Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
Move collective algorithms to separate folders
[simgrid.git] / src / smpi / colls / allgather / allgather-bruck.cpp
diff --git a/src/smpi/colls/allgather/allgather-bruck.cpp b/src/smpi/colls/allgather/allgather-bruck.cpp
new file mode 100644 (file)
index 0000000..68edee2
--- /dev/null
@@ -0,0 +1,135 @@
+/* Copyright (c) 2013-2014. The SimGrid Team.
+ * All rights reserved.                                                     */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#include "../colls_private.h"
+
+/*****************************************************************************
+
+Copyright (c) 2006, Ahmad Faraj & Xin Yuan,
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+  * Redistributions of source code must retain the above copyright notice,
+    this list of conditions and the following disclaimer.
+
+  * Redistributions in binary form must reproduce the above copyright notice,
+    this list of conditions and the following disclaimer in the documentation
+    and/or other materials provided with the distribution.
+
+  * Neither the name of the Florida State University nor the names of its
+    contributors may be used to endorse or promote products derived from this
+    software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+  *************************************************************************
+  *     Any results obtained from executing this software require the     *
+  *     acknowledgment and citation of the software and its owners.       *
+  *     The full citation is given below:                                 *
+  *                                                                       *
+  *     A. Faraj and X. Yuan. "Automatic Generation and Tuning of MPI     *
+  *     Collective Communication Routines." The 19th ACM International    *
+  *     Conference on Supercomputing (ICS), Cambridge, Massachusetts,     *
+  *     June 20-22, 2005.                                                 *
+  *************************************************************************
+
+*****************************************************************************/
+
+
+/*****************************************************************************
+ * Function: allgather_bruck
+ * return: int
+ * inputs:
+ *   send_buff: send input buffer
+ *   send_count: number of elements to send
+ *   send_type: data type of elements being sent
+ *   recv_buff: receive output buffer
+ *   recv_count: number of elements to received
+ *   recv_type: data type of elements being received
+ *   comm: communication
+ * Descrp: Function realizes the allgather operation using the bruck
+ *         algorithm.
+ * Auther: MPICH
+ * Comment: Original bruck algorithm from MPICH is slightly modified by
+ *          Ahmad Faraj.  
+ ****************************************************************************/
+int smpi_coll_tuned_allgather_bruck(void *send_buff, int send_count,
+                                    MPI_Datatype send_type, void *recv_buff,
+                                    int recv_count, MPI_Datatype recv_type,
+                                    MPI_Comm comm)
+{
+  // MPI variables
+  MPI_Status status;
+  MPI_Aint recv_extent;
+
+  // local int variables
+  int src, dst, rank, num_procs, count, remainder;
+  int tag = COLL_TAG_ALLGATHER;
+  int pof2 = 1;
+
+  // local string variables
+  char *tmp_buff;
+  char *send_ptr = (char *) send_buff;
+  char *recv_ptr = (char *) recv_buff;
+
+  // get size of the communicator, followed by rank 
+  num_procs = comm->size();
+  rank = comm->rank();
+
+  // get size of single element's type for recv buffer
+  recv_extent = recv_type->get_extent();
+
+  count = recv_count;
+
+  tmp_buff = (char *) smpi_get_tmp_sendbuffer(num_procs * recv_count * recv_extent);
+
+  // perform a local copy
+  Datatype::copy(send_ptr, send_count, send_type,
+                    tmp_buff, recv_count, recv_type);
+  while (pof2 <= (num_procs / 2)) {
+    src = (rank + pof2) % num_procs;
+    dst = (rank - pof2 + num_procs) % num_procs;
+
+    Request::sendrecv(tmp_buff, count, recv_type, dst, tag,
+                  tmp_buff + count * recv_extent, count, recv_type,
+                  src, tag, comm, &status);
+    count *= 2;
+    pof2 *= 2;
+  }
+
+  remainder = num_procs - pof2;
+  if (remainder) {
+    src = (rank + pof2) % num_procs;
+    dst = (rank - pof2 + num_procs) % num_procs;
+
+    Request::sendrecv(tmp_buff, remainder * recv_count, recv_type, dst, tag,
+                  tmp_buff + count * recv_extent, remainder * recv_count,
+                  recv_type, src, tag, comm, &status);
+  }
+
+  Request::sendrecv(tmp_buff, (num_procs - rank) * recv_count, recv_type, rank,
+                tag, recv_ptr + rank * recv_count * recv_extent,
+                (num_procs - rank) * recv_count, recv_type, rank, tag, comm,
+                &status);
+
+  if (rank)
+    Request::sendrecv(tmp_buff + (num_procs - rank) * recv_count * recv_extent,
+                  rank * recv_count, recv_type, rank, tag, recv_ptr,
+                  rank * recv_count, recv_type, rank, tag, comm, &status);
+  smpi_free_tmp_buffer(tmp_buff);
+  return MPI_SUCCESS;
+}