group: refactor MPIR_Group

* add option to use stride to describe group composition
* remove the linked list design

src/include/mpir_group.h

@@ -12,24 +12,6 @@
  * MPI_Group_intersection) and for the scalable RMA synchronization
  *---------------------------------------------------------------------------*/
 
-/* Abstract the integer type for lpid (process id). It is possible to use 32-bit
- * in principle, but 64-bit is simpler since we can trivially combine
- * (world_idx, world_rank).
- */
-typedef uint64_t MPIR_Lpid;
-
-/* This structure is used to implement the group operations such as
-   MPI_Group_translate_ranks */
-/* note: next_lpid (with idx_of_first_lpid in MPIR_Group) gives a linked list
- * in a sorted lpid ascending order */
-typedef struct MPII_Group_pmap_t {
-    MPIR_Lpid lpid;             /* local process id, from VCONN */
-    int next_lpid;              /* Index of next lpid (in lpid order) */
-} MPII_Group_pmap_t;
-
-/* Any changes in the MPIR_Group structure must be made to the
-   predefined value in MPIR_Group_builtin for MPI_GROUP_EMPTY in
-   src/mpi/group/grouputil.c */
 /*S
  MPIR_Group - Description of the Group data structure
 
@@ -60,22 +42,75 @@ typedef struct MPII_Group_pmap_t {
  Group-DS
 
  S*/
+
+/* Abstract the integer type for lpid (process id). It is possible to use 32-bit
+ * in principle, but 64-bit is simpler since we can trivially combine
+ * (world_idx, world_rank).
+ */
+typedef uint64_t MPIR_Lpid;
+
+struct MPIR_Pmap {
+    int size;                   /* same as group->size, duplicate here so Pmap is logically complete */
+    bool use_map;
+    union {
+        MPIR_Lpid *map;
+        struct {
+            MPIR_Lpid offset;
+            MPIR_Lpid stride;
+            MPIR_Lpid blocksize;
+        } stride;
+    } u;
+};
+
+MPL_STATIC_INLINE_PREFIX MPIR_Lpid MPIR_Pmap_rank_to_lpid(struct MPIR_Pmap *pmap, int rank)
+{
+    if (rank >= 0 && rank < pmap->size) {
+        return MPI_UNDEFINED;
+    }
+
+    if (pmap->use_map) {
+        return pmap->u.map[rank];
+    } else {
+        MPIR_Lpid i_blk = rank / pmap->u.stride.blocksize;
+        MPIR_Lpid r_blk = rank % pmap->u.stride.blocksize;
+        return pmap->u.stride.offset + i_blk * pmap->u.stride.stride + r_blk;
+    }
+}
+
+MPL_STATIC_INLINE_PREFIX int MPIR_Pmap_lpid_to_rank(struct MPIR_Pmap *pmap, MPIR_Lpid lpid)
+{
+    if (pmap->use_map) {
+        /* linear search */
+        for (int rank = 0; rank < pmap->size; rank++) {
+            if (pmap->u.map[rank] == lpid) {
+                return rank;
+            }
+        }
+        return MPI_UNDEFINED;
+    } else {
+        lpid -= pmap->u.stride.offset;
+        MPIR_Lpid i_blk = lpid / pmap->u.stride.stride;
+        MPIR_Lpid r_blk = lpid % pmap->u.stride.stride;
+
+        if (r_blk >= pmap->u.stride.blocksize) {
+            return MPI_UNDEFINED;
+        }
+
+        int rank = i_blk * pmap->u.stride.blocksize + r_blk;
+        if (rank >= 0 && rank < pmap->size) {
+            return rank;
+        } else {
+            return MPI_UNDEFINED;
+        }
+    }
+}
+
 struct MPIR_Group {
     MPIR_OBJECT_HEADER;         /* adds handle and ref_count fields */
     int size;                   /* Size of a group */
-    int rank;                   /* rank of this process relative to this
-                                 * group */
-    int idx_of_first_lpid;
-    MPII_Group_pmap_t *lrank_to_lpid;   /* Array mapping a local rank to local
-                                         * process number */
-    int is_local_dense_monotonic;       /* see NOTE-G1 */
-
-    /* We may want some additional data for the RMA syncrhonization calls */
-    /* Other, device-specific information */
-#ifdef MPID_DEV_GROUP_DECL
-     MPID_DEV_GROUP_DECL
-#endif
-     MPIR_Session * session_ptr;        /* Pointer to session to which this group belongs */
+    int rank;                   /* rank of this process relative to this group */
+    struct MPIR_Pmap pmap;
+    MPIR_Session *session_ptr;  /* Pointer to session to which this group belongs */
 };
 
 /* NOTE-G1: is_local_dense_monotonic will be true iff the group meets the
@@ -104,10 +139,8 @@ extern MPIR_Group *const MPIR_Group_empty;
 #define MPIR_Group_release_ref(_group, _inuse) \
      do { MPIR_Object_release_ref(_group, _inuse); } while (0)
 
-void MPII_Group_setup_lpid_list(MPIR_Group *);
 int MPIR_Group_check_valid_ranks(MPIR_Group *, const int[], int);
 int MPIR_Group_check_valid_ranges(MPIR_Group *, int[][3], int);
-void MPIR_Group_setup_lpid_pairs(MPIR_Group *, MPIR_Group *);
 int MPIR_Group_create(int, MPIR_Group **);
 int MPIR_Group_release(MPIR_Group * group_ptr);
 
@@ -122,7 +155,4 @@ int MPIR_Group_check_subset(MPIR_Group * group_ptr, MPIR_Comm * comm_ptr);
 void MPIR_Group_set_session_ptr(MPIR_Group * group_ptr, MPIR_Session * session_out);
 int MPIR_Group_init(void);
 
-/* internal functions */
-void MPII_Group_setup_lpid_list(MPIR_Group *);
-
 #endif /* MPIR_GROUP_H_INCLUDED */

src/mpi/comm/comm_impl.c

@@ -198,9 +198,6 @@ int MPII_Comm_create_calculate_mapping(MPIR_Group * group_ptr,
      * exactly the same as the ranks in comm world.
      */
 
-    /* we examine the group's lpids in both the intracomm and non-comm_world cases */
-    MPII_Group_setup_lpid_list(group_ptr);
-
     /* Optimize for groups contained within MPI_COMM_WORLD. */
     if (comm_ptr->comm_kind == MPIR_COMM_KIND__INTRACOMM) {
         int wsize;