separate memory model support (WIP)

configure.ac

@@ -126,7 +126,7 @@ if test "$ac_cv_prog_cc_c99" = "no" ; then
    AC_ERROR([C99 not supported by the compiler])
 fi
 
-AC_CHECK_HEADERS([execinfo.h string.h strings.h stdint.h inttypes.h unistd.h errno.h time.h sys/time.h])
+AC_CHECK_HEADERS([execinfo.h string.h strings.h stdint.h stdbool.h inttypes.h unistd.h errno.h time.h sys/time.h])
 AC_TYPE_UINT8_T
 
 # asynchronous progress

src/armci_internals.h

@@ -105,6 +105,7 @@ typedef struct {
   int           progress_usleep;        /* Argument to usleep() to throttling polling                           */
 #endif
   int           use_win_allocate;       /* Use win_allocate or win_create (or special memory...)                */
+  int           msg_barrier_syncs;      /* Call MPI_Win_sync in armci_msg_barrier                               */
   int           explicit_nb_progress;   /* Poke the MPI progress engine at the end of nonblocking (NB) calls    */
   int           use_alloc_shm;          /* Pass alloc_shm info to win_allocate / alloc_mem                      */
   int           rma_atomicity;          /* Use Accumulate and Get_accumulate for Put and Get                    */
@@ -230,4 +231,7 @@ void ARMCII_Buf_finish_write_vec(void **orig_bufs, void **new_bufs, int count, i
 int  ARMCII_Buf_acc_is_scaled(int datatype, void *scale);
 void ARMCII_Buf_acc_scale(void *buf_in, void *buf_out, int size, int datatype, void *scale);
 
+int ARMCII_Is_win_unified(MPI_Win win);
+void ARMCII_Sync(void);
+
 #endif /* HAVE_ARMCI_INTERNALS_H */

src/gmr-extras.c

@@ -190,13 +190,18 @@ int gmr_flushall(gmr_t *mreg, int local_only) {
   * @param[in] mreg         Memory region
   * @return                 0 on success, non-zero on failure
   */
-int gmr_sync(gmr_t *mreg) {
-  int grp_me   = ARMCII_Translate_absolute_to_group(&mreg->group, ARMCI_GROUP_WORLD.rank);
-
+int gmr_sync(gmr_t *mreg)
+{
+#if 0
+  // what is the point of this?
+  int grp_me = ARMCII_Translate_absolute_to_group(&mreg->group, ARMCI_GROUP_WORLD.rank);
   ARMCII_Assert(grp_me >= 0);
+#endif
   ARMCII_Assert_msg(mreg->window != MPI_WIN_NULL, "A non-null mreg contains a null window.");
 
-  MPI_Win_sync(mreg->window);
+  if (!(mreg->unified)) {
+      MPI_Win_sync(mreg->window);
+  }
 
   return 0;
 }

src/gmr.c

@@ -72,6 +72,7 @@ gmr_t *gmr_create(gmr_size_t local_size, void **base_ptrs, ARMCI_Group *group) {
   mreg->nslices        = world_nproc;
   mreg->prev           = NULL;
   mreg->next           = NULL;
+  mreg->unified        = false;
 
   /* Allocate my slice of the GMR */
   alloc_slices[alloc_me].size = local_size;
@@ -213,7 +214,8 @@ gmr_t *gmr_create(gmr_size_t local_size, void **base_ptrs, ARMCI_Group *group) {
                    mreg->window);
 
   {
-    int unified = 0;
+#if 0
+    int unified = false;
     void    *attr_ptr;
     int     *attr_val;
     int      attr_flag;
@@ -224,23 +226,38 @@ gmr_t *gmr_create(gmr_size_t local_size, void **base_ptrs, ARMCI_Group *group) {
       if (world_me==0) {
         if ( (*attr_val)==MPI_WIN_SEPARATE ) {
           printf("MPI_WIN_MODEL = MPI_WIN_SEPARATE \n" );
-          unified = 0;
+          unified = false;
         } else if ( (*attr_val)==MPI_WIN_UNIFIED ) {
 #ifdef DEBUG
           printf("MPI_WIN_MODEL = MPI_WIN_UNIFIED \n" );
 #endif
-          unified = 1;
+          unified = true;
         } else {
           printf("MPI_WIN_MODEL = %d (not UNIFIED or SEPARATE) \n", *attr_val );
-          unified = 0;
+          unified = false;
         }
       }
     } else {
       if (world_me==0) {
         printf("MPI_WIN_MODEL attribute missing \n");
       }
+      unified = false;
     }
-    if (!unified && (ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_NOGUARD) ) {
+#else
+    int unified = ARMCII_Is_win_unified(mreg->window);
+    int print = ARMCII_GLOBAL_STATE.verbose;
+    if (unified == 1) {
+        mreg->unified = true;
+        if (print) printf("MPI_WIN_MODEL = MPI_WIN_UNIFIED\n");
+    } else if (unified == 0) {
+        mreg->unified = false;
+        if (print) printf("MPI_WIN_MODEL = MPI_WIN_SEPARATE\n");
+    } else {
+        mreg->unified = false;
+        if (print) printf("MPI_WIN_MODEL not available\n");
+    }
+#endif
+    if (!(mreg->unified) && (ARMCII_GLOBAL_STATE.shr_buf_method == ARMCII_SHR_BUF_NOGUARD) ) {
       if (world_me==0) {
         printf("Please re-run with ARMCI_SHR_BUF_METHOD=COPY\n");
       }

src/gmr.h

@@ -10,6 +10,14 @@
 #include <armci.h>
 #include <armcix.h>
 
+#ifdef HAVE_STDBOOL_H
+#include <stdbool.h>
+#else
+typedef int bool;
+#define false 0
+#define true  1
+#endif
+
 typedef armci_size_t gmr_size_t;
 #define GMR_MPI_SIZE_T ARMCII_MPI_SIZE_T
 
@@ -26,6 +34,7 @@ typedef struct gmr_s {
   struct gmr_s           *next;
   gmr_slice_t            *slices;         /* Array of GMR slices for this allocation                        */
   int                     nslices;
+  bool                    unified;        /* separate/unified attribute of the window                       */
 } gmr_t;
 
 extern gmr_t *gmr_list;

src/init_finalize.c

@@ -357,7 +357,10 @@ int PARMCI_Init_thread_comm(int armci_requested, MPI_Comm comm) {
   }
 
   /* Use win_allocate or not, to work around MPI-3 RMA implementation bugs. */
-  ARMCII_GLOBAL_STATE.use_win_allocate=ARMCII_Getenv_int("ARMCI_USE_WIN_ALLOCATE", 1);
+  ARMCII_GLOBAL_STATE.use_win_allocate = ARMCII_Getenv_bool("ARMCI_USE_WIN_ALLOCATE", 1);
+
+  /* Do MPI_Win_sync in armci_msg_barrier */
+  ARMCII_GLOBAL_STATE.msg_barrier_syncs = ARMCII_Getenv_bool("ARMCI_MSG_BARRIER_SYNCS", 0);
 
   /* Equivalent to ARMCI_Set_shm_limit - determines the size of:
    * - MPI_Win_allocate slab in the case of slab allocation
@@ -538,6 +541,7 @@ int PARMCI_Init_thread_comm(int armci_requested, MPI_Comm comm) {
       printf("  RMA_ATOMICITY          = %s\n", ARMCII_GLOBAL_STATE.rma_atomicity          ? "TRUE" : "FALSE");
       printf("  NO_FLUSH_LOCAL         = %s\n", ARMCII_GLOBAL_STATE.end_to_end_flush       ? "TRUE" : "FALSE");
       printf("  RMA_NOCHECK            = %s\n", ARMCII_GLOBAL_STATE.rma_nocheck            ? "TRUE" : "FALSE");
+      printf("  MSG_BARRIER_SYNCS      = %s\n", ARMCII_GLOBAL_STATE.msg_barrier_syncs      ? "TRUE" : "FALSE");
 
       /* MPI info set on window */
       printf("  USE_ALLOC_SHM          = %s\n", ARMCII_GLOBAL_STATE.use_alloc_shm          ? "TRUE" : "FALSE");

src/internals.c

@@ -56,19 +56,23 @@ int ARMCII_Translate_absolute_to_group(ARMCI_Group *group, int world_rank) {
   MPI_Group world_group, sub_group;
 
   ARMCII_Assert(world_rank >= 0 && world_rank < ARMCI_GROUP_WORLD.size);
-
+#if 0
+  // this is redundant - the assert checks this
   if (!(0 <= world_rank && world_rank < ARMCI_GROUP_WORLD.size)) {
       ARMCII_Warning("world_rank (%d) is not in the range [0,ARMCI_GROUP_WORLD.size=%d)!\n",
                       world_rank, ARMCI_GROUP_WORLD.size);
   }
+#endif
 
   /* Check if group is the world group */
   if (group->comm == ARMCI_GROUP_WORLD.comm) {
     group_rank = world_rank;
+    return group_rank;
   }
   /* Check for translation cache */
   else if (group->grp_to_abs != NULL) {
     group_rank = group->abs_to_grp[world_rank];
+    return group_rank;
   }
   else {
     /* Translate the rank */

src/message.c

@@ -108,6 +108,11 @@ void armci_msg_bcast_scope(int scope, void *buffer, int len, int root) {
   */
 void parmci_msg_barrier(void) {
   MPI_Barrier(ARMCI_GROUP_WORLD.comm);
+
+  if (ARMCII_GLOBAL_STATE.msg_barrier_syncs) {
+    ARMCII_Sync();
+    MPI_Barrier(ARMCI_GROUP_WORLD.comm);
+  }
 }
 
 
@@ -127,6 +132,11 @@ void parmci_msg_barrier(void) {
   */
 void parmci_msg_group_barrier(ARMCI_Group *group) {
   MPI_Barrier(group->comm);
+
+  if (ARMCII_GLOBAL_STATE.msg_barrier_syncs) {
+    ARMCII_Sync();
+    MPI_Barrier(group->comm);
+  }
 }
 
 

src/strided.c

@@ -137,11 +137,10 @@ int PARMCI_PutS(void *src_ptr, int src_stride_ar[/*stride_levels*/],
       }
     }
     else {
-      /* Jeff: WIN_UNIFIED should allow overlap to work but we
-       *       do a memory barrier here to be safe. */
       gmr_loc = gmr_lookup(src_ptr, ARMCI_GROUP_WORLD.rank);
-      if (gmr_loc != NULL)
+      if (gmr_loc != NULL) {
           gmr_sync(gmr_loc);
+      }
     }
 
     /* NOGUARD: If src_buf hasn't been assigned to a copy, the strided source
@@ -238,11 +237,10 @@ int PARMCI_GetS(void *src_ptr, int src_stride_ar[/*stride_levels*/],
       }
     }
     else {
-      /* Jeff: WIN_UNIFIED should allow overlap to work but we
-       *       do a memory barrier here to be safe. */
       gmr_loc = gmr_lookup(dst_ptr, ARMCI_GROUP_WORLD.rank);
-      if (gmr_loc != NULL)
+      if (gmr_loc != NULL) {
           gmr_sync(gmr_loc);
+      }
     }
 
     /* NOGUARD: If dst_buf hasn't been assigned to a copy, the strided source
@@ -376,11 +374,10 @@ int PARMCI_AccS(int datatype, void *scale,
       }
     }
     else {
-      /* Jeff: WIN_UNIFIED should allow overlap to work but we
-       *       do a memory barrier here to be safe. */
       gmr_loc = gmr_lookup(src_ptr, ARMCI_GROUP_WORLD.rank);
-      if (gmr_loc != NULL)
+      if (gmr_loc != NULL) {
           gmr_sync(gmr_loc);
+      }
     }
 
     /* NOGUARD: If src_buf hasn't been assigned to a copy, the strided source

src/strided_nb.c

@@ -66,11 +66,10 @@ int PARMCI_NbPutS(void *src_ptr, int src_stride_ar[/*stride_levels*/],
       }
     }
     else {
-      /* Jeff: WIN_UNIFIED should allow overlap to work but we
-       *       do a memory barrier here to be safe. */
       gmr_loc = gmr_lookup(src_ptr, ARMCI_GROUP_WORLD.rank);
-      if (gmr_loc != NULL)
+      if (gmr_loc != NULL) {
           gmr_sync(gmr_loc);
+      }
     }
 
     /* NOGUARD: If src_buf hasn't been assigned to a copy, the strided source
@@ -174,11 +173,10 @@ int PARMCI_NbGetS(void *src_ptr, int src_stride_ar[/*stride_levels*/],
       }
     }
     else {
-      /* Jeff: WIN_UNIFIED should allow overlap to work but we
-       *       do a memory barrier here to be safe. */
       gmr_loc = gmr_lookup(dst_ptr, ARMCI_GROUP_WORLD.rank);
-      if (gmr_loc != NULL)
+      if (gmr_loc != NULL) {
           gmr_sync(gmr_loc);
+      }
     }
 
     /* NOGUARD: If dst_buf hasn't been assigned to a copy, the strided source
@@ -318,11 +316,10 @@ int PARMCI_NbAccS(int datatype, void *scale,
       }
     }
     else {
-      /* Jeff: WIN_UNIFIED should allow overlap to work but we
-       *       do a memory barrier here to be safe. */
       gmr_loc = gmr_lookup(src_ptr, ARMCI_GROUP_WORLD.rank);
-      if (gmr_loc != NULL)
+      if (gmr_loc != NULL) {
           gmr_sync(gmr_loc);
+      }
     }
 
     /* NOGUARD: If src_buf hasn't been assigned to a copy, the strided source

src/util.c

@@ -37,15 +37,9 @@ void ARMCI_Error(const char *msg, int code) {
   * group!).
   */
 void PARMCI_Barrier(void) {
-  gmr_t *cur_mreg = gmr_list;
-
   PARMCI_AllFence();
   MPI_Barrier(ARMCI_GROUP_WORLD.comm);
-
-  while (cur_mreg) {
-    gmr_sync(cur_mreg);
-    cur_mreg = cur_mreg->next;
-  }
+  ARMCII_Sync();
 }
 
 /* -- begin weak symbols block -- */
@@ -59,8 +53,6 @@ void PARMCI_Barrier(void) {
 /* -- end weak symbols block -- */
 
 /** Wait for remote completion on one-sided operations targeting process proc.
-  * In MPI-2, this is a no-op since get/put/acc already guarantee remote
-  * completion.
   *
   * @param[in] proc Process to target
   */
@@ -85,8 +77,7 @@ void PARMCI_Fence(int proc) {
 #endif
 /* -- end weak symbols block -- */
 
-/** Wait for remote completion on all one-sided operations.  In MPI-2, this is
-  * a no-op since get/put/acc already guarantee remote completion.
+/** Wait for remote completion on all one-sided operations.
   */
 void PARMCI_AllFence(void) {
   gmr_t *cur_mreg = gmr_list;
@@ -122,7 +113,6 @@ int ARMCI_Uses_shm_grp(ARMCI_Group *group) {
   return 0;
 }
 
-
 /** Copy local data.
   *
   * @param[in]  src  Source buffer
@@ -228,3 +218,14 @@ int ARMCII_Is_win_unified(MPI_Win win)
     return -1;
   }
 }
+
+/** Sync all windows
+  */
+void ARMCII_Sync(void) {
+  gmr_t *cur_mreg = gmr_list;
+
+  while (cur_mreg) {
+    gmr_sync(cur_mreg);
+    cur_mreg = cur_mreg->next;
+  }
+}