdr,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_send_hdr_reply(MPIR_Context_id_t context_id, int src_rank,
+                                                        int handler_id, const void *am_hdr,
                                                         size_t am_hdr_sz) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_isend_reply(MPIR_Context_id_t context_id, int src_rank,
                                                      int handler_id, const void *am_hdr,
@@ -550,423 +567,448 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_create_intercomm_from_lpids(MPIR_Comm * ne
                                                                   int size,
                                                                   const int lpids[])
     MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_create_hook(MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_free_hook(MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_create_hook(MPIR_Comm *
+                                                           comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_free_hook(MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX void MPIDI_NM_am_request_init(MPIR_Request * req) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX void MPIDI_NM_am_request_finalize(MPIR_Request *
                                                            req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_send(const void *buf, int count, MPI_Datatype datatype,
-                                           int rank, int tag, MPIR_Comm * comm,
-                                           int context_offset,
-                                           MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ssend(const void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_startall(int count,
-                                               MPIR_Request * requests[]) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_send_init(const void *buf, int count,
-                                                MPI_Datatype datatype, int rank, int tag,
-                                                MPIR_Comm * comm, int context_offset,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_send(const void *buf, int count, MPI_Datatype datatype,
+                                               int rank, int tag, MPIR_Comm * comm,
+                                               int context_offset,
+                                               MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ssend(const void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset,
                                                 MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ssend_init(const void *buf, int count,
-                                                 MPI_Datatype datatype, int rank, int tag,
-                                                 MPIR_Comm * comm, int context_offset,
-                                                 MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rsend_init(const void *buf, int count,
-                                                 MPI_Datatype datatype, int rank, int tag,
-                                                 MPIR_Comm * comm, int context_offset,
-                                                 MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_bsend_init(const void *buf, int count,
-                                                 MPI_Datatype datatype, int rank, int tag,
-                                                 MPIR_Comm * comm, int context_offset,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_startall(int count,
+                                                   MPIR_Request *
+                                                   requests[]) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_send_init(const void *buf, int count,
+                                                    MPI_Datatype datatype, int rank, int tag,
+                                                    MPIR_Comm * comm, int context_offset,
+                                                    MPIR_Request **
+                                                    request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ssend_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request **
+                                                     request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rsend_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request **
+                                                     request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_bsend_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request **
+                                                     request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_isend(const void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset,
+                                                MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_issend(const void *buf, int count, MPI_Datatype datatype,
+                                                 int rank, int tag, MPIR_Comm * comm,
+                                                 int context_offset,
                                                  MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_isend(const void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_issend(const void *buf, int count, MPI_Datatype datatype,
-                                             int rank, int tag, MPIR_Comm * comm,
-                                             int context_offset,
-                                             MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_cancel_send(MPIR_Request * sreq) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_recv_init(void *buf, int count, MPI_Datatype datatype,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
+                                                    int rank, int tag, MPIR_Comm * comm,
+                                                    int context_offset,
+                                                    MPIR_Request **
+                                                    request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv(void *buf, int count, MPI_Datatype datatype,
+                                               int rank, int tag, MPIR_Comm * comm,
+                                               int context_offset, MPI_Status * status,
+                                               MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
                                                 int rank, int tag, MPIR_Comm * comm,
                                                 int context_offset,
                                                 MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_recv(void *buf, int count, MPI_Datatype datatype,
-                                           int rank, int tag, MPIR_Comm * comm,
-                                           int context_offset, MPI_Status * status,
-                                           MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_irecv(void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_imrecv(void *buf, int count, MPI_Datatype datatype,
-                                             MPIR_Request * message,
-                                             MPIR_Request ** rreqp) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_cancel_recv(MPIR_Request * rreq) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX void *MPIDI_NM_alloc_mem(size_t size,
-                                                  MPIR_Info * info_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_free_mem(void *ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_improbe(int source, int tag, MPIR_Comm * comm,
-                                              int context_offset, int *flag,
-                                              MPIR_Request ** message,
-                                              MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iprobe(int source, int tag, MPIR_Comm * comm,
-                                             int context_offset, int *flag,
-                                             MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_set_info(MPIR_Win * win,
-                                                   MPIR_Info * info) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_shared_query(MPIR_Win * win, int rank,
-                                                       MPI_Aint * size, int *disp_unit,
-                                                       void *baseptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_put(const void *origin_addr, int origin_count,
-                                          MPI_Datatype origin_datatype, int target_rank,
-                                          MPI_Aint target_disp, int target_count,
-                                          MPI_Datatype target_datatype,
-                                          MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_start(MPIR_Group * group, int assert,
-                                                MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_complete(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_post(MPIR_Group * group, int assert,
-                                               MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_wait(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_test(MPIR_Win * win, int *flag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_lock(int lock_type, int rank, int assert,
-                                               MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_unlock(int rank, MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_get_info(MPIR_Win * win,
-                                                   MPIR_Info ** info_p_p) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_get(void *origin_addr, int origin_count,
-                                          MPI_Datatype origin_datatype, int target_rank,
-                                          MPI_Aint target_disp, int target_count,
-                                          MPI_Datatype target_datatype,
-                                          MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_free(MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_fence(int assert,
-                                                MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_create(void *base, MPI_Aint length, int disp_unit,
-                                                 MPIR_Info * info, MPIR_Comm * comm_ptr,
-                                                 MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_accumulate(const void *origin_addr, int origin_count,
-                                                 MPI_Datatype origin_datatype, int target_rank,
-                                                 MPI_Aint target_disp, int target_count,
-                                                 MPI_Datatype target_datatype, MPI_Op op,
-                                                 MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_attach(MPIR_Win * win, void *base,
-                                                 MPI_Aint size) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_allocate_shared(MPI_Aint size, int disp_unit,
-                                                          MPIR_Info * info_ptr,
-                                                          MPIR_Comm * comm_ptr,
-                                                          void **base_ptr,
-                                                          MPIR_Win **
-                                                          win_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rput(const void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPIR_Win * win,
-                                           MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush_local(int rank,
-                                                      MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_detach(MPIR_Win * win,
-                                                 const void *base) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                                       const void *compare_addr,
-                                                       void *result_addr,
-                                                       MPI_Datatype datatype, int target_rank,
-                                                       MPI_Aint target_disp,
-                                                       MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_raccumulate(const void *origin_addr, int origin_count,
-                                                  MPI_Datatype origin_datatype,
-                                                  int target_rank, MPI_Aint target_disp,
-                                                  int target_count,
-                                                  MPI_Datatype target_datatype, MPI_Op op,
-                                                  MPIR_Win * win,
-                                                  MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rget_accumulate(const void *origin_addr,
-                                                      int origin_count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_imrecv(void *buf, int count, MPI_Datatype datatype,
+                                                 MPIR_Request * message,
+                                                 MPIR_Request ** rreqp) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX void *MPIDI_NM_mpi_alloc_mem(size_t size,
+                                                      MPIR_Info *
+                                                      info_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_free_mem(void *ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_improbe(int source, int tag, MPIR_Comm * comm,
+                                                  int context_offset, int *flag,
+                                                  MPIR_Request ** message,
+                                                  MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iprobe(int source, int tag, MPIR_Comm * comm,
+                                                 int context_offset, int *flag,
+                                                 MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_set_info(MPIR_Win * win,
+                                                       MPIR_Info * info) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win, int rank,
+                                                           MPI_Aint * size, int *disp_unit,
+                                                           void *baseptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_put(const void *origin_addr, int origin_count,
+                                              MPI_Datatype origin_datatype, int target_rank,
+                                              MPI_Aint target_disp, int target_count,
+                                              MPI_Datatype target_datatype,
+                                              MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert,
+                                                    MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_complete(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert,
+                                                   MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_wait(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_test(MPIR_Win * win,
+                                                   int *flag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert,
+                                                   MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_unlock(int rank,
+                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_get_info(MPIR_Win * win,
+                                                       MPIR_Info **
+                                                       info_p_p) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_get(void *origin_addr, int origin_count,
+                                              MPI_Datatype origin_datatype, int target_rank,
+                                              MPI_Aint target_disp, int target_count,
+                                              MPI_Datatype target_datatype,
+                                              MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_fence(int assert,
+                                                    MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_create(void *base, MPI_Aint length, int disp_unit,
+                                                     MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                                     MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_accumulate(const void *origin_addr, int origin_count,
+                                                     MPI_Datatype origin_datatype, int target_rank,
+                                                     MPI_Aint target_disp, int target_count,
+                                                     MPI_Datatype target_datatype, MPI_Op op,
+                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base,
+                                                     MPI_Aint size) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size, int disp_unit,
+                                                              MPIR_Info * info_ptr,
+                                                              MPIR_Comm * comm_ptr,
+                                                              void **base_ptr,
+                                                              MPIR_Win **
+                                                              win_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rput(const void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype, MPIR_Win * win,
+                                               MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush_local(int rank,
+                                                          MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_detach(MPIR_Win * win,
+                                                     const void *base) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                           const void *compare_addr,
+                                                           void *result_addr,
+                                                           MPI_Datatype datatype, int target_rank,
+                                                           MPI_Aint target_disp,
+                                                           MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_raccumulate(const void *origin_addr, int origin_count,
                                                       MPI_Datatype origin_datatype,
-                                                      void *result_addr, int result_count,
-                                                      MPI_Datatype result_datatype,
                                                       int target_rank, MPI_Aint target_disp,
                                                       int target_count,
                                                       MPI_Datatype target_datatype, MPI_Op op,
                                                       MPIR_Win * win,
                                                       MPIR_Request **
                                                       request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_fetch_and_op(const void *origin_addr, void *result_addr,
-                                                   MPI_Datatype datatype, int target_rank,
-                                                   MPI_Aint target_disp, MPI_Op op,
-                                                   MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_allocate(MPI_Aint size, int disp_unit,
-                                                   MPIR_Info * info, MPIR_Comm * comm,
-                                                   void *baseptr,
-                                                   MPIR_Win ** win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush(int rank, MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush_local_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_unlock_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
-                                                         MPIR_Win ** win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rget(void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPIR_Win * win,
-                                           MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_sync(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_get_accumulate(const void *origin_addr, int origin_count,
-                                                     MPI_Datatype origin_datatype,
-                                                     void *result_addr, int result_count,
-                                                     MPI_Datatype result_datatype,
-                                                     int target_rank, MPI_Aint target_disp,
-                                                     int target_count,
-                                                     MPI_Datatype target_datatype, MPI_Op op,
-                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_lock_all(int assert,
-                                                   MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr, int origin_count,
+                                                          MPI_Datatype origin_datatype,
+                                                          void *result_addr, int result_count,
+                                                          MPI_Datatype result_datatype,
+                                                          int target_rank, MPI_Aint target_disp,
+                                                          int target_count,
+                                                          MPI_Datatype target_datatype, MPI_Op op,
+                                                          MPIR_Win * win,
+                                                          MPIR_Request **
+                                                          request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr, void *result_addr,
+                                                       MPI_Datatype datatype, int target_rank,
+                                                       MPI_Aint target_disp, MPI_Op op,
+                                                       MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_allocate(MPI_Aint size, int disp_unit,
+                                                       MPIR_Info * info, MPIR_Comm * comm,
+                                                       void *baseptr,
+                                                       MPIR_Win ** win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush(int rank,
+                                                    MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win *
+                                                              win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                             MPIR_Win **
+                                                             win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rget(void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype, MPIR_Win * win,
+                                               MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_sync(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_get_accumulate(const void *origin_addr, int origin_count,
+                                                         MPI_Datatype origin_datatype,
+                                                         void *result_addr, int result_count,
+                                                         MPI_Datatype result_datatype,
+                                                         int target_rank, MPI_Aint target_disp,
+                                                         int target_count,
+                                                         MPI_Datatype target_datatype, MPI_Op op,
+                                                         MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_lock_all(int assert,
+                                                       MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rank_is_local(int target,
                                                     MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_barrier(MPIR_Comm * comm,
-                                              MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                            int root, MPIR_Comm * comm,
-                                            MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                                MPI_Datatype datatype, MPI_Op op,
-                                                MPIR_Comm * comm,
-                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_allgather(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_barrier(MPIR_Comm * comm,
+                                                  MPIR_Errflag_t *
+                                                  errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                                int root, MPIR_Comm * comm,
                                                 MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                                    MPI_Datatype datatype, MPI_Op op,
+                                                    MPIR_Comm * comm,
+                                                    MPIR_Errflag_t *
+                                                    errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_allgather(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm,
+                                                    MPIR_Errflag_t *
+                                                    errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_allgatherv(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     const int *recvcounts, const int *displs,
+                                                     MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                     MPIR_Errflag_t *
+                                                     errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_scatter(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm,
+                                                  MPIR_Errflag_t *
+                                                  errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                                   const int *displs, MPI_Datatype sendtype,
+                                                   void *recvbuf, int recvcount,
+                                                   MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm_ptr,
+                                                   MPIR_Errflag_t *
+                                                   errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_gather(const void *sendbuf, int sendcount,
                                                  MPI_Datatype sendtype, void *recvbuf,
-                                                 const int *recvcounts, const int *displs,
-                                                 MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                 int recvcount, MPI_Datatype recvtype, int root,
+                                                 MPIR_Comm * comm,
                                                  MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_scatter(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm,
-                                              MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
-                                               const int *displs, MPI_Datatype sendtype,
-                                               void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm_ptr,
-                                               MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_gather(const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf,
-                                             int recvcount, MPI_Datatype recvtype, int root,
-                                             MPIR_Comm * comm,
-                                             MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_gatherv(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              const int *recvcounts, const int *displs,
-                                              MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm,
-                                              MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_alltoall(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype,
-                                               MPIR_Comm * comm,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_gatherv(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  const int *recvcounts, const int *displs,
+                                                  MPI_Datatype recvtype, int root, MPIR_Comm * comm,
+                                                  MPIR_Errflag_t *
+                                                  errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_alltoall(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm,
+                                                   MPIR_Errflag_t *
+                                                   errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                                    const int *sdispls, MPI_Datatype sendtype,
+                                                    void *recvbuf, const int *recvcounts,
+                                                    const int *rdispls, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm,
+                                                    MPIR_Errflag_t *
+                                                    errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_alltoallw(const void *sendbuf, const int *sendcounts,
+                                                    const int *sdispls,
+                                                    const MPI_Datatype sendtypes[], void *recvbuf,
+                                                    const int *recvcounts, const int *rdispls,
+                                                    const MPI_Datatype recvtypes[],
+                                                    MPIR_Comm * comm,
+                                                    MPIR_Errflag_t *
+                                                    errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                                 MPI_Datatype datatype, MPI_Op op, int root,
+                                                 MPIR_Comm * comm_ptr,
+                                                 MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                                         const int *recvcounts,
+                                                         MPI_Datatype datatype, MPI_Op op,
+                                                         MPIR_Comm * comm_ptr,
+                                                         MPIR_Errflag_t *
+                                                         errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                               int recvcount, MPI_Datatype datatype,
+                                                               MPI_Op op, MPIR_Comm * comm_ptr,
+                                                               MPIR_Errflag_t *
+                                                               errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                               MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                                const int *sdispls, MPI_Datatype sendtype,
-                                                void *recvbuf, const int *recvcounts,
-                                                const int *rdispls, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm,
-                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_alltoallw(const void *sendbuf, const int *sendcounts,
-                                                const int *sdispls,
-                                                const MPI_Datatype sendtypes[], void *recvbuf,
-                                                const int *recvcounts, const int *rdispls,
-                                                const MPI_Datatype recvtypes[],
-                                                MPIR_Comm * comm,
-                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                             MPI_Datatype datatype, MPI_Op op, int root,
-                                             MPIR_Comm * comm_ptr,
-                                             MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                                     const int *recvcounts,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                                 MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             int recvcount, MPI_Datatype recvtype,
+                                                             MPIR_Comm *
+                                                             comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const int *displs,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm *
+                                                              comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_alltoallv(const void *sendbuf,
+                                                             const int *sendcounts,
+                                                             const int *sdispls,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             const int *recvcounts,
+                                                             const int *rdispls,
+                                                             MPI_Datatype recvtype,
+                                                             MPIR_Comm *
+                                                             comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_alltoallw(const void *sendbuf,
+                                                             const int *sendcounts,
+                                                             const MPI_Aint * sdispls,
+                                                             const MPI_Datatype * sendtypes,
+                                                             void *recvbuf, const int *recvcounts,
+                                                             const MPI_Aint * rdispls,
+                                                             const MPI_Datatype * recvtypes,
+                                                             MPIR_Comm *
+                                                             comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                            MPI_Datatype sendtype, void *recvbuf,
+                                                            int recvcount, MPI_Datatype recvtype,
+                                                            MPIR_Comm *
+                                                            comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              int recvcount, MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPI_Request *
+                                                              req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                               MPI_Datatype sendtype, void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const int *displs,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm,
+                                                               MPI_Request *
+                                                               req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             int recvcount, MPI_Datatype recvtype,
+                                                             MPIR_Comm * comm,
+                                                             MPI_Request *
+                                                             req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_alltoallv(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const int *sdispls,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const int *rdispls,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPI_Request *
+                                                              req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_alltoallw(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const MPI_Aint * sdispls,
+                                                              const MPI_Datatype * sendtypes,
+                                                              void *recvbuf, const int *recvcounts,
+                                                              const MPI_Aint * rdispls,
+                                                              const MPI_Datatype * recvtypes,
+                                                              MPIR_Comm * comm,
+                                                              MPI_Request *
+                                                              req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ibarrier(MPIR_Comm * comm,
+                                                   MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                                 int root, MPIR_Comm * comm,
+                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iallgather(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     int recvcount, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm,
+                                                     MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                                      MPI_Datatype sendtype, void *recvbuf,
+                                                      const int *recvcounts, const int *displs,
+                                                      MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                      MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
                                                      MPI_Datatype datatype, MPI_Op op,
-                                                     MPIR_Comm * comm_ptr,
-                                                     MPIR_Errflag_t *
-                                                     errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                           int recvcount,
-                                                           MPI_Datatype datatype, MPI_Op op,
-                                                           MPIR_Comm * comm_ptr,
-                                                           MPIR_Errflag_t *
-                                                           errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
-                                           MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                           MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                             MPI_Datatype datatype, MPI_Op op,
-                                             MPIR_Comm * comm,
-                                             MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         int recvcount, MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype, void *recvbuf,
-                                                          const int *recvcounts,
-                                                          const int *displs,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm *
-                                                          comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_alltoallv(const void *sendbuf,
-                                                         const int *sendcounts,
-                                                         const int *sdispls,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         const int *recvcounts,
-                                                         const int *rdispls,
-                                                         MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_alltoallw(const void *sendbuf,
-                                                         const int *sendcounts,
-                                                         const MPI_Aint * sdispls,
-                                                         const MPI_Datatype * sendtypes,
-                                                         void *recvbuf, const int *recvcounts,
-                                                         const MPI_Aint * rdispls,
-                                                         const MPI_Datatype * recvtypes,
-                                                         MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                                        MPI_Datatype sendtype, void *recvbuf,
-                                                        int recvcount, MPI_Datatype recvtype,
-                                                        MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype, void *recvbuf,
-                                                          int recvcount, MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPI_Request *
-                                                          req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const int *displs,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm,
-                                                           MPI_Request *
-                                                           req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         int recvcount, MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm,
-                                                         MPI_Request *
-                                                         req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const int *sdispls,
-                                                          MPI_Datatype sendtype, void *recvbuf,
+                                                     MPIR_Comm * comm,
+                                                     MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ialltoall(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm,
+                                                    MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls, MPI_Datatype sendtype,
+                                                     void *recvbuf, const int *recvcounts,
+                                                     const int *rdispls, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm,
+                                                     MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls,
+                                                     const MPI_Datatype sendtypes[], void *recvbuf,
+                                                     const int *recvcounts, const int *rdispls,
+                                                     const MPI_Datatype recvtypes[],
+                                                     MPIR_Comm * comm,
+                                                     MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op,
+                                                  MPIR_Comm * comm,
+                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_igather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm,
+                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_igatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int *recvcounts, const int *displs,
+                                                   MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm,
+                                                   MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                                int recvcount,
+                                                                MPI_Datatype datatype, MPI_Op op,
+                                                                MPIR_Comm * comm,
+                                                                MPI_Request *
+                                                                req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
                                                           const int *recvcounts,
-                                                          const int *rdispls,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPI_Request *
-                                                          req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const MPI_Aint * sdispls,
-                                                          const MPI_Datatype * sendtypes,
-                                                          void *recvbuf, const int *recvcounts,
-                                                          const MPI_Aint * rdispls,
-                                                          const MPI_Datatype * recvtypes,
+                                                          MPI_Datatype datatype, MPI_Op op,
                                                           MPIR_Comm * comm,
                                                           MPI_Request *
                                                           req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ibarrier(MPIR_Comm * comm,
-                                               MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                             int root, MPIR_Comm * comm,
-                                             MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iallgather(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 int recvcount, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm,
-                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount,
-                                                  MPI_Datatype sendtype, void *recvbuf,
-                                                  const int *recvcounts, const int *displs,
-                                                  MPI_Datatype recvtype, MPIR_Comm * comm,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op, int root,
+                                                  MPIR_Comm * comm_ptr,
                                                   MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                                 MPI_Datatype datatype, MPI_Op op,
-                                                 MPIR_Comm * comm,
-                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm,
-                                                MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls, MPI_Datatype sendtype,
-                                                 void *recvbuf, const int *recvcounts,
-                                                 const int *rdispls, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm,
-                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls,
-                                                 const MPI_Datatype sendtypes[], void *recvbuf,
-                                                 const int *recvcounts, const int *rdispls,
-                                                 const MPI_Datatype recvtypes[],
-                                                 MPIR_Comm * comm,
-                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op,
-                                              MPIR_Comm * comm,
-                                              MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_igather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm,
-                                              MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_igatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int *recvcounts, const int *displs,
-                                               MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm,
-                                               MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                            int recvcount,
-                                                            MPI_Datatype datatype, MPI_Op op,
-                                                            MPIR_Comm * comm,
-                                                            MPI_Request *
-                                                            req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                                      const int *recvcounts,
-                                                      MPI_Datatype datatype, MPI_Op op,
-                                                      MPIR_Comm * comm,
-                                                      MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op, int root,
-                                              MPIR_Comm * comm_ptr,
-                                              MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                            MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                            MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm,
-                                               MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                                const int *displs, MPI_Datatype sendtype,
-                                                void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, int root,
-                                                MPIR_Comm * comm_ptr,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_type_create_hook(MPIR_Datatype *
-                                                       datatype_p) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_type_free_hook(MPIR_Datatype *
-                                                     datatype_p) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_op_create_hook(MPIR_Op * op_p) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_op_free_hook(MPIR_Op * op_p) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm,
+                                                   MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                                    const int *displs, MPI_Datatype sendtype,
+                                                    void *recvbuf, int recvcount,
+                                                    MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm_ptr,
+                                                    MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_type_create_hook(MPIR_Datatype *
+                                                           datatype_p) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_type_free_hook(MPIR_Datatype *
+                                                         datatype_p) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_op_create_hook(MPIR_Op * op_p) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_op_free_hook(MPIR_Op * op_p) MPL_STATIC_INLINE_SUFFIX;
 
 #endif
diff --git a/src/mpid/ch4/netmod/include/netmod_impl.h b/src/mpid/ch4/netmod/include/netmod_impl.h
index 4ba34a1..ad16dbc 100644
--- a/src/mpid/ch4/netmod/include/netmod_impl.h
+++ b/src/mpid/ch4/netmod/include/netmod_impl.h
@@ -16,17 +16,18 @@
 #ifndef NETMOD_DIRECT
 #ifndef NETMOD_DISABLE_INLINES
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_init(int rank, int size, int appnum, int *tag_ub,
-                                           MPIR_Comm * comm_world, MPIR_Comm * comm_self,
-                                           int spawned, int num_contexts, void **netmod_contexts)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_init(int rank, int size, int appnum, int *tag_ub,
+                                               MPIR_Comm * comm_world, MPIR_Comm * comm_self,
+                                               int spawned, int num_contexts,
+                                               void **netmod_contexts)
 {
-    return MPIDI_NM_func->init(rank, size, appnum, tag_ub, comm_world, comm_self, spawned,
-                               num_contexts, netmod_contexts);
+    return MPIDI_NM_func->mpi_init(rank, size, appnum, tag_ub, comm_world, comm_self, spawned,
+                                   num_contexts, netmod_contexts);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_finalize(void)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_finalize(void)
 {
-    return MPIDI_NM_func->finalize();
+    return MPIDI_NM_func->mpi_finalize();
 };
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_progress(void *netmod_context, int blocking)
@@ -43,33 +44,33 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_reg_handler(int handler_id,
     return MPIDI_NM_func->am_reg_handler(handler_id, origin_handler_fn, target_handler_fn);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_connect(const char *port_name, MPIR_Info * info,
-                                                   int root, MPIR_Comm * comm,
-                                                   MPIR_Comm ** newcomm_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_connect(const char *port_name, MPIR_Info * info,
+                                                       int root, MPIR_Comm * comm,
+                                                       MPIR_Comm ** newcomm_ptr)
 {
-    return MPIDI_NM_func->comm_connect(port_name, info, root, comm, newcomm_ptr);
+    return MPIDI_NM_func->mpi_comm_connect(port_name, info, root, comm, newcomm_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
-    return MPIDI_NM_func->comm_disconnect(comm_ptr);
+    return MPIDI_NM_func->mpi_comm_disconnect(comm_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
-    return MPIDI_NM_func->open_port(info_ptr, port_name);
+    return MPIDI_NM_func->mpi_open_port(info_ptr, port_name);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_close_port(const char *port_name)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_close_port(const char *port_name)
 {
-    return MPIDI_NM_func->close_port(port_name);
+    return MPIDI_NM_func->mpi_close_port(port_name);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_accept(const char *port_name, MPIR_Info * info,
-                                                  int root, MPIR_Comm * comm,
-                                                  MPIR_Comm ** newcomm_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_accept(const char *port_name, MPIR_Info * info,
+                                                      int root, MPIR_Comm * comm,
+                                                      MPIR_Comm ** newcomm_ptr)
 {
-    return MPIDI_NM_func->comm_accept(port_name, info, root, comm, newcomm_ptr);
+    return MPIDI_NM_func->mpi_comm_accept(port_name, info, root, comm, newcomm_ptr);
 };
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_am_send_hdr(int rank, MPIR_Comm * comm, int handler_id,
@@ -154,14 +155,14 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_create_intercomm_from_lpids(MPIR_Comm * ne
     return MPIDI_NM_func->create_intercomm_from_lpids(newcomm_ptr, size, lpids);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
-    return MPIDI_NM_func->comm_create_hook(comm);
+    return MPIDI_NM_func->mpi_comm_create_hook(comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_comm_free_hook(MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
-    return MPIDI_NM_func->comm_free_hook(comm);
+    return MPIDI_NM_func->mpi_comm_free_hook(comm);
 };
 
 MPL_STATIC_INLINE_PREFIX void MPIDI_NM_am_request_init(MPIR_Request * req)
@@ -174,386 +175,390 @@ MPL_STATIC_INLINE_PREFIX void MPIDI_NM_am_request_finalize(MPIR_Request * req)
     return MPIDI_NM_func->am_request_finalize(req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_send(const void *buf, int count, MPI_Datatype datatype,
-                                           int rank, int tag, MPIR_Comm * comm,
-                                           int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_send(const void *buf, int count, MPI_Datatype datatype,
+                                               int rank, int tag, MPIR_Comm * comm,
+                                               int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->send(buf, count, datatype, rank, tag, comm, context_offset,
-                                      request);
+    return MPIDI_NM_native_func->mpi_send(buf, count, datatype, rank, tag, comm, context_offset,
+                                          request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ssend(const void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset, MPIR_Request ** request)
-{
-    return MPIDI_NM_native_func->ssend(buf, count, datatype, rank, tag, comm, context_offset,
-                                       request);
-};
-
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ssend(const void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->startall(count, requests);
+    return MPIDI_NM_native_func->mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset,
+                                           request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_send_init(const void *buf, int count,
-                                                MPI_Datatype datatype, int rank, int tag,
-                                                MPIR_Comm * comm, int context_offset,
-                                                MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_NM_native_func->send_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                           request);
+    return MPIDI_NM_native_func->mpi_startall(count, requests);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ssend_init(const void *buf, int count,
-                                                 MPI_Datatype datatype, int rank, int tag,
-                                                 MPIR_Comm * comm, int context_offset,
-                                                 MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_send_init(const void *buf, int count,
+                                                    MPI_Datatype datatype, int rank, int tag,
+                                                    MPIR_Comm * comm, int context_offset,
+                                                    MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                            request);
+    return MPIDI_NM_native_func->mpi_send_init(buf, count, datatype, rank, tag, comm,
+                                               context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rsend_init(const void *buf, int count,
-                                                 MPI_Datatype datatype, int rank, int tag,
-                                                 MPIR_Comm * comm, int context_offset,
-                                                 MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ssend_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                            request);
+    return MPIDI_NM_native_func->mpi_ssend_init(buf, count, datatype, rank, tag, comm,
+                                                context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_bsend_init(const void *buf, int count,
-                                                 MPI_Datatype datatype, int rank, int tag,
-                                                 MPIR_Comm * comm, int context_offset,
-                                                 MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rsend_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                            request);
+    return MPIDI_NM_native_func->mpi_rsend_init(buf, count, datatype, rank, tag, comm,
+                                                context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_isend(const void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_bsend_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->isend(buf, count, datatype, rank, tag, comm, context_offset,
-                                       request);
+    return MPIDI_NM_native_func->mpi_bsend_init(buf, count, datatype, rank, tag, comm,
+                                                context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_issend(const void *buf, int count, MPI_Datatype datatype,
-                                             int rank, int tag, MPIR_Comm * comm,
-                                             int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_isend(const void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->issend(buf, count, datatype, rank, tag, comm, context_offset,
-                                        request);
+    return MPIDI_NM_native_func->mpi_isend(buf, count, datatype, rank, tag, comm, context_offset,
+                                           request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_issend(const void *buf, int count, MPI_Datatype datatype,
+                                                 int rank, int tag, MPIR_Comm * comm,
+                                                 int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->cancel_send(sreq);
+    return MPIDI_NM_native_func->mpi_issend(buf, count, datatype, rank, tag, comm, context_offset,
+                                            request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_recv_init(void *buf, int count, MPI_Datatype datatype,
-                                                int rank, int tag, MPIR_Comm * comm,
-                                                int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
-    return MPIDI_NM_native_func->recv_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                           request);
+    return MPIDI_NM_native_func->mpi_cancel_send(sreq);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_recv(void *buf, int count, MPI_Datatype datatype,
-                                           int rank, int tag, MPIR_Comm * comm,
-                                           int context_offset, MPI_Status * status,
-                                           MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
+                                                    int rank, int tag, MPIR_Comm * comm,
+                                                    int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->recv(buf, count, datatype, rank, tag, comm, context_offset, status,
-                                      request);
+    return MPIDI_NM_native_func->mpi_recv_init(buf, count, datatype, rank, tag, comm,
+                                               context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_irecv(void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_recv(void *buf, int count, MPI_Datatype datatype,
+                                               int rank, int tag, MPIR_Comm * comm,
+                                               int context_offset, MPI_Status * status,
+                                               MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->irecv(buf, count, datatype, rank, tag, comm, context_offset,
-                                       request);
+    return MPIDI_NM_native_func->mpi_recv(buf, count, datatype, rank, tag, comm, context_offset,
+                                          status, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_imrecv(void *buf, int count, MPI_Datatype datatype,
-                                             MPIR_Request * message, MPIR_Request ** rreqp)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->imrecv(buf, count, datatype, message, rreqp);
+    return MPIDI_NM_native_func->mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset,
+                                           request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_imrecv(void *buf, int count, MPI_Datatype datatype,
+                                                 MPIR_Request * message, MPIR_Request ** rreqp)
 {
-    return MPIDI_NM_native_func->cancel_recv(rreq);
+    return MPIDI_NM_native_func->mpi_imrecv(buf, count, datatype, message, rreqp);
 };
 
-MPL_STATIC_INLINE_PREFIX void *MPIDI_NM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
-    return MPIDI_NM_native_func->alloc_mem(size, info_ptr);
+    return MPIDI_NM_native_func->mpi_cancel_recv(rreq);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_free_mem(void *ptr)
+MPL_STATIC_INLINE_PREFIX void *MPIDI_NM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
-    return MPIDI_NM_native_func->free_mem(ptr);
+    return MPIDI_NM_native_func->mpi_alloc_mem(size, info_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_improbe(int source, int tag, MPIR_Comm * comm,
-                                              int context_offset, int *flag,
-                                              MPIR_Request ** message, MPI_Status * status)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_free_mem(void *ptr)
 {
-    return MPIDI_NM_native_func->improbe(source, tag, comm, context_offset, flag, message, status);
+    return MPIDI_NM_native_func->mpi_free_mem(ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iprobe(int source, int tag, MPIR_Comm * comm,
-                                             int context_offset, int *flag, MPI_Status * status)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_improbe(int source, int tag, MPIR_Comm * comm,
+                                                  int context_offset, int *flag,
+                                                  MPIR_Request ** message, MPI_Status * status)
 {
-    return MPIDI_NM_native_func->iprobe(source, tag, comm, context_offset, flag, status);
+    return MPIDI_NM_native_func->mpi_improbe(source, tag, comm, context_offset, flag, message,
+                                             status);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iprobe(int source, int tag, MPIR_Comm * comm,
+                                                 int context_offset, int *flag, MPI_Status * status)
 {
-    return MPIDI_NM_native_func->win_set_info(win, info);
+    return MPIDI_NM_native_func->mpi_iprobe(source, tag, comm, context_offset, flag, status);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_shared_query(MPIR_Win * win, int rank,
-                                                       MPI_Aint * size, int *disp_unit,
-                                                       void *baseptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_NM_native_func->win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_NM_native_func->mpi_win_set_info(win, info);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_put(const void *origin_addr, int origin_count,
-                                          MPI_Datatype origin_datatype, int target_rank,
-                                          MPI_Aint target_disp, int target_count,
-                                          MPI_Datatype target_datatype, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win, int rank,
+                                                           MPI_Aint * size, int *disp_unit,
+                                                           void *baseptr)
 {
-    return MPIDI_NM_native_func->put(origin_addr, origin_count, origin_datatype, target_rank,
-                                     target_disp, target_count, target_datatype, win);
+    return MPIDI_NM_native_func->mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_put(const void *origin_addr, int origin_count,
+                                              MPI_Datatype origin_datatype, int target_rank,
+                                              MPI_Aint target_disp, int target_count,
+                                              MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_start(group, assert, win);
+    return MPIDI_NM_native_func->mpi_put(origin_addr, origin_count, origin_datatype, target_rank,
+                                         target_disp, target_count, target_datatype, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_complete(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_complete(win);
+    return MPIDI_NM_native_func->mpi_win_start(group, assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_post(group, assert, win);
+    return MPIDI_NM_native_func->mpi_win_complete(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_wait(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_wait(win);
+    return MPIDI_NM_native_func->mpi_win_post(group, assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_test(win, flag);
+    return MPIDI_NM_native_func->mpi_win_wait(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_NM_native_func->win_lock(lock_type, rank, assert, win);
+    return MPIDI_NM_native_func->mpi_win_test(win, flag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert,
+                                                   MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_unlock(rank, win);
+    return MPIDI_NM_native_func->mpi_win_lock(lock_type, rank, assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_get_info(win, info_p_p);
+    return MPIDI_NM_native_func->mpi_win_unlock(rank, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_get(void *origin_addr, int origin_count,
-                                          MPI_Datatype origin_datatype, int target_rank,
-                                          MPI_Aint target_disp, int target_count,
-                                          MPI_Datatype target_datatype, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_NM_native_func->get(origin_addr, origin_count, origin_datatype, target_rank,
-                                     target_disp, target_count, target_datatype, win);
+    return MPIDI_NM_native_func->mpi_win_get_info(win, info_p_p);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_get(void *origin_addr, int origin_count,
+                                              MPI_Datatype origin_datatype, int target_rank,
+                                              MPI_Aint target_disp, int target_count,
+                                              MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_free(win_ptr);
+    return MPIDI_NM_native_func->mpi_get(origin_addr, origin_count, origin_datatype, target_rank,
+                                         target_disp, target_count, target_datatype, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
-    return MPIDI_NM_native_func->win_fence(assert, win);
+    return MPIDI_NM_native_func->mpi_win_free(win_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_create(void *base, MPI_Aint length, int disp_unit,
-                                                 MPIR_Info * info, MPIR_Comm * comm_ptr,
-                                                 MPIR_Win ** win_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_fence(int assert, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    return MPIDI_NM_native_func->mpi_win_fence(assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_accumulate(const void *origin_addr, int origin_count,
-                                                 MPI_Datatype origin_datatype, int target_rank,
-                                                 MPI_Aint target_disp, int target_count,
-                                                 MPI_Datatype target_datatype, MPI_Op op,
-                                                 MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_create(void *base, MPI_Aint length, int disp_unit,
+                                                     MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                                     MPIR_Win ** win_ptr)
 {
-    return MPIDI_NM_native_func->accumulate(origin_addr, origin_count, origin_datatype, target_rank,
-                                            target_disp, target_count, target_datatype, op, win);
+    return MPIDI_NM_native_func->mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_accumulate(const void *origin_addr, int origin_count,
+                                                     MPI_Datatype origin_datatype, int target_rank,
+                                                     MPI_Aint target_disp, int target_count,
+                                                     MPI_Datatype target_datatype, MPI_Op op,
+                                                     MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_attach(win, base, size);
+    return MPIDI_NM_native_func->mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                                target_rank, target_disp, target_count,
+                                                target_datatype, op, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_allocate_shared(MPI_Aint size, int disp_unit,
-                                                          MPIR_Info * info_ptr,
-                                                          MPIR_Comm * comm_ptr,
-                                                          void **base_ptr, MPIR_Win ** win_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_NM_native_func->win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
-                                                     win_ptr);
+    return MPIDI_NM_native_func->mpi_win_attach(win, base, size);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rput(const void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPIR_Win * win,
-                                           MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size, int disp_unit,
+                                                              MPIR_Info * info_ptr,
+                                                              MPIR_Comm * comm_ptr,
+                                                              void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_NM_native_func->rput(origin_addr, origin_count, origin_datatype, target_rank,
-                                      target_disp, target_count, target_datatype, win, request);
+    return MPIDI_NM_native_func->mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr,
+                                                         base_ptr, win_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rput(const void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype, MPIR_Win * win,
+                                               MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->win_flush_local(rank, win);
+    return MPIDI_NM_native_func->mpi_rput(origin_addr, origin_count, origin_datatype, target_rank,
+                                          target_disp, target_count, target_datatype, win, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_detach(win, base);
+    return MPIDI_NM_native_func->mpi_win_flush_local(rank, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                                       const void *compare_addr,
-                                                       void *result_addr,
-                                                       MPI_Datatype datatype, int target_rank,
-                                                       MPI_Aint target_disp, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_NM_native_func->compare_and_swap(origin_addr, compare_addr, result_addr, datatype,
-                                                  target_rank, target_disp, win);
+    return MPIDI_NM_native_func->mpi_win_detach(win, base);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_raccumulate(const void *origin_addr, int origin_count,
-                                                  MPI_Datatype origin_datatype,
-                                                  int target_rank, MPI_Aint target_disp,
-                                                  int target_count,
-                                                  MPI_Datatype target_datatype, MPI_Op op,
-                                                  MPIR_Win * win, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                           const void *compare_addr,
+                                                           void *result_addr,
+                                                           MPI_Datatype datatype, int target_rank,
+                                                           MPI_Aint target_disp, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->raccumulate(origin_addr, origin_count, origin_datatype,
-                                             target_rank, target_disp, target_count,
-                                             target_datatype, op, win, request);
+    return MPIDI_NM_native_func->mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                                      datatype, target_rank, target_disp, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rget_accumulate(const void *origin_addr,
-                                                      int origin_count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_raccumulate(const void *origin_addr, int origin_count,
                                                       MPI_Datatype origin_datatype,
-                                                      void *result_addr, int result_count,
-                                                      MPI_Datatype result_datatype,
                                                       int target_rank, MPI_Aint target_disp,
                                                       int target_count,
                                                       MPI_Datatype target_datatype, MPI_Op op,
                                                       MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                                 result_addr, result_count, result_datatype,
+    return MPIDI_NM_native_func->mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                                  target_rank, target_disp, target_count,
                                                  target_datatype, op, win, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_fetch_and_op(const void *origin_addr, void *result_addr,
-                                                   MPI_Datatype datatype, int target_rank,
-                                                   MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                                          int origin_count,
+                                                          MPI_Datatype origin_datatype,
+                                                          void *result_addr, int result_count,
+                                                          MPI_Datatype result_datatype,
+                                                          int target_rank, MPI_Aint target_disp,
+                                                          int target_count,
+                                                          MPI_Datatype target_datatype, MPI_Op op,
+                                                          MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->fetch_and_op(origin_addr, result_addr, datatype, target_rank,
-                                              target_disp, op, win);
+    return MPIDI_NM_native_func->mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                                     result_addr, result_count, result_datatype,
+                                                     target_rank, target_disp, target_count,
+                                                     target_datatype, op, win, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_allocate(MPI_Aint size, int disp_unit,
-                                                   MPIR_Info * info, MPIR_Comm * comm,
-                                                   void *baseptr, MPIR_Win ** win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr, void *result_addr,
+                                                       MPI_Datatype datatype, int target_rank,
+                                                       MPI_Aint target_disp, MPI_Op op,
+                                                       MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_allocate(size, disp_unit, info, comm, baseptr, win);
+    return MPIDI_NM_native_func->mpi_fetch_and_op(origin_addr, result_addr, datatype, target_rank,
+                                                  target_disp, op, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_allocate(MPI_Aint size, int disp_unit,
+                                                       MPIR_Info * info, MPIR_Comm * comm,
+                                                       void *baseptr, MPIR_Win ** win)
 {
-    return MPIDI_NM_native_func->win_flush(rank, win);
+    return MPIDI_NM_native_func->mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_flush_local_all(win);
+    return MPIDI_NM_native_func->mpi_win_flush(rank, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_unlock_all(win);
+    return MPIDI_NM_native_func->mpi_win_flush_local_all(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
-                                                         MPIR_Win ** win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_create_dynamic(info, comm, win);
+    return MPIDI_NM_native_func->mpi_win_unlock_all(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rget(void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPIR_Win * win,
-                                           MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                             MPIR_Win ** win)
 {
-    return MPIDI_NM_native_func->rget(origin_addr, origin_count, origin_datatype, target_rank,
-                                      target_disp, target_count, target_datatype, win, request);
+    return MPIDI_NM_native_func->mpi_win_create_dynamic(info, comm, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_sync(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_rget(void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype, MPIR_Win * win,
+                                               MPIR_Request ** request)
 {
-    return MPIDI_NM_native_func->win_sync(win);
+    return MPIDI_NM_native_func->mpi_rget(origin_addr, origin_count, origin_datatype, target_rank,
+                                          target_disp, target_count, target_datatype, win, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_flush_all(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_flush_all(win);
+    return MPIDI_NM_native_func->mpi_win_sync(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_get_accumulate(const void *origin_addr, int origin_count,
-                                                     MPI_Datatype origin_datatype,
-                                                     void *result_addr, int result_count,
-                                                     MPI_Datatype result_datatype,
-                                                     int target_rank, MPI_Aint target_disp,
-                                                     int target_count,
-                                                     MPI_Datatype target_datatype, MPI_Op op,
-                                                     MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->get_accumulate(origin_addr, origin_count, origin_datatype,
-                                                result_addr, result_count, result_datatype,
-                                                target_rank, target_disp, target_count,
-                                                target_datatype, op, win);
+    return MPIDI_NM_native_func->mpi_win_flush_all(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_get_accumulate(const void *origin_addr, int origin_count,
+                                                         MPI_Datatype origin_datatype,
+                                                         void *result_addr, int result_count,
+                                                         MPI_Datatype result_datatype,
+                                                         int target_rank, MPI_Aint target_disp,
+                                                         int target_count,
+                                                         MPI_Datatype target_datatype, MPI_Op op,
+                                                         MPIR_Win * win)
 {
-    return MPIDI_NM_native_func->win_lock_all(assert, win);
+    return MPIDI_NM_native_func->mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                                    result_addr, result_count, result_datatype,
+                                                    target_rank, target_disp, target_count,
+                                                    target_datatype, op, win);
+};
+
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
+{
+    return MPIDI_NM_native_func->mpi_win_lock_all(assert, win);
 };
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rank_is_local(int target, MPIR_Comm * comm)
@@ -561,426 +566,433 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_NM_rank_is_local(int target, MPIR_Comm * comm
     return MPIDI_NM_native_func->rank_is_local(target, comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_barrier(MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_barrier(MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->barrier(comm, errflag);
+    return MPIDI_NM_native_func->mpi_barrier(comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                            int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                                int root, MPIR_Comm * comm,
+                                                MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->bcast(buffer, count, datatype, root, comm, errflag);
+    return MPIDI_NM_native_func->mpi_bcast(buffer, count, datatype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                                MPI_Datatype datatype, MPI_Op op,
-                                                MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                                    MPI_Datatype datatype, MPI_Op op,
+                                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->allreduce(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_NM_native_func->mpi_allreduce(sendbuf, recvbuf, count, datatype, op, comm,
+                                               errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_allgather(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_allgather(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                           recvtype, comm, errflag);
+    return MPIDI_NM_native_func->mpi_allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                               recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 const int *recvcounts, const int *displs,
-                                                 MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                 MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_allgatherv(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     const int *recvcounts, const int *displs,
+                                                     MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                     MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
-                                            displs, recvtype, comm, errflag);
+    return MPIDI_NM_native_func->mpi_allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                                displs, recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_scatter(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_scatter(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
-                                         root, comm, errflag);
+    return MPIDI_NM_native_func->mpi_scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                             recvtype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
-                                               const int *displs, MPI_Datatype sendtype,
-                                               void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                                   const int *displs, MPI_Datatype sendtype,
+                                                   void *recvbuf, int recvcount,
+                                                   MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount,
-                                          recvtype, root, comm_ptr, errflag);
+    return MPIDI_NM_native_func->mpi_scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
+                                              recvcount, recvtype, root, comm_ptr, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_gather(const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf,
-                                             int recvcount, MPI_Datatype recvtype, int root,
-                                             MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_gather(const void *sendbuf, int sendcount,
+                                                 MPI_Datatype sendtype, void *recvbuf,
+                                                 int recvcount, MPI_Datatype recvtype, int root,
+                                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->gather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
-                                        root, comm, errflag);
+    return MPIDI_NM_native_func->mpi_gather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                            recvtype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_gatherv(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              const int *recvcounts, const int *displs,
-                                              MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_gatherv(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  const int *recvcounts, const int *displs,
+                                                  MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs,
-                                         recvtype, root, comm, errflag);
+    return MPIDI_NM_native_func->mpi_gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                             displs, recvtype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_alltoall(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype,
-                                               MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_alltoall(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                          recvtype, comm, errflag);
+    return MPIDI_NM_native_func->mpi_alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                              recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                                const int *sdispls, MPI_Datatype sendtype,
-                                                void *recvbuf, const int *recvcounts,
-                                                const int *rdispls, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                                    const int *sdispls, MPI_Datatype sendtype,
+                                                    void *recvbuf, const int *recvcounts,
+                                                    const int *rdispls, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->alltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
-                                           recvcounts, rdispls, recvtype, comm, errflag);
+    return MPIDI_NM_native_func->mpi_alltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
+                                               recvcounts, rdispls, recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_alltoallw(const void *sendbuf, const int *sendcounts,
-                                                const int *sdispls,
-                                                const MPI_Datatype sendtypes[], void *recvbuf,
-                                                const int *recvcounts, const int *rdispls,
-                                                const MPI_Datatype recvtypes[],
-                                                MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_alltoallw(const void *sendbuf, const int *sendcounts,
+                                                    const int *sdispls,
+                                                    const MPI_Datatype sendtypes[], void *recvbuf,
+                                                    const int *recvcounts, const int *rdispls,
+                                                    const MPI_Datatype recvtypes[],
+                                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->alltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
-                                           recvcounts, rdispls, recvtypes, comm, errflag);
+    return MPIDI_NM_native_func->mpi_alltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
+                                               recvcounts, rdispls, recvtypes, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                             MPI_Datatype datatype, MPI_Op op, int root,
-                                             MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                                 MPI_Datatype datatype, MPI_Op op, int root,
+                                                 MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->reduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
-                                        errflag);
+    return MPIDI_NM_native_func->mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
+                                            errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                                     const int *recvcounts,
-                                                     MPI_Datatype datatype, MPI_Op op,
-                                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                                         const int *recvcounts,
+                                                         MPI_Datatype datatype, MPI_Op op,
+                                                         MPIR_Comm * comm_ptr,
+                                                         MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->reduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op,
-                                                comm_ptr, errflag);
+    return MPIDI_NM_native_func->mpi_reduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op,
+                                                    comm_ptr, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                           int recvcount,
-                                                           MPI_Datatype datatype, MPI_Op op,
-                                                           MPIR_Comm * comm_ptr,
-                                                           MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                               int recvcount,
+                                                               MPI_Datatype datatype, MPI_Op op,
+                                                               MPIR_Comm * comm_ptr,
+                                                               MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op,
-                                                      comm_ptr, errflag);
+    return MPIDI_NM_native_func->mpi_reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op,
+                                                          comm_ptr, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
-                                           MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                           MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                               MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                               MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->scan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_NM_native_func->mpi_scan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                             MPI_Datatype datatype, MPI_Op op,
-                                             MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                                 MPI_Datatype datatype, MPI_Op op,
+                                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_native_func->exscan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_NM_native_func->mpi_exscan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         int recvcount, MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             int recvcount, MPI_Datatype recvtype,
+                                                             MPIR_Comm * comm)
 {
-    return MPIDI_NM_native_func->neighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
-                                                    recvcount, recvtype, comm);
+    return MPIDI_NM_native_func->mpi_neighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
+                                                        recvcount, recvtype, comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype, void *recvbuf,
-                                                          const int *recvcounts,
-                                                          const int *displs,
-                                                          MPI_Datatype recvtype, MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const int *displs,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm)
 {
-    return MPIDI_NM_native_func->neighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
-                                                     recvcounts, displs, recvtype, comm);
+    return MPIDI_NM_native_func->mpi_neighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                                         recvcounts, displs, recvtype, comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_alltoallv(const void *sendbuf,
-                                                         const int *sendcounts,
-                                                         const int *sdispls,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         const int *recvcounts,
-                                                         const int *rdispls,
-                                                         MPI_Datatype recvtype, MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_alltoallv(const void *sendbuf,
+                                                             const int *sendcounts,
+                                                             const int *sdispls,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             const int *recvcounts,
+                                                             const int *rdispls,
+                                                             MPI_Datatype recvtype,
+                                                             MPIR_Comm * comm)
 {
-    return MPIDI_NM_native_func->neighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
-                                                    recvcounts, rdispls, recvtype, comm);
+    return MPIDI_NM_native_func->mpi_neighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
+                                                        recvbuf, recvcounts, rdispls, recvtype,
+                                                        comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_alltoallw(const void *sendbuf,
-                                                         const int *sendcounts,
-                                                         const MPI_Aint * sdispls,
-                                                         const MPI_Datatype * sendtypes,
-                                                         void *recvbuf, const int *recvcounts,
-                                                         const MPI_Aint * rdispls,
-                                                         const MPI_Datatype * recvtypes,
-                                                         MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_alltoallw(const void *sendbuf,
+                                                             const int *sendcounts,
+                                                             const MPI_Aint * sdispls,
+                                                             const MPI_Datatype * sendtypes,
+                                                             void *recvbuf, const int *recvcounts,
+                                                             const MPI_Aint * rdispls,
+                                                             const MPI_Datatype * recvtypes,
+                                                             MPIR_Comm * comm)
 {
-    return MPIDI_NM_native_func->neighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
-                                                    recvbuf, recvcounts, rdispls, recvtypes, comm);
+    return MPIDI_NM_native_func->mpi_neighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
+                                                        recvbuf, recvcounts, rdispls, recvtypes,
+                                                        comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                                        MPI_Datatype sendtype, void *recvbuf,
-                                                        int recvcount, MPI_Datatype recvtype,
-                                                        MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                            MPI_Datatype sendtype, void *recvbuf,
+                                                            int recvcount, MPI_Datatype recvtype,
+                                                            MPIR_Comm * comm)
 {
-    return MPIDI_NM_native_func->neighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                                   recvtype, comm);
+    return MPIDI_NM_native_func->mpi_neighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
+                                                       recvcount, recvtype, comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype, void *recvbuf,
-                                                          int recvcount, MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              int recvcount, MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ineighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
-                                                     recvcount, recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_ineighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
+                                                         recvcount, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const int *displs,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                               MPI_Datatype sendtype,
+                                                               void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const int *displs,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ineighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
-                                                      recvcounts, displs, recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_ineighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                                          recvcounts, displs, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         int recvcount, MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             int recvcount, MPI_Datatype recvtype,
+                                                             MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ineighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
-                                                    recvcount, recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_ineighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
+                                                        recvcount, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const int *sdispls,
-                                                          MPI_Datatype sendtype, void *recvbuf,
-                                                          const int *recvcounts,
-                                                          const int *rdispls,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_alltoallv(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const int *sdispls,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const int *rdispls,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ineighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
-                                                     recvbuf, recvcounts, rdispls, recvtype, comm,
-                                                     req);
+    return MPIDI_NM_native_func->mpi_ineighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
+                                                         recvbuf, recvcounts, rdispls, recvtype,
+                                                         comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const MPI_Aint * sdispls,
-                                                          const MPI_Datatype * sendtypes,
-                                                          void *recvbuf, const int *recvcounts,
-                                                          const MPI_Aint * rdispls,
-                                                          const MPI_Datatype * recvtypes,
-                                                          MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ineighbor_alltoallw(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const MPI_Aint * sdispls,
+                                                              const MPI_Datatype * sendtypes,
+                                                              void *recvbuf, const int *recvcounts,
+                                                              const MPI_Aint * rdispls,
+                                                              const MPI_Datatype * recvtypes,
+                                                              MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ineighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
-                                                     recvbuf, recvcounts, rdispls, recvtypes, comm,
-                                                     req);
+    return MPIDI_NM_native_func->mpi_ineighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
+                                                         recvbuf, recvcounts, rdispls, recvtypes,
+                                                         comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ibarrier(MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ibarrier(MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ibarrier(comm, req);
+    return MPIDI_NM_native_func->mpi_ibarrier(comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                             int root, MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                                 int root, MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ibcast(buffer, count, datatype, root, comm, req);
+    return MPIDI_NM_native_func->mpi_ibcast(buffer, count, datatype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iallgather(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 int recvcount, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iallgather(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     int recvcount, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                            recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                                recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount,
-                                                  MPI_Datatype sendtype, void *recvbuf,
-                                                  const int *recvcounts, const int *displs,
-                                                  MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                  MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                                      MPI_Datatype sendtype, void *recvbuf,
+                                                      const int *recvcounts, const int *displs,
+                                                      MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                      MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
-                                             displs, recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                                 displs, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                                 MPI_Datatype datatype, MPI_Op op,
-                                                 MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                                     MPI_Datatype datatype, MPI_Op op,
+                                                     MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iallreduce(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_NM_native_func->mpi_iallreduce(sendbuf, recvbuf, count, datatype, op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ialltoall(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                           recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                               recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls, MPI_Datatype sendtype,
-                                                 void *recvbuf, const int *recvcounts,
-                                                 const int *rdispls, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls, MPI_Datatype sendtype,
+                                                     void *recvbuf, const int *recvcounts,
+                                                     const int *rdispls, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ialltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
-                                            recvcounts, rdispls, recvtype, comm, req);
+    return MPIDI_NM_native_func->mpi_ialltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
+                                                recvcounts, rdispls, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls,
-                                                 const MPI_Datatype sendtypes[], void *recvbuf,
-                                                 const int *recvcounts, const int *rdispls,
-                                                 const MPI_Datatype recvtypes[],
-                                                 MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls,
+                                                     const MPI_Datatype sendtypes[], void *recvbuf,
+                                                     const int *recvcounts, const int *rdispls,
+                                                     const MPI_Datatype recvtypes[],
+                                                     MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ialltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
-                                            recvcounts, rdispls, recvtypes, comm, req);
+    return MPIDI_NM_native_func->mpi_ialltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
+                                                recvcounts, rdispls, recvtypes, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op,
-                                              MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op,
+                                                  MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iexscan(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_NM_native_func->mpi_iexscan(sendbuf, recvbuf, count, datatype, op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_igather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_igather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->igather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
-                                         root, comm, req);
+    return MPIDI_NM_native_func->mpi_igather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                             recvtype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_igatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int *recvcounts, const int *displs,
-                                               MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_igatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int *recvcounts, const int *displs,
+                                                   MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs,
-                                          recvtype, root, comm, req);
+    return MPIDI_NM_native_func->mpi_igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                              displs, recvtype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                            int recvcount,
-                                                            MPI_Datatype datatype, MPI_Op op,
-                                                            MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                                int recvcount,
+                                                                MPI_Datatype datatype, MPI_Op op,
+                                                                MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op,
-                                                       comm, req);
+    return MPIDI_NM_native_func->mpi_ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype,
+                                                           op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                                      const int *recvcounts,
-                                                      MPI_Datatype datatype, MPI_Op op,
-                                                      MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                                          const int *recvcounts,
+                                                          MPI_Datatype datatype, MPI_Op op,
+                                                          MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm,
-                                                 req);
+    return MPIDI_NM_native_func->mpi_ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op,
+                                                     comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op, int root,
-                                              MPIR_Comm * comm_ptr, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op, int root,
+                                                  MPIR_Comm * comm_ptr, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->ireduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
-                                         req);
+    return MPIDI_NM_native_func->mpi_ireduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
+                                             req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                            MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                            MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                                MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iscan(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_NM_native_func->mpi_iscan(sendbuf, recvbuf, count, datatype, op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                          recvtype, root, comm, req);
+    return MPIDI_NM_native_func->mpi_iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                              recvtype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                                const int *displs, MPI_Datatype sendtype,
-                                                void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, int root,
-                                                MPIR_Comm * comm_ptr, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                                    const int *displs, MPI_Datatype sendtype,
+                                                    void *recvbuf, int recvcount,
+                                                    MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
-    return MPIDI_NM_native_func->iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
-                                           recvcount, recvtype, root, comm_ptr, req);
+    return MPIDI_NM_native_func->mpi_iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
+                                               recvcount, recvtype, root, comm_ptr, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_type_create_hook(MPIR_Datatype * datatype_p)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_type_create_hook(MPIR_Datatype * datatype_p)
 {
-    return MPIDI_NM_native_func->type_create_hook(datatype_p);
+    return MPIDI_NM_native_func->mpi_type_create_hook(datatype_p);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_type_free_hook(MPIR_Datatype * datatype_p)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_type_free_hook(MPIR_Datatype * datatype_p)
 {
-    return MPIDI_NM_native_func->type_free_hook(datatype_p);
+    return MPIDI_NM_native_func->mpi_type_free_hook(datatype_p);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_op_create_hook(MPIR_Op * op_p)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_op_create_hook(MPIR_Op * op_p)
 {
-    return MPIDI_NM_native_func->op_create_hook(op_p);
+    return MPIDI_NM_native_func->mpi_op_create_hook(op_p);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_NM_op_free_hook(MPIR_Op * op_p)
+MPL_STATIC_INLINE_PREFIX int MPIDI_NM_mpi_op_free_hook(MPIR_Op * op_p)
 {
-    return MPIDI_NM_native_func->op_free_hook(op_p);
+    return MPIDI_NM_native_func->mpi_op_free_hook(op_p);
 };
 
 #endif /* NETMOD_DISABLE_INLINES  */
diff --git a/src/mpid/ch4/netmod/ofi/func_table.c b/src/mpid/ch4/netmod/ofi/func_table.c
index f1586d4..f8789f2 100644
--- a/src/mpid/ch4/netmod/ofi/func_table.c
+++ b/src/mpid/ch4/netmod/ofi/func_table.c
@@ -14,21 +14,21 @@
 #include <mpidimpl.h>
 #include "netmod_direct.h"
 MPIDI_NM_funcs_t MPIDI_NM_ofi_funcs = {
-    MPIDI_NM_init,
-    MPIDI_NM_finalize,
+    MPIDI_NM_mpi_init,
+    MPIDI_NM_mpi_finalize,
     MPIDI_NM_progress,
-    MPIDI_NM_comm_connect,
-    MPIDI_NM_comm_disconnect,
-    MPIDI_NM_open_port,
-    MPIDI_NM_close_port,
-    MPIDI_NM_comm_accept,
+    MPIDI_NM_mpi_comm_connect,
+    MPIDI_NM_mpi_comm_disconnect,
+    MPIDI_NM_mpi_open_port,
+    MPIDI_NM_mpi_close_port,
+    MPIDI_NM_mpi_comm_accept,
     MPIDI_NM_comm_get_lpid,
     MPIDI_NM_gpid_get,
     MPIDI_NM_getallincomm,
     MPIDI_NM_gpid_tolpidarray,
     MPIDI_NM_create_intercomm_from_lpids,
-    MPIDI_NM_comm_create_hook,
-    MPIDI_NM_comm_free_hook,
+    MPIDI_NM_mpi_comm_create_hook,
+    MPIDI_NM_mpi_comm_free_hook,
     MPIDI_NM_am_request_init,
     MPIDI_NM_am_request_finalize,
     MPIDI_NM_am_reg_handler,
@@ -42,108 +42,108 @@ MPIDI_NM_funcs_t MPIDI_NM_ofi_funcs = {
 };
 
 MPIDI_NM_native_funcs_t MPIDI_NM_native_ofi_funcs = {
-    MPIDI_NM_send,
-    MPIDI_NM_ssend,
-    MPIDI_NM_startall,
-    MPIDI_NM_send_init,
-    MPIDI_NM_ssend_init,
-    MPIDI_NM_rsend_init,
-    MPIDI_NM_bsend_init,
-    MPIDI_NM_isend,
-    MPIDI_NM_issend,
-    MPIDI_NM_cancel_send,
-    MPIDI_NM_recv_init,
-    MPIDI_NM_recv,
-    MPIDI_NM_irecv,
-    MPIDI_NM_imrecv,
-    MPIDI_NM_cancel_recv,
-    MPIDI_NM_alloc_mem,
-    MPIDI_NM_free_mem,
-    MPIDI_NM_improbe,
-    MPIDI_NM_iprobe,
-    MPIDI_NM_win_set_info,
-    MPIDI_NM_win_shared_query,
-    MPIDI_NM_put,
-    MPIDI_NM_win_start,
-    MPIDI_NM_win_complete,
-    MPIDI_NM_win_post,
-    MPIDI_NM_win_wait,
-    MPIDI_NM_win_test,
-    MPIDI_NM_win_lock,
-    MPIDI_NM_win_unlock,
-    MPIDI_NM_win_get_info,
-    MPIDI_NM_get,
-    MPIDI_NM_win_free,
-    MPIDI_NM_win_fence,
-    MPIDI_NM_win_create,
-    MPIDI_NM_accumulate,
-    MPIDI_NM_win_attach,
-    MPIDI_NM_win_allocate_shared,
-    MPIDI_NM_rput,
-    MPIDI_NM_win_flush_local,
-    MPIDI_NM_win_detach,
-    MPIDI_NM_compare_and_swap,
-    MPIDI_NM_raccumulate,
-    MPIDI_NM_rget_accumulate,
-    MPIDI_NM_fetch_and_op,
-    MPIDI_NM_win_allocate,
-    MPIDI_NM_win_flush,
-    MPIDI_NM_win_flush_local_all,
-    MPIDI_NM_win_unlock_all,
-    MPIDI_NM_win_create_dynamic,
-    MPIDI_NM_rget,
-    MPIDI_NM_win_sync,
-    MPIDI_NM_win_flush_all,
-    MPIDI_NM_get_accumulate,
-    MPIDI_NM_win_lock_all,
+    MPIDI_NM_mpi_send,
+    MPIDI_NM_mpi_ssend,
+    MPIDI_NM_mpi_startall,
+    MPIDI_NM_mpi_send_init,
+    MPIDI_NM_mpi_ssend_init,
+    MPIDI_NM_mpi_rsend_init,
+    MPIDI_NM_mpi_bsend_init,
+    MPIDI_NM_mpi_isend,
+    MPIDI_NM_mpi_issend,
+    MPIDI_NM_mpi_cancel_send,
+    MPIDI_NM_mpi_recv_init,
+    MPIDI_NM_mpi_recv,
+    MPIDI_NM_mpi_irecv,
+    MPIDI_NM_mpi_imrecv,
+    MPIDI_NM_mpi_cancel_recv,
+    MPIDI_NM_mpi_alloc_mem,
+    MPIDI_NM_mpi_free_mem,
+    MPIDI_NM_mpi_improbe,
+    MPIDI_NM_mpi_iprobe,
+    MPIDI_NM_mpi_win_set_info,
+    MPIDI_NM_mpi_win_shared_query,
+    MPIDI_NM_mpi_put,
+    MPIDI_NM_mpi_win_start,
+    MPIDI_NM_mpi_win_complete,
+    MPIDI_NM_mpi_win_post,
+    MPIDI_NM_mpi_win_wait,
+    MPIDI_NM_mpi_win_test,
+    MPIDI_NM_mpi_win_lock,
+    MPIDI_NM_mpi_win_unlock,
+    MPIDI_NM_mpi_win_get_info,
+    MPIDI_NM_mpi_get,
+    MPIDI_NM_mpi_win_free,
+    MPIDI_NM_mpi_win_fence,
+    MPIDI_NM_mpi_win_create,
+    MPIDI_NM_mpi_accumulate,
+    MPIDI_NM_mpi_win_attach,
+    MPIDI_NM_mpi_win_allocate_shared,
+    MPIDI_NM_mpi_rput,
+    MPIDI_NM_mpi_win_flush_local,
+    MPIDI_NM_mpi_win_detach,
+    MPIDI_NM_mpi_compare_and_swap,
+    MPIDI_NM_mpi_raccumulate,
+    MPIDI_NM_mpi_rget_accumulate,
+    MPIDI_NM_mpi_fetch_and_op,
+    MPIDI_NM_mpi_win_allocate,
+    MPIDI_NM_mpi_win_flush,
+    MPIDI_NM_mpi_win_flush_local_all,
+    MPIDI_NM_mpi_win_unlock_all,
+    MPIDI_NM_mpi_win_create_dynamic,
+    MPIDI_NM_mpi_rget,
+    MPIDI_NM_mpi_win_sync,
+    MPIDI_NM_mpi_win_flush_all,
+    MPIDI_NM_mpi_get_accumulate,
+    MPIDI_NM_mpi_win_lock_all,
     MPIDI_NM_rank_is_local,
-    MPIDI_NM_barrier,
-    MPIDI_NM_bcast,
-    MPIDI_NM_allreduce,
-    MPIDI_NM_allgather,
-    MPIDI_NM_allgatherv,
-    MPIDI_NM_scatter,
-    MPIDI_NM_scatterv,
-    MPIDI_NM_gather,
-    MPIDI_NM_gatherv,
-    MPIDI_NM_alltoall,
-    MPIDI_NM_alltoallv,
-    MPIDI_NM_alltoallw,
-    MPIDI_NM_reduce,
-    MPIDI_NM_reduce_scatter,
-    MPIDI_NM_reduce_scatter_block,
-    MPIDI_NM_scan,
-    MPIDI_NM_exscan,
-    MPIDI_NM_neighbor_allgather,
-    MPIDI_NM_neighbor_allgatherv,
-    MPIDI_NM_neighbor_alltoall,
-    MPIDI_NM_neighbor_alltoallv,
-    MPIDI_NM_neighbor_alltoallw,
-    MPIDI_NM_ineighbor_allgather,
-    MPIDI_NM_ineighbor_allgatherv,
-    MPIDI_NM_ineighbor_alltoall,
-    MPIDI_NM_ineighbor_alltoallv,
-    MPIDI_NM_ineighbor_alltoallw,
-    MPIDI_NM_ibarrier,
-    MPIDI_NM_ibcast,
-    MPIDI_NM_iallgather,
-    MPIDI_NM_iallgatherv,
-    MPIDI_NM_iallreduce,
-    MPIDI_NM_ialltoall,
-    MPIDI_NM_ialltoallv,
-    MPIDI_NM_ialltoallw,
-    MPIDI_NM_iexscan,
-    MPIDI_NM_igather,
-    MPIDI_NM_igatherv,
-    MPIDI_NM_ireduce_scatter_block,
-    MPIDI_NM_ireduce_scatter,
-    MPIDI_NM_ireduce,
-    MPIDI_NM_iscan,
-    MPIDI_NM_iscatter,
-    MPIDI_NM_iscatterv,
-    MPIDI_NM_type_create_hook,
-    MPIDI_NM_type_free_hook,
-    MPIDI_NM_op_create_hook,
-    MPIDI_NM_op_free_hook,
+    MPIDI_NM_mpi_barrier,
+    MPIDI_NM_mpi_bcast,
+    MPIDI_NM_mpi_allreduce,
+    MPIDI_NM_mpi_allgather,
+    MPIDI_NM_mpi_allgatherv,
+    MPIDI_NM_mpi_scatter,
+    MPIDI_NM_mpi_scatterv,
+    MPIDI_NM_mpi_gather,
+    MPIDI_NM_mpi_gatherv,
+    MPIDI_NM_mpi_alltoall,
+    MPIDI_NM_mpi_alltoallv,
+    MPIDI_NM_mpi_alltoallw,
+    MPIDI_NM_mpi_reduce,
+    MPIDI_NM_mpi_reduce_scatter,
+    MPIDI_NM_mpi_reduce_scatter_block,
+    MPIDI_NM_mpi_scan,
+    MPIDI_NM_mpi_exscan,
+    MPIDI_NM_mpi_neighbor_allgather,
+    MPIDI_NM_mpi_neighbor_allgatherv,
+    MPIDI_NM_mpi_neighbor_alltoall,
+    MPIDI_NM_mpi_neighbor_alltoallv,
+    MPIDI_NM_mpi_neighbor_alltoallw,
+    MPIDI_NM_mpi_ineighbor_allgather,
+    MPIDI_NM_mpi_ineighbor_allgatherv,
+    MPIDI_NM_mpi_ineighbor_alltoall,
+    MPIDI_NM_mpi_ineighbor_alltoallv,
+    MPIDI_NM_mpi_ineighbor_alltoallw,
+    MPIDI_NM_mpi_ibarrier,
+    MPIDI_NM_mpi_ibcast,
+    MPIDI_NM_mpi_iallgather,
+    MPIDI_NM_mpi_iallgatherv,
+    MPIDI_NM_mpi_iallreduce,
+    MPIDI_NM_mpi_ialltoall,
+    MPIDI_NM_mpi_ialltoallv,
+    MPIDI_NM_mpi_ialltoallw,
+    MPIDI_NM_mpi_iexscan,
+    MPIDI_NM_mpi_igather,
+    MPIDI_NM_mpi_igatherv,
+    MPIDI_NM_mpi_ireduce_scatter_block,
+    MPIDI_NM_mpi_ireduce_scatter,
+    MPIDI_NM_mpi_ireduce,
+    MPIDI_NM_mpi_iscan,
+    MPIDI_NM_mpi_iscatter,
+    MPIDI_NM_mpi_iscatterv,
+    MPIDI_NM_mpi_type_create_hook,
+    MPIDI_NM_mpi_type_free_hook,
+    MPIDI_NM_mpi_op_create_hook,
+    MPIDI_NM_mpi_op_free_hook,
 };
 #endif
diff --git a/src/mpid/ch4/netmod/ofi/ofi_am_probe.h b/src/mpid/ch4/netmod/ofi/ofi_am_probe.h
index e8575ec..58a750c 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_am_probe.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_am_probe.h
@@ -19,21 +19,21 @@ static inline int MPIDI_NM_probe(int source,
     return MPIDI_CH4U_probe(source, tag, comm, context_offset, status);
 }
 
-static inline int MPIDI_NM_improbe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset,
-                                   int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_NM_mpi_improbe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset,
+                                       int *flag, MPIR_Request ** message, MPI_Status * status)
 {
-    return MPIDI_CH4U_improbe(source, tag, comm, context_offset, flag, message, status);
+    return MPIDI_CH4U_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
 }
 
-static inline int MPIDI_NM_iprobe(int source,
-                                  int tag,
-                                  MPIR_Comm * comm,
-                                  int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_NM_mpi_iprobe(int source,
+                                      int tag,
+                                      MPIR_Comm * comm,
+                                      int context_offset, int *flag, MPI_Status * status)
 {
-    return MPIDI_CH4U_iprobe(source, tag, comm, context_offset, flag, status);
+    return MPIDI_CH4U_mpi_iprobe(source, tag, comm, context_offset, flag, status);
 }
 
 #endif /* NETMOD_AM_OFI_PROBE_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ofi/ofi_am_recv.h b/src/mpid/ch4/netmod/ofi/ofi_am_recv.h
index 0e844db..45c1932 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_am_recv.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_am_recv.h
@@ -13,49 +13,52 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_recv(void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm,
-                                int context_offset, MPI_Status * status, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv(void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm,
+                                    int context_offset, MPI_Status * status,
+                                    MPIR_Request ** request)
 {
-    return MPIDI_CH4U_recv(buf, count, datatype, rank, tag, comm, context_offset, status, request);
+    return MPIDI_CH4U_mpi_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
+                               request);
 }
 
-static inline int MPIDI_NM_recv_init(void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv_init(void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_imrecv(void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  MPIR_Request * message, MPIR_Request ** rreqp)
+static inline int MPIDI_NM_mpi_imrecv(void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      MPIR_Request * message, MPIR_Request ** rreqp)
 {
-    return MPIDI_CH4U_imrecv(buf, count, datatype, message, rreqp);
+    return MPIDI_CH4U_mpi_imrecv(buf, count, datatype, message, rreqp);
 }
 
-static inline int MPIDI_NM_irecv(void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_irecv(void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 
 }
 
-static inline int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+static inline int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
-    return MPIDI_CH4U_cancel_recv(rreq);
+    return MPIDI_CH4U_mpi_cancel_recv(rreq);
 }
 
 #endif /* NETMOD_AM_OFI_RECV_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ofi/ofi_am_rma.h b/src/mpid/ch4/netmod/ofi/ofi_am_rma.h
index e93b892..d04235a 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_am_rma.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_am_rma.h
@@ -13,136 +13,141 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_put(const void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_put(const void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_put(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_put(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_get(void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get(void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_get(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_rput(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rput(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rput(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rput(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                            const void *compare_addr,
-                                            void *result_addr,
-                                            MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win)
-{
-    return MPIDI_CH4U_compare_and_swap(origin_addr, compare_addr, result_addr,
-                                       datatype, target_rank, target_disp, win);
-}
-
-static inline int MPIDI_NM_raccumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype,
-                                       MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                const void *compare_addr,
+                                                void *result_addr,
+                                                MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win)
 {
-    return MPIDI_CH4U_raccumulate(origin_addr, origin_count, origin_datatype,
-                                  target_rank, target_disp, target_count,
-                                  target_datatype, op, win, request);
+    return MPIDI_CH4U_mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                           datatype, target_rank, target_disp, win);
 }
 
-static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_raccumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
                                            MPI_Datatype target_datatype,
                                            MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                      result_addr, result_count, result_datatype,
+    return MPIDI_CH4U_mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                       target_rank, target_disp, target_count,
                                       target_datatype, op, win, request);
 }
 
-static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
-                                        void *result_addr,
-                                        MPI_Datatype datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+{
+    return MPIDI_CH4U_mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                          result_addr, result_count, result_datatype,
+                                          target_rank, target_disp, target_count,
+                                          target_datatype, op, win, request);
+}
+
+static inline int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr,
+                                            void *result_addr,
+                                            MPI_Datatype datatype,
+                                            int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_fetch_and_op(origin_addr, result_addr, datatype,
-                                   target_rank, target_disp, op, win);
+    return MPIDI_CH4U_mpi_fetch_and_op(origin_addr, result_addr, datatype,
+                                       target_rank, target_disp, op, win);
 }
 
 
-static inline int MPIDI_NM_rget(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rget(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_get_accumulate(const void *origin_addr,
+                                              int origin_count,
+                                              MPI_Datatype origin_datatype,
+                                              void *result_addr,
+                                              int result_count,
+                                              MPI_Datatype result_datatype,
+                                              int target_rank,
+                                              MPI_Aint target_disp,
+                                              int target_count,
+                                              MPI_Datatype target_datatype, MPI_Op op,
+                                              MPIR_Win * win)
+{
+    return MPIDI_CH4U_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                         result_addr, result_count, result_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win);
+}
+
+static inline int MPIDI_NM_mpi_accumulate(const void *origin_addr,
                                           int origin_count,
                                           MPI_Datatype origin_datatype,
-                                          void *result_addr,
-                                          int result_count,
-                                          MPI_Datatype result_datatype,
                                           int target_rank,
                                           MPI_Aint target_disp,
                                           int target_count,
                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                     result_addr, result_count, result_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win);
-}
-
-static inline int MPIDI_NM_accumulate(const void *origin_addr,
-                                      int origin_count,
-                                      MPI_Datatype origin_datatype,
-                                      int target_rank,
-                                      MPI_Aint target_disp,
-                                      int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    return MPIDI_CH4U_accumulate(origin_addr, origin_count, origin_datatype,
-                                 target_rank, target_disp, target_count, target_datatype, op, win);
+    return MPIDI_CH4U_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                     target_rank, target_disp, target_count, target_datatype, op,
+                                     win);
 }
 
 #endif /* NETMOD_AM_OFI_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ofi/ofi_am_send.h b/src/mpid/ch4/netmod/ofi/ofi_am_send.h
index 777377a..3165a33 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_am_send.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_am_send.h
@@ -13,14 +13,14 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_send(const void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send(const void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
 static inline int MPIDI_NM_rsend(const void *buf,
@@ -45,84 +45,91 @@ static inline int MPIDI_NM_irsend(const void *buf,
     return MPIDI_CH4U_irsend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_CH4U_startall(count, requests);
+    return MPIDI_CH4U_mpi_startall(count, requests);
 }
 
-static inline int MPIDI_NM_send_init(const void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send_init(const void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_bsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_bsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_rsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_isend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_isend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_issend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_issend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
-    return MPIDI_CH4U_cancel_send(sreq);
+    return MPIDI_CH4U_mpi_cancel_send(sreq);
 }
 
 #endif /* NETMOD_AM_OFI_SEND_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ofi/ofi_am_spawn.h b/src/mpid/ch4/netmod/ofi/ofi_am_spawn.h
index 75a5ef5..ba6249d 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_am_spawn.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_am_spawn.h
@@ -13,35 +13,35 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_comm_connect(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_NM_mpi_comm_connect(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_close_port(const char *port_name)
+static inline int MPIDI_NM_mpi_close_port(const char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_comm_accept(const char *port_name,
-                                       MPIR_Info * info,
-                                       int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_NM_mpi_comm_accept(const char *port_name,
+                                           MPIR_Info * info,
+                                           int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_am_win.h b/src/mpid/ch4/netmod/ofi/ofi_am_win.h
index 89e5081..0670648 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_am_win.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_am_win.h
@@ -13,147 +13,150 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_CH4R_win_set_info(win, info);
+    return MPIDI_CH4R_mpi_win_set_info(win, info);
 }
 
 
-static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_start(group, assert, win);
+    return MPIDI_CH4R_mpi_win_start(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_complete(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_complete(win);
+    return MPIDI_CH4R_mpi_win_complete(win);
 }
 
-static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_post(group, assert, win);
+    return MPIDI_CH4R_mpi_win_post(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_wait(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_wait(win);
+    return MPIDI_CH4R_mpi_win_wait(win);
 }
 
 
-static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_CH4R_win_test(win, flag);
+    return MPIDI_CH4R_mpi_win_test(win, flag);
 }
 
-static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock(lock_type, rank, assert, win);
+    return MPIDI_CH4R_mpi_win_lock(lock_type, rank, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock(rank, win);
+    return MPIDI_CH4R_mpi_win_unlock(rank, win);
 }
 
-static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_CH4R_win_get_info(win, info_p_p);
+    return MPIDI_CH4R_mpi_win_get_info(win, info_p_p);
 }
 
 
-static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_free(win_ptr);
+    return MPIDI_CH4R_mpi_win_free(win_ptr);
 }
 
-static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_fence(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_fence(assert, win);
+    return MPIDI_CH4R_mpi_win_fence(assert, win);
 }
 
-static inline int MPIDI_NM_win_create(void *base,
-                                      MPI_Aint length,
-                                      int disp_unit,
-                                      MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create(void *base,
+                                          MPI_Aint length,
+                                          int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
 }
 
-static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_CH4R_win_attach(win, base, size);
+    return MPIDI_CH4R_mpi_win_attach(win, base, size);
 }
 
-static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
-                                               int disp_unit,
-                                               MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr,
-                                               void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size,
+                                                   int disp_unit,
+                                                   MPIR_Info * info_ptr,
+                                                   MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
+                                              win_ptr);
 }
 
-static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_CH4R_win_detach(win, base);
+    return MPIDI_CH4R_mpi_win_detach(win, base);
 }
 
-static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
-                                            int rank,
-                                            MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win,
+                                                int rank,
+                                                MPI_Aint * size, int *disp_unit, void *baseptr)
 {
-    return MPIDI_CH4R_win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_CH4R_mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 }
 
-static inline int MPIDI_NM_win_allocate(MPI_Aint size,
-                                        int disp_unit,
-                                        MPIR_Info * info,
-                                        MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_allocate(MPI_Aint size,
+                                            int disp_unit,
+                                            MPIR_Info * info,
+                                            MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_allocate(size, disp_unit, info, comm, baseptr, win);
+    return MPIDI_CH4R_mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
 }
 
-static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush(rank, win);
+    return MPIDI_CH4R_mpi_win_flush(rank, win);
 }
 
-static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local_all(win);
+    return MPIDI_CH4R_mpi_win_flush_local_all(win);
 }
 
-static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock_all(win);
+    return MPIDI_CH4R_mpi_win_unlock_all(win);
 }
 
-static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                  MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_create_dynamic(info, comm, win);
+    return MPIDI_CH4R_mpi_win_create_dynamic(info, comm, win);
 }
 
-static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local(rank, win);
+    return MPIDI_CH4R_mpi_win_flush_local(rank, win);
 }
 
-static inline int MPIDI_NM_win_sync(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_sync(win);
+    return MPIDI_CH4R_mpi_win_sync(win);
 }
 
-static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_all(win);
+    return MPIDI_CH4R_mpi_win_flush_all(win);
 }
 
-static inline int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock_all(assert, win);
+    return MPIDI_CH4R_mpi_win_lock_all(assert, win);
 }
 
 
diff --git a/src/mpid/ch4/netmod/ofi/ofi_coll.h b/src/mpid/ch4/netmod/ofi/ofi_coll.h
index 1f87481..a6ee413 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_coll.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_coll.h
@@ -14,10 +14,10 @@
 #include "ofi_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_barrier
+#define FUNCNAME MPIDI_NM_mpi_barrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BARRIER);
@@ -30,11 +30,11 @@ static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errfla
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_bcast
+#define FUNCNAME MPIDI_NM_mpi_bcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                 int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                     int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BCAST);
@@ -47,12 +47,12 @@ static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allreduce
+#define FUNCNAME MPIDI_NM_mpi_allreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                     MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                         MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                         MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLREDUCE);
@@ -65,12 +65,12 @@ static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int cou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgather
+#define FUNCNAME MPIDI_NM_mpi_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHER);
@@ -84,13 +84,13 @@ static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts, const int *displs,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                      MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts, const int *displs,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                          MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHERV);
@@ -104,12 +104,12 @@ static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gather
+#define FUNCNAME MPIDI_NM_mpi_gather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                  void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                  int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                      int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHER);
@@ -123,13 +123,13 @@ static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Dataty
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gatherv
+#define FUNCNAME MPIDI_NM_mpi_gatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, const int *recvcounts, const int *displs,
-                                   MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, const int *recvcounts, const int *displs,
+                                       MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHERV);
@@ -143,12 +143,12 @@ static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatter
+#define FUNCNAME MPIDI_NM_mpi_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTER);
@@ -162,13 +162,13 @@ static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatterv
+#define FUNCNAME MPIDI_NM_mpi_scatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
-                                    const int *displs, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                        const int *displs, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTERV);
@@ -182,12 +182,12 @@ static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoall
+#define FUNCNAME MPIDI_NM_mpi_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALL);
@@ -201,14 +201,14 @@ static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                     const int *sdispls, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts,
-                                     const int *rdispls, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                         const int *sdispls, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts,
+                                         const int *rdispls, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLV);
@@ -221,14 +221,14 @@ static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[],
-                                     const int sdispls[], const MPI_Datatype sendtypes[],
-                                     void *recvbuf, const int recvcounts[],
-                                     const int rdispls[], const MPI_Datatype recvtypes[],
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallw(const void *sendbuf, const int sendcounts[],
+                                         const int sdispls[], const MPI_Datatype sendtypes[],
+                                         void *recvbuf, const int recvcounts[],
+                                         const int rdispls[], const MPI_Datatype recvtypes[],
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLW);
@@ -242,12 +242,12 @@ static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[]
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce
+#define FUNCNAME MPIDI_NM_mpi_reduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, int root,
-                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, int root,
+                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE);
@@ -260,12 +260,13 @@ static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                          const int recvcounts[], MPI_Datatype datatype,
-                                          MPI_Op op, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                              const int recvcounts[], MPI_Datatype datatype,
+                                              MPI_Op op, MPIR_Comm * comm_ptr,
+                                              MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER);
@@ -278,13 +279,13 @@ static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                int recvcount, MPI_Datatype datatype,
-                                                MPI_Op op, MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                    int recvcount, MPI_Datatype datatype,
+                                                    MPI_Op op, MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER_BLOCK);
@@ -298,12 +299,12 @@ static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvb
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scan
+#define FUNCNAME MPIDI_NM_mpi_scan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
-                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCAN);
@@ -316,12 +317,12 @@ static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_exscan
+#define FUNCNAME MPIDI_NM_mpi_exscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                  MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                      MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_EXSCAN);
@@ -334,12 +335,13 @@ static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHER);
@@ -354,13 +356,13 @@ static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int recvcounts[], const int displs[],
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int recvcounts[], const int displs[],
+                                                   MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHERV);
@@ -374,12 +376,13 @@ static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                             MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                 MPI_Datatype sendtype, void *recvbuf,
+                                                 int recvcount, MPI_Datatype recvtype,
+                                                 MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALL);
@@ -393,14 +396,14 @@ static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                              const int sdispls[], MPI_Datatype sendtype,
-                                              void *recvbuf, const int recvcounts[],
-                                              const int rdispls[], MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                  const int sdispls[], MPI_Datatype sendtype,
+                                                  void *recvbuf, const int recvcounts[],
+                                                  const int rdispls[], MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLV);
@@ -414,14 +417,15 @@ static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                              const MPI_Aint sdispls[],
-                                              const MPI_Datatype sendtypes[], void *recvbuf,
-                                              const int recvcounts[], const MPI_Aint rdispls[],
-                                              const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                  const MPI_Aint sdispls[],
+                                                  const MPI_Datatype sendtypes[], void *recvbuf,
+                                                  const int recvcounts[], const MPI_Aint rdispls[],
+                                                  const MPI_Datatype recvtypes[],
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLW);
@@ -435,13 +439,13 @@ static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHER);
@@ -455,14 +459,14 @@ static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int recvcounts[], const int displs[],
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int recvcounts[], const int displs[],
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                    MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHERV);
@@ -477,13 +481,13 @@ static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALL);
@@ -497,14 +501,14 @@ static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                               const int sdispls[], MPI_Datatype sendtype,
-                                               void *recvbuf, const int recvcounts[],
-                                               const int rdispls[], MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                   const int sdispls[], MPI_Datatype sendtype,
+                                                   void *recvbuf, const int recvcounts[],
+                                                   const int rdispls[], MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLV);
@@ -519,15 +523,15 @@ static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                               const MPI_Aint sdispls[],
-                                               const MPI_Datatype sendtypes[], void *recvbuf,
-                                               const int recvcounts[], const MPI_Aint rdispls[],
-                                               const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                   const MPI_Aint sdispls[],
+                                                   const MPI_Datatype sendtypes[], void *recvbuf,
+                                                   const int recvcounts[], const MPI_Aint rdispls[],
+                                                   const MPI_Datatype recvtypes[],
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLW);
@@ -542,10 +546,10 @@ static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibarrier
+#define FUNCNAME MPIDI_NM_mpi_ibarrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBARRIER);
@@ -558,11 +562,11 @@ static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibcast
+#define FUNCNAME MPIDI_NM_mpi_ibcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                  int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                      int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBCAST);
@@ -575,12 +579,12 @@ static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgather
+#define FUNCNAME MPIDI_NM_mpi_iallgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHER);
@@ -594,13 +598,14 @@ static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgatherv
+#define FUNCNAME MPIDI_NM_mpi_iallgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                       MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHERV);
@@ -614,12 +619,12 @@ static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallreduce
+#define FUNCNAME MPIDI_NM_mpi_iallreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                      MPI_Request * request)
+static inline int MPIDI_NM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                          MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLREDUCE);
@@ -632,12 +637,12 @@ static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int co
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoall
+#define FUNCNAME MPIDI_NM_mpi_ialltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALL);
@@ -651,14 +656,14 @@ static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallv
+#define FUNCNAME MPIDI_NM_mpi_ialltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLV);
@@ -673,14 +678,14 @@ static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallw
+#define FUNCNAME MPIDI_NM_mpi_ialltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, const MPI_Datatype recvtypes[],
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, const MPI_Datatype recvtypes[],
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLW);
@@ -695,12 +700,12 @@ static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iexscan
+#define FUNCNAME MPIDI_NM_mpi_iexscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                   MPI_Request * req)
+static inline int MPIDI_NM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                       MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IEXSCAN);
@@ -713,12 +718,12 @@ static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igather
+#define FUNCNAME MPIDI_NM_mpi_igather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHER);
@@ -732,13 +737,13 @@ static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igatherv
+#define FUNCNAME MPIDI_NM_mpi_igatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPI_Request * req)
+static inline int MPIDI_NM_mpi_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                        MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHERV);
@@ -752,12 +757,13 @@ static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                 int recvcount, MPI_Datatype datatype,
-                                                 MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER_BLOCK);
@@ -771,12 +777,12 @@ static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recv
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                           const int recvcounts[], MPI_Datatype datatype,
-                                           MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                               const int recvcounts[], MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER);
@@ -790,12 +796,12 @@ static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce
+#define FUNCNAME MPIDI_NM_mpi_ireduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root,
-                                   MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE);
@@ -808,12 +814,12 @@ static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscan
+#define FUNCNAME MPIDI_NM_mpi_iscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                 MPI_Request * req)
+static inline int MPIDI_NM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCAN);
@@ -826,13 +832,13 @@ static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatter
+#define FUNCNAME MPIDI_NM_mpi_iscatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
-                                    MPI_Datatype sendtype, void *recvbuf,
-                                    int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                        MPI_Datatype sendtype, void *recvbuf,
+                                        int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTER);
@@ -846,14 +852,14 @@ static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatterv
+#define FUNCNAME MPIDI_NM_mpi_iscatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                     const int *displs, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount,
-                                     MPI_Datatype recvtype, int root,
-                                     MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount,
+                                         MPI_Datatype recvtype, int root,
+                                         MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTERV);
diff --git a/src/mpid/ch4/netmod/ofi/ofi_comm.h b/src/mpid/ch4/netmod/ofi/ofi_comm.h
index bf17dab..3c7ed98 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_comm.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_comm.h
@@ -15,10 +15,10 @@
 #include "mpl_utlist.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_create_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_COMM_CREATE);
@@ -42,10 +42,10 @@ static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_free_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_free_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_COMM_DESTROY);
diff --git a/src/mpid/ch4/netmod/ofi/ofi_datatype.h b/src/mpid/ch4/netmod/ofi/ofi_datatype.h
index 7a12b23..9e6d964 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_datatype.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_datatype.h
@@ -13,12 +13,12 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_type_free_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_free_hook(MPIR_Datatype * datatype_p)
 {
     return 0;
 }
 
-static inline int MPIDI_NM_type_create_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_create_hook(MPIR_Datatype * datatype_p)
 {
     return 0;
 }
diff --git a/src/mpid/ch4/netmod/ofi/ofi_init.h b/src/mpid/ch4/netmod/ofi/ofi_init.h
index 149f159..9046bf2 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_init.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_init.h
@@ -392,7 +392,7 @@ static inline int MPIDI_OFI_init_generic(int rank,
         MPIR_Assert(MPIDI_OFI_DEFAULT_SHORT_SEND_SIZE <= MPIDI_Global.max_send);
         MPIDI_Global.am_buf_pool =
             MPIDI_CH4U_create_buf_pool(MPIDI_OFI_BUF_POOL_NUM, MPIDI_OFI_BUF_POOL_SIZE);
-        mpi_errno = MPIDI_CH4U_init(comm_world, comm_self, num_contexts, netmod_contexts);
+        mpi_errno = MPIDI_CH4U_mpi_init(comm_world, comm_self, num_contexts, netmod_contexts);
 
         if (mpi_errno)
             MPIR_ERR_POP(mpi_errno);
@@ -482,16 +482,16 @@ static inline int MPIDI_OFI_init_generic(int rank,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_init
+#define FUNCNAME MPIDI_NM_mpi_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_init(int rank,
-                                int size,
-                                int appnum,
-                                int *tag_ub,
-                                MPIR_Comm * comm_world,
-                                MPIR_Comm * comm_self,
-                                int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init(int rank,
+                                    int size,
+                                    int appnum,
+                                    int *tag_ub,
+                                    MPIR_Comm * comm_world,
+                                    MPIR_Comm * comm_self,
+                                    int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno;
     mpi_errno = MPIDI_OFI_init_generic(rank, size, appnum, tag_ub, comm_world,
@@ -562,7 +562,7 @@ static inline int MPIDI_OFI_finalize_generic(int do_scalable_ep, int do_am, int
     comm = MPIR_Process.comm_self;
     MPIR_Comm_release_always(comm);
 
-    MPIDI_CH4U_finalize();
+    MPIDI_CH4U_mpi_finalize();
 
     MPIDI_OFI_map_destroy(MPIDI_Global.win_map);
 
@@ -589,13 +589,13 @@ static inline int MPIDI_OFI_finalize_generic(int do_scalable_ep, int do_am, int
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_finalize(void)
+static inline int MPIDI_NM_mpi_finalize(void)
 {
     return MPIDI_OFI_finalize_generic(MPIDI_OFI_ENABLE_SCALABLE_ENDPOINTS,
                                       MPIDI_OFI_ENABLE_AM, MPIDI_OFI_ENABLE_STX_RMA);
 }
 
-static inline void *MPIDI_NM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+static inline void *MPIDI_NM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
 
     void *ap;
@@ -603,7 +603,7 @@ static inline void *MPIDI_NM_alloc_mem(size_t size, MPIR_Info * info_ptr)
     return ap;
 }
 
-static inline int MPIDI_NM_free_mem(void *ptr)
+static inline int MPIDI_NM_mpi_free_mem(void *ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPL_free(ptr);
diff --git a/src/mpid/ch4/netmod/ofi/ofi_op.h b/src/mpid/ch4/netmod/ofi/ofi_op.h
index 7fb4904..f21c99b 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_op.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_op.h
@@ -13,12 +13,12 @@
 
 #include "ofi_impl.h"
 
-static inline int MPIDI_NM_op_free_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_free_hook(MPIR_Op * op_p)
 {
     return 0;
 }
 
-static inline int MPIDI_NM_op_create_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_create_hook(MPIR_Op * op_p)
 {
     return 0;
 }
diff --git a/src/mpid/ch4/netmod/ofi/ofi_probe.h b/src/mpid/ch4/netmod/ofi/ofi_probe.h
index b2eaff4..e7b57c1 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_probe.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_probe.h
@@ -124,14 +124,14 @@ static inline int MPIDI_NM_probe(int source,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_improbe
+#define FUNCNAME MPIDI_NM_mpi_improbe
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_improbe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset,
-                                   int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_NM_mpi_improbe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset,
+                                       int *flag, MPIR_Request ** message, MPI_Status * status)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_NETMOD_IMPROBE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_NETMOD_IMPROBE);
@@ -149,13 +149,13 @@ static inline int MPIDI_NM_improbe(int source,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iprobe
+#define FUNCNAME MPIDI_NM_mpi_iprobe
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iprobe(int source,
-                                  int tag,
-                                  MPIR_Comm * comm,
-                                  int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_NM_mpi_iprobe(int source,
+                                      int tag,
+                                      MPIR_Comm * comm,
+                                      int context_offset, int *flag, MPI_Status * status)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_NETMOD_IPROBE);
diff --git a/src/mpid/ch4/netmod/ofi/ofi_recv.h b/src/mpid/ch4/netmod/ofi/ofi_recv.h
index bed547d..375b73d 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_recv.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_recv.h
@@ -117,17 +117,17 @@ __ALWAYS_INLINE__ int MPIDI_OFI_do_irecv(void *buf,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_recv
+#define FUNCNAME MPIDI_NM_mpi_recv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_recv(void *buf,
-                                    int count,
-                                    MPI_Datatype datatype,
-                                    int rank,
-                                    int tag,
-                                    MPIR_Comm * comm,
-                                    int context_offset,
-                                    MPI_Status * status, MPIR_Request ** request)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv(void *buf,
+                                        int count,
+                                        MPI_Datatype datatype,
+                                        int rank,
+                                        int tag,
+                                        MPIR_Comm * comm,
+                                        int context_offset,
+                                        MPI_Status * status, MPIR_Request ** request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RECV);
@@ -139,16 +139,16 @@ __ALWAYS_INLINE__ int MPIDI_NM_recv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_recv_init
+#define FUNCNAME MPIDI_NM_mpi_recv_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_recv_init(void *buf,
-                                         int count,
-                                         MPI_Datatype datatype,
-                                         int rank,
-                                         int tag,
-                                         MPIR_Comm * comm,
-                                         int context_offset, MPIR_Request ** request)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv_init(void *buf,
+                                             int count,
+                                             MPI_Datatype datatype,
+                                             int rank,
+                                             int tag,
+                                             MPIR_Comm * comm,
+                                             int context_offset, MPIR_Request ** request)
 {
     MPIR_Request *rreq;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RECV_INIT);
@@ -184,13 +184,13 @@ __ALWAYS_INLINE__ int MPIDI_NM_recv_init(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_imrecv
+#define FUNCNAME MPIDI_NM_mpi_imrecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_imrecv(void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      MPIR_Request * message, MPIR_Request ** rreqp)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_imrecv(void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          MPIR_Request * message, MPIR_Request ** rreqp)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq;
@@ -220,15 +220,16 @@ __ALWAYS_INLINE__ int MPIDI_NM_imrecv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_irecv
+#define FUNCNAME MPIDI_NM_mpi_irecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_irecv(void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_irecv(void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_IRECV);
@@ -240,10 +241,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_irecv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_cancel_recv
+#define FUNCNAME MPIDI_NM_mpi_cancel_recv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
 
     int mpi_errno = MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_rma.h b/src/mpid/ch4/netmod/ofi/ofi_rma.h
index 0675b18..531c96e 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_rma.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_rma.h
@@ -377,15 +377,15 @@ static inline int MPIDI_OFI_do_put(const void *origin_addr,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_put
+#define FUNCNAME MPIDI_NM_mpi_put
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_put(const void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_put(const void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_PUT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_PUT);
@@ -525,15 +525,15 @@ static inline int MPIDI_OFI_do_get(void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_get
+#define FUNCNAME MPIDI_NM_mpi_get
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_get(void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get(void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     int origin_contig, target_contig, mpi_errno = MPI_SUCCESS;
     MPIDI_OFI_win_datatype_t origin_dt, target_dt;
@@ -605,17 +605,17 @@ static inline int MPIDI_NM_get(void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_rput
+#define FUNCNAME MPIDI_NM_mpi_rput
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_rput(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rput(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RPUT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_RPUT);
@@ -660,14 +660,15 @@ static inline int MPIDI_NM_rput(const void *origin_addr,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_compare_and_swap
+#define FUNCNAME MPIDI_NM_mpi_compare_and_swap
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                            const void *compare_addr,
-                                            void *result_addr,
-                                            MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                const void *compare_addr,
+                                                void *result_addr,
+                                                MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     enum fi_op fi_op;
@@ -894,12 +895,13 @@ static inline int MPIDI_OFI_do_accumulate(const void *origin_addr,
   am_fallback:
     /* Fall back to active message */
     MPIDI_OFI_win_request_complete(req);
-    return MPIDI_CH4U_accumulate(origin_addr, origin_count, origin_datatype,
-                                 target_rank, target_disp, target_count, target_datatype, op, win);
+    return MPIDI_CH4U_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                     target_rank, target_disp, target_count, target_datatype, op,
+                                     win);
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_get_accumulate
+#define FUNCNAME MPIDI_NM_mpi_get_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
 static inline int MPIDI_OFI_do_get_accumulate(const void *origin_addr,
@@ -1095,25 +1097,25 @@ static inline int MPIDI_OFI_do_get_accumulate(const void *origin_addr,
     goto fn_exit;
   am_fallback:
     MPIDI_OFI_win_request_complete(req);
-    return MPIDI_CH4U_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                     result_addr, result_count, result_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win);
+    return MPIDI_CH4U_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                         result_addr, result_count, result_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win);
 }
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_raccumulate
+#define FUNCNAME MPIDI_NM_mpi_raccumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_raccumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype,
-                                       MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_raccumulate(const void *origin_addr,
+                                           int origin_count,
+                                           MPI_Datatype origin_datatype,
+                                           int target_rank,
+                                           MPI_Aint target_disp,
+                                           int target_count,
+                                           MPI_Datatype target_datatype,
+                                           MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_RACCUMULATE);
@@ -1134,20 +1136,20 @@ static inline int MPIDI_NM_raccumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_rget_accumulate
+#define FUNCNAME MPIDI_NM_mpi_rget_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
-                                           int origin_count,
-                                           MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
-                                           int target_rank,
-                                           MPI_Aint target_disp,
-                                           int target_count,
-                                           MPI_Datatype target_datatype,
-                                           MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq;
@@ -1164,14 +1166,14 @@ static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_fetch_and_op
+#define FUNCNAME MPIDI_NM_mpi_fetch_and_op
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
-                                        void *result_addr,
-                                        MPI_Datatype datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr,
+                                            void *result_addr,
+                                            MPI_Datatype datatype,
+                                            int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_FETCH_AND_OP);
@@ -1189,17 +1191,17 @@ static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_rget
+#define FUNCNAME MPIDI_NM_mpi_rget
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_rget(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RGET);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_RGET);
@@ -1242,19 +1244,20 @@ static inline int MPIDI_NM_rget(void *origin_addr,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_get_accumulate
+#define FUNCNAME MPIDI_NM_mpi_get_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
-                                          int origin_count,
-                                          MPI_Datatype origin_datatype,
-                                          void *result_addr,
-                                          int result_count,
-                                          MPI_Datatype result_datatype,
-                                          int target_rank,
-                                          MPI_Aint target_disp,
-                                          int target_count,
-                                          MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get_accumulate(const void *origin_addr,
+                                              int origin_count,
+                                              MPI_Datatype origin_datatype,
+                                              void *result_addr,
+                                              int result_count,
+                                              MPI_Datatype result_datatype,
+                                              int target_rank,
+                                              MPI_Aint target_disp,
+                                              int target_count,
+                                              MPI_Datatype target_datatype, MPI_Op op,
+                                              MPIR_Win * win)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_GET_ACCUMULATE);
@@ -1268,16 +1271,16 @@ static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_accumulate
+#define FUNCNAME MPIDI_NM_mpi_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_accumulate(const void *origin_addr,
-                                      int origin_count,
-                                      MPI_Datatype origin_datatype,
-                                      int target_rank,
-                                      MPI_Aint target_disp,
-                                      int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_accumulate(const void *origin_addr,
+                                          int origin_count,
+                                          MPI_Datatype origin_datatype,
+                                          int target_rank,
+                                          MPI_Aint target_disp,
+                                          int target_count,
+                                          MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_ACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_ACCUMULATE);
diff --git a/src/mpid/ch4/netmod/ofi/ofi_send.h b/src/mpid/ch4/netmod/ofi/ofi_send.h
index 8cc943d..43a7e4c 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_send.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_send.h
@@ -331,10 +331,10 @@ __ALWAYS_INLINE__ int MPIDI_OFI_persistent_send(MPIDI_OFI_SENDPARAMS)
   }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_send
+#define FUNCNAME MPIDI_NM_mpi_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_send(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_send(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SEND);
@@ -374,10 +374,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_irsend(MPIDI_OFI_SENDPARAMS)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ssend
+#define FUNCNAME MPIDI_NM_mpi_ssend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_ssend(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_ssend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SSEND);
@@ -389,10 +389,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_ssend(MPIDI_OFI_SENDPARAMS)
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_isend
+#define FUNCNAME MPIDI_NM_mpi_isend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_isend(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_isend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_ISEND);
@@ -403,10 +403,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_isend(MPIDI_OFI_SENDPARAMS)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_issend
+#define FUNCNAME MPIDI_NM_mpi_issend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_issend(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_issend(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_ISSEND);
@@ -418,10 +418,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_issend(MPIDI_OFI_SENDPARAMS)
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_startall
+#define FUNCNAME MPIDI_NM_mpi_startall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
     int rc = MPI_SUCCESS, i;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_STARTALL);
@@ -432,13 +432,13 @@ __ALWAYS_INLINE__ int MPIDI_NM_startall(int count, MPIR_Request * requests[])
 
         switch (MPIDI_OFI_REQUEST(preq, util.persist.type)) {
 #ifdef MPIDI_BUILD_CH4_SHM
-            STARTALL_CASE(MPIDI_PTYPE_RECV, MPIDI_NM_irecv, preq->comm->recvcontext_id);
+            STARTALL_CASE(MPIDI_PTYPE_RECV, MPIDI_NM_mpi_irecv, preq->comm->recvcontext_id);
 #else
             STARTALL_CASE(MPIDI_PTYPE_RECV, MPIDI_Irecv, preq->comm->recvcontext_id);
 #endif
 
 #ifdef MPIDI_BUILD_CH4_SHM
-            STARTALL_CASE(MPIDI_PTYPE_SEND, MPIDI_NM_isend, preq->comm->context_id);
+            STARTALL_CASE(MPIDI_PTYPE_SEND, MPIDI_NM_mpi_isend, preq->comm->context_id);
 #else
             STARTALL_CASE(MPIDI_PTYPE_SEND, MPIDI_Isend, preq->comm->context_id);
 #endif
@@ -489,10 +489,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_startall(int count, MPIR_Request * requests[])
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_send_init
+#define FUNCNAME MPIDI_NM_mpi_send_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_send_init(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_send_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SEND_INIT);
@@ -504,10 +504,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_send_init(MPIDI_OFI_SENDPARAMS)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ssend_init
+#define FUNCNAME MPIDI_NM_mpi_ssend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_ssend_init(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_ssend_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_SSEND_INIT);
@@ -519,10 +519,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_ssend_init(MPIDI_OFI_SENDPARAMS)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_bsend_init
+#define FUNCNAME MPIDI_NM_mpi_bsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_bsend_init(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_bsend_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_BSEND_INIT);
@@ -534,10 +534,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_bsend_init(MPIDI_OFI_SENDPARAMS)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_rsend_init
+#define FUNCNAME MPIDI_NM_mpi_rsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_rsend_init(MPIDI_OFI_SENDPARAMS)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_rsend_init(MPIDI_OFI_SENDPARAMS)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_RSEND_INIT);
@@ -549,10 +549,10 @@ __ALWAYS_INLINE__ int MPIDI_NM_rsend_init(MPIDI_OFI_SENDPARAMS)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_cancel_send
+#define FUNCNAME MPIDI_NM_mpi_cancel_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-__ALWAYS_INLINE__ int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_CANCEL_SEND);
diff --git a/src/mpid/ch4/netmod/ofi/ofi_spawn.h b/src/mpid/ch4/netmod/ofi/ofi_spawn.h
index 9b109d8..a0c0272 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_spawn.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_spawn.h
@@ -380,12 +380,12 @@ static inline int MPIDI_OFI_dynproc_exchange_map(int root,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_connect
+#define FUNCNAME MPIDI_NM_mpi_comm_connect
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_connect(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
+static inline int MPIDI_NM_mpi_comm_connect(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
 {
     int entries, mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_COMM_CONNECT);
@@ -437,10 +437,10 @@ static inline int MPIDI_NM_comm_connect(const char *port_name,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_disconnect
+#define FUNCNAME MPIDI_NM_mpi_comm_disconnect
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -462,7 +462,7 @@ static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
 #define FUNCNAME MPIDI_NM_comm_open_port
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     int mpi_errno = MPI_SUCCESS;
     int str_errno = MPL_STR_SUCCESS;
@@ -488,7 +488,7 @@ static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
 #define FUNCNAME MPIDI_NM_comm_close_port
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_close_port(const char *port_name)
+static inline int MPIDI_NM_mpi_close_port(const char *port_name)
 {
     int mpi_errno = MPI_SUCCESS;
     int port_name_tag;
@@ -507,9 +507,9 @@ static inline int MPIDI_NM_close_port(const char *port_name)
 #define FUNCNAME MPIDI_NM_comm_close_port
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_accept(const char *port_name,
-                                       MPIR_Info * info,
-                                       int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
+static inline int MPIDI_NM_mpi_comm_accept(const char *port_name,
+                                           MPIR_Info * info,
+                                           int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
 {
     int entries, mpi_errno = MPI_SUCCESS;
     char *child_addr_table = NULL;
diff --git a/src/mpid/ch4/netmod/ofi/ofi_win.h b/src/mpid/ch4/netmod/ofi/ofi_win.h
index 32e7520..5f75f71 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_win.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_win.h
@@ -304,16 +304,16 @@ static inline int MPIDI_Win_progress_fence(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_set_info
+#define FUNCNAME MPIDI_NM_mpi_win_set_info
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_SET_INFO);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_WIN_SET_INFO);
 
-    mpi_errno = MPIDI_CH4R_win_set_info(win, info);
+    mpi_errno = MPIDI_CH4R_mpi_win_set_info(win, info);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 
@@ -325,10 +325,10 @@ static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_start
+#define FUNCNAME MPIDI_NM_mpi_win_start
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -357,10 +357,10 @@ static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win *
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_complete
+#define FUNCNAME MPIDI_NM_mpi_win_complete
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_complete(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -400,10 +400,10 @@ static inline int MPIDI_NM_win_complete(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_post
+#define FUNCNAME MPIDI_NM_mpi_win_post
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     int peer, index, mpi_errno = MPI_SUCCESS;
     MPIDI_OFI_win_control_t msg;
@@ -440,10 +440,10 @@ static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * w
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_wait
+#define FUNCNAME MPIDI_NM_mpi_win_wait
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_wait(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -472,10 +472,10 @@ static inline int MPIDI_NM_win_wait(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_test
+#define FUNCNAME MPIDI_NM_mpi_win_test
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -507,10 +507,10 @@ static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_lock
+#define FUNCNAME MPIDI_NM_mpi_win_lock
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -546,10 +546,10 @@ static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Wi
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_unlock
+#define FUNCNAME MPIDI_NM_mpi_win_unlock
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_UNLOCK);
@@ -583,16 +583,16 @@ static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_get_info
+#define FUNCNAME MPIDI_NM_mpi_win_get_info
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_GET_INFO);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_OFI_WIN_GET_INFO);
 
-    mpi_errno = MPIDI_CH4R_win_get_info(win, info_p_p);
+    mpi_errno = MPIDI_CH4R_mpi_win_get_info(win, info_p_p);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 
@@ -605,10 +605,10 @@ static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_free
+#define FUNCNAME MPIDI_NM_mpi_win_free
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -650,10 +650,10 @@ static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_fence
+#define FUNCNAME MPIDI_NM_mpi_win_fence
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_fence(int massert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_fence(int massert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -688,13 +688,14 @@ static inline int MPIDI_NM_win_fence(int massert, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_create
+#define FUNCNAME MPIDI_NM_mpi_win_create
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_create(void *base,
-                                      MPI_Aint length,
-                                      int disp_unit,
-                                      MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create(void *base,
+                                          MPI_Aint length,
+                                          int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -731,10 +732,10 @@ static inline int MPIDI_NM_win_create(void *base,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_attach
+#define FUNCNAME MPIDI_NM_mpi_win_attach
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_ATTACH);
@@ -750,14 +751,14 @@ static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_allocate_shared
+#define FUNCNAME MPIDI_NM_mpi_win_allocate_shared
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
-                                               int disp_unit,
-                                               MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr,
-                                               void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size,
+                                                   int disp_unit,
+                                                   MPIR_Info * info_ptr,
+                                                   MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr)
 {
     int i = 0, fd = -1, rc, first = 0, mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -912,10 +913,10 @@ static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_detach
+#define FUNCNAME MPIDI_NM_mpi_win_detach
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_DETACH);
@@ -930,12 +931,12 @@ static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_shared_query
+#define FUNCNAME MPIDI_NM_mpi_win_shared_query
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
-                                            int rank,
-                                            MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win,
+                                                int rank,
+                                                MPI_Aint * size, int *disp_unit, void *baseptr)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -966,13 +967,13 @@ static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_allocate
+#define FUNCNAME MPIDI_NM_mpi_win_allocate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_allocate(MPI_Aint size,
-                                        int disp_unit,
-                                        MPIR_Info * info,
-                                        MPIR_Comm * comm, void *baseptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate(MPI_Aint size,
+                                            int disp_unit,
+                                            MPIR_Info * info,
+                                            MPIR_Comm * comm, void *baseptr, MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -1014,10 +1015,10 @@ static inline int MPIDI_NM_win_allocate(MPI_Aint size,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_flush
+#define FUNCNAME MPIDI_NM_mpi_win_flush
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_FLUSH);
@@ -1035,10 +1036,10 @@ static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_flush_local_all
+#define FUNCNAME MPIDI_NM_mpi_win_flush_local_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_FLUSH_LOCAL_ALL);
@@ -1055,10 +1056,10 @@ static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_unlock_all
+#define FUNCNAME MPIDI_NM_mpi_win_unlock_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_UNLOCK_ALL);
@@ -1102,11 +1103,11 @@ static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_create_dynamic
+#define FUNCNAME MPIDI_NM_mpi_win_create_dynamic
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info,
-                                              MPIR_Comm * comm, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info,
+                                                  MPIR_Comm * comm, MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     int rc = MPI_SUCCESS;
@@ -1139,10 +1140,10 @@ static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_flush_local
+#define FUNCNAME MPIDI_NM_mpi_win_flush_local
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_FLUSH_LOCAL);
@@ -1161,10 +1162,10 @@ static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_sync
+#define FUNCNAME MPIDI_NM_mpi_win_sync
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_sync(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_SYNC);
@@ -1182,10 +1183,10 @@ static inline int MPIDI_NM_win_sync(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_flush_all
+#define FUNCNAME MPIDI_NM_mpi_win_flush_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_OFI_WIN_FLUSH_ALL);
@@ -1202,10 +1203,10 @@ static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_win_lock_all
+#define FUNCNAME MPIDI_NM_mpi_win_lock_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIDI_CH4U_win_lock_info *lockQ;
diff --git a/src/mpid/ch4/netmod/portals4/func_table.c b/src/mpid/ch4/netmod/portals4/func_table.c
index 59cfe2b..e707315 100644
--- a/src/mpid/ch4/netmod/portals4/func_table.c
+++ b/src/mpid/ch4/netmod/portals4/func_table.c
@@ -14,21 +14,21 @@
 #include "netmod_direct.h"
 
 MPIDI_NM_funcs_t MPIDI_NM_portals4_funcs = {
-    MPIDI_NM_init,
-    MPIDI_NM_finalize,
+    MPIDI_NM_mpi_init,
+    MPIDI_NM_mpi_finalize,
     MPIDI_NM_progress,
-    MPIDI_NM_comm_connect,
-    MPIDI_NM_comm_disconnect,
-    MPIDI_NM_open_port,
-    MPIDI_NM_close_port,
-    MPIDI_NM_comm_accept,
+    MPIDI_NM_mpi_comm_connect,
+    MPIDI_NM_mpi_comm_disconnect,
+    MPIDI_NM_mpi_open_port,
+    MPIDI_NM_mpi_close_port,
+    MPIDI_NM_mpi_comm_accept,
     MPIDI_NM_comm_get_lpid,
     MPIDI_NM_gpid_get,
     MPIDI_NM_getallincomm,
     MPIDI_NM_gpid_tolpidarray,
     MPIDI_NM_create_intercomm_from_lpids,
-    MPIDI_NM_comm_create_hook,
-    MPIDI_NM_comm_free_hook,
+    MPIDI_NM_mpi_comm_create_hook,
+    MPIDI_NM_mpi_comm_free_hook,
     MPIDI_NM_am_request_init,
     MPIDI_NM_am_request_finalize,
     MPIDI_NM_am_reg_handler,
@@ -42,108 +42,108 @@ MPIDI_NM_funcs_t MPIDI_NM_portals4_funcs = {
 };
 
 MPIDI_NM_native_funcs_t MPIDI_NM_native_portals4_funcs = {
-    MPIDI_NM_send,
-    MPIDI_NM_ssend,
-    MPIDI_NM_startall,
-    MPIDI_NM_send_init,
-    MPIDI_NM_ssend_init,
-    MPIDI_NM_rsend_init,
-    MPIDI_NM_bsend_init,
-    MPIDI_NM_isend,
-    MPIDI_NM_issend,
-    MPIDI_NM_cancel_send,
-    MPIDI_NM_recv_init,
-    MPIDI_NM_recv,
-    MPIDI_NM_irecv,
-    MPIDI_NM_imrecv,
-    MPIDI_NM_cancel_recv,
-    MPIDI_NM_alloc_mem,
-    MPIDI_NM_free_mem,
-    MPIDI_NM_improbe,
-    MPIDI_NM_iprobe,
-    MPIDI_NM_win_set_info,
-    MPIDI_NM_win_shared_query,
-    MPIDI_NM_put,
-    MPIDI_NM_win_start,
-    MPIDI_NM_win_complete,
-    MPIDI_NM_win_post,
-    MPIDI_NM_win_wait,
-    MPIDI_NM_win_test,
-    MPIDI_NM_win_lock,
-    MPIDI_NM_win_unlock,
-    MPIDI_NM_win_get_info,
-    MPIDI_NM_get,
-    MPIDI_NM_win_free,
-    MPIDI_NM_win_fence,
-    MPIDI_NM_win_create,
-    MPIDI_NM_accumulate,
-    MPIDI_NM_win_attach,
-    MPIDI_NM_win_allocate_shared,
-    MPIDI_NM_rput,
-    MPIDI_NM_win_flush_local,
-    MPIDI_NM_win_detach,
-    MPIDI_NM_compare_and_swap,
-    MPIDI_NM_raccumulate,
-    MPIDI_NM_rget_accumulate,
-    MPIDI_NM_fetch_and_op,
-    MPIDI_NM_win_allocate,
-    MPIDI_NM_win_flush,
-    MPIDI_NM_win_flush_local_all,
-    MPIDI_NM_win_unlock_all,
-    MPIDI_NM_win_create_dynamic,
-    MPIDI_NM_rget,
-    MPIDI_NM_win_sync,
-    MPIDI_NM_win_flush_all,
-    MPIDI_NM_get_accumulate,
-    MPIDI_NM_win_lock_all,
+    MPIDI_NM_mpi_send,
+    MPIDI_NM_mpi_ssend,
+    MPIDI_NM_mpi_startall,
+    MPIDI_NM_mpi_send_init,
+    MPIDI_NM_mpi_ssend_init,
+    MPIDI_NM_mpi_rsend_init,
+    MPIDI_NM_mpi_bsend_init,
+    MPIDI_NM_mpi_isend,
+    MPIDI_NM_mpi_issend,
+    MPIDI_NM_mpi_cancel_send,
+    MPIDI_NM_mpi_recv_init,
+    MPIDI_NM_mpi_recv,
+    MPIDI_NM_mpi_irecv,
+    MPIDI_NM_mpi_imrecv,
+    MPIDI_NM_mpi_cancel_recv,
+    MPIDI_NM_mpi_alloc_mem,
+    MPIDI_NM_mpi_free_mem,
+    MPIDI_NM_mpi_improbe,
+    MPIDI_NM_mpi_iprobe,
+    MPIDI_NM_mpi_win_set_info,
+    MPIDI_NM_mpi_win_shared_query,
+    MPIDI_NM_mpi_put,
+    MPIDI_NM_mpi_win_start,
+    MPIDI_NM_mpi_win_complete,
+    MPIDI_NM_mpi_win_post,
+    MPIDI_NM_mpi_win_wait,
+    MPIDI_NM_mpi_win_test,
+    MPIDI_NM_mpi_win_lock,
+    MPIDI_NM_mpi_win_unlock,
+    MPIDI_NM_mpi_win_get_info,
+    MPIDI_NM_mpi_get,
+    MPIDI_NM_mpi_win_free,
+    MPIDI_NM_mpi_win_fence,
+    MPIDI_NM_mpi_win_create,
+    MPIDI_NM_mpi_accumulate,
+    MPIDI_NM_mpi_win_attach,
+    MPIDI_NM_mpi_win_allocate_shared,
+    MPIDI_NM_mpi_rput,
+    MPIDI_NM_mpi_win_flush_local,
+    MPIDI_NM_mpi_win_detach,
+    MPIDI_NM_mpi_compare_and_swap,
+    MPIDI_NM_mpi_raccumulate,
+    MPIDI_NM_mpi_rget_accumulate,
+    MPIDI_NM_mpi_fetch_and_op,
+    MPIDI_NM_mpi_win_allocate,
+    MPIDI_NM_mpi_win_flush,
+    MPIDI_NM_mpi_win_flush_local_all,
+    MPIDI_NM_mpi_win_unlock_all,
+    MPIDI_NM_mpi_win_create_dynamic,
+    MPIDI_NM_mpi_rget,
+    MPIDI_NM_mpi_win_sync,
+    MPIDI_NM_mpi_win_flush_all,
+    MPIDI_NM_mpi_get_accumulate,
+    MPIDI_NM_mpi_win_lock_all,
     MPIDI_NM_rank_is_local,
-    MPIDI_NM_barrier,
-    MPIDI_NM_bcast,
-    MPIDI_NM_allreduce,
-    MPIDI_NM_allgather,
-    MPIDI_NM_allgatherv,
-    MPIDI_NM_scatter,
-    MPIDI_NM_scatterv,
-    MPIDI_NM_gather,
-    MPIDI_NM_gatherv,
-    MPIDI_NM_alltoall,
-    MPIDI_NM_alltoallv,
-    MPIDI_NM_alltoallw,
-    MPIDI_NM_reduce,
-    MPIDI_NM_reduce_scatter,
-    MPIDI_NM_reduce_scatter_block,
-    MPIDI_NM_scan,
-    MPIDI_NM_exscan,
-    MPIDI_NM_neighbor_allgather,
-    MPIDI_NM_neighbor_allgatherv,
-    MPIDI_NM_neighbor_alltoall,
-    MPIDI_NM_neighbor_alltoallv,
-    MPIDI_NM_neighbor_alltoallw,
-    MPIDI_NM_ineighbor_allgather,
-    MPIDI_NM_ineighbor_allgatherv,
-    MPIDI_NM_ineighbor_alltoall,
-    MPIDI_NM_ineighbor_alltoallv,
-    MPIDI_NM_ineighbor_alltoallw,
-    MPIDI_NM_ibarrier,
-    MPIDI_NM_ibcast,
-    MPIDI_NM_iallgather,
-    MPIDI_NM_iallgatherv,
-    MPIDI_NM_iallreduce,
-    MPIDI_NM_ialltoall,
-    MPIDI_NM_ialltoallv,
-    MPIDI_NM_ialltoallw,
-    MPIDI_NM_iexscan,
-    MPIDI_NM_igather,
-    MPIDI_NM_igatherv,
-    MPIDI_NM_ireduce_scatter_block,
-    MPIDI_NM_ireduce_scatter,
-    MPIDI_NM_ireduce,
-    MPIDI_NM_iscan,
-    MPIDI_NM_iscatter,
-    MPIDI_NM_iscatterv,
-    MPIDI_NM_type_create_hook,
-    MPIDI_NM_type_free_hook,
-    MPIDI_NM_op_create_hook,
-    MPIDI_NM_op_free_hook
+    MPIDI_NM_mpi_barrier,
+    MPIDI_NM_mpi_bcast,
+    MPIDI_NM_mpi_allreduce,
+    MPIDI_NM_mpi_allgather,
+    MPIDI_NM_mpi_allgatherv,
+    MPIDI_NM_mpi_scatter,
+    MPIDI_NM_mpi_scatterv,
+    MPIDI_NM_mpi_gather,
+    MPIDI_NM_mpi_gatherv,
+    MPIDI_NM_mpi_alltoall,
+    MPIDI_NM_mpi_alltoallv,
+    MPIDI_NM_mpi_alltoallw,
+    MPIDI_NM_mpi_reduce,
+    MPIDI_NM_mpi_reduce_scatter,
+    MPIDI_NM_mpi_reduce_scatter_block,
+    MPIDI_NM_mpi_scan,
+    MPIDI_NM_mpi_exscan,
+    MPIDI_NM_mpi_neighbor_allgather,
+    MPIDI_NM_mpi_neighbor_allgatherv,
+    MPIDI_NM_mpi_neighbor_alltoall,
+    MPIDI_NM_mpi_neighbor_alltoallv,
+    MPIDI_NM_mpi_neighbor_alltoallw,
+    MPIDI_NM_mpi_ineighbor_allgather,
+    MPIDI_NM_mpi_ineighbor_allgatherv,
+    MPIDI_NM_mpi_ineighbor_alltoall,
+    MPIDI_NM_mpi_ineighbor_alltoallv,
+    MPIDI_NM_mpi_ineighbor_alltoallw,
+    MPIDI_NM_mpi_ibarrier,
+    MPIDI_NM_mpi_ibcast,
+    MPIDI_NM_mpi_iallgather,
+    MPIDI_NM_mpi_iallgatherv,
+    MPIDI_NM_mpi_iallreduce,
+    MPIDI_NM_mpi_ialltoall,
+    MPIDI_NM_mpi_ialltoallv,
+    MPIDI_NM_mpi_ialltoallw,
+    MPIDI_NM_mpi_iexscan,
+    MPIDI_NM_mpi_igather,
+    MPIDI_NM_mpi_igatherv,
+    MPIDI_NM_mpi_ireduce_scatter_block,
+    MPIDI_NM_mpi_ireduce_scatter,
+    MPIDI_NM_mpi_ireduce,
+    MPIDI_NM_mpi_iscan,
+    MPIDI_NM_mpi_iscatter,
+    MPIDI_NM_mpi_iscatterv,
+    MPIDI_NM_mpi_type_create_hook,
+    MPIDI_NM_mpi_type_free_hook,
+    MPIDI_NM_mpi_op_create_hook,
+    MPIDI_NM_mpi_op_free_hook
 };
 #endif
diff --git a/src/mpid/ch4/netmod/portals4/ptl_coll.h b/src/mpid/ch4/netmod/portals4/ptl_coll.h
index 5d7073f..ae1695c 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_coll.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_coll.h
@@ -15,10 +15,10 @@
 #include "ptl_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_barrier
+#define FUNCNAME MPIDI_NM_mpi_barrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BARRIER);
@@ -31,11 +31,11 @@ static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errfla
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_bcast
+#define FUNCNAME MPIDI_NM_mpi_bcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                 int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                     int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BCAST);
@@ -48,12 +48,12 @@ static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allreduce
+#define FUNCNAME MPIDI_NM_mpi_allreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                     MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                         MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                         MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLREDUCE);
@@ -66,12 +66,12 @@ static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int cou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgather
+#define FUNCNAME MPIDI_NM_mpi_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHER);
@@ -85,13 +85,13 @@ static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts, const int *displs,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                      MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts, const int *displs,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                          MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHERV);
@@ -105,12 +105,12 @@ static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gather
+#define FUNCNAME MPIDI_NM_mpi_gather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                  void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                  int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                      int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHER);
@@ -124,13 +124,13 @@ static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Dataty
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gatherv
+#define FUNCNAME MPIDI_NM_mpi_gatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, const int *recvcounts, const int *displs,
-                                   MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, const int *recvcounts, const int *displs,
+                                       MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHERV);
@@ -144,12 +144,12 @@ static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatter
+#define FUNCNAME MPIDI_NM_mpi_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTER);
@@ -163,13 +163,13 @@ static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatterv
+#define FUNCNAME MPIDI_NM_mpi_scatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
-                                    const int *displs, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                        const int *displs, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTERV);
@@ -183,12 +183,12 @@ static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoall
+#define FUNCNAME MPIDI_NM_mpi_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALL);
@@ -202,14 +202,14 @@ static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                     const int *sdispls, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts,
-                                     const int *rdispls, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                         const int *sdispls, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts,
+                                         const int *rdispls, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLV);
@@ -222,14 +222,14 @@ static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[],
-                                     const int sdispls[], const MPI_Datatype sendtypes[],
-                                     void *recvbuf, const int recvcounts[],
-                                     const int rdispls[], const MPI_Datatype recvtypes[],
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallw(const void *sendbuf, const int sendcounts[],
+                                         const int sdispls[], const MPI_Datatype sendtypes[],
+                                         void *recvbuf, const int recvcounts[],
+                                         const int rdispls[], const MPI_Datatype recvtypes[],
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLW);
@@ -243,12 +243,12 @@ static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[]
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce
+#define FUNCNAME MPIDI_NM_mpi_reduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, int root,
-                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, int root,
+                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE);
@@ -261,12 +261,13 @@ static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                          const int recvcounts[], MPI_Datatype datatype,
-                                          MPI_Op op, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                              const int recvcounts[], MPI_Datatype datatype,
+                                              MPI_Op op, MPIR_Comm * comm_ptr,
+                                              MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER);
@@ -279,13 +280,13 @@ static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                int recvcount, MPI_Datatype datatype,
-                                                MPI_Op op, MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                    int recvcount, MPI_Datatype datatype,
+                                                    MPI_Op op, MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER_BLOCK);
@@ -299,12 +300,12 @@ static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvb
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scan
+#define FUNCNAME MPIDI_NM_mpi_scan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
-                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCAN);
@@ -317,12 +318,12 @@ static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_exscan
+#define FUNCNAME MPIDI_NM_mpi_exscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                  MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                      MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_EXSCAN);
@@ -335,12 +336,13 @@ static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHER);
@@ -355,13 +357,13 @@ static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int recvcounts[], const int displs[],
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int recvcounts[], const int displs[],
+                                                   MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHERV);
@@ -375,12 +377,13 @@ static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                             MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                 MPI_Datatype sendtype, void *recvbuf,
+                                                 int recvcount, MPI_Datatype recvtype,
+                                                 MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALL);
@@ -394,14 +397,14 @@ static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                              const int sdispls[], MPI_Datatype sendtype,
-                                              void *recvbuf, const int recvcounts[],
-                                              const int rdispls[], MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                  const int sdispls[], MPI_Datatype sendtype,
+                                                  void *recvbuf, const int recvcounts[],
+                                                  const int rdispls[], MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLV);
@@ -415,14 +418,15 @@ static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                              const MPI_Aint sdispls[],
-                                              const MPI_Datatype sendtypes[], void *recvbuf,
-                                              const int recvcounts[], const MPI_Aint rdispls[],
-                                              const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                  const MPI_Aint sdispls[],
+                                                  const MPI_Datatype sendtypes[], void *recvbuf,
+                                                  const int recvcounts[], const MPI_Aint rdispls[],
+                                                  const MPI_Datatype recvtypes[],
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLW);
@@ -436,13 +440,13 @@ static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHER);
@@ -456,14 +460,14 @@ static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int recvcounts[], const int displs[],
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int recvcounts[], const int displs[],
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                    MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHERV);
@@ -478,13 +482,13 @@ static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALL);
@@ -498,14 +502,14 @@ static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                               const int sdispls[], MPI_Datatype sendtype,
-                                               void *recvbuf, const int recvcounts[],
-                                               const int rdispls[], MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                   const int sdispls[], MPI_Datatype sendtype,
+                                                   void *recvbuf, const int recvcounts[],
+                                                   const int rdispls[], MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLV);
@@ -520,15 +524,15 @@ static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                               const MPI_Aint sdispls[],
-                                               const MPI_Datatype sendtypes[], void *recvbuf,
-                                               const int recvcounts[], const MPI_Aint rdispls[],
-                                               const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                   const MPI_Aint sdispls[],
+                                                   const MPI_Datatype sendtypes[], void *recvbuf,
+                                                   const int recvcounts[], const MPI_Aint rdispls[],
+                                                   const MPI_Datatype recvtypes[],
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLW);
@@ -543,10 +547,10 @@ static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibarrier
+#define FUNCNAME MPIDI_NM_mpi_ibarrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBARRIER);
@@ -559,11 +563,11 @@ static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibcast
+#define FUNCNAME MPIDI_NM_mpi_ibcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                  int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                      int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBCAST);
@@ -576,12 +580,12 @@ static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgather
+#define FUNCNAME MPIDI_NM_mpi_iallgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHER);
@@ -595,12 +599,12 @@ static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallreduce
+#define FUNCNAME MPIDI_NM_mpi_iallreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE);
@@ -614,13 +618,14 @@ static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int co
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgatherv
+#define FUNCNAME MPIDI_NM_mpi_iallgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                       MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHERV);
@@ -634,12 +639,12 @@ static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoall
+#define FUNCNAME MPIDI_NM_mpi_ialltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALL);
@@ -653,14 +658,14 @@ static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallv
+#define FUNCNAME MPIDI_NM_mpi_ialltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLV);
@@ -675,14 +680,14 @@ static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallw
+#define FUNCNAME MPIDI_NM_mpi_ialltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, const MPI_Datatype recvtypes[],
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, const MPI_Datatype recvtypes[],
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLW);
@@ -697,12 +702,12 @@ static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iexscan
+#define FUNCNAME MPIDI_NM_mpi_iexscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                   MPI_Request * req)
+static inline int MPIDI_NM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                       MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IEXSCAN);
@@ -715,12 +720,12 @@ static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igather
+#define FUNCNAME MPIDI_NM_mpi_igather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHER);
@@ -734,13 +739,13 @@ static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igatherv
+#define FUNCNAME MPIDI_NM_mpi_igatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPI_Request * req)
+static inline int MPIDI_NM_mpi_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                        MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHERV);
@@ -754,12 +759,13 @@ static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                 int recvcount, MPI_Datatype datatype,
-                                                 MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER_BLOCK);
@@ -773,12 +779,12 @@ static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recv
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                           const int recvcounts[], MPI_Datatype datatype,
-                                           MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                               const int recvcounts[], MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER);
@@ -792,12 +798,12 @@ static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce
+#define FUNCNAME MPIDI_NM_mpi_ireduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root,
-                                   MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE);
@@ -810,12 +816,12 @@ static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscan
+#define FUNCNAME MPIDI_NM_mpi_iscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                 MPI_Request * req)
+static inline int MPIDI_NM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCAN);
@@ -828,13 +834,13 @@ static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatter
+#define FUNCNAME MPIDI_NM_mpi_iscatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
-                                    MPI_Datatype sendtype, void *recvbuf,
-                                    int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                        MPI_Datatype sendtype, void *recvbuf,
+                                        int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTER);
@@ -848,14 +854,14 @@ static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatterv
+#define FUNCNAME MPIDI_NM_mpi_iscatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                     const int *displs, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount,
-                                     MPI_Datatype recvtype, int root,
-                                     MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount,
+                                         MPI_Datatype recvtype, int root,
+                                         MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTERV);
diff --git a/src/mpid/ch4/netmod/portals4/ptl_comm.h b/src/mpid/ch4/netmod/portals4/ptl_comm.h
index 4edbcd9..c15e2f7 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_comm.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_comm.h
@@ -14,10 +14,10 @@
 #include "ptl_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_create_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     mpi_errno = MPIDI_CH4U_init_comm(comm);
@@ -25,10 +25,10 @@ static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_free_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_free_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     mpi_errno = MPIDI_CH4U_destroy_comm(comm);
diff --git a/src/mpid/ch4/netmod/portals4/ptl_datatype.h b/src/mpid/ch4/netmod/portals4/ptl_datatype.h
index 8f1cc2e..3d2c4d2 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_datatype.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_datatype.h
@@ -9,12 +9,12 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_type_free_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_free_hook(MPIR_Datatype * datatype_p)
 {
     return 0;
 }
 
-static inline int MPIDI_NM_type_create_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_create_hook(MPIR_Datatype * datatype_p)
 {
     return 0;
 }
diff --git a/src/mpid/ch4/netmod/portals4/ptl_init.h b/src/mpid/ch4/netmod/portals4/ptl_init.h
index 4f2d592..06b1f51 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_init.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_init.h
@@ -40,16 +40,16 @@ static inline int MPIDI_PTL_append_overflow(size_t i)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_init
+#define FUNCNAME MPIDI_NM_mpi_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_init(int rank,
-                                int size,
-                                int appnum,
-                                int *tag_ub,
-                                MPIR_Comm * comm_world,
-                                MPIR_Comm * comm_self,
-                                int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init(int rank,
+                                    int size,
+                                    int appnum,
+                                    int *tag_ub,
+                                    MPIR_Comm * comm_world,
+                                    MPIR_Comm * comm_self,
+                                    int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS;
     int ret;
@@ -163,7 +163,7 @@ static inline int MPIDI_NM_init(int rank,
     }
 
     /* Setup CH4R Active Messages */
-    MPIDI_CH4U_init(comm_world, comm_self, num_contexts, netmod_contexts);
+    MPIDI_CH4U_mpi_init(comm_world, comm_self, num_contexts, netmod_contexts);
     for (i = 0; i < MPIDI_PTL_NUM_OVERFLOW_BUFFERS; i++) {
         MPIDI_PTL_global.overflow_bufs[i] = MPL_malloc(MPIDI_PTL_OVERFLOW_BUFFER_SZ);
         MPIDI_PTL_append_overflow(i);
@@ -183,7 +183,7 @@ static inline int MPIDI_NM_init(int rank,
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_finalize(void)
+static inline int MPIDI_NM_mpi_finalize(void)
 {
     int mpi_errno = MPI_SUCCESS;
     int ret, i;
@@ -191,7 +191,7 @@ static inline int MPIDI_NM_finalize(void)
     MPIR_Comm_release(MPIR_Process.comm_world);
     MPIR_Comm_release(MPIR_Process.comm_self);
 
-    MPIDI_CH4U_finalize();
+    MPIDI_CH4U_mpi_finalize();
 
     for (i = 0; i < MPIDI_PTL_NUM_OVERFLOW_BUFFERS; i++) {
         ret = PtlMEUnlink(MPIDI_PTL_global.overflow_me_handles[i]);
@@ -245,14 +245,14 @@ static inline int MPIDI_NM_create_intercomm_from_lpids(MPIR_Comm * newcomm_ptr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_free_mem(void *ptr)
+static inline int MPIDI_NM_mpi_free_mem(void *ptr)
 {
-    return MPIDI_CH4U_free_mem(ptr);
+    return MPIDI_CH4U_mpi_free_mem(ptr);
 }
 
-static inline void *MPIDI_NM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+static inline void *MPIDI_NM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
-    return MPIDI_CH4U_alloc_mem(size, info_ptr);
+    return MPIDI_CH4U_mpi_alloc_mem(size, info_ptr);
 }
 
 
diff --git a/src/mpid/ch4/netmod/portals4/ptl_op.h b/src/mpid/ch4/netmod/portals4/ptl_op.h
index a79ad6e..7a690da 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_op.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_op.h
@@ -9,13 +9,13 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_op_free_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_free_hook(MPIR_Op * op_p)
 {
     MPIR_Assert(0);
     return 0;
 }
 
-static inline int MPIDI_NM_op_create_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_create_hook(MPIR_Op * op_p)
 {
     MPIR_Assert(0);
     return 0;
diff --git a/src/mpid/ch4/netmod/portals4/ptl_probe.h b/src/mpid/ch4/netmod/portals4/ptl_probe.h
index df4eae2..7f23bd9 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_probe.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_probe.h
@@ -19,21 +19,21 @@ static inline int MPIDI_NM_probe(int source,
     return MPIDI_CH4U_probe(source, tag, comm, context_offset, status);
 }
 
-static inline int MPIDI_NM_improbe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset,
-                                   int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_NM_mpi_improbe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset,
+                                       int *flag, MPIR_Request ** message, MPI_Status * status)
 {
-    return MPIDI_CH4U_improbe(source, tag, comm, context_offset, flag, message, status);
+    return MPIDI_CH4U_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
 }
 
-static inline int MPIDI_NM_iprobe(int source,
-                                  int tag,
-                                  MPIR_Comm * comm,
-                                  int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_NM_mpi_iprobe(int source,
+                                      int tag,
+                                      MPIR_Comm * comm,
+                                      int context_offset, int *flag, MPI_Status * status)
 {
-    return MPIDI_CH4U_iprobe(source, tag, comm, context_offset, flag, status);
+    return MPIDI_CH4U_mpi_iprobe(source, tag, comm, context_offset, flag, status);
 }
 
 #endif /* NETMOD_PTL_PROBE_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/portals4/ptl_recv.h b/src/mpid/ch4/netmod/portals4/ptl_recv.h
index a239a3a..5b29727 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_recv.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_recv.h
@@ -13,48 +13,51 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_recv(void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm,
-                                int context_offset, MPI_Status * status, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv(void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm,
+                                    int context_offset, MPI_Status * status,
+                                    MPIR_Request ** request)
 {
-    return MPIDI_CH4U_recv(buf, count, datatype, rank, tag, comm, context_offset, status, request);
+    return MPIDI_CH4U_mpi_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
+                               request);
 }
 
-static inline int MPIDI_NM_recv_init(void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv_init(void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_imrecv(void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  MPIR_Request * message, MPIR_Request ** rreqp)
+static inline int MPIDI_NM_mpi_imrecv(void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      MPIR_Request * message, MPIR_Request ** rreqp)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_irecv(void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_irecv(void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+static inline int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/portals4/ptl_rma.h b/src/mpid/ch4/netmod/portals4/ptl_rma.h
index 8be0450..e455697 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_rma.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_rma.h
@@ -13,136 +13,141 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_put(const void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_put(const void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_put(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_put(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_get(void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get(void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_get(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_rput(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rput(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rput(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rput(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                            const void *compare_addr,
-                                            void *result_addr,
-                                            MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win)
-{
-    return MPIDI_CH4U_compare_and_swap(origin_addr, compare_addr, result_addr,
-                                       datatype, target_rank, target_disp, win);
-}
-
-static inline int MPIDI_NM_raccumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype,
-                                       MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                const void *compare_addr,
+                                                void *result_addr,
+                                                MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win)
 {
-    return MPIDI_CH4U_raccumulate(origin_addr, origin_count, origin_datatype,
-                                  target_rank, target_disp, target_count,
-                                  target_datatype, op, win, request);
+    return MPIDI_CH4U_mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                           datatype, target_rank, target_disp, win);
 }
 
-static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_raccumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
                                            MPI_Datatype target_datatype,
                                            MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                      result_addr, result_count, result_datatype,
+    return MPIDI_CH4U_mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                       target_rank, target_disp, target_count,
                                       target_datatype, op, win, request);
 }
 
-static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
-                                        void *result_addr,
-                                        MPI_Datatype datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+{
+    return MPIDI_CH4U_mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                          result_addr, result_count, result_datatype,
+                                          target_rank, target_disp, target_count,
+                                          target_datatype, op, win, request);
+}
+
+static inline int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr,
+                                            void *result_addr,
+                                            MPI_Datatype datatype,
+                                            int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_fetch_and_op(origin_addr, result_addr, datatype,
-                                   target_rank, target_disp, op, win);
+    return MPIDI_CH4U_mpi_fetch_and_op(origin_addr, result_addr, datatype,
+                                       target_rank, target_disp, op, win);
 }
 
 
-static inline int MPIDI_NM_rget(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rget(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_get_accumulate(const void *origin_addr,
+                                              int origin_count,
+                                              MPI_Datatype origin_datatype,
+                                              void *result_addr,
+                                              int result_count,
+                                              MPI_Datatype result_datatype,
+                                              int target_rank,
+                                              MPI_Aint target_disp,
+                                              int target_count,
+                                              MPI_Datatype target_datatype, MPI_Op op,
+                                              MPIR_Win * win)
+{
+    return MPIDI_CH4U_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                         result_addr, result_count, result_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win);
+}
+
+static inline int MPIDI_NM_mpi_accumulate(const void *origin_addr,
                                           int origin_count,
                                           MPI_Datatype origin_datatype,
-                                          void *result_addr,
-                                          int result_count,
-                                          MPI_Datatype result_datatype,
                                           int target_rank,
                                           MPI_Aint target_disp,
                                           int target_count,
                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                     result_addr, result_count, result_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win);
-}
-
-static inline int MPIDI_NM_accumulate(const void *origin_addr,
-                                      int origin_count,
-                                      MPI_Datatype origin_datatype,
-                                      int target_rank,
-                                      MPI_Aint target_disp,
-                                      int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    return MPIDI_CH4U_accumulate(origin_addr, origin_count, origin_datatype,
-                                 target_rank, target_disp, target_count, target_datatype, op, win);
+    return MPIDI_CH4U_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                     target_rank, target_disp, target_count, target_datatype, op,
+                                     win);
 }
 
 #endif /* NETMOD_PTL_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/portals4/ptl_send.h b/src/mpid/ch4/netmod/portals4/ptl_send.h
index d910222..0419e44 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_send.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_send.h
@@ -13,14 +13,14 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_send(const void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send(const void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
 static inline int MPIDI_NM_rsend(const void *buf,
@@ -45,84 +45,91 @@ static inline int MPIDI_NM_irsend(const void *buf,
     return MPIDI_CH4U_irsend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_CH4U_startall(count, requests);
+    return MPIDI_CH4U_mpi_startall(count, requests);
 }
 
-static inline int MPIDI_NM_send_init(const void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send_init(const void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_bsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_bsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_rsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_isend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_isend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_issend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_issend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
-    return MPIDI_CH4U_cancel_send(sreq);
+    return MPIDI_CH4U_mpi_cancel_send(sreq);
 }
 
 #endif /* NETMOD_PTL_SEND_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/portals4/ptl_spawn.h b/src/mpid/ch4/netmod/portals4/ptl_spawn.h
index b0ef76a..905d9ec 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_spawn.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_spawn.h
@@ -13,35 +13,35 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_comm_connect(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_NM_mpi_comm_connect(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_close_port(const char *port_name)
+static inline int MPIDI_NM_mpi_close_port(const char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_comm_accept(const char *port_name,
-                                       MPIR_Info * info,
-                                       int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_NM_mpi_comm_accept(const char *port_name,
+                                           MPIR_Info * info,
+                                           int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/portals4/ptl_win.h b/src/mpid/ch4/netmod/portals4/ptl_win.h
index c325315..06d7d96 100644
--- a/src/mpid/ch4/netmod/portals4/ptl_win.h
+++ b/src/mpid/ch4/netmod/portals4/ptl_win.h
@@ -13,147 +13,150 @@
 
 #include "ptl_impl.h"
 
-static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_CH4R_win_set_info(win, info);
+    return MPIDI_CH4R_mpi_win_set_info(win, info);
 }
 
 
-static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_start(group, assert, win);
+    return MPIDI_CH4R_mpi_win_start(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_complete(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_complete(win);
+    return MPIDI_CH4R_mpi_win_complete(win);
 }
 
-static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_post(group, assert, win);
+    return MPIDI_CH4R_mpi_win_post(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_wait(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_wait(win);
+    return MPIDI_CH4R_mpi_win_wait(win);
 }
 
 
-static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_CH4R_win_test(win, flag);
+    return MPIDI_CH4R_mpi_win_test(win, flag);
 }
 
-static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock(lock_type, rank, assert, win);
+    return MPIDI_CH4R_mpi_win_lock(lock_type, rank, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock(rank, win);
+    return MPIDI_CH4R_mpi_win_unlock(rank, win);
 }
 
-static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_CH4R_win_get_info(win, info_p_p);
+    return MPIDI_CH4R_mpi_win_get_info(win, info_p_p);
 }
 
 
-static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_free(win_ptr);
+    return MPIDI_CH4R_mpi_win_free(win_ptr);
 }
 
-static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_fence(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_fence(assert, win);
+    return MPIDI_CH4R_mpi_win_fence(assert, win);
 }
 
-static inline int MPIDI_NM_win_create(void *base,
-                                      MPI_Aint length,
-                                      int disp_unit,
-                                      MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create(void *base,
+                                          MPI_Aint length,
+                                          int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
 }
 
-static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_CH4R_win_attach(win, base, size);
+    return MPIDI_CH4R_mpi_win_attach(win, base, size);
 }
 
-static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
-                                               int disp_unit,
-                                               MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr,
-                                               void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size,
+                                                   int disp_unit,
+                                                   MPIR_Info * info_ptr,
+                                                   MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
+                                              win_ptr);
 }
 
-static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_CH4R_win_detach(win, base);
+    return MPIDI_CH4R_mpi_win_detach(win, base);
 }
 
-static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
-                                            int rank,
-                                            MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win,
+                                                int rank,
+                                                MPI_Aint * size, int *disp_unit, void *baseptr)
 {
-    return MPIDI_CH4R_win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_CH4R_mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 }
 
-static inline int MPIDI_NM_win_allocate(MPI_Aint size,
-                                        int disp_unit,
-                                        MPIR_Info * info,
-                                        MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_allocate(MPI_Aint size,
+                                            int disp_unit,
+                                            MPIR_Info * info,
+                                            MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_allocate(size, disp_unit, info, comm, baseptr, win);
+    return MPIDI_CH4R_mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
 }
 
-static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush(rank, win);
+    return MPIDI_CH4R_mpi_win_flush(rank, win);
 }
 
-static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local_all(win);
+    return MPIDI_CH4R_mpi_win_flush_local_all(win);
 }
 
-static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock_all(win);
+    return MPIDI_CH4R_mpi_win_unlock_all(win);
 }
 
-static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                  MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_create_dynamic(info, comm, win);
+    return MPIDI_CH4R_mpi_win_create_dynamic(info, comm, win);
 }
 
-static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local(rank, win);
+    return MPIDI_CH4R_mpi_win_flush_local(rank, win);
 }
 
-static inline int MPIDI_NM_win_sync(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_sync(win);
+    return MPIDI_CH4R_mpi_win_sync(win);
 }
 
-static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_all(win);
+    return MPIDI_CH4R_mpi_win_flush_all(win);
 }
 
-static inline int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock_all(assert, win);
+    return MPIDI_CH4R_mpi_win_lock_all(assert, win);
 }
 
 
diff --git a/src/mpid/ch4/netmod/stubnm/globals.c b/src/mpid/ch4/netmod/stubnm/globals.c
index 190d561..f87b80b 100644
--- a/src/mpid/ch4/netmod/stubnm/globals.c
+++ b/src/mpid/ch4/netmod/stubnm/globals.c
@@ -14,21 +14,21 @@
 #include <mpidimpl.h>
 #include "netmod_direct.h"
 MPIDI_NM_funcs_t MPIDI_NM_stubnm_funcs = {
-    MPIDI_NM_init,
-    MPIDI_NM_finalize,
+    MPIDI_NM_mpi_init,
+    MPIDI_NM_mpi_finalize,
     MPIDI_NM_progress,
-    MPIDI_NM_comm_connect,
-    MPIDI_NM_comm_disconnect,
-    MPIDI_NM_open_port,
-    MPIDI_NM_close_port,
-    MPIDI_NM_comm_accept,
+    MPIDI_NM_mpi_comm_connect,
+    MPIDI_NM_mpi_comm_disconnect,
+    MPIDI_NM_mpi_open_port,
+    MPIDI_NM_mpi_close_port,
+    MPIDI_NM_mpi_comm_accept,
     MPIDI_NM_comm_get_lpid,
     MPIDI_NM_gpid_get,
     MPIDI_NM_getallincomm,
     MPIDI_NM_gpid_tolpidarray,
     MPIDI_NM_create_intercomm_from_lpids,
-    MPIDI_NM_comm_create_hook,
-    MPIDI_NM_comm_free_hook,
+    MPIDI_NM_mpi_comm_create_hook,
+    MPIDI_NM_mpi_comm_free_hook,
     MPIDI_NM_am_request_init,
     MPIDI_NM_am_request_finalize,
     MPIDI_NM_am_reg_handler,
@@ -42,108 +42,109 @@ MPIDI_NM_funcs_t MPIDI_NM_stubnm_funcs = {
 };
 
 MPIDI_NM_native_funcs_t MPIDI_NM_native_stubnm_funcs = {
-    MPIDI_NM_send,
-    MPIDI_NM_ssend,
-    MPIDI_NM_startall,
-    MPIDI_NM_send_init,
-    MPIDI_NM_ssend_init,
-    MPIDI_NM_rsend_init,
-    MPIDI_NM_bsend_init,
-    MPIDI_NM_isend,
-    MPIDI_NM_issend,
-    MPIDI_NM_cancel_send,
-    MPIDI_NM_recv_init,
-    MPIDI_NM_recv,
-    MPIDI_NM_irecv,
-    MPIDI_NM_imrecv,
-    MPIDI_NM_cancel_recv,
-    MPIDI_NM_alloc_mem,
-    MPIDI_NM_free_mem,
-    MPIDI_NM_improbe,
-    MPIDI_NM_iprobe,
-    MPIDI_NM_win_set_info,
-    MPIDI_NM_win_shared_query,
-    MPIDI_NM_put,
-    MPIDI_NM_win_start,
-    MPIDI_NM_win_complete,
-    MPIDI_NM_win_post,
-    MPIDI_NM_win_wait,
-    MPIDI_NM_win_test,
-    MPIDI_NM_win_lock,
-    MPIDI_NM_win_unlock,
-    MPIDI_NM_win_get_info,
-    MPIDI_NM_get,
-    MPIDI_NM_win_free,
-    MPIDI_NM_win_fence,
-    MPIDI_NM_win_create,
-    MPIDI_NM_accumulate,
-    MPIDI_NM_win_attach,
-    MPIDI_NM_win_allocate_shared,
-    MPIDI_NM_rput,
-    MPIDI_NM_win_flush_local,
-    MPIDI_NM_win_detach,
-    MPIDI_NM_compare_and_swap,
-    MPIDI_NM_raccumulate,
-    MPIDI_NM_rget_accumulate,
-    MPIDI_NM_fetch_and_op,
-    MPIDI_NM_win_allocate,
-    MPIDI_NM_win_flush,
-    MPIDI_NM_win_flush_local_all,
-    MPIDI_NM_win_unlock_all,
-    MPIDI_NM_win_create_dynamic,
-    MPIDI_NM_rget,
-    MPIDI_NM_win_sync,
-    MPIDI_NM_win_flush_all,
-    MPIDI_NM_get_accumulate,
-    MPIDI_NM_win_lock_all,
+    MPIDI_NM_mpi_send,
+    MPIDI_NM_mpi_ssend,
+    MPIDI_NM_mpi_startall,
+    MPIDI_NM_mpi_send_init,
+    MPIDI_NM_mpi_ssend_init,
+    MPIDI_NM_mpi_rsend_init,
+    MPIDI_NM_mpi_bsend_init,
+    MPIDI_NM_mpi_isend,
+    MPIDI_NM_mpi_issend,
+    MPIDI_NM_mpi_cancel_send,
+    MPIDI_NM_mpi_recv_init,
+    MPIDI_NM_mpi_recv,
+    MPIDI_NM_mpi_irecv,
+    MPIDI_NM_mpi_imrecv,
+    MPIDI_NM_mpi_cancel_recv,
+    MPIDI_NM_mpi_alloc_mem,
+    MPIDI_NM_mpi_free_mem,
+    MPIDI_NM_mpi_improbe,
+    MPIDI_NM_mpi_iprobe,
+    MPIDI_NM_mpi_win_set_info,
+    MPIDI_NM_mpi_win_shared_query,
+    MPIDI_NM_mpi_put,
+    MPIDI_NM_mpi_win_start,
+    MPIDI_NM_mpi_win_complete,
+    MPIDI_NM_mpi_win_post,
+    MPIDI_NM_mpi_win_wait,
+    MPIDI_NM_mpi_win_test,
+    MPIDI_NM_mpi_win_lock,
+    MPIDI_NM_mpi_win_unlock,
+    MPIDI_NM_mpi_win_get_info,
+    MPIDI_NM_mpi_get,
+    MPIDI_NM_mpi_win_free,
+    MPIDI_NM_mpi_win_fence,
+    MPIDI_NM_mpi_win_create,
+    MPIDI_NM_mpi_accumulate,
+    MPIDI_NM_mpi_win_attach,
+    MPIDI_NM_mpi_win_allocate_shared,
+    MPIDI_NM_mpi_rput,
+    MPIDI_NM_mpi_win_flush_local,
+    MPIDI_NM_mpi_win_detach,
+    MPIDI_NM_mpi_compare_and_swap,
+    MPIDI_NM_mpi_raccumulate,
+    MPIDI_NM_mpi_rget_accumulate,
+    MPIDI_NM_mpi_fetch_and_op,
+    MPIDI_NM_mpi_win_allocate,
+    MPIDI_NM_mpi_win_flush,
+    MPIDI_NM_mpi_win_flush_local_all,
+    MPIDI_NM_mpi_win_unlock_all,
+    MPIDI_NM_mpi_win_create_dynamic,
+    MPIDI_NM_mpi_rget,
+    MPIDI_NM_mpi_win_sync,
+    MPIDI_NM_mpi_win_flush_all,
+    MPIDI_NM_mpi_get_accumulate,
+    MPIDI_NM_mpi_win_lock_all,
     MPIDI_NM_rank_is_local,
-    MPIDI_NM_barrier,
-    MPIDI_NM_bcast,
-    MPIDI_NM_allreduce,
-    MPIDI_NM_allgather,
-    MPIDI_NM_allgatherv,
-    MPIDI_NM_scatter,
-    MPIDI_NM_scatterv,
-    MPIDI_NM_gather,
-    MPIDI_NM_gatherv,
-    MPIDI_NM_alltoall,
-    MPIDI_NM_alltoallv,
-    MPIDI_NM_alltoallw,
-    MPIDI_NM_reduce,
-    MPIDI_NM_reduce_scatter,
-    MPIDI_NM_reduce_scatter_block,
-    MPIDI_NM_scan,
-    MPIDI_NM_exscan,
-    MPIDI_NM_neighbor_allgather,
-    MPIDI_NM_neighbor_allgatherv,
-    MPIDI_NM_neighbor_alltoall,
-    MPIDI_NM_neighbor_alltoallv,
-    MPIDI_NM_neighbor_alltoallw,
-    MPIDI_NM_ineighbor_allgather,
-    MPIDI_NM_ineighbor_allgatherv,
-    MPIDI_NM_ineighbor_alltoall,
-    MPIDI_NM_ineighbor_alltoallv,
-    MPIDI_NM_ineighbor_alltoallw,
-    MPIDI_NM_ibarrier,
-    MPIDI_NM_ibcast,
-    MPIDI_NM_iallgather,
-    MPIDI_NM_iallgatherv,
-    MPIDI_NM_iallreduce,
-    MPIDI_NM_ialltoall,
-    MPIDI_NM_ialltoallv,
-    MPIDI_NM_ialltoallw,
-    MPIDI_NM_iexscan,
-    MPIDI_NM_igather,
-    MPIDI_NM_igatherv,
-    MPIDI_NM_ireduce_scatter_block,
-    MPIDI_NM_ireduce_scatter,
-    MPIDI_NM_ireduce,
-    MPIDI_NM_iscan,
-    MPIDI_NM_iscatter,
-    MPIDI_NM_iscatterv,
-    MPIDI_NM_type_create_hook,
-    MPIDI_NM_type_free_hook,
-    MPIDI_NM_op_create_hook,
-    MPIDI_NM_op_free_hook,
+    MPIDI_NM_mpi_barrier,
+    MPIDI_NM_mpi_bcast,
+    MPIDI_NM_mpi_allreduce,
+    MPIDI_NM_mpi_allgather,
+    MPIDI_NM_mpi_allgatherv,
+    MPIDI_NM_mpi_scatter,
+    MPIDI_NM_mpi_scatterv,
+    MPIDI_NM_mpi_gather,
+    MPIDI_NM_mpi_gatherv,
+    MPIDI_NM_mpi_alltoall,
+    MPIDI_NM_mpi_alltoallv,
+    MPIDI_NM_mpi_alltoallw,
+    MPIDI_NM_mpi_reduce,
+    MPIDI_NM_mpi_reduce_scatter,
+    MPIDI_NM_mpi_reduce_scatter_block,
+    MPIDI_NM_mpi_scan,
+    MPIDI_NM_mpi_exscan,
+    MPIDI_NM_mpi_neighbor_allgather,
+    MPIDI_NM_mpi_neighbor_allgatherv,
+    MPIDI_NM_mpi_neighbor_alltoall,
+    MPIDI_NM_mpi_neighbor_alltoallv,
+    MPIDI_NM_mpi_neighbor_alltoallw,
+    MPIDI_NM_mpi_ineighbor_allgather,
+    MPIDI_NM_mpi_ineighbor_allgatherv,
+    MPIDI_NM_mpi_ineighbor_alltoall,
+    MPIDI_NM_mpi_ineighbor_alltoallv,
+    MPIDI_NM_mpi_ineighbor_alltoallw,
+    MPIDI_NM_mpi_ibarrier,
+    MPIDI_NM_mpi_ibcast,
+    MPIDI_NM_mpi_iallgather,
+    MPIDI_NM_mpi_iallgatherv,
+    MPIDI_NM_mpi_iallreduce,
+    MPIDI_NM_mpi_ialltoall,
+    MPIDI_NM_mpi_ialltoallv,
+    MPIDI_NM_mpi_ialltoallw,
+    MPIDI_NM_mpi_iexscan,
+    MPIDI_NM_mpi_igather,
+    MPIDI_NM_mpi_igatherv,
+    MPIDI_NM_mpi_ireduce_scatter_block,
+    MPIDI_NM_mpi_ireduce_scatter,
+    MPIDI_NM_mpi_ireduce,
+    MPIDI_NM_mpi_iscan,
+    MPIDI_NM_mpi_iscatter,
+    MPIDI_NM_mpi_iscatterv,
+    MPIDI_NM_mpi_type_create_hook,
+    MPIDI_NM_mpi_type_dup_hook,
+    MPIDI_NM_mpi_type_free_hook,
+    MPIDI_NM_mpi_op_create_hook,
+    MPIDI_NM_mpi_op_free_hook,
 };
 #endif
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_coll.h b/src/mpid/ch4/netmod/stubnm/stubnm_coll.h
index cf76ab6..4ffb542 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_coll.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_coll.h
@@ -15,10 +15,10 @@
 #include "stubnm_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_barrier
+#define FUNCNAME MPIDI_NM_mpi_barrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BARRIER);
@@ -31,11 +31,11 @@ static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errfla
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_bcast
+#define FUNCNAME MPIDI_NM_mpi_bcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                 int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                     int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BCAST);
@@ -48,12 +48,12 @@ static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allreduce
+#define FUNCNAME MPIDI_NM_mpi_allreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                     MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                         MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                         MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLREDUCE);
@@ -66,12 +66,12 @@ static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int cou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgather
+#define FUNCNAME MPIDI_NM_mpi_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHER);
@@ -85,13 +85,13 @@ static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts, const int *displs,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                      MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts, const int *displs,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                          MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHERV);
@@ -105,12 +105,12 @@ static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gather
+#define FUNCNAME MPIDI_NM_mpi_gather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                  void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                  int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                      int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHER);
@@ -124,13 +124,13 @@ static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Dataty
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gatherv
+#define FUNCNAME MPIDI_NM_mpi_gatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, const int *recvcounts, const int *displs,
-                                   MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, const int *recvcounts, const int *displs,
+                                       MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHERV);
@@ -144,12 +144,12 @@ static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatter
+#define FUNCNAME MPIDI_NM_mpi_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTER);
@@ -163,13 +163,13 @@ static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatterv
+#define FUNCNAME MPIDI_NM_mpi_scatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
-                                    const int *displs, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                        const int *displs, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTERV);
@@ -183,12 +183,12 @@ static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoall
+#define FUNCNAME MPIDI_NM_mpi_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALL);
@@ -202,14 +202,14 @@ static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                     const int *sdispls, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts,
-                                     const int *rdispls, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                         const int *sdispls, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts,
+                                         const int *rdispls, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLV);
@@ -222,14 +222,14 @@ static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[],
-                                     const int sdispls[], const MPI_Datatype sendtypes[],
-                                     void *recvbuf, const int recvcounts[],
-                                     const int rdispls[], const MPI_Datatype recvtypes[],
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallw(const void *sendbuf, const int sendcounts[],
+                                         const int sdispls[], const MPI_Datatype sendtypes[],
+                                         void *recvbuf, const int recvcounts[],
+                                         const int rdispls[], const MPI_Datatype recvtypes[],
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLW);
@@ -243,12 +243,12 @@ static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[]
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce
+#define FUNCNAME MPIDI_NM_mpi_reduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, int root,
-                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, int root,
+                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE);
@@ -261,12 +261,13 @@ static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                          const int recvcounts[], MPI_Datatype datatype,
-                                          MPI_Op op, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                              const int recvcounts[], MPI_Datatype datatype,
+                                              MPI_Op op, MPIR_Comm * comm_ptr,
+                                              MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER);
@@ -279,13 +280,13 @@ static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                int recvcount, MPI_Datatype datatype,
-                                                MPI_Op op, MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                    int recvcount, MPI_Datatype datatype,
+                                                    MPI_Op op, MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER_BLOCK);
@@ -299,12 +300,12 @@ static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvb
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scan
+#define FUNCNAME MPIDI_NM_mpi_scan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
-                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCAN);
@@ -317,12 +318,12 @@ static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_exscan
+#define FUNCNAME MPIDI_NM_mpi_exscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                  MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                      MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_EXSCAN);
@@ -335,12 +336,13 @@ static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHER);
@@ -355,13 +357,13 @@ static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int recvcounts[], const int displs[],
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int recvcounts[], const int displs[],
+                                                   MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHERV);
@@ -375,12 +377,13 @@ static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                             MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                 MPI_Datatype sendtype, void *recvbuf,
+                                                 int recvcount, MPI_Datatype recvtype,
+                                                 MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALL);
@@ -394,14 +397,14 @@ static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                              const int sdispls[], MPI_Datatype sendtype,
-                                              void *recvbuf, const int recvcounts[],
-                                              const int rdispls[], MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                  const int sdispls[], MPI_Datatype sendtype,
+                                                  void *recvbuf, const int recvcounts[],
+                                                  const int rdispls[], MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLV);
@@ -415,14 +418,15 @@ static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                              const MPI_Aint sdispls[],
-                                              const MPI_Datatype sendtypes[], void *recvbuf,
-                                              const int recvcounts[], const MPI_Aint rdispls[],
-                                              const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                  const MPI_Aint sdispls[],
+                                                  const MPI_Datatype sendtypes[], void *recvbuf,
+                                                  const int recvcounts[], const MPI_Aint rdispls[],
+                                                  const MPI_Datatype recvtypes[],
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLW);
@@ -436,13 +440,13 @@ static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHER);
@@ -456,14 +460,14 @@ static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int recvcounts[], const int displs[],
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int recvcounts[], const int displs[],
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                    MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHERV);
@@ -478,13 +482,13 @@ static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALL);
@@ -498,14 +502,14 @@ static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                               const int sdispls[], MPI_Datatype sendtype,
-                                               void *recvbuf, const int recvcounts[],
-                                               const int rdispls[], MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                   const int sdispls[], MPI_Datatype sendtype,
+                                                   void *recvbuf, const int recvcounts[],
+                                                   const int rdispls[], MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLV);
@@ -520,15 +524,15 @@ static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                               const MPI_Aint sdispls[],
-                                               const MPI_Datatype sendtypes[], void *recvbuf,
-                                               const int recvcounts[], const MPI_Aint rdispls[],
-                                               const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                   const MPI_Aint sdispls[],
+                                                   const MPI_Datatype sendtypes[], void *recvbuf,
+                                                   const int recvcounts[], const MPI_Aint rdispls[],
+                                                   const MPI_Datatype recvtypes[],
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLW);
@@ -543,10 +547,10 @@ static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibarrier
+#define FUNCNAME MPIDI_NM_mpi_ibarrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBARRIER);
@@ -559,11 +563,11 @@ static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibcast
+#define FUNCNAME MPIDI_NM_mpi_ibcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                  int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                      int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBCAST);
@@ -576,12 +580,12 @@ static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgather
+#define FUNCNAME MPIDI_NM_mpi_iallgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHER);
@@ -595,12 +599,12 @@ static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallreduce
+#define FUNCNAME MPIDI_NM_mpi_iallreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE);
@@ -614,13 +618,14 @@ static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int co
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgatherv
+#define FUNCNAME MPIDI_NM_mpi_iallgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                       MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHERV);
@@ -634,12 +639,12 @@ static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoall
+#define FUNCNAME MPIDI_NM_mpi_ialltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALL);
@@ -653,14 +658,14 @@ static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallv
+#define FUNCNAME MPIDI_NM_mpi_ialltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLV);
@@ -675,14 +680,14 @@ static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallw
+#define FUNCNAME MPIDI_NM_mpi_ialltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, const MPI_Datatype recvtypes[],
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, const MPI_Datatype recvtypes[],
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLW);
@@ -697,12 +702,12 @@ static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iexscan
+#define FUNCNAME MPIDI_NM_mpi_iexscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                   MPI_Request * req)
+static inline int MPIDI_NM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                       MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IEXSCAN);
@@ -715,12 +720,12 @@ static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igather
+#define FUNCNAME MPIDI_NM_mpi_igather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHER);
@@ -734,13 +739,13 @@ static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igatherv
+#define FUNCNAME MPIDI_NM_mpi_igatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPI_Request * req)
+static inline int MPIDI_NM_mpi_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                        MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHERV);
@@ -754,12 +759,13 @@ static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                 int recvcount, MPI_Datatype datatype,
-                                                 MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER_BLOCK);
@@ -773,12 +779,12 @@ static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recv
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                           const int recvcounts[], MPI_Datatype datatype,
-                                           MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                               const int recvcounts[], MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER);
@@ -792,12 +798,12 @@ static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce
+#define FUNCNAME MPIDI_NM_mpi_ireduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root,
-                                   MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE);
@@ -810,12 +816,12 @@ static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscan
+#define FUNCNAME MPIDI_NM_mpi_iscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                 MPI_Request * req)
+static inline int MPIDI_NM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCAN);
@@ -828,13 +834,13 @@ static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatter
+#define FUNCNAME MPIDI_NM_mpi_iscatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
-                                    MPI_Datatype sendtype, void *recvbuf,
-                                    int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                        MPI_Datatype sendtype, void *recvbuf,
+                                        int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTER);
@@ -848,14 +854,14 @@ static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatterv
+#define FUNCNAME MPIDI_NM_mpi_iscatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                     const int *displs, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount,
-                                     MPI_Datatype recvtype, int root,
-                                     MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount,
+                                         MPI_Datatype recvtype, int root,
+                                         MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTERV);
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_comm.h b/src/mpid/ch4/netmod/stubnm/stubnm_comm.h
index da5ef9a..1d3877f 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_comm.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_comm.h
@@ -14,10 +14,10 @@
 #include "stubnm_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_create_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -25,10 +25,10 @@ static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_free_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_free_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_datatype.h b/src/mpid/ch4/netmod/stubnm/stubnm_datatype.h
index cac199f..8a416a0 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_datatype.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_datatype.h
@@ -13,13 +13,13 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_type_free_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_free_hook(MPIR_Datatype * datatype_p)
 {
     MPIR_Assert(0);
     return 0;
 }
 
-static inline int MPIDI_NM_type_create_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_create_hook(MPIR_Datatype * datatype_p)
 {
     MPIR_Assert(0);
     return 0;
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_init.h b/src/mpid/ch4/netmod/stubnm/stubnm_init.h
index 8b1a78d..fbe2d18 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_init.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_init.h
@@ -13,13 +13,13 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_init(int rank,
-                                int size,
-                                int appnum,
-                                int *tag_ub,
-                                MPIR_Comm * comm_world,
-                                MPIR_Comm * comm_self,
-                                int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init(int rank,
+                                    int size,
+                                    int appnum,
+                                    int *tag_ub,
+                                    MPIR_Comm * comm_world,
+                                    MPIR_Comm * comm_self,
+                                    int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -28,7 +28,7 @@ static inline int MPIDI_NM_init(int rank,
     return mpi_errno;
 }
 
-static inline int MPIDI_NM_finalize(void)
+static inline int MPIDI_NM_mpi_finalize(void)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -69,14 +69,14 @@ static inline int MPIDI_NM_create_intercomm_from_lpids(MPIR_Comm * newcomm_ptr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_free_mem(void *ptr)
+static inline int MPIDI_NM_mpi_free_mem(void *ptr)
 {
-    return MPIDI_CH4U_free_mem(ptr);
+    return MPIDI_CH4U_mpi_free_mem(ptr);
 }
 
-static inline void *MPIDI_NM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+static inline void *MPIDI_NM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
-    return MPIDI_CH4U_alloc_mem(size, info_ptr);
+    return MPIDI_CH4U_mpi_alloc_mem(size, info_ptr);
 }
 
 
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_op.h b/src/mpid/ch4/netmod/stubnm/stubnm_op.h
index 3c37b0d..de79fac 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_op.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_op.h
@@ -13,13 +13,13 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_op_free_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_free_hook(MPIR_Op * op_p)
 {
     MPIR_Assert(0);
     return;
 }
 
-static inline int MPIDI_NM_op_create_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_create_hook(MPIR_Op * op_p)
 {
     MPIR_Assert(0);
     return;
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_probe.h b/src/mpid/ch4/netmod/stubnm/stubnm_probe.h
index 6442fb6..15f8f62 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_probe.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_probe.h
@@ -19,21 +19,21 @@ static inline int MPIDI_NM_probe(int source,
     return MPIDI_CH4U_probe(source, tag, comm, context_offset, status);
 }
 
-static inline int MPIDI_NM_improbe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset,
-                                   int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_NM_mpi_improbe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset,
+                                       int *flag, MPIR_Request ** message, MPI_Status * status)
 {
-    return MPIDI_CH4U_improbe(source, tag, comm, context_offset, flag, message, status);
+    return MPIDI_CH4U_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
 }
 
-static inline int MPIDI_NM_iprobe(int source,
-                                  int tag,
-                                  MPIR_Comm * comm,
-                                  int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_NM_mpi_iprobe(int source,
+                                      int tag,
+                                      MPIR_Comm * comm,
+                                      int context_offset, int *flag, MPI_Status * status)
 {
-    return MPIDI_CH4U_iprobe(source, tag, comm, context_offset, flag, status);
+    return MPIDI_CH4U_mpi_iprobe(source, tag, comm, context_offset, flag, status);
 }
 
 #endif /* NETMOD_STUBNM_PROBE_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_recv.h b/src/mpid/ch4/netmod/stubnm/stubnm_recv.h
index 18724cc..3916157 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_recv.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_recv.h
@@ -13,52 +13,54 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_recv(void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm,
-                                int context_offset, MPI_Status * status, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv(void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm,
+                                    int context_offset, MPI_Status * status,
+                                    MPIR_Request ** request)
 {
     int err = MPI_SUCCESS;
     MPIR_Assert(0);
     return err;
 }
 
-static inline int MPIDI_NM_recv_init(void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv_init(void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_imrecv(void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  MPIR_Request * message, MPIR_Request ** rreqp)
+static inline int MPIDI_NM_mpi_imrecv(void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      MPIR_Request * message, MPIR_Request ** rreqp)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_irecv(void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_irecv(void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int err = MPI_SUCCESS;
     MPIR_Assert(0);
     return err;
 }
 
-static inline int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+static inline int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_rma.h b/src/mpid/ch4/netmod/stubnm/stubnm_rma.h
index de8172f..d78f373 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_rma.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_rma.h
@@ -13,136 +13,141 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_put(const void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_put(const void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_put(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_put(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_get(void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get(void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_get(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_rput(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rput(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rput(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rput(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                            const void *compare_addr,
-                                            void *result_addr,
-                                            MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win)
-{
-    return MPIDI_CH4U_compare_and_swap(origin_addr, compare_addr, result_addr,
-                                       datatype, target_rank, target_disp, win);
-}
-
-static inline int MPIDI_NM_raccumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype,
-                                       MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                const void *compare_addr,
+                                                void *result_addr,
+                                                MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win)
 {
-    return MPIDI_CH4U_raccumulate(origin_addr, origin_count, origin_datatype,
-                                  target_rank, target_disp, target_count,
-                                  target_datatype, op, win, request);
+    return MPIDI_CH4U_mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                           datatype, target_rank, target_disp, win);
 }
 
-static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_raccumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
                                            MPI_Datatype target_datatype,
                                            MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                      result_addr, result_count, result_datatype,
+    return MPIDI_CH4U_mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                       target_rank, target_disp, target_count,
                                       target_datatype, op, win, request);
 }
 
-static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
-                                        void *result_addr,
-                                        MPI_Datatype datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+{
+    return MPIDI_CH4U_mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                          result_addr, result_count, result_datatype,
+                                          target_rank, target_disp, target_count,
+                                          target_datatype, op, win, request);
+}
+
+static inline int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr,
+                                            void *result_addr,
+                                            MPI_Datatype datatype,
+                                            int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_fetch_and_op(origin_addr, result_addr, datatype,
-                                   target_rank, target_disp, op, win);
+    return MPIDI_CH4U_mpi_fetch_and_op(origin_addr, result_addr, datatype,
+                                       target_rank, target_disp, op, win);
 }
 
 
-static inline int MPIDI_NM_rget(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rget(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_get_accumulate(const void *origin_addr,
+                                              int origin_count,
+                                              MPI_Datatype origin_datatype,
+                                              void *result_addr,
+                                              int result_count,
+                                              MPI_Datatype result_datatype,
+                                              int target_rank,
+                                              MPI_Aint target_disp,
+                                              int target_count,
+                                              MPI_Datatype target_datatype, MPI_Op op,
+                                              MPIR_Win * win)
+{
+    return MPIDI_CH4U_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                         result_addr, result_count, result_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win);
+}
+
+static inline int MPIDI_NM_mpi_accumulate(const void *origin_addr,
                                           int origin_count,
                                           MPI_Datatype origin_datatype,
-                                          void *result_addr,
-                                          int result_count,
-                                          MPI_Datatype result_datatype,
                                           int target_rank,
                                           MPI_Aint target_disp,
                                           int target_count,
                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                     result_addr, result_count, result_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win);
-}
-
-static inline int MPIDI_NM_accumulate(const void *origin_addr,
-                                      int origin_count,
-                                      MPI_Datatype origin_datatype,
-                                      int target_rank,
-                                      MPI_Aint target_disp,
-                                      int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    return MPIDI_CH4U_accumulate(origin_addr, origin_count, origin_datatype,
-                                 target_rank, target_disp, target_count, target_datatype, op, win);
+    return MPIDI_CH4U_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                     target_rank, target_disp, target_count, target_datatype, op,
+                                     win);
 }
 
 #endif /* NETMOD_STUBNM_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_send.h b/src/mpid/ch4/netmod/stubnm/stubnm_send.h
index c83867d..b543c42 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_send.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_send.h
@@ -13,14 +13,14 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_send(const void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send(const void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
 static inline int MPIDI_NM_rsend(const void *buf,
@@ -45,84 +45,91 @@ static inline int MPIDI_NM_irsend(const void *buf,
     return MPIDI_CH4U_irsend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_CH4U_startall(count, requests);
+    return MPIDI_CH4U_mpi_startall(count, requests);
 }
 
-static inline int MPIDI_NM_send_init(const void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send_init(const void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_bsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_bsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_rsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_isend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_isend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_issend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_issend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
-    return MPIDI_CH4U_cancel_send(sreq);
+    return MPIDI_CH4U_mpi_cancel_send(sreq);
 }
 
 #endif /* NETMOD_STUBNM_SEND_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_spawn.h b/src/mpid/ch4/netmod/stubnm/stubnm_spawn.h
index 2f70a59..1c6725b 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_spawn.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_spawn.h
@@ -13,35 +13,35 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_comm_connect(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_NM_mpi_comm_connect(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_close_port(const char *port_name)
+static inline int MPIDI_NM_mpi_close_port(const char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_comm_accept(const char *port_name,
-                                       MPIR_Info * info,
-                                       int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_NM_mpi_comm_accept(const char *port_name,
+                                           MPIR_Info * info,
+                                           int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/netmod/stubnm/stubnm_win.h b/src/mpid/ch4/netmod/stubnm/stubnm_win.h
index 5c8a112..fd99d1a 100644
--- a/src/mpid/ch4/netmod/stubnm/stubnm_win.h
+++ b/src/mpid/ch4/netmod/stubnm/stubnm_win.h
@@ -13,147 +13,150 @@
 
 #include "stubnm_impl.h"
 
-static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_CH4R_win_set_info(win, info);
+    return MPIDI_CH4R_mpi_win_set_info(win, info);
 }
 
 
-static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_start(group, assert, win);
+    return MPIDI_CH4R_mpi_win_start(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_complete(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_complete(win);
+    return MPIDI_CH4R_mpi_win_complete(win);
 }
 
-static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_post(group, assert, win);
+    return MPIDI_CH4R_mpi_win_post(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_wait(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_wait(win);
+    return MPIDI_CH4R_mpi_win_wait(win);
 }
 
 
-static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_CH4R_win_test(win, flag);
+    return MPIDI_CH4R_mpi_win_test(win, flag);
 }
 
-static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock(lock_type, rank, assert, win);
+    return MPIDI_CH4R_mpi_win_lock(lock_type, rank, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock(rank, win);
+    return MPIDI_CH4R_mpi_win_unlock(rank, win);
 }
 
-static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_CH4R_win_get_info(win, info_p_p);
+    return MPIDI_CH4R_mpi_win_get_info(win, info_p_p);
 }
 
 
-static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_free(win_ptr);
+    return MPIDI_CH4R_mpi_win_free(win_ptr);
 }
 
-static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_fence(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_fence(assert, win);
+    return MPIDI_CH4R_mpi_win_fence(assert, win);
 }
 
-static inline int MPIDI_NM_win_create(void *base,
-                                      MPI_Aint length,
-                                      int disp_unit,
-                                      MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create(void *base,
+                                          MPI_Aint length,
+                                          int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
 }
 
-static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_CH4R_win_attach(win, base, size);
+    return MPIDI_CH4R_mpi_win_attach(win, base, size);
 }
 
-static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
-                                               int disp_unit,
-                                               MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr,
-                                               void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size,
+                                                   int disp_unit,
+                                                   MPIR_Info * info_ptr,
+                                                   MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
+                                              win_ptr);
 }
 
-static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_CH4R_win_detach(win, base);
+    return MPIDI_CH4R_mpi_win_detach(win, base);
 }
 
-static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
-                                            int rank,
-                                            MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win,
+                                                int rank,
+                                                MPI_Aint * size, int *disp_unit, void *baseptr)
 {
-    return MPIDI_CH4R_win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_CH4R_mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 }
 
-static inline int MPIDI_NM_win_allocate(MPI_Aint size,
-                                        int disp_unit,
-                                        MPIR_Info * info,
-                                        MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_allocate(MPI_Aint size,
+                                            int disp_unit,
+                                            MPIR_Info * info,
+                                            MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_allocate(size, disp_unit, info, comm, baseptr, win);
+    return MPIDI_CH4R_mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
 }
 
-static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush(rank, win);
+    return MPIDI_CH4R_mpi_win_flush(rank, win);
 }
 
-static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local_all(win);
+    return MPIDI_CH4R_mpi_win_flush_local_all(win);
 }
 
-static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock_all(win);
+    return MPIDI_CH4R_mpi_win_unlock_all(win);
 }
 
-static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                  MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_create_dynamic(info, comm, win);
+    return MPIDI_CH4R_mpi_win_create_dynamic(info, comm, win);
 }
 
-static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local(rank, win);
+    return MPIDI_CH4R_mpi_win_flush_local(rank, win);
 }
 
-static inline int MPIDI_NM_win_sync(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_sync(win);
+    return MPIDI_CH4R_mpi_win_sync(win);
 }
 
-static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_all(win);
+    return MPIDI_CH4R_mpi_win_flush_all(win);
 }
 
-static inline int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock_all(assert, win);
+    return MPIDI_CH4R_mpi_win_lock_all(assert, win);
 }
 
 
diff --git a/src/mpid/ch4/netmod/ucx/func_table.c b/src/mpid/ch4/netmod/ucx/func_table.c
index 01c9ec5..aa2a904 100644
--- a/src/mpid/ch4/netmod/ucx/func_table.c
+++ b/src/mpid/ch4/netmod/ucx/func_table.c
@@ -12,21 +12,21 @@
 #include <mpidimpl.h>
 #include "netmod_direct.h"
 MPIDI_NM_funcs_t MPIDI_NM_ucx_funcs = {
-    MPIDI_NM_init,
-    MPIDI_NM_finalize,
+    MPIDI_NM_mpi_init,
+    MPIDI_NM_mpi_finalize,
     MPIDI_NM_progress,
-    MPIDI_NM_comm_connect,
-    MPIDI_NM_comm_disconnect,
-    MPIDI_NM_open_port,
-    MPIDI_NM_close_port,
-    MPIDI_NM_comm_accept,
+    MPIDI_NM_mpi_comm_connect,
+    MPIDI_NM_mpi_comm_disconnect,
+    MPIDI_NM_mpi_open_port,
+    MPIDI_NM_mpi_close_port,
+    MPIDI_NM_mpi_comm_accept,
     MPIDI_NM_comm_get_lpid,
     MPIDI_NM_gpid_get,
     MPIDI_NM_getallincomm,
     MPIDI_NM_gpid_tolpidarray,
     MPIDI_NM_create_intercomm_from_lpids,
-    MPIDI_NM_comm_create_hook,
-    MPIDI_NM_comm_free_hook,
+    MPIDI_NM_mpi_comm_create_hook,
+    MPIDI_NM_mpi_comm_free_hook,
     MPIDI_NM_am_request_init,
     MPIDI_NM_am_request_finalize,
     MPIDI_NM_am_reg_handler,
@@ -40,108 +40,108 @@ MPIDI_NM_funcs_t MPIDI_NM_ucx_funcs = {
 };
 
 MPIDI_NM_native_funcs_t MPIDI_NM_native_ucx_funcs = {
-    MPIDI_NM_send,
-    MPIDI_NM_ssend,
-    MPIDI_NM_startall,
-    MPIDI_NM_send_init,
-    MPIDI_NM_ssend_init,
-    MPIDI_NM_rsend_init,
-    MPIDI_NM_bsend_init,
-    MPIDI_NM_isend,
-    MPIDI_NM_issend,
-    MPIDI_NM_cancel_send,
-    MPIDI_NM_recv_init,
-    MPIDI_NM_recv,
-    MPIDI_NM_irecv,
-    MPIDI_NM_imrecv,
-    MPIDI_NM_cancel_recv,
-    MPIDI_NM_alloc_mem,
-    MPIDI_NM_free_mem,
-    MPIDI_NM_improbe,
-    MPIDI_NM_iprobe,
-    MPIDI_NM_win_set_info,
-    MPIDI_NM_win_shared_query,
-    MPIDI_NM_put,
-    MPIDI_NM_win_start,
-    MPIDI_NM_win_complete,
-    MPIDI_NM_win_post,
-    MPIDI_NM_win_wait,
-    MPIDI_NM_win_test,
-    MPIDI_NM_win_lock,
-    MPIDI_NM_win_unlock,
-    MPIDI_NM_win_get_info,
-    MPIDI_NM_get,
-    MPIDI_NM_win_free,
-    MPIDI_NM_win_fence,
-    MPIDI_NM_win_create,
-    MPIDI_NM_accumulate,
-    MPIDI_NM_win_attach,
-    MPIDI_NM_win_allocate_shared,
-    MPIDI_NM_rput,
-    MPIDI_NM_win_flush_local,
-    MPIDI_NM_win_detach,
-    MPIDI_NM_compare_and_swap,
-    MPIDI_NM_raccumulate,
-    MPIDI_NM_rget_accumulate,
-    MPIDI_NM_fetch_and_op,
-    MPIDI_NM_win_allocate,
-    MPIDI_NM_win_flush,
-    MPIDI_NM_win_flush_local_all,
-    MPIDI_NM_win_unlock_all,
-    MPIDI_NM_win_create_dynamic,
-    MPIDI_NM_rget,
-    MPIDI_NM_win_sync,
-    MPIDI_NM_win_flush_all,
-    MPIDI_NM_get_accumulate,
-    MPIDI_NM_win_lock_all,
+    MPIDI_NM_mpi_send,
+    MPIDI_NM_mpi_ssend,
+    MPIDI_NM_mpi_startall,
+    MPIDI_NM_mpi_send_init,
+    MPIDI_NM_mpi_ssend_init,
+    MPIDI_NM_mpi_rsend_init,
+    MPIDI_NM_mpi_bsend_init,
+    MPIDI_NM_mpi_isend,
+    MPIDI_NM_mpi_issend,
+    MPIDI_NM_mpi_cancel_send,
+    MPIDI_NM_mpi_recv_init,
+    MPIDI_NM_mpi_recv,
+    MPIDI_NM_mpi_irecv,
+    MPIDI_NM_mpi_imrecv,
+    MPIDI_NM_mpi_cancel_recv,
+    MPIDI_NM_mpi_alloc_mem,
+    MPIDI_NM_mpi_free_mem,
+    MPIDI_NM_mpi_improbe,
+    MPIDI_NM_mpi_iprobe,
+    MPIDI_NM_mpi_win_set_info,
+    MPIDI_NM_mpi_win_shared_query,
+    MPIDI_NM_mpi_put,
+    MPIDI_NM_mpi_win_start,
+    MPIDI_NM_mpi_win_complete,
+    MPIDI_NM_mpi_win_post,
+    MPIDI_NM_mpi_win_wait,
+    MPIDI_NM_mpi_win_test,
+    MPIDI_NM_mpi_win_lock,
+    MPIDI_NM_mpi_win_unlock,
+    MPIDI_NM_mpi_win_get_info,
+    MPIDI_NM_mpi_get,
+    MPIDI_NM_mpi_win_free,
+    MPIDI_NM_mpi_win_fence,
+    MPIDI_NM_mpi_win_create,
+    MPIDI_NM_mpi_accumulate,
+    MPIDI_NM_mpi_win_attach,
+    MPIDI_NM_mpi_win_allocate_shared,
+    MPIDI_NM_mpi_rput,
+    MPIDI_NM_mpi_win_flush_local,
+    MPIDI_NM_mpi_win_detach,
+    MPIDI_NM_mpi_compare_and_swap,
+    MPIDI_NM_mpi_raccumulate,
+    MPIDI_NM_mpi_rget_accumulate,
+    MPIDI_NM_mpi_fetch_and_op,
+    MPIDI_NM_mpi_win_allocate,
+    MPIDI_NM_mpi_win_flush,
+    MPIDI_NM_mpi_win_flush_local_all,
+    MPIDI_NM_mpi_win_unlock_all,
+    MPIDI_NM_mpi_win_create_dynamic,
+    MPIDI_NM_mpi_rget,
+    MPIDI_NM_mpi_win_sync,
+    MPIDI_NM_mpi_win_flush_all,
+    MPIDI_NM_mpi_get_accumulate,
+    MPIDI_NM_mpi_win_lock_all,
     MPIDI_NM_rank_is_local,
-    MPIDI_NM_barrier,
-    MPIDI_NM_bcast,
-    MPIDI_NM_allreduce,
-    MPIDI_NM_allgather,
-    MPIDI_NM_allgatherv,
-    MPIDI_NM_scatter,
-    MPIDI_NM_scatterv,
-    MPIDI_NM_gather,
-    MPIDI_NM_gatherv,
-    MPIDI_NM_alltoall,
-    MPIDI_NM_alltoallv,
-    MPIDI_NM_alltoallw,
-    MPIDI_NM_reduce,
-    MPIDI_NM_reduce_scatter,
-    MPIDI_NM_reduce_scatter_block,
-    MPIDI_NM_scan,
-    MPIDI_NM_exscan,
-    MPIDI_NM_neighbor_allgather,
-    MPIDI_NM_neighbor_allgatherv,
-    MPIDI_NM_neighbor_alltoall,
-    MPIDI_NM_neighbor_alltoallv,
-    MPIDI_NM_neighbor_alltoallw,
-    MPIDI_NM_ineighbor_allgather,
-    MPIDI_NM_ineighbor_allgatherv,
-    MPIDI_NM_ineighbor_alltoall,
-    MPIDI_NM_ineighbor_alltoallv,
-    MPIDI_NM_ineighbor_alltoallw,
-    MPIDI_NM_ibarrier,
-    MPIDI_NM_ibcast,
-    MPIDI_NM_iallgather,
-    MPIDI_NM_iallgatherv,
-    MPIDI_NM_iallreduce,
-    MPIDI_NM_ialltoall,
-    MPIDI_NM_ialltoallv,
-    MPIDI_NM_ialltoallw,
-    MPIDI_NM_iexscan,
-    MPIDI_NM_igather,
-    MPIDI_NM_igatherv,
-    MPIDI_NM_ireduce_scatter_block,
-    MPIDI_NM_ireduce_scatter,
-    MPIDI_NM_ireduce,
-    MPIDI_NM_iscan,
-    MPIDI_NM_iscatter,
-    MPIDI_NM_iscatterv,
-    MPIDI_NM_type_create_hook,
-    MPIDI_NM_type_free_hook,
-    MPIDI_NM_op_create_hook,
-    MPIDI_NM_op_free_hook
+    MPIDI_NM_mpi_barrier,
+    MPIDI_NM_mpi_bcast,
+    MPIDI_NM_mpi_allreduce,
+    MPIDI_NM_mpi_allgather,
+    MPIDI_NM_mpi_allgatherv,
+    MPIDI_NM_mpi_scatter,
+    MPIDI_NM_mpi_scatterv,
+    MPIDI_NM_mpi_gather,
+    MPIDI_NM_mpi_gatherv,
+    MPIDI_NM_mpi_alltoall,
+    MPIDI_NM_mpi_alltoallv,
+    MPIDI_NM_mpi_alltoallw,
+    MPIDI_NM_mpi_reduce,
+    MPIDI_NM_mpi_reduce_scatter,
+    MPIDI_NM_mpi_reduce_scatter_block,
+    MPIDI_NM_mpi_scan,
+    MPIDI_NM_mpi_exscan,
+    MPIDI_NM_mpi_neighbor_allgather,
+    MPIDI_NM_mpi_neighbor_allgatherv,
+    MPIDI_NM_mpi_neighbor_alltoall,
+    MPIDI_NM_mpi_neighbor_alltoallv,
+    MPIDI_NM_mpi_neighbor_alltoallw,
+    MPIDI_NM_mpi_ineighbor_allgather,
+    MPIDI_NM_mpi_ineighbor_allgatherv,
+    MPIDI_NM_mpi_ineighbor_alltoall,
+    MPIDI_NM_mpi_ineighbor_alltoallv,
+    MPIDI_NM_mpi_ineighbor_alltoallw,
+    MPIDI_NM_mpi_ibarrier,
+    MPIDI_NM_mpi_ibcast,
+    MPIDI_NM_mpi_iallgather,
+    MPIDI_NM_mpi_iallgatherv,
+    MPIDI_NM_mpi_iallreduce,
+    MPIDI_NM_mpi_ialltoall,
+    MPIDI_NM_mpi_ialltoallv,
+    MPIDI_NM_mpi_ialltoallw,
+    MPIDI_NM_mpi_iexscan,
+    MPIDI_NM_mpi_igather,
+    MPIDI_NM_mpi_igatherv,
+    MPIDI_NM_mpi_ireduce_scatter_block,
+    MPIDI_NM_mpi_ireduce_scatter,
+    MPIDI_NM_mpi_ireduce,
+    MPIDI_NM_mpi_iscan,
+    MPIDI_NM_mpi_iscatter,
+    MPIDI_NM_mpi_iscatterv,
+    MPIDI_NM_mpi_type_create_hook,
+    MPIDI_NM_mpi_type_free_hook,
+    MPIDI_NM_mpi_op_create_hook,
+    MPIDI_NM_mpi_op_free_hook
 };
 #endif
diff --git a/src/mpid/ch4/netmod/ucx/ucx_am_recv.h b/src/mpid/ch4/netmod/ucx/ucx_am_recv.h
index 2f4f821..f3d3305 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_am_recv.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_am_recv.h
@@ -13,48 +13,51 @@
 
 #include "ucx_impl.h"
 
-static inline int MPIDI_NM_recv(void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm,
-                                int context_offset, MPI_Status * status, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv(void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm,
+                                    int context_offset, MPI_Status * status,
+                                    MPIR_Request ** request)
 {
-    return MPIDI_CH4U_recv(buf, count, datatype, rank, tag, comm, context_offset, status, request);
+    return MPIDI_CH4U_mpi_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
+                               request);
 }
 
-static inline int MPIDI_NM_recv_init(void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_recv_init(void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_imrecv(void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  MPIR_Request * message, MPIR_Request ** rreqp)
+static inline int MPIDI_NM_mpi_imrecv(void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      MPIR_Request * message, MPIR_Request ** rreqp)
 {
-    return MPIDI_CH4U_imrecv(buf, count, datatype, message, rreqp);
+    return MPIDI_CH4U_mpi_imrecv(buf, count, datatype, message, rreqp);
 }
 
-static inline int MPIDI_NM_irecv(void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_irecv(void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+static inline int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
-    return MPIDI_CH4U_cancel_recv(rreq);
+    return MPIDI_CH4U_mpi_cancel_recv(rreq);
 }
 
 #endif /* NETMOD_UCX_RECV_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ucx/ucx_am_rma.h b/src/mpid/ch4/netmod/ucx/ucx_am_rma.h
index 7863464..42168e8 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_am_rma.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_am_rma.h
@@ -14,136 +14,141 @@
 
 #include "ucx_impl.h"
 
-static inline int MPIDI_NM_put(const void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_put(const void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_put(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_put(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_get(void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get(void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get(origin_addr, origin_count, origin_datatype,
-                          target_rank, target_disp, target_count, target_datatype, win);
+    return MPIDI_CH4U_mpi_get(origin_addr, origin_count, origin_datatype,
+                              target_rank, target_disp, target_count, target_datatype, win);
 }
 
-static inline int MPIDI_NM_rput(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rput(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rput(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rput(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                            const void *compare_addr,
-                                            void *result_addr,
-                                            MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win)
-{
-    return MPIDI_CH4U_compare_and_swap(origin_addr, compare_addr, result_addr,
-                                       datatype, target_rank, target_disp, win);
-}
-
-static inline int MPIDI_NM_raccumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype,
-                                       MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                const void *compare_addr,
+                                                void *result_addr,
+                                                MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win)
 {
-    return MPIDI_CH4U_raccumulate(origin_addr, origin_count, origin_datatype,
-                                  target_rank, target_disp, target_count,
-                                  target_datatype, op, win, request);
+    return MPIDI_CH4U_mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                           datatype, target_rank, target_disp, win);
 }
 
-static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_raccumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
                                            MPI_Datatype target_datatype,
                                            MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                      result_addr, result_count, result_datatype,
+    return MPIDI_CH4U_mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                       target_rank, target_disp, target_count,
                                       target_datatype, op, win, request);
 }
 
-static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
-                                        void *result_addr,
-                                        MPI_Datatype datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+{
+    return MPIDI_CH4U_mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                          result_addr, result_count, result_datatype,
+                                          target_rank, target_disp, target_count,
+                                          target_datatype, op, win, request);
+}
+
+static inline int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr,
+                                            void *result_addr,
+                                            MPI_Datatype datatype,
+                                            int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_fetch_and_op(origin_addr, result_addr, datatype,
-                                   target_rank, target_disp, op, win);
+    return MPIDI_CH4U_mpi_fetch_and_op(origin_addr, result_addr, datatype,
+                                       target_rank, target_disp, op, win);
 }
 
 
-static inline int MPIDI_NM_rget(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rget(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_get_accumulate(const void *origin_addr,
+                                              int origin_count,
+                                              MPI_Datatype origin_datatype,
+                                              void *result_addr,
+                                              int result_count,
+                                              MPI_Datatype result_datatype,
+                                              int target_rank,
+                                              MPI_Aint target_disp,
+                                              int target_count,
+                                              MPI_Datatype target_datatype, MPI_Op op,
+                                              MPIR_Win * win)
+{
+    return MPIDI_CH4U_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                         result_addr, result_count, result_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win);
+}
+
+static inline int MPIDI_NM_mpi_accumulate(const void *origin_addr,
                                           int origin_count,
                                           MPI_Datatype origin_datatype,
-                                          void *result_addr,
-                                          int result_count,
-                                          MPI_Datatype result_datatype,
                                           int target_rank,
                                           MPI_Aint target_disp,
                                           int target_count,
                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                     result_addr, result_count, result_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win);
-}
-
-static inline int MPIDI_NM_accumulate(const void *origin_addr,
-                                      int origin_count,
-                                      MPI_Datatype origin_datatype,
-                                      int target_rank,
-                                      MPI_Aint target_disp,
-                                      int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    return MPIDI_CH4U_accumulate(origin_addr, origin_count, origin_datatype,
-                                 target_rank, target_disp, target_count, target_datatype, op, win);
+    return MPIDI_CH4U_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                     target_rank, target_disp, target_count, target_datatype, op,
+                                     win);
 }
 
 #endif /* NETMOD_UCX_AM_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ucx/ucx_am_send.h b/src/mpid/ch4/netmod/ucx/ucx_am_send.h
index 9fd977b..3c9ec9b 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_am_send.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_am_send.h
@@ -13,14 +13,14 @@
 
 #include "ucx_impl.h"
 
-static inline int MPIDI_NM_send(const void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send(const void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
 static inline int MPIDI_NM_rsend(const void *buf,
@@ -45,84 +45,91 @@ static inline int MPIDI_NM_irsend(const void *buf,
     return MPIDI_CH4U_irsend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_CH4U_startall(count, requests);
+    return MPIDI_CH4U_mpi_startall(count, requests);
 }
 
-static inline int MPIDI_NM_send_init(const void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send_init(const void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_ssend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_bsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_bsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_rsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
-static inline int MPIDI_NM_isend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_isend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_issend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_issend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-static inline int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
-    return MPIDI_CH4U_cancel_send(sreq);
+    return MPIDI_CH4U_mpi_cancel_send(sreq);
 }
 
 #endif /* NETMOD_UCX_SEND_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ucx/ucx_am_win.h b/src/mpid/ch4/netmod/ucx/ucx_am_win.h
index 2bf5da9..ea142a5 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_am_win.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_am_win.h
@@ -13,147 +13,150 @@
 
 #include "ucx_impl.h"
 
-static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_CH4R_win_set_info(win, info);
+    return MPIDI_CH4R_mpi_win_set_info(win, info);
 }
 
 
-static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_start(group, assert, win);
+    return MPIDI_CH4R_mpi_win_start(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_complete(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_complete(win);
+    return MPIDI_CH4R_mpi_win_complete(win);
 }
 
-static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_post(group, assert, win);
+    return MPIDI_CH4R_mpi_win_post(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_wait(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_wait(win);
+    return MPIDI_CH4R_mpi_win_wait(win);
 }
 
 
-static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_CH4R_win_test(win, flag);
+    return MPIDI_CH4R_mpi_win_test(win, flag);
 }
 
-static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock(lock_type, rank, assert, win);
+    return MPIDI_CH4R_mpi_win_lock(lock_type, rank, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock(rank, win);
+    return MPIDI_CH4R_mpi_win_unlock(rank, win);
 }
 
-static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_CH4R_win_get_info(win, info_p_p);
+    return MPIDI_CH4R_mpi_win_get_info(win, info_p_p);
 }
 
 
-static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_free(win_ptr);
+    return MPIDI_CH4R_mpi_win_free(win_ptr);
 }
 
-static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_fence(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_fence(assert, win);
+    return MPIDI_CH4R_mpi_win_fence(assert, win);
 }
 
-static inline int MPIDI_NM_win_create(void *base,
-                                      MPI_Aint length,
-                                      int disp_unit,
-                                      MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create(void *base,
+                                          MPI_Aint length,
+                                          int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
 }
 
-static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_CH4R_win_attach(win, base, size);
+    return MPIDI_CH4R_mpi_win_attach(win, base, size);
 }
 
-static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
-                                               int disp_unit,
-                                               MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr,
-                                               void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size,
+                                                   int disp_unit,
+                                                   MPIR_Info * info_ptr,
+                                                   MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
+                                              win_ptr);
 }
 
-static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_CH4R_win_detach(win, base);
+    return MPIDI_CH4R_mpi_win_detach(win, base);
 }
 
-static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
-                                            int rank,
-                                            MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win,
+                                                int rank,
+                                                MPI_Aint * size, int *disp_unit, void *baseptr)
 {
-    return MPIDI_CH4R_win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_CH4R_mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 }
 
-static inline int MPIDI_NM_win_allocate(MPI_Aint size,
-                                        int disp_unit,
-                                        MPIR_Info * info,
-                                        MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_allocate(MPI_Aint size,
+                                            int disp_unit,
+                                            MPIR_Info * info,
+                                            MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_allocate(size, disp_unit, info, comm, baseptr, win);
+    return MPIDI_CH4R_mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
 }
 
-static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush(rank, win);
+    return MPIDI_CH4R_mpi_win_flush(rank, win);
 }
 
-static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local_all(win);
+    return MPIDI_CH4R_mpi_win_flush_local_all(win);
 }
 
-static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_unlock_all(win);
+    return MPIDI_CH4R_mpi_win_unlock_all(win);
 }
 
-static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                  MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_create_dynamic(info, comm, win);
+    return MPIDI_CH4R_mpi_win_create_dynamic(info, comm, win);
 }
 
-static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_local(rank, win);
+    return MPIDI_CH4R_mpi_win_flush_local(rank, win);
 }
 
-static inline int MPIDI_NM_win_sync(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_sync(win);
+    return MPIDI_CH4R_mpi_win_sync(win);
 }
 
-static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_flush_all(win);
+    return MPIDI_CH4R_mpi_win_flush_all(win);
 }
 
-static inline int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock_all(assert, win);
+    return MPIDI_CH4R_mpi_win_lock_all(assert, win);
 }
 
 
diff --git a/src/mpid/ch4/netmod/ucx/ucx_coll.h b/src/mpid/ch4/netmod/ucx/ucx_coll.h
index 185c5b9..6d1d37c 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_coll.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_coll.h
@@ -12,10 +12,10 @@
 #include "ucx_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_barrier
+#define FUNCNAME MPIDI_NM_mpi_barrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BARRIER);
@@ -28,11 +28,11 @@ static inline int MPIDI_NM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errfla
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_bcast
+#define FUNCNAME MPIDI_NM_mpi_bcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                 int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                     int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_BCAST);
@@ -45,12 +45,12 @@ static inline int MPIDI_NM_bcast(void *buffer, int count, MPI_Datatype datatype,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allreduce
+#define FUNCNAME MPIDI_NM_mpi_allreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                     MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                         MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                         MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLREDUCE);
@@ -63,12 +63,12 @@ static inline int MPIDI_NM_allreduce(const void *sendbuf, void *recvbuf, int cou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgather
+#define FUNCNAME MPIDI_NM_mpi_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHER);
@@ -82,13 +82,13 @@ static inline int MPIDI_NM_allgather(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts, const int *displs,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                      MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts, const int *displs,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                          MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLGATHERV);
@@ -102,12 +102,12 @@ static inline int MPIDI_NM_allgatherv(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gather
+#define FUNCNAME MPIDI_NM_mpi_gather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                  void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                  int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                      int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHER);
@@ -121,13 +121,13 @@ static inline int MPIDI_NM_gather(const void *sendbuf, int sendcount, MPI_Dataty
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_gatherv
+#define FUNCNAME MPIDI_NM_mpi_gatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, const int *recvcounts, const int *displs,
-                                   MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, const int *recvcounts, const int *displs,
+                                       MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_GATHERV);
@@ -141,12 +141,12 @@ static inline int MPIDI_NM_gatherv(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatter
+#define FUNCNAME MPIDI_NM_mpi_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTER);
@@ -160,13 +160,13 @@ static inline int MPIDI_NM_scatter(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scatterv
+#define FUNCNAME MPIDI_NM_mpi_scatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
-                                    const int *displs, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                        const int *displs, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCATTERV);
@@ -180,12 +180,12 @@ static inline int MPIDI_NM_scatterv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoall
+#define FUNCNAME MPIDI_NM_mpi_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALL);
@@ -199,14 +199,14 @@ static inline int MPIDI_NM_alltoall(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                     const int *sdispls, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts,
-                                     const int *rdispls, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                         const int *sdispls, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts,
+                                         const int *rdispls, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLV);
@@ -219,14 +219,14 @@ static inline int MPIDI_NM_alltoallv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[],
-                                     const int sdispls[], const MPI_Datatype sendtypes[],
-                                     void *recvbuf, const int recvcounts[],
-                                     const int rdispls[], const MPI_Datatype recvtypes[],
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_alltoallw(const void *sendbuf, const int sendcounts[],
+                                         const int sdispls[], const MPI_Datatype sendtypes[],
+                                         void *recvbuf, const int recvcounts[],
+                                         const int rdispls[], const MPI_Datatype recvtypes[],
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ALLTOALLW);
@@ -240,12 +240,12 @@ static inline int MPIDI_NM_alltoallw(const void *sendbuf, const int sendcounts[]
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce
+#define FUNCNAME MPIDI_NM_mpi_reduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, int root,
-                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, int root,
+                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE);
@@ -258,12 +258,13 @@ static inline int MPIDI_NM_reduce(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                          const int recvcounts[], MPI_Datatype datatype,
-                                          MPI_Op op, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                              const int recvcounts[], MPI_Datatype datatype,
+                                              MPI_Op op, MPIR_Comm * comm_ptr,
+                                              MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER);
@@ -276,13 +277,13 @@ static inline int MPIDI_NM_reduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_reduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_reduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                int recvcount, MPI_Datatype datatype,
-                                                MPI_Op op, MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                    int recvcount, MPI_Datatype datatype,
+                                                    MPI_Op op, MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_REDUCE_SCATTER_BLOCK);
@@ -296,12 +297,12 @@ static inline int MPIDI_NM_reduce_scatter_block(const void *sendbuf, void *recvb
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_scan
+#define FUNCNAME MPIDI_NM_mpi_scan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
-                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_SCAN);
@@ -314,12 +315,12 @@ static inline int MPIDI_NM_scan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_exscan
+#define FUNCNAME MPIDI_NM_mpi_exscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                  MPIR_Errflag_t * errflag)
+static inline int MPIDI_NM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                      MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_EXSCAN);
@@ -332,12 +333,13 @@ static inline int MPIDI_NM_exscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHER);
@@ -352,13 +354,13 @@ static inline int MPIDI_NM_neighbor_allgather(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int recvcounts[], const int displs[],
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int recvcounts[], const int displs[],
+                                                   MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLGATHERV);
@@ -372,12 +374,13 @@ static inline int MPIDI_NM_neighbor_allgatherv(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                             MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                             MPI_Datatype recvtype, MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                 MPI_Datatype sendtype, void *recvbuf,
+                                                 int recvcount, MPI_Datatype recvtype,
+                                                 MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALL);
@@ -391,14 +394,14 @@ static inline int MPIDI_NM_neighbor_alltoall(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                              const int sdispls[], MPI_Datatype sendtype,
-                                              void *recvbuf, const int recvcounts[],
-                                              const int rdispls[], MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                  const int sdispls[], MPI_Datatype sendtype,
+                                                  void *recvbuf, const int recvcounts[],
+                                                  const int rdispls[], MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLV);
@@ -412,14 +415,15 @@ static inline int MPIDI_NM_neighbor_alltoallv(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_neighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_neighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                              const MPI_Aint sdispls[],
-                                              const MPI_Datatype sendtypes[], void *recvbuf,
-                                              const int recvcounts[], const MPI_Aint rdispls[],
-                                              const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                  const MPI_Aint sdispls[],
+                                                  const MPI_Datatype sendtypes[], void *recvbuf,
+                                                  const int recvcounts[], const MPI_Aint rdispls[],
+                                                  const MPI_Datatype recvtypes[],
+                                                  MPIR_Comm * comm_ptr)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_NEIGHBOR_ALLTOALLW);
@@ -433,13 +437,13 @@ static inline int MPIDI_NM_neighbor_alltoallw(const void *sendbuf, const int sen
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgather
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHER);
@@ -453,14 +457,14 @@ static inline int MPIDI_NM_ineighbor_allgather(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_allgatherv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int recvcounts[], const int displs[],
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int recvcounts[], const int displs[],
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                    MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLGATHERV);
@@ -475,13 +479,13 @@ static inline int MPIDI_NM_ineighbor_allgatherv(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoall
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype,
-                                              MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALL);
@@ -495,14 +499,14 @@ static inline int MPIDI_NM_ineighbor_alltoall(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallv
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                               const int sdispls[], MPI_Datatype sendtype,
-                                               void *recvbuf, const int recvcounts[],
-                                               const int rdispls[], MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                   const int sdispls[], MPI_Datatype sendtype,
+                                                   void *recvbuf, const int recvcounts[],
+                                                   const int rdispls[], MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLV);
@@ -517,15 +521,15 @@ static inline int MPIDI_NM_ineighbor_alltoallv(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ineighbor_alltoallw
+#define FUNCNAME MPIDI_NM_mpi_ineighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                               const MPI_Aint sdispls[],
-                                               const MPI_Datatype sendtypes[], void *recvbuf,
-                                               const int recvcounts[], const MPI_Aint rdispls[],
-                                               const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr,
-                                               MPI_Request * req)
+static inline int MPIDI_NM_mpi_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                   const MPI_Aint sdispls[],
+                                                   const MPI_Datatype sendtypes[], void *recvbuf,
+                                                   const int recvcounts[], const MPI_Aint rdispls[],
+                                                   const MPI_Datatype recvtypes[],
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_INEIGHBOR_ALLTOALLW);
@@ -540,10 +544,10 @@ static inline int MPIDI_NM_ineighbor_alltoallw(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibarrier
+#define FUNCNAME MPIDI_NM_mpi_ibarrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBARRIER);
@@ -556,11 +560,11 @@ static inline int MPIDI_NM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ibcast
+#define FUNCNAME MPIDI_NM_mpi_ibcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                  int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                      int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IBCAST);
@@ -573,12 +577,12 @@ static inline int MPIDI_NM_ibcast(void *buffer, int count, MPI_Datatype datatype
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgather
+#define FUNCNAME MPIDI_NM_mpi_iallgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHER);
@@ -592,13 +596,14 @@ static inline int MPIDI_NM_iallgather(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallgatherv
+#define FUNCNAME MPIDI_NM_mpi_iallgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                       MPI_Request * req)
+static inline int MPIDI_NM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLGATHERV);
@@ -612,12 +617,12 @@ static inline int MPIDI_NM_iallgatherv(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iallreduce
+#define FUNCNAME MPIDI_NM_mpi_iallreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                      MPI_Request * request)
+static inline int MPIDI_NM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                          MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLREDUCE);
@@ -630,12 +635,12 @@ static inline int MPIDI_NM_iallreduce(const void *sendbuf, void *recvbuf, int co
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoall
+#define FUNCNAME MPIDI_NM_mpi_ialltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALL);
@@ -649,14 +654,14 @@ static inline int MPIDI_NM_ialltoall(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallv
+#define FUNCNAME MPIDI_NM_mpi_ialltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLV);
@@ -671,14 +676,14 @@ static inline int MPIDI_NM_ialltoallv(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ialltoallw
+#define FUNCNAME MPIDI_NM_mpi_ialltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, const MPI_Datatype recvtypes[],
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, const MPI_Datatype recvtypes[],
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IALLTOALLW);
@@ -693,12 +698,12 @@ static inline int MPIDI_NM_ialltoallw(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iexscan
+#define FUNCNAME MPIDI_NM_mpi_iexscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                   MPI_Request * req)
+static inline int MPIDI_NM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                       MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IEXSCAN);
@@ -711,12 +716,12 @@ static inline int MPIDI_NM_iexscan(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igather
+#define FUNCNAME MPIDI_NM_mpi_igather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHER);
@@ -730,13 +735,13 @@ static inline int MPIDI_NM_igather(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_igatherv
+#define FUNCNAME MPIDI_NM_mpi_igatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPI_Request * req)
+static inline int MPIDI_NM_mpi_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                        MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IGATHERV);
@@ -750,12 +755,13 @@ static inline int MPIDI_NM_igatherv(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter_block
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                 int recvcount, MPI_Datatype datatype,
-                                                 MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER_BLOCK);
@@ -769,12 +775,12 @@ static inline int MPIDI_NM_ireduce_scatter_block(const void *sendbuf, void *recv
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce_scatter
+#define FUNCNAME MPIDI_NM_mpi_ireduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                           const int recvcounts[], MPI_Datatype datatype,
-                                           MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                               const int recvcounts[], MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE_SCATTER);
@@ -788,12 +794,12 @@ static inline int MPIDI_NM_ireduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_ireduce
+#define FUNCNAME MPIDI_NM_mpi_ireduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root,
-                                   MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_NM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_IREDUCE);
@@ -806,12 +812,12 @@ static inline int MPIDI_NM_ireduce(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscan
+#define FUNCNAME MPIDI_NM_mpi_iscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                 MPI_Request * req)
+static inline int MPIDI_NM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCAN);
@@ -824,13 +830,13 @@ static inline int MPIDI_NM_iscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatter
+#define FUNCNAME MPIDI_NM_mpi_iscatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
-                                    MPI_Datatype sendtype, void *recvbuf,
-                                    int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                        MPI_Datatype sendtype, void *recvbuf,
+                                        int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTER);
@@ -844,14 +850,14 @@ static inline int MPIDI_NM_iscatter(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_iscatterv
+#define FUNCNAME MPIDI_NM_mpi_iscatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                     const int *displs, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount,
-                                     MPI_Datatype recvtype, int root,
-                                     MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_NM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount,
+                                         MPI_Datatype recvtype, int root,
+                                         MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NM_ISCATTERV);
diff --git a/src/mpid/ch4/netmod/ucx/ucx_comm.h b/src/mpid/ch4/netmod/ucx/ucx_comm.h
index 5b4368c..85e1ce0 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_comm.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_comm.h
@@ -12,10 +12,10 @@
 #include "ucx_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_create_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_UCX_COMM_CREATE);
@@ -29,10 +29,10 @@ static inline int MPIDI_NM_comm_create_hook(MPIR_Comm * comm)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_free_hook
+#define FUNCNAME MPIDI_NM_mpi_comm_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_free_hook(MPIR_Comm * comm)
+static inline int MPIDI_NM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_UCX_COMM_DESTROY);
diff --git a/src/mpid/ch4/netmod/ucx/ucx_datatype.h b/src/mpid/ch4/netmod/ucx/ucx_datatype.h
index ce4f933..bfcd91d 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_datatype.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_datatype.h
@@ -107,7 +107,7 @@ static ucp_generic_dt_ops_t MPIDI_UCX_datatype_ops = {
 };
 
 
-static inline int MPIDI_NM_type_free_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_free_hook(MPIR_Datatype * datatype_p)
 {
 
 
@@ -119,7 +119,7 @@ static inline int MPIDI_NM_type_free_hook(MPIR_Datatype * datatype_p)
     return 0;
 }
 
-static inline int MPIDI_NM_type_create_hook(MPIR_Datatype * datatype_p)
+static inline int MPIDI_NM_mpi_type_create_hook(MPIR_Datatype * datatype_p)
 {
     ucp_datatype_t ucp_datatype;
     ucs_status_t status;
diff --git a/src/mpid/ch4/netmod/ucx/ucx_init.h b/src/mpid/ch4/netmod/ucx/ucx_init.h
index 52374b0..acbd31f 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_init.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_init.h
@@ -15,16 +15,16 @@
 #include "pmi.h"
 #include <ucp/api/ucp.h>
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_init
+#define FUNCNAME MPIDI_NM_mpi_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_init(int rank,
-                                int size,
-                                int appnum,
-                                int *tag_ub,
-                                MPIR_Comm * comm_world,
-                                MPIR_Comm * comm_self,
-                                int spawned, int num_contexts, void **netmod_contexts)
+static inline int MPIDI_NM_mpi_init(int rank,
+                                    int size,
+                                    int appnum,
+                                    int *tag_ub,
+                                    MPIR_Comm * comm_world,
+                                    MPIR_Comm * comm_self,
+                                    int spawned, int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS, thr_err, pmi_errno;
     int str_errno = MPL_STR_SUCCESS;
@@ -108,7 +108,7 @@ static inline int MPIDI_NM_init(int rank,
         memset(remote_addr, 0x0, maxlen);
     }
 
-    MPIDI_CH4U_init(comm_world, comm_self, num_contexts, netmod_contexts);
+    MPIDI_CH4U_mpi_init(comm_world, comm_self, num_contexts, netmod_contexts);
 
     mpi_errno = MPIR_Datatype_init_names();
     MPIDI_CH4_UCX_MPI_ERROR(mpi_errno);
@@ -127,7 +127,7 @@ static inline int MPIDI_NM_init(int rank,
 
 }
 
-static inline int MPIDI_NM_finalize(void)
+static inline int MPIDI_NM_mpi_finalize(void)
 {
     int mpi_errno = MPI_SUCCESS, thr_err, pmi_errno;
     int i, j, max_n_avts;
@@ -157,7 +157,7 @@ static inline int MPIDI_NM_finalize(void)
     if (MPIDI_UCX_global.pmi_addr_table)
         MPL_free(MPIDI_UCX_global.pmi_addr_table);
 
-    MPIDI_CH4U_finalize();
+    MPIDI_CH4U_mpi_finalize();
     PMI_Finalize();
 
   fn_exit:
@@ -305,14 +305,14 @@ static inline int MPIDI_NM_create_intercomm_from_lpids(MPIR_Comm * newcomm_ptr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_NM_free_mem(void *ptr)
+static inline int MPIDI_NM_mpi_free_mem(void *ptr)
 {
-    return MPIDI_CH4U_free_mem(ptr);
+    return MPIDI_CH4U_mpi_free_mem(ptr);
 }
 
-static inline void *MPIDI_NM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+static inline void *MPIDI_NM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
-    return MPIDI_CH4U_alloc_mem(size, info_ptr);
+    return MPIDI_CH4U_mpi_alloc_mem(size, info_ptr);
 }
 
 #endif /* NETMOD_UCX_INIT_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ucx/ucx_op.h b/src/mpid/ch4/netmod/ucx/ucx_op.h
index fb58231..13e9b30 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_op.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_op.h
@@ -11,12 +11,12 @@
 
 #include "ucx_impl.h"
 
-static inline int MPIDI_NM_op_free_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_free_hook(MPIR_Op * op_p)
 {
     return 0;
 }
 
-static inline int MPIDI_NM_op_create_hook(MPIR_Op * op_p)
+static inline int MPIDI_NM_mpi_op_create_hook(MPIR_Op * op_p)
 {
     return 0;
 }
diff --git a/src/mpid/ch4/netmod/ucx/ucx_probe.h b/src/mpid/ch4/netmod/ucx/ucx_probe.h
index 8806146..b1d9e72 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_probe.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_probe.h
@@ -43,11 +43,11 @@ static inline int ucx_do_iprobe(int source,
 
 }
 
-static inline int MPIDI_NM_improbe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset,
-                                   int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_NM_mpi_improbe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset,
+                                       int *flag, MPIR_Request ** message, MPI_Status * status)
 {
 
 
@@ -85,10 +85,10 @@ static inline int MPIDI_NM_improbe(int source,
 }
 
 
-static inline int MPIDI_NM_iprobe(int source,
-                                  int tag,
-                                  MPIR_Comm * comm,
-                                  int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_NM_mpi_iprobe(int source,
+                                      int tag,
+                                      MPIR_Comm * comm,
+                                      int context_offset, int *flag, MPI_Status * status)
 {
     int mpi_errno = MPI_SUCCESS;
 
diff --git a/src/mpid/ch4/netmod/ucx/ucx_recv.h b/src/mpid/ch4/netmod/ucx/ucx_recv.h
index 11eb7cf..e166096 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_recv.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_recv.h
@@ -134,34 +134,34 @@ static inline int do_irecv(void *buf,
 
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_recv(void *buf,
-                                    int count,
-                                    MPI_Datatype datatype,
-                                    int rank,
-                                    int tag,
-                                    MPIR_Comm * comm,
-                                    int context_offset,
-                                    MPI_Status * status, MPIR_Request ** request)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv(void *buf,
+                                        int count,
+                                        MPI_Datatype datatype,
+                                        int rank,
+                                        int tag,
+                                        MPIR_Comm * comm,
+                                        int context_offset,
+                                        MPI_Status * status, MPIR_Request ** request)
 {
 
     return do_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_recv_init(void *buf,
-                                         int count,
-                                         MPI_Datatype datatype,
-                                         int rank,
-                                         int tag,
-                                         MPIR_Comm * comm,
-                                         int context_offset, MPIR_Request ** request)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_recv_init(void *buf,
+                                             int count,
+                                             MPI_Datatype datatype,
+                                             int rank,
+                                             int tag,
+                                             MPIR_Comm * comm,
+                                             int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_imrecv(void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      MPIR_Request * message, MPIR_Request ** rreqp)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_imrecv(void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          MPIR_Request * message, MPIR_Request ** rreqp)
 {
     ucp_tag_message_h message_handler;
     int mpi_errno = MPI_SUCCESS;
@@ -218,12 +218,13 @@ __ALWAYS_INLINE__ int MPIDI_NM_imrecv(void *buf,
     goto fn_exit;
 }
 
-__ALWAYS_INLINE__ int MPIDI_NM_irecv(void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+__ALWAYS_INLINE__ int MPIDI_NM_mpi_irecv(void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
 
 
@@ -232,7 +233,7 @@ __ALWAYS_INLINE__ int MPIDI_NM_irecv(void *buf,
 
 }
 
-static inline int MPIDI_NM_cancel_recv(MPIR_Request * rreq)
+static inline int MPIDI_NM_mpi_cancel_recv(MPIR_Request * rreq)
 {
 
     if (MPIDI_UCX_REQ(rreq).a.ucp_request) {
diff --git a/src/mpid/ch4/netmod/ucx/ucx_rma.h b/src/mpid/ch4/netmod/ucx/ucx_rma.h
index 88aa9af..3c0046c 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_rma.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_rma.h
@@ -78,12 +78,12 @@ static inline int MPIDI_UCX_contig_get(void *origin_addr,
 
 }
 
-static inline int MPIDI_NM_put(const void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_put(const void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_NETMOD_UCX_PUT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_UCX_PUT);
@@ -92,8 +92,8 @@ static inline int MPIDI_NM_put(const void *origin_addr,
     MPI_Aint origin_true_lb, target_true_lb;
     size_t offset;
     if (win->create_flavor == MPI_WIN_FLAVOR_DYNAMIC || win->create_flavor == MPI_WIN_FLAVOR_SHARED)
-        return MPIDI_CH4U_put(origin_addr, origin_count, origin_datatype,
-                              target_rank, target_disp, target_count, target_datatype, win);
+        return MPIDI_CH4U_mpi_put(origin_addr, origin_count, origin_datatype,
+                                  target_rank, target_disp, target_count, target_datatype, win);
 
 
 
@@ -109,8 +109,8 @@ static inline int MPIDI_NM_put(const void *origin_addr,
     if (unlikely((origin_bytes == 0) || (target_rank == MPI_PROC_NULL)))
         goto fn_exit;
     if (!target_contig || !origin_contig || MPIDI_UCX_WIN_INFO(win, target_rank).rkey == NULL)
-        return MPIDI_CH4U_put(origin_addr, origin_count, origin_datatype,
-                              target_rank, target_disp, target_count, target_datatype, win);
+        return MPIDI_CH4U_mpi_put(origin_addr, origin_count, origin_datatype,
+                                  target_rank, target_disp, target_count, target_datatype, win);
 
     MPIDI_CH4U_EPOCH_CHECK_SYNC(win, mpi_errno, goto fn_fail);
 
@@ -132,12 +132,12 @@ static inline int MPIDI_NM_put(const void *origin_addr,
 
 }
 
-static inline int MPIDI_NM_get(void *origin_addr,
-                               int origin_count,
-                               MPI_Datatype origin_datatype,
-                               int target_rank,
-                               MPI_Aint target_disp,
-                               int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_get(void *origin_addr,
+                                   int origin_count,
+                                   MPI_Datatype origin_datatype,
+                                   int target_rank,
+                                   MPI_Aint target_disp,
+                                   int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
 
 
@@ -150,8 +150,8 @@ static inline int MPIDI_NM_get(void *origin_addr,
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_NETMOD_UCX_GET);
 
     if (win->create_flavor == MPI_WIN_FLAVOR_DYNAMIC || win->create_flavor == MPI_WIN_FLAVOR_SHARED)
-        return MPIDI_CH4U_get(origin_addr, origin_count, origin_datatype,
-                              target_rank, target_disp, target_count, target_datatype, win);
+        return MPIDI_CH4U_mpi_get(origin_addr, origin_count, origin_datatype,
+                                  target_rank, target_disp, target_count, target_datatype, win);
 
     MPIDI_Datatype_check_contig_size_lb(target_datatype, target_count,
                                         target_contig, target_bytes, target_true_lb);
@@ -164,8 +164,8 @@ static inline int MPIDI_NM_get(void *origin_addr,
 
 
     if (!origin_contig || !target_contig || MPIDI_UCX_WIN_INFO(win, target_rank).rkey == NULL)
-        return MPIDI_CH4U_get(origin_addr, origin_count, origin_datatype,
-                              target_rank, target_disp, target_count, target_datatype, win);
+        return MPIDI_CH4U_mpi_get(origin_addr, origin_count, origin_datatype,
+                                  target_rank, target_disp, target_count, target_datatype, win);
 
     MPIDI_CH4U_EPOCH_CHECK_SYNC(win, mpi_errno, goto fn_fail);
 
@@ -186,114 +186,119 @@ static inline int MPIDI_NM_get(void *origin_addr,
 
 }
 
-static inline int MPIDI_NM_rput(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rput(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rput(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rput(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_compare_and_swap(const void *origin_addr,
-                                            const void *compare_addr,
-                                            void *result_addr,
-                                            MPI_Datatype datatype,
-                                            int target_rank, MPI_Aint target_disp, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_compare_and_swap(const void *origin_addr,
+                                                const void *compare_addr,
+                                                void *result_addr,
+                                                MPI_Datatype datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                MPIR_Win * win)
 {
-    return MPIDI_CH4U_compare_and_swap(origin_addr, compare_addr, result_addr,
-                                       datatype, target_rank, target_disp, win);
+    return MPIDI_CH4U_mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                           datatype, target_rank, target_disp, win);
 }
 
-static inline int MPIDI_NM_raccumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype,
-                                       MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
-{
-    return MPIDI_CH4U_raccumulate(origin_addr, origin_count, origin_datatype,
-                                  target_rank, target_disp, target_count,
-                                  target_datatype, op, win, request);
-}
-
-static inline int MPIDI_NM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_raccumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
                                            MPI_Datatype target_datatype,
                                            MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                      result_addr, result_count, result_datatype,
+    return MPIDI_CH4U_mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                       target_rank, target_disp, target_count,
                                       target_datatype, op, win, request);
 }
 
-static inline int MPIDI_NM_fetch_and_op(const void *origin_addr,
-                                        void *result_addr,
-                                        MPI_Datatype datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_rget_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_fetch_and_op(origin_addr, result_addr, datatype,
-                                   target_rank, target_disp, op, win);
+    return MPIDI_CH4U_mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                          result_addr, result_count, result_datatype,
+                                          target_rank, target_disp, target_count,
+                                          target_datatype, op, win, request);
+}
+
+static inline int MPIDI_NM_mpi_fetch_and_op(const void *origin_addr,
+                                            void *result_addr,
+                                            MPI_Datatype datatype,
+                                            int target_rank,
+                                            MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+{
+    return MPIDI_CH4U_mpi_fetch_and_op(origin_addr, result_addr, datatype,
+                                       target_rank, target_disp, op, win);
 }
 
 
-static inline int MPIDI_NM_rget(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count,
-                                MPI_Datatype target_datatype,
-                                MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rget(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count,
+                                    MPI_Datatype target_datatype,
+                                    MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rget(origin_addr, origin_count, origin_datatype,
-                           target_rank, target_disp, target_count, target_datatype, win, request);
+    return MPIDI_CH4U_mpi_rget(origin_addr, origin_count, origin_datatype,
+                               target_rank, target_disp, target_count, target_datatype, win,
+                               request);
 }
 
 
-static inline int MPIDI_NM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_NM_mpi_get_accumulate(const void *origin_addr,
+                                              int origin_count,
+                                              MPI_Datatype origin_datatype,
+                                              void *result_addr,
+                                              int result_count,
+                                              MPI_Datatype result_datatype,
+                                              int target_rank,
+                                              MPI_Aint target_disp,
+                                              int target_count,
+                                              MPI_Datatype target_datatype, MPI_Op op,
+                                              MPIR_Win * win)
+{
+    return MPIDI_CH4U_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                         result_addr, result_count, result_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win);
+}
+
+static inline int MPIDI_NM_mpi_accumulate(const void *origin_addr,
                                           int origin_count,
                                           MPI_Datatype origin_datatype,
-                                          void *result_addr,
-                                          int result_count,
-                                          MPI_Datatype result_datatype,
                                           int target_rank,
                                           MPI_Aint target_disp,
                                           int target_count,
                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_CH4U_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                     result_addr, result_count, result_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win);
-}
-
-static inline int MPIDI_NM_accumulate(const void *origin_addr,
-                                      int origin_count,
-                                      MPI_Datatype origin_datatype,
-                                      int target_rank,
-                                      MPI_Aint target_disp,
-                                      int target_count,
-                                      MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    return MPIDI_CH4U_accumulate(origin_addr, origin_count, origin_datatype,
-                                 target_rank, target_disp, target_count, target_datatype, op, win);
+    return MPIDI_CH4U_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                     target_rank, target_disp, target_count, target_datatype, op,
+                                     win);
 }
 
 #endif /* NETMOD_UCX_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ucx/ucx_send.h b/src/mpid/ch4/netmod/ucx/ucx_send.h
index 6b2f41c..0217ca4 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_send.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_send.h
@@ -320,12 +320,12 @@ static inline int ucx_sync_send(const void *buf,
 #define FUNCNAME MPIDI_netmode_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_send(const void *buf,
-                                int count,
-                                MPI_Datatype datatype,
-                                int rank,
-                                int tag,
-                                MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send(const void *buf,
+                                    int count,
+                                    MPI_Datatype datatype,
+                                    int rank,
+                                    int tag,
+                                    MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
 
     int mpi_errno;
@@ -370,87 +370,94 @@ static inline int MPIDI_netmod_irsend(const void *buf,
 #define FUNCNAME MPIDI_netmode_ssend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ssend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     return ucx_sync_send(buf, count, datatype, rank, tag, comm, context_offset, request, 0);
 }
 
-static inline int MPIDI_NM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_NM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_CH4U_startall(count, requests);
+    return MPIDI_CH4U_mpi_startall(count, requests);
 }
 
 #undef FUNCNAME
 #define FUNCNAME MPIDI_netmode_send_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_send_init(const void *buf,
-                                     int count,
-                                     MPI_Datatype datatype,
-                                     int rank,
-                                     int tag,
-                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_send_init(const void *buf,
+                                         int count,
+                                         MPI_Datatype datatype,
+                                         int rank,
+                                         int tag,
+                                         MPIR_Comm * comm, int context_offset,
+                                         MPIR_Request ** request)
 {
-    return MPIDI_CH4U_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 }
 
 #undef FUNCNAME
 #define FUNCNAME MPIDI_netmode_ssend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_ssend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_ssend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
 #undef FUNCNAME
 #define FUNCNAME MPIDI_netmode_bsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_bsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_bsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
 #undef FUNCNAME
 #define FUNCNAME MPIDI_netmode_rsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_rsend_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_rsend_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
-    return MPIDI_CH4U_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    return MPIDI_CH4U_mpi_rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                     request);
 }
 
 #undef FUNCNAME
 #define FUNCNAME MPIDI_netmode_isend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_isend(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_isend(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
 
     return ucx_send(buf, count, datatype, rank, tag, comm, context_offset, request, 1);
@@ -461,12 +468,12 @@ static inline int MPIDI_NM_isend(const void *buf,
 #define FUNCNAME MPIDI_netmode_issend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_issend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_NM_mpi_issend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
 
     return ucx_sync_send(buf, count, datatype, rank, tag, comm, context_offset, request, 1);
@@ -477,7 +484,7 @@ static inline int MPIDI_NM_issend(const void *buf,
 #define FUNCNAME MPIDI_netmode_cancel_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_NM_mpi_cancel_send(MPIR_Request * sreq)
 {
     if (MPIDI_UCX_REQ(sreq).a.ucp_request) {
         ucp_request_cancel(MPIDI_UCX_global.worker, MPIDI_UCX_REQ(sreq).a.ucp_request);
diff --git a/src/mpid/ch4/netmod/ucx/ucx_spawn.h b/src/mpid/ch4/netmod/ucx/ucx_spawn.h
index 037b348..cdc8bae 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_spawn.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_spawn.h
@@ -15,7 +15,7 @@
 #define FUNCNAME MPIDI_netmod_comm_open_port
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_NM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     int mpi_errno = MPI_SUCCESS;
   fn_exit:
@@ -29,7 +29,7 @@ static inline int MPIDI_NM_open_port(MPIR_Info * info_ptr, char *port_name)
 #define FUNCNAME MPIDI_netmod_comm_close_port
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_close_port(const char *port_name)
+static inline int MPIDI_NM_mpi_close_port(const char *port_name)
 {
     int mpi_errno = MPI_SUCCESS;
   fn_exit:
@@ -40,12 +40,12 @@ static inline int MPIDI_NM_close_port(const char *port_name)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_connect
+#define FUNCNAME MPIDI_NM_mpi_comm_connect
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_connect(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
+static inline int MPIDI_NM_mpi_comm_connect(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -59,10 +59,10 @@ static inline int MPIDI_NM_comm_connect(const char *port_name,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_NM_comm_disconnect
+#define FUNCNAME MPIDI_NM_mpi_comm_disconnect
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_NM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -83,9 +83,9 @@ static inline int MPIDI_NM_comm_disconnect(MPIR_Comm * comm_ptr)
 #define FUNCNAME MPIDI_netmod_comm_close_port
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_NM_comm_accept(const char *port_name,
-                                       MPIR_Info * info,
-                                       int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
+static inline int MPIDI_NM_mpi_comm_accept(const char *port_name,
+                                           MPIR_Info * info,
+                                           int root, MPIR_Comm * comm_ptr, MPIR_Comm ** newcomm)
 {
     int mpi_errno = MPI_SUCCESS;
 
diff --git a/src/mpid/ch4/netmod/ucx/ucx_win.h b/src/mpid/ch4/netmod/ucx/ucx_win.h
index 8ae3fad..16ccbdd 100644
--- a/src/mpid/ch4/netmod/ucx/ucx_win.h
+++ b/src/mpid/ch4/netmod/ucx/ucx_win.h
@@ -159,51 +159,51 @@ static inline int MPIDI_UCX_Win_init(MPI_Aint length,
 
 }
 
-static inline int MPIDI_NM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_NM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_CH4R_win_set_info(win, info);
+    return MPIDI_CH4R_mpi_win_set_info(win, info);
 }
 
 
-static inline int MPIDI_NM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_start(group, assert, win);
+    return MPIDI_CH4R_mpi_win_start(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_complete(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_complete(MPIR_Win * win)
 {
 
     ucs_status_t ucp_status;
     ucp_status = ucp_worker_flush(MPIDI_UCX_global.worker);
-    return MPIDI_CH4R_win_complete(win);
+    return MPIDI_CH4R_mpi_win_complete(win);
 }
 
-static inline int MPIDI_NM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
 
-    return MPIDI_CH4R_win_post(group, assert, win);
+    return MPIDI_CH4R_mpi_win_post(group, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_wait(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_wait(win);
+    return MPIDI_CH4R_mpi_win_wait(win);
 }
 
 
-static inline int MPIDI_NM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_NM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_CH4R_win_test(win, flag);
+    return MPIDI_CH4R_mpi_win_test(win, flag);
 }
 
-static inline int MPIDI_NM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock(lock_type, rank, assert, win);
+    return MPIDI_CH4R_mpi_win_lock(lock_type, rank, assert, win);
 }
 
 
-static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
 
     int mpi_errno = MPI_SUCCESS;
@@ -212,20 +212,20 @@ static inline int MPIDI_NM_win_unlock(int rank, MPIR_Win * win)
     /* make sure all operations are completed  */
     ucp_status = ucp_ep_flush(ep);
     MPIDI_UCX_CHK_STATUS(ucp_status, ucp_worker_fence);
-    mpi_errno = MPIDI_CH4R_win_unlock(rank, win);
+    mpi_errno = MPIDI_CH4R_mpi_win_unlock(rank, win);
   fn_exit:
     return mpi_errno;
   fn_fail:
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_NM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_CH4R_win_get_info(win, info_p_p);
+    return MPIDI_CH4R_mpi_win_get_info(win, info_p_p);
 }
 
 
-static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_free(MPIR_Win ** win_ptr)
 {
 
     int mpi_errno = MPI_SUCCESS;
@@ -250,7 +250,7 @@ static inline int MPIDI_NM_win_free(MPIR_Win ** win_ptr)
 
 }
 
-static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_fence(int assert, MPIR_Win * win)
 {
     int mpi_errno;
     ucs_status_t ucp_status;
@@ -259,7 +259,7 @@ static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
     ucp_status = ucp_worker_flush(MPIDI_UCX_global.worker);
 
 
-    mpi_errno = MPIDI_CH4R_win_fence(assert, win);
+    mpi_errno = MPIDI_CH4R_mpi_win_fence(assert, win);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 
@@ -270,10 +270,11 @@ static inline int MPIDI_NM_win_fence(int assert, MPIR_Win * win)
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_win_create(void *base,
-                                      MPI_Aint length,
-                                      int disp_unit,
-                                      MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_create(void *base,
+                                          MPI_Aint length,
+                                          int disp_unit,
+                                          MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                          MPIR_Win ** win_ptr)
 {
 
     int mpi_errno = MPI_SUCCESS;
@@ -310,36 +311,38 @@ static inline int MPIDI_NM_win_create(void *base,
 
 }
 
-static inline int MPIDI_NM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_NM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_CH4R_win_attach(win, base, size);
+    return MPIDI_CH4R_mpi_win_attach(win, base, size);
 }
 
-static inline int MPIDI_NM_win_allocate_shared(MPI_Aint size,
-                                               int disp_unit,
-                                               MPIR_Info * info_ptr,
-                                               MPIR_Comm * comm_ptr,
-                                               void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate_shared(MPI_Aint size,
+                                                   int disp_unit,
+                                                   MPIR_Info * info_ptr,
+                                                   MPIR_Comm * comm_ptr,
+                                                   void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_CH4R_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr, win_ptr);
+    return MPIDI_CH4R_mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
+                                              win_ptr);
 }
 
-static inline int MPIDI_NM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_NM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_CH4R_win_detach(win, base);
+    return MPIDI_CH4R_mpi_win_detach(win, base);
 }
 
-static inline int MPIDI_NM_win_shared_query(MPIR_Win * win,
-                                            int rank,
-                                            MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_NM_mpi_win_shared_query(MPIR_Win * win,
+                                                int rank,
+                                                MPI_Aint * size, int *disp_unit, void *baseptr)
 {
-    return MPIDI_CH4R_win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_CH4R_mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 }
 
-static inline int MPIDI_NM_win_allocate(MPI_Aint length,
-                                        int disp_unit,
-                                        MPIR_Info * info,
-                                        MPIR_Comm * comm_ptr, void *baseptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_NM_mpi_win_allocate(MPI_Aint length,
+                                            int disp_unit,
+                                            MPIR_Info * info,
+                                            MPIR_Comm * comm_ptr, void *baseptr,
+                                            MPIR_Win ** win_ptr)
 {
 
     int mpi_errno = MPI_SUCCESS;
@@ -377,7 +380,7 @@ static inline int MPIDI_NM_win_allocate(MPI_Aint length,
 
 }
 
-static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush(int rank, MPIR_Win * win)
 {
 
     int mpi_errno;
@@ -385,7 +388,7 @@ static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
 
     ucp_ep_h ep = MPIDI_UCX_COMM_TO_EP(win->comm_ptr, rank);
 
-    mpi_errno = MPIDI_CH4R_win_flush(rank, win);
+    mpi_errno = MPIDI_CH4R_mpi_win_flush(rank, win);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 /* only flush the endpoint */
@@ -400,11 +403,11 @@ static inline int MPIDI_NM_win_flush(int rank, MPIR_Win * win)
 
 }
 
-static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     ucs_status_t ucp_status;
-    mpi_errno = MPIDI_CH4R_win_flush_local_all(win);
+    mpi_errno = MPIDI_CH4R_mpi_win_flush_local_all(win);
 
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
@@ -423,7 +426,7 @@ static inline int MPIDI_NM_win_flush_local_all(MPIR_Win * win)
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_unlock_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     ucs_status_t ucp_status;
@@ -431,23 +434,24 @@ static inline int MPIDI_NM_win_unlock_all(MPIR_Win * win)
     /*first we have to make sure that all operations are completed */
     ucp_status = ucp_worker_flush(MPIDI_UCX_global.worker);
     MPIDI_UCX_CHK_STATUS(ucp_status, ucp_worker_fence);
-    mpi_errno = MPIDI_CH4R_win_unlock_all(win);
+    mpi_errno = MPIDI_CH4R_mpi_win_unlock_all(win);
   fn_exit:
     return mpi_errno;
   fn_fail:
     goto fn_exit;
 }
 
-static inline int MPIDI_NM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_NM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                  MPIR_Win ** win)
 {
-    return MPIDI_CH4R_win_create_dynamic(info, comm, win);
+    return MPIDI_CH4R_mpi_win_create_dynamic(info, comm, win);
 }
 
-static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     ucs_status_t ucp_status;
-    mpi_errno = MPIDI_CH4R_win_flush_local(rank, win);
+    mpi_errno = MPIDI_CH4R_mpi_win_flush_local(rank, win);
 
     ucp_ep_h ep = MPIDI_UCX_COMM_TO_EP(win->comm_ptr, rank);
     if (mpi_errno)
@@ -468,18 +472,18 @@ static inline int MPIDI_NM_win_flush_local(int rank, MPIR_Win * win)
 
 }
 
-static inline int MPIDI_NM_win_sync(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_sync(win);
+    return MPIDI_CH4R_mpi_win_sync(win);
 }
 
-static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_flush_all(MPIR_Win * win)
 {
 
 /*maybe we just flush all eps here? More efficient for smaller communicators...*/
     int mpi_errno = MPI_SUCCESS;
     ucs_status_t ucp_status;
-    mpi_errno = MPIDI_CH4R_win_flush_all(win);
+    mpi_errno = MPIDI_CH4R_mpi_win_flush_all(win);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 
@@ -494,9 +498,9 @@ static inline int MPIDI_NM_win_flush_all(MPIR_Win * win)
 
 }
 
-static inline int MPIDI_NM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_NM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
-    return MPIDI_CH4R_win_lock_all(assert, win);
+    return MPIDI_CH4R_mpi_win_lock_all(assert, win);
 }
 
 
diff --git a/src/mpid/ch4/shm/include/shm.h b/src/mpid/ch4/shm/include/shm.h
index 7b90ecb..597922e 100644
--- a/src/mpid/ch4/shm/include/shm.h
+++ b/src/mpid/ch4/shm/include/shm.h
@@ -27,19 +27,19 @@ typedef int (*MPIDI_SHM_am_target_handler_fn)
   MPIDI_SHM_am_completion_handler_fn * cmpl_handler_fn, /* completion handler */
   MPIR_Request ** req);         /* if allocated, need pointer to completion function */
 
-typedef int (*MPIDI_SHM_init_t) (int rank, int size);
-typedef int (*MPIDI_SHM_finalize_t) (void);
+typedef int (*MPIDI_SHM_mpi_init_t) (int rank, int size);
+typedef int (*MPIDI_SHM_mpi_finalize_t) (void);
 typedef int (*MPIDI_SHM_progress_t) (int blocking);
 typedef int (*MPIDI_SHM_am_reg_handler_t) (int handler_id,
                                            MPIDI_SHM_am_origin_handler_fn origin_handler_fn,
                                            MPIDI_SHM_am_target_handler_fn target_handler_fn);
-typedef int (*MPIDI_SHM_comm_connect_t) (const char *port_name, MPIR_Info * info, int root,
-                                         MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
-typedef int (*MPIDI_SHM_comm_disconnect_t) (MPIR_Comm * comm_ptr);
-typedef int (*MPIDI_SHM_open_port_t) (MPIR_Info * info_ptr, char *port_name);
-typedef int (*MPIDI_SHM_close_port_t) (const char *port_name);
-typedef int (*MPIDI_SHM_comm_accept_t) (const char *port_name, MPIR_Info * info, int root,
-                                        MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
+typedef int (*MPIDI_SHM_mpi_comm_connect_t) (const char *port_name, MPIR_Info * info, int root,
+                                             MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
+typedef int (*MPIDI_SHM_mpi_comm_disconnect_t) (MPIR_Comm * comm_ptr);
+typedef int (*MPIDI_SHM_mpi_open_port_t) (MPIR_Info * info_ptr, char *port_name);
+typedef int (*MPIDI_SHM_mpi_close_port_t) (const char *port_name);
+typedef int (*MPIDI_SHM_mpi_comm_accept_t) (const char *port_name, MPIR_Info * info, int root,
+                                            MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr);
 typedef int (*MPIDI_SHM_am_send_hdr_t) (int rank, MPIR_Comm * comm, int handler_id,
                                         const void *am_hdr, size_t am_hdr_sz, void *shm_context);
 typedef int (*MPIDI_SHM_am_isend_t) (int rank, MPIR_Comm * comm, int handler_id, const void *am_hdr,
@@ -82,295 +82,311 @@ typedef int (*MPIDI_SHM_getallincomm_t) (MPIR_Comm * comm_ptr, int local_size,
 typedef int (*MPIDI_SHM_gpid_tolpidarray_t) (int size, MPIR_Gpid gpid[], int lpid[]);
 typedef int (*MPIDI_SHM_create_intercomm_from_lpids_t) (MPIR_Comm * newcomm_ptr, int size,
                                                         const int lpids[]);
-typedef int (*MPIDI_SHM_comm_create_hook_t) (MPIR_Comm * comm);
-typedef int (*MPIDI_SHM_comm_free_hook_t) (MPIR_Comm * comm);
-typedef int (*MPIDI_SHM_type_create_hook_t) (MPIR_Datatype * type);
-typedef int (*MPIDI_SHM_type_free_hook_t) (MPIR_Datatype * type);
-typedef int (*MPIDI_SHM_op_create_hook_t) (MPIR_Op * op);
-typedef int (*MPIDI_SHM_op_free_hook_t) (MPIR_Op * op);
+typedef int (*MPIDI_SHM_mpi_comm_create_hook_t) (MPIR_Comm * comm);
+typedef int (*MPIDI_SHM_mpi_comm_free_hook_t) (MPIR_Comm * comm);
+typedef int (*MPIDI_SHM_mpi_type_create_hook_t) (MPIR_Datatype * type);
+typedef int (*MPIDI_SHM_mpi_type_free_hook_t) (MPIR_Datatype * type);
+typedef int (*MPIDI_SHM_mpi_op_create_hook_t) (MPIR_Op * op);
+typedef int (*MPIDI_SHM_mpi_op_free_hook_t) (MPIR_Op * op);
 typedef void (*MPIDI_SHM_am_request_init_t) (MPIR_Request * req);
 typedef void (*MPIDI_SHM_am_request_finalize_t) (MPIR_Request * req);
-typedef int (*MPIDI_SHM_send_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                 int tag, MPIR_Comm * comm, int context_offset,
-                                 MPIR_Request ** request);
-typedef int (*MPIDI_SHM_ssend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                  int tag, MPIR_Comm * comm, int context_offset,
-                                  MPIR_Request ** request);
-typedef int (*MPIDI_SHM_startall_t) (int count, MPIR_Request * requests[]);
-typedef int (*MPIDI_SHM_send_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+typedef int (*MPIDI_SHM_mpi_send_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+                                     int tag, MPIR_Comm * comm, int context_offset,
+                                     MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_ssend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
                                       int tag, MPIR_Comm * comm, int context_offset,
                                       MPIR_Request ** request);
-typedef int (*MPIDI_SHM_ssend_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                       int tag, MPIR_Comm * comm, int context_offset,
-                                       MPIR_Request ** request);
-typedef int (*MPIDI_SHM_rsend_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                       int tag, MPIR_Comm * comm, int context_offset,
-                                       MPIR_Request ** request);
-typedef int (*MPIDI_SHM_bsend_init_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+typedef int (*MPIDI_SHM_mpi_startall_t) (int count, MPIR_Request * requests[]);
+typedef int (*MPIDI_SHM_mpi_send_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                          int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_ssend_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                           int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                           MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_rsend_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                           int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                           MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_bsend_init_t) (const void *buf, int count, MPI_Datatype datatype,
+                                           int rank, int tag, MPIR_Comm * comm, int context_offset,
+                                           MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_isend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
+                                      int tag, MPIR_Comm * comm, int context_offset,
+                                      MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_issend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
                                        int tag, MPIR_Comm * comm, int context_offset,
                                        MPIR_Request ** request);
-typedef int (*MPIDI_SHM_isend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                  int tag, MPIR_Comm * comm, int context_offset,
-                                  MPIR_Request ** request);
-typedef int (*MPIDI_SHM_issend_t) (const void *buf, int count, MPI_Datatype datatype, int rank,
-                                   int tag, MPIR_Comm * comm, int context_offset,
-                                   MPIR_Request ** request);
-typedef int (*MPIDI_SHM_cancel_send_t) (MPIR_Request * sreq);
-typedef int (*MPIDI_SHM_recv_init_t) (void *buf, int count, MPI_Datatype datatype, int rank,
+typedef int (*MPIDI_SHM_mpi_cancel_send_t) (MPIR_Request * sreq);
+typedef int (*MPIDI_SHM_mpi_recv_init_t) (void *buf, int count, MPI_Datatype datatype, int rank,
+                                          int tag, MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_recv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
+                                     MPIR_Comm * comm, int context_offset, MPI_Status * status,
+                                     MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_irecv_t) (void *buf, int count, MPI_Datatype datatype, int rank,
                                       int tag, MPIR_Comm * comm, int context_offset,
                                       MPIR_Request ** request);
-typedef int (*MPIDI_SHM_recv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
-                                 MPIR_Comm * comm, int context_offset, MPI_Status * status,
-                                 MPIR_Request ** request);
-typedef int (*MPIDI_SHM_irecv_t) (void *buf, int count, MPI_Datatype datatype, int rank, int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request);
-typedef int (*MPIDI_SHM_imrecv_t) (void *buf, int count, MPI_Datatype datatype,
-                                   MPIR_Request * message, MPIR_Request ** rreqp);
-typedef int (*MPIDI_SHM_cancel_recv_t) (MPIR_Request * rreq);
-typedef void *(*MPIDI_SHM_alloc_mem_t) (size_t size, MPIR_Info * info_ptr);
-typedef int (*MPIDI_SHM_free_mem_t) (void *ptr);
-typedef int (*MPIDI_SHM_improbe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
-                                    int *flag, MPIR_Request ** message, MPI_Status * status);
-typedef int (*MPIDI_SHM_iprobe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
-                                   int *flag, MPI_Status * status);
-typedef int (*MPIDI_SHM_win_set_info_t) (MPIR_Win * win, MPIR_Info * info);
-typedef int (*MPIDI_SHM_win_shared_query_t) (MPIR_Win * win, int rank, MPI_Aint * size,
-                                             int *disp_unit, void *baseptr);
-typedef int (*MPIDI_SHM_put_t) (const void *origin_addr, int origin_count,
-                                MPI_Datatype origin_datatype, int target_rank, MPI_Aint target_disp,
-                                int target_count, MPI_Datatype target_datatype, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_start_t) (MPIR_Group * group, int assert, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_complete_t) (MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_post_t) (MPIR_Group * group, int assert, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_wait_t) (MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_test_t) (MPIR_Win * win, int *flag);
-typedef int (*MPIDI_SHM_win_lock_t) (int lock_type, int rank, int assert, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_unlock_t) (int rank, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_get_info_t) (MPIR_Win * win, MPIR_Info ** info_p_p);
-typedef int (*MPIDI_SHM_get_t) (void *origin_addr, int origin_count, MPI_Datatype origin_datatype,
-                                int target_rank, MPI_Aint target_disp, int target_count,
-                                MPI_Datatype target_datatype, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_free_t) (MPIR_Win ** win_ptr);
-typedef int (*MPIDI_SHM_win_fence_t) (int assert, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_create_t) (void *base, MPI_Aint length, int disp_unit, MPIR_Info * info,
-                                       MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr);
-typedef int (*MPIDI_SHM_accumulate_t) (const void *origin_addr, int origin_count,
-                                       MPI_Datatype origin_datatype, int target_rank,
-                                       MPI_Aint target_disp, int target_count,
-                                       MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_attach_t) (MPIR_Win * win, void *base, MPI_Aint size);
-typedef int (*MPIDI_SHM_win_allocate_shared_t) (MPI_Aint size, int disp_unit, MPIR_Info * info_ptr,
-                                                MPIR_Comm * comm_ptr, void **base_ptr,
-                                                MPIR_Win ** win_ptr);
-typedef int (*MPIDI_SHM_rput_t) (const void *origin_addr, int origin_count,
-                                 MPI_Datatype origin_datatype, int target_rank,
-                                 MPI_Aint target_disp, int target_count,
-                                 MPI_Datatype target_datatype, MPIR_Win * win,
-                                 MPIR_Request ** request);
-typedef int (*MPIDI_SHM_win_flush_local_t) (int rank, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_detach_t) (MPIR_Win * win, const void *base);
-typedef int (*MPIDI_SHM_compare_and_swap_t) (const void *origin_addr, const void *compare_addr,
-                                             void *result_addr, MPI_Datatype datatype,
-                                             int target_rank, MPI_Aint target_disp, MPIR_Win * win);
-typedef int (*MPIDI_SHM_raccumulate_t) (const void *origin_addr, int origin_count,
-                                        MPI_Datatype origin_datatype, int target_rank,
-                                        MPI_Aint target_disp, int target_count,
-                                        MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win,
-                                        MPIR_Request ** request);
-typedef int (*MPIDI_SHM_rget_accumulate_t) (const void *origin_addr, int origin_count,
-                                            MPI_Datatype origin_datatype, void *result_addr,
-                                            int result_count, MPI_Datatype result_datatype,
-                                            int target_rank, MPI_Aint target_disp, int target_count,
+typedef int (*MPIDI_SHM_mpi_imrecv_t) (void *buf, int count, MPI_Datatype datatype,
+                                       MPIR_Request * message, MPIR_Request ** rreqp);
+typedef int (*MPIDI_SHM_mpi_cancel_recv_t) (MPIR_Request * rreq);
+typedef void *(*MPIDI_SHM_mpi_alloc_mem_t) (size_t size, MPIR_Info * info_ptr);
+typedef int (*MPIDI_SHM_mpi_free_mem_t) (void *ptr);
+typedef int (*MPIDI_SHM_mpi_improbe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
+                                        int *flag, MPIR_Request ** message, MPI_Status * status);
+typedef int (*MPIDI_SHM_mpi_iprobe_t) (int source, int tag, MPIR_Comm * comm, int context_offset,
+                                       int *flag, MPI_Status * status);
+typedef int (*MPIDI_SHM_mpi_win_set_info_t) (MPIR_Win * win, MPIR_Info * info);
+typedef int (*MPIDI_SHM_mpi_win_shared_query_t) (MPIR_Win * win, int rank, MPI_Aint * size,
+                                                 int *disp_unit, void *baseptr);
+typedef int (*MPIDI_SHM_mpi_put_t) (const void *origin_addr, int origin_count,
+                                    MPI_Datatype origin_datatype, int target_rank,
+                                    MPI_Aint target_disp, int target_count,
+                                    MPI_Datatype target_datatype, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_start_t) (MPIR_Group * group, int assert, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_complete_t) (MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_post_t) (MPIR_Group * group, int assert, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_wait_t) (MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_test_t) (MPIR_Win * win, int *flag);
+typedef int (*MPIDI_SHM_mpi_win_lock_t) (int lock_type, int rank, int assert, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_unlock_t) (int rank, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_get_info_t) (MPIR_Win * win, MPIR_Info ** info_p_p);
+typedef int (*MPIDI_SHM_mpi_get_t) (void *origin_addr, int origin_count,
+                                    MPI_Datatype origin_datatype, int target_rank,
+                                    MPI_Aint target_disp, int target_count,
+                                    MPI_Datatype target_datatype, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_free_t) (MPIR_Win ** win_ptr);
+typedef int (*MPIDI_SHM_mpi_win_fence_t) (int assert, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_create_t) (void *base, MPI_Aint length, int disp_unit,
+                                           MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                           MPIR_Win ** win_ptr);
+typedef int (*MPIDI_SHM_mpi_accumulate_t) (const void *origin_addr, int origin_count,
+                                           MPI_Datatype origin_datatype, int target_rank,
+                                           MPI_Aint target_disp, int target_count,
+                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_attach_t) (MPIR_Win * win, void *base, MPI_Aint size);
+typedef int (*MPIDI_SHM_mpi_win_allocate_shared_t) (MPI_Aint size, int disp_unit,
+                                                    MPIR_Info * info_ptr, MPIR_Comm * comm_ptr,
+                                                    void **base_ptr, MPIR_Win ** win_ptr);
+typedef int (*MPIDI_SHM_mpi_rput_t) (const void *origin_addr, int origin_count,
+                                     MPI_Datatype origin_datatype, int target_rank,
+                                     MPI_Aint target_disp, int target_count,
+                                     MPI_Datatype target_datatype, MPIR_Win * win,
+                                     MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_win_flush_local_t) (int rank, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_detach_t) (MPIR_Win * win, const void *base);
+typedef int (*MPIDI_SHM_mpi_compare_and_swap_t) (const void *origin_addr, const void *compare_addr,
+                                                 void *result_addr, MPI_Datatype datatype,
+                                                 int target_rank, MPI_Aint target_disp,
+                                                 MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_raccumulate_t) (const void *origin_addr, int origin_count,
+                                            MPI_Datatype origin_datatype, int target_rank,
+                                            MPI_Aint target_disp, int target_count,
                                             MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win,
                                             MPIR_Request ** request);
-typedef int (*MPIDI_SHM_fetch_and_op_t) (const void *origin_addr, void *result_addr,
-                                         MPI_Datatype datatype, int target_rank,
-                                         MPI_Aint target_disp, MPI_Op op, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_allocate_t) (MPI_Aint size, int disp_unit, MPIR_Info * info,
-                                         MPIR_Comm * comm, void *baseptr, MPIR_Win ** win);
-typedef int (*MPIDI_SHM_win_flush_t) (int rank, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_flush_local_all_t) (MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_unlock_all_t) (MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_create_dynamic_t) (MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win);
-typedef int (*MPIDI_SHM_rget_t) (void *origin_addr, int origin_count, MPI_Datatype origin_datatype,
-                                 int target_rank, MPI_Aint target_disp, int target_count,
-                                 MPI_Datatype target_datatype, MPIR_Win * win,
-                                 MPIR_Request ** request);
-typedef int (*MPIDI_SHM_win_sync_t) (MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_flush_all_t) (MPIR_Win * win);
-typedef int (*MPIDI_SHM_get_accumulate_t) (const void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, void *result_addr,
-                                           int result_count, MPI_Datatype result_datatype,
-                                           int target_rank, MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win);
-typedef int (*MPIDI_SHM_win_lock_all_t) (int assert, MPIR_Win * win);
-typedef int (*MPIDI_SHM_barrier_t) (MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_bcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
-                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_allreduce_t) (const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                      MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_allgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
+typedef int (*MPIDI_SHM_mpi_rget_accumulate_t) (const void *origin_addr, int origin_count,
+                                                MPI_Datatype origin_datatype, void *result_addr,
+                                                int result_count, MPI_Datatype result_datatype,
+                                                int target_rank, MPI_Aint target_disp,
+                                                int target_count, MPI_Datatype target_datatype,
+                                                MPI_Op op, MPIR_Win * win, MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_fetch_and_op_t) (const void *origin_addr, void *result_addr,
+                                             MPI_Datatype datatype, int target_rank,
+                                             MPI_Aint target_disp, MPI_Op op, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_allocate_t) (MPI_Aint size, int disp_unit, MPIR_Info * info,
+                                             MPIR_Comm * comm, void *baseptr, MPIR_Win ** win);
+typedef int (*MPIDI_SHM_mpi_win_flush_t) (int rank, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_flush_local_all_t) (MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_unlock_all_t) (MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_create_dynamic_t) (MPIR_Info * info, MPIR_Comm * comm,
+                                                   MPIR_Win ** win);
+typedef int (*MPIDI_SHM_mpi_rget_t) (void *origin_addr, int origin_count,
+                                     MPI_Datatype origin_datatype, int target_rank,
+                                     MPI_Aint target_disp, int target_count,
+                                     MPI_Datatype target_datatype, MPIR_Win * win,
+                                     MPIR_Request ** request);
+typedef int (*MPIDI_SHM_mpi_win_sync_t) (MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_flush_all_t) (MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_get_accumulate_t) (const void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, void *result_addr,
+                                               int result_count, MPI_Datatype result_datatype,
+                                               int target_rank, MPI_Aint target_disp,
+                                               int target_count, MPI_Datatype target_datatype,
+                                               MPI_Op op, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_win_lock_all_t) (int assert, MPIR_Win * win);
+typedef int (*MPIDI_SHM_mpi_barrier_t) (MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_bcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
                                       MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_allgatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm,
-                                       MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_scatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_scatterv_t) (const void *sendbuf, const int *sendcounts, const int *displs,
-                                     MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                     MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                     MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_gather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_gatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm,
-                                    MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_alltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_alltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
-                                      const int *recvcounts, const int *rdispls,
-                                      MPI_Datatype recvtype, MPIR_Comm * comm,
-                                      MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_alltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int *recvcounts, const int *rdispls,
-                                      const MPI_Datatype recvtypes[], MPIR_Comm * comm,
-                                      MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_reduce_t) (const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_reduce_scatter_t) (const void *sendbuf, void *recvbuf,
-                                           const int *recvcounts, MPI_Datatype datatype, MPI_Op op,
-                                           MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_reduce_scatter_block_t) (const void *sendbuf, void *recvbuf, int recvcount,
-                                                 MPI_Datatype datatype, MPI_Op op,
-                                                 MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_scan_t) (const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                 MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_exscan_t) (const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                   MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_neighbor_allgather_t) (const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm,
-                                               MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_neighbor_allgatherv_t) (const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int *recvcounts, const int *displs,
-                                                MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_neighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                               const int *sdispls, MPI_Datatype sendtype,
-                                               void *recvbuf, const int *recvcounts,
-                                               const int *rdispls, MPI_Datatype recvtype,
-                                               MPIR_Comm * comm, MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_neighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                               const MPI_Aint * sdispls,
-                                               const MPI_Datatype * sendtypes, void *recvbuf,
-                                               const int *recvcounts, const MPI_Aint * rdispls,
-                                               const MPI_Datatype * recvtypes, MPIR_Comm * comm,
-                                               MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_neighbor_alltoall_t) (const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm,
-                                              MPIR_Errflag_t * errflag);
-typedef int (*MPIDI_SHM_ineighbor_allgather_t) (const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                MPI_Request * req);
-typedef int (*MPIDI_SHM_ineighbor_allgatherv_t) (const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 const int *recvcounts, const int *displs,
-                                                 MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                 MPI_Request * req);
-typedef int (*MPIDI_SHM_ineighbor_alltoall_t) (const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm,
-                                               MPI_Request * req);
-typedef int (*MPIDI_SHM_ineighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                                const int *sdispls, MPI_Datatype sendtype,
-                                                void *recvbuf, const int *recvcounts,
-                                                const int *rdispls, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_ineighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                                const MPI_Aint * sdispls,
-                                                const MPI_Datatype * sendtypes, void *recvbuf,
-                                                const int *recvcounts, const MPI_Aint * rdispls,
-                                                const MPI_Datatype * recvtypes, MPIR_Comm * comm,
-                                                MPI_Request * req);
-typedef int (*MPIDI_SHM_ibarrier_t) (MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_ibcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
-                                   MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_iallgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+typedef int (*MPIDI_SHM_mpi_allreduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                          MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_allgather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_allgatherv_t) (const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm,
+                                           MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_scatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_scatterv_t) (const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype, void *recvbuf,
+                                         int recvcount, MPI_Datatype recvtype, int root,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_gather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                       MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_iallgatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_gatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
                                         void *recvbuf, const int *recvcounts, const int *displs,
-                                        MPI_Datatype recvtype, MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_iallreduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm,
+                                        MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_alltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_alltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
+                                          const int *recvcounts, const int *rdispls,
+                                          MPI_Datatype recvtype, MPIR_Comm * comm,
+                                          MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_alltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int *recvcounts, const int *rdispls,
+                                          const MPI_Datatype recvtypes[], MPIR_Comm * comm,
+                                          MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_reduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_reduce_scatter_t) (const void *sendbuf, void *recvbuf,
+                                               const int *recvcounts, MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr,
+                                               MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_reduce_scatter_block_t) (const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_scan_t) (const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                     MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_exscan_t) (const void *sendbuf, void *recvbuf, int count,
                                        MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                       MPI_Request * req);
-typedef int (*MPIDI_SHM_ialltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_ialltoallv_t) (const void *sendbuf, const int *sendcounts,
-                                       const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
-                                       const int *recvcounts, const int *rdispls,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_ialltoallw_t) (const void *sendbuf, const int *sendcounts,
-                                       const int *sdispls, const MPI_Datatype sendtypes[],
-                                       void *recvbuf, const int *recvcounts, const int *rdispls,
-                                       const MPI_Datatype recvtypes[], MPIR_Comm * comm,
-                                       MPI_Request * req);
-typedef int (*MPIDI_SHM_iexscan_t) (const void *sendbuf, void *recvbuf, int count,
-                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                    MPI_Request * req);
-typedef int (*MPIDI_SHM_igather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                    MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_igatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts, const int *displs,
-                                     MPI_Datatype recvtype, int root, MPIR_Comm * comm,
-                                     MPI_Request * req);
-typedef int (*MPIDI_SHM_ireduce_scatter_block_t) (const void *sendbuf, void *recvbuf, int recvcount,
-                                                  MPI_Datatype datatype, MPI_Op op,
-                                                  MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_ireduce_scatter_t) (const void *sendbuf, void *recvbuf,
-                                            const int *recvcounts, MPI_Datatype datatype, MPI_Op op,
-                                            MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_ireduce_t) (const void *sendbuf, void *recvbuf, int count,
-                                    MPI_Datatype datatype, MPI_Op op, int root,
-                                    MPIR_Comm * comm_ptr, MPI_Request * req);
-typedef int (*MPIDI_SHM_iscan_t) (const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
-                                  MPI_Request * req);
-typedef int (*MPIDI_SHM_iscatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype, int root,
-                                     MPIR_Comm * comm, MPI_Request * req);
-typedef int (*MPIDI_SHM_iscatterv_t) (const void *sendbuf, const int *sendcounts, const int *displs,
-                                      MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                      MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_neighbor_allgather_t) (const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_neighbor_allgatherv_t) (const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int *recvcounts, const int *displs,
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                    MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_neighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                                   const int *sdispls, MPI_Datatype sendtype,
+                                                   void *recvbuf, const int *recvcounts,
+                                                   const int *rdispls, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_neighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                                   const MPI_Aint * sdispls,
+                                                   const MPI_Datatype * sendtypes, void *recvbuf,
+                                                   const int *recvcounts, const MPI_Aint * rdispls,
+                                                   const MPI_Datatype * recvtypes, MPIR_Comm * comm,
+                                                   MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_neighbor_alltoall_t) (const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag);
+typedef int (*MPIDI_SHM_mpi_ineighbor_allgather_t) (const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ineighbor_allgatherv_t) (const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     const int *recvcounts, const int *displs,
+                                                     MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                     MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ineighbor_alltoall_t) (const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ineighbor_alltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                                    const int *sdispls, MPI_Datatype sendtype,
+                                                    void *recvbuf, const int *recvcounts,
+                                                    const int *rdispls, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ineighbor_alltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                                    const MPI_Aint * sdispls,
+                                                    const MPI_Datatype * sendtypes, void *recvbuf,
+                                                    const int *recvcounts, const MPI_Aint * rdispls,
+                                                    const MPI_Datatype * recvtypes,
+                                                    MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ibarrier_t) (MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ibcast_t) (void *buffer, int count, MPI_Datatype datatype, int root,
+                                       MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iallgather_t) (const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf, int recvcount,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm,
+                                           MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iallgatherv_t) (const void *sendbuf, int sendcount,
+                                            MPI_Datatype sendtype, void *recvbuf,
+                                            const int *recvcounts, const int *displs,
+                                            MPI_Datatype recvtype, MPIR_Comm * comm,
+                                            MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iallreduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                           MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                           MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ialltoall_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ialltoallv_t) (const void *sendbuf, const int *sendcounts,
+                                           const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *rdispls,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm,
+                                           MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ialltoallw_t) (const void *sendbuf, const int *sendcounts,
+                                           const int *sdispls, const MPI_Datatype sendtypes[],
+                                           void *recvbuf, const int *recvcounts, const int *rdispls,
+                                           const MPI_Datatype recvtypes[], MPIR_Comm * comm,
+                                           MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iexscan_t) (const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                        MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_igather_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_igatherv_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts, const int *displs,
+                                         MPI_Datatype recvtype, int root, MPIR_Comm * comm,
+                                         MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ireduce_scatter_block_t) (const void *sendbuf, void *recvbuf,
+                                                      int recvcount, MPI_Datatype datatype,
+                                                      MPI_Op op, MPIR_Comm * comm,
+                                                      MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ireduce_scatter_t) (const void *sendbuf, void *recvbuf,
+                                                const int *recvcounts, MPI_Datatype datatype,
+                                                MPI_Op op, MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_ireduce_t) (const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype datatype, MPI_Op op, int root,
+                                        MPIR_Comm * comm_ptr, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iscan_t) (const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                       MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iscatter_t) (const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         int root, MPIR_Comm * comm, MPI_Request * req);
+typedef int (*MPIDI_SHM_mpi_iscatterv_t) (const void *sendbuf, const int *sendcounts,
+                                          const int *displs, MPI_Datatype sendtype, void *recvbuf,
+                                          int recvcount, MPI_Datatype recvtype, int root,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req);
 
 typedef struct MPIDI_SHM_funcs {
-    MPIDI_SHM_init_t init;
-    MPIDI_SHM_finalize_t finalize;
+    MPIDI_SHM_mpi_init_t mpi_init;
+    MPIDI_SHM_mpi_finalize_t mpi_finalize;
     MPIDI_SHM_progress_t progress;
     MPIDI_SHM_am_reg_handler_t am_reg_handler;
-    MPIDI_SHM_comm_connect_t comm_connect;
-    MPIDI_SHM_comm_disconnect_t comm_disconnect;
-    MPIDI_SHM_open_port_t open_port;
-    MPIDI_SHM_close_port_t close_port;
-    MPIDI_SHM_comm_accept_t comm_accept;
+    MPIDI_SHM_mpi_comm_connect_t mpi_comm_connect;
+    MPIDI_SHM_mpi_comm_disconnect_t mpi_comm_disconnect;
+    MPIDI_SHM_mpi_open_port_t mpi_open_port;
+    MPIDI_SHM_mpi_close_port_t mpi_close_port;
+    MPIDI_SHM_mpi_comm_accept_t mpi_comm_accept;
     MPIDI_SHM_am_send_hdr_t am_send_hdr;
     MPIDI_SHM_am_isend_t am_isend;
     MPIDI_SHM_inject_am_t inject_am;
@@ -391,117 +407,117 @@ typedef struct MPIDI_SHM_funcs {
     MPIDI_SHM_getallincomm_t getallincomm;
     MPIDI_SHM_gpid_tolpidarray_t gpid_tolpidarray;
     MPIDI_SHM_create_intercomm_from_lpids_t create_intercomm_from_lpids;
-    MPIDI_SHM_comm_create_hook_t comm_create_hook;
-    MPIDI_SHM_comm_free_hook_t comm_free_hook;
-    MPIDI_SHM_type_create_hook_t type_create_hook;
-    MPIDI_SHM_type_free_hook_t type_free_hook;
-    MPIDI_SHM_op_create_hook_t op_create_hook;
-    MPIDI_SHM_op_free_hook_t op_free_hook;
+    MPIDI_SHM_mpi_comm_create_hook_t mpi_comm_create_hook;
+    MPIDI_SHM_mpi_comm_free_hook_t mpi_comm_free_hook;
+    MPIDI_SHM_mpi_type_create_hook_t mpi_type_create_hook;
+    MPIDI_SHM_mpi_type_free_hook_t mpi_type_free_hook;
+    MPIDI_SHM_mpi_op_create_hook_t mpi_op_create_hook;
+    MPIDI_SHM_mpi_op_free_hook_t mpi_op_free_hook;
     /* Request allocation routines */
     MPIDI_SHM_am_request_init_t am_request_init;
     MPIDI_SHM_am_request_finalize_t am_request_finalize;
 } MPIDI_SHM_funcs_t;
 
 typedef struct MPIDI_SHM_native_funcs {
-    MPIDI_SHM_send_t send;
-    MPIDI_SHM_ssend_t ssend;
-    MPIDI_SHM_startall_t startall;
-    MPIDI_SHM_send_init_t send_init;
-    MPIDI_SHM_ssend_init_t ssend_init;
-    MPIDI_SHM_rsend_init_t rsend_init;
-    MPIDI_SHM_bsend_init_t bsend_init;
-    MPIDI_SHM_isend_t isend;
-    MPIDI_SHM_issend_t issend;
-    MPIDI_SHM_cancel_send_t cancel_send;
-    MPIDI_SHM_recv_init_t recv_init;
-    MPIDI_SHM_recv_t recv;
-    MPIDI_SHM_irecv_t irecv;
-    MPIDI_SHM_imrecv_t imrecv;
-    MPIDI_SHM_cancel_recv_t cancel_recv;
-    MPIDI_SHM_alloc_mem_t alloc_mem;
-    MPIDI_SHM_free_mem_t free_mem;
-    MPIDI_SHM_improbe_t improbe;
-    MPIDI_SHM_iprobe_t iprobe;
-    MPIDI_SHM_win_set_info_t win_set_info;
-    MPIDI_SHM_win_shared_query_t win_shared_query;
-    MPIDI_SHM_put_t put;
-    MPIDI_SHM_win_start_t win_start;
-    MPIDI_SHM_win_complete_t win_complete;
-    MPIDI_SHM_win_post_t win_post;
-    MPIDI_SHM_win_wait_t win_wait;
-    MPIDI_SHM_win_test_t win_test;
-    MPIDI_SHM_win_lock_t win_lock;
-    MPIDI_SHM_win_unlock_t win_unlock;
-    MPIDI_SHM_win_get_info_t win_get_info;
-    MPIDI_SHM_get_t get;
-    MPIDI_SHM_win_free_t win_free;
-    MPIDI_SHM_win_fence_t win_fence;
-    MPIDI_SHM_win_create_t win_create;
-    MPIDI_SHM_accumulate_t accumulate;
-    MPIDI_SHM_win_attach_t win_attach;
-    MPIDI_SHM_win_allocate_shared_t win_allocate_shared;
-    MPIDI_SHM_rput_t rput;
-    MPIDI_SHM_win_flush_local_t win_flush_local;
-    MPIDI_SHM_win_detach_t win_detach;
-    MPIDI_SHM_compare_and_swap_t compare_and_swap;
-    MPIDI_SHM_raccumulate_t raccumulate;
-    MPIDI_SHM_rget_accumulate_t rget_accumulate;
-    MPIDI_SHM_fetch_and_op_t fetch_and_op;
-    MPIDI_SHM_win_allocate_t win_allocate;
-    MPIDI_SHM_win_flush_t win_flush;
-    MPIDI_SHM_win_flush_local_all_t win_flush_local_all;
-    MPIDI_SHM_win_unlock_all_t win_unlock_all;
-    MPIDI_SHM_win_create_dynamic_t win_create_dynamic;
-    MPIDI_SHM_rget_t rget;
-    MPIDI_SHM_win_sync_t win_sync;
-    MPIDI_SHM_win_flush_all_t win_flush_all;
-    MPIDI_SHM_get_accumulate_t get_accumulate;
-    MPIDI_SHM_win_lock_all_t win_lock_all;
+    MPIDI_SHM_mpi_send_t mpi_send;
+    MPIDI_SHM_mpi_ssend_t mpi_ssend;
+    MPIDI_SHM_mpi_startall_t mpi_startall;
+    MPIDI_SHM_mpi_send_init_t mpi_send_init;
+    MPIDI_SHM_mpi_ssend_init_t mpi_ssend_init;
+    MPIDI_SHM_mpi_rsend_init_t mpi_rsend_init;
+    MPIDI_SHM_mpi_bsend_init_t mpi_bsend_init;
+    MPIDI_SHM_mpi_isend_t mpi_isend;
+    MPIDI_SHM_mpi_issend_t mpi_issend;
+    MPIDI_SHM_mpi_cancel_send_t mpi_cancel_send;
+    MPIDI_SHM_mpi_recv_init_t mpi_recv_init;
+    MPIDI_SHM_mpi_recv_t mpi_recv;
+    MPIDI_SHM_mpi_irecv_t mpi_irecv;
+    MPIDI_SHM_mpi_imrecv_t mpi_imrecv;
+    MPIDI_SHM_mpi_cancel_recv_t mpi_cancel_recv;
+    MPIDI_SHM_mpi_alloc_mem_t mpi_alloc_mem;
+    MPIDI_SHM_mpi_free_mem_t mpi_free_mem;
+    MPIDI_SHM_mpi_improbe_t mpi_improbe;
+    MPIDI_SHM_mpi_iprobe_t mpi_iprobe;
+    MPIDI_SHM_mpi_win_set_info_t mpi_win_set_info;
+    MPIDI_SHM_mpi_win_shared_query_t mpi_win_shared_query;
+    MPIDI_SHM_mpi_put_t mpi_put;
+    MPIDI_SHM_mpi_win_start_t mpi_win_start;
+    MPIDI_SHM_mpi_win_complete_t mpi_win_complete;
+    MPIDI_SHM_mpi_win_post_t mpi_win_post;
+    MPIDI_SHM_mpi_win_wait_t mpi_win_wait;
+    MPIDI_SHM_mpi_win_test_t mpi_win_test;
+    MPIDI_SHM_mpi_win_lock_t mpi_win_lock;
+    MPIDI_SHM_mpi_win_unlock_t mpi_win_unlock;
+    MPIDI_SHM_mpi_win_get_info_t mpi_win_get_info;
+    MPIDI_SHM_mpi_get_t mpi_get;
+    MPIDI_SHM_mpi_win_free_t mpi_win_free;
+    MPIDI_SHM_mpi_win_fence_t mpi_win_fence;
+    MPIDI_SHM_mpi_win_create_t mpi_win_create;
+    MPIDI_SHM_mpi_accumulate_t mpi_accumulate;
+    MPIDI_SHM_mpi_win_attach_t mpi_win_attach;
+    MPIDI_SHM_mpi_win_allocate_shared_t mpi_win_allocate_shared;
+    MPIDI_SHM_mpi_rput_t mpi_rput;
+    MPIDI_SHM_mpi_win_flush_local_t mpi_win_flush_local;
+    MPIDI_SHM_mpi_win_detach_t mpi_win_detach;
+    MPIDI_SHM_mpi_compare_and_swap_t mpi_compare_and_swap;
+    MPIDI_SHM_mpi_raccumulate_t mpi_raccumulate;
+    MPIDI_SHM_mpi_rget_accumulate_t mpi_rget_accumulate;
+    MPIDI_SHM_mpi_fetch_and_op_t mpi_fetch_and_op;
+    MPIDI_SHM_mpi_win_allocate_t mpi_win_allocate;
+    MPIDI_SHM_mpi_win_flush_t mpi_win_flush;
+    MPIDI_SHM_mpi_win_flush_local_all_t mpi_win_flush_local_all;
+    MPIDI_SHM_mpi_win_unlock_all_t mpi_win_unlock_all;
+    MPIDI_SHM_mpi_win_create_dynamic_t mpi_win_create_dynamic;
+    MPIDI_SHM_mpi_rget_t mpi_rget;
+    MPIDI_SHM_mpi_win_sync_t mpi_win_sync;
+    MPIDI_SHM_mpi_win_flush_all_t mpi_win_flush_all;
+    MPIDI_SHM_mpi_get_accumulate_t mpi_get_accumulate;
+    MPIDI_SHM_mpi_win_lock_all_t mpi_win_lock_all;
     /* Collectives */
-    MPIDI_SHM_barrier_t barrier;
-    MPIDI_SHM_bcast_t bcast;
-    MPIDI_SHM_allreduce_t allreduce;
-    MPIDI_SHM_allgather_t allgather;
-    MPIDI_SHM_allgatherv_t allgatherv;
-    MPIDI_SHM_scatter_t scatter;
-    MPIDI_SHM_scatterv_t scatterv;
-    MPIDI_SHM_gather_t gather;
-    MPIDI_SHM_gatherv_t gatherv;
-    MPIDI_SHM_alltoall_t alltoall;
-    MPIDI_SHM_alltoallv_t alltoallv;
-    MPIDI_SHM_alltoallw_t alltoallw;
-    MPIDI_SHM_reduce_t reduce;
-    MPIDI_SHM_reduce_scatter_t reduce_scatter;
-    MPIDI_SHM_reduce_scatter_block_t reduce_scatter_block;
-    MPIDI_SHM_scan_t scan;
-    MPIDI_SHM_exscan_t exscan;
-    MPIDI_SHM_neighbor_allgather_t neighbor_allgather;
-    MPIDI_SHM_neighbor_allgatherv_t neighbor_allgatherv;
-    MPIDI_SHM_neighbor_alltoall_t neighbor_alltoall;
-    MPIDI_SHM_neighbor_alltoallv_t neighbor_alltoallv;
-    MPIDI_SHM_neighbor_alltoallw_t neighbor_alltoallw;
-    MPIDI_SHM_ineighbor_allgather_t ineighbor_allgather;
-    MPIDI_SHM_ineighbor_allgatherv_t ineighbor_allgatherv;
-    MPIDI_SHM_ineighbor_alltoall_t ineighbor_alltoall;
-    MPIDI_SHM_ineighbor_alltoallv_t ineighbor_alltoallv;
-    MPIDI_SHM_ineighbor_alltoallw_t ineighbor_alltoallw;
-    MPIDI_SHM_ibarrier_t ibarrier;
-    MPIDI_SHM_ibcast_t ibcast;
-    MPIDI_SHM_iallgather_t iallgather;
-    MPIDI_SHM_iallgatherv_t iallgatherv;
-    MPIDI_SHM_iallreduce_t iallreduce;
-    MPIDI_SHM_ialltoall_t ialltoall;
-    MPIDI_SHM_ialltoallv_t ialltoallv;
-    MPIDI_SHM_ialltoallw_t ialltoallw;
-    MPIDI_SHM_iexscan_t iexscan;
-    MPIDI_SHM_igather_t igather;
-    MPIDI_SHM_igatherv_t igatherv;
-    MPIDI_SHM_ireduce_scatter_block_t ireduce_scatter_block;
-    MPIDI_SHM_ireduce_scatter_t ireduce_scatter;
-    MPIDI_SHM_ireduce_t ireduce;
-    MPIDI_SHM_iscan_t iscan;
-    MPIDI_SHM_iscatter_t iscatter;
-    MPIDI_SHM_iscatterv_t iscatterv;
+    MPIDI_SHM_mpi_barrier_t mpi_barrier;
+    MPIDI_SHM_mpi_bcast_t mpi_bcast;
+    MPIDI_SHM_mpi_allreduce_t mpi_allreduce;
+    MPIDI_SHM_mpi_allgather_t mpi_allgather;
+    MPIDI_SHM_mpi_allgatherv_t mpi_allgatherv;
+    MPIDI_SHM_mpi_scatter_t mpi_scatter;
+    MPIDI_SHM_mpi_scatterv_t mpi_scatterv;
+    MPIDI_SHM_mpi_gather_t mpi_gather;
+    MPIDI_SHM_mpi_gatherv_t mpi_gatherv;
+    MPIDI_SHM_mpi_alltoall_t mpi_alltoall;
+    MPIDI_SHM_mpi_alltoallv_t mpi_alltoallv;
+    MPIDI_SHM_mpi_alltoallw_t mpi_alltoallw;
+    MPIDI_SHM_mpi_reduce_t mpi_reduce;
+    MPIDI_SHM_mpi_reduce_scatter_t mpi_reduce_scatter;
+    MPIDI_SHM_mpi_reduce_scatter_block_t mpi_reduce_scatter_block;
+    MPIDI_SHM_mpi_scan_t mpi_scan;
+    MPIDI_SHM_mpi_exscan_t mpi_exscan;
+    MPIDI_SHM_mpi_neighbor_allgather_t mpi_neighbor_allgather;
+    MPIDI_SHM_mpi_neighbor_allgatherv_t mpi_neighbor_allgatherv;
+    MPIDI_SHM_mpi_neighbor_alltoall_t mpi_neighbor_alltoall;
+    MPIDI_SHM_mpi_neighbor_alltoallv_t mpi_neighbor_alltoallv;
+    MPIDI_SHM_mpi_neighbor_alltoallw_t mpi_neighbor_alltoallw;
+    MPIDI_SHM_mpi_ineighbor_allgather_t mpi_ineighbor_allgather;
+    MPIDI_SHM_mpi_ineighbor_allgatherv_t mpi_ineighbor_allgatherv;
+    MPIDI_SHM_mpi_ineighbor_alltoall_t mpi_ineighbor_alltoall;
+    MPIDI_SHM_mpi_ineighbor_alltoallv_t mpi_ineighbor_alltoallv;
+    MPIDI_SHM_mpi_ineighbor_alltoallw_t mpi_ineighbor_alltoallw;
+    MPIDI_SHM_mpi_ibarrier_t mpi_ibarrier;
+    MPIDI_SHM_mpi_ibcast_t mpi_ibcast;
+    MPIDI_SHM_mpi_iallgather_t mpi_iallgather;
+    MPIDI_SHM_mpi_iallgatherv_t mpi_iallgatherv;
+    MPIDI_SHM_mpi_iallreduce_t mpi_iallreduce;
+    MPIDI_SHM_mpi_ialltoall_t mpi_ialltoall;
+    MPIDI_SHM_mpi_ialltoallv_t mpi_ialltoallv;
+    MPIDI_SHM_mpi_ialltoallw_t mpi_ialltoallw;
+    MPIDI_SHM_mpi_iexscan_t mpi_iexscan;
+    MPIDI_SHM_mpi_igather_t mpi_igather;
+    MPIDI_SHM_mpi_igatherv_t mpi_igatherv;
+    MPIDI_SHM_mpi_ireduce_scatter_block_t mpi_ireduce_scatter_block;
+    MPIDI_SHM_mpi_ireduce_scatter_t mpi_ireduce_scatter;
+    MPIDI_SHM_mpi_ireduce_t mpi_ireduce;
+    MPIDI_SHM_mpi_iscan_t mpi_iscan;
+    MPIDI_SHM_mpi_iscatter_t mpi_iscatter;
+    MPIDI_SHM_mpi_iscatterv_t mpi_iscatterv;
 } MPIDI_SHM_native_funcs_t;
 
 extern MPIDI_SHM_funcs_t *MPIDI_SHM_funcs[];
@@ -511,30 +527,30 @@ extern MPIDI_SHM_native_funcs_t *MPIDI_SHM_native_func;
 extern int MPIDI_num_shms;
 extern char MPIDI_SHM_strings[][MPIDI_MAX_SHM_STRING_LEN];
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_init(int rank, int size) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_finalize(void) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_init(int rank, int size) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_finalize(void) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_progress(int blocking) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_reg_handler(int handler_id,
                                                       MPIDI_SHM_am_origin_handler_fn
                                                       origin_handler_fn,
                                                       MPIDI_SHM_am_target_handler_fn
                                                       target_handler_fn) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_connect(const char *port_name, MPIR_Info * info,
-                                                    int root, MPIR_Comm * comm,
-                                                    MPIR_Comm **
-                                                    newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_disconnect(MPIR_Comm *
-                                                       comm_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_open_port(MPIR_Info * info_ptr,
-                                                 char *port_name) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_close_port(const char *port_name) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_accept(const char *port_name, MPIR_Info * info,
-                                                   int root, MPIR_Comm * comm,
-                                                   MPIR_Comm **
-                                                   newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_send_hdr(int rank, MPIR_Comm * comm,
-                                                   int handler_id, const void *am_hdr,
-                                                   size_t am_hdr_sz,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_connect(const char *port_name, MPIR_Info * info,
+                                                        int root, MPIR_Comm * comm,
+                                                        MPIR_Comm **
+                                                        newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_disconnect(MPIR_Comm *
+                                                           comm_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_open_port(MPIR_Info * info_ptr,
+                                                     char *port_name) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_close_port(const char *port_name)
+    MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_accept(const char *port_name, MPIR_Info * info,
+                                                       int root, MPIR_Comm * comm,
+                                                       MPIR_Comm **
+                                                       newcomm_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_send_hdr(int rank, MPIR_Comm * comm, int handler_id,
+                                                   const void *am_hdr, size_t am_hdr_sz,
                                                    void *shm_context) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_isend(int rank, MPIR_Comm * comm, int handler_id,
                                                 const void *am_hdr, size_t am_hdr_sz,
@@ -556,26 +572,24 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_inject_amv(int rank, MPIR_Comm * comm, in
                                                   const void *data, MPI_Count count,
                                                   MPI_Datatype datatype,
                                                   void *shm_context) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_send_hdr_reply(MPIR_Context_id_t context_id,
-                                                         int src_rank, int handler_id,
-                                                         const void *am_hdr,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_send_hdr_reply(MPIR_Context_id_t context_id, int src_rank,
+                                                         int handler_id, const void *am_hdr,
                                                          size_t am_hdr_sz) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_isend_reply(MPIR_Context_id_t context_id,
-                                                      int src_rank, int handler_id,
-                                                      const void *am_hdr, size_t am_hdr_sz,
-                                                      const void *data, MPI_Count count,
-                                                      MPI_Datatype datatype,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_isend_reply(MPIR_Context_id_t context_id, int src_rank,
+                                                      int handler_id, const void *am_hdr,
+                                                      size_t am_hdr_sz, const void *data,
+                                                      MPI_Count count, MPI_Datatype datatype,
                                                       MPIR_Request * sreq) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_inject_am_reply(MPIR_Context_id_t context_id,
-                                                       int src_rank, int handler_id,
-                                                       const void *am_hdr, size_t am_hdr_sz,
-                                                       const void *data, MPI_Count count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_inject_am_reply(MPIR_Context_id_t context_id, int src_rank,
+                                                       int handler_id, const void *am_hdr,
+                                                       size_t am_hdr_sz, const void *data,
+                                                       MPI_Count count,
                                                        MPI_Datatype datatype)
     MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_inject_amv_reply(MPIR_Context_id_t context_id,
-                                                        int src_rank, int handler_id,
-                                                        struct iovec *am_hdrs, size_t iov_len,
-                                                        const void *data, MPI_Count count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_inject_amv_reply(MPIR_Context_id_t context_id, int src_rank,
+                                                        int handler_id, struct iovec *am_hdrs,
+                                                        size_t iov_len, const void *data,
+                                                        MPI_Count count,
                                                         MPI_Datatype datatype)
     MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX size_t MPIDI_SHM_am_hdr_max_sz(void) MPL_STATIC_INLINE_SUFFIX;
@@ -600,450 +614,466 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_create_intercomm_from_lpids(MPIR_Comm * n
                                                                    int size,
                                                                    const int lpids[])
     MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_create_hook(MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_free_hook(MPIR_Comm * comm) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_type_create_hook(MPIR_Datatype *
-                                                        type) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_type_free_hook(MPIR_Datatype *
-                                                      type) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_op_create_hook(MPIR_Op * op) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_op_free_hook(MPIR_Op * op) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_create_hook(MPIR_Comm *
+                                                            comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_free_hook(MPIR_Comm *
+                                                          comm) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_type_create_hook(MPIR_Datatype *
+                                                            type) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_type_free_hook(MPIR_Datatype *
+                                                          type) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_op_create_hook(MPIR_Op * op) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_op_free_hook(MPIR_Op * op) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX void MPIDI_SHM_am_request_init(MPIR_Request *
                                                         req) MPL_STATIC_INLINE_SUFFIX;
 MPL_STATIC_INLINE_PREFIX void MPIDI_SHM_am_request_finalize(MPIR_Request *
                                                             req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_send(const void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ssend(const void *buf, int count,
-                                             MPI_Datatype datatype, int rank, int tag,
-                                             MPIR_Comm * comm, int context_offset,
-                                             MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_startall(int count,
-                                                MPIR_Request * requests[]) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_send_init(const void *buf, int count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_send(const void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset,
+                                                MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ssend(const void *buf, int count,
                                                  MPI_Datatype datatype, int rank, int tag,
                                                  MPIR_Comm * comm, int context_offset,
                                                  MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ssend_init(const void *buf, int count,
-                                                  MPI_Datatype datatype, int rank, int tag,
-                                                  MPIR_Comm * comm, int context_offset,
-                                                  MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rsend_init(const void *buf, int count,
-                                                  MPI_Datatype datatype, int rank, int tag,
-                                                  MPIR_Comm * comm, int context_offset,
-                                                  MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_bsend_init(const void *buf, int count,
-                                                  MPI_Datatype datatype, int rank, int tag,
-                                                  MPIR_Comm * comm, int context_offset,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_startall(int count,
+                                                    MPIR_Request *
+                                                    requests[]) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_send_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request **
+                                                     request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ssend_init(const void *buf, int count,
+                                                      MPI_Datatype datatype, int rank, int tag,
+                                                      MPIR_Comm * comm, int context_offset,
+                                                      MPIR_Request **
+                                                      request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rsend_init(const void *buf, int count,
+                                                      MPI_Datatype datatype, int rank, int tag,
+                                                      MPIR_Comm * comm, int context_offset,
+                                                      MPIR_Request **
+                                                      request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_bsend_init(const void *buf, int count,
+                                                      MPI_Datatype datatype, int rank, int tag,
+                                                      MPIR_Comm * comm, int context_offset,
+                                                      MPIR_Request **
+                                                      request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_isend(const void *buf, int count, MPI_Datatype datatype,
+                                                 int rank, int tag, MPIR_Comm * comm,
+                                                 int context_offset,
+                                                 MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_issend(const void *buf, int count, MPI_Datatype datatype,
+                                                  int rank, int tag, MPIR_Comm * comm,
+                                                  int context_offset,
                                                   MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_isend(const void *buf, int count,
-                                             MPI_Datatype datatype, int rank, int tag,
-                                             MPIR_Comm * comm, int context_offset,
-                                             MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_issend(const void *buf, int count,
-                                              MPI_Datatype datatype, int rank, int tag,
-                                              MPIR_Comm * comm, int context_offset,
-                                              MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_cancel_send(MPIR_Request * sreq) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_recv_init(void *buf, int count, MPI_Datatype datatype,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_cancel_send(MPIR_Request *
+                                                       sreq) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
+                                                     int rank, int tag, MPIR_Comm * comm,
+                                                     int context_offset,
+                                                     MPIR_Request **
+                                                     request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_recv(void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset, MPI_Status * status,
+                                                MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
                                                  int rank, int tag, MPIR_Comm * comm,
                                                  int context_offset,
                                                  MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_recv(void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset, MPI_Status * status,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_irecv(void *buf, int count, MPI_Datatype datatype,
-                                             int rank, int tag, MPIR_Comm * comm,
-                                             int context_offset,
-                                             MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_imrecv(void *buf, int count, MPI_Datatype datatype,
-                                              MPIR_Request * message,
-                                              MPIR_Request ** rreqp) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_cancel_recv(MPIR_Request * rreq) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX void *MPIDI_SHM_alloc_mem(size_t size,
-                                                   MPIR_Info * info_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_free_mem(void *ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_improbe(int source, int tag, MPIR_Comm * comm,
-                                               int context_offset, int *flag,
-                                               MPIR_Request ** message,
-                                               MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iprobe(int source, int tag, MPIR_Comm * comm,
-                                              int context_offset, int *flag,
-                                              MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_set_info(MPIR_Win * win,
-                                                    MPIR_Info * info) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_shared_query(MPIR_Win * win, int rank,
-                                                        MPI_Aint * size, int *disp_unit,
-                                                        void *baseptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_put(const void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype,
-                                           MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_start(MPIR_Group * group, int assert,
-                                                 MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_complete(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_post(MPIR_Group * group, int assert,
-                                                MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_wait(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_test(MPIR_Win * win, int *flag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_lock(int lock_type, int rank, int assert,
-                                                MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_unlock(int rank,
-                                                  MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_get_info(MPIR_Win * win,
-                                                    MPIR_Info ** info_p_p) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_get(void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype,
-                                           MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_free(MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_fence(int assert,
-                                                 MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_create(void *base, MPI_Aint length, int disp_unit,
-                                                  MPIR_Info * info, MPIR_Comm * comm_ptr,
-                                                  MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_accumulate(const void *origin_addr, int origin_count,
-                                                  MPI_Datatype origin_datatype,
-                                                  int target_rank, MPI_Aint target_disp,
-                                                  int target_count,
-                                                  MPI_Datatype target_datatype, MPI_Op op,
-                                                  MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_attach(MPIR_Win * win, void *base,
-                                                  MPI_Aint size) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_allocate_shared(MPI_Aint size, int disp_unit,
-                                                           MPIR_Info * info_ptr,
-                                                           MPIR_Comm * comm_ptr,
-                                                           void **base_ptr,
-                                                           MPIR_Win **
-                                                           win_ptr) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rput(const void *origin_addr, int origin_count,
-                                            MPI_Datatype origin_datatype, int target_rank,
-                                            MPI_Aint target_disp, int target_count,
-                                            MPI_Datatype target_datatype, MPIR_Win * win,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush_local(int rank,
-                                                       MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_detach(MPIR_Win * win,
-                                                  const void *base) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_compare_and_swap(const void *origin_addr,
-                                                        const void *compare_addr,
-                                                        void *result_addr,
-                                                        MPI_Datatype datatype,
-                                                        int target_rank, MPI_Aint target_disp,
-                                                        MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_raccumulate(const void *origin_addr, int origin_count,
-                                                   MPI_Datatype origin_datatype,
-                                                   int target_rank, MPI_Aint target_disp,
-                                                   int target_count,
-                                                   MPI_Datatype target_datatype, MPI_Op op,
-                                                   MPIR_Win * win,
-                                                   MPIR_Request **
-                                                   request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rget_accumulate(const void *origin_addr,
-                                                       int origin_count,
-                                                       MPI_Datatype origin_datatype,
-                                                       void *result_addr, int result_count,
-                                                       MPI_Datatype result_datatype,
-                                                       int target_rank, MPI_Aint target_disp,
-                                                       int target_count,
-                                                       MPI_Datatype target_datatype,
-                                                       MPI_Op op, MPIR_Win * win,
-                                                       MPIR_Request **
-                                                       request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_fetch_and_op(const void *origin_addr,
-                                                    void *result_addr, MPI_Datatype datatype,
-                                                    int target_rank, MPI_Aint target_disp,
-                                                    MPI_Op op,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_imrecv(void *buf, int count, MPI_Datatype datatype,
+                                                  MPIR_Request * message,
+                                                  MPIR_Request ** rreqp) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_cancel_recv(MPIR_Request *
+                                                       rreq) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX void *MPIDI_SHM_mpi_alloc_mem(size_t size,
+                                                       MPIR_Info *
+                                                       info_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_free_mem(void *ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_improbe(int source, int tag, MPIR_Comm * comm,
+                                                   int context_offset, int *flag,
+                                                   MPIR_Request ** message,
+                                                   MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iprobe(int source, int tag, MPIR_Comm * comm,
+                                                  int context_offset, int *flag,
+                                                  MPI_Status * status) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_set_info(MPIR_Win * win,
+                                                        MPIR_Info * info) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_shared_query(MPIR_Win * win, int rank,
+                                                            MPI_Aint * size, int *disp_unit,
+                                                            void *baseptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_put(const void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_start(MPIR_Group * group, int assert,
+                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_complete(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_post(MPIR_Group * group, int assert,
                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_allocate(MPI_Aint size, int disp_unit,
-                                                    MPIR_Info * info, MPIR_Comm * comm,
-                                                    void *baseptr,
-                                                    MPIR_Win ** win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush(int rank, MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush_local_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_unlock_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
-                                                          MPIR_Win ** win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rget(void *origin_addr, int origin_count,
-                                            MPI_Datatype origin_datatype, int target_rank,
-                                            MPI_Aint target_disp, int target_count,
-                                            MPI_Datatype target_datatype, MPIR_Win * win,
-                                            MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_sync(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_get_accumulate(const void *origin_addr,
-                                                      int origin_count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_wait(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_test(MPIR_Win * win,
+                                                    int *flag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_lock(int lock_type, int rank, int assert,
+                                                    MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_unlock(int rank,
+                                                      MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_get_info(MPIR_Win * win,
+                                                        MPIR_Info **
+                                                        info_p_p) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_get(void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype,
+                                               MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_free(MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_fence(int assert,
+                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_create(void *base, MPI_Aint length, int disp_unit,
+                                                      MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                                      MPIR_Win ** win_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_accumulate(const void *origin_addr, int origin_count,
                                                       MPI_Datatype origin_datatype,
-                                                      void *result_addr, int result_count,
-                                                      MPI_Datatype result_datatype,
                                                       int target_rank, MPI_Aint target_disp,
                                                       int target_count,
                                                       MPI_Datatype target_datatype, MPI_Op op,
                                                       MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_lock_all(int assert,
-                                                    MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_barrier(MPIR_Comm * comm,
-                                               MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                             int root, MPIR_Comm * comm,
-                                             MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_allreduce(const void *sendbuf, void *recvbuf,
-                                                 int count, MPI_Datatype datatype, MPI_Op op,
-                                                 MPIR_Comm * comm,
-                                                 MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_allgather(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 int recvcount, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_attach(MPIR_Win * win, void *base,
+                                                      MPI_Aint size) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_allocate_shared(MPI_Aint size, int disp_unit,
+                                                               MPIR_Info * info_ptr,
+                                                               MPIR_Comm * comm_ptr,
+                                                               void **base_ptr,
+                                                               MPIR_Win **
+                                                               win_ptr) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rput(const void *origin_addr, int origin_count,
+                                                MPI_Datatype origin_datatype, int target_rank,
+                                                MPI_Aint target_disp, int target_count,
+                                                MPI_Datatype target_datatype, MPIR_Win * win,
+                                                MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush_local(int rank,
+                                                           MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_detach(MPIR_Win * win,
+                                                      const void *base) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_compare_and_swap(const void *origin_addr,
+                                                            const void *compare_addr,
+                                                            void *result_addr,
+                                                            MPI_Datatype datatype,
+                                                            int target_rank, MPI_Aint target_disp,
+                                                            MPIR_Win *
+                                                            win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_raccumulate(const void *origin_addr, int origin_count,
+                                                       MPI_Datatype origin_datatype,
+                                                       int target_rank, MPI_Aint target_disp,
+                                                       int target_count,
+                                                       MPI_Datatype target_datatype, MPI_Op op,
+                                                       MPIR_Win * win,
+                                                       MPIR_Request **
+                                                       request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rget_accumulate(const void *origin_addr,
+                                                           int origin_count,
+                                                           MPI_Datatype origin_datatype,
+                                                           void *result_addr, int result_count,
+                                                           MPI_Datatype result_datatype,
+                                                           int target_rank, MPI_Aint target_disp,
+                                                           int target_count,
+                                                           MPI_Datatype target_datatype, MPI_Op op,
+                                                           MPIR_Win * win,
+                                                           MPIR_Request **
+                                                           request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_fetch_and_op(const void *origin_addr, void *result_addr,
+                                                        MPI_Datatype datatype, int target_rank,
+                                                        MPI_Aint target_disp, MPI_Op op,
+                                                        MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_allocate(MPI_Aint size, int disp_unit,
+                                                        MPIR_Info * info, MPIR_Comm * comm,
+                                                        void *baseptr,
+                                                        MPIR_Win ** win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush(int rank,
+                                                     MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush_local_all(MPIR_Win *
+                                                               win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_unlock_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                              MPIR_Win **
+                                                              win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rget(void *origin_addr, int origin_count,
+                                                MPI_Datatype origin_datatype, int target_rank,
+                                                MPI_Aint target_disp, int target_count,
+                                                MPI_Datatype target_datatype, MPIR_Win * win,
+                                                MPIR_Request ** request) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_sync(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush_all(MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_get_accumulate(const void *origin_addr,
+                                                          int origin_count,
+                                                          MPI_Datatype origin_datatype,
+                                                          void *result_addr, int result_count,
+                                                          MPI_Datatype result_datatype,
+                                                          int target_rank, MPI_Aint target_disp,
+                                                          int target_count,
+                                                          MPI_Datatype target_datatype, MPI_Op op,
+                                                          MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_lock_all(int assert,
+                                                        MPIR_Win * win) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_barrier(MPIR_Comm * comm,
+                                                   MPIR_Errflag_t *
+                                                   errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                                 int root, MPIR_Comm * comm,
                                                  MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_allgatherv(const void *sendbuf, int sendcount,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                                     MPI_Datatype datatype, MPI_Op op,
+                                                     MPIR_Comm * comm,
+                                                     MPIR_Errflag_t *
+                                                     errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_allgather(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     int recvcount, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm,
+                                                     MPIR_Errflag_t *
+                                                     errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_allgatherv(const void *sendbuf, int sendcount,
+                                                      MPI_Datatype sendtype, void *recvbuf,
+                                                      const int *recvcounts, const int *displs,
+                                                      MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                      MPIR_Errflag_t *
+                                                      errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_scatter(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm,
+                                                   MPIR_Errflag_t *
+                                                   errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                                    const int *displs, MPI_Datatype sendtype,
+                                                    void *recvbuf, int recvcount,
+                                                    MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t *
+                                                    errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_gather(const void *sendbuf, int sendcount,
                                                   MPI_Datatype sendtype, void *recvbuf,
-                                                  const int *recvcounts, const int *displs,
-                                                  MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                  int recvcount, MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm,
                                                   MPIR_Errflag_t *
                                                   errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_scatter(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm,
-                                               MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_scatterv(const void *sendbuf, const int *sendcounts,
-                                                const int *displs, MPI_Datatype sendtype,
-                                                void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, int root,
-                                                MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_gather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm,
-                                              MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_gatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int *recvcounts, const int *displs,
-                                               MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm,
-                                               MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_alltoall(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm,
-                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls, MPI_Datatype sendtype,
-                                                 void *recvbuf, const int *recvcounts,
-                                                 const int *rdispls, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm,
-                                                 MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_alltoallw(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls,
-                                                 const MPI_Datatype sendtypes[],
-                                                 void *recvbuf, const int *recvcounts,
-                                                 const int *rdispls,
-                                                 const MPI_Datatype recvtypes[],
-                                                 MPIR_Comm * comm,
-                                                 MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op, int root,
-                                              MPIR_Comm * comm_ptr,
-                                              MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                                      const int *recvcounts,
-                                                      MPI_Datatype datatype, MPI_Op op,
-                                                      MPIR_Comm * comm_ptr,
-                                                      MPIR_Errflag_t *
-                                                      errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_reduce_scatter_block(const void *sendbuf,
-                                                            void *recvbuf, int recvcount,
-                                                            MPI_Datatype datatype, MPI_Op op,
-                                                            MPIR_Comm * comm_ptr,
-                                                            MPIR_Errflag_t *
-                                                            errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_scan(const void *sendbuf, void *recvbuf, int count,
-                                            MPI_Datatype datatype, MPI_Op op,
-                                            MPIR_Comm * comm,
-                                            MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op,
-                                              MPIR_Comm * comm,
-                                              MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype,
-                                                          void *recvbuf, int recvcount,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPIR_Errflag_t *
-                                                          errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const int *displs,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm,
-                                                           MPIR_Errflag_t *
-                                                           errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_alltoallv(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const int *sdispls,
-                                                          MPI_Datatype sendtype,
-                                                          void *recvbuf,
-                                                          const int *recvcounts,
-                                                          const int *rdispls,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPIR_Errflag_t *
-                                                          errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_alltoallw(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const MPI_Aint * sdispls,
-                                                          const MPI_Datatype * sendtypes,
-                                                          void *recvbuf,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_gatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int *recvcounts, const int *displs,
+                                                   MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm,
+                                                   MPIR_Errflag_t *
+                                                   errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_alltoall(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm,
+                                                    MPIR_Errflag_t *
+                                                    errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls, MPI_Datatype sendtype,
+                                                     void *recvbuf, const int *recvcounts,
+                                                     const int *rdispls, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm,
+                                                     MPIR_Errflag_t *
+                                                     errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_alltoallw(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls,
+                                                     const MPI_Datatype sendtypes[], void *recvbuf,
+                                                     const int *recvcounts, const int *rdispls,
+                                                     const MPI_Datatype recvtypes[],
+                                                     MPIR_Comm * comm,
+                                                     MPIR_Errflag_t *
+                                                     errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op, int root,
+                                                  MPIR_Comm * comm_ptr,
+                                                  MPIR_Errflag_t *
+                                                  errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
                                                           const int *recvcounts,
-                                                          const MPI_Aint * rdispls,
-                                                          const MPI_Datatype * recvtypes,
-                                                          MPIR_Comm * comm,
+                                                          MPI_Datatype datatype, MPI_Op op,
+                                                          MPIR_Comm * comm_ptr,
                                                           MPIR_Errflag_t *
                                                           errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         int recvcount, MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm,
-                                                         MPIR_Errflag_t *
-                                                         errflag) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf, int recvcount,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm,
-                                                           MPI_Request *
-                                                           req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_allgatherv(const void *sendbuf,
-                                                            int sendcount,
-                                                            MPI_Datatype sendtype,
-                                                            void *recvbuf,
-                                                            const int *recvcounts,
-                                                            const int *displs,
-                                                            MPI_Datatype recvtype,
-                                                            MPIR_Comm * comm,
-                                                            MPI_Request *
-                                                            req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype,
-                                                          void *recvbuf, int recvcount,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPI_Request *
-                                                          req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_alltoallv(const void *sendbuf,
-                                                           const int *sendcounts,
-                                                           const int *sdispls,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const int *rdispls,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm,
-                                                           MPI_Request *
-                                                           req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_alltoallw(const void *sendbuf,
-                                                           const int *sendcounts,
-                                                           const MPI_Aint * sdispls,
-                                                           const MPI_Datatype * sendtypes,
-                                                           void *recvbuf,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                                int recvcount,
+                                                                MPI_Datatype datatype, MPI_Op op,
+                                                                MPIR_Comm * comm_ptr,
+                                                                MPIR_Errflag_t *
+                                                                errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                                MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
+                                                MPIR_Errflag_t * errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op,
+                                                  MPIR_Comm * comm,
+                                                  MPIR_Errflag_t *
+                                                  errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              int recvcount, MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPIR_Errflag_t *
+                                                              errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                               MPI_Datatype sendtype, void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const int *displs,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm,
+                                                               MPIR_Errflag_t *
+                                                               errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_alltoallv(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const int *sdispls,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const int *rdispls,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPIR_Errflag_t *
+                                                              errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_alltoallw(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const MPI_Aint * sdispls,
+                                                              const MPI_Datatype * sendtypes,
+                                                              void *recvbuf, const int *recvcounts,
+                                                              const MPI_Aint * rdispls,
+                                                              const MPI_Datatype * recvtypes,
+                                                              MPIR_Comm * comm,
+                                                              MPIR_Errflag_t *
+                                                              errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             int recvcount, MPI_Datatype recvtype,
+                                                             MPIR_Comm * comm,
+                                                             MPIR_Errflag_t *
+                                                             errflag) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                               MPI_Datatype sendtype, void *recvbuf,
+                                                               int recvcount, MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm,
+                                                               MPI_Request *
+                                                               req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                                MPI_Datatype sendtype,
+                                                                void *recvbuf,
+                                                                const int *recvcounts,
+                                                                const int *displs,
+                                                                MPI_Datatype recvtype,
+                                                                MPIR_Comm * comm,
+                                                                MPI_Request *
+                                                                req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype, void *recvbuf,
+                                                              int recvcount, MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPI_Request *
+                                                              req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_alltoallv(const void *sendbuf,
+                                                               const int *sendcounts,
+                                                               const int *sdispls,
+                                                               MPI_Datatype sendtype, void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const int *rdispls,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm,
+                                                               MPI_Request *
+                                                               req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_alltoallw(const void *sendbuf,
+                                                               const int *sendcounts,
+                                                               const MPI_Aint * sdispls,
+                                                               const MPI_Datatype * sendtypes,
+                                                               void *recvbuf, const int *recvcounts,
+                                                               const MPI_Aint * rdispls,
+                                                               const MPI_Datatype * recvtypes,
+                                                               MPIR_Comm * comm,
+                                                               MPI_Request *
+                                                               req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ibarrier(MPIR_Comm * comm,
+                                                    MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                                  int root, MPIR_Comm * comm,
+                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iallgather(const void *sendbuf, int sendcount,
+                                                      MPI_Datatype sendtype, void *recvbuf,
+                                                      int recvcount, MPI_Datatype recvtype,
+                                                      MPIR_Comm * comm,
+                                                      MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                                       MPI_Datatype sendtype, void *recvbuf,
+                                                       const int *recvcounts, const int *displs,
+                                                       MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                       MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                                      MPI_Datatype datatype, MPI_Op op,
+                                                      MPIR_Comm * comm,
+                                                      MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ialltoall(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     int recvcount, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm,
+                                                     MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                                      const int *sdispls, MPI_Datatype sendtype,
+                                                      void *recvbuf, const int *recvcounts,
+                                                      const int *rdispls, MPI_Datatype recvtype,
+                                                      MPIR_Comm * comm,
+                                                      MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                                      const int *sdispls,
+                                                      const MPI_Datatype sendtypes[], void *recvbuf,
+                                                      const int *recvcounts, const int *rdispls,
+                                                      const MPI_Datatype recvtypes[],
+                                                      MPIR_Comm * comm,
+                                                      MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                                   MPI_Datatype datatype, MPI_Op op,
+                                                   MPIR_Comm * comm,
+                                                   MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_igather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm,
+                                                   MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_igatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int *recvcounts, const int *displs,
+                                                    MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm,
+                                                    MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                                 int recvcount,
+                                                                 MPI_Datatype datatype, MPI_Op op,
+                                                                 MPIR_Comm * comm,
+                                                                 MPI_Request *
+                                                                 req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
                                                            const int *recvcounts,
-                                                           const MPI_Aint * rdispls,
-                                                           const MPI_Datatype * recvtypes,
+                                                           MPI_Datatype datatype, MPI_Op op,
                                                            MPIR_Comm * comm,
                                                            MPI_Request *
                                                            req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ibarrier(MPIR_Comm * comm,
-                                                MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                              int root, MPIR_Comm * comm,
-                                              MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iallgather(const void *sendbuf, int sendcount,
-                                                  MPI_Datatype sendtype, void *recvbuf,
-                                                  int recvcount, MPI_Datatype recvtype,
-                                                  MPIR_Comm * comm,
-                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iallgatherv(const void *sendbuf, int sendcount,
-                                                   MPI_Datatype sendtype, void *recvbuf,
-                                                   const int *recvcounts, const int *displs,
-                                                   MPI_Datatype recvtype, MPIR_Comm * comm,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                                   MPI_Datatype datatype, MPI_Op op, int root,
+                                                   MPIR_Comm * comm_ptr,
                                                    MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iallreduce(const void *sendbuf, void *recvbuf,
-                                                  int count, MPI_Datatype datatype, MPI_Op op,
-                                                  MPIR_Comm * comm,
-                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ialltoall(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 int recvcount, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm,
-                                                 MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                                  const int *sdispls, MPI_Datatype sendtype,
-                                                  void *recvbuf, const int *recvcounts,
-                                                  const int *rdispls, MPI_Datatype recvtype,
-                                                  MPIR_Comm * comm,
-                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                                  const int *sdispls,
-                                                  const MPI_Datatype sendtypes[],
-                                                  void *recvbuf, const int *recvcounts,
-                                                  const int *rdispls,
-                                                  const MPI_Datatype recvtypes[],
-                                                  MPIR_Comm * comm,
-                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                               MPI_Datatype datatype, MPI_Op op,
-                                               MPIR_Comm * comm,
-                                               MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_igather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm,
-                                               MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_igatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int *recvcounts, const int *displs,
-                                                MPI_Datatype recvtype, int root,
-                                                MPIR_Comm * comm,
-                                                MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ireduce_scatter_block(const void *sendbuf,
-                                                             void *recvbuf, int recvcount,
-                                                             MPI_Datatype datatype, MPI_Op op,
-                                                             MPIR_Comm * comm,
-                                                             MPI_Request *
-                                                             req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                                       const int *recvcounts,
-                                                       MPI_Datatype datatype, MPI_Op op,
-                                                       MPIR_Comm * comm,
-                                                       MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                               MPI_Datatype datatype, MPI_Op op, int root,
-                                               MPIR_Comm * comm_ptr,
-                                               MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                             MPI_Datatype datatype, MPI_Op op,
-                                             MPIR_Comm * comm,
-                                             MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iscatter(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                int root, MPIR_Comm * comm,
-                                                MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                                 const int *displs, MPI_Datatype sendtype,
-                                                 void *recvbuf, int recvcount,
-                                                 MPI_Datatype recvtype, int root,
-                                                 MPIR_Comm * comm_ptr,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                                  MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm,
+                                                    MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                                     const int *displs, MPI_Datatype sendtype,
+                                                     void *recvbuf, int recvcount,
+                                                     MPI_Datatype recvtype, int root,
+                                                     MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req) MPL_STATIC_INLINE_SUFFIX;
 
 #endif /* SHM_PROTOTYPES_H_INCLUDED */
diff --git a/src/mpid/ch4/shm/include/shm_impl.h b/src/mpid/ch4/shm/include/shm_impl.h
index b563ae2..ad23af0 100644
--- a/src/mpid/ch4/shm/include/shm_impl.h
+++ b/src/mpid/ch4/shm/include/shm_impl.h
@@ -15,14 +15,14 @@
 #ifndef SHM_DIRECT
 #ifndef SHM_DISABLE_INLINES
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_init(int rank, int size)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_init(int rank, int size)
 {
-    return MPIDI_SHM_func->init(rank, size);
+    return MPIDI_SHM_func->mpi_init(rank, size);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_finalize(void)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_finalize(void)
 {
-    return MPIDI_SHM_func->finalize();
+    return MPIDI_SHM_func->mpi_finalize();
 };
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_progress(int blocking)
@@ -39,33 +39,33 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_reg_handler(int handler_id,
     return MPIDI_SHM_func->am_reg_handler(handler_id, origin_handler_fn, target_handler_fn);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_connect(const char *port_name, MPIR_Info * info,
-                                                    int root, MPIR_Comm * comm,
-                                                    MPIR_Comm ** newcomm_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_connect(const char *port_name, MPIR_Info * info,
+                                                        int root, MPIR_Comm * comm,
+                                                        MPIR_Comm ** newcomm_ptr)
 {
-    return MPIDI_SHM_func->comm_connect(port_name, info, root, comm, newcomm_ptr);
+    return MPIDI_SHM_func->mpi_comm_connect(port_name, info, root, comm, newcomm_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_disconnect(MPIR_Comm * comm_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
-    return MPIDI_SHM_func->comm_disconnect(comm_ptr);
+    return MPIDI_SHM_func->mpi_comm_disconnect(comm_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_open_port(MPIR_Info * info_ptr, char *port_name)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
-    return MPIDI_SHM_func->open_port(info_ptr, port_name);
+    return MPIDI_SHM_func->mpi_open_port(info_ptr, port_name);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_close_port(const char *port_name)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_close_port(const char *port_name)
 {
-    return MPIDI_SHM_func->close_port(port_name);
+    return MPIDI_SHM_func->mpi_close_port(port_name);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_accept(const char *port_name, MPIR_Info * info,
-                                                   int root, MPIR_Comm * comm,
-                                                   MPIR_Comm ** newcomm_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_accept(const char *port_name, MPIR_Info * info,
+                                                       int root, MPIR_Comm * comm,
+                                                       MPIR_Comm ** newcomm_ptr)
 {
-    return MPIDI_SHM_func->comm_accept(port_name, info, root, comm, newcomm_ptr);
+    return MPIDI_SHM_func->mpi_comm_accept(port_name, info, root, comm, newcomm_ptr);
 };
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_am_send_hdr(int rank, MPIR_Comm * comm,
@@ -204,34 +204,34 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_create_intercomm_from_lpids(MPIR_Comm * n
     return MPIDI_SHM_func->create_intercomm_from_lpids(newcomm_ptr, size, lpids);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_create_hook(MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
-    return MPIDI_SHM_func->comm_create_hook(comm);
+    return MPIDI_SHM_func->mpi_comm_create_hook(comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_comm_free_hook(MPIR_Comm * comm)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
-    return MPIDI_SHM_func->comm_free_hook(comm);
+    return MPIDI_SHM_func->mpi_comm_free_hook(comm);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_type_create_hook(MPIR_Datatype * type)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_type_create_hook(MPIR_Datatype * type)
 {
-    return MPIDI_SHM_func->type_create_hook(type);
+    return MPIDI_SHM_func->mpi_type_create_hook(type);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_type_free_hook(MPIR_Datatype * type)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_type_free_hook(MPIR_Datatype * type)
 {
-    return MPIDI_SHM_func->type_free_hook(type);
+    return MPIDI_SHM_func->mpi_type_free_hook(type);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_op_create_hook(MPIR_Op * op)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_op_create_hook(MPIR_Op * op)
 {
-    return MPIDI_SHM_func->op_create_hook(op);
+    return MPIDI_SHM_func->mpi_op_create_hook(op);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_op_free_hook(MPIR_Op * op)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_op_free_hook(MPIR_Op * op)
 {
-    return MPIDI_SHM_func->op_free_hook(op);
+    return MPIDI_SHM_func->mpi_op_free_hook(op);
 };
 
 MPL_STATIC_INLINE_PREFIX void MPIDI_SHM_am_request_init(MPIR_Request * req)
@@ -244,817 +244,827 @@ MPL_STATIC_INLINE_PREFIX void MPIDI_SHM_am_request_finalize(MPIR_Request * req)
     return MPIDI_SHM_func->am_request_finalize(req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_send(const void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_send(const void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->send(buf, count, datatype, rank, tag, comm, context_offset,
-                                       request);
+    return MPIDI_SHM_native_func->mpi_send(buf, count, datatype, rank, tag, comm, context_offset,
+                                           request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ssend(const void *buf, int count,
-                                             MPI_Datatype datatype, int rank, int tag,
-                                             MPIR_Comm * comm, int context_offset,
-                                             MPIR_Request ** request)
-{
-    return MPIDI_SHM_native_func->ssend(buf, count, datatype, rank, tag, comm, context_offset,
-                                        request);
-};
-
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_startall(int count, MPIR_Request * requests[])
-{
-    return MPIDI_SHM_native_func->startall(count, requests);
-};
-
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_send_init(const void *buf, int count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ssend(const void *buf, int count,
                                                  MPI_Datatype datatype, int rank, int tag,
                                                  MPIR_Comm * comm, int context_offset,
                                                  MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->send_init(buf, count, datatype, rank, tag, comm, context_offset,
+    return MPIDI_SHM_native_func->mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset,
                                             request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ssend_init(const void *buf, int count,
-                                                  MPI_Datatype datatype, int rank, int tag,
-                                                  MPIR_Comm * comm, int context_offset,
-                                                  MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_startall(int count, MPIR_Request * requests[])
 {
-    return MPIDI_SHM_native_func->ssend_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                             request);
+    return MPIDI_SHM_native_func->mpi_startall(count, requests);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rsend_init(const void *buf, int count,
-                                                  MPI_Datatype datatype, int rank, int tag,
-                                                  MPIR_Comm * comm, int context_offset,
-                                                  MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_send_init(const void *buf, int count,
+                                                     MPI_Datatype datatype, int rank, int tag,
+                                                     MPIR_Comm * comm, int context_offset,
+                                                     MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->rsend_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                             request);
+    return MPIDI_SHM_native_func->mpi_send_init(buf, count, datatype, rank, tag, comm,
+                                                context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_bsend_init(const void *buf, int count,
-                                                  MPI_Datatype datatype, int rank, int tag,
-                                                  MPIR_Comm * comm, int context_offset,
-                                                  MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ssend_init(const void *buf, int count,
+                                                      MPI_Datatype datatype, int rank, int tag,
+                                                      MPIR_Comm * comm, int context_offset,
+                                                      MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->bsend_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                             request);
+    return MPIDI_SHM_native_func->mpi_ssend_init(buf, count, datatype, rank, tag, comm,
+                                                 context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_isend(const void *buf, int count,
-                                             MPI_Datatype datatype, int rank, int tag,
-                                             MPIR_Comm * comm, int context_offset,
-                                             MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rsend_init(const void *buf, int count,
+                                                      MPI_Datatype datatype, int rank, int tag,
+                                                      MPIR_Comm * comm, int context_offset,
+                                                      MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->isend(buf, count, datatype, rank, tag, comm, context_offset,
-                                        request);
+    return MPIDI_SHM_native_func->mpi_rsend_init(buf, count, datatype, rank, tag, comm,
+                                                 context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_issend(const void *buf, int count,
-                                              MPI_Datatype datatype, int rank, int tag,
-                                              MPIR_Comm * comm, int context_offset,
-                                              MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_bsend_init(const void *buf, int count,
+                                                      MPI_Datatype datatype, int rank, int tag,
+                                                      MPIR_Comm * comm, int context_offset,
+                                                      MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->issend(buf, count, datatype, rank, tag, comm, context_offset,
-                                         request);
+    return MPIDI_SHM_native_func->mpi_bsend_init(buf, count, datatype, rank, tag, comm,
+                                                 context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_cancel_send(MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_isend(const void *buf, int count,
+                                                 MPI_Datatype datatype, int rank, int tag,
+                                                 MPIR_Comm * comm, int context_offset,
+                                                 MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->cancel_send(sreq);
+    return MPIDI_SHM_native_func->mpi_isend(buf, count, datatype, rank, tag, comm, context_offset,
+                                            request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_recv_init(void *buf, int count, MPI_Datatype datatype,
-                                                 int rank, int tag, MPIR_Comm * comm,
-                                                 int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_issend(const void *buf, int count,
+                                                  MPI_Datatype datatype, int rank, int tag,
+                                                  MPIR_Comm * comm, int context_offset,
+                                                  MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->recv_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                            request);
+    return MPIDI_SHM_native_func->mpi_issend(buf, count, datatype, rank, tag, comm, context_offset,
+                                             request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_recv(void *buf, int count, MPI_Datatype datatype,
-                                            int rank, int tag, MPIR_Comm * comm,
-                                            int context_offset, MPI_Status * status,
-                                            MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_cancel_send(MPIR_Request * sreq)
 {
-    return MPIDI_SHM_native_func->recv(buf, count, datatype, rank, tag, comm, context_offset,
-                                       status, request);
+    return MPIDI_SHM_native_func->mpi_cancel_send(sreq);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_irecv(void *buf, int count, MPI_Datatype datatype,
-                                             int rank, int tag, MPIR_Comm * comm,
-                                             int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
+                                                     int rank, int tag, MPIR_Comm * comm,
+                                                     int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->irecv(buf, count, datatype, rank, tag, comm, context_offset,
-                                        request);
+    return MPIDI_SHM_native_func->mpi_recv_init(buf, count, datatype, rank, tag, comm,
+                                                context_offset, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_imrecv(void *buf, int count, MPI_Datatype datatype,
-                                              MPIR_Request * message, MPIR_Request ** rreqp)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_recv(void *buf, int count, MPI_Datatype datatype,
+                                                int rank, int tag, MPIR_Comm * comm,
+                                                int context_offset, MPI_Status * status,
+                                                MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->imrecv(buf, count, datatype, message, rreqp);
+    return MPIDI_SHM_native_func->mpi_recv(buf, count, datatype, rank, tag, comm, context_offset,
+                                           status, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_cancel_recv(MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_irecv(void *buf, int count, MPI_Datatype datatype,
+                                                 int rank, int tag, MPIR_Comm * comm,
+                                                 int context_offset, MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->cancel_recv(rreq);
+    return MPIDI_SHM_native_func->mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset,
+                                            request);
 };
 
-MPL_STATIC_INLINE_PREFIX void *MPIDI_SHM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_imrecv(void *buf, int count, MPI_Datatype datatype,
+                                                  MPIR_Request * message, MPIR_Request ** rreqp)
 {
-    return MPIDI_SHM_native_func->alloc_mem(size, info_ptr);
+    return MPIDI_SHM_native_func->mpi_imrecv(buf, count, datatype, message, rreqp);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_free_mem(void *ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_cancel_recv(MPIR_Request * rreq)
 {
-    return MPIDI_SHM_native_func->free_mem(ptr);
+    return MPIDI_SHM_native_func->mpi_cancel_recv(rreq);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_improbe(int source, int tag, MPIR_Comm * comm,
-                                               int context_offset, int *flag,
-                                               MPIR_Request ** message, MPI_Status * status)
+MPL_STATIC_INLINE_PREFIX void *MPIDI_SHM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
-    return MPIDI_SHM_native_func->improbe(source, tag, comm, context_offset, flag, message, status);
+    return MPIDI_SHM_native_func->mpi_alloc_mem(size, info_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iprobe(int source, int tag, MPIR_Comm * comm,
-                                              int context_offset, int *flag, MPI_Status * status)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_free_mem(void *ptr)
 {
-    return MPIDI_SHM_native_func->iprobe(source, tag, comm, context_offset, flag, status);
+    return MPIDI_SHM_native_func->mpi_free_mem(ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_improbe(int source, int tag, MPIR_Comm * comm,
+                                                   int context_offset, int *flag,
+                                                   MPIR_Request ** message, MPI_Status * status)
 {
-    return MPIDI_SHM_native_func->win_set_info(win, info);
+    return MPIDI_SHM_native_func->mpi_improbe(source, tag, comm, context_offset, flag, message,
+                                              status);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_shared_query(MPIR_Win * win, int rank,
-                                                        MPI_Aint * size, int *disp_unit,
-                                                        void *baseptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iprobe(int source, int tag, MPIR_Comm * comm,
+                                                  int context_offset, int *flag,
+                                                  MPI_Status * status)
 {
-    return MPIDI_SHM_native_func->win_shared_query(win, rank, size, disp_unit, baseptr);
+    return MPIDI_SHM_native_func->mpi_iprobe(source, tag, comm, context_offset, flag, status);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_put(const void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
-    return MPIDI_SHM_native_func->put(origin_addr, origin_count, origin_datatype, target_rank,
-                                      target_disp, target_count, target_datatype, win);
+    return MPIDI_SHM_native_func->mpi_win_set_info(win, info);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_shared_query(MPIR_Win * win, int rank,
+                                                            MPI_Aint * size, int *disp_unit,
+                                                            void *baseptr)
 {
-    return MPIDI_SHM_native_func->win_start(group, assert, win);
+    return MPIDI_SHM_native_func->mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_complete(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_put(const void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_complete(win);
+    return MPIDI_SHM_native_func->mpi_put(origin_addr, origin_count, origin_datatype, target_rank,
+                                          target_disp, target_count, target_datatype, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_post(group, assert, win);
+    return MPIDI_SHM_native_func->mpi_win_start(group, assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_wait(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_complete(MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_wait(win);
+    return MPIDI_SHM_native_func->mpi_win_complete(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_test(MPIR_Win * win, int *flag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_test(win, flag);
+    return MPIDI_SHM_native_func->mpi_win_post(group, assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_wait(MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_lock(lock_type, rank, assert, win);
+    return MPIDI_SHM_native_func->mpi_win_wait(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_unlock(int rank, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_test(MPIR_Win * win, int *flag)
 {
-    return MPIDI_SHM_native_func->win_unlock(rank, win);
+    return MPIDI_SHM_native_func->mpi_win_test(win, flag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_lock(int lock_type, int rank, int assert,
+                                                    MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_get_info(win, info_p_p);
+    return MPIDI_SHM_native_func->mpi_win_lock(lock_type, rank, assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_get(void *origin_addr, int origin_count,
-                                           MPI_Datatype origin_datatype, int target_rank,
-                                           MPI_Aint target_disp, int target_count,
-                                           MPI_Datatype target_datatype, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->get(origin_addr, origin_count, origin_datatype, target_rank,
-                                      target_disp, target_count, target_datatype, win);
+    return MPIDI_SHM_native_func->mpi_win_unlock(rank, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_free(MPIR_Win ** win_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
-    return MPIDI_SHM_native_func->win_free(win_ptr);
+    return MPIDI_SHM_native_func->mpi_win_get_info(win, info_p_p);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_fence(int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_get(void *origin_addr, int origin_count,
+                                               MPI_Datatype origin_datatype, int target_rank,
+                                               MPI_Aint target_disp, int target_count,
+                                               MPI_Datatype target_datatype, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_fence(assert, win);
+    return MPIDI_SHM_native_func->mpi_get(origin_addr, origin_count, origin_datatype, target_rank,
+                                          target_disp, target_count, target_datatype, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_create(void *base, MPI_Aint length, int disp_unit,
-                                                  MPIR_Info * info, MPIR_Comm * comm_ptr,
-                                                  MPIR_Win ** win_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_free(MPIR_Win ** win_ptr)
 {
-    return MPIDI_SHM_native_func->win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    return MPIDI_SHM_native_func->mpi_win_free(win_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_accumulate(const void *origin_addr, int origin_count,
-                                                  MPI_Datatype origin_datatype,
-                                                  int target_rank, MPI_Aint target_disp,
-                                                  int target_count,
-                                                  MPI_Datatype target_datatype, MPI_Op op,
-                                                  MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_fence(int assert, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->accumulate(origin_addr, origin_count, origin_datatype,
-                                             target_rank, target_disp, target_count,
-                                             target_datatype, op, win);
+    return MPIDI_SHM_native_func->mpi_win_fence(assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_create(void *base, MPI_Aint length, int disp_unit,
+                                                      MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                                      MPIR_Win ** win_ptr)
 {
-    return MPIDI_SHM_native_func->win_attach(win, base, size);
+    return MPIDI_SHM_native_func->mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_allocate_shared(MPI_Aint size, int disp_unit,
-                                                           MPIR_Info * info_ptr,
-                                                           MPIR_Comm * comm_ptr,
-                                                           void **base_ptr, MPIR_Win ** win_ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_accumulate(const void *origin_addr, int origin_count,
+                                                      MPI_Datatype origin_datatype,
+                                                      int target_rank, MPI_Aint target_disp,
+                                                      int target_count,
+                                                      MPI_Datatype target_datatype, MPI_Op op,
+                                                      MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, base_ptr,
-                                                      win_ptr);
+    return MPIDI_SHM_native_func->mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                                 target_rank, target_disp, target_count,
+                                                 target_datatype, op, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rput(const void *origin_addr, int origin_count,
-                                            MPI_Datatype origin_datatype, int target_rank,
-                                            MPI_Aint target_disp, int target_count,
-                                            MPI_Datatype target_datatype, MPIR_Win * win,
-                                            MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
-    return MPIDI_SHM_native_func->rput(origin_addr, origin_count, origin_datatype, target_rank,
-                                       target_disp, target_count, target_datatype, win, request);
+    return MPIDI_SHM_native_func->mpi_win_attach(win, base, size);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush_local(int rank, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_allocate_shared(MPI_Aint size, int disp_unit,
+                                                               MPIR_Info * info_ptr,
+                                                               MPIR_Comm * comm_ptr,
+                                                               void **base_ptr, MPIR_Win ** win_ptr)
 {
-    return MPIDI_SHM_native_func->win_flush_local(rank, win);
+    return MPIDI_SHM_native_func->mpi_win_allocate_shared(size, disp_unit, info_ptr, comm_ptr,
+                                                          base_ptr, win_ptr);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_detach(MPIR_Win * win, const void *base)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rput(const void *origin_addr, int origin_count,
+                                                MPI_Datatype origin_datatype, int target_rank,
+                                                MPI_Aint target_disp, int target_count,
+                                                MPI_Datatype target_datatype, MPIR_Win * win,
+                                                MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->win_detach(win, base);
+    return MPIDI_SHM_native_func->mpi_rput(origin_addr, origin_count, origin_datatype, target_rank,
+                                           target_disp, target_count, target_datatype, win,
+                                           request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_compare_and_swap(const void *origin_addr,
-                                                        const void *compare_addr,
-                                                        void *result_addr,
-                                                        MPI_Datatype datatype,
-                                                        int target_rank, MPI_Aint target_disp,
-                                                        MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush_local(int rank, MPIR_Win * win)
+{
+    return MPIDI_SHM_native_func->mpi_win_flush_local(rank, win);
+};
+
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
-    return MPIDI_SHM_native_func->compare_and_swap(origin_addr, compare_addr, result_addr, datatype,
-                                                   target_rank, target_disp, win);
+    return MPIDI_SHM_native_func->mpi_win_detach(win, base);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_raccumulate(const void *origin_addr, int origin_count,
-                                                   MPI_Datatype origin_datatype,
-                                                   int target_rank, MPI_Aint target_disp,
-                                                   int target_count,
-                                                   MPI_Datatype target_datatype, MPI_Op op,
-                                                   MPIR_Win * win, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_compare_and_swap(const void *origin_addr,
+                                                            const void *compare_addr,
+                                                            void *result_addr,
+                                                            MPI_Datatype datatype,
+                                                            int target_rank, MPI_Aint target_disp,
+                                                            MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->raccumulate(origin_addr, origin_count, origin_datatype,
-                                              target_rank, target_disp, target_count,
-                                              target_datatype, op, win, request);
+    return MPIDI_SHM_native_func->mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                                       datatype, target_rank, target_disp, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rget_accumulate(const void *origin_addr,
-                                                       int origin_count,
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_raccumulate(const void *origin_addr, int origin_count,
                                                        MPI_Datatype origin_datatype,
-                                                       void *result_addr, int result_count,
-                                                       MPI_Datatype result_datatype,
                                                        int target_rank, MPI_Aint target_disp,
                                                        int target_count,
-                                                       MPI_Datatype target_datatype,
-                                                       MPI_Op op, MPIR_Win * win,
-                                                       MPIR_Request ** request)
+                                                       MPI_Datatype target_datatype, MPI_Op op,
+                                                       MPIR_Win * win, MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                                  result_addr, result_count, result_datatype,
+    return MPIDI_SHM_native_func->mpi_raccumulate(origin_addr, origin_count, origin_datatype,
                                                   target_rank, target_disp, target_count,
                                                   target_datatype, op, win, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_fetch_and_op(const void *origin_addr,
-                                                    void *result_addr, MPI_Datatype datatype,
-                                                    int target_rank, MPI_Aint target_disp,
-                                                    MPI_Op op, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rget_accumulate(const void *origin_addr,
+                                                           int origin_count,
+                                                           MPI_Datatype origin_datatype,
+                                                           void *result_addr, int result_count,
+                                                           MPI_Datatype result_datatype,
+                                                           int target_rank, MPI_Aint target_disp,
+                                                           int target_count,
+                                                           MPI_Datatype target_datatype,
+                                                           MPI_Op op, MPIR_Win * win,
+                                                           MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->fetch_and_op(origin_addr, result_addr, datatype, target_rank,
-                                               target_disp, op, win);
+    return MPIDI_SHM_native_func->mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                                      result_addr, result_count, result_datatype,
+                                                      target_rank, target_disp, target_count,
+                                                      target_datatype, op, win, request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_allocate(MPI_Aint size, int disp_unit,
-                                                    MPIR_Info * info, MPIR_Comm * comm,
-                                                    void *baseptr, MPIR_Win ** win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_fetch_and_op(const void *origin_addr,
+                                                        void *result_addr, MPI_Datatype datatype,
+                                                        int target_rank, MPI_Aint target_disp,
+                                                        MPI_Op op, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_allocate(size, disp_unit, info, comm, baseptr, win);
+    return MPIDI_SHM_native_func->mpi_fetch_and_op(origin_addr, result_addr, datatype, target_rank,
+                                                   target_disp, op, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush(int rank, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_allocate(MPI_Aint size, int disp_unit,
+                                                        MPIR_Info * info, MPIR_Comm * comm,
+                                                        void *baseptr, MPIR_Win ** win)
 {
-    return MPIDI_SHM_native_func->win_flush(rank, win);
+    return MPIDI_SHM_native_func->mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush_local_all(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush(int rank, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_flush_local_all(win);
+    return MPIDI_SHM_native_func->mpi_win_flush(rank, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_unlock_all(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush_local_all(MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_unlock_all(win);
+    return MPIDI_SHM_native_func->mpi_win_flush_local_all(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
-                                                          MPIR_Win ** win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_unlock_all(MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_create_dynamic(info, comm, win);
+    return MPIDI_SHM_native_func->mpi_win_unlock_all(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_rget(void *origin_addr, int origin_count,
-                                            MPI_Datatype origin_datatype, int target_rank,
-                                            MPI_Aint target_disp, int target_count,
-                                            MPI_Datatype target_datatype, MPIR_Win * win,
-                                            MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                              MPIR_Win ** win)
 {
-    return MPIDI_SHM_native_func->rget(origin_addr, origin_count, origin_datatype, target_rank,
-                                       target_disp, target_count, target_datatype, win, request);
+    return MPIDI_SHM_native_func->mpi_win_create_dynamic(info, comm, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_sync(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_rget(void *origin_addr, int origin_count,
+                                                MPI_Datatype origin_datatype, int target_rank,
+                                                MPI_Aint target_disp, int target_count,
+                                                MPI_Datatype target_datatype, MPIR_Win * win,
+                                                MPIR_Request ** request)
 {
-    return MPIDI_SHM_native_func->win_sync(win);
+    return MPIDI_SHM_native_func->mpi_rget(origin_addr, origin_count, origin_datatype, target_rank,
+                                           target_disp, target_count, target_datatype, win,
+                                           request);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_flush_all(MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_sync(MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_flush_all(win);
+    return MPIDI_SHM_native_func->mpi_win_sync(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_get_accumulate(const void *origin_addr,
-                                                      int origin_count,
-                                                      MPI_Datatype origin_datatype,
-                                                      void *result_addr, int result_count,
-                                                      MPI_Datatype result_datatype,
-                                                      int target_rank, MPI_Aint target_disp,
-                                                      int target_count,
-                                                      MPI_Datatype target_datatype, MPI_Op op,
-                                                      MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_flush_all(MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->get_accumulate(origin_addr, origin_count, origin_datatype,
-                                                 result_addr, result_count, result_datatype,
-                                                 target_rank, target_disp, target_count,
-                                                 target_datatype, op, win);
+    return MPIDI_SHM_native_func->mpi_win_flush_all(win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_win_lock_all(int assert, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_get_accumulate(const void *origin_addr,
+                                                          int origin_count,
+                                                          MPI_Datatype origin_datatype,
+                                                          void *result_addr, int result_count,
+                                                          MPI_Datatype result_datatype,
+                                                          int target_rank, MPI_Aint target_disp,
+                                                          int target_count,
+                                                          MPI_Datatype target_datatype, MPI_Op op,
+                                                          MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->win_lock_all(assert, win);
+    return MPIDI_SHM_native_func->mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                                     result_addr, result_count, result_datatype,
+                                                     target_rank, target_disp, target_count,
+                                                     target_datatype, op, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_barrier(MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
-    return MPIDI_SHM_native_func->barrier(comm, errflag);
+    return MPIDI_SHM_native_func->mpi_win_lock_all(assert, win);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                             int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_barrier(MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->bcast(buffer, count, datatype, root, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_barrier(comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_allreduce(const void *sendbuf, void *recvbuf,
-                                                 int count, MPI_Datatype datatype, MPI_Op op,
-                                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                                 int root, MPIR_Comm * comm,
+                                                 MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->allreduce(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_bcast(buffer, count, datatype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_allgather(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 int recvcount, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_allreduce(const void *sendbuf, void *recvbuf,
+                                                     int count, MPI_Datatype datatype, MPI_Op op,
+                                                     MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                            recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_allreduce(sendbuf, recvbuf, count, datatype, op, comm,
+                                                errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_allgatherv(const void *sendbuf, int sendcount,
-                                                  MPI_Datatype sendtype, void *recvbuf,
-                                                  const int *recvcounts, const int *displs,
-                                                  MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                  MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_allgather(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     int recvcount, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
-                                             displs, recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                                recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_scatter(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_allgatherv(const void *sendbuf, int sendcount,
+                                                      MPI_Datatype sendtype, void *recvbuf,
+                                                      const int *recvcounts, const int *displs,
+                                                      MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                      MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                          recvtype, root, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                                 displs, recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_scatterv(const void *sendbuf, const int *sendcounts,
-                                                const int *displs, MPI_Datatype sendtype,
-                                                void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, int root,
-                                                MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_scatter(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
-                                           recvcount, recvtype, root, comm_ptr, errflag);
+    return MPIDI_SHM_native_func->mpi_scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                              recvtype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_gather(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf,
-                                              int recvcount, MPI_Datatype recvtype, int root,
-                                              MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                                    const int *displs, MPI_Datatype sendtype,
+                                                    void *recvbuf, int recvcount,
+                                                    MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->gather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
-                                         root, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
+                                               recvcount, recvtype, root, comm_ptr, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_gatherv(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               const int *recvcounts, const int *displs,
-                                               MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_gather(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype, int root,
+                                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs,
-                                          recvtype, root, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_gather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                             recvtype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_alltoall(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_gatherv(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   const int *recvcounts, const int *displs,
+                                                   MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                           recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                              displs, recvtype, root, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls, MPI_Datatype sendtype,
-                                                 void *recvbuf, const int *recvcounts,
-                                                 const int *rdispls, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_alltoall(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->alltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
-                                            recvcounts, rdispls, recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                               recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_alltoallw(const void *sendbuf, const int *sendcounts,
-                                                 const int *sdispls,
-                                                 const MPI_Datatype sendtypes[],
-                                                 void *recvbuf, const int *recvcounts,
-                                                 const int *rdispls,
-                                                 const MPI_Datatype recvtypes[],
-                                                 MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls, MPI_Datatype sendtype,
+                                                     void *recvbuf, const int *recvcounts,
+                                                     const int *rdispls, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->alltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
-                                            recvcounts, rdispls, recvtypes, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_alltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
+                                                recvcounts, rdispls, recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op, int root,
-                                              MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_alltoallw(const void *sendbuf, const int *sendcounts,
+                                                     const int *sdispls,
+                                                     const MPI_Datatype sendtypes[],
+                                                     void *recvbuf, const int *recvcounts,
+                                                     const int *rdispls,
+                                                     const MPI_Datatype recvtypes[],
+                                                     MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->reduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
-                                         errflag);
+    return MPIDI_SHM_native_func->mpi_alltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
+                                                recvcounts, rdispls, recvtypes, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                                      const int *recvcounts,
-                                                      MPI_Datatype datatype, MPI_Op op,
-                                                      MPIR_Comm * comm_ptr,
-                                                      MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op, int root,
+                                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->reduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op,
-                                                 comm_ptr, errflag);
+    return MPIDI_SHM_native_func->mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
+                                             errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_reduce_scatter_block(const void *sendbuf,
-                                                            void *recvbuf, int recvcount,
-                                                            MPI_Datatype datatype, MPI_Op op,
-                                                            MPIR_Comm * comm_ptr,
-                                                            MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                                          const int *recvcounts,
+                                                          MPI_Datatype datatype, MPI_Op op,
+                                                          MPIR_Comm * comm_ptr,
+                                                          MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op,
-                                                       comm_ptr, errflag);
+    return MPIDI_SHM_native_func->mpi_reduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op,
+                                                     comm_ptr, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_scan(const void *sendbuf, void *recvbuf, int count,
-                                            MPI_Datatype datatype, MPI_Op op,
-                                            MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_reduce_scatter_block(const void *sendbuf,
+                                                                void *recvbuf, int recvcount,
+                                                                MPI_Datatype datatype, MPI_Op op,
+                                                                MPIR_Comm * comm_ptr,
+                                                                MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->scan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_reduce_scatter_block(sendbuf, recvbuf, recvcount, datatype,
+                                                           op, comm_ptr, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                              MPI_Datatype datatype, MPI_Op op,
-                                              MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                                MPI_Datatype datatype, MPI_Op op,
+                                                MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->exscan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_scan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype,
-                                                          void *recvbuf, int recvcount,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                                  MPI_Datatype datatype, MPI_Op op,
+                                                  MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->neighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
-                                                     recvcount, recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_exscan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const int *displs,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm,
-                                                           MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype,
+                                                              void *recvbuf, int recvcount,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->neighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
-                                                      recvcounts, displs, recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_neighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
+                                                         recvcount, recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_alltoallv(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const int *sdispls,
-                                                          MPI_Datatype sendtype,
-                                                          void *recvbuf,
-                                                          const int *recvcounts,
-                                                          const int *rdispls,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm,
-                                                          MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                               MPI_Datatype sendtype,
+                                                               void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const int *displs,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm,
+                                                               MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->neighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
-                                                     recvbuf, recvcounts, rdispls, recvtype, comm,
-                                                     errflag);
+    return MPIDI_SHM_native_func->mpi_neighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                                          recvcounts, displs, recvtype, comm,
+                                                          errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_alltoallw(const void *sendbuf,
-                                                          const int *sendcounts,
-                                                          const MPI_Aint * sdispls,
-                                                          const MPI_Datatype * sendtypes,
-                                                          void *recvbuf,
-                                                          const int *recvcounts,
-                                                          const MPI_Aint * rdispls,
-                                                          const MPI_Datatype * recvtypes,
-                                                          MPIR_Comm * comm,
-                                                          MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_alltoallv(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const int *sdispls,
+                                                              MPI_Datatype sendtype,
+                                                              void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const int *rdispls,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm,
+                                                              MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->neighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
-                                                     recvbuf, recvcounts, rdispls, recvtypes, comm,
-                                                     errflag);
+    return MPIDI_SHM_native_func->mpi_neighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
+                                                         recvbuf, recvcounts, rdispls, recvtype,
+                                                         comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                                         MPI_Datatype sendtype, void *recvbuf,
-                                                         int recvcount, MPI_Datatype recvtype,
-                                                         MPIR_Comm * comm, MPIR_Errflag_t * errflag)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_alltoallw(const void *sendbuf,
+                                                              const int *sendcounts,
+                                                              const MPI_Aint * sdispls,
+                                                              const MPI_Datatype * sendtypes,
+                                                              void *recvbuf,
+                                                              const int *recvcounts,
+                                                              const MPI_Aint * rdispls,
+                                                              const MPI_Datatype * recvtypes,
+                                                              MPIR_Comm * comm,
+                                                              MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->neighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
-                                                    recvcount, recvtype, comm, errflag);
+    return MPIDI_SHM_native_func->mpi_neighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
+                                                         recvbuf, recvcounts, rdispls, recvtypes,
+                                                         comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf, int recvcount,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                             MPI_Datatype sendtype, void *recvbuf,
+                                                             int recvcount, MPI_Datatype recvtype,
+                                                             MPIR_Comm * comm,
+                                                             MPIR_Errflag_t * errflag)
 {
-    return MPIDI_SHM_native_func->ineighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
-                                                      recvcount, recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_neighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
+                                                        recvcount, recvtype, comm, errflag);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_allgatherv(const void *sendbuf,
-                                                            int sendcount,
-                                                            MPI_Datatype sendtype,
-                                                            void *recvbuf,
-                                                            const int *recvcounts,
-                                                            const int *displs,
-                                                            MPI_Datatype recvtype,
-                                                            MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                               MPI_Datatype sendtype,
+                                                               void *recvbuf, int recvcount,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ineighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
-                                                       recvcounts, displs, recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_ineighbor_allgather(sendbuf, sendcount, sendtype, recvbuf,
+                                                          recvcount, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                                          MPI_Datatype sendtype,
-                                                          void *recvbuf, int recvcount,
-                                                          MPI_Datatype recvtype,
-                                                          MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_allgatherv(const void *sendbuf,
+                                                                int sendcount,
+                                                                MPI_Datatype sendtype,
+                                                                void *recvbuf,
+                                                                const int *recvcounts,
+                                                                const int *displs,
+                                                                MPI_Datatype recvtype,
+                                                                MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ineighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
-                                                     recvcount, recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_ineighbor_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                                           recvcounts, displs, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_alltoallv(const void *sendbuf,
-                                                           const int *sendcounts,
-                                                           const int *sdispls,
-                                                           MPI_Datatype sendtype,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const int *rdispls,
-                                                           MPI_Datatype recvtype,
-                                                           MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                              MPI_Datatype sendtype,
+                                                              void *recvbuf, int recvcount,
+                                                              MPI_Datatype recvtype,
+                                                              MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ineighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
-                                                      recvbuf, recvcounts, rdispls, recvtype, comm,
-                                                      req);
+    return MPIDI_SHM_native_func->mpi_ineighbor_alltoall(sendbuf, sendcount, sendtype, recvbuf,
+                                                         recvcount, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ineighbor_alltoallw(const void *sendbuf,
-                                                           const int *sendcounts,
-                                                           const MPI_Aint * sdispls,
-                                                           const MPI_Datatype * sendtypes,
-                                                           void *recvbuf,
-                                                           const int *recvcounts,
-                                                           const MPI_Aint * rdispls,
-                                                           const MPI_Datatype * recvtypes,
-                                                           MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_alltoallv(const void *sendbuf,
+                                                               const int *sendcounts,
+                                                               const int *sdispls,
+                                                               MPI_Datatype sendtype,
+                                                               void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const int *rdispls,
+                                                               MPI_Datatype recvtype,
+                                                               MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ineighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
-                                                      recvbuf, recvcounts, rdispls, recvtypes, comm,
-                                                      req);
+    return MPIDI_SHM_native_func->mpi_ineighbor_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
+                                                          recvbuf, recvcounts, rdispls, recvtype,
+                                                          comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ibarrier(MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ineighbor_alltoallw(const void *sendbuf,
+                                                               const int *sendcounts,
+                                                               const MPI_Aint * sdispls,
+                                                               const MPI_Datatype * sendtypes,
+                                                               void *recvbuf,
+                                                               const int *recvcounts,
+                                                               const MPI_Aint * rdispls,
+                                                               const MPI_Datatype * recvtypes,
+                                                               MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ibarrier(comm, req);
+    return MPIDI_SHM_native_func->mpi_ineighbor_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
+                                                          recvbuf, recvcounts, rdispls, recvtypes,
+                                                          comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                              int root, MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ibarrier(MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ibcast(buffer, count, datatype, root, comm, req);
+    return MPIDI_SHM_native_func->mpi_ibarrier(comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iallgather(const void *sendbuf, int sendcount,
-                                                  MPI_Datatype sendtype, void *recvbuf,
-                                                  int recvcount, MPI_Datatype recvtype,
-                                                  MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                                  int root, MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                             recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_ibcast(buffer, count, datatype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iallgatherv(const void *sendbuf, int sendcount,
-                                                   MPI_Datatype sendtype, void *recvbuf,
-                                                   const int *recvcounts, const int *displs,
-                                                   MPI_Datatype recvtype, MPIR_Comm * comm,
-                                                   MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iallgather(const void *sendbuf, int sendcount,
+                                                      MPI_Datatype sendtype, void *recvbuf,
+                                                      int recvcount, MPI_Datatype recvtype,
+                                                      MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
-                                              displs, recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                                 recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iallreduce(const void *sendbuf, void *recvbuf,
-                                                  int count, MPI_Datatype datatype, MPI_Op op,
-                                                  MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                                       MPI_Datatype sendtype, void *recvbuf,
+                                                       const int *recvcounts, const int *displs,
+                                                       MPI_Datatype recvtype, MPIR_Comm * comm,
+                                                       MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iallreduce(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_SHM_native_func->mpi_iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                                  displs, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ialltoall(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 int recvcount, MPI_Datatype recvtype,
-                                                 MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iallreduce(const void *sendbuf, void *recvbuf,
+                                                      int count, MPI_Datatype datatype, MPI_Op op,
+                                                      MPIR_Comm * comm, MPI_Request * req)
+{
+    return MPIDI_SHM_native_func->mpi_iallreduce(sendbuf, recvbuf, count, datatype, op, comm, req);
+};
+
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ialltoall(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     int recvcount, MPI_Datatype recvtype,
+                                                     MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                            recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                                recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                                  const int *sdispls, MPI_Datatype sendtype,
-                                                  void *recvbuf, const int *recvcounts,
-                                                  const int *rdispls, MPI_Datatype recvtype,
-                                                  MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                                      const int *sdispls, MPI_Datatype sendtype,
+                                                      void *recvbuf, const int *recvcounts,
+                                                      const int *rdispls, MPI_Datatype recvtype,
+                                                      MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ialltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
-                                             recvcounts, rdispls, recvtype, comm, req);
+    return MPIDI_SHM_native_func->mpi_ialltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf,
+                                                 recvcounts, rdispls, recvtype, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                                  const int *sdispls,
-                                                  const MPI_Datatype sendtypes[],
-                                                  void *recvbuf, const int *recvcounts,
-                                                  const int *rdispls,
-                                                  const MPI_Datatype recvtypes[],
-                                                  MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                                      const int *sdispls,
+                                                      const MPI_Datatype sendtypes[],
+                                                      void *recvbuf, const int *recvcounts,
+                                                      const int *rdispls,
+                                                      const MPI_Datatype recvtypes[],
+                                                      MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ialltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
-                                             recvcounts, rdispls, recvtypes, comm, req);
+    return MPIDI_SHM_native_func->mpi_ialltoallw(sendbuf, sendcounts, sdispls, sendtypes, recvbuf,
+                                                 recvcounts, rdispls, recvtypes, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                               MPI_Datatype datatype, MPI_Op op,
-                                               MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                                   MPI_Datatype datatype, MPI_Op op,
+                                                   MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iexscan(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_SHM_native_func->mpi_iexscan(sendbuf, recvbuf, count, datatype, op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_igather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype, int root,
-                                               MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_igather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype, int root,
+                                                   MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->igather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                          recvtype, root, comm, req);
+    return MPIDI_SHM_native_func->mpi_igather(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                              recvtype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_igatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int *recvcounts, const int *displs,
-                                                MPI_Datatype recvtype, int root,
-                                                MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_igatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int *recvcounts, const int *displs,
+                                                    MPI_Datatype recvtype, int root,
+                                                    MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
-                                           displs, recvtype, root, comm, req);
+    return MPIDI_SHM_native_func->mpi_igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts,
+                                               displs, recvtype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ireduce_scatter_block(const void *sendbuf,
-                                                             void *recvbuf, int recvcount,
-                                                             MPI_Datatype datatype, MPI_Op op,
-                                                             MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ireduce_scatter_block(const void *sendbuf,
+                                                                 void *recvbuf, int recvcount,
+                                                                 MPI_Datatype datatype, MPI_Op op,
+                                                                 MPIR_Comm * comm,
+                                                                 MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op,
-                                                        comm, req);
+    return MPIDI_SHM_native_func->mpi_ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype,
+                                                            op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                                       const int *recvcounts,
-                                                       MPI_Datatype datatype, MPI_Op op,
-                                                       MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                                           const int *recvcounts,
+                                                           MPI_Datatype datatype, MPI_Op op,
+                                                           MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm,
-                                                  req);
+    return MPIDI_SHM_native_func->mpi_ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op,
+                                                      comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                               MPI_Datatype datatype, MPI_Op op, int root,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                                   MPI_Datatype datatype, MPI_Op op, int root,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->ireduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
-                                          req);
+    return MPIDI_SHM_native_func->mpi_ireduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr,
+                                              req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                             MPI_Datatype datatype, MPI_Op op,
-                                             MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                                 MPI_Datatype datatype, MPI_Op op,
+                                                 MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iscan(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_SHM_native_func->mpi_iscan(sendbuf, recvbuf, count, datatype, op, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iscatter(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                int recvcount, MPI_Datatype recvtype,
-                                                int root, MPIR_Comm * comm, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    int root, MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
-                                           recvtype, root, comm, req);
+    return MPIDI_SHM_native_func->mpi_iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount,
+                                               recvtype, root, comm, req);
 };
 
-MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                                 const int *displs, MPI_Datatype sendtype,
-                                                 void *recvbuf, int recvcount,
-                                                 MPI_Datatype recvtype, int root,
-                                                 MPIR_Comm * comm_ptr, MPI_Request * req)
+MPL_STATIC_INLINE_PREFIX int MPIDI_SHM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                                     const int *displs, MPI_Datatype sendtype,
+                                                     void *recvbuf, int recvcount,
+                                                     MPI_Datatype recvtype, int root,
+                                                     MPIR_Comm * comm_ptr, MPI_Request * req)
 {
-    return MPIDI_SHM_native_func->iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
-                                            recvcount, recvtype, root, comm_ptr, req);
+    return MPIDI_SHM_native_func->mpi_iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
+                                                recvcount, recvtype, root, comm_ptr, req);
 };
 
 #endif /* SHM_DISABLE_INLINES  */
diff --git a/src/mpid/ch4/shm/posix/func_table.c b/src/mpid/ch4/shm/posix/func_table.c
index 3319e3a..24e9709 100644
--- a/src/mpid/ch4/shm/posix/func_table.c
+++ b/src/mpid/ch4/shm/posix/func_table.c
@@ -13,15 +13,15 @@
 #include <mpidimpl.h>
 #include "shm_direct.h"
 MPIDI_SHM_funcs_t MPIDI_SHM_posix_funcs = {
-    MPIDI_SHM_init,
-    MPIDI_SHM_finalize,
+    MPIDI_SHM_mpi_init,
+    MPIDI_SHM_mpi_finalize,
     MPIDI_SHM_progress,
     MPIDI_SHM_am_reg_handler,
-    MPIDI_SHM_comm_connect,
-    MPIDI_SHM_comm_disconnect,
-    MPIDI_SHM_open_port,
-    MPIDI_SHM_close_port,
-    MPIDI_SHM_comm_accept,
+    MPIDI_SHM_mpi_comm_connect,
+    MPIDI_SHM_mpi_comm_disconnect,
+    MPIDI_SHM_mpi_open_port,
+    MPIDI_SHM_mpi_close_port,
+    MPIDI_SHM_mpi_comm_accept,
     MPIDI_SHM_am_send_hdr,
     MPIDI_SHM_am_isend,
     MPIDI_SHM_inject_am,
@@ -41,113 +41,113 @@ MPIDI_SHM_funcs_t MPIDI_SHM_posix_funcs = {
     MPIDI_SHM_getallincomm,
     MPIDI_SHM_gpid_tolpidarray,
     MPIDI_SHM_create_intercomm_from_lpids,
-    MPIDI_SHM_comm_create_hook,
-    MPIDI_SHM_comm_free_hook,
-    MPIDI_SHM_type_create_hook,
-    MPIDI_SHM_type_free_hook,
-    MPIDI_SHM_op_create_hook,
-    MPIDI_SHM_op_free_hook,
+    MPIDI_SHM_mpi_comm_create_hook,
+    MPIDI_SHM_mpi_comm_free_hook,
+    MPIDI_SHM_mpi_type_create_hook,
+    MPIDI_SHM_mpi_type_free_hook,
+    MPIDI_SHM_mpi_op_create_hook,
+    MPIDI_SHM_mpi_op_free_hook,
     MPIDI_SHM_am_request_init,
 };
 
 MPIDI_SHM_native_funcs_t MPIDI_SHM_native_posix_funcs = {
-    MPIDI_SHM_send,
-    MPIDI_SHM_ssend,
-    MPIDI_SHM_startall,
-    MPIDI_SHM_send_init,
-    MPIDI_SHM_ssend_init,
-    MPIDI_SHM_rsend_init,
-    MPIDI_SHM_bsend_init,
-    MPIDI_SHM_isend,
-    MPIDI_SHM_issend,
-    MPIDI_SHM_cancel_send,
-    MPIDI_SHM_recv_init,
-    MPIDI_SHM_recv,
-    MPIDI_SHM_irecv,
-    MPIDI_SHM_imrecv,
-    MPIDI_SHM_cancel_recv,
-    MPIDI_SHM_alloc_mem,
-    MPIDI_SHM_free_mem,
-    MPIDI_SHM_improbe,
-    MPIDI_SHM_iprobe,
-    MPIDI_SHM_win_set_info,
-    MPIDI_SHM_win_shared_query,
-    MPIDI_SHM_put,
-    MPIDI_SHM_win_start,
-    MPIDI_SHM_win_complete,
-    MPIDI_SHM_win_post,
-    MPIDI_SHM_win_wait,
-    MPIDI_SHM_win_test,
-    MPIDI_SHM_win_lock,
-    MPIDI_SHM_win_unlock,
-    MPIDI_SHM_win_get_info,
-    MPIDI_SHM_get,
-    MPIDI_SHM_win_free,
-    MPIDI_SHM_win_fence,
-    MPIDI_SHM_win_create,
-    MPIDI_SHM_accumulate,
-    MPIDI_SHM_win_attach,
-    MPIDI_SHM_win_allocate_shared,
-    MPIDI_SHM_rput,
-    MPIDI_SHM_win_flush_local,
-    MPIDI_SHM_win_detach,
-    MPIDI_SHM_compare_and_swap,
-    MPIDI_SHM_raccumulate,
-    MPIDI_SHM_rget_accumulate,
-    MPIDI_SHM_fetch_and_op,
-    MPIDI_SHM_win_allocate,
-    MPIDI_SHM_win_flush,
-    MPIDI_SHM_win_flush_local_all,
-    MPIDI_SHM_win_unlock_all,
-    MPIDI_SHM_win_create_dynamic,
-    MPIDI_SHM_rget,
-    MPIDI_SHM_win_sync,
-    MPIDI_SHM_win_flush_all,
-    MPIDI_SHM_get_accumulate,
-    MPIDI_SHM_win_lock_all,
-    MPIDI_SHM_barrier,
-    MPIDI_SHM_bcast,
-    MPIDI_SHM_allreduce,
-    MPIDI_SHM_allgather,
-    MPIDI_SHM_allgatherv,
-    MPIDI_SHM_scatter,
-    MPIDI_SHM_scatterv,
-    MPIDI_SHM_gather,
-    MPIDI_SHM_gatherv,
-    MPIDI_SHM_alltoall,
-    MPIDI_SHM_alltoallv,
-    MPIDI_SHM_alltoallw,
-    MPIDI_SHM_reduce,
-    MPIDI_SHM_reduce_scatter,
-    MPIDI_SHM_reduce_scatter_block,
-    MPIDI_SHM_scan,
-    MPIDI_SHM_exscan,
-    MPIDI_SHM_neighbor_allgather,
-    MPIDI_SHM_neighbor_allgatherv,
-    MPIDI_SHM_neighbor_alltoall,
-    MPIDI_SHM_neighbor_alltoallv,
-    MPIDI_SHM_neighbor_alltoallw,
-    MPIDI_SHM_ineighbor_allgather,
-    MPIDI_SHM_ineighbor_allgatherv,
-    MPIDI_SHM_ineighbor_alltoall,
-    MPIDI_SHM_ineighbor_alltoallv,
-    MPIDI_SHM_ineighbor_alltoallw,
-    MPIDI_SHM_ibarrier,
-    MPIDI_SHM_ibcast,
-    MPIDI_SHM_iallgather,
-    MPIDI_SHM_iallgatherv,
-    MPIDI_SHM_iallreduce,
-    MPIDI_SHM_ialltoall,
-    MPIDI_SHM_ialltoallv,
-    MPIDI_SHM_ialltoallw,
-    MPIDI_SHM_iexscan,
-    MPIDI_SHM_igather,
-    MPIDI_SHM_igatherv,
-    MPIDI_SHM_ireduce_scatter_block,
-    MPIDI_SHM_ireduce_scatter,
-    MPIDI_SHM_ireduce,
-    MPIDI_SHM_iscan,
-    MPIDI_SHM_iscatter,
-    MPIDI_SHM_iscatterv,
+    MPIDI_SHM_mpi_send,
+    MPIDI_SHM_mpi_ssend,
+    MPIDI_SHM_mpi_startall,
+    MPIDI_SHM_mpi_send_init,
+    MPIDI_SHM_mpi_ssend_init,
+    MPIDI_SHM_mpi_rsend_init,
+    MPIDI_SHM_mpi_bsend_init,
+    MPIDI_SHM_mpi_isend,
+    MPIDI_SHM_mpi_issend,
+    MPIDI_SHM_mpi_cancel_send,
+    MPIDI_SHM_mpi_recv_init,
+    MPIDI_SHM_mpi_recv,
+    MPIDI_SHM_mpi_irecv,
+    MPIDI_SHM_mpi_imrecv,
+    MPIDI_SHM_mpi_cancel_recv,
+    MPIDI_SHM_mpi_alloc_mem,
+    MPIDI_SHM_mpi_free_mem,
+    MPIDI_SHM_mpi_improbe,
+    MPIDI_SHM_mpi_iprobe,
+    MPIDI_SHM_mpi_win_set_info,
+    MPIDI_SHM_mpi_win_shared_query,
+    MPIDI_SHM_mpi_put,
+    MPIDI_SHM_mpi_win_start,
+    MPIDI_SHM_mpi_win_complete,
+    MPIDI_SHM_mpi_win_post,
+    MPIDI_SHM_mpi_win_wait,
+    MPIDI_SHM_mpi_win_test,
+    MPIDI_SHM_mpi_win_lock,
+    MPIDI_SHM_mpi_win_unlock,
+    MPIDI_SHM_mpi_win_get_info,
+    MPIDI_SHM_mpi_get,
+    MPIDI_SHM_mpi_win_free,
+    MPIDI_SHM_mpi_win_fence,
+    MPIDI_SHM_mpi_win_create,
+    MPIDI_SHM_mpi_accumulate,
+    MPIDI_SHM_mpi_win_attach,
+    MPIDI_SHM_mpi_win_allocate_shared,
+    MPIDI_SHM_mpi_rput,
+    MPIDI_SHM_mpi_win_flush_local,
+    MPIDI_SHM_mpi_win_detach,
+    MPIDI_SHM_mpi_compare_and_swap,
+    MPIDI_SHM_mpi_raccumulate,
+    MPIDI_SHM_mpi_rget_accumulate,
+    MPIDI_SHM_mpi_fetch_and_op,
+    MPIDI_SHM_mpi_win_allocate,
+    MPIDI_SHM_mpi_win_flush,
+    MPIDI_SHM_mpi_win_flush_local_all,
+    MPIDI_SHM_mpi_win_unlock_all,
+    MPIDI_SHM_mpi_win_create_dynamic,
+    MPIDI_SHM_mpi_rget,
+    MPIDI_SHM_mpi_win_sync,
+    MPIDI_SHM_mpi_win_flush_all,
+    MPIDI_SHM_mpi_get_accumulate,
+    MPIDI_SHM_mpi_win_lock_all,
+    MPIDI_SHM_mpi_barrier,
+    MPIDI_SHM_mpi_bcast,
+    MPIDI_SHM_mpi_allreduce,
+    MPIDI_SHM_mpi_allgather,
+    MPIDI_SHM_mpi_allgatherv,
+    MPIDI_SHM_mpi_scatter,
+    MPIDI_SHM_mpi_scatterv,
+    MPIDI_SHM_mpi_gather,
+    MPIDI_SHM_mpi_gatherv,
+    MPIDI_SHM_mpi_alltoall,
+    MPIDI_SHM_mpi_alltoallv,
+    MPIDI_SHM_mpi_alltoallw,
+    MPIDI_SHM_mpi_reduce,
+    MPIDI_SHM_mpi_reduce_scatter,
+    MPIDI_SHM_mpi_reduce_scatter_block,
+    MPIDI_SHM_mpi_scan,
+    MPIDI_SHM_mpi_exscan,
+    MPIDI_SHM_mpi_neighbor_allgather,
+    MPIDI_SHM_mpi_neighbor_allgatherv,
+    MPIDI_SHM_mpi_neighbor_alltoall,
+    MPIDI_SHM_mpi_neighbor_alltoallv,
+    MPIDI_SHM_mpi_neighbor_alltoallw,
+    MPIDI_SHM_mpi_ineighbor_allgather,
+    MPIDI_SHM_mpi_ineighbor_allgatherv,
+    MPIDI_SHM_mpi_ineighbor_alltoall,
+    MPIDI_SHM_mpi_ineighbor_alltoallv,
+    MPIDI_SHM_mpi_ineighbor_alltoallw,
+    MPIDI_SHM_mpi_ibarrier,
+    MPIDI_SHM_mpi_ibcast,
+    MPIDI_SHM_mpi_iallgather,
+    MPIDI_SHM_mpi_iallgatherv,
+    MPIDI_SHM_mpi_iallreduce,
+    MPIDI_SHM_mpi_ialltoall,
+    MPIDI_SHM_mpi_ialltoallv,
+    MPIDI_SHM_mpi_ialltoallw,
+    MPIDI_SHM_mpi_iexscan,
+    MPIDI_SHM_mpi_igather,
+    MPIDI_SHM_mpi_igatherv,
+    MPIDI_SHM_mpi_ireduce_scatter_block,
+    MPIDI_SHM_mpi_ireduce_scatter,
+    MPIDI_SHM_mpi_ireduce,
+    MPIDI_SHM_mpi_iscan,
+    MPIDI_SHM_mpi_iscatter,
+    MPIDI_SHM_mpi_iscatterv,
 };
 #endif
diff --git a/src/mpid/ch4/shm/posix/posix_coll.h b/src/mpid/ch4/shm/posix/posix_coll.h
index 52e3228..bc754a5 100644
--- a/src/mpid/ch4/shm/posix/posix_coll.h
+++ b/src/mpid/ch4/shm/posix/posix_coll.h
@@ -15,10 +15,10 @@
 #include "ch4_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_barrier
+#define FUNCNAME MPIDI_SHM_mpi_barrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_BARRIER);
@@ -31,11 +31,11 @@ static inline int MPIDI_SHM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errfl
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_bcast
+#define FUNCNAME MPIDI_SHM_mpi_bcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                  int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                      int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_BCAST);
@@ -48,12 +48,12 @@ static inline int MPIDI_SHM_bcast(void *buffer, int count, MPI_Datatype datatype
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_allreduce
+#define FUNCNAME MPIDI_SHM_mpi_allreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                      MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                          MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ALLREDUCE);
@@ -66,12 +66,12 @@ static inline int MPIDI_SHM_allreduce(const void *sendbuf, void *recvbuf, int co
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_allgather
+#define FUNCNAME MPIDI_SHM_mpi_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ALLGATHER);
@@ -85,13 +85,14 @@ static inline int MPIDI_SHM_allgather(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_allgatherv
+#define FUNCNAME MPIDI_SHM_mpi_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                       MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_allgatherv(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ALLGATHERV);
@@ -105,12 +106,12 @@ static inline int MPIDI_SHM_allgatherv(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_gather
+#define FUNCNAME MPIDI_SHM_mpi_gather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_GATHER);
@@ -124,13 +125,13 @@ static inline int MPIDI_SHM_gather(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_gatherv
+#define FUNCNAME MPIDI_SHM_mpi_gatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                        MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_GATHERV);
@@ -144,12 +145,12 @@ static inline int MPIDI_SHM_gatherv(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_scatter
+#define FUNCNAME MPIDI_SHM_mpi_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_SCATTER);
@@ -163,13 +164,13 @@ static inline int MPIDI_SHM_scatter(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_scatterv
+#define FUNCNAME MPIDI_SHM_mpi_scatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_scatterv(const void *sendbuf, const int *sendcounts,
-                                     const int *displs, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_SCATTERV);
@@ -183,12 +184,12 @@ static inline int MPIDI_SHM_scatterv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_alltoall
+#define FUNCNAME MPIDI_SHM_mpi_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ALLTOALL);
@@ -202,14 +203,14 @@ static inline int MPIDI_SHM_alltoall(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_alltoallv
+#define FUNCNAME MPIDI_SHM_mpi_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ALLTOALLV);
@@ -222,14 +223,14 @@ static inline int MPIDI_SHM_alltoallv(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_alltoallw
+#define FUNCNAME MPIDI_SHM_mpi_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_alltoallw(const void *sendbuf, const int sendcounts[],
-                                      const int sdispls[], const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int recvcounts[],
-                                      const int rdispls[], const MPI_Datatype recvtypes[],
-                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_alltoallw(const void *sendbuf, const int sendcounts[],
+                                          const int sdispls[], const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int recvcounts[],
+                                          const int rdispls[], const MPI_Datatype recvtypes[],
+                                          MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ALLTOALLW);
@@ -243,12 +244,12 @@ static inline int MPIDI_SHM_alltoallw(const void *sendbuf, const int sendcounts[
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_reduce
+#define FUNCNAME MPIDI_SHM_mpi_reduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root,
-                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_REDUCE);
@@ -261,13 +262,13 @@ static inline int MPIDI_SHM_reduce(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_reduce_scatter
+#define FUNCNAME MPIDI_SHM_mpi_reduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                           const int recvcounts[], MPI_Datatype datatype,
-                                           MPI_Op op, MPIR_Comm * comm_ptr,
-                                           MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                               const int recvcounts[], MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr,
+                                               MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_REDUCE_SCATTER);
@@ -280,13 +281,13 @@ static inline int MPIDI_SHM_reduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_reduce_scatter_block
+#define FUNCNAME MPIDI_SHM_mpi_reduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                 int recvcount, MPI_Datatype datatype,
-                                                 MPI_Op op, MPIR_Comm * comm_ptr,
-                                                 MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_REDUCE_SCATTER_BLOCK);
@@ -300,12 +301,12 @@ static inline int MPIDI_SHM_reduce_scatter_block(const void *sendbuf, void *recv
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_scan
+#define FUNCNAME MPIDI_SHM_mpi_scan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_scan(const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                 MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                     MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_SCAN);
@@ -318,12 +319,12 @@ static inline int MPIDI_SHM_scan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_exscan
+#define FUNCNAME MPIDI_SHM_mpi_exscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_EXSCAN);
@@ -336,13 +337,13 @@ static inline int MPIDI_SHM_exscan(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_allgather
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                               MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_NEIGHBOR_ALLGATHER);
@@ -357,14 +358,14 @@ static inline int MPIDI_SHM_neighbor_allgather(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_allgatherv
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int recvcounts[], const int displs[],
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int recvcounts[], const int displs[],
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_NEIGHBOR_ALLGATHERV);
@@ -378,13 +379,13 @@ static inline int MPIDI_SHM_neighbor_allgatherv(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_alltoall
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                              MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_NEIGHBOR_ALLTOALL);
@@ -398,14 +399,14 @@ static inline int MPIDI_SHM_neighbor_alltoall(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_alltoallv
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                               const int sdispls[], MPI_Datatype sendtype,
-                                               void *recvbuf, const int recvcounts[],
-                                               const int rdispls[], MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                   const int sdispls[], MPI_Datatype sendtype,
+                                                   void *recvbuf, const int recvcounts[],
+                                                   const int rdispls[], MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_NEIGHBOR_ALLTOALLV);
@@ -419,15 +420,15 @@ static inline int MPIDI_SHM_neighbor_alltoallv(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_alltoallw
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                               const MPI_Aint sdispls[],
-                                               const MPI_Datatype sendtypes[], void *recvbuf,
-                                               const int recvcounts[], const MPI_Aint rdispls[],
-                                               const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr,
-                                               MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                   const MPI_Aint sdispls[],
+                                                   const MPI_Datatype sendtypes[], void *recvbuf,
+                                                   const int recvcounts[], const MPI_Aint rdispls[],
+                                                   const MPI_Datatype recvtypes[],
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_NEIGHBOR_ALLTOALLW);
@@ -441,13 +442,13 @@ static inline int MPIDI_SHM_neighbor_alltoallw(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_allgather
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_INEIGHBOR_ALLGATHER);
@@ -461,14 +462,14 @@ static inline int MPIDI_SHM_ineighbor_allgather(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_allgatherv
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 const int recvcounts[], const int displs[],
-                                                 MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                 MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     const int recvcounts[], const int displs[],
+                                                     MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_INEIGHBOR_ALLGATHERV);
@@ -483,13 +484,13 @@ static inline int MPIDI_SHM_ineighbor_allgatherv(const void *sendbuf, int sendco
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_alltoall
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_INEIGHBOR_ALLTOALL);
@@ -503,14 +504,14 @@ static inline int MPIDI_SHM_ineighbor_alltoall(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_alltoallv
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                                const int sdispls[], MPI_Datatype sendtype,
-                                                void *recvbuf, const int recvcounts[],
-                                                const int rdispls[], MPI_Datatype recvtype,
-                                                MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                    const int sdispls[], MPI_Datatype sendtype,
+                                                    void *recvbuf, const int recvcounts[],
+                                                    const int rdispls[], MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_INEIGHBOR_ALLTOALLV);
@@ -525,15 +526,16 @@ static inline int MPIDI_SHM_ineighbor_alltoallv(const void *sendbuf, const int s
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_alltoallw
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                                const MPI_Aint sdispls[],
-                                                const MPI_Datatype sendtypes[], void *recvbuf,
-                                                const int recvcounts[], const MPI_Aint rdispls[],
-                                                const MPI_Datatype recvtypes[],
-                                                MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                    const MPI_Aint sdispls[],
+                                                    const MPI_Datatype sendtypes[], void *recvbuf,
+                                                    const int recvcounts[],
+                                                    const MPI_Aint rdispls[],
+                                                    const MPI_Datatype recvtypes[],
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_INEIGHBOR_ALLTOALLW);
@@ -548,10 +550,10 @@ static inline int MPIDI_SHM_ineighbor_alltoallw(const void *sendbuf, const int s
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ibarrier
+#define FUNCNAME MPIDI_SHM_mpi_ibarrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IBARRIER);
@@ -564,11 +566,11 @@ static inline int MPIDI_SHM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ibcast
+#define FUNCNAME MPIDI_SHM_mpi_ibcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                   int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                       int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IBCAST);
@@ -581,12 +583,13 @@ static inline int MPIDI_SHM_ibcast(void *buffer, int count, MPI_Datatype datatyp
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iallgather
+#define FUNCNAME MPIDI_SHM_mpi_iallgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iallgather(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf, int recvcount,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IALLGATHER);
@@ -600,13 +603,14 @@ static inline int MPIDI_SHM_iallgather(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iallgatherv
+#define FUNCNAME MPIDI_SHM_mpi_iallgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iallgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                        void *recvbuf, const int *recvcounts, const int *displs,
-                                        MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                        MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                            MPI_Datatype sendtype, void *recvbuf,
+                                            const int *recvcounts, const int *displs,
+                                            MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                            MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IALLGATHERV);
@@ -620,12 +624,12 @@ static inline int MPIDI_SHM_iallgatherv(const void *sendbuf, int sendcount, MPI_
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ialltoall
+#define FUNCNAME MPIDI_SHM_mpi_ialltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IALLTOALL);
@@ -639,14 +643,14 @@ static inline int MPIDI_SHM_ialltoall(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ialltoallv
+#define FUNCNAME MPIDI_SHM_mpi_ialltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                       const int *sdispls, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts,
-                                       const int *rdispls, MPI_Datatype recvtype,
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                           const int *sdispls, MPI_Datatype sendtype,
+                                           void *recvbuf, const int *recvcounts,
+                                           const int *rdispls, MPI_Datatype recvtype,
+                                           MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IALLTOALLV);
@@ -661,14 +665,14 @@ static inline int MPIDI_SHM_ialltoallv(const void *sendbuf, const int *sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ialltoallw
+#define FUNCNAME MPIDI_SHM_mpi_ialltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                       const int *sdispls, const MPI_Datatype sendtypes[],
-                                       void *recvbuf, const int *recvcounts,
-                                       const int *rdispls, const MPI_Datatype recvtypes[],
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                           const int *sdispls, const MPI_Datatype sendtypes[],
+                                           void *recvbuf, const int *recvcounts,
+                                           const int *rdispls, const MPI_Datatype recvtypes[],
+                                           MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IALLTOALLW);
@@ -683,12 +687,12 @@ static inline int MPIDI_SHM_ialltoallw(const void *sendbuf, const int *sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iexscan
+#define FUNCNAME MPIDI_SHM_mpi_iexscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                    MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                        MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IEXSCAN);
@@ -701,12 +705,12 @@ static inline int MPIDI_SHM_iexscan(const void *sendbuf, void *recvbuf, int coun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_igather
+#define FUNCNAME MPIDI_SHM_mpi_igather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IGATHER);
@@ -720,13 +724,13 @@ static inline int MPIDI_SHM_igather(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_igatherv
+#define FUNCNAME MPIDI_SHM_mpi_igatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts, const int *displs,
-                                     MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                     MPI_Request * req)
+static inline int MPIDI_SHM_mpi_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts, const int *displs,
+                                         MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                         MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IGATHERV);
@@ -740,13 +744,13 @@ static inline int MPIDI_SHM_igatherv(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ireduce_scatter_block
+#define FUNCNAME MPIDI_SHM_mpi_ireduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                  int recvcount, MPI_Datatype datatype,
-                                                  MPI_Op op, MPIR_Comm * comm_ptr,
-                                                  MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                      int recvcount, MPI_Datatype datatype,
+                                                      MPI_Op op, MPIR_Comm * comm_ptr,
+                                                      MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IREDUCE_SCATTER_BLOCK);
@@ -760,12 +764,12 @@ static inline int MPIDI_SHM_ireduce_scatter_block(const void *sendbuf, void *rec
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ireduce_scatter
+#define FUNCNAME MPIDI_SHM_mpi_ireduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                            const int recvcounts[], MPI_Datatype datatype,
-                                            MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                                const int recvcounts[], MPI_Datatype datatype,
+                                                MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IREDUCE_SCATTER);
@@ -779,12 +783,12 @@ static inline int MPIDI_SHM_ireduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ireduce
+#define FUNCNAME MPIDI_SHM_mpi_ireduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                    MPI_Datatype datatype, MPI_Op op, int root,
-                                    MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype datatype, MPI_Op op, int root,
+                                        MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IREDUCE);
@@ -797,12 +801,12 @@ static inline int MPIDI_SHM_ireduce(const void *sendbuf, void *recvbuf, int coun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iallreduce
+#define FUNCNAME MPIDI_SHM_mpi_iallreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                       MPI_Datatype datatype, MPI_Op op,
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                           MPI_Datatype datatype, MPI_Op op,
+                                           MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_IALLREDUCE);
@@ -815,12 +819,12 @@ static inline int MPIDI_SHM_iallreduce(const void *sendbuf, void *recvbuf, int c
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iscan
+#define FUNCNAME MPIDI_SHM_mpi_iscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                  MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                      MPI_Request * req)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ISCAN);
@@ -833,13 +837,13 @@ static inline int MPIDI_SHM_iscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iscatter
+#define FUNCNAME MPIDI_SHM_mpi_iscatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iscatter(const void *sendbuf, int sendcount,
-                                     MPI_Datatype sendtype, void *recvbuf,
-                                     int recvcount, MPI_Datatype recvtype,
-                                     int root, MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_SHM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                         MPI_Datatype sendtype, void *recvbuf,
+                                         int recvcount, MPI_Datatype recvtype,
+                                         int root, MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ISCATTER);
@@ -853,14 +857,14 @@ static inline int MPIDI_SHM_iscatter(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iscatterv
+#define FUNCNAME MPIDI_SHM_mpi_iscatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                      const int *displs, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount,
-                                      MPI_Datatype recvtype, int root,
-                                      MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_SHM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                          const int *displs, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount,
+                                          MPI_Datatype recvtype, int root,
+                                          MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_ISCATTERV);
diff --git a/src/mpid/ch4/shm/posix/posix_comm.h b/src/mpid/ch4/shm/posix/posix_comm.h
index 0aaf6ef..f0637c3 100644
--- a/src/mpid/ch4/shm/posix/posix_comm.h
+++ b/src/mpid/ch4/shm/posix/posix_comm.h
@@ -14,10 +14,10 @@
 #include "mpl_utlist.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_comm_create_hook
+#define FUNCNAME MPIDI_SHM_mpi_comm_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_comm_create_hook(MPIR_Comm * comm)
+static inline int MPIDI_SHM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_COMM_CREATE);
@@ -28,10 +28,10 @@ static inline int MPIDI_SHM_comm_create_hook(MPIR_Comm * comm)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_comm_free_hook
+#define FUNCNAME MPIDI_SHM_mpi_comm_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_comm_free_hook(MPIR_Comm * comm)
+static inline int MPIDI_SHM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_COMM_DESTROY);
diff --git a/src/mpid/ch4/shm/posix/posix_init.h b/src/mpid/ch4/shm/posix/posix_init.h
index c07cbb9..1f2754b 100644
--- a/src/mpid/ch4/shm/posix/posix_init.h
+++ b/src/mpid/ch4/shm/posix/posix_init.h
@@ -22,8 +22,8 @@ extern MPIDI_POSIX_mem_region_t MPIDI_POSIX_mem_region;
 extern char *MPIDI_POSIX_asym_base_addr;
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_init)
-static inline int MPIDI_SHM_init(int rank, int size)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_init)
+static inline int MPIDI_SHM_mpi_init(int rank, int size)
 {
     int mpi_errno = MPI_SUCCESS;
     int num_local = 0;
@@ -229,8 +229,8 @@ static inline int MPIDI_SHM_init(int rank, int size)
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_finalize)
-static inline int MPIDI_SHM_finalize(void)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_finalize)
+static inline int MPIDI_SHM_mpi_finalize(void)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_SHM_FINALIZE);
@@ -264,13 +264,13 @@ static inline int MPIDI_SHM_finalize(void)
     goto fn_exit;
 }
 
-static inline void *MPIDI_SHM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+static inline void *MPIDI_SHM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
     MPIR_Assert(0);
     return NULL;
 }
 
-static inline int MPIDI_SHM_free_mem(void *ptr)
+static inline int MPIDI_SHM_mpi_free_mem(void *ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
@@ -322,10 +322,10 @@ static inline int MPIDI_SHM_create_intercomm_from_lpids(MPIR_Comm * newcomm_ptr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_type_create_hook
+#define FUNCNAME MPIDI_SHM_mpi_type_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_type_create_hook(MPIR_Datatype * type)
+static inline int MPIDI_SHM_mpi_type_create_hook(MPIR_Datatype * type)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_TYPE_CREATE_HOOK);
@@ -336,10 +336,10 @@ static inline int MPIDI_SHM_type_create_hook(MPIR_Datatype * type)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_type_free_hook
+#define FUNCNAME MPIDI_SHM_mpi_type_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_type_free_hook(MPIR_Datatype * type)
+static inline int MPIDI_SHM_mpi_type_free_hook(MPIR_Datatype * type)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_TYPE_FREE_HOOK);
@@ -350,10 +350,10 @@ static inline int MPIDI_SHM_type_free_hook(MPIR_Datatype * type)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_op_create_hook
+#define FUNCNAME MPIDI_SHM_mpi_op_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_op_create_hook(MPIR_Op * op)
+static inline int MPIDI_SHM_mpi_op_create_hook(MPIR_Op * op)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_OP_CREATE_HOOK);
@@ -364,10 +364,10 @@ static inline int MPIDI_SHM_op_create_hook(MPIR_Op * op)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_op_free_hook
+#define FUNCNAME MPIDI_SHM_mpi_op_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_op_free_hook(MPIR_Op * op)
+static inline int MPIDI_SHM_mpi_op_free_hook(MPIR_Op * op)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_OP_FREE_HOOK);
diff --git a/src/mpid/ch4/shm/posix/posix_probe.h b/src/mpid/ch4/shm/posix/posix_probe.h
index 5fe8e11..757c5e4 100644
--- a/src/mpid/ch4/shm/posix/posix_probe.h
+++ b/src/mpid/ch4/shm/posix/posix_probe.h
@@ -14,11 +14,11 @@
 #include "posix_impl.h"
 
 
-static inline int MPIDI_SHM_improbe(int source,
-                                    int tag,
-                                    MPIR_Comm * comm,
-                                    int context_offset,
-                                    int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_SHM_mpi_improbe(int source,
+                                        int tag,
+                                        MPIR_Comm * comm,
+                                        int context_offset,
+                                        int *flag, MPIR_Request ** message, MPI_Status * status)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *req, *matched_req = NULL;
@@ -94,10 +94,10 @@ static inline int MPIDI_SHM_improbe(int source,
     return mpi_errno;
 }
 
-static inline int MPIDI_SHM_iprobe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_SHM_mpi_iprobe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset, int *flag, MPI_Status * status)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *req, *matched_req = NULL;
diff --git a/src/mpid/ch4/shm/posix/posix_recv.h b/src/mpid/ch4/shm/posix/posix_recv.h
index d51e8e4..a0c43d4 100644
--- a/src/mpid/ch4/shm/posix/posix_recv.h
+++ b/src/mpid/ch4/shm/posix/posix_recv.h
@@ -93,14 +93,15 @@ static inline int MPIDI_POSIX_do_irecv(void *buf,
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_recv)
-static inline int MPIDI_SHM_recv(void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm,
-                                 int context_offset, MPI_Status * status, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_recv)
+static inline int MPIDI_SHM_mpi_recv(void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm,
+                                     int context_offset, MPI_Status * status,
+                                     MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS, dt_contig __attribute__ ((__unused__));
     size_t data_sz __attribute__ ((__unused__));
@@ -118,13 +119,14 @@ static inline int MPIDI_SHM_recv(void *buf,
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_recv)
-static inline int MPIDI_SHM_recv_init(void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_recv)
+static inline int MPIDI_SHM_mpi_recv_init(void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq = NULL;
@@ -150,10 +152,10 @@ static inline int MPIDI_SHM_recv_init(void *buf,
 }
 
 
-static inline int MPIDI_SHM_imrecv(void *buf,
-                                   int count,
-                                   MPI_Datatype datatype,
-                                   MPIR_Request * message, MPIR_Request ** rreqp)
+static inline int MPIDI_SHM_mpi_imrecv(void *buf,
+                                       int count,
+                                       MPI_Datatype datatype,
+                                       MPIR_Request * message, MPIR_Request ** rreqp)
 {
     int mpi_errno = MPI_SUCCESS;
     int dt_contig;
@@ -293,13 +295,13 @@ static inline int MPIDI_SHM_imrecv(void *buf,
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_irecv)
-static inline int MPIDI_SHM_irecv(void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_irecv)
+static inline int MPIDI_SHM_mpi_irecv(void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPIDI_SHM_IRECV);
@@ -314,8 +316,8 @@ static inline int MPIDI_SHM_irecv(void *buf,
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_cancel_recv)
-static inline int MPIDI_SHM_cancel_recv(MPIR_Request * rreq)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_cancel_recv)
+static inline int MPIDI_SHM_mpi_cancel_recv(MPIR_Request * rreq)
 {
     MPID_THREAD_CS_ENTER(POBJ, MPIDI_POSIX_SHM_MUTEX);
     MPIR_Request *req = MPIDI_POSIX_recvq_posted.head;
diff --git a/src/mpid/ch4/shm/posix/posix_rma.h b/src/mpid/ch4/shm/posix/posix_rma.h
index 8ea78bf..321ed73 100644
--- a/src/mpid/ch4/shm/posix/posix_rma.h
+++ b/src/mpid/ch4/shm/posix/posix_rma.h
@@ -13,71 +13,56 @@
 
 #include "posix_impl.h"
 
-static inline int MPIDI_SHM_put(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_put(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_get(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_get(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_rput(const void *origin_addr,
-                                 int origin_count,
-                                 MPI_Datatype origin_datatype,
-                                 int target_rank,
-                                 MPI_Aint target_disp,
-                                 int target_count,
-                                 MPI_Datatype target_datatype,
-                                 MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_rput(const void *origin_addr,
+                                     int origin_count,
+                                     MPI_Datatype origin_datatype,
+                                     int target_rank,
+                                     MPI_Aint target_disp,
+                                     int target_count,
+                                     MPI_Datatype target_datatype,
+                                     MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_compare_and_swap(const void *origin_addr,
-                                             const void *compare_addr,
-                                             void *result_addr,
-                                             MPI_Datatype datatype,
-                                             int target_rank, MPI_Aint target_disp, MPIR_Win * win)
-{
-    MPIR_Assert(0);
-    return MPI_SUCCESS;
-}
-
-static inline int MPIDI_SHM_raccumulate(const void *origin_addr,
-                                        int origin_count,
-                                        MPI_Datatype origin_datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp,
-                                        int target_count,
-                                        MPI_Datatype target_datatype,
-                                        MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_compare_and_swap(const void *origin_addr,
+                                                 const void *compare_addr,
+                                                 void *result_addr,
+                                                 MPI_Datatype datatype,
+                                                 int target_rank, MPI_Aint target_disp,
+                                                 MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_SHM_mpi_raccumulate(const void *origin_addr,
                                             int origin_count,
                                             MPI_Datatype origin_datatype,
-                                            void *result_addr,
-                                            int result_count,
-                                            MPI_Datatype result_datatype,
                                             int target_rank,
                                             MPI_Aint target_disp,
                                             int target_count,
@@ -88,37 +73,66 @@ static inline int MPIDI_SHM_rget_accumulate(const void *origin_addr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_fetch_and_op(const void *origin_addr,
-                                         void *result_addr,
-                                         MPI_Datatype datatype,
-                                         int target_rank,
-                                         MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_rget_accumulate(const void *origin_addr,
+                                                int origin_count,
+                                                MPI_Datatype origin_datatype,
+                                                void *result_addr,
+                                                int result_count,
+                                                MPI_Datatype result_datatype,
+                                                int target_rank,
+                                                MPI_Aint target_disp,
+                                                int target_count,
+                                                MPI_Datatype target_datatype,
+                                                MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+{
+    MPIR_Assert(0);
+    return MPI_SUCCESS;
+}
+
+static inline int MPIDI_SHM_mpi_fetch_and_op(const void *origin_addr,
+                                             void *result_addr,
+                                             MPI_Datatype datatype,
+                                             int target_rank,
+                                             MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_rget(void *origin_addr,
-                                 int origin_count,
-                                 MPI_Datatype origin_datatype,
-                                 int target_rank,
-                                 MPI_Aint target_disp,
-                                 int target_count,
-                                 MPI_Datatype target_datatype,
-                                 MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_rget(void *origin_addr,
+                                     int origin_count,
+                                     MPI_Datatype origin_datatype,
+                                     int target_rank,
+                                     MPI_Aint target_disp,
+                                     int target_count,
+                                     MPI_Datatype target_datatype,
+                                     MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_SHM_mpi_get_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype, MPI_Op op,
+                                               MPIR_Win * win)
+{
+    MPIR_Assert(0);
+    return MPI_SUCCESS;
+}
+
+static inline int MPIDI_SHM_mpi_accumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
@@ -128,16 +142,4 @@ static inline int MPIDI_SHM_get_accumulate(const void *origin_addr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_accumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    MPIR_Assert(0);
-    return MPI_SUCCESS;
-}
-
 #endif /* SHM_POSIX_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/shm/posix/posix_send.h b/src/mpid/ch4/shm/posix/posix_send.h
index afde708..0e9baf2 100644
--- a/src/mpid/ch4/shm/posix/posix_send.h
+++ b/src/mpid/ch4/shm/posix/posix_send.h
@@ -90,12 +90,12 @@ static inline int MPIDI_POSIX_do_isend(const void *buf,
 
 #undef FCNAME
 #define FCNAME DECL_FUNC(MPIDI_SHM_SEND)
-static inline int MPIDI_SHM_send(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_send(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int dt_contig __attribute__ ((__unused__)), mpi_errno = MPI_SUCCESS;
     MPI_Aint dt_true_lb;
@@ -167,12 +167,12 @@ static inline int MPIDI_SHM_irsend(const void *buf,
 
 #undef FCNAME
 #define FCNAME DECL_FUNC(MPIDI_SHM_SSEND)
-static inline int MPIDI_SHM_ssend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_ssend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_SHM_SSEND);
@@ -189,7 +189,7 @@ static inline int MPIDI_SHM_ssend(const void *buf,
 
 #undef FCNAME
 #define FCNAME DECL_FUNC(MPIDI_SHM_STARTALL)
-static inline int MPIDI_SHM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_SHM_mpi_startall(int count, MPIR_Request * requests[])
 {
     int i, mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_SHM_STARTALL);
@@ -264,12 +264,13 @@ static inline int MPIDI_SHM_startall(int count, MPIR_Request * requests[])
 
 #undef FCNAME
 #define FCNAME DECL_FUNC(MPIDI_SHM_SEND_INIT)
-static inline int MPIDI_SHM_send_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_send_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq = NULL;
@@ -297,13 +298,13 @@ static inline int MPIDI_SHM_send_init(const void *buf,
 
 #undef FCNAME
 #define FCNAME DECL_FUNC(MPIDI_SHM_SSEND_INIT)
-static inline int MPIDI_SHM_ssend_init(const void *buf,
-                                       int count,
-                                       MPI_Datatype datatype,
-                                       int rank,
-                                       int tag,
-                                       MPIR_Comm * comm,
-                                       int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_ssend_init(const void *buf,
+                                           int count,
+                                           MPI_Datatype datatype,
+                                           int rank,
+                                           int tag,
+                                           MPIR_Comm * comm,
+                                           int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq = NULL;
@@ -328,13 +329,13 @@ static inline int MPIDI_SHM_ssend_init(const void *buf,
     return mpi_errno;
 }
 
-static inline int MPIDI_SHM_bsend_init(const void *buf,
-                                       int count,
-                                       MPI_Datatype datatype,
-                                       int rank,
-                                       int tag,
-                                       MPIR_Comm * comm,
-                                       int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_bsend_init(const void *buf,
+                                           int count,
+                                           MPI_Datatype datatype,
+                                           int rank,
+                                           int tag,
+                                           MPIR_Comm * comm,
+                                           int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq = NULL;
@@ -359,13 +360,13 @@ static inline int MPIDI_SHM_bsend_init(const void *buf,
     return mpi_errno;
 }
 
-static inline int MPIDI_SHM_rsend_init(const void *buf,
-                                       int count,
-                                       MPI_Datatype datatype,
-                                       int rank,
-                                       int tag,
-                                       MPIR_Comm * comm,
-                                       int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_rsend_init(const void *buf,
+                                           int count,
+                                           MPI_Datatype datatype,
+                                           int rank,
+                                           int tag,
+                                           MPIR_Comm * comm,
+                                           int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq = NULL;
@@ -392,13 +393,13 @@ static inline int MPIDI_SHM_rsend_init(const void *buf,
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_isend)
-static inline int MPIDI_SHM_isend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_isend)
+static inline int MPIDI_SHM_mpi_isend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_SHM_ISEND);
@@ -413,12 +414,13 @@ static inline int MPIDI_SHM_isend(const void *buf,
     return mpi_errno;
 }
 
-static inline int MPIDI_SHM_issend(const void *buf,
-                                   int count,
-                                   MPI_Datatype datatype,
-                                   int rank,
-                                   int tag,
-                                   MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_issend(const void *buf,
+                                       int count,
+                                       MPI_Datatype datatype,
+                                       int rank,
+                                       int tag,
+                                       MPIR_Comm * comm, int context_offset,
+                                       MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_SHM_ISSEND);
@@ -433,7 +435,7 @@ static inline int MPIDI_SHM_issend(const void *buf,
     return mpi_errno;
 }
 
-static inline int MPIDI_SHM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_SHM_mpi_cancel_send(MPIR_Request * sreq)
 {
     MPID_THREAD_CS_ENTER(POBJ, MPIDI_POSIX_SHM_MUTEX);
     MPIR_Request *req = MPIDI_POSIX_sendq.head;
diff --git a/src/mpid/ch4/shm/posix/posix_spawn.h b/src/mpid/ch4/shm/posix/posix_spawn.h
index 2a2e001..e7862d6 100644
--- a/src/mpid/ch4/shm/posix/posix_spawn.h
+++ b/src/mpid/ch4/shm/posix/posix_spawn.h
@@ -13,35 +13,35 @@
 
 #include "posix_impl.h"
 
-static inline int MPIDI_SHM_comm_connect(const char *port_name,
-                                         MPIR_Info * info,
-                                         int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_SHM_mpi_comm_connect(const char *port_name,
+                                             MPIR_Info * info,
+                                             int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_SHM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_SHM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_close_port(const char *port_name)
+static inline int MPIDI_SHM_mpi_close_port(const char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_comm_accept(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_SHM_mpi_comm_accept(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/shm/posix/posix_win.h b/src/mpid/ch4/shm/posix/posix_win.h
index aeac908..2b86292 100644
--- a/src/mpid/ch4/shm/posix/posix_win.h
+++ b/src/mpid/ch4/shm/posix/posix_win.h
@@ -13,169 +13,171 @@
 
 #include "posix_impl.h"
 
-static inline int MPIDI_SHM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_SHM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_complete(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_complete(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_wait(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_wait(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_SHM_mpi_win_test(MPIR_Win * win, int *flag)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_SHM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_SHM_mpi_win_free(MPIR_Win ** win_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_fence(int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_create(void *base,
-                                       MPI_Aint length,
-                                       int disp_unit,
-                                       MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_SHM_mpi_win_create(void *base,
+                                           MPI_Aint length,
+                                           int disp_unit,
+                                           MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                           MPIR_Win ** win_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_SHM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_allocate_shared(MPI_Aint size,
-                                                int disp_unit,
-                                                MPIR_Info * info_ptr,
-                                                MPIR_Comm * comm_ptr,
-                                                void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_SHM_mpi_win_allocate_shared(MPI_Aint size,
+                                                    int disp_unit,
+                                                    MPIR_Info * info_ptr,
+                                                    MPIR_Comm * comm_ptr,
+                                                    void **base_ptr, MPIR_Win ** win_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_SHM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_shared_query(MPIR_Win * win,
-                                             int rank,
-                                             MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_SHM_mpi_win_shared_query(MPIR_Win * win,
+                                                 int rank,
+                                                 MPI_Aint * size, int *disp_unit, void *baseptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_allocate(MPI_Aint size,
-                                         int disp_unit,
-                                         MPIR_Info * info,
-                                         MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
+static inline int MPIDI_SHM_mpi_win_allocate(MPI_Aint size,
+                                             int disp_unit,
+                                             MPIR_Info * info,
+                                             MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush(int rank, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush_local_all(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_unlock_all(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_SHM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                   MPIR_Win ** win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_sync(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_sync(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush_all(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/shm/stubshm/func_table.c b/src/mpid/ch4/shm/stubshm/func_table.c
index 44a89c5..acb9672 100644
--- a/src/mpid/ch4/shm/stubshm/func_table.c
+++ b/src/mpid/ch4/shm/stubshm/func_table.c
@@ -13,15 +13,15 @@
 #include <mpidimpl.h>
 #include "shm_direct.h"
 MPIDI_SHM_funcs_t MPIDI_SHM_stubshm_funcs = {
-    MPIDI_SHM_init,
-    MPIDI_SHM_finalize,
+    MPIDI_SHM_mpi_init,
+    MPIDI_SHM_mpi_finalize,
     MPIDI_SHM_progress,
     MPIDI_SHM_am_reg_handler,
-    MPIDI_SHM_comm_connect,
-    MPIDI_SHM_comm_disconnect,
-    MPIDI_SHM_open_port,
-    MPIDI_SHM_close_port,
-    MPIDI_SHM_comm_accept,
+    MPIDI_SHM_mpi_comm_connect,
+    MPIDI_SHM_mpi_comm_disconnect,
+    MPIDI_SHM_mpi_open_port,
+    MPIDI_SHM_mpi_close_port,
+    MPIDI_SHM_mpi_comm_accept,
     MPIDI_SHM_am_send_hdr,
     MPIDI_SHM_am_isend,
     MPIDI_SHM_inject_am,
@@ -41,113 +41,113 @@ MPIDI_SHM_funcs_t MPIDI_SHM_stubshm_funcs = {
     MPIDI_SHM_getallincomm,
     MPIDI_SHM_gpid_tolpidarray,
     MPIDI_SHM_create_intercomm_from_lpids,
-    MPIDI_SHM_comm_create_hook,
-    MPIDI_SHM_comm_free_hook,
-    MPIDI_SHM_type_create_hook,
-    MPIDI_SHM_type_free_hook,
-    MPIDI_SHM_op_create_hook,
-    MPIDI_SHM_op_free_hook,
+    MPIDI_SHM_mpi_comm_create_hook,
+    MPIDI_SHM_mpi_comm_free_hook,
+    MPIDI_SHM_mpi_type_create_hook,
+    MPIDI_SHM_mpi_type_free_hook,
+    MPIDI_SHM_mpi_op_create_hook,
+    MPIDI_SHM_mpi_op_free_hook,
     MPIDI_SHM_am_request_init,
 };
 
 MPIDI_SHM_native_funcs_t MPIDI_SHM_native_stubshm_funcs = {
-    MPIDI_SHM_send,
-    MPIDI_SHM_ssend,
-    MPIDI_SHM_startall,
-    MPIDI_SHM_send_init,
-    MPIDI_SHM_ssend_init,
-    MPIDI_SHM_rsend_init,
-    MPIDI_SHM_bsend_init,
-    MPIDI_SHM_isend,
-    MPIDI_SHM_issend,
-    MPIDI_SHM_cancel_send,
-    MPIDI_SHM_recv_init,
-    MPIDI_SHM_recv,
-    MPIDI_SHM_irecv,
-    MPIDI_SHM_imrecv,
-    MPIDI_SHM_cancel_recv,
-    MPIDI_SHM_alloc_mem,
-    MPIDI_SHM_free_mem,
-    MPIDI_SHM_improbe,
-    MPIDI_SHM_iprobe,
-    MPIDI_SHM_win_set_info,
-    MPIDI_SHM_win_shared_query,
-    MPIDI_SHM_put,
-    MPIDI_SHM_win_start,
-    MPIDI_SHM_win_complete,
-    MPIDI_SHM_win_post,
-    MPIDI_SHM_win_wait,
-    MPIDI_SHM_win_test,
-    MPIDI_SHM_win_lock,
-    MPIDI_SHM_win_unlock,
-    MPIDI_SHM_win_get_info,
-    MPIDI_SHM_get,
-    MPIDI_SHM_win_free,
-    MPIDI_SHM_win_fence,
-    MPIDI_SHM_win_create,
-    MPIDI_SHM_accumulate,
-    MPIDI_SHM_win_attach,
-    MPIDI_SHM_win_allocate_shared,
-    MPIDI_SHM_rput,
-    MPIDI_SHM_win_flush_local,
-    MPIDI_SHM_win_detach,
-    MPIDI_SHM_compare_and_swap,
-    MPIDI_SHM_raccumulate,
-    MPIDI_SHM_rget_accumulate,
-    MPIDI_SHM_fetch_and_op,
-    MPIDI_SHM_win_allocate,
-    MPIDI_SHM_win_flush,
-    MPIDI_SHM_win_flush_local_all,
-    MPIDI_SHM_win_unlock_all,
-    MPIDI_SHM_win_create_dynamic,
-    MPIDI_SHM_rget,
-    MPIDI_SHM_win_sync,
-    MPIDI_SHM_win_flush_all,
-    MPIDI_SHM_get_accumulate,
-    MPIDI_SHM_win_lock_all,
-    MPIDI_SHM_barrier,
-    MPIDI_SHM_bcast,
-    MPIDI_SHM_allreduce,
-    MPIDI_SHM_allgather,
-    MPIDI_SHM_allgatherv,
-    MPIDI_SHM_scatter,
-    MPIDI_SHM_scatterv,
-    MPIDI_SHM_gather,
-    MPIDI_SHM_gatherv,
-    MPIDI_SHM_alltoall,
-    MPIDI_SHM_alltoallv,
-    MPIDI_SHM_alltoallw,
-    MPIDI_SHM_reduce,
-    MPIDI_SHM_reduce_scatter,
-    MPIDI_SHM_reduce_scatter_block,
-    MPIDI_SHM_scan,
-    MPIDI_SHM_exscan,
-    MPIDI_SHM_neighbor_allgather,
-    MPIDI_SHM_neighbor_allgatherv,
-    MPIDI_SHM_neighbor_alltoall,
-    MPIDI_SHM_neighbor_alltoallv,
-    MPIDI_SHM_neighbor_alltoallw,
-    MPIDI_SHM_ineighbor_allgather,
-    MPIDI_SHM_ineighbor_allgatherv,
-    MPIDI_SHM_ineighbor_alltoall,
-    MPIDI_SHM_ineighbor_alltoallv,
-    MPIDI_SHM_ineighbor_alltoallw,
-    MPIDI_SHM_ibarrier,
-    MPIDI_SHM_ibcast,
-    MPIDI_SHM_iallgather,
-    MPIDI_SHM_iallgatherv,
-    MPIDI_SHM_iallreduce,
-    MPIDI_SHM_ialltoall,
-    MPIDI_SHM_ialltoallv,
-    MPIDI_SHM_ialltoallw,
-    MPIDI_SHM_iexscan,
-    MPIDI_SHM_igather,
-    MPIDI_SHM_igatherv,
-    MPIDI_SHM_ireduce_scatter_block,
-    MPIDI_SHM_ireduce_scatter,
-    MPIDI_SHM_ireduce,
-    MPIDI_SHM_iscan,
-    MPIDI_SHM_iscatter,
-    MPIDI_SHM_iscatterv,
+    MPIDI_SHM_mpi_send,
+    MPIDI_SHM_mpi_ssend,
+    MPIDI_SHM_mpi_startall,
+    MPIDI_SHM_mpi_send_init,
+    MPIDI_SHM_mpi_ssend_init,
+    MPIDI_SHM_mpi_rsend_init,
+    MPIDI_SHM_mpi_bsend_init,
+    MPIDI_SHM_mpi_isend,
+    MPIDI_SHM_mpi_issend,
+    MPIDI_SHM_mpi_cancel_send,
+    MPIDI_SHM_mpi_recv_init,
+    MPIDI_SHM_mpi_recv,
+    MPIDI_SHM_mpi_irecv,
+    MPIDI_SHM_mpi_imrecv,
+    MPIDI_SHM_mpi_cancel_recv,
+    MPIDI_SHM_mpi_alloc_mem,
+    MPIDI_SHM_mpi_free_mem,
+    MPIDI_SHM_mpi_improbe,
+    MPIDI_SHM_mpi_iprobe,
+    MPIDI_SHM_mpi_win_set_info,
+    MPIDI_SHM_mpi_win_shared_query,
+    MPIDI_SHM_mpi_put,
+    MPIDI_SHM_mpi_win_start,
+    MPIDI_SHM_mpi_win_complete,
+    MPIDI_SHM_mpi_win_post,
+    MPIDI_SHM_mpi_win_wait,
+    MPIDI_SHM_mpi_win_test,
+    MPIDI_SHM_mpi_win_lock,
+    MPIDI_SHM_mpi_win_unlock,
+    MPIDI_SHM_mpi_win_get_info,
+    MPIDI_SHM_mpi_get,
+    MPIDI_SHM_mpi_win_free,
+    MPIDI_SHM_mpi_win_fence,
+    MPIDI_SHM_mpi_win_create,
+    MPIDI_SHM_mpi_accumulate,
+    MPIDI_SHM_mpi_win_attach,
+    MPIDI_SHM_mpi_win_allocate_shared,
+    MPIDI_SHM_mpi_rput,
+    MPIDI_SHM_mpi_win_flush_local,
+    MPIDI_SHM_mpi_win_detach,
+    MPIDI_SHM_mpi_compare_and_swap,
+    MPIDI_SHM_mpi_raccumulate,
+    MPIDI_SHM_mpi_rget_accumulate,
+    MPIDI_SHM_mpi_fetch_and_op,
+    MPIDI_SHM_mpi_win_allocate,
+    MPIDI_SHM_mpi_win_flush,
+    MPIDI_SHM_mpi_win_flush_local_all,
+    MPIDI_SHM_mpi_win_unlock_all,
+    MPIDI_SHM_mpi_win_create_dynamic,
+    MPIDI_SHM_mpi_rget,
+    MPIDI_SHM_mpi_win_sync,
+    MPIDI_SHM_mpi_win_flush_all,
+    MPIDI_SHM_mpi_get_accumulate,
+    MPIDI_SHM_mpi_win_lock_all,
+    MPIDI_SHM_mpi_barrier,
+    MPIDI_SHM_mpi_bcast,
+    MPIDI_SHM_mpi_allreduce,
+    MPIDI_SHM_mpi_allgather,
+    MPIDI_SHM_mpi_allgatherv,
+    MPIDI_SHM_mpi_scatter,
+    MPIDI_SHM_mpi_scatterv,
+    MPIDI_SHM_mpi_gather,
+    MPIDI_SHM_mpi_gatherv,
+    MPIDI_SHM_mpi_alltoall,
+    MPIDI_SHM_mpi_alltoallv,
+    MPIDI_SHM_mpi_alltoallw,
+    MPIDI_SHM_mpi_reduce,
+    MPIDI_SHM_mpi_reduce_scatter,
+    MPIDI_SHM_mpi_reduce_scatter_block,
+    MPIDI_SHM_mpi_scan,
+    MPIDI_SHM_mpi_exscan,
+    MPIDI_SHM_mpi_neighbor_allgather,
+    MPIDI_SHM_mpi_neighbor_allgatherv,
+    MPIDI_SHM_mpi_neighbor_alltoall,
+    MPIDI_SHM_mpi_neighbor_alltoallv,
+    MPIDI_SHM_mpi_neighbor_alltoallw,
+    MPIDI_SHM_mpi_ineighbor_allgather,
+    MPIDI_SHM_mpi_ineighbor_allgatherv,
+    MPIDI_SHM_mpi_ineighbor_alltoall,
+    MPIDI_SHM_mpi_ineighbor_alltoallv,
+    MPIDI_SHM_mpi_ineighbor_alltoallw,
+    MPIDI_SHM_mpi_ibarrier,
+    MPIDI_SHM_mpi_ibcast,
+    MPIDI_SHM_mpi_iallgather,
+    MPIDI_SHM_mpi_iallgatherv,
+    MPIDI_SHM_mpi_iallreduce,
+    MPIDI_SHM_mpi_ialltoall,
+    MPIDI_SHM_mpi_ialltoallv,
+    MPIDI_SHM_mpi_ialltoallw,
+    MPIDI_SHM_mpi_iexscan,
+    MPIDI_SHM_mpi_igather,
+    MPIDI_SHM_mpi_igatherv,
+    MPIDI_SHM_mpi_ireduce_scatter_block,
+    MPIDI_SHM_mpi_ireduce_scatter,
+    MPIDI_SHM_mpi_ireduce,
+    MPIDI_SHM_mpi_iscan,
+    MPIDI_SHM_mpi_iscatter,
+    MPIDI_SHM_mpi_iscatterv,
 };
 #endif
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_coll.h b/src/mpid/ch4/shm/stubshm/stubshm_coll.h
index 73ef4e8..3866a4d 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_coll.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_coll.h
@@ -15,10 +15,10 @@
 #include "ch4_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_barrier
+#define FUNCNAME MPIDI_SHM_mpi_barrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -28,11 +28,11 @@ static inline int MPIDI_SHM_barrier(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errfl
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_bcast
+#define FUNCNAME MPIDI_SHM_mpi_bcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_bcast(void *buffer, int count, MPI_Datatype datatype,
-                                  int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_bcast(void *buffer, int count, MPI_Datatype datatype,
+                                      int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -42,12 +42,12 @@ static inline int MPIDI_SHM_bcast(void *buffer, int count, MPI_Datatype datatype
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_allreduce
+#define FUNCNAME MPIDI_SHM_mpi_allreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_allreduce(const void *sendbuf, void *recvbuf, int count,
-                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                      MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_allreduce(const void *sendbuf, void *recvbuf, int count,
+                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                          MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -57,12 +57,12 @@ static inline int MPIDI_SHM_allreduce(const void *sendbuf, void *recvbuf, int co
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_allgather
+#define FUNCNAME MPIDI_SHM_mpi_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -72,13 +72,14 @@ static inline int MPIDI_SHM_allgather(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_allgatherv
+#define FUNCNAME MPIDI_SHM_mpi_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts, const int *displs,
-                                       MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                       MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_allgatherv(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf,
+                                           const int *recvcounts, const int *displs,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -88,12 +89,12 @@ static inline int MPIDI_SHM_allgatherv(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_gather
+#define FUNCNAME MPIDI_SHM_mpi_gather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                   void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                   int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                       int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -103,13 +104,13 @@ static inline int MPIDI_SHM_gather(const void *sendbuf, int sendcount, MPI_Datat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_gatherv
+#define FUNCNAME MPIDI_SHM_mpi_gatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, const int *recvcounts, const int *displs,
-                                    MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                    MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, const int *recvcounts, const int *displs,
+                                        MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                        MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -119,12 +120,12 @@ static inline int MPIDI_SHM_gatherv(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_scatter
+#define FUNCNAME MPIDI_SHM_mpi_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -134,13 +135,13 @@ static inline int MPIDI_SHM_scatter(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_scatterv
+#define FUNCNAME MPIDI_SHM_mpi_scatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_scatterv(const void *sendbuf, const int *sendcounts,
-                                     const int *displs, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_scatterv(const void *sendbuf, const int *sendcounts,
+                                         const int *displs, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -150,12 +151,12 @@ static inline int MPIDI_SHM_scatterv(const void *sendbuf, const int *sendcounts,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_alltoall
+#define FUNCNAME MPIDI_SHM_mpi_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                     MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_alltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                         MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -165,14 +166,14 @@ static inline int MPIDI_SHM_alltoall(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_alltoallv
+#define FUNCNAME MPIDI_SHM_mpi_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_alltoallv(const void *sendbuf, const int *sendcounts,
-                                      const int *sdispls, MPI_Datatype sendtype,
-                                      void *recvbuf, const int *recvcounts,
-                                      const int *rdispls, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_alltoallv(const void *sendbuf, const int *sendcounts,
+                                          const int *sdispls, MPI_Datatype sendtype,
+                                          void *recvbuf, const int *recvcounts,
+                                          const int *rdispls, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -181,14 +182,14 @@ static inline int MPIDI_SHM_alltoallv(const void *sendbuf, const int *sendcounts
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_alltoallw
+#define FUNCNAME MPIDI_SHM_mpi_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_alltoallw(const void *sendbuf, const int sendcounts[],
-                                      const int sdispls[], const MPI_Datatype sendtypes[],
-                                      void *recvbuf, const int recvcounts[],
-                                      const int rdispls[], const MPI_Datatype recvtypes[],
-                                      MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_alltoallw(const void *sendbuf, const int sendcounts[],
+                                          const int sdispls[], const MPI_Datatype sendtypes[],
+                                          void *recvbuf, const int recvcounts[],
+                                          const int rdispls[], const MPI_Datatype recvtypes[],
+                                          MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -196,12 +197,12 @@ static inline int MPIDI_SHM_alltoallw(const void *sendbuf, const int sendcounts[
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_reduce
+#define FUNCNAME MPIDI_SHM_mpi_reduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_reduce(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, int root,
-                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_reduce(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, int root,
+                                       MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -211,13 +212,13 @@ static inline int MPIDI_SHM_reduce(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_reduce_scatter
+#define FUNCNAME MPIDI_SHM_mpi_reduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_reduce_scatter(const void *sendbuf, void *recvbuf,
-                                           const int recvcounts[], MPI_Datatype datatype,
-                                           MPI_Op op, MPIR_Comm * comm_ptr,
-                                           MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_reduce_scatter(const void *sendbuf, void *recvbuf,
+                                               const int recvcounts[], MPI_Datatype datatype,
+                                               MPI_Op op, MPIR_Comm * comm_ptr,
+                                               MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -225,13 +226,13 @@ static inline int MPIDI_SHM_reduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_reduce_scatter_block
+#define FUNCNAME MPIDI_SHM_mpi_reduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_reduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                 int recvcount, MPI_Datatype datatype,
-                                                 MPI_Op op, MPIR_Comm * comm_ptr,
-                                                 MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_reduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                     int recvcount, MPI_Datatype datatype,
+                                                     MPI_Op op, MPIR_Comm * comm_ptr,
+                                                     MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -239,12 +240,12 @@ static inline int MPIDI_SHM_reduce_scatter_block(const void *sendbuf, void *recv
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_scan
+#define FUNCNAME MPIDI_SHM_mpi_scan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_scan(const void *sendbuf, void *recvbuf, int count,
-                                 MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                 MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_scan(const void *sendbuf, void *recvbuf, int count,
+                                     MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                     MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -254,12 +255,12 @@ static inline int MPIDI_SHM_scan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_exscan
+#define FUNCNAME MPIDI_SHM_mpi_exscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_exscan(const void *sendbuf, void *recvbuf, int count,
-                                   MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                   MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_exscan(const void *sendbuf, void *recvbuf, int count,
+                                       MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                       MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -269,13 +270,13 @@ static inline int MPIDI_SHM_exscan(const void *sendbuf, void *recvbuf, int count
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_allgather
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_allgather(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                               MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                               MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_allgather(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -285,14 +286,14 @@ static inline int MPIDI_SHM_neighbor_allgather(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_allgatherv
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf,
-                                                const int recvcounts[], const int displs[],
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    const int recvcounts[], const int displs[],
+                                                    MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                    MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -302,13 +303,13 @@ static inline int MPIDI_SHM_neighbor_allgatherv(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_alltoall
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_alltoall(const void *sendbuf, int sendcount,
-                                              MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                              MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                              MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_alltoall(const void *sendbuf, int sendcount,
+                                                  MPI_Datatype sendtype, void *recvbuf,
+                                                  int recvcount, MPI_Datatype recvtype,
+                                                  MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -318,14 +319,14 @@ static inline int MPIDI_SHM_neighbor_alltoall(const void *sendbuf, int sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_alltoallv
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                               const int sdispls[], MPI_Datatype sendtype,
-                                               void *recvbuf, const int recvcounts[],
-                                               const int rdispls[], MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                   const int sdispls[], MPI_Datatype sendtype,
+                                                   void *recvbuf, const int recvcounts[],
+                                                   const int rdispls[], MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -335,15 +336,15 @@ static inline int MPIDI_SHM_neighbor_alltoallv(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_neighbor_alltoallw
+#define FUNCNAME MPIDI_SHM_mpi_neighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                               const MPI_Aint sdispls[],
-                                               const MPI_Datatype sendtypes[], void *recvbuf,
-                                               const int recvcounts[], const MPI_Aint rdispls[],
-                                               const MPI_Datatype recvtypes[], MPIR_Comm * comm_ptr,
-                                               MPIR_Errflag_t * errflag)
+static inline int MPIDI_SHM_mpi_neighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                   const MPI_Aint sdispls[],
+                                                   const MPI_Datatype sendtypes[], void *recvbuf,
+                                                   const int recvcounts[], const MPI_Aint rdispls[],
+                                                   const MPI_Datatype recvtypes[],
+                                                   MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -353,13 +354,13 @@ static inline int MPIDI_SHM_neighbor_alltoallw(const void *sendbuf, const int se
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_allgather
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_allgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_allgather(const void *sendbuf, int sendcount,
-                                                MPI_Datatype sendtype, void *recvbuf, int recvcount,
-                                                MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_allgather(const void *sendbuf, int sendcount,
+                                                    MPI_Datatype sendtype, void *recvbuf,
+                                                    int recvcount, MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -369,14 +370,14 @@ static inline int MPIDI_SHM_ineighbor_allgather(const void *sendbuf, int sendcou
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_allgatherv
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_allgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_allgatherv(const void *sendbuf, int sendcount,
-                                                 MPI_Datatype sendtype, void *recvbuf,
-                                                 const int recvcounts[], const int displs[],
-                                                 MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                                 MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_allgatherv(const void *sendbuf, int sendcount,
+                                                     MPI_Datatype sendtype, void *recvbuf,
+                                                     const int recvcounts[], const int displs[],
+                                                     MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                                     MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -386,13 +387,13 @@ static inline int MPIDI_SHM_ineighbor_allgatherv(const void *sendbuf, int sendco
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_alltoall
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_alltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_alltoall(const void *sendbuf, int sendcount,
-                                               MPI_Datatype sendtype, void *recvbuf,
-                                               int recvcount, MPI_Datatype recvtype,
-                                               MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_alltoall(const void *sendbuf, int sendcount,
+                                                   MPI_Datatype sendtype, void *recvbuf,
+                                                   int recvcount, MPI_Datatype recvtype,
+                                                   MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -402,14 +403,14 @@ static inline int MPIDI_SHM_ineighbor_alltoall(const void *sendbuf, int sendcoun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_alltoallv
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_alltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
-                                                const int sdispls[], MPI_Datatype sendtype,
-                                                void *recvbuf, const int recvcounts[],
-                                                const int rdispls[], MPI_Datatype recvtype,
-                                                MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_alltoallv(const void *sendbuf, const int sendcounts[],
+                                                    const int sdispls[], MPI_Datatype sendtype,
+                                                    void *recvbuf, const int recvcounts[],
+                                                    const int rdispls[], MPI_Datatype recvtype,
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -419,15 +420,16 @@ static inline int MPIDI_SHM_ineighbor_alltoallv(const void *sendbuf, const int s
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ineighbor_alltoallw
+#define FUNCNAME MPIDI_SHM_mpi_ineighbor_alltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
-                                                const MPI_Aint sdispls[],
-                                                const MPI_Datatype sendtypes[], void *recvbuf,
-                                                const int recvcounts[], const MPI_Aint rdispls[],
-                                                const MPI_Datatype recvtypes[],
-                                                MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ineighbor_alltoallw(const void *sendbuf, const int sendcounts[],
+                                                    const MPI_Aint sdispls[],
+                                                    const MPI_Datatype sendtypes[], void *recvbuf,
+                                                    const int recvcounts[],
+                                                    const MPI_Aint rdispls[],
+                                                    const MPI_Datatype recvtypes[],
+                                                    MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -437,10 +439,10 @@ static inline int MPIDI_SHM_ineighbor_alltoallw(const void *sendbuf, const int s
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ibarrier
+#define FUNCNAME MPIDI_SHM_mpi_ibarrier
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -450,11 +452,11 @@ static inline int MPIDI_SHM_ibarrier(MPIR_Comm * comm_ptr, MPI_Request * req)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ibcast
+#define FUNCNAME MPIDI_SHM_mpi_ibcast
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ibcast(void *buffer, int count, MPI_Datatype datatype,
-                                   int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ibcast(void *buffer, int count, MPI_Datatype datatype,
+                                       int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -464,12 +466,13 @@ static inline int MPIDI_SHM_ibcast(void *buffer, int count, MPI_Datatype datatyp
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iallgather
+#define FUNCNAME MPIDI_SHM_mpi_iallgather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iallgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                       void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iallgather(const void *sendbuf, int sendcount,
+                                           MPI_Datatype sendtype, void *recvbuf, int recvcount,
+                                           MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                           MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -479,13 +482,14 @@ static inline int MPIDI_SHM_iallgather(const void *sendbuf, int sendcount, MPI_D
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iallgatherv
+#define FUNCNAME MPIDI_SHM_mpi_iallgatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iallgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                        void *recvbuf, const int *recvcounts, const int *displs,
-                                        MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
-                                        MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iallgatherv(const void *sendbuf, int sendcount,
+                                            MPI_Datatype sendtype, void *recvbuf,
+                                            const int *recvcounts, const int *displs,
+                                            MPI_Datatype recvtype, MPIR_Comm * comm_ptr,
+                                            MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -495,12 +499,12 @@ static inline int MPIDI_SHM_iallgatherv(const void *sendbuf, int sendcount, MPI_
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ialltoall
+#define FUNCNAME MPIDI_SHM_mpi_ialltoall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                      MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ialltoall(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                          MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -510,14 +514,14 @@ static inline int MPIDI_SHM_ialltoall(const void *sendbuf, int sendcount, MPI_Da
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ialltoallv
+#define FUNCNAME MPIDI_SHM_mpi_ialltoallv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ialltoallv(const void *sendbuf, const int *sendcounts,
-                                       const int *sdispls, MPI_Datatype sendtype,
-                                       void *recvbuf, const int *recvcounts,
-                                       const int *rdispls, MPI_Datatype recvtype,
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ialltoallv(const void *sendbuf, const int *sendcounts,
+                                           const int *sdispls, MPI_Datatype sendtype,
+                                           void *recvbuf, const int *recvcounts,
+                                           const int *rdispls, MPI_Datatype recvtype,
+                                           MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -527,14 +531,14 @@ static inline int MPIDI_SHM_ialltoallv(const void *sendbuf, const int *sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ialltoallw
+#define FUNCNAME MPIDI_SHM_mpi_ialltoallw
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ialltoallw(const void *sendbuf, const int *sendcounts,
-                                       const int *sdispls, const MPI_Datatype sendtypes[],
-                                       void *recvbuf, const int *recvcounts,
-                                       const int *rdispls, const MPI_Datatype recvtypes[],
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ialltoallw(const void *sendbuf, const int *sendcounts,
+                                           const int *sdispls, const MPI_Datatype sendtypes[],
+                                           void *recvbuf, const int *recvcounts,
+                                           const int *rdispls, const MPI_Datatype recvtypes[],
+                                           MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -542,12 +546,12 @@ static inline int MPIDI_SHM_ialltoallw(const void *sendbuf, const int *sendcount
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iexscan
+#define FUNCNAME MPIDI_SHM_mpi_iexscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iexscan(const void *sendbuf, void *recvbuf, int count,
-                                    MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                    MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iexscan(const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                        MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -557,12 +561,12 @@ static inline int MPIDI_SHM_iexscan(const void *sendbuf, void *recvbuf, int coun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_igather
+#define FUNCNAME MPIDI_SHM_mpi_igather
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                    void *recvbuf, int recvcount, MPI_Datatype recvtype,
-                                    int root, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                        void *recvbuf, int recvcount, MPI_Datatype recvtype,
+                                        int root, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -572,13 +576,13 @@ static inline int MPIDI_SHM_igather(const void *sendbuf, int sendcount, MPI_Data
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_igatherv
+#define FUNCNAME MPIDI_SHM_mpi_igatherv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
-                                     void *recvbuf, const int *recvcounts, const int *displs,
-                                     MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
-                                     MPI_Request * req)
+static inline int MPIDI_SHM_mpi_igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
+                                         void *recvbuf, const int *recvcounts, const int *displs,
+                                         MPI_Datatype recvtype, int root, MPIR_Comm * comm_ptr,
+                                         MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -588,13 +592,13 @@ static inline int MPIDI_SHM_igatherv(const void *sendbuf, int sendcount, MPI_Dat
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ireduce_scatter_block
+#define FUNCNAME MPIDI_SHM_mpi_ireduce_scatter_block
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
-                                                  int recvcount, MPI_Datatype datatype,
-                                                  MPI_Op op, MPIR_Comm * comm_ptr,
-                                                  MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ireduce_scatter_block(const void *sendbuf, void *recvbuf,
+                                                      int recvcount, MPI_Datatype datatype,
+                                                      MPI_Op op, MPIR_Comm * comm_ptr,
+                                                      MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -604,12 +608,12 @@ static inline int MPIDI_SHM_ireduce_scatter_block(const void *sendbuf, void *rec
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ireduce_scatter
+#define FUNCNAME MPIDI_SHM_mpi_ireduce_scatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ireduce_scatter(const void *sendbuf, void *recvbuf,
-                                            const int recvcounts[], MPI_Datatype datatype,
-                                            MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ireduce_scatter(const void *sendbuf, void *recvbuf,
+                                                const int recvcounts[], MPI_Datatype datatype,
+                                                MPI_Op op, MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -619,12 +623,12 @@ static inline int MPIDI_SHM_ireduce_scatter(const void *sendbuf, void *recvbuf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_ireduce
+#define FUNCNAME MPIDI_SHM_mpi_ireduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_ireduce(const void *sendbuf, void *recvbuf, int count,
-                                    MPI_Datatype datatype, MPI_Op op, int root,
-                                    MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_ireduce(const void *sendbuf, void *recvbuf, int count,
+                                        MPI_Datatype datatype, MPI_Op op, int root,
+                                        MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -634,12 +638,12 @@ static inline int MPIDI_SHM_ireduce(const void *sendbuf, void *recvbuf, int coun
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iallreduce
+#define FUNCNAME MPIDI_SHM_mpi_iallreduce
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iallreduce(const void *sendbuf, void *recvbuf, int count,
-                                       MPI_Datatype datatype, MPI_Op op,
-                                       MPIR_Comm * comm_ptr, MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iallreduce(const void *sendbuf, void *recvbuf, int count,
+                                           MPI_Datatype datatype, MPI_Op op,
+                                           MPIR_Comm * comm_ptr, MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -649,12 +653,12 @@ static inline int MPIDI_SHM_iallreduce(const void *sendbuf, void *recvbuf, int c
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iscan
+#define FUNCNAME MPIDI_SHM_mpi_iscan
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iscan(const void *sendbuf, void *recvbuf, int count,
-                                  MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
-                                  MPI_Request * req)
+static inline int MPIDI_SHM_mpi_iscan(const void *sendbuf, void *recvbuf, int count,
+                                      MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm_ptr,
+                                      MPI_Request * req)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -664,13 +668,13 @@ static inline int MPIDI_SHM_iscan(const void *sendbuf, void *recvbuf, int count,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iscatter
+#define FUNCNAME MPIDI_SHM_mpi_iscatter
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iscatter(const void *sendbuf, int sendcount,
-                                     MPI_Datatype sendtype, void *recvbuf,
-                                     int recvcount, MPI_Datatype recvtype,
-                                     int root, MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_SHM_mpi_iscatter(const void *sendbuf, int sendcount,
+                                         MPI_Datatype sendtype, void *recvbuf,
+                                         int recvcount, MPI_Datatype recvtype,
+                                         int root, MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -678,14 +682,14 @@ static inline int MPIDI_SHM_iscatter(const void *sendbuf, int sendcount,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_iscatterv
+#define FUNCNAME MPIDI_SHM_mpi_iscatterv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_iscatterv(const void *sendbuf, const int *sendcounts,
-                                      const int *displs, MPI_Datatype sendtype,
-                                      void *recvbuf, int recvcount,
-                                      MPI_Datatype recvtype, int root,
-                                      MPIR_Comm * comm, MPI_Request * request)
+static inline int MPIDI_SHM_mpi_iscatterv(const void *sendbuf, const int *sendcounts,
+                                          const int *displs, MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount,
+                                          MPI_Datatype recvtype, int root,
+                                          MPIR_Comm * comm, MPI_Request * request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_comm.h b/src/mpid/ch4/shm/stubshm/stubshm_comm.h
index 99ae077..498e4b4 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_comm.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_comm.h
@@ -11,14 +11,14 @@
 #define SHM_STUBSHM_COMM_H_INCLUDED
 
 #include "stubshm_impl.h"
-static inline int MPIDI_SHM_comm_create_hook(MPIR_Comm * comm)
+static inline int MPIDI_SHM_mpi_comm_create_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
     return mpi_errno;
 }
 
-static inline int MPIDI_SHM_comm_free_hook(MPIR_Comm * comm)
+static inline int MPIDI_SHM_mpi_comm_free_hook(MPIR_Comm * comm)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Assert(0);
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_init.h b/src/mpid/ch4/shm/stubshm/stubshm_init.h
index f101951..6b61d99 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_init.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_init.h
@@ -13,25 +13,25 @@
 
 #include "stubshm_impl.h"
 
-static inline int MPIDI_SHM_init(int rank, int size)
+static inline int MPIDI_SHM_mpi_init(int rank, int size)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_finalize(void)
+static inline int MPIDI_SHM_mpi_finalize(void)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline void *MPIDI_SHM_alloc_mem(size_t size, MPIR_Info * info_ptr)
+static inline void *MPIDI_SHM_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
     MPIR_Assert(0);
     return NULL;
 }
 
-static inline int MPIDI_SHM_free_mem(void *ptr)
+static inline int MPIDI_SHM_mpi_free_mem(void *ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
@@ -83,10 +83,10 @@ static inline int MPIDI_SHM_create_intercomm_from_lpids(MPIR_Comm * newcomm_ptr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_type_create_hook
+#define FUNCNAME MPIDI_SHM_mpi_type_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_type_create_hook(MPIR_Datatype * type)
+static inline int MPIDI_SHM_mpi_type_create_hook(MPIR_Datatype * type)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_TYPE_CREATE_HOOK);
@@ -97,10 +97,10 @@ static inline int MPIDI_SHM_type_create_hook(MPIR_Datatype * type)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_type_free_hook
+#define FUNCNAME MPIDI_SHM_mpi_type_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_type_free_hook(MPIR_Datatype * type)
+static inline int MPIDI_SHM_mpi_type_free_hook(MPIR_Datatype * type)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_TYPE_FREE_HOOK);
@@ -111,10 +111,10 @@ static inline int MPIDI_SHM_type_free_hook(MPIR_Datatype * type)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_op_create_hook
+#define FUNCNAME MPIDI_SHM_mpi_op_create_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_op_create_hook(MPIR_Op * op)
+static inline int MPIDI_SHM_mpi_op_create_hook(MPIR_Op * op)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_OP_CREATE_HOOK);
@@ -125,10 +125,10 @@ static inline int MPIDI_SHM_op_create_hook(MPIR_Op * op)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_SHM_op_free_hook
+#define FUNCNAME MPIDI_SHM_mpi_op_free_hook
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_SHM_op_free_hook(MPIR_Op * op)
+static inline int MPIDI_SHM_mpi_op_free_hook(MPIR_Op * op)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_SHM_OP_FREE_HOOK);
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_probe.h b/src/mpid/ch4/shm/stubshm/stubshm_probe.h
index 0ba72f6..726735f 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_probe.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_probe.h
@@ -14,20 +14,20 @@
 #include "stubshm_impl.h"
 
 
-static inline int MPIDI_SHM_improbe(int source,
-                                    int tag,
-                                    MPIR_Comm * comm,
-                                    int context_offset,
-                                    int *flag, MPIR_Request ** message, MPI_Status * status)
+static inline int MPIDI_SHM_mpi_improbe(int source,
+                                        int tag,
+                                        MPIR_Comm * comm,
+                                        int context_offset,
+                                        int *flag, MPIR_Request ** message, MPI_Status * status)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_iprobe(int source,
-                                   int tag,
-                                   MPIR_Comm * comm,
-                                   int context_offset, int *flag, MPI_Status * status)
+static inline int MPIDI_SHM_mpi_iprobe(int source,
+                                       int tag,
+                                       MPIR_Comm * comm,
+                                       int context_offset, int *flag, MPI_Status * status)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_recv.h b/src/mpid/ch4/shm/stubshm/stubshm_recv.h
index 63eaf7a..eb16842 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_recv.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_recv.h
@@ -14,54 +14,56 @@
 #include "stubshm_impl.h"
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_recv)
-static inline int MPIDI_SHM_recv(void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm,
-                                 int context_offset, MPI_Status * status, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_recv)
+static inline int MPIDI_SHM_mpi_recv(void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm,
+                                     int context_offset, MPI_Status * status,
+                                     MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_recv_init(void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_recv_init(void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_imrecv(void *buf,
-                                   int count,
-                                   MPI_Datatype datatype,
-                                   MPIR_Request * message, MPIR_Request ** rreqp)
+static inline int MPIDI_SHM_mpi_imrecv(void *buf,
+                                       int count,
+                                       MPI_Datatype datatype,
+                                       MPIR_Request * message, MPIR_Request ** rreqp)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_irecv)
-static inline int MPIDI_SHM_irecv(void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_irecv)
+static inline int MPIDI_SHM_mpi_irecv(void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_cancel_recv(MPIR_Request * rreq)
+static inline int MPIDI_SHM_mpi_cancel_recv(MPIR_Request * rreq)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_rma.h b/src/mpid/ch4/shm/stubshm/stubshm_rma.h
index 34256d1..447b0f1 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_rma.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_rma.h
@@ -13,71 +13,56 @@
 
 #include "stubshm_impl.h"
 
-static inline int MPIDI_SHM_put(const void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_put(const void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_get(void *origin_addr,
-                                int origin_count,
-                                MPI_Datatype origin_datatype,
-                                int target_rank,
-                                MPI_Aint target_disp,
-                                int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_get(void *origin_addr,
+                                    int origin_count,
+                                    MPI_Datatype origin_datatype,
+                                    int target_rank,
+                                    MPI_Aint target_disp,
+                                    int target_count, MPI_Datatype target_datatype, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_rput(const void *origin_addr,
-                                 int origin_count,
-                                 MPI_Datatype origin_datatype,
-                                 int target_rank,
-                                 MPI_Aint target_disp,
-                                 int target_count,
-                                 MPI_Datatype target_datatype,
-                                 MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_rput(const void *origin_addr,
+                                     int origin_count,
+                                     MPI_Datatype origin_datatype,
+                                     int target_rank,
+                                     MPI_Aint target_disp,
+                                     int target_count,
+                                     MPI_Datatype target_datatype,
+                                     MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_compare_and_swap(const void *origin_addr,
-                                             const void *compare_addr,
-                                             void *result_addr,
-                                             MPI_Datatype datatype,
-                                             int target_rank, MPI_Aint target_disp, MPIR_Win * win)
-{
-    MPIR_Assert(0);
-    return MPI_SUCCESS;
-}
-
-static inline int MPIDI_SHM_raccumulate(const void *origin_addr,
-                                        int origin_count,
-                                        MPI_Datatype origin_datatype,
-                                        int target_rank,
-                                        MPI_Aint target_disp,
-                                        int target_count,
-                                        MPI_Datatype target_datatype,
-                                        MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_compare_and_swap(const void *origin_addr,
+                                                 const void *compare_addr,
+                                                 void *result_addr,
+                                                 MPI_Datatype datatype,
+                                                 int target_rank, MPI_Aint target_disp,
+                                                 MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_rget_accumulate(const void *origin_addr,
+static inline int MPIDI_SHM_mpi_raccumulate(const void *origin_addr,
                                             int origin_count,
                                             MPI_Datatype origin_datatype,
-                                            void *result_addr,
-                                            int result_count,
-                                            MPI_Datatype result_datatype,
                                             int target_rank,
                                             MPI_Aint target_disp,
                                             int target_count,
@@ -88,37 +73,66 @@ static inline int MPIDI_SHM_rget_accumulate(const void *origin_addr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_fetch_and_op(const void *origin_addr,
-                                         void *result_addr,
-                                         MPI_Datatype datatype,
-                                         int target_rank,
-                                         MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_rget_accumulate(const void *origin_addr,
+                                                int origin_count,
+                                                MPI_Datatype origin_datatype,
+                                                void *result_addr,
+                                                int result_count,
+                                                MPI_Datatype result_datatype,
+                                                int target_rank,
+                                                MPI_Aint target_disp,
+                                                int target_count,
+                                                MPI_Datatype target_datatype,
+                                                MPI_Op op, MPIR_Win * win, MPIR_Request ** request)
+{
+    MPIR_Assert(0);
+    return MPI_SUCCESS;
+}
+
+static inline int MPIDI_SHM_mpi_fetch_and_op(const void *origin_addr,
+                                             void *result_addr,
+                                             MPI_Datatype datatype,
+                                             int target_rank,
+                                             MPI_Aint target_disp, MPI_Op op, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_rget(void *origin_addr,
-                                 int origin_count,
-                                 MPI_Datatype origin_datatype,
-                                 int target_rank,
-                                 MPI_Aint target_disp,
-                                 int target_count,
-                                 MPI_Datatype target_datatype,
-                                 MPIR_Win * win, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_rget(void *origin_addr,
+                                     int origin_count,
+                                     MPI_Datatype origin_datatype,
+                                     int target_rank,
+                                     MPI_Aint target_disp,
+                                     int target_count,
+                                     MPI_Datatype target_datatype,
+                                     MPIR_Win * win, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_get_accumulate(const void *origin_addr,
+static inline int MPIDI_SHM_mpi_get_accumulate(const void *origin_addr,
+                                               int origin_count,
+                                               MPI_Datatype origin_datatype,
+                                               void *result_addr,
+                                               int result_count,
+                                               MPI_Datatype result_datatype,
+                                               int target_rank,
+                                               MPI_Aint target_disp,
+                                               int target_count,
+                                               MPI_Datatype target_datatype, MPI_Op op,
+                                               MPIR_Win * win)
+{
+    MPIR_Assert(0);
+    return MPI_SUCCESS;
+}
+
+static inline int MPIDI_SHM_mpi_accumulate(const void *origin_addr,
                                            int origin_count,
                                            MPI_Datatype origin_datatype,
-                                           void *result_addr,
-                                           int result_count,
-                                           MPI_Datatype result_datatype,
                                            int target_rank,
                                            MPI_Aint target_disp,
                                            int target_count,
@@ -128,16 +142,4 @@ static inline int MPIDI_SHM_get_accumulate(const void *origin_addr,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_accumulate(const void *origin_addr,
-                                       int origin_count,
-                                       MPI_Datatype origin_datatype,
-                                       int target_rank,
-                                       MPI_Aint target_disp,
-                                       int target_count,
-                                       MPI_Datatype target_datatype, MPI_Op op, MPIR_Win * win)
-{
-    MPIR_Assert(0);
-    return MPI_SUCCESS;
-}
-
 #endif /* SHM_STUBSHM_RMA_H_INCLUDED */
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_send.h b/src/mpid/ch4/shm/stubshm/stubshm_send.h
index 446959f..1d0e1d0 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_send.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_send.h
@@ -13,12 +13,12 @@
 
 #include "stubshm_impl.h"
 
-static inline int MPIDI_SHM_send(const void *buf,
-                                 int count,
-                                 MPI_Datatype datatype,
-                                 int rank,
-                                 int tag,
-                                 MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_send(const void *buf,
+                                     int count,
+                                     MPI_Datatype datatype,
+                                     int rank,
+                                     int tag,
+                                     MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int err = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -39,12 +39,12 @@ static inline int MPIDI_SHM_irsend(const void *buf,
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_ssend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_ssend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int err = MPI_SUCCESS;
     MPIR_Assert(0);
@@ -52,86 +52,88 @@ static inline int MPIDI_SHM_ssend(const void *buf,
     return err;
 }
 
-static inline int MPIDI_SHM_startall(int count, MPIR_Request * requests[])
+static inline int MPIDI_SHM_mpi_startall(int count, MPIR_Request * requests[])
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_send_init(const void *buf,
-                                      int count,
-                                      MPI_Datatype datatype,
-                                      int rank,
-                                      int tag,
-                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_send_init(const void *buf,
+                                          int count,
+                                          MPI_Datatype datatype,
+                                          int rank,
+                                          int tag,
+                                          MPIR_Comm * comm, int context_offset,
+                                          MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_ssend_init(const void *buf,
-                                       int count,
-                                       MPI_Datatype datatype,
-                                       int rank,
-                                       int tag,
-                                       MPIR_Comm * comm,
-                                       int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_ssend_init(const void *buf,
+                                           int count,
+                                           MPI_Datatype datatype,
+                                           int rank,
+                                           int tag,
+                                           MPIR_Comm * comm,
+                                           int context_offset, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_bsend_init(const void *buf,
-                                       int count,
-                                       MPI_Datatype datatype,
-                                       int rank,
-                                       int tag,
-                                       MPIR_Comm * comm,
-                                       int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_bsend_init(const void *buf,
+                                           int count,
+                                           MPI_Datatype datatype,
+                                           int rank,
+                                           int tag,
+                                           MPIR_Comm * comm,
+                                           int context_offset, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_rsend_init(const void *buf,
-                                       int count,
-                                       MPI_Datatype datatype,
-                                       int rank,
-                                       int tag,
-                                       MPIR_Comm * comm,
-                                       int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_rsend_init(const void *buf,
+                                           int count,
+                                           MPI_Datatype datatype,
+                                           int rank,
+                                           int tag,
+                                           MPIR_Comm * comm,
+                                           int context_offset, MPIR_Request ** request)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 #undef FCNAME
-#define FCNAME DECL_FUNC(MPIDI_SHM_isend)
-static inline int MPIDI_SHM_isend(const void *buf,
-                                  int count,
-                                  MPI_Datatype datatype,
-                                  int rank,
-                                  int tag,
-                                  MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+#define FCNAME DECL_FUNC(MPIDI_SHM_mpi_isend)
+static inline int MPIDI_SHM_mpi_isend(const void *buf,
+                                      int count,
+                                      MPI_Datatype datatype,
+                                      int rank,
+                                      int tag,
+                                      MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
 {
     int err = MPI_SUCCESS;
     MPIR_Assert(0);
     return err;
 }
 
-static inline int MPIDI_SHM_issend(const void *buf,
-                                   int count,
-                                   MPI_Datatype datatype,
-                                   int rank,
-                                   int tag,
-                                   MPIR_Comm * comm, int context_offset, MPIR_Request ** request)
+static inline int MPIDI_SHM_mpi_issend(const void *buf,
+                                       int count,
+                                       MPI_Datatype datatype,
+                                       int rank,
+                                       int tag,
+                                       MPIR_Comm * comm, int context_offset,
+                                       MPIR_Request ** request)
 {
     int err = MPI_SUCCESS;
     MPIR_Assert(0);
     return err;
 }
 
-static inline int MPIDI_SHM_cancel_send(MPIR_Request * sreq)
+static inline int MPIDI_SHM_mpi_cancel_send(MPIR_Request * sreq)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_spawn.h b/src/mpid/ch4/shm/stubshm/stubshm_spawn.h
index cedb3f4..5995d0e 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_spawn.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_spawn.h
@@ -13,35 +13,35 @@
 
 #include "stubshm_impl.h"
 
-static inline int MPIDI_SHM_comm_connect(const char *port_name,
-                                         MPIR_Info * info,
-                                         int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_SHM_mpi_comm_connect(const char *port_name,
+                                             MPIR_Info * info,
+                                             int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_comm_disconnect(MPIR_Comm * comm_ptr)
+static inline int MPIDI_SHM_mpi_comm_disconnect(MPIR_Comm * comm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_open_port(MPIR_Info * info_ptr, char *port_name)
+static inline int MPIDI_SHM_mpi_open_port(MPIR_Info * info_ptr, char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_close_port(const char *port_name)
+static inline int MPIDI_SHM_mpi_close_port(const char *port_name)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_comm_accept(const char *port_name,
-                                        MPIR_Info * info,
-                                        int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
+static inline int MPIDI_SHM_mpi_comm_accept(const char *port_name,
+                                            MPIR_Info * info,
+                                            int root, MPIR_Comm * comm, MPIR_Comm ** newcomm_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/shm/stubshm/stubshm_win.h b/src/mpid/ch4/shm/stubshm/stubshm_win.h
index 24a9b1b..b5ce0c7 100644
--- a/src/mpid/ch4/shm/stubshm/stubshm_win.h
+++ b/src/mpid/ch4/shm/stubshm/stubshm_win.h
@@ -13,169 +13,171 @@
 
 #include "stubshm_impl.h"
 
-static inline int MPIDI_SHM_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_SHM_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_complete(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_complete(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_wait(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_wait(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_SHM_mpi_win_test(MPIR_Win * win, int *flag)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_unlock(int rank, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_SHM_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
 
-static inline int MPIDI_SHM_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_SHM_mpi_win_free(MPIR_Win ** win_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_fence(int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_fence(int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_create(void *base,
-                                       MPI_Aint length,
-                                       int disp_unit,
-                                       MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_SHM_mpi_win_create(void *base,
+                                           MPI_Aint length,
+                                           int disp_unit,
+                                           MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                           MPIR_Win ** win_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_SHM_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_allocate_shared(MPI_Aint size,
-                                                int disp_unit,
-                                                MPIR_Info * info_ptr,
-                                                MPIR_Comm * comm_ptr,
-                                                void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_SHM_mpi_win_allocate_shared(MPI_Aint size,
+                                                    int disp_unit,
+                                                    MPIR_Info * info_ptr,
+                                                    MPIR_Comm * comm_ptr,
+                                                    void **base_ptr, MPIR_Win ** win_ptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_SHM_mpi_win_detach(MPIR_Win * win, const void *base)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_shared_query(MPIR_Win * win,
-                                             int rank,
-                                             MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_SHM_mpi_win_shared_query(MPIR_Win * win,
+                                                 int rank,
+                                                 MPI_Aint * size, int *disp_unit, void *baseptr)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_allocate(MPI_Aint size,
-                                         int disp_unit,
-                                         MPIR_Info * info,
-                                         MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
+static inline int MPIDI_SHM_mpi_win_allocate(MPI_Aint size,
+                                             int disp_unit,
+                                             MPIR_Info * info,
+                                             MPIR_Comm * comm, void *baseptr, MPIR_Win ** win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush(int rank, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush_local_all(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_unlock_all(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm, MPIR_Win ** win)
+static inline int MPIDI_SHM_mpi_win_create_dynamic(MPIR_Info * info, MPIR_Comm * comm,
+                                                   MPIR_Win ** win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_sync(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_sync(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_flush_all(MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
 }
 
-static inline int MPIDI_SHM_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_SHM_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
     MPIR_Assert(0);
     return MPI_SUCCESS;
diff --git a/src/mpid/ch4/src/ch4_coll.h b/src/mpid/ch4/src/ch4_coll.h
index 595b176..3d9bbaf 100644
--- a/src/mpid/ch4/src/ch4_coll.h
+++ b/src/mpid/ch4/src/ch4_coll.h
@@ -16,20 +16,20 @@
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Barrier(MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_barrier(comm, errflag);
+    return MPIDI_NM_mpi_barrier(comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Bcast(void *buffer, int count, MPI_Datatype datatype,
                                          int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_bcast(buffer, count, datatype, root, comm, errflag);
+    return MPIDI_NM_mpi_bcast(buffer, count, datatype, root, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Allreduce(const void *sendbuf, void *recvbuf, int count,
                                              MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                              MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_allreduce(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_NM_mpi_allreduce(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Allgather(const void *sendbuf, int sendcount,
@@ -37,8 +37,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Allgather(const void *sendbuf, int sendcount,
                                              MPI_Datatype recvtype, MPIR_Comm * comm,
                                              MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_allgather(sendbuf, sendcount, sendtype, recvbuf,
-                              recvcount, recvtype, comm, errflag);
+    return MPIDI_NM_mpi_allgather(sendbuf, sendcount, sendtype, recvbuf,
+                                  recvcount, recvtype, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Allgatherv(const void *sendbuf, int sendcount,
@@ -47,8 +47,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Allgatherv(const void *sendbuf, int sendcount
                                               MPI_Datatype recvtype, MPIR_Comm * comm,
                                               MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
-                               recvcounts, displs, recvtype, comm, errflag);
+    return MPIDI_NM_mpi_allgatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                   recvcounts, displs, recvtype, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Scatter(const void *sendbuf, int sendcount,
@@ -56,8 +56,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Scatter(const void *sendbuf, int sendcount,
                                            MPI_Datatype recvtype, int root, MPIR_Comm * comm,
                                            MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_scatter(sendbuf, sendcount, sendtype, recvbuf,
-                            recvcount, recvtype, root, comm, errflag);
+    return MPIDI_NM_mpi_scatter(sendbuf, sendcount, sendtype, recvbuf,
+                                recvcount, recvtype, root, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Scatterv(const void *sendbuf, const int *sendcounts,
@@ -66,16 +66,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Scatterv(const void *sendbuf, const int *send
                                             int root, MPIR_Comm * comm_ptr,
                                             MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_scatterv(sendbuf, sendcounts, displs, sendtype,
-                             recvbuf, recvcount, recvtype, root, comm_ptr, errflag);
+    return MPIDI_NM_mpi_scatterv(sendbuf, sendcounts, displs, sendtype,
+                                 recvbuf, recvcount, recvtype, root, comm_ptr, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
                                           void *recvbuf, int recvcount, MPI_Datatype recvtype,
                                           int root, MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_gather(sendbuf, sendcount, sendtype, recvbuf,
-                           recvcount, recvtype, root, comm, errflag);
+    return MPIDI_NM_mpi_gather(sendbuf, sendcount, sendtype, recvbuf,
+                               recvcount, recvtype, root, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Gatherv(const void *sendbuf, int sendcount,
@@ -84,8 +84,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Gatherv(const void *sendbuf, int sendcount,
                                            MPI_Datatype recvtype, int root, MPIR_Comm * comm,
                                            MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_gatherv(sendbuf, sendcount, sendtype, recvbuf,
-                            recvcounts, displs, recvtype, root, comm, errflag);
+    return MPIDI_NM_mpi_gatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                recvcounts, displs, recvtype, root, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Alltoall(const void *sendbuf, int sendcount,
@@ -93,8 +93,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Alltoall(const void *sendbuf, int sendcount,
                                             MPI_Datatype recvtype, MPIR_Comm * comm,
                                             MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_alltoall(sendbuf, sendcount, sendtype, recvbuf,
-                             recvcount, recvtype, comm, errflag);
+    return MPIDI_NM_mpi_alltoall(sendbuf, sendcount, sendtype, recvbuf,
+                                 recvcount, recvtype, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Alltoallv(const void *sendbuf, const int *sendcounts,
@@ -103,8 +103,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Alltoallv(const void *sendbuf, const int *sen
                                              const int *rdispls, MPI_Datatype recvtype,
                                              MPIR_Comm * comm, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
-                              recvbuf, recvcounts, rdispls, recvtype, comm, errflag);
+    return MPIDI_NM_mpi_alltoallv(sendbuf, sendcounts, sdispls, sendtype,
+                                  recvbuf, recvcounts, rdispls, recvtype, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Alltoallw(const void *sendbuf, const int sendcounts[],
@@ -113,15 +113,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Alltoallw(const void *sendbuf, const int send
                                              const int rdispls[], const MPI_Datatype recvtypes[],
                                              MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
-                              recvbuf, recvcounts, rdispls, recvtypes, comm_ptr, errflag);
+    return MPIDI_NM_mpi_alltoallw(sendbuf, sendcounts, sdispls, sendtypes,
+                                  recvbuf, recvcounts, rdispls, recvtypes, comm_ptr, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Reduce(const void *sendbuf, void *recvbuf,
                                           int count, MPI_Datatype datatype, MPI_Op op,
                                           int root, MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_reduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr, errflag);
+    return MPIDI_NM_mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm_ptr, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Reduce_scatter(const void *sendbuf, void *recvbuf,
@@ -129,7 +129,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Reduce_scatter(const void *sendbuf, void *rec
                                                   MPI_Op op, MPIR_Comm * comm_ptr,
                                                   MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_reduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm_ptr, errflag);
+    return MPIDI_NM_mpi_reduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm_ptr,
+                                       errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Reduce_scatter_block(const void *sendbuf, void *recvbuf,
@@ -137,22 +138,22 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Reduce_scatter_block(const void *sendbuf, voi
                                                         MPI_Op op, MPIR_Comm * comm_ptr,
                                                         MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_reduce_scatter_block(sendbuf, recvbuf, recvcount,
-                                         datatype, op, comm_ptr, errflag);
+    return MPIDI_NM_mpi_reduce_scatter_block(sendbuf, recvbuf, recvcount,
+                                             datatype, op, comm_ptr, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Scan(const void *sendbuf, void *recvbuf, int count,
                                         MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                         MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_scan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_NM_mpi_scan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Exscan(const void *sendbuf, void *recvbuf, int count,
                                           MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                           MPIR_Errflag_t * errflag)
 {
-    return MPIDI_NM_exscan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
+    return MPIDI_NM_mpi_exscan(sendbuf, recvbuf, count, datatype, op, comm, errflag);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_allgather(const void *sendbuf, int sendcount,
@@ -160,8 +161,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_allgather(const void *sendbuf, int s
                                                       int recvcount, MPI_Datatype recvtype,
                                                       MPIR_Comm * comm)
 {
-    return MPIDI_NM_neighbor_allgather(sendbuf, sendcount, sendtype,
-                                       recvbuf, recvcount, recvtype, comm);
+    return MPIDI_NM_mpi_neighbor_allgather(sendbuf, sendcount, sendtype,
+                                           recvbuf, recvcount, recvtype, comm);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_allgatherv(const void *sendbuf, int sendcount,
@@ -169,8 +170,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_allgatherv(const void *sendbuf, int
                                                        const int *recvcounts, const int *displs,
                                                        MPI_Datatype recvtype, MPIR_Comm * comm)
 {
-    return MPIDI_NM_neighbor_allgatherv(sendbuf, sendcount, sendtype,
-                                        recvbuf, recvcounts, displs, recvtype, comm);
+    return MPIDI_NM_mpi_neighbor_allgatherv(sendbuf, sendcount, sendtype,
+                                            recvbuf, recvcounts, displs, recvtype, comm);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_alltoallv(const void *sendbuf, const int *sendcounts,
@@ -179,8 +180,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_alltoallv(const void *sendbuf, const
                                                       const int *rdispls, MPI_Datatype recvtype,
                                                       MPIR_Comm * comm)
 {
-    return MPIDI_NM_neighbor_alltoallv(sendbuf, sendcounts, sdispls,
-                                       sendtype, recvbuf, recvcounts, rdispls, recvtype, comm);
+    return MPIDI_NM_mpi_neighbor_alltoallv(sendbuf, sendcounts, sdispls,
+                                           sendtype, recvbuf, recvcounts, rdispls, recvtype, comm);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_alltoallw(const void *sendbuf, const int *sendcounts,
@@ -191,8 +192,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_alltoallw(const void *sendbuf, const
                                                       const MPI_Datatype * recvtypes,
                                                       MPIR_Comm * comm)
 {
-    return MPIDI_NM_neighbor_alltoallw(sendbuf, sendcounts, sdispls,
-                                       sendtypes, recvbuf, recvcounts, rdispls, recvtypes, comm);
+    return MPIDI_NM_mpi_neighbor_alltoallw(sendbuf, sendcounts, sdispls,
+                                           sendtypes, recvbuf, recvcounts, rdispls, recvtypes,
+                                           comm);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_alltoall(const void *sendbuf, int sendcount,
@@ -200,8 +202,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Neighbor_alltoall(const void *sendbuf, int se
                                                      int recvcount, MPI_Datatype recvtype,
                                                      MPIR_Comm * comm)
 {
-    return MPIDI_NM_neighbor_alltoall(sendbuf, sendcount, sendtype,
-                                      recvbuf, recvcount, recvtype, comm);
+    return MPIDI_NM_mpi_neighbor_alltoall(sendbuf, sendcount, sendtype,
+                                          recvbuf, recvcount, recvtype, comm);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_allgather(const void *sendbuf, int sendcount,
@@ -209,8 +211,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_allgather(const void *sendbuf, int
                                                        int recvcount, MPI_Datatype recvtype,
                                                        MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ineighbor_allgather(sendbuf, sendcount, sendtype,
-                                        recvbuf, recvcount, recvtype, comm, req);
+    return MPIDI_NM_mpi_ineighbor_allgather(sendbuf, sendcount, sendtype,
+                                            recvbuf, recvcount, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_allgatherv(const void *sendbuf, int sendcount,
@@ -219,8 +221,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_allgatherv(const void *sendbuf, int
                                                         MPI_Datatype recvtype, MPIR_Comm * comm,
                                                         MPI_Request * req)
 {
-    return MPIDI_NM_ineighbor_allgatherv(sendbuf, sendcount, sendtype,
-                                         recvbuf, recvcounts, displs, recvtype, comm, req);
+    return MPIDI_NM_mpi_ineighbor_allgatherv(sendbuf, sendcount, sendtype,
+                                             recvbuf, recvcounts, displs, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_alltoall(const void *sendbuf, int sendcount,
@@ -228,8 +230,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_alltoall(const void *sendbuf, int s
                                                       int recvcount, MPI_Datatype recvtype,
                                                       MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ineighbor_alltoall(sendbuf, sendcount, sendtype,
-                                       recvbuf, recvcount, recvtype, comm, req);
+    return MPIDI_NM_mpi_ineighbor_alltoall(sendbuf, sendcount, sendtype,
+                                           recvbuf, recvcount, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_alltoallv(const void *sendbuf, const int *sendcounts,
@@ -238,9 +240,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_alltoallv(const void *sendbuf, cons
                                                        const int *rdispls, MPI_Datatype recvtype,
                                                        MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ineighbor_alltoallv(sendbuf, sendcounts, sdispls,
-                                        sendtype, recvbuf, recvcounts, rdispls, recvtype, comm,
-                                        req);
+    return MPIDI_NM_mpi_ineighbor_alltoallv(sendbuf, sendcounts, sdispls,
+                                            sendtype, recvbuf, recvcounts, rdispls, recvtype, comm,
+                                            req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_alltoallw(const void *sendbuf, const int *sendcounts,
@@ -251,20 +253,20 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ineighbor_alltoallw(const void *sendbuf, cons
                                                        const MPI_Datatype * recvtypes,
                                                        MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ineighbor_alltoallw(sendbuf, sendcounts, sdispls,
-                                        sendtypes, recvbuf, recvcounts, rdispls, recvtypes, comm,
-                                        req);
+    return MPIDI_NM_mpi_ineighbor_alltoallw(sendbuf, sendcounts, sdispls,
+                                            sendtypes, recvbuf, recvcounts, rdispls, recvtypes,
+                                            comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ibarrier(MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ibarrier(comm, req);
+    return MPIDI_NM_mpi_ibarrier(comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ibcast(void *buffer, int count, MPI_Datatype datatype,
                                           int root, MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ibcast(buffer, count, datatype, root, comm, req);
+    return MPIDI_NM_mpi_ibcast(buffer, count, datatype, root, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iallgather(const void *sendbuf, int sendcount,
@@ -272,8 +274,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Iallgather(const void *sendbuf, int sendcount
                                               MPI_Datatype recvtype, MPIR_Comm * comm,
                                               MPI_Request * req)
 {
-    return MPIDI_NM_iallgather(sendbuf, sendcount, sendtype, recvbuf,
-                               recvcount, recvtype, comm, req);
+    return MPIDI_NM_mpi_iallgather(sendbuf, sendcount, sendtype, recvbuf,
+                                   recvcount, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iallgatherv(const void *sendbuf, int sendcount,
@@ -282,15 +284,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Iallgatherv(const void *sendbuf, int sendcoun
                                                MPI_Datatype recvtype, MPIR_Comm * comm,
                                                MPI_Request * req)
 {
-    return MPIDI_NM_iallgatherv(sendbuf, sendcount, sendtype, recvbuf,
-                                recvcounts, displs, recvtype, comm, req);
+    return MPIDI_NM_mpi_iallgatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                    recvcounts, displs, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iallreduce(const void *sendbuf, void *recvbuf, int count,
                                               MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                               MPI_Request * req)
 {
-    return MPIDI_NM_iallreduce(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_NM_mpi_iallreduce(sendbuf, recvbuf, count, datatype, op, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ialltoall(const void *sendbuf, int sendcount,
@@ -298,8 +300,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ialltoall(const void *sendbuf, int sendcount,
                                              MPI_Datatype recvtype, MPIR_Comm * comm,
                                              MPI_Request * req)
 {
-    return MPIDI_NM_ialltoall(sendbuf, sendcount, sendtype, recvbuf,
-                              recvcount, recvtype, comm, req);
+    return MPIDI_NM_mpi_ialltoall(sendbuf, sendcount, sendtype, recvbuf,
+                                  recvcount, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ialltoallv(const void *sendbuf, const int *sendcounts,
@@ -308,8 +310,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ialltoallv(const void *sendbuf, const int *se
                                               const int *rdispls, MPI_Datatype recvtype,
                                               MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ialltoallv(sendbuf, sendcounts, sdispls, sendtype,
-                               recvbuf, recvcounts, rdispls, recvtype, comm, req);
+    return MPIDI_NM_mpi_ialltoallv(sendbuf, sendcounts, sdispls, sendtype,
+                                   recvbuf, recvcounts, rdispls, recvtype, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ialltoallw(const void *sendbuf, const int *sendcounts,
@@ -318,15 +320,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ialltoallw(const void *sendbuf, const int *se
                                               const int *rdispls, const MPI_Datatype * recvtypes,
                                               MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ialltoallw(sendbuf, sendcounts, sdispls, sendtypes,
-                               recvbuf, recvcounts, rdispls, recvtypes, comm, req);
+    return MPIDI_NM_mpi_ialltoallw(sendbuf, sendcounts, sdispls, sendtypes,
+                                   recvbuf, recvcounts, rdispls, recvtypes, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iexscan(const void *sendbuf, void *recvbuf, int count,
                                            MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                            MPI_Request * req)
 {
-    return MPIDI_NM_iexscan(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_NM_mpi_iexscan(sendbuf, recvbuf, count, datatype, op, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Igather(const void *sendbuf, int sendcount,
@@ -334,8 +336,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Igather(const void *sendbuf, int sendcount,
                                            MPI_Datatype recvtype, int root, MPIR_Comm * comm,
                                            MPI_Request * req)
 {
-    return MPIDI_NM_igather(sendbuf, sendcount, sendtype, recvbuf,
-                            recvcount, recvtype, root, comm, req);
+    return MPIDI_NM_mpi_igather(sendbuf, sendcount, sendtype, recvbuf,
+                                recvcount, recvtype, root, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Igatherv(const void *sendbuf, int sendcount,
@@ -344,8 +346,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Igatherv(const void *sendbuf, int sendcount,
                                             MPI_Datatype recvtype, int root, MPIR_Comm * comm,
                                             MPI_Request * req)
 {
-    return MPIDI_NM_igatherv(sendbuf, sendcount, sendtype, recvbuf,
-                             recvcounts, displs, recvtype, root, comm, req);
+    return MPIDI_NM_mpi_igatherv(sendbuf, sendcount, sendtype, recvbuf,
+                                 recvcounts, displs, recvtype, root, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ireduce_scatter_block(const void *sendbuf, void *recvbuf,
@@ -353,28 +355,28 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ireduce_scatter_block(const void *sendbuf, vo
                                                          MPI_Op op, MPIR_Comm * comm,
                                                          MPI_Request * req)
 {
-    return MPIDI_NM_ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm, req);
+    return MPIDI_NM_mpi_ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ireduce_scatter(const void *sendbuf, void *recvbuf,
                                                    const int *recvcounts, MPI_Datatype datatype,
                                                    MPI_Op op, MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm, req);
+    return MPIDI_NM_mpi_ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Ireduce(const void *sendbuf, void *recvbuf, int count,
                                            MPI_Datatype datatype, MPI_Op op, int root,
                                            MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, req);
+    return MPIDI_NM_mpi_ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iscan(const void *sendbuf, void *recvbuf, int count,
                                          MPI_Datatype datatype, MPI_Op op, MPIR_Comm * comm,
                                          MPI_Request * req)
 {
-    return MPIDI_NM_iscan(sendbuf, recvbuf, count, datatype, op, comm, req);
+    return MPIDI_NM_mpi_iscan(sendbuf, recvbuf, count, datatype, op, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iscatter(const void *sendbuf, int sendcount,
@@ -382,8 +384,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Iscatter(const void *sendbuf, int sendcount,
                                             MPI_Datatype recvtype, int root, MPIR_Comm * comm,
                                             MPI_Request * req)
 {
-    return MPIDI_NM_iscatter(sendbuf, sendcount, sendtype, recvbuf,
-                             recvcount, recvtype, root, comm, req);
+    return MPIDI_NM_mpi_iscatter(sendbuf, sendcount, sendtype, recvbuf,
+                                 recvcount, recvtype, root, comm, req);
 }
 
 MPL_STATIC_INLINE_PREFIX int MPIDI_Iscatterv(const void *sendbuf, const int *sendcounts,
@@ -391,8 +393,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Iscatterv(const void *sendbuf, const int *sen
                                              void *recvbuf, int recvcount, MPI_Datatype recvtype,
                                              int root, MPIR_Comm * comm, MPI_Request * req)
 {
-    return MPIDI_NM_iscatterv(sendbuf, sendcounts, displs, sendtype,
-                              recvbuf, recvcount, recvtype, root, comm, req);
+    return MPIDI_NM_mpi_iscatterv(sendbuf, sendcounts, displs, sendtype,
+                                  recvbuf, recvcount, recvtype, root, comm, req);
 }
 
 #endif /* CH4_COLL_H_INCLUDED */
diff --git a/src/mpid/ch4/src/ch4_comm.h b/src/mpid/ch4/src/ch4_comm.h
index f849329..6ae26b1 100644
--- a/src/mpid/ch4/src/ch4_comm.h
+++ b/src/mpid/ch4/src/ch4_comm.h
@@ -106,12 +106,12 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Comm_create_hook(MPIR_Comm * comm)
     int max_n_avts;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_COMM_CREATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_COMM_CREATE);
-    mpi_errno = MPIDI_NM_comm_create_hook(comm);
+    mpi_errno = MPIDI_NM_mpi_comm_create_hook(comm);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
 #if defined(MPIDI_BUILD_CH4_SHM)
-    mpi_errno = MPIDI_SHM_comm_create_hook(comm);
+    mpi_errno = MPIDI_SHM_mpi_comm_create_hook(comm);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -217,12 +217,12 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Comm_free_hook(MPIR_Comm * comm)
         MPIDIU_avt_release_ref(MPIDII_COMM(comm, local_map).avtid);
     }
 
-    mpi_errno = MPIDI_NM_comm_free_hook(comm);
+    mpi_errno = MPIDI_NM_mpi_comm_free_hook(comm);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
 #if defined(MPIDI_BUILD_CH4_SHM)
-    mpi_errno = MPIDI_SHM_comm_free_hook(comm);
+    mpi_errno = MPIDI_SHM_mpi_comm_free_hook(comm);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
diff --git a/src/mpid/ch4/src/ch4_init.h b/src/mpid/ch4/src/ch4_init.h
index ca88320..6613e82 100644
--- a/src/mpid/ch4/src/ch4_init.h
+++ b/src/mpid/ch4/src/ch4_init.h
@@ -248,9 +248,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Init(int *argc,
     MPIR_Process.attrs.tag_ub = (1ULL << MPIDI_CH4U_TAG_SHIFT) - 1;
     /* discuss */
 
-    mpi_errno = MPIDI_NM_init(rank, size, appnum, &MPIR_Process.attrs.tag_ub,
-                              MPIR_Process.comm_world,
-                              MPIR_Process.comm_self, has_parent, 1, &netmod_contexts);
+    mpi_errno = MPIDI_NM_mpi_init(rank, size, appnum, &MPIR_Process.attrs.tag_ub,
+                                  MPIR_Process.comm_world,
+                                  MPIR_Process.comm_self, has_parent, 1, &netmod_contexts);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POPFATAL(mpi_errno);
     }
@@ -279,7 +279,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Init(int *argc,
 #endif
 
 #ifdef MPIDI_BUILD_CH4_SHM
-    mpi_errno = MPIDI_SHM_init(rank, size);
+    mpi_errno = MPIDI_SHM_mpi_init(rank, size);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POPFATAL(mpi_errno);
@@ -346,11 +346,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Finalize(void)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_FINALIZE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_FINALIZE);
 
-    mpi_errno = MPIDI_NM_finalize();
+    mpi_errno = MPIDI_NM_mpi_finalize();
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 #ifdef MPIDI_BUILD_CH4_SHM
-    mpi_errno = MPIDI_SHM_finalize();
+    mpi_errno = MPIDI_SHM_mpi_finalize();
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 #endif
@@ -502,7 +502,7 @@ MPL_STATIC_INLINE_PREFIX void *MPIDI_Alloc_mem(size_t size, MPIR_Info * info_ptr
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_ALLOC_MEM);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_ALLOC_MEM);
 
-    p = MPIDI_NM_alloc_mem(size, info_ptr);
+    p = MPIDI_NM_mpi_alloc_mem(size, info_ptr);
 
     MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_CH4_ALLOC_MEM);
     return p;
@@ -517,7 +517,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Free_mem(void *ptr)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_FREE_MEM);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_FREE_MEM);
-    mpi_errno = MPIDI_NM_free_mem(ptr);
+    mpi_errno = MPIDI_NM_mpi_free_mem(ptr);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
@@ -761,13 +761,13 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Type_create_hook(MPIR_Datatype * type)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_TYPE_CREATE_HOOK);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_TYPE_CREATE_HOOK);
 
-    mpi_errno = MPIDI_NM_type_create_hook(type);
+    mpi_errno = MPIDI_NM_mpi_type_create_hook(type);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
 
 #if defined(MPIDI_BUILD_CH4_SHM)
-    mpi_errno = MPIDI_SHM_type_create_hook(type);
+    mpi_errno = MPIDI_SHM_mpi_type_create_hook(type);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -791,13 +791,13 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Type_free_hook(MPIR_Datatype * type)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_TYPE_FREE_HOOK);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_TYPE_FREE_HOOK);
 
-    mpi_errno = MPIDI_NM_type_free_hook(type);
+    mpi_errno = MPIDI_NM_mpi_type_free_hook(type);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
 
 #if defined(MPIDI_BUILD_CH4_SHM)
-    mpi_errno = MPIDI_SHM_type_free_hook(type);
+    mpi_errno = MPIDI_SHM_mpi_type_free_hook(type);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -821,13 +821,13 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Op_create_hook(MPIR_Op * op)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_OP_CREATE_HOOK);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_OP_CREATE_HOOK);
 
-    mpi_errno = MPIDI_NM_op_create_hook(op);
+    mpi_errno = MPIDI_NM_mpi_op_create_hook(op);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
 
 #if defined(MPIDI_BUILD_CH4_SHM)
-    mpi_errno = MPIDI_SHM_op_create_hook(op);
+    mpi_errno = MPIDI_SHM_mpi_op_create_hook(op);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -851,13 +851,13 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Op_free_hook(MPIR_Op * op)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_OP_FREE_HOOK);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_OP_FREE_HOOK);
 
-    mpi_errno = MPIDI_NM_op_free_hook(op);
+    mpi_errno = MPIDI_NM_mpi_op_free_hook(op);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
 
 #if defined(MPIDI_BUILD_CH4_SHM)
-    mpi_errno = MPIDI_SHM_op_free_hook(op);
+    mpi_errno = MPIDI_SHM_mpi_op_free_hook(op);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
diff --git a/src/mpid/ch4/src/ch4_probe.h b/src/mpid/ch4/src/ch4_probe.h
index 2e43468..9a44e77 100644
--- a/src/mpid/ch4/src/ch4_probe.h
+++ b/src/mpid/ch4/src/ch4_probe.h
@@ -33,17 +33,17 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Probe(int source,
 
     while (!flag) {
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-        mpi_errno = MPIDI_NM_iprobe(source, tag, comm, context_offset, &flag, status);
+        mpi_errno = MPIDI_NM_mpi_iprobe(source, tag, comm, context_offset, &flag, status);
 #else
         if (unlikely(source == MPI_ANY_SOURCE)) {
-            mpi_errno = MPIDI_SHM_iprobe(source, tag, comm, context_offset, &flag, status);
+            mpi_errno = MPIDI_SHM_mpi_iprobe(source, tag, comm, context_offset, &flag, status);
             if (!flag)
-                mpi_errno = MPIDI_NM_iprobe(source, tag, comm, context_offset, &flag, status);
+                mpi_errno = MPIDI_NM_mpi_iprobe(source, tag, comm, context_offset, &flag, status);
         }
         else if (MPIDI_CH4_rank_is_local(source, comm))
-            mpi_errno = MPIDI_SHM_iprobe(source, tag, comm, context_offset, &flag, status);
+            mpi_errno = MPIDI_SHM_mpi_iprobe(source, tag, comm, context_offset, &flag, status);
         else
-            mpi_errno = MPIDI_NM_iprobe(source, tag, comm, context_offset, &flag, status);
+            mpi_errno = MPIDI_NM_mpi_iprobe(source, tag, comm, context_offset, &flag, status);
 #endif
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
@@ -92,20 +92,21 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Mprobe(int source,
 
     while (!flag) {
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-        mpi_errno = MPIDI_NM_improbe(source, tag, comm, context_offset, &flag, message, status);
+        mpi_errno = MPIDI_NM_mpi_improbe(source, tag, comm, context_offset, &flag, message, status);
 #else
         if (unlikely(source == MPI_ANY_SOURCE)) {
             mpi_errno =
-                MPIDI_SHM_improbe(source, tag, comm, context_offset, &flag, message, status);
+                MPIDI_SHM_mpi_improbe(source, tag, comm, context_offset, &flag, message, status);
             if (!flag)
                 mpi_errno =
-                    MPIDI_NM_improbe(source, tag, comm, context_offset, &flag, message, status);
+                    MPIDI_NM_mpi_improbe(source, tag, comm, context_offset, &flag, message, status);
         }
         else if (MPIDI_CH4_rank_is_local(source, comm))
             mpi_errno =
-                MPIDI_SHM_improbe(source, tag, comm, context_offset, &flag, message, status);
+                MPIDI_SHM_mpi_improbe(source, tag, comm, context_offset, &flag, message, status);
         else
-            mpi_errno = MPIDI_NM_improbe(source, tag, comm, context_offset, &flag, message, status);
+            mpi_errno =
+                MPIDI_NM_mpi_improbe(source, tag, comm, context_offset, &flag, message, status);
 #endif
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
@@ -153,17 +154,18 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Improbe(int source,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_improbe(source, tag, comm, context_offset, flag, message, status);
+    mpi_errno = MPIDI_NM_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
 #else
     if (unlikely(source == MPI_ANY_SOURCE)) {
-        mpi_errno = MPIDI_SHM_improbe(source, tag, comm, context_offset, flag, message, status);
+        mpi_errno = MPIDI_SHM_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
         if (!*flag)
-            mpi_errno = MPIDI_NM_improbe(source, tag, comm, context_offset, flag, message, status);
+            mpi_errno =
+                MPIDI_NM_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
     }
     else if (MPIDI_CH4_rank_is_local(source, comm))
-        mpi_errno = MPIDI_SHM_improbe(source, tag, comm, context_offset, flag, message, status);
+        mpi_errno = MPIDI_SHM_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
     else
-        mpi_errno = MPIDI_NM_improbe(source, tag, comm, context_offset, flag, message, status);
+        mpi_errno = MPIDI_NM_mpi_improbe(source, tag, comm, context_offset, flag, message, status);
 #endif
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
@@ -197,17 +199,17 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Iprobe(int source,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_iprobe(source, tag, comm, context_offset, flag, status);
+    mpi_errno = MPIDI_NM_mpi_iprobe(source, tag, comm, context_offset, flag, status);
 #else
     if (unlikely(source == MPI_ANY_SOURCE)) {
-        mpi_errno = MPIDI_SHM_iprobe(source, tag, comm, context_offset, flag, status);
+        mpi_errno = MPIDI_SHM_mpi_iprobe(source, tag, comm, context_offset, flag, status);
         if (!*flag)
-            mpi_errno = MPIDI_NM_iprobe(source, tag, comm, context_offset, flag, status);
+            mpi_errno = MPIDI_NM_mpi_iprobe(source, tag, comm, context_offset, flag, status);
     }
     else if (MPIDI_CH4_rank_is_local(source, comm))
-        mpi_errno = MPIDI_SHM_iprobe(source, tag, comm, context_offset, flag, status);
+        mpi_errno = MPIDI_SHM_mpi_iprobe(source, tag, comm, context_offset, flag, status);
     else
-        mpi_errno = MPIDI_NM_iprobe(source, tag, comm, context_offset, flag, status);
+        mpi_errno = MPIDI_NM_mpi_iprobe(source, tag, comm, context_offset, flag, status);
 #endif
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
diff --git a/src/mpid/ch4/src/ch4_recv.h b/src/mpid/ch4/src/ch4_recv.h
index 74aafb3..a660536 100644
--- a/src/mpid/ch4/src/ch4_recv.h
+++ b/src/mpid/ch4/src/ch4_recv.h
@@ -43,17 +43,18 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Recv(void *buf,
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
     mpi_errno =
-        MPIDI_NM_recv(buf, count, datatype, rank, tag, comm, context_offset, status, request);
+        MPIDI_NM_mpi_recv(buf, count, datatype, rank, tag, comm, context_offset, status, request);
 #else
     if (unlikely(rank == MPI_ANY_SOURCE)) {
-        mpi_errno = MPIDI_SHM_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
         }
 
-        mpi_errno = MPIDI_NM_irecv(buf, count, datatype, rank, tag, comm, context_offset,
-                                   &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)));
+        mpi_errno = MPIDI_NM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset,
+                                       &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)));
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
@@ -61,7 +62,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Recv(void *buf,
         /* cancel the shm request if netmod/am handles the request from unexpected queue. */
         else if (*request) {
             if (MPIR_Request_is_complete(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request))) {
-                mpi_errno = MPIDI_SHM_cancel_recv(*request);
+                mpi_errno = MPIDI_SHM_mpi_cancel_recv(*request);
                 if (MPIR_STATUS_GET_CANCEL_BIT((*request)->status)) {
                     (*request)->status = MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)->status;
                 }
@@ -79,12 +80,12 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Recv(void *buf,
         int r;
         if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
             mpi_errno =
-                MPIDI_SHM_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
-                               request);
+                MPIDI_SHM_mpi_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
+                                   request);
         else
             mpi_errno =
-                MPIDI_NM_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
-                              request);
+                MPIDI_NM_mpi_recv(buf, count, datatype, rank, tag, comm, context_offset, status,
+                                  request);
         if (mpi_errno == MPI_SUCCESS && *request) {
             MPIDI_CH4I_REQUEST(*request, is_local) = r;
             MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -118,20 +119,21 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Recv_init(void *buf,
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_RECV_INIT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_RECV_INIT);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno =
+        MPIDI_NM_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
     MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_CH4_RECV_INIT);
     return mpi_errno;
 #else
     if (unlikely(rank == MPI_ANY_SOURCE)) {
         mpi_errno =
-            MPIDI_SHM_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+            MPIDI_SHM_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
         }
 
-        mpi_errno = MPIDI_NM_recv_init(buf, count, datatype, rank, tag, comm, context_offset,
-                                       &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)));
+        mpi_errno = MPIDI_NM_mpi_recv_init(buf, count, datatype, rank, tag, comm, context_offset,
+                                           &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)));
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
@@ -146,11 +148,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Recv_init(void *buf,
     else {
         int r;
         if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-            mpi_errno = MPIDI_SHM_recv_init(buf, count, datatype, rank, tag,
-                                            comm, context_offset, request);
+            mpi_errno = MPIDI_SHM_mpi_recv_init(buf, count, datatype, rank, tag,
+                                                comm, context_offset, request);
         else
-            mpi_errno = MPIDI_NM_recv_init(buf, count, datatype, rank, tag,
-                                           comm, context_offset, request);
+            mpi_errno = MPIDI_NM_mpi_recv_init(buf, count, datatype, rank, tag,
+                                               comm, context_offset, request);
         if (mpi_errno == MPI_SUCCESS) {
             MPIDI_CH4I_REQUEST(*request, is_local) = r;
             MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -200,19 +202,19 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Mrecv(void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_imrecv(buf, count, datatype, message, &rreq);
+    mpi_errno = MPIDI_NM_mpi_imrecv(buf, count, datatype, message, &rreq);
 #else
 
     if (unlikely(message->status.MPI_SOURCE == MPI_ANY_SOURCE)) {
-        mpi_errno = MPIDI_SHM_imrecv(buf, count, datatype, message, &rreq);
+        mpi_errno = MPIDI_SHM_mpi_imrecv(buf, count, datatype, message, &rreq);
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
         }
 
         mpi_errno =
-            MPIDI_NM_imrecv(buf, count, datatype, message,
-                            &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(rreq)));
+            MPIDI_NM_mpi_imrecv(buf, count, datatype, message,
+                                &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(rreq)));
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
@@ -225,9 +227,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Mrecv(void *buf,
     else {
         int local = MPIDI_CH4_rank_is_local(message->status.MPI_SOURCE, message->comm);
         if (local)
-            mpi_errno = MPIDI_SHM_imrecv(buf, count, datatype, message, &rreq);
+            mpi_errno = MPIDI_SHM_mpi_imrecv(buf, count, datatype, message, &rreq);
         else
-            mpi_errno = MPIDI_NM_imrecv(buf, count, datatype, message, &rreq);
+            mpi_errno = MPIDI_NM_mpi_imrecv(buf, count, datatype, message, &rreq);
 
         if (mpi_errno == MPI_SUCCESS) {
             MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(rreq) = NULL;
@@ -287,14 +289,14 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Imrecv(void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_imrecv(buf, count, datatype, message, rreqp);
+    mpi_errno = MPIDI_NM_mpi_imrecv(buf, count, datatype, message, rreqp);
 #else
     {
         int local = MPIDI_CH4_rank_is_local(message->status.MPI_SOURCE, message->comm);
         if (local)
-            mpi_errno = MPIDI_SHM_imrecv(buf, count, datatype, message, rreqp);
+            mpi_errno = MPIDI_SHM_mpi_imrecv(buf, count, datatype, message, rreqp);
         else
-            mpi_errno = MPIDI_NM_imrecv(buf, count, datatype, message, rreqp);
+            mpi_errno = MPIDI_NM_mpi_imrecv(buf, count, datatype, message, rreqp);
         if (mpi_errno == MPI_SUCCESS) {
             MPIDI_CH4I_REQUEST(*rreqp, is_local) = local;
             MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*rreqp) = NULL;
@@ -339,17 +341,18 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Irecv(void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     if (unlikely(rank == MPI_ANY_SOURCE)) {
-        mpi_errno = MPIDI_SHM_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
         }
 
-        mpi_errno = MPIDI_NM_irecv(buf, count, datatype, rank, tag, comm, context_offset,
-                                   &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)));
+        mpi_errno = MPIDI_NM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset,
+                                       &(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request)));
 
         if (mpi_errno != MPI_SUCCESS) {
             MPIR_ERR_POP(mpi_errno);
@@ -366,10 +369,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Irecv(void *buf,
         int r;
         if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
             mpi_errno =
-                MPIDI_SHM_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+                MPIDI_SHM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
         else
             mpi_errno =
-                MPIDI_NM_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
+                MPIDI_NM_mpi_irecv(buf, count, datatype, rank, tag, comm, context_offset, request);
         if (mpi_errno == MPI_SUCCESS && *request) {
             MPIDI_CH4I_REQUEST(*request, is_local) = r;
             MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -396,12 +399,12 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Cancel_recv(MPIR_Request * rreq)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_CANCEL_RECV);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_CANCEL_RECV);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_cancel_recv(rreq);
+    mpi_errno = MPIDI_NM_mpi_cancel_recv(rreq);
 #else
     if (MPIDI_CH4I_REQUEST(rreq, is_local))
-        mpi_errno = MPIDI_SHM_cancel_recv(rreq);
+        mpi_errno = MPIDI_SHM_mpi_cancel_recv(rreq);
     else
-        mpi_errno = MPIDI_NM_cancel_recv(rreq);
+        mpi_errno = MPIDI_NM_mpi_cancel_recv(rreq);
 #endif
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
diff --git a/src/mpid/ch4/src/ch4_rma.h b/src/mpid/ch4/src/ch4_rma.h
index 7eabd9d..6534ee0 100644
--- a/src/mpid/ch4/src/ch4_rma.h
+++ b/src/mpid/ch4/src/ch4_rma.h
@@ -28,8 +28,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Put(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_PUT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_PUT);
-    mpi_errno = MPIDI_NM_put(origin_addr, origin_count, origin_datatype,
-                             target_rank, target_disp, target_count, target_datatype, win);
+    mpi_errno = MPIDI_NM_mpi_put(origin_addr, origin_count, origin_datatype,
+                                 target_rank, target_disp, target_count, target_datatype, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -55,8 +55,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Get(void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_GET);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_GET);
-    mpi_errno = MPIDI_NM_get(origin_addr, origin_count, origin_datatype,
-                             target_rank, target_disp, target_count, target_datatype, win);
+    mpi_errno = MPIDI_NM_mpi_get(origin_addr, origin_count, origin_datatype,
+                                 target_rank, target_disp, target_count, target_datatype, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -83,9 +83,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Accumulate(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_ACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_ACCUMULATE);
-    mpi_errno = MPIDI_NM_accumulate(origin_addr, origin_count, origin_datatype,
-                                    target_rank, target_disp, target_count,
-                                    target_datatype, op, win);
+    mpi_errno = MPIDI_NM_mpi_accumulate(origin_addr, origin_count, origin_datatype,
+                                        target_rank, target_disp, target_count,
+                                        target_datatype, op, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -110,8 +110,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Compare_and_swap(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_COMPARE_AND_SWAP);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_COMPARE_AND_SWAP);
-    mpi_errno = MPIDI_NM_compare_and_swap(origin_addr, compare_addr, result_addr,
-                                          datatype, target_rank, target_disp, win);
+    mpi_errno = MPIDI_NM_mpi_compare_and_swap(origin_addr, compare_addr, result_addr,
+                                              datatype, target_rank, target_disp, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -138,9 +138,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Raccumulate(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_RACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_RACCUMULATE);
-    mpi_errno = MPIDI_NM_raccumulate(origin_addr, origin_count, origin_datatype,
-                                     target_rank, target_disp, target_count,
-                                     target_datatype, op, win, request);
+    mpi_errno = MPIDI_NM_mpi_raccumulate(origin_addr, origin_count, origin_datatype,
+                                         target_rank, target_disp, target_count,
+                                         target_datatype, op, win, request);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -171,10 +171,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Rget_accumulate(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_RGET_ACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_RGET_ACCUMULATE);
-    mpi_errno = MPIDI_NM_rget_accumulate(origin_addr, origin_count, origin_datatype,
-                                         result_addr, result_count, result_datatype,
-                                         target_rank, target_disp, target_count,
-                                         target_datatype, op, win, request);
+    mpi_errno = MPIDI_NM_mpi_rget_accumulate(origin_addr, origin_count, origin_datatype,
+                                             result_addr, result_count, result_datatype,
+                                             target_rank, target_disp, target_count,
+                                             target_datatype, op, win, request);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -198,8 +198,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Fetch_and_op(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_FETCH_AND_OP);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_FETCH_AND_OP);
-    mpi_errno = MPIDI_NM_fetch_and_op(origin_addr, result_addr,
-                                      datatype, target_rank, target_disp, op, win);
+    mpi_errno = MPIDI_NM_mpi_fetch_and_op(origin_addr, result_addr,
+                                          datatype, target_rank, target_disp, op, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -227,9 +227,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Rget(void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_RGET);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_RGET);
-    mpi_errno = MPIDI_NM_rget(origin_addr, origin_count, origin_datatype,
-                              target_rank, target_disp, target_count,
-                              target_datatype, win, request);
+    mpi_errno = MPIDI_NM_mpi_rget(origin_addr, origin_count, origin_datatype,
+                                  target_rank, target_disp, target_count,
+                                  target_datatype, win, request);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -256,9 +256,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Rput(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_RPUT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_RPUT);
-    mpi_errno = MPIDI_NM_rput(origin_addr, origin_count, origin_datatype,
-                              target_rank, target_disp, target_count,
-                              target_datatype, win, request);
+    mpi_errno = MPIDI_NM_mpi_rput(origin_addr, origin_count, origin_datatype,
+                                  target_rank, target_disp, target_count,
+                                  target_datatype, win, request);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -288,10 +288,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Get_accumulate(const void *origin_addr,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_GET_ACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_GET_ACCUMULATE);
-    mpi_errno = MPIDI_NM_get_accumulate(origin_addr, origin_count, origin_datatype,
-                                        result_addr, result_count, result_datatype,
-                                        target_rank, target_disp, target_count, target_datatype,
-                                        op, win);
+    mpi_errno = MPIDI_NM_mpi_get_accumulate(origin_addr, origin_count, origin_datatype,
+                                            result_addr, result_count, result_datatype,
+                                            target_rank, target_disp, target_count, target_datatype,
+                                            op, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
diff --git a/src/mpid/ch4/src/ch4_send.h b/src/mpid/ch4/src/ch4_send.h
index c719115..ac5c9e9 100644
--- a/src/mpid/ch4/src/ch4_send.h
+++ b/src/mpid/ch4/src/ch4_send.h
@@ -39,13 +39,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Send(const void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_NM_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS && *request)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
 #endif
@@ -85,13 +87,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Isend(const void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_NM_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
 #endif
@@ -132,13 +136,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Rsend(const void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_send(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_NM_mpi_send(buf, count, datatype, rank, tag, comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS && *request)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
 #endif
@@ -179,13 +185,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Irsend(const void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_NM_mpi_isend(buf, count, datatype, rank, tag, comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
 #endif
@@ -225,13 +233,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ssend(const void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_SHM_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_NM_mpi_ssend(buf, count, datatype, rank, tag, comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS && *request)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
 #endif
@@ -271,14 +281,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Issend(const void *buf,
     }
 
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno = MPIDI_NM_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
         mpi_errno =
-            MPIDI_SHM_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+            MPIDI_SHM_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
+        mpi_errno =
+            MPIDI_NM_mpi_issend(buf, count, datatype, rank, tag, comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
 #endif
@@ -302,16 +313,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Startall(int count, MPIR_Request * requests[]
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_STARTALL);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_STARTALL);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_startall(count, requests);
+    mpi_errno = MPIDI_NM_mpi_startall(count, requests);
 #else
     int i;
     for (i = 0; i < count; i++) {
         /* This is sub-optimal, can we do better? */
         if (MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(requests[i])) {
-            mpi_errno = MPIDI_SHM_startall(1, &requests[i]);
+            mpi_errno = MPIDI_SHM_mpi_startall(1, &requests[i]);
             if (mpi_errno == MPI_SUCCESS) {
                 mpi_errno =
-                    MPIDI_NM_startall(1, &MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(requests[i]));
+                    MPIDI_NM_mpi_startall(1, &MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(requests[i]));
                 MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(requests[i]->u.persist.real_request) =
                     MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(requests[i])->u.persist.real_request;
                 MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER
@@ -320,9 +331,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Startall(int count, MPIR_Request * requests[]
             }
         }
         else if (MPIDI_CH4I_REQUEST(requests[i], is_local))
-            mpi_errno = MPIDI_SHM_startall(1, &requests[i]);
+            mpi_errno = MPIDI_SHM_mpi_startall(1, &requests[i]);
         else
-            mpi_errno = MPIDI_NM_startall(1, &requests[i]);
+            mpi_errno = MPIDI_NM_mpi_startall(1, &requests[i]);
     }
 #endif
     if (mpi_errno != MPI_SUCCESS) {
@@ -351,15 +362,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Send_init(const void *buf,
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_SEND_INIT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_SEND_INIT);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno =
+        MPIDI_NM_mpi_send_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_send_init(buf, count, datatype, rank, tag,
-                                        comm, context_offset, request);
+        mpi_errno = MPIDI_SHM_mpi_send_init(buf, count, datatype, rank, tag,
+                                            comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_send_init(buf, count, datatype, rank, tag,
-                                       comm, context_offset, request);
+        mpi_errno = MPIDI_NM_mpi_send_init(buf, count, datatype, rank, tag,
+                                           comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS)
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
     MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -390,15 +402,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Ssend_init(const void *buf,
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_SSEND_INIT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_SSEND_INIT);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno =
+        MPIDI_NM_mpi_ssend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_ssend_init(buf, count, datatype, rank, tag,
-                                         comm, context_offset, request);
+        mpi_errno = MPIDI_SHM_mpi_ssend_init(buf, count, datatype, rank, tag,
+                                             comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_ssend_init(buf, count, datatype, rank, tag,
-                                        comm, context_offset, request);
+        mpi_errno = MPIDI_NM_mpi_ssend_init(buf, count, datatype, rank, tag,
+                                            comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS && *request) {
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
         MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -430,15 +443,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Bsend_init(const void *buf,
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_BSEND_INIT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_BSEND_INIT);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno =
+        MPIDI_NM_mpi_bsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_bsend_init(buf, count, datatype, rank, tag,
-                                         comm, context_offset, request);
+        mpi_errno = MPIDI_SHM_mpi_bsend_init(buf, count, datatype, rank, tag,
+                                             comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_bsend_init(buf, count, datatype, rank, tag,
-                                        comm, context_offset, request);
+        mpi_errno = MPIDI_NM_mpi_bsend_init(buf, count, datatype, rank, tag,
+                                            comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS && *request) {
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
         MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -470,15 +484,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Rsend_init(const void *buf,
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_RSEND_INIT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_RSEND_INIT);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
+    mpi_errno =
+        MPIDI_NM_mpi_rsend_init(buf, count, datatype, rank, tag, comm, context_offset, request);
 #else
     int r;
     if ((r = MPIDI_CH4_rank_is_local(rank, comm)))
-        mpi_errno = MPIDI_SHM_rsend_init(buf, count, datatype, rank, tag,
-                                         comm, context_offset, request);
+        mpi_errno = MPIDI_SHM_mpi_rsend_init(buf, count, datatype, rank, tag,
+                                             comm, context_offset, request);
     else
-        mpi_errno = MPIDI_NM_rsend_init(buf, count, datatype, rank, tag,
-                                        comm, context_offset, request);
+        mpi_errno = MPIDI_NM_mpi_rsend_init(buf, count, datatype, rank, tag,
+                                            comm, context_offset, request);
     if (mpi_errno == MPI_SUCCESS && *request) {
         MPIDI_CH4I_REQUEST(*request, is_local) = r;
         MPIDI_CH4I_REQUEST_ANYSOURCE_PARTNER(*request) = NULL;
@@ -505,12 +520,12 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Cancel_send(MPIR_Request * sreq)
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_CANCEL_SEND);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_CANCEL_SEND);
 #ifndef MPIDI_CH4_EXCLUSIVE_SHM
-    mpi_errno = MPIDI_NM_cancel_send(sreq);
+    mpi_errno = MPIDI_NM_mpi_cancel_send(sreq);
 #else
     if (MPIDI_CH4I_REQUEST(sreq, is_local))
-        mpi_errno = MPIDI_SHM_cancel_send(sreq);
+        mpi_errno = MPIDI_SHM_mpi_cancel_send(sreq);
     else
-        mpi_errno = MPIDI_NM_cancel_send(sreq);
+        mpi_errno = MPIDI_NM_mpi_cancel_send(sreq);
 #endif
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
diff --git a/src/mpid/ch4/src/ch4_spawn.h b/src/mpid/ch4/src/ch4_spawn.h
index c7d7921..c92c0bf 100644
--- a/src/mpid/ch4/src/ch4_spawn.h
+++ b/src/mpid/ch4/src/ch4_spawn.h
@@ -214,7 +214,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Comm_connect(const char *port_name,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_COMM_CONNECT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_COMM_CONNECT);
-    mpi_errno = MPIDI_NM_comm_connect(port_name, info, root, comm, newcomm_ptr);
+    mpi_errno = MPIDI_NM_mpi_comm_connect(port_name, info, root, comm, newcomm_ptr);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
@@ -236,7 +236,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Comm_disconnect(MPIR_Comm * comm_ptr)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_COMM_DISCONNECT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_COMM_DISCONNECT);
-    mpi_errno = MPIDI_NM_comm_disconnect(comm_ptr);
+    mpi_errno = MPIDI_NM_mpi_comm_disconnect(comm_ptr);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
@@ -258,7 +258,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Open_port(MPIR_Info * info_ptr, char *port_na
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_OPEN_PORT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_OPEN_PORT);
-    mpi_errno = MPIDI_NM_open_port(info_ptr, port_name);
+    mpi_errno = MPIDI_NM_mpi_open_port(info_ptr, port_name);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
@@ -280,7 +280,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Close_port(const char *port_name)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_CLOSE_PORT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_CLOSE_PORT);
-    mpi_errno = MPIDI_NM_close_port(port_name);
+    mpi_errno = MPIDI_NM_mpi_close_port(port_name);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
@@ -304,7 +304,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Comm_accept(const char *port_name,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_COMM_ACCEPT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_COMM_ACCEPT);
-    mpi_errno = MPIDI_NM_comm_accept(port_name, info, root, comm, newcomm_ptr);
+    mpi_errno = MPIDI_NM_mpi_comm_accept(port_name, info, root, comm, newcomm_ptr);
 
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
diff --git a/src/mpid/ch4/src/ch4_win.h b/src/mpid/ch4/src/ch4_win.h
index 12f899d..9e0861f 100644
--- a/src/mpid/ch4/src/ch4_win.h
+++ b/src/mpid/ch4/src/ch4_win.h
@@ -22,7 +22,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_set_info(MPIR_Win * win, MPIR_Info * info
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_SET_INFO);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_SET_INFO);
-    mpi_errno = MPIDI_NM_win_set_info(win, info);
+    mpi_errno = MPIDI_NM_mpi_win_set_info(win, info);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -42,7 +42,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_start(MPIR_Group * group, int assert, MPI
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_START);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_START);
-    mpi_errno = MPIDI_NM_win_start(group, assert, win);
+    mpi_errno = MPIDI_NM_mpi_win_start(group, assert, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -62,7 +62,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_complete(MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_COMPLETE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_COMPLETE);
-    mpi_errno = MPIDI_NM_win_complete(win);
+    mpi_errno = MPIDI_NM_mpi_win_complete(win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -82,7 +82,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_post(MPIR_Group * group, int assert, MPIR
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_POST);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_POST);
-    mpi_errno = MPIDI_NM_win_post(group, assert, win);
+    mpi_errno = MPIDI_NM_mpi_win_post(group, assert, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -102,7 +102,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_wait(MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_WAIT);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_WAIT);
-    mpi_errno = MPIDI_NM_win_wait(win);
+    mpi_errno = MPIDI_NM_mpi_win_wait(win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -123,7 +123,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_test(MPIR_Win * win, int *flag)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_TEST);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_TEST);
-    mpi_errno = MPIDI_NM_win_test(win, flag);
+    mpi_errno = MPIDI_NM_mpi_win_test(win, flag);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -143,7 +143,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_lock(int lock_type, int rank, int assert,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_LOCK);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_LOCK);
-    mpi_errno = MPIDI_NM_win_lock(lock_type, rank, assert, win);
+    mpi_errno = MPIDI_NM_mpi_win_lock(lock_type, rank, assert, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -163,7 +163,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_unlock(int rank, MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_UNLOCK);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_UNLOCK);
-    mpi_errno = MPIDI_NM_win_unlock(rank, win);
+    mpi_errno = MPIDI_NM_mpi_win_unlock(rank, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -183,7 +183,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_get_info(MPIR_Win * win, MPIR_Info ** inf
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_GET_INFO);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_GET_INFO);
-    mpi_errno = MPIDI_NM_win_get_info(win, info_p_p);
+    mpi_errno = MPIDI_NM_mpi_win_get_info(win, info_p_p);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -203,7 +203,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_free(MPIR_Win ** win_ptr)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_FREE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_FREE);
-    mpi_errno = MPIDI_NM_win_free(win_ptr);
+    mpi_errno = MPIDI_NM_mpi_win_free(win_ptr);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -223,7 +223,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_fence(int assert, MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_FENCE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_FENCE);
-    mpi_errno = MPIDI_NM_win_fence(assert, win);
+    mpi_errno = MPIDI_NM_mpi_win_fence(assert, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -247,7 +247,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_create(void *base,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_CREATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_CREATE);
-    mpi_errno = MPIDI_NM_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
+    mpi_errno = MPIDI_NM_mpi_win_create(base, length, disp_unit, info, comm_ptr, win_ptr);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -267,7 +267,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_attach(MPIR_Win * win, void *base, MPI_Ai
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_ATTACH);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_ATTACH);
-    mpi_errno = MPIDI_NM_win_attach(win, base, size);
+    mpi_errno = MPIDI_NM_mpi_win_attach(win, base, size);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -291,8 +291,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_allocate_shared(MPI_Aint size,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_ALLOCATE_SHARED);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_ALLOCATE_SHARED);
-    mpi_errno = MPIDI_NM_win_allocate_shared(size, disp_unit,
-                                             info_ptr, comm_ptr, base_ptr, win_ptr);
+    mpi_errno = MPIDI_NM_mpi_win_allocate_shared(size, disp_unit,
+                                                 info_ptr, comm_ptr, base_ptr, win_ptr);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -312,7 +312,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_flush_local(int rank, MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_FLUSH_LOCAL);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_FLUSH_LOCAL);
-    mpi_errno = MPIDI_NM_win_flush_local(rank, win);
+    mpi_errno = MPIDI_NM_mpi_win_flush_local(rank, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -332,7 +332,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_detach(MPIR_Win * win, const void *base)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_DETACH);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_DETACH);
-    mpi_errno = MPIDI_NM_win_detach(win, base);
+    mpi_errno = MPIDI_NM_mpi_win_detach(win, base);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -355,7 +355,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_shared_query(MPIR_Win * win,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_SHARED_QUERY);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_SHARED_QUERY);
-    mpi_errno = MPIDI_NM_win_shared_query(win, rank, size, disp_unit, baseptr);
+    mpi_errno = MPIDI_NM_mpi_win_shared_query(win, rank, size, disp_unit, baseptr);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -378,7 +378,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_allocate(MPI_Aint size,
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_ALLOCATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_ALLOCATE);
-    mpi_errno = MPIDI_NM_win_allocate(size, disp_unit, info, comm, baseptr, win);
+    mpi_errno = MPIDI_NM_mpi_win_allocate(size, disp_unit, info, comm, baseptr, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -398,7 +398,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_flush(int rank, MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_FLUSH);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_FLUSH);
-    mpi_errno = MPIDI_NM_win_flush(rank, win);
+    mpi_errno = MPIDI_NM_mpi_win_flush(rank, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -418,7 +418,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_flush_local_all(MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_FLUSH_LOCAL_ALL);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_FLUSH_LOCAL_ALL);
-    mpi_errno = MPIDI_NM_win_flush_local_all(win);
+    mpi_errno = MPIDI_NM_mpi_win_flush_local_all(win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -438,7 +438,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_unlock_all(MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_UNLOCK_ALL);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_UNLOCK_ALL);
-    mpi_errno = MPIDI_NM_win_unlock_all(win);
+    mpi_errno = MPIDI_NM_mpi_win_unlock_all(win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -459,7 +459,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_create_dynamic(MPIR_Info * info, MPIR_Com
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_CREATE_DYNAMIC);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_CREATE_DYNAMIC);
-    mpi_errno = MPIDI_NM_win_create_dynamic(info, comm, win);
+    mpi_errno = MPIDI_NM_mpi_win_create_dynamic(info, comm, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -479,7 +479,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_sync(MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_SYNC);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_SYNC);
-    mpi_errno = MPIDI_NM_win_sync(win);
+    mpi_errno = MPIDI_NM_mpi_win_sync(win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -499,7 +499,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_flush_all(MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_FLUSH_ALL);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_FLUSH_ALL);
-    mpi_errno = MPIDI_NM_win_flush_all(win);
+    mpi_errno = MPIDI_NM_mpi_win_flush_all(win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
@@ -519,7 +519,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Win_lock_all(int assert, MPIR_Win * win)
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_WIN_LOCK_ALL);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_WIN_LOCK_ALL);
-    mpi_errno = MPIDI_NM_win_lock_all(assert, win);
+    mpi_errno = MPIDI_NM_mpi_win_lock_all(assert, win);
     if (mpi_errno != MPI_SUCCESS) {
         MPIR_ERR_POP(mpi_errno);
     }
diff --git a/src/mpid/ch4/src/ch4r_init.h b/src/mpid/ch4/src/ch4r_init.h
index 0089861..5861605 100644
--- a/src/mpid/ch4/src/ch4r_init.h
+++ b/src/mpid/ch4/src/ch4r_init.h
@@ -32,7 +32,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_init_comm(MPIR_Comm * comm)
     /*
      * Prevents double initialization of some special communicators.
      *
-     * comm_world and comm_self may exhibit this function twice, first during MPIDI_CH4U_init
+     * comm_world and comm_self may exhibit this function twice, first during MPIDI_CH4U_mpi_init
      * and the second during MPIR_Comm_commit in MPIDI_Init.
      * If there is an early arrival of an unexpected message before the second visit,
      * the following code will wipe out the unexpected queue andthe message is lost forever.
@@ -101,11 +101,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_destroy_comm(MPIR_Comm * comm)
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_init
+#define FUNCNAME MPIDI_CH4U_mpi_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_init(MPIR_Comm * comm_world, MPIR_Comm * comm_self,
-                                             int num_contexts, void **netmod_contexts)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_init(MPIR_Comm * comm_world, MPIR_Comm * comm_self,
+                                                 int num_contexts, void **netmod_contexts)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_INIT);
@@ -319,10 +319,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_init(MPIR_Comm * comm_world, MPIR_Comm *
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_init
+#define FUNCNAME MPIDI_CH4U_mpi_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX void MPIDI_CH4U_finalize()
+MPL_STATIC_INLINE_PREFIX void MPIDI_CH4U_mpi_finalize()
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_FINALIZE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4U_FINALIZE);
@@ -334,10 +334,10 @@ MPL_STATIC_INLINE_PREFIX void MPIDI_CH4U_finalize()
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_alloc_mem
+#define FUNCNAME MPIDI_CH4U_mpi_alloc_mem
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX void *MPIDI_CH4U_alloc_mem(size_t size, MPIR_Info * info_ptr)
+MPL_STATIC_INLINE_PREFIX void *MPIDI_CH4U_mpi_alloc_mem(size_t size, MPIR_Info * info_ptr)
 {
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_ALLOC_MEM);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4U_ALLOC_MEM);
@@ -348,10 +348,10 @@ MPL_STATIC_INLINE_PREFIX void *MPIDI_CH4U_alloc_mem(size_t size, MPIR_Info * inf
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_free_mem
+#define FUNCNAME MPIDI_CH4U_mpi_free_mem
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_free_mem(void *ptr)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_free_mem(void *ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_FREE_MEM);
diff --git a/src/mpid/ch4/src/ch4r_probe.h b/src/mpid/ch4/src/ch4r_probe.h
index 015a274..a473ddb 100644
--- a/src/mpid/ch4/src/ch4r_probe.h
+++ b/src/mpid/ch4/src/ch4r_probe.h
@@ -14,13 +14,14 @@
 #include "ch4_impl.h"
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_iprobe
+#define FUNCNAME MPIDI_CH4U_mpi_iprobe
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_iprobe(int source,
-                                               int tag,
-                                               MPIR_Comm * comm,
-                                               int context_offset, int *flag, MPI_Status * status)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_iprobe(int source,
+                                                   int tag,
+                                                   MPIR_Comm * comm,
+                                                   int context_offset, int *flag,
+                                                   MPI_Status * status)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Comm *root_comm;
@@ -79,7 +80,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_probe(int source,
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4U_PROBE);
 
     while (!flag) {
-        mpi_errno = MPIDI_CH4U_iprobe(source, tag, comm, context_offset, &flag, status);
+        mpi_errno = MPIDI_CH4U_mpi_iprobe(source, tag, comm, context_offset, &flag, status);
         if (mpi_errno)
             MPIR_ERR_POP(mpi_errno);
     }
@@ -92,15 +93,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_probe(int source,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_improbe
+#define FUNCNAME MPIDI_CH4U_mpi_improbe
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_improbe(int source,
-                                                int tag,
-                                                MPIR_Comm * comm,
-                                                int context_offset,
-                                                int *flag, MPIR_Request ** message,
-                                                MPI_Status * status)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_improbe(int source,
+                                                    int tag,
+                                                    MPIR_Comm * comm,
+                                                    int context_offset,
+                                                    int *flag, MPIR_Request ** message,
+                                                    MPI_Status * status)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Comm *root_comm;
@@ -169,7 +170,8 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mprobe(int source,
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4_MPROBE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4_MPROBE);
     while (!flag) {
-        mpi_errno = MPIDI_CH4U_improbe(source, tag, comm, context_offset, &flag, message, status);
+        mpi_errno =
+            MPIDI_CH4U_mpi_improbe(source, tag, comm, context_offset, &flag, message, status);
     }
     MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_CH4_MPROBE);
     return mpi_errno;
diff --git a/src/mpid/ch4/src/ch4r_recv.h b/src/mpid/ch4/src/ch4r_recv.h
index 4672198..addb8fa 100644
--- a/src/mpid/ch4/src/ch4r_recv.h
+++ b/src/mpid/ch4/src/ch4r_recv.h
@@ -210,17 +210,17 @@ static inline int MPIDI_CH4I_do_irecv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_recv
+#define FUNCNAME MPIDI_CH4U_mpi_recv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_recv(void *buf,
-                                             int count,
-                                             MPI_Datatype datatype,
-                                             int rank,
-                                             int tag,
-                                             MPIR_Comm * comm,
-                                             int context_offset, MPI_Status * status,
-                                             MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_recv(void *buf,
+                                                 int count,
+                                                 MPI_Datatype datatype,
+                                                 int rank,
+                                                 int tag,
+                                                 MPIR_Comm * comm,
+                                                 int context_offset, MPI_Status * status,
+                                                 MPIR_Request ** request)
 {
     int mpi_errno;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_RECV);
@@ -239,16 +239,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_recv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_recv_init
+#define FUNCNAME MPIDI_CH4U_mpi_recv_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_recv_init(void *buf,
-                                                  int count,
-                                                  MPI_Datatype datatype,
-                                                  int rank,
-                                                  int tag,
-                                                  MPIR_Comm * comm,
-                                                  int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_recv_init(void *buf,
+                                                      int count,
+                                                      MPI_Datatype datatype,
+                                                      int rank,
+                                                      int tag,
+                                                      MPIR_Comm * comm,
+                                                      int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq;
@@ -277,13 +277,13 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_recv_init(void *buf,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_imrecv
+#define FUNCNAME MPIDI_CH4U_mpi_imrecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_imrecv(void *buf,
-                                               int count,
-                                               MPI_Datatype datatype,
-                                               MPIR_Request * message, MPIR_Request ** rreqp)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_imrecv(void *buf,
+                                                   int count,
+                                                   MPI_Datatype datatype,
+                                                   MPIR_Request * message, MPIR_Request ** rreqp)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *rreq;
@@ -367,16 +367,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mrecv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_irecv
+#define FUNCNAME MPIDI_CH4U_mpi_irecv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_irecv(void *buf,
-                                              int count,
-                                              MPI_Datatype datatype,
-                                              int rank,
-                                              int tag,
-                                              MPIR_Comm * comm, int context_offset,
-                                              MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_irecv(void *buf,
+                                                  int count,
+                                                  MPI_Datatype datatype,
+                                                  int rank,
+                                                  int tag,
+                                                  MPIR_Comm * comm, int context_offset,
+                                                  MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_IRECV);
@@ -394,10 +394,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_irecv(void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_cancel_recv
+#define FUNCNAME MPIDI_CH4U_mpi_cancel_recv
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_cancel_recv(MPIR_Request * rreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_cancel_recv(MPIR_Request * rreq)
 {
     int mpi_errno = MPI_SUCCESS, found;
     MPIR_Comm *root_comm;
diff --git a/src/mpid/ch4/src/ch4r_request.h b/src/mpid/ch4/src/ch4r_request.h
index 0308b09..08ba7e4 100644
--- a/src/mpid/ch4/src/ch4r_request.h
+++ b/src/mpid/ch4/src/ch4r_request.h
@@ -98,7 +98,7 @@ static inline int MPIDI_CH4R_anysource_matched(MPIR_Request * rreq, int caller,
 
     if (MPIDI_CH4R_NETMOD == caller) {
 #ifdef MPIDI_BUILD_CH4_SHM
-        mpi_errno = MPIDI_SHM_cancel_recv(rreq);
+        mpi_errno = MPIDI_SHM_mpi_cancel_recv(rreq);
 
         /* If the netmod is cancelling the request, then shared memory will
          * just copy the status from the shared memory side because the netmod
@@ -112,7 +112,7 @@ static inline int MPIDI_CH4R_anysource_matched(MPIR_Request * rreq, int caller,
         *continue_matching = 0;
     }
     else if (MPIDI_CH4R_SHM == caller) {
-        mpi_errno = MPIDI_NM_cancel_recv(rreq);
+        mpi_errno = MPIDI_NM_mpi_cancel_recv(rreq);
 
         /* If the netmod has already matched this request, shared memory will
          * lose and should stop matching this request */
diff --git a/src/mpid/ch4/src/ch4r_rma.h b/src/mpid/ch4/src/ch4r_rma.h
index 39bf2cf..eb55e4f 100644
--- a/src/mpid/ch4/src/ch4r_rma.h
+++ b/src/mpid/ch4/src/ch4r_rma.h
@@ -250,16 +250,16 @@ static inline int MPIDI_CH4I_do_get(void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_put
+#define FUNCNAME MPIDI_CH4U_mpi_put
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_put(const void *origin_addr,
-                                            int origin_count,
-                                            MPI_Datatype origin_datatype,
-                                            int target_rank,
-                                            MPI_Aint target_disp,
-                                            int target_count, MPI_Datatype target_datatype,
-                                            MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_put(const void *origin_addr,
+                                                int origin_count,
+                                                MPI_Datatype origin_datatype,
+                                                int target_rank,
+                                                MPI_Aint target_disp,
+                                                int target_count, MPI_Datatype target_datatype,
+                                                MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_PUT);
@@ -278,17 +278,17 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_put(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_rput
+#define FUNCNAME MPIDI_CH4U_mpi_rput
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rput(const void *origin_addr,
-                                             int origin_count,
-                                             MPI_Datatype origin_datatype,
-                                             int target_rank,
-                                             MPI_Aint target_disp,
-                                             int target_count,
-                                             MPI_Datatype target_datatype,
-                                             MPIR_Win * win, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_rput(const void *origin_addr,
+                                                 int origin_count,
+                                                 MPI_Datatype origin_datatype,
+                                                 int target_rank,
+                                                 MPI_Aint target_disp,
+                                                 int target_count,
+                                                 MPI_Datatype target_datatype,
+                                                 MPIR_Win * win, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_RPUT);
@@ -308,16 +308,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rput(const void *origin_addr,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_get
+#define FUNCNAME MPIDI_CH4U_mpi_get
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_get(void *origin_addr,
-                                            int origin_count,
-                                            MPI_Datatype origin_datatype,
-                                            int target_rank,
-                                            MPI_Aint target_disp,
-                                            int target_count, MPI_Datatype target_datatype,
-                                            MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_get(void *origin_addr,
+                                                int origin_count,
+                                                MPI_Datatype origin_datatype,
+                                                int target_rank,
+                                                MPI_Aint target_disp,
+                                                int target_count, MPI_Datatype target_datatype,
+                                                MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_GET);
@@ -336,17 +336,17 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_get(void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_rget
+#define FUNCNAME MPIDI_CH4U_mpi_rget
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rget(void *origin_addr,
-                                             int origin_count,
-                                             MPI_Datatype origin_datatype,
-                                             int target_rank,
-                                             MPI_Aint target_disp,
-                                             int target_count,
-                                             MPI_Datatype target_datatype,
-                                             MPIR_Win * win, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_rget(void *origin_addr,
+                                                 int origin_count,
+                                                 MPI_Datatype origin_datatype,
+                                                 int target_rank,
+                                                 MPI_Aint target_disp,
+                                                 int target_count,
+                                                 MPI_Datatype target_datatype,
+                                                 MPIR_Win * win, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_RGET);
@@ -509,18 +509,18 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4I_do_accumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_raccumulate
+#define FUNCNAME MPIDI_CH4U_mpi_raccumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_raccumulate(const void *origin_addr,
-                                                    int origin_count,
-                                                    MPI_Datatype origin_datatype,
-                                                    int target_rank,
-                                                    MPI_Aint target_disp,
-                                                    int target_count,
-                                                    MPI_Datatype target_datatype,
-                                                    MPI_Op op, MPIR_Win * win,
-                                                    MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_raccumulate(const void *origin_addr,
+                                                        int origin_count,
+                                                        MPI_Datatype origin_datatype,
+                                                        int target_rank,
+                                                        MPI_Aint target_disp,
+                                                        int target_count,
+                                                        MPI_Datatype target_datatype,
+                                                        MPI_Op op, MPIR_Win * win,
+                                                        MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq;
@@ -551,27 +551,28 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_raccumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_accumulate
+#define FUNCNAME MPIDI_CH4U_mpi_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_accumulate(const void *origin_addr,
-                                                   int origin_count,
-                                                   MPI_Datatype origin_datatype,
-                                                   int target_rank,
-                                                   MPI_Aint target_disp,
-                                                   int target_count,
-                                                   MPI_Datatype target_datatype, MPI_Op op,
-                                                   MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_accumulate(const void *origin_addr,
+                                                       int origin_count,
+                                                       MPI_Datatype origin_datatype,
+                                                       int target_rank,
+                                                       MPI_Aint target_disp,
+                                                       int target_count,
+                                                       MPI_Datatype target_datatype, MPI_Op op,
+                                                       MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_ACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4U_ACCUMULATE);
 
-    mpi_errno = MPIDI_CH4U_raccumulate(origin_addr,
-                                       origin_count,
-                                       origin_datatype,
-                                       target_rank,
-                                       target_disp, target_count, target_datatype, op, win, NULL);
+    mpi_errno = MPIDI_CH4U_mpi_raccumulate(origin_addr,
+                                           origin_count,
+                                           origin_datatype,
+                                           target_rank,
+                                           target_disp, target_count, target_datatype, op, win,
+                                           NULL);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
 
@@ -583,21 +584,21 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_accumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_rget_accumulate
+#define FUNCNAME MPIDI_CH4U_mpi_rget_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rget_accumulate(const void *origin_addr,
-                                                        int origin_count,
-                                                        MPI_Datatype origin_datatype,
-                                                        void *result_addr,
-                                                        int result_count,
-                                                        MPI_Datatype result_datatype,
-                                                        int target_rank,
-                                                        MPI_Aint target_disp,
-                                                        int target_count,
-                                                        MPI_Datatype target_datatype,
-                                                        MPI_Op op, MPIR_Win * win,
-                                                        MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_rget_accumulate(const void *origin_addr,
+                                                            int origin_count,
+                                                            MPI_Datatype origin_datatype,
+                                                            void *result_addr,
+                                                            int result_count,
+                                                            MPI_Datatype result_datatype,
+                                                            int target_rank,
+                                                            MPI_Aint target_disp,
+                                                            int target_count,
+                                                            MPI_Datatype target_datatype,
+                                                            MPI_Op op, MPIR_Win * win,
+                                                            MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Request *sreq;
@@ -634,34 +635,34 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rget_accumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_get_accumulate
+#define FUNCNAME MPIDI_CH4U_mpi_get_accumulate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_get_accumulate(const void *origin_addr,
-                                                       int origin_count,
-                                                       MPI_Datatype origin_datatype,
-                                                       void *result_addr,
-                                                       int result_count,
-                                                       MPI_Datatype result_datatype,
-                                                       int target_rank,
-                                                       MPI_Aint target_disp,
-                                                       int target_count,
-                                                       MPI_Datatype target_datatype,
-                                                       MPI_Op op, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_get_accumulate(const void *origin_addr,
+                                                           int origin_count,
+                                                           MPI_Datatype origin_datatype,
+                                                           void *result_addr,
+                                                           int result_count,
+                                                           MPI_Datatype result_datatype,
+                                                           int target_rank,
+                                                           MPI_Aint target_disp,
+                                                           int target_count,
+                                                           MPI_Datatype target_datatype,
+                                                           MPI_Op op, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_GET_ACCUMULATE);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4U_GET_ACCUMULATE);
 
-    mpi_errno = MPIDI_CH4U_rget_accumulate(origin_addr,
-                                           origin_count,
-                                           origin_datatype,
-                                           result_addr,
-                                           result_count,
-                                           result_datatype,
-                                           target_rank,
-                                           target_disp,
-                                           target_count, target_datatype, op, win, NULL);
+    mpi_errno = MPIDI_CH4U_mpi_rget_accumulate(origin_addr,
+                                               origin_count,
+                                               origin_datatype,
+                                               result_addr,
+                                               result_count,
+                                               result_datatype,
+                                               target_rank,
+                                               target_disp,
+                                               target_count, target_datatype, op, win, NULL);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
   fn_exit:
@@ -672,15 +673,15 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_get_accumulate(const void *origin_addr,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_compare_and_swap
+#define FUNCNAME MPIDI_CH4U_mpi_compare_and_swap
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_compare_and_swap(const void *origin_addr,
-                                                         const void *compare_addr,
-                                                         void *result_addr,
-                                                         MPI_Datatype datatype,
-                                                         int target_rank,
-                                                         MPI_Aint target_disp, MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_compare_and_swap(const void *origin_addr,
+                                                             const void *compare_addr,
+                                                             void *result_addr,
+                                                             MPI_Datatype datatype,
+                                                             int target_rank,
+                                                             MPI_Aint target_disp, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS, c;
     MPIR_Request *sreq = NULL;
@@ -740,23 +741,23 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_compare_and_swap(const void *origin_addr
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_fetch_and_op
+#define FUNCNAME MPIDI_CH4U_mpi_fetch_and_op
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_fetch_and_op(const void *origin_addr,
-                                                     void *result_addr,
-                                                     MPI_Datatype datatype,
-                                                     int target_rank,
-                                                     MPI_Aint target_disp, MPI_Op op,
-                                                     MPIR_Win * win)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_fetch_and_op(const void *origin_addr,
+                                                         void *result_addr,
+                                                         MPI_Datatype datatype,
+                                                         int target_rank,
+                                                         MPI_Aint target_disp, MPI_Op op,
+                                                         MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_FETCH_AND_OP);
     MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_CH4U_FETCH_AND_OP);
 
-    mpi_errno = MPIDI_CH4U_get_accumulate(origin_addr, 1, datatype,
-                                          result_addr, 1, datatype,
-                                          target_rank, target_disp, 1, datatype, op, win);
+    mpi_errno = MPIDI_CH4U_mpi_get_accumulate(origin_addr, 1, datatype,
+                                              result_addr, 1, datatype,
+                                              target_rank, target_disp, 1, datatype, op, win);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
   fn_exit:
diff --git a/src/mpid/ch4/src/ch4r_send.h b/src/mpid/ch4/src/ch4r_send.h
index 57f50cb..236ad88 100644
--- a/src/mpid/ch4/src/ch4r_send.h
+++ b/src/mpid/ch4/src/ch4r_send.h
@@ -137,16 +137,16 @@ static inline int MPIDI_CH4I_psend(const void *buf,
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_send
+#define FUNCNAME MPIDI_CH4U_mpi_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_send(const void *buf,
-                                             int count,
-                                             MPI_Datatype datatype,
-                                             int rank,
-                                             int tag,
-                                             MPIR_Comm * comm, int context_offset,
-                                             MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_send(const void *buf,
+                                                 int count,
+                                                 MPI_Datatype datatype,
+                                                 int rank,
+                                                 int tag,
+                                                 MPIR_Comm * comm, int context_offset,
+                                                 MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_SEND);
@@ -158,16 +158,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_send(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_isend
+#define FUNCNAME MPIDI_CH4U_mpi_isend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_isend(const void *buf,
-                                              int count,
-                                              MPI_Datatype datatype,
-                                              int rank,
-                                              int tag,
-                                              MPIR_Comm * comm, int context_offset,
-                                              MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_isend(const void *buf,
+                                                  int count,
+                                                  MPI_Datatype datatype,
+                                                  int rank,
+                                                  int tag,
+                                                  MPIR_Comm * comm, int context_offset,
+                                                  MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_ISEND);
@@ -224,16 +224,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_irsend(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_ssend
+#define FUNCNAME MPIDI_CH4U_mpi_ssend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_ssend(const void *buf,
-                                              int count,
-                                              MPI_Datatype datatype,
-                                              int rank,
-                                              int tag,
-                                              MPIR_Comm * comm, int context_offset,
-                                              MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_ssend(const void *buf,
+                                                  int count,
+                                                  MPI_Datatype datatype,
+                                                  int rank,
+                                                  int tag,
+                                                  MPIR_Comm * comm, int context_offset,
+                                                  MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_SSEND);
@@ -245,16 +245,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_ssend(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_issend
+#define FUNCNAME MPIDI_CH4U_mpi_issend
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_issend(const void *buf,
-                                               int count,
-                                               MPI_Datatype datatype,
-                                               int rank,
-                                               int tag,
-                                               MPIR_Comm * comm, int context_offset,
-                                               MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_issend(const void *buf,
+                                                   int count,
+                                                   MPI_Datatype datatype,
+                                                   int rank,
+                                                   int tag,
+                                                   MPIR_Comm * comm, int context_offset,
+                                                   MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_ISSEND);
@@ -266,10 +266,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_issend(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_startall
+#define FUNCNAME MPIDI_CH4U_mpi_startall
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_startall(int count, MPIR_Request * requests[])
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_startall(int count, MPIR_Request * requests[])
 {
     int mpi_errno = MPI_SUCCESS, i;
     int rank, tag, context_offset;
@@ -294,10 +294,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_startall(int count, MPIR_Request * reque
 
         case MPIDI_PTYPE_RECV:
 #ifdef MPIDI_BUILD_CH4_SHM
-            mpi_errno = MPIDI_NM_irecv(MPIDI_CH4U_REQUEST(preq, buffer),
-                                       MPIDI_CH4U_REQUEST(preq, count),
-                                       datatype, rank, tag,
-                                       preq->comm, context_offset, &preq->u.persist.real_request);
+            mpi_errno = MPIDI_NM_mpi_irecv(MPIDI_CH4U_REQUEST(preq, buffer),
+                                           MPIDI_CH4U_REQUEST(preq, count),
+                                           datatype, rank, tag,
+                                           preq->comm, context_offset,
+                                           &preq->u.persist.real_request);
 #else
             mpi_errno = MPIDI_Irecv(MPIDI_CH4U_REQUEST(preq, buffer),
                                     MPIDI_CH4U_REQUEST(preq, count),
@@ -308,10 +309,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_startall(int count, MPIR_Request * reque
 
         case MPIDI_PTYPE_SEND:
 #ifdef MPIDI_BUILD_CH4_SHM
-            mpi_errno = MPIDI_NM_isend(MPIDI_CH4U_REQUEST(preq, buffer),
-                                       MPIDI_CH4U_REQUEST(preq, count),
-                                       datatype, rank, tag,
-                                       preq->comm, context_offset, &preq->u.persist.real_request);
+            mpi_errno = MPIDI_NM_mpi_isend(MPIDI_CH4U_REQUEST(preq, buffer),
+                                           MPIDI_CH4U_REQUEST(preq, count),
+                                           datatype, rank, tag,
+                                           preq->comm, context_offset,
+                                           &preq->u.persist.real_request);
 #else
             mpi_errno = MPIDI_Isend(MPIDI_CH4U_REQUEST(preq, buffer),
                                     MPIDI_CH4U_REQUEST(preq, count),
@@ -322,10 +324,11 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_startall(int count, MPIR_Request * reque
 
         case MPIDI_PTYPE_SSEND:
 #ifdef MPIDI_BUILD_CH4_SHM
-            mpi_errno = MPIDI_NM_issend(MPIDI_CH4U_REQUEST(preq, buffer),
-                                        MPIDI_CH4U_REQUEST(preq, count),
-                                        datatype, rank, tag,
-                                        preq->comm, context_offset, &preq->u.persist.real_request);
+            mpi_errno = MPIDI_NM_mpi_issend(MPIDI_CH4U_REQUEST(preq, buffer),
+                                            MPIDI_CH4U_REQUEST(preq, count),
+                                            datatype, rank, tag,
+                                            preq->comm, context_offset,
+                                            &preq->u.persist.real_request);
 #else
             mpi_errno = MPIDI_Issend(MPIDI_CH4U_REQUEST(preq, buffer),
                                      MPIDI_CH4U_REQUEST(preq, count),
@@ -374,16 +377,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_startall(int count, MPIR_Request * reque
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_send_init
+#define FUNCNAME MPIDI_CH4U_mpi_send_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_send_init(const void *buf,
-                                                  int count,
-                                                  MPI_Datatype datatype,
-                                                  int rank,
-                                                  int tag,
-                                                  MPIR_Comm * comm,
-                                                  int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_send_init(const void *buf,
+                                                      int count,
+                                                      MPI_Datatype datatype,
+                                                      int rank,
+                                                      int tag,
+                                                      MPIR_Comm * comm,
+                                                      int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_SEND_INIT);
@@ -395,16 +398,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_send_init(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_ssend_init
+#define FUNCNAME MPIDI_CH4U_mpi_ssend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_ssend_init(const void *buf,
-                                                   int count,
-                                                   MPI_Datatype datatype,
-                                                   int rank,
-                                                   int tag,
-                                                   MPIR_Comm * comm,
-                                                   int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_ssend_init(const void *buf,
+                                                       int count,
+                                                       MPI_Datatype datatype,
+                                                       int rank,
+                                                       int tag,
+                                                       MPIR_Comm * comm,
+                                                       int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_SSEND_INIT);
@@ -416,16 +419,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_ssend_init(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_bsend_init
+#define FUNCNAME MPIDI_CH4U_mpi_bsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_bsend_init(const void *buf,
-                                                   int count,
-                                                   MPI_Datatype datatype,
-                                                   int rank,
-                                                   int tag,
-                                                   MPIR_Comm * comm,
-                                                   int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_bsend_init(const void *buf,
+                                                       int count,
+                                                       MPI_Datatype datatype,
+                                                       int rank,
+                                                       int tag,
+                                                       MPIR_Comm * comm,
+                                                       int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_BSEND_INIT);
@@ -437,16 +440,16 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_bsend_init(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_rsend_init
+#define FUNCNAME MPIDI_CH4U_mpi_rsend_init
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rsend_init(const void *buf,
-                                                   int count,
-                                                   MPI_Datatype datatype,
-                                                   int rank,
-                                                   int tag,
-                                                   MPIR_Comm * comm,
-                                                   int context_offset, MPIR_Request ** request)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_rsend_init(const void *buf,
+                                                       int count,
+                                                       MPI_Datatype datatype,
+                                                       int rank,
+                                                       int tag,
+                                                       MPIR_Comm * comm,
+                                                       int context_offset, MPIR_Request ** request)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_RSEND_INIT);
@@ -458,10 +461,10 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_rsend_init(const void *buf,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4U_cancel_send
+#define FUNCNAME MPIDI_CH4U_mpi_cancel_send
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_cancel_send(MPIR_Request * sreq)
+MPL_STATIC_INLINE_PREFIX int MPIDI_CH4U_mpi_cancel_send(MPIR_Request * sreq)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4U_CANCEL_SEND);
diff --git a/src/mpid/ch4/src/ch4r_win.h b/src/mpid/ch4/src/ch4r_win.h
index 064fa89..12b078a 100644
--- a/src/mpid/ch4/src/ch4r_win.h
+++ b/src/mpid/ch4/src/ch4r_win.h
@@ -21,10 +21,10 @@
 #endif /* HAVE_SYS_MMAN_H */
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_set_info
+#define FUNCNAME MPIDI_CH4R_mpi_win_set_info
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_set_info(MPIR_Win * win, MPIR_Info * info)
+static inline int MPIDI_CH4R_mpi_win_set_info(MPIR_Win * win, MPIR_Info * info)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -127,7 +127,7 @@ static inline int MPIDI_CH4R_win_init(MPI_Aint length,
     win->copySize = 0;
     MPIDI_CH4U_WIN(win, shared_table) = NULL;
     if ((info != NULL) && ((int *) info != (int *) MPI_INFO_NULL)) {
-        mpi_errno = MPIDI_CH4R_win_set_info(win, info);
+        mpi_errno = MPIDI_CH4R_mpi_win_set_info(win, info);
         MPIR_Assert(mpi_errno == 0);
     }
 
@@ -218,10 +218,10 @@ static inline int MPIDI_CH4I_progress_win_fence(MPIR_Win * win)
 
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_start
+#define FUNCNAME MPIDI_CH4R_mpi_win_start
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_start(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -248,10 +248,10 @@ static inline int MPIDI_CH4R_win_start(MPIR_Group * group, int assert, MPIR_Win
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_complete
+#define FUNCNAME MPIDI_CH4R_mpi_win_complete
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_complete(MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_complete(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIDI_CH4U_win_cntrl_msg_t msg;
@@ -302,10 +302,10 @@ static inline int MPIDI_CH4R_win_complete(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_post
+#define FUNCNAME MPIDI_CH4R_mpi_win_post
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_post(MPIR_Group * group, int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIDI_CH4U_win_cntrl_msg_t msg;
@@ -352,10 +352,10 @@ static inline int MPIDI_CH4R_win_post(MPIR_Group * group, int assert, MPIR_Win *
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_wait
+#define FUNCNAME MPIDI_CH4R_mpi_win_wait
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_wait(MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_wait(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Group *group;
@@ -380,10 +380,10 @@ static inline int MPIDI_CH4R_win_wait(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_test
+#define FUNCNAME MPIDI_CH4R_mpi_win_test
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_test(MPIR_Win * win, int *flag)
+static inline int MPIDI_CH4R_mpi_win_test(MPIR_Win * win, int *flag)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -415,10 +415,10 @@ static inline int MPIDI_CH4R_win_test(MPIR_Win * win, int *flag)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_lock
+#define FUNCNAME MPIDI_CH4R_mpi_win_lock
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_lock(int lock_type, int rank, int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     unsigned locked;
@@ -454,10 +454,10 @@ static inline int MPIDI_CH4R_win_lock(int lock_type, int rank, int assert, MPIR_
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_unlock
+#define FUNCNAME MPIDI_CH4R_mpi_win_unlock
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_unlock(int rank, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_unlock(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     unsigned unlocked;
@@ -500,10 +500,10 @@ static inline int MPIDI_CH4R_win_unlock(int rank, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_get_info
+#define FUNCNAME MPIDI_CH4R_mpi_win_get_info
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
+static inline int MPIDI_CH4R_mpi_win_get_info(MPIR_Win * win, MPIR_Info ** info_p_p)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_GET_INFO);
@@ -616,10 +616,10 @@ static inline int MPIDI_CH4R_win_finalize(MPIR_Win ** win_ptr)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_free
+#define FUNCNAME MPIDI_CH4R_mpi_win_free
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_free(MPIR_Win ** win_ptr)
+static inline int MPIDI_CH4R_mpi_win_free(MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -641,10 +641,10 @@ static inline int MPIDI_CH4R_win_free(MPIR_Win ** win_ptr)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_fence
+#define FUNCNAME MPIDI_CH4R_mpi_win_fence
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_fence(int massert, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_fence(int massert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -679,13 +679,14 @@ static inline int MPIDI_CH4R_win_fence(int massert, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_create
+#define FUNCNAME MPIDI_CH4R_mpi_win_create
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_create(void *base,
-                                        MPI_Aint length,
-                                        int disp_unit,
-                                        MPIR_Info * info, MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_CH4R_mpi_win_create(void *base,
+                                            MPI_Aint length,
+                                            int disp_unit,
+                                            MPIR_Info * info, MPIR_Comm * comm_ptr,
+                                            MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -718,10 +719,10 @@ static inline int MPIDI_CH4R_win_create(void *base,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_attach
+#define FUNCNAME MPIDI_CH4R_mpi_win_attach
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
+static inline int MPIDI_CH4R_mpi_win_attach(MPIR_Win * win, void *base, MPI_Aint size)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_ATTACH);
@@ -737,14 +738,14 @@ static inline int MPIDI_CH4R_win_attach(MPIR_Win * win, void *base, MPI_Aint siz
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_allocate_shared
+#define FUNCNAME MPIDI_CH4R_mpi_win_allocate_shared
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_allocate_shared(MPI_Aint size,
-                                                 int disp_unit,
-                                                 MPIR_Info * info_ptr,
-                                                 MPIR_Comm * comm_ptr,
-                                                 void **base_ptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_CH4R_mpi_win_allocate_shared(MPI_Aint size,
+                                                     int disp_unit,
+                                                     MPIR_Info * info_ptr,
+                                                     MPIR_Comm * comm_ptr,
+                                                     void **base_ptr, MPIR_Win ** win_ptr)
 {
     int i = 0, fd = -1, rc, first = 0, mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -892,10 +893,10 @@ static inline int MPIDI_CH4R_win_allocate_shared(MPI_Aint size,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_detach
+#define FUNCNAME MPIDI_CH4R_mpi_win_detach
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_detach(MPIR_Win * win, const void *base)
+static inline int MPIDI_CH4R_mpi_win_detach(MPIR_Win * win, const void *base)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_DETACH);
@@ -910,12 +911,12 @@ static inline int MPIDI_CH4R_win_detach(MPIR_Win * win, const void *base)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_shared_query
+#define FUNCNAME MPIDI_CH4R_mpi_win_shared_query
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_shared_query(MPIR_Win * win,
-                                              int rank,
-                                              MPI_Aint * size, int *disp_unit, void *baseptr)
+static inline int MPIDI_CH4R_mpi_win_shared_query(MPIR_Win * win,
+                                                  int rank,
+                                                  MPI_Aint * size, int *disp_unit, void *baseptr)
 {
     int mpi_errno = MPI_SUCCESS;
     uintptr_t base = (uintptr_t) MPIDI_CH4U_WIN(win, mmap_addr);
@@ -942,13 +943,13 @@ static inline int MPIDI_CH4R_win_shared_query(MPIR_Win * win,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_allocate
+#define FUNCNAME MPIDI_CH4R_mpi_win_allocate
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_allocate(MPI_Aint size,
-                                          int disp_unit,
-                                          MPIR_Info * info,
-                                          MPIR_Comm * comm, void *baseptr, MPIR_Win ** win_ptr)
+static inline int MPIDI_CH4R_mpi_win_allocate(MPI_Aint size,
+                                              int disp_unit,
+                                              MPIR_Info * info,
+                                              MPIR_Comm * comm, void *baseptr, MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_Errflag_t errflag = MPIR_ERR_NONE;
@@ -986,10 +987,10 @@ static inline int MPIDI_CH4R_win_allocate(MPI_Aint size,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_flush
+#define FUNCNAME MPIDI_CH4R_mpi_win_flush
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_flush(int rank, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_flush(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_FLUSH);
@@ -1008,10 +1009,10 @@ static inline int MPIDI_CH4R_win_flush(int rank, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_flush_local_all
+#define FUNCNAME MPIDI_CH4R_mpi_win_flush_local_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_flush_local_all(MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_flush_local_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_FLUSH_LOCAL_ALL);
@@ -1031,10 +1032,10 @@ static inline int MPIDI_CH4R_win_flush_local_all(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_unlock_all
+#define FUNCNAME MPIDI_CH4R_mpi_win_unlock_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_unlock_all(MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_unlock_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_UNLOCK_ALL);
@@ -1083,11 +1084,11 @@ static inline int MPIDI_CH4R_win_unlock_all(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_create_dynamic
+#define FUNCNAME MPIDI_CH4R_mpi_win_create_dynamic
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_create_dynamic(MPIR_Info * info,
-                                                MPIR_Comm * comm, MPIR_Win ** win_ptr)
+static inline int MPIDI_CH4R_mpi_win_create_dynamic(MPIR_Info * info,
+                                                    MPIR_Comm * comm, MPIR_Win ** win_ptr)
 {
     int mpi_errno = MPI_SUCCESS;
     int rc = MPI_SUCCESS;
@@ -1120,10 +1121,10 @@ static inline int MPIDI_CH4R_win_create_dynamic(MPIR_Info * info,
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_flush_local
+#define FUNCNAME MPIDI_CH4R_mpi_win_flush_local
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_flush_local(int rank, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_flush_local(int rank, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_FLUSH_LOCAL);
@@ -1143,10 +1144,10 @@ static inline int MPIDI_CH4R_win_flush_local(int rank, MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_sync
+#define FUNCNAME MPIDI_CH4R_mpi_win_sync
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_sync(MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_sync(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_SYNC);
@@ -1163,10 +1164,10 @@ static inline int MPIDI_CH4R_win_sync(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_flush_all
+#define FUNCNAME MPIDI_CH4R_mpi_win_flush_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_flush_all(MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_flush_all(MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_CH4I_WIN_FLUSH_ALL);
@@ -1186,10 +1187,10 @@ static inline int MPIDI_CH4R_win_flush_all(MPIR_Win * win)
 }
 
 #undef FUNCNAME
-#define FUNCNAME MPIDI_CH4R_win_lock_all
+#define FUNCNAME MPIDI_CH4R_mpi_win_lock_all
 #undef FCNAME
 #define FCNAME MPL_QUOTE(FUNCNAME)
-static inline int MPIDI_CH4R_win_lock_all(int assert, MPIR_Win * win)
+static inline int MPIDI_CH4R_mpi_win_lock_all(int assert, MPIR_Win * win)
 {
     int mpi_errno = MPI_SUCCESS;
 

http://git.mpich.org/mpich.git/commitdiff/7adf2a94c124f6b5a5ecbd0f77c9ca70a8ed8451

commit 7adf2a94c124f6b5a5ecbd0f77c9ca70a8ed8451
Author: Pavan Balaji <balaji at anl.gov>
Date:   Fri Sep 9 11:00:21 2016 -0500

    CH4: Whitespace cleanup
    
    Signed-off-by: Ken Raffenetti <raffenet at mcs.anl.gov>

diff --git a/src/mpid/ch4/netmod/ofi/ofi_init.h b/src/mpid/ch4/netmod/ofi/ofi_init.h
index 9ea0a18..149f159 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_init.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_init.h
@@ -289,7 +289,7 @@ static inline int MPIDI_OFI_init_generic(int rank,
     else {
         av_attr.type = FI_AV_MAP;
     }
-    mapped_table           = (fi_addr_t *) MPL_malloc(size * sizeof(fi_addr_t));
+    mapped_table = (fi_addr_t *) MPL_malloc(size * sizeof(fi_addr_t));
 
     av_attr.rx_ctx_bits = MPIDI_OFI_MAX_ENDPOINTS_BITS;
 
diff --git a/src/mpid/ch4/netmod/ofi/ofi_win.h b/src/mpid/ch4/netmod/ofi/ofi_win.h
index 2bb045d..32e7520 100644
--- a/src/mpid/ch4/netmod/ofi/ofi_win.h
+++ b/src/mpid/ch4/netmod/ofi/ofi_win.h
@@ -154,9 +154,7 @@ static inline int MPIDI_OFI_win_init_generic(MPI_Aint length,
         MPIR_Assert(finfo);
         finfo->ep_attr->tx_ctx_cnt = FI_SHARED_CONTEXT; /* Request a shared context */
         MPIDI_OFI_CALL_RETURN(fi_endpoint(MPIDI_Global.domain,
-                                          finfo,
-                                          &MPIDI_OFI_WIN(win).ep,
-                                          NULL), ret);
+                                          finfo, &MPIDI_OFI_WIN(win).ep, NULL), ret);
         if (ret < 0) {
             MPL_DBG_MSG(MPIDI_CH4_DBG_GENERAL, VERBOSE,
                         "Failed to create per-window EP (with completion), "
@@ -183,9 +181,7 @@ static inline int MPIDI_OFI_win_init_generic(MPI_Aint length,
         MPIDI_OFI_CALL(fi_ep_bind(MPIDI_OFI_WIN(win).ep, &MPIDI_Global.av->fid, 0), bind);
 
         MPIDI_OFI_CALL_RETURN(fi_endpoint(MPIDI_Global.domain,
-                                          finfo,
-                                          &MPIDI_OFI_WIN(win).ep_nocmpl,
-                                          NULL), ret);
+                                          finfo, &MPIDI_OFI_WIN(win).ep_nocmpl, NULL), ret);
         fi_freeinfo(finfo);
         if (ret < 0) {
             MPL_DBG_MSG(MPIDI_CH4_DBG_GENERAL, VERBOSE,
@@ -199,10 +195,8 @@ static inline int MPIDI_OFI_win_init_generic(MPI_Aint length,
         MPIDI_OFI_CALL(fi_ep_bind(MPIDI_OFI_WIN(win).ep_nocmpl,
                                   &MPIDI_Global.stx_ctx->fid, 0), bind);
         MPIDI_OFI_CALL(fi_ep_bind(MPIDI_OFI_WIN(win).ep_nocmpl,
-                                  &MPIDI_OFI_WIN(win).cmpl_cntr->fid,
-                                  FI_READ | FI_WRITE), bind);
-        MPIDI_OFI_CALL(fi_ep_bind(MPIDI_OFI_WIN(win).ep_nocmpl,
-                                  &MPIDI_Global.av->fid, 0), bind);
+                                  &MPIDI_OFI_WIN(win).cmpl_cntr->fid, FI_READ | FI_WRITE), bind);
+        MPIDI_OFI_CALL(fi_ep_bind(MPIDI_OFI_WIN(win).ep_nocmpl, &MPIDI_Global.av->fid, 0), bind);
 
         MPIDI_OFI_CALL(fi_enable(MPIDI_OFI_WIN(win).ep), ep_enable);
         MPIDI_OFI_CALL(fi_enable(MPIDI_OFI_WIN(win).ep_nocmpl), ep_enable);
diff --git a/src/mpid/ch4/src/ch4_spawn.h b/src/mpid/ch4/src/ch4_spawn.h
index 07290cc..c7d7921 100644
--- a/src/mpid/ch4/src/ch4_spawn.h
+++ b/src/mpid/ch4/src/ch4_spawn.h
@@ -129,7 +129,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_Comm_spawn_multiple(int count,
         preput_keyval_vector.val = port_name;
         pmi_errno = PMI_Spawn_multiple(count, (const char **)
                                        commands,
-                                       (const char ** *) argvs,
+                                       (const char ***) argvs,
                                        maxprocs, info_keyval_sizes, (const PMI_keyval_t **)
                                        info_keyval_vectors, 1, &preput_keyval_vector, pmi_errcodes);
 

-----------------------------------------------------------------------

Summary of changes:
 src/include/mpir_request.h                   |    4 +-
 src/mpid/ch3/include/mpidpre.h               |    4 +-
 src/mpid/ch3/src/ch3u_request.c              |    6 +-
 src/mpid/ch4/include/mpid_ticketlock.h       |   22 +-
 src/mpid/ch4/include/mpidpost.h              |    4 +-
 src/mpid/ch4/include/mpidpre.h               |    6 +-
 src/mpid/ch4/netmod/include/netmod.h         | 1623 +++++++++++++------------
 src/mpid/ch4/netmod/include/netmod_impl.h    | 1050 ++++++++--------
 src/mpid/ch4/netmod/ofi/func_table.c         |  222 ++--
 src/mpid/ch4/netmod/ofi/ofi_am_probe.h       |   22 +-
 src/mpid/ch4/netmod/ofi/ofi_am_recv.h        |   61 +-
 src/mpid/ch4/netmod/ofi/ofi_am_rma.h         |  185 ++--
 src/mpid/ch4/netmod/ofi/ofi_am_send.h        |  127 +-
 src/mpid/ch4/netmod/ofi/ofi_am_spawn.h       |   18 +-
 src/mpid/ch4/netmod/ofi/ofi_am_win.h         |  127 +-
 src/mpid/ch4/netmod/ofi/ofi_coll.h           |  408 ++++---
 src/mpid/ch4/netmod/ofi/ofi_comm.h           |    8 +-
 src/mpid/ch4/netmod/ofi/ofi_datatype.h       |    4 +-
 src/mpid/ch4/netmod/ofi/ofi_events.h         |   73 +-
 src/mpid/ch4/netmod/ofi/ofi_impl.h           |   57 +-
 src/mpid/ch4/netmod/ofi/ofi_init.h           |   28 +-
 src/mpid/ch4/netmod/ofi/ofi_op.h             |    4 +-
 src/mpid/ch4/netmod/ofi/ofi_probe.h          |   22 +-
 src/mpid/ch4/netmod/ofi/ofi_progress.h       |    2 +-
 src/mpid/ch4/netmod/ofi/ofi_recv.h           |   79 +-
 src/mpid/ch4/netmod/ofi/ofi_rma.h            |  255 ++--
 src/mpid/ch4/netmod/ofi/ofi_send.h           |   85 +-
 src/mpid/ch4/netmod/ofi/ofi_spawn.h          |   22 +-
 src/mpid/ch4/netmod/ofi/ofi_win.h            |  145 ++--
 src/mpid/ch4/netmod/portals4/func_table.c    |  222 ++--
 src/mpid/ch4/netmod/portals4/ptl_coll.h      |  408 ++++---
 src/mpid/ch4/netmod/portals4/ptl_comm.h      |    8 +-
 src/mpid/ch4/netmod/portals4/ptl_datatype.h  |    4 +-
 src/mpid/ch4/netmod/portals4/ptl_init.h      |   30 +-
 src/mpid/ch4/netmod/portals4/ptl_op.h        |    4 +-
 src/mpid/ch4/netmod/portals4/ptl_probe.h     |   22 +-
 src/mpid/ch4/netmod/portals4/ptl_recv.h      |   55 +-
 src/mpid/ch4/netmod/portals4/ptl_rma.h       |  185 ++--
 src/mpid/ch4/netmod/portals4/ptl_send.h      |  127 +-
 src/mpid/ch4/netmod/portals4/ptl_spawn.h     |   18 +-
 src/mpid/ch4/netmod/portals4/ptl_win.h       |  127 +-
 src/mpid/ch4/netmod/stubnm/globals.c         |  223 ++--
 src/mpid/ch4/netmod/stubnm/stubnm_coll.h     |  408 ++++---
 src/mpid/ch4/netmod/stubnm/stubnm_comm.h     |    8 +-
 src/mpid/ch4/netmod/stubnm/stubnm_datatype.h |    4 +-
 src/mpid/ch4/netmod/stubnm/stubnm_init.h     |   24 +-
 src/mpid/ch4/netmod/stubnm/stubnm_op.h       |    4 +-
 src/mpid/ch4/netmod/stubnm/stubnm_probe.h    |   22 +-
 src/mpid/ch4/netmod/stubnm/stubnm_recv.h     |   50 +-
 src/mpid/ch4/netmod/stubnm/stubnm_rma.h      |  185 ++--
 src/mpid/ch4/netmod/stubnm/stubnm_send.h     |  127 +-
 src/mpid/ch4/netmod/stubnm/stubnm_spawn.h    |   18 +-
 src/mpid/ch4/netmod/stubnm/stubnm_win.h      |  127 +-
 src/mpid/ch4/netmod/ucx/func_table.c         |  222 ++--
 src/mpid/ch4/netmod/ucx/ucx_am_recv.h        |   61 +-
 src/mpid/ch4/netmod/ucx/ucx_am_rma.h         |  185 ++--
 src/mpid/ch4/netmod/ucx/ucx_am_send.h        |  127 +-
 src/mpid/ch4/netmod/ucx/ucx_am_win.h         |  127 +-
 src/mpid/ch4/netmod/ucx/ucx_coll.h           |  408 ++++---
 src/mpid/ch4/netmod/ucx/ucx_comm.h           |    8 +-
 src/mpid/ch4/netmod/ucx/ucx_datatype.h       |    4 +-
 src/mpid/ch4/netmod/ucx/ucx_init.h           |   30 +-
 src/mpid/ch4/netmod/ucx/ucx_op.h             |    4 +-
 src/mpid/ch4/netmod/ucx/ucx_probe.h          |   18 +-
 src/mpid/ch4/netmod/ucx/ucx_recv.h           |   81 +-
 src/mpid/ch4/netmod/ucx/ucx_rma.h            |  193 ++--
 src/mpid/ch4/netmod/ucx/ucx_send.h           |  169 ++--
 src/mpid/ch4/netmod/ucx/ucx_spawn.h          |   22 +-
 src/mpid/ch4/netmod/ucx/ucx_win.h            |  122 +-
 src/mpid/ch4/shm/include/shm.h               | 1694 +++++++++++++-------------
 src/mpid/ch4/shm/include/shm_impl.h          | 1102 +++++++++---------
 src/mpid/ch4/shm/posix/func_table.c          |  222 ++--
 src/mpid/ch4/shm/posix/posix_coll.h          |  418 ++++----
 src/mpid/ch4/shm/posix/posix_comm.h          |    8 +-
 src/mpid/ch4/shm/posix/posix_init.h          |   28 +-
 src/mpid/ch4/shm/posix/posix_probe.h         |   18 +-
 src/mpid/ch4/shm/posix/posix_recv.h          |   58 +-
 src/mpid/ch4/shm/posix/posix_rma.h           |  144 ++--
 src/mpid/ch4/shm/posix/posix_send.h          |  110 +-
 src/mpid/ch4/shm/posix/posix_spawn.h         |   18 +-
 src/mpid/ch4/shm/posix/posix_win.h           |   76 +-
 src/mpid/ch4/shm/stubshm/func_table.c        |  222 ++--
 src/mpid/ch4/shm/stubshm/stubshm_coll.h      |  418 ++++----
 src/mpid/ch4/shm/stubshm/stubshm_comm.h      |    4 +-
 src/mpid/ch4/shm/stubshm/stubshm_init.h      |   24 +-
 src/mpid/ch4/shm/stubshm/stubshm_probe.h     |   18 +-
 src/mpid/ch4/shm/stubshm/stubshm_recv.h      |   54 +-
 src/mpid/ch4/shm/stubshm/stubshm_rma.h       |  144 ++--
 src/mpid/ch4/shm/stubshm/stubshm_send.h      |  110 +-
 src/mpid/ch4/shm/stubshm/stubshm_spawn.h     |   18 +-
 src/mpid/ch4/shm/stubshm/stubshm_win.h       |   76 +-
 src/mpid/ch4/src/ch4_coll.h                  |  152 ++--
 src/mpid/ch4/src/ch4_comm.h                  |    8 +-
 src/mpid/ch4/src/ch4_init.h                  |   32 +-
 src/mpid/ch4/src/ch4_probe.h                 |   42 +-
 src/mpid/ch4/src/ch4_recv.h                  |   73 +-
 src/mpid/ch4/src/ch4_rma.h                   |   56 +-
 src/mpid/ch4/src/ch4_send.h                  |  107 +-
 src/mpid/ch4/src/ch4_spawn.h                 |   12 +-
 src/mpid/ch4/src/ch4_win.h                   |   52 +-
 src/mpid/ch4/src/ch4r_init.h                 |   20 +-
 src/mpid/ch4/src/ch4r_probe.h                |   30 +-
 src/mpid/ch4/src/ch4r_recv.h                 |   64 +-
 src/mpid/ch4/src/ch4r_request.h              |    4 +-
 src/mpid/ch4/src/ch4r_rma.h                  |  219 ++--
 src/mpid/ch4/src/ch4r_send.h                 |  163 ++--
 src/mpid/ch4/src/ch4r_win.h                  |  129 +-
 107 files changed, 7646 insertions(+), 7365 deletions(-)


hooks/post-receive
-- 
MPICH primary repository


More information about the commits mailing list