coll: merge barrier dissemination to the high radix dissemination

Make the optimized version of dissemination algorithm to be a
special case of the high radix dissemination algorithm.

maint/tuning/coll/mpir/generic.json

@@ -421,7 +421,9 @@
             },
             "comm_hierarchy=any":
             {
-                "algorithm=MPIR_Barrier_intra_dissemination":{}
+                "algorithm=MPIR_Barrier_intra_k_dissemination":{
+                  "k=2": {}
+                }
             }
         },
         "comm_type=inter":

src/include/mpir_coll.h

@@ -53,6 +53,8 @@ int MPIC_Waitall(int numreq, MPIR_Request * requests[], MPI_Status statuses[],
 int MPIR_Reduce_local(const void *inbuf, void *inoutbuf, MPI_Aint count, MPI_Datatype datatype,
                       MPI_Op op);
 
+int MPIR_Barrier_intra_dissemination(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag);
+
 /* TSP auto */
 int MPIR_TSP_Iallreduce_sched_intra_tsp_auto(const void *sendbuf, void *recvbuf, MPI_Aint count,
                                              MPI_Datatype datatype, MPI_Op op,

src/mpi/coll/barrier/Makefile.mk

@@ -9,7 +9,6 @@
 
 mpi_core_sources +=										\
     src/mpi/coll/barrier/barrier_allcomm_nb.c	\
-    src/mpi/coll/barrier/barrier_intra_dissemination.c	\
     src/mpi/coll/barrier/barrier_intra_k_dissemination.c	\
     src/mpi/coll/barrier/barrier_intra_recexch.c	\
     src/mpi/coll/barrier/barrier_intra_smp.c				\

src/mpi/coll/barrier/barrier_intra_dissemination.c

@@ -1,50 +0,0 @@
-/*
- * Copyright (C) by Argonne National Laboratory
- *     See COPYRIGHT in top-level directory
- */
-
-#include "mpiimpl.h"
-
-/* Algorithm: MPI_Barrier
- *
- * We use the dissemination algorithm described in:
- * Debra Hensgen, Raphael Finkel, and Udi Manbet, "Two Algorithms for
- * Barrier Synchronization," International Journal of Parallel
- * Programming, 17(1):1-17, 1988.
- *
- * It uses ceiling(lgp) steps. In step k, 0 <= k <= (ceiling(lgp)-1),
- * process i sends to process (i + 2^k) % p and receives from process
- * (i - 2^k + p) % p.
- */
-int MPIR_Barrier_intra_dissemination(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
-{
-    int size, rank, src, dst, mask, mpi_errno = MPI_SUCCESS;
-    int mpi_errno_ret = MPI_SUCCESS;
-
-    size = comm_ptr->local_size;
-    rank = comm_ptr->rank;
-
-    mask = 0x1;
-    while (mask < size) {
-        dst = (rank + mask) % size;
-        src = (rank - mask + size) % size;
-        mpi_errno = MPIC_Sendrecv(NULL, 0, MPI_BYTE, dst,
-                                  MPIR_BARRIER_TAG, NULL, 0, MPI_BYTE,
-                                  src, MPIR_BARRIER_TAG, comm_ptr, MPI_STATUS_IGNORE, errflag);
-        if (mpi_errno) {
-            /* for communication errors, just record the error but continue */
-            *errflag =
-                MPIX_ERR_PROC_FAILED ==
-                MPIR_ERR_GET_CLASS(mpi_errno) ? MPIR_ERR_PROC_FAILED : MPIR_ERR_OTHER;
-            MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
-            MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
-        }
-        mask <<= 1;
-    }
-
-    if (mpi_errno_ret)
-        mpi_errno = mpi_errno_ret;
-    else if (*errflag != MPIR_ERR_NONE)
-        MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
-    return mpi_errno;
-}

src/mpi/coll/barrier/barrier_intra_k_dissemination.c

@@ -5,6 +5,42 @@
 
 #include "mpiimpl.h"
 
+/* Algorithm: MPI_Barrier
+ *
+ * We use the dissemination algorithm described in:
+ * Debra Hensgen, Raphael Finkel, and Udi Manbet, "Two Algorithms for
+ * Barrier Synchronization," International Journal of Parallel
+ * Programming, 17(1):1-17, 1988.
+ *
+ * It uses ceiling(lgp) steps. In step k, 0 <= k <= (ceiling(lgp)-1),
+ * process i sends to process (i + 2^k) % p and receives from process
+ * (i - 2^k + p) % p.
+ */
+int MPIR_Barrier_intra_dissemination(MPIR_Comm * comm_ptr, MPIR_Errflag_t * errflag)
+{
+    int size, rank, src, dst, mask, mpi_errno = MPI_SUCCESS;
+    int mpi_errno_ret = MPI_SUCCESS;
+
+    size = comm_ptr->local_size;
+    rank = comm_ptr->rank;
+
+    mask = 0x1;
+    while (mask < size) {
+        dst = (rank + mask) % size;
+        src = (rank - mask + size) % size;
+        mpi_errno = MPIC_Sendrecv(NULL, 0, MPI_BYTE, dst,
+                                  MPIR_BARRIER_TAG, NULL, 0, MPI_BYTE,
+                                  src, MPIR_BARRIER_TAG, comm_ptr, MPI_STATUS_IGNORE, errflag);
+        MPIR_ERR_COLL_CHECKANDCONT(mpi_errno, *errflag);
+        mask <<= 1;
+    }
+
+    if (mpi_errno_ret)
+        mpi_errno = mpi_errno_ret;
+    else if (*errflag != MPIR_ERR_NONE)
+        MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
+    return mpi_errno;
+}
 
 /* Algorithm: high radix dissemination
  * Similar to dissemination algorithm, but generalized with high radix k
@@ -28,6 +64,10 @@ int MPIR_Barrier_intra_k_dissemination(MPIR_Comm * comm, int k, MPIR_Errflag_t *
     if (nranks < k)
         k = nranks;
 
+    if (k == 2) {
+        return MPIR_Barrier_intra_dissemination(comm, errflag);
+    }
+
     /* If k value is greater than the maximum radix defined by MAX_RADIX macro,
      * we allocate memory for requests here. Otherwise we use the requests defined
      * in the communicator for allreduce/barrier recexch */

src/mpi/coll/coll_algorithms.txt

@@ -41,7 +41,6 @@
 #         Some algorithm use another algorithm or use a different function name.
 
 barrier-intra:
-    dissemination
     k_dissemination
         extra_params: k
         cvar_params: DISSEM_KVAL

src/mpi/coll/cvars.txt

@@ -19,7 +19,6 @@ cvars:
         auto - Internal algorithm selection (can be overridden with MPIR_CVAR_COLL_SELECTION_TUNING_JSON_FILE)
         nb            - Force nonblocking algorithm
         smp           - Force smp algorithm
-        dissemination - Force dissemination algorithm
         k_dissemination - Force high radix dissemination algorithm
         recexch       - Force recursive exchange algorithm
 

src/mpi/coll/include/csel_container.h

@@ -44,7 +44,6 @@ typedef enum {
     MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Alltoallw_intra_scattered,
     MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Alltoallw_inter_pairwise_exchange,
     MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Alltoallw_allcomm_nb,
-    MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Barrier_intra_dissemination,
     MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Barrier_intra_k_dissemination,
     MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Barrier_intra_recexch,
     MPII_CSEL_CONTAINER_TYPE__ALGORITHM__MPIR_Barrier_intra_smp,

test/mpi/maint/coll_cvars.txt

@@ -308,7 +308,6 @@ algorithms:
         inter-nonblocking:
             sched_bcast
         intra-blocking:
-            dissemination
             k_dissemination
                 .MPIR_CVAR_BARRIER_DISSEM_KVAL=2,3,4,8
             recexch