Merge pull request #3490 from jain-surabhi-23/intra_node_colls

ch4/posix: shared memory based intra-node collectives

configure.ac

@@ -1173,6 +1173,7 @@ fi
 if test "$izem_atomic" = "yes" ; then
     AC_DEFINE(ENABLE_IZEM_ATOMIC,1,[Define to enable using Izem CPU atomics])
 fi
+AM_CONDITIONAL([ENABLE_IZEM_ATOMIC], [test x$izem_atomic = xyes])
 
 AC_ARG_VAR([ZMLIBNAME],[can be used to override the name of the Izem library (default: "zm")])
 ZMLIBNAME=${ZMLIBNAME:-"zm"}

maint/extracterrmsgs

@@ -731,7 +731,7 @@ sub ProcessFile
 		    !($args[$errClassLoc] =~ /^MPIDI_CH3I_SOCK_ERR_/) &&
 		    !($args[$errClassLoc] =~ /^MPIX_ERR_/) &&
 		    !($args[$errClassLoc] =~ /^errclass/) &&
-                    !($args[$errClassLoc] =~ /^\*\(errflag_\)/) &&
+                    !($args[$errClassLoc] =~ /^errflag/) &&
                     !($args[$errClassLoc] =~ /^\*errflag/)) {
 		    $bad_syntax_in_file{$filename} = 1;
 		    print STDERR "Invalid argument $args[$errClassLoc] for the MPI Error class in $routineName in $filename\n";

src/mpi/coll/algorithms/common/algo_common.h

@@ -22,19 +22,19 @@ static int MPII_Algo_compare_int(const void *a, const void *b)
 }
 
 /* Avoid unused function warning in certain configurations */
-static int MPII_Algo_calculate_pipeline_chunk_info(int maxbytes, int type_size, int count,
-                                                   int *num_segments, int *segsize_floor,
-                                                   int *segsize_ceil) ATTRIBUTE((unused));
-static int MPII_Algo_calculate_pipeline_chunk_info(int maxbytes,
-                                                   int type_size, int count,
-                                                   int *num_segments,
-                                                   int *segsize_floor, int *segsize_ceil)
+static inline int MPIR_Algo_calculate_pipeline_chunk_info(int maxbytes, int type_size, int count,
+                                                          int *num_segments, int *segsize_floor,
+                                                          int *segsize_ceil) ATTRIBUTE((unused));
+static inline int MPIR_Algo_calculate_pipeline_chunk_info(int maxbytes,
+                                                          int type_size, int count,
+                                                          int *num_segments,
+                                                          int *segsize_floor, int *segsize_ceil)
 {
     int maxelems;
     int mpi_errno = MPI_SUCCESS;
 
-    MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPII_ALGO_CALCULATE_PIPELINE_CHUNK_INFO);
-    MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPII_ALGO_CALCULATE_PIPELINE_CHUNK_INFO);
+    MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIR_ALGO_CALCULATE_PIPELINE_CHUNK_INFO);
+    MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIR_ALGO_CALCULATE_PIPELINE_CHUNK_INFO);
 
     if (count == 0 || type_size == 0) {
         *num_segments = *segsize_floor = *segsize_ceil = 0;
@@ -57,7 +57,7 @@ static int MPII_Algo_calculate_pipeline_chunk_info(int maxbytes,
 
     MPL_DBG_MSG_FMT(MPIR_DBG_COLL, VERBOSE, (MPL_DBG_FDEST, "num_segments %d", *num_segments));
 
-    MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPII_ALGO_CALCULATE_PIPELINE_CHUNK_INFO);
+    MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIR_ALGO_CALCULATE_PIPELINE_CHUNK_INFO);
 
   fn_exit:
     return mpi_errno;

src/mpi/coll/algorithms/treealgo/treealgo.c

@@ -35,13 +35,13 @@ int MPII_Treealgo_comm_cleanup(MPIR_Comm * comm)
 }
 
 
-int MPII_Treealgo_tree_create(int rank, int nranks, int tree_type, int k, int root,
-                              MPII_Treealgo_tree_t * ct)
+int MPIR_Treealgo_tree_create(int rank, int nranks, int tree_type, int k, int root,
+                              MPIR_Treealgo_tree_t * ct)
 {
     int mpi_errno = MPI_SUCCESS;
 
-    MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPII_TREEALGO_TREE_INIT);
-    MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPII_TREEALGO_TREE_INIT);
+    MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIR_TREEALGO_TREE_INIT);
+    MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIR_TREEALGO_TREE_INIT);
 
     switch (tree_type) {
         case MPIR_TREE_TYPE_KARY:
@@ -68,7 +68,7 @@ int MPII_Treealgo_tree_create(int rank, int nranks, int tree_type, int k, int ro
             break;
     }
 
-    MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPII_TREEALGO_TREE_INIT);
+    MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIR_TREEALGO_TREE_INIT);
 
   fn_exit:
     return mpi_errno;
@@ -78,7 +78,7 @@ int MPII_Treealgo_tree_create(int rank, int nranks, int tree_type, int k, int ro
 }
 
 
-void MPII_Treealgo_tree_free(MPII_Treealgo_tree_t * tree)
+void MPIR_Treealgo_tree_free(MPIR_Treealgo_tree_t * tree)
 {
     utarray_free(tree->children);
 }

src/mpi/coll/algorithms/treealgo/treealgo.h

@@ -16,8 +16,8 @@
 int MPII_Treealgo_init(void);
 int MPII_Treealgo_comm_init(MPIR_Comm * comm);
 int MPII_Treealgo_comm_cleanup(MPIR_Comm * comm);
-int MPII_Treealgo_tree_create(int rank, int nranks, int tree_type, int k, int root,
-                              MPII_Treealgo_tree_t * ct);
-void MPII_Treealgo_tree_free(MPII_Treealgo_tree_t * tree);
+int MPIR_Treealgo_tree_create(int rank, int nranks, int tree_type, int k, int root,
+                              MPIR_Treealgo_tree_t * ct);
+void MPIR_Treealgo_tree_free(MPIR_Treealgo_tree_t * tree);
 
 #endif /* TREEALGO_H_INCLUDED */

src/mpi/coll/algorithms/treealgo/treealgo_types.h

@@ -18,6 +18,6 @@ typedef struct {
     int parent;
     int num_children;
     UT_array *children;
-} MPII_Treealgo_tree_t;
+} MPIR_Treealgo_tree_t;
 
 #endif /* TREEALGO_TYPES_H_INCLUDED */

src/mpi/coll/algorithms/treealgo/treeutil.c

@@ -15,7 +15,7 @@
 #include "treeutil.h"
 #include "mpiimpl.h"
 
-static int tree_add_child(MPII_Treealgo_tree_t * t, int rank)
+static int tree_add_child(MPIR_Treealgo_tree_t * t, int rank)
 {
     int mpi_errno = MPI_SUCCESS;
 
@@ -26,7 +26,7 @@ static int tree_add_child(MPII_Treealgo_tree_t * t, int rank)
 }
 
 
-int MPII_Treeutil_tree_kary_init(int rank, int nranks, int k, int root, MPII_Treealgo_tree_t * ct)
+int MPII_Treeutil_tree_kary_init(int rank, int nranks, int k, int root, MPIR_Treealgo_tree_t * ct)
 {
     int lrank, child;
     int mpi_errno = MPI_SUCCESS;
@@ -76,7 +76,7 @@ int MPII_Treeutil_tree_kary_init(int rank, int nranks, int k, int root, MPII_Tre
  *                     3
  */
 int MPII_Treeutil_tree_knomial_1_init(int rank, int nranks, int k, int root,
-                                      MPII_Treealgo_tree_t * ct)
+                                      MPIR_Treealgo_tree_t * ct)
 {
     int lrank, i, j, maxtime, tmp, time, parent, current_rank, running_rank, crank;
     int mpi_errno = MPI_SUCCESS;
@@ -171,7 +171,7 @@ int MPII_Treeutil_tree_knomial_1_init(int rank, int nranks, int k, int root,
  *                      7
  */
 int MPII_Treeutil_tree_knomial_2_init(int rank, int nranks, int k, int root,
-                                      MPII_Treealgo_tree_t * ct)
+                                      MPIR_Treealgo_tree_t * ct)
 {
     int mpi_errno = MPI_SUCCESS;
     int lrank, i, j, depth;

src/mpi/coll/algorithms/treealgo/treeutil.h

@@ -13,14 +13,14 @@
 #define TREEUTIL_H_INCLUDED
 
 /* Generate kary tree information for rank 'rank' */
-int MPII_Treeutil_tree_kary_init(int rank, int nranks, int k, int root, MPII_Treealgo_tree_t * ct);
+int MPII_Treeutil_tree_kary_init(int rank, int nranks, int k, int root, MPIR_Treealgo_tree_t * ct);
 
 /* Generate knomial_1 tree information for rank 'rank' */
 int MPII_Treeutil_tree_knomial_1_init(int rank, int nranks, int k, int root,
-                                      MPII_Treealgo_tree_t * ct);
+                                      MPIR_Treealgo_tree_t * ct);
 
 /* Generate knomial_2 tree information for rank 'rank' */
 int MPII_Treeutil_tree_knomial_2_init(int rank, int nranks, int k, int root,
-                                      MPII_Treealgo_tree_t * ct);
+                                      MPIR_Treealgo_tree_t * ct);
 
 #endif /* TREEUTIL_H_INCLUDED */

src/mpi/coll/iallreduce/iallreduce_tsp_tree_algos.h

@@ -33,7 +33,7 @@ int MPIR_TSP_Iallreduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int
     int size;
     int rank;
     int num_children = 0;
-    MPII_Treealgo_tree_t my_tree;
+    MPIR_Treealgo_tree_t my_tree;
     void **child_buffer;        /* Buffer array in which data from children is received */
     void *reduce_buffer;        /* Buffer in which allreduced data is present */
     int *vtcs = NULL, *recv_id = NULL, *reduce_id = NULL;       /* Arrays to store graph vertex ids */
@@ -61,7 +61,7 @@ int MPIR_TSP_Iallreduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int
 
 
     /* calculate chunking information for pipelining */
-    MPII_Algo_calculate_pipeline_chunk_info(maxbytes, type_size, count, &num_chunks,
+    MPIR_Algo_calculate_pipeline_chunk_info(maxbytes, type_size, count, &num_chunks,
                                             &chunk_size_floor, &chunk_size_ceil);
     /* print chunking information */
     MPL_DBG_MSG_FMT(MPIR_DBG_COLL, VERBOSE, (MPL_DBG_FDEST,
@@ -76,7 +76,7 @@ int MPIR_TSP_Iallreduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int
 
     /* initialize the tree */
     my_tree.children = NULL;
-    mpi_errno = MPII_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
+    mpi_errno = MPIR_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
     num_children = my_tree.num_children;
@@ -216,7 +216,7 @@ int MPIR_TSP_Iallreduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int
         offset += msgsize;
     }
 
-    MPII_Treealgo_tree_free(&my_tree);
+    MPIR_Treealgo_tree_free(&my_tree);
 
   fn_exit:
     MPL_free(vtcs);

src/mpi/coll/ibcast/ibcast_tsp_tree_algos.h

@@ -32,7 +32,7 @@ int MPIR_TSP_Ibcast_sched_intra_tree(void *buffer, int count, MPI_Datatype datat
     int rank;
     int recv_id;
     int num_children;
-    MPII_Treealgo_tree_t my_tree;
+    MPIR_Treealgo_tree_t my_tree;
     int tag;
 
     MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIR_TSP_IBCAST_SCHED_INTRA_TREE);
@@ -51,15 +51,15 @@ int MPIR_TSP_Ibcast_sched_intra_tree(void *buffer, int count, MPI_Datatype datat
     extent = MPL_MAX(extent, true_extent);
 
     /* calculate chunking information for pipelining */
-    MPII_Algo_calculate_pipeline_chunk_info(maxbytes, type_size, count, &num_chunks,
+    MPIR_Algo_calculate_pipeline_chunk_info(maxbytes, type_size, count, &num_chunks,
                                             &chunk_size_floor, &chunk_size_ceil);
     /* print chunking information */
     MPL_DBG_MSG_FMT(MPIR_DBG_COLL, VERBOSE, (MPL_DBG_FDEST,
                                              "Broadcast pipeline info: maxbytes=%d count=%d num_chunks=%d chunk_size_floor=%d chunk_size_ceil=%d",
                                              maxbytes, count, num_chunks,
                                              chunk_size_floor, chunk_size_ceil));
 
-    mpi_errno = MPII_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
+    mpi_errno = MPIR_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
     num_children = my_tree.num_children;
@@ -93,7 +93,7 @@ int MPIR_TSP_Ibcast_sched_intra_tree(void *buffer, int count, MPI_Datatype datat
         offset += msgsize;
     }
 
-    MPII_Treealgo_tree_free(&my_tree);
+    MPIR_Treealgo_tree_free(&my_tree);
 
   fn_exit:
     MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIR_TSP_IBCAST_SCHED_INTRA_TREE);

src/mpi/coll/igather/igather_tsp_tree_algos.h

@@ -32,7 +32,7 @@ int MPIR_TSP_Igather_sched_intra_tree(const void *sendbuf, int sendcount,
     void *tmp_buf = NULL;
     const void *data_buf = NULL;
     int tree_type;
-    MPII_Treealgo_tree_t my_tree, parents_tree;
+    MPIR_Treealgo_tree_t my_tree, parents_tree;
     int next_child, num_children, *child_subtree_size = NULL, *child_data_offset = NULL;
     int offset, recv_size, num_dependencies;
 
@@ -48,7 +48,7 @@ int MPIR_TSP_Igather_sched_intra_tree(const void *sendbuf, int sendcount,
         is_inplace = (sendbuf == MPI_IN_PLACE); /* For gather, MPI_IN_PLACE is significant only at root */
 
     tree_type = MPIR_TREE_TYPE_KNOMIAL_1;       /* currently only tree_type=MPIR_TREE_TYPE_KNOMIAL_1 is supported for gather */
-    mpi_errno = MPII_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
+    mpi_errno = MPIR_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
     num_children = my_tree.num_children;
@@ -83,7 +83,7 @@ int MPIR_TSP_Igather_sched_intra_tree(const void *sendbuf, int sendcount,
 
     /* get tree information of the parent */
     if (my_tree.parent != -1) {
-        MPII_Treealgo_tree_create(my_tree.parent, size, tree_type, k, root, &parents_tree);
+        MPIR_Treealgo_tree_create(my_tree.parent, size, tree_type, k, root, &parents_tree);
     } else {    /* initialize an empty children array */
         utarray_new(parents_tree.children, &ut_int_icd, MPL_MEM_COLL);
         parents_tree.num_children = 0;
@@ -116,7 +116,7 @@ int MPIR_TSP_Igather_sched_intra_tree(const void *sendbuf, int sendcount,
         recv_size += child_subtree_size[i];
     }
 
-    MPII_Treealgo_tree_free(&parents_tree);
+    MPIR_Treealgo_tree_free(&parents_tree);
 
     recv_size *= (lrank == 0) ? recvcount : sendcount;
     offset = (lrank == 0) ? recvcount : sendcount;
@@ -184,7 +184,7 @@ int MPIR_TSP_Igather_sched_intra_tree(const void *sendbuf, int sendcount,
 
     }
 
-    MPII_Treealgo_tree_free(&my_tree);
+    MPIR_Treealgo_tree_free(&my_tree);
 
   fn_exit:
     MPL_free(child_subtree_size);

src/mpi/coll/ireduce/ireduce_tsp_tree_algos.h

@@ -40,7 +40,7 @@ int MPIR_TSP_Ireduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int co
                                  * rank 0 sends the reduced data to the root of the ecollective op. */
     int is_tree_root, is_tree_leaf, is_tree_intermediate;       /* Variables to store location of this rank in the tree */
     int is_root;
-    MPII_Treealgo_tree_t my_tree;
+    MPIR_Treealgo_tree_t my_tree;
     void **child_buffer = NULL; /* Buffer array in which data from children is received */
     void *reduce_buffer;        /* Buffer in which reduced data is present */
     int *vtcs = NULL, *recv_id = NULL, *reduce_id = NULL;       /* Arrays to store graph vertex ids */
@@ -66,7 +66,7 @@ int MPIR_TSP_Ireduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int co
     is_commutative = MPIR_Op_is_commutative(op);
 
     /* calculate chunking information for pipelining */
-    MPII_Algo_calculate_pipeline_chunk_info(maxbytes, type_size, count, &num_chunks,
+    MPIR_Algo_calculate_pipeline_chunk_info(maxbytes, type_size, count, &num_chunks,
                                             &chunk_size_floor, &chunk_size_ceil);
     /* print chunking information */
     MPL_DBG_MSG_FMT(MPIR_DBG_COLL, VERBOSE, (MPL_DBG_FDEST,
@@ -85,7 +85,7 @@ int MPIR_TSP_Ireduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int co
 
     /* initialize the tree */
     my_tree.children = NULL;
-    mpi_errno = MPII_Treealgo_tree_create(rank, size, tree_type, k, tree_root, &my_tree);
+    mpi_errno = MPIR_Treealgo_tree_create(rank, size, tree_type, k, tree_root, &my_tree);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
     num_children = my_tree.num_children;
@@ -240,7 +240,7 @@ int MPIR_TSP_Ireduce_sched_intra_tree(const void *sendbuf, void *recvbuf, int co
         offset += msgsize;
     }
 
-    MPII_Treealgo_tree_free(&my_tree);
+    MPIR_Treealgo_tree_free(&my_tree);
 
   fn_exit:
     MPL_free(vtcs);

src/mpi/coll/iscatter/iscatter_tsp_tree_algos.h

@@ -35,7 +35,7 @@ int MPIR_TSP_Iscatter_sched_intra_tree(const void *sendbuf, int sendcount,
     void *tmp_buf = NULL;
     int recv_id;
     int tree_type;
-    MPII_Treealgo_tree_t my_tree, parents_tree;
+    MPIR_Treealgo_tree_t my_tree, parents_tree;
     int next_child;
     int num_children, *child_subtree_size = NULL, *child_data_offset = NULL;
     int offset, recv_size;
@@ -50,7 +50,7 @@ int MPIR_TSP_Iscatter_sched_intra_tree(const void *sendbuf, int sendcount,
         is_inplace = (recvbuf == MPI_IN_PLACE); /* For scatter, MPI_IN_PLACE is significant only at root */
 
     tree_type = MPIR_TREE_TYPE_KNOMIAL_1;       /* currently only tree_type=MPIR_TREE_TYPE_KNOMIAL_1 is supported for scatter */
-    mpi_errno = MPII_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
+    mpi_errno = MPIR_Treealgo_tree_create(rank, size, tree_type, k, root, &my_tree);
     if (mpi_errno)
         MPIR_ERR_POP(mpi_errno);
     num_children = my_tree.num_children;
@@ -85,7 +85,7 @@ int MPIR_TSP_Iscatter_sched_intra_tree(const void *sendbuf, int sendcount,
 
     /* get tree information of the parent */
     if (my_tree.parent != -1) {
-        MPII_Treealgo_tree_create(my_tree.parent, size, tree_type, k, root, &parents_tree);
+        MPIR_Treealgo_tree_create(my_tree.parent, size, tree_type, k, root, &parents_tree);
     } else {    /* initialize an empty children array */
         utarray_new(parents_tree.children, &ut_int_icd, MPL_MEM_COLL);
         parents_tree.num_children = 0;
@@ -118,7 +118,7 @@ int MPIR_TSP_Iscatter_sched_intra_tree(const void *sendbuf, int sendcount,
         recv_size += child_subtree_size[i];
     }
 
-    MPII_Treealgo_tree_free(&parents_tree);
+    MPIR_Treealgo_tree_free(&parents_tree);
 
     recv_size *= (lrank == 0) ? sendcount : recvcount;
     offset = (lrank == 0) ? sendcount : recvcount;
@@ -177,7 +177,7 @@ int MPIR_TSP_Iscatter_sched_intra_tree(const void *sendbuf, int sendcount,
                                  recvcount, recvtype, sched, 0, NULL);
     }
 
-    MPII_Treealgo_tree_free(&my_tree);
+    MPIR_Treealgo_tree_free(&my_tree);
 
   fn_exit:
     MPL_free(child_subtree_size);

src/mpi/errhan/errnames.txt

@@ -213,6 +213,9 @@ MPI_TYPECLASS_INTEGER, or MPI_TYPECLASS_COMPLEX
 **treetype: Invalid tree type used for initializing Tree algorithms
 **treetype %d: Invalid tree type (%d) used for initializing Tree algorithms
 
+**noizem: release_gather based intra-node collectives cannot be used without izem submodule. \
+Reconfigure mpich with --enable-izem=atomic --with-zm-prefix=yes
+
 # -- FIXME: Some (but not all) of the messages below this line have been used
 #---- The messages below this line haven't been used yet.
 #