X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d38fb65942fcf1a4bd05cb86c0b86b7bbbd3a47f..e8f8350f07c1b13dd8628f56d5a5050da18264d3:/src/smpi/smpi_coll.c

diff --git a/src/smpi/smpi_coll.c b/src/smpi/smpi_coll.c
index 54f1e2db8c..a12de9c56e 100644
--- a/src/smpi/smpi_coll.c
+++ b/src/smpi/smpi_coll.c
@@ -1,19 +1,16 @@
-/* $Id$tag */
-
 /* smpi_coll.c -- various optimized routing for collectives                   */
 
-/* Copyright (c) 2009 Stephane Genaud.                                        */
-/* All rights reserved.                                                       */
+/* Copyright (c) 2009, 2010. 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. */
+ * under the terms of the license (GNU LGPL) which comes with this package. */
 
 #include <stdio.h>
 #include <string.h>
 #include <assert.h>
 
 #include "private.h"
-#include "smpi_coll_private.h"
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
                                 "Logging specific to SMPI (coll)");
@@ -21,27 +18,28 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
 struct s_proc_tree {
   int PROCTREE_A;
   int numChildren;
-  int * child;
+  int *child;
   int parent;
   int me;
   int root;
   int isRoot;
 };
-typedef struct s_proc_tree * proc_tree_t;
+typedef struct s_proc_tree *proc_tree_t;
 
 /**
  * alloc and init
  **/
-static proc_tree_t alloc_tree(int arity) {
+static proc_tree_t alloc_tree(int arity)
+{
   proc_tree_t tree;
   int i;
 
   tree = xbt_new(struct s_proc_tree, 1);
   tree->PROCTREE_A = arity;
-  tree->isRoot = 0; 
+  tree->isRoot = 0;
   tree->numChildren = 0;
   tree->child = xbt_new(int, arity);
-  for(i = 0; i < arity; i++) {
+  for (i = 0; i < arity; i++) {
     tree->child[i] = -1;
   }
   tree->root = -1;
@@ -52,8 +50,9 @@ static proc_tree_t alloc_tree(int arity) {
 /**
  * free
  **/
-static void free_tree(proc_tree_t tree) {
-  xbt_free(tree->child );
+static void free_tree(proc_tree_t tree)
+{
+  xbt_free(tree->child);
   xbt_free(tree);
 }
 
@@ -62,22 +61,23 @@ static void free_tree(proc_tree_t tree) {
  * @param index the rank of the calling process
  * @param extent the total number of processes
  **/
-static void build_tree(int index, int extent, proc_tree_t* tree) {
+static void build_tree(int index, int extent, proc_tree_t * tree)
+{
   int places = (*tree)->PROCTREE_A * index;
   int i, ch, pr;
 
   (*tree)->me = index;
-  (*tree)->root = 0 ;
-  for(i = 1; i <= (*tree)->PROCTREE_A; i++) {
+  (*tree)->root = 0;
+  for (i = 1; i <= (*tree)->PROCTREE_A; i++) {
     ++places;
     ch = (*tree)->PROCTREE_A * index + i + (*tree)->root;
     ch %= extent;
-    if(places < extent) {
+    if (places < extent) {
       (*tree)->child[i - 1] = ch;
       (*tree)->numChildren++;
     }
   }
-  if(index == (*tree)->root) {
+  if (index == (*tree)->root) {
     (*tree)->isRoot = 1;
   } else {
     (*tree)->isRoot = 0;
@@ -89,29 +89,37 @@ static void build_tree(int index, int extent, proc_tree_t* tree) {
 /**
  * bcast
  **/
-static void tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, proc_tree_t tree) {
-  int system_tag = 999;  // used negative int but smpi_create_request() declares this illegal (to be checked)
+static void tree_bcast(void *buf, int count, MPI_Datatype datatype,
+                       int root, MPI_Comm comm, proc_tree_t tree)
+{
+  int system_tag = 999;         // used negative int but smpi_create_request() declares this illegal (to be checked)
   int rank, i;
-  MPI_Request* requests;
+  MPI_Request *requests;
 
   rank = smpi_comm_rank(comm);
   /* wait for data from my parent in the tree */
-  if(!tree->isRoot) {
-    DEBUG3("<%d> tree_bcast(): i am not root: recv from %d, tag=%d)",
+  if (!tree->isRoot) {
+    XBT_DEBUG("<%d> tree_bcast(): i am not root: recv from %d, tag=%d)",
            rank, tree->parent, system_tag + rank);
-    smpi_mpi_recv(buf, count, datatype, tree->parent, system_tag + rank, comm, MPI_STATUS_IGNORE);
+    smpi_mpi_recv(buf, count, datatype, tree->parent, system_tag + rank,
+                  comm, MPI_STATUS_IGNORE);
   }
   requests = xbt_new(MPI_Request, tree->numChildren);
-  DEBUG2("<%d> creates %d requests (1 per child)\n", rank, tree->numChildren);
+  XBT_DEBUG("<%d> creates %d requests (1 per child)", rank,
+         tree->numChildren);
   /* iniates sends to ranks lower in the tree */
-  for(i = 0; i < tree->numChildren; i++) {
-    if(tree->child[i] == -1) {
+  for (i = 0; i < tree->numChildren; i++) {
+    if (tree->child[i] == -1) {
       requests[i] = MPI_REQUEST_NULL;
     } else {
-      DEBUG3("<%d> send to <%d>, tag=%d", rank, tree->child[i], system_tag + tree->child[i]);
-      requests[i] = smpi_mpi_isend(buf, count, datatype, tree->child[i], system_tag + tree->child[i], comm);
+      XBT_DEBUG("<%d> send to <%d>, tag=%d", rank, tree->child[i],
+             system_tag + tree->child[i]);
+      requests[i] =
+          smpi_isend_init(buf, count, datatype, tree->child[i],
+                          system_tag + tree->child[i], comm);
     }
   }
+  smpi_mpi_startall(tree->numChildren, requests);
   smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE);
   xbt_free(requests);
 }
@@ -119,38 +127,48 @@ static void tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MP
 /**
  * anti-bcast
  **/
-static void tree_antibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, proc_tree_t tree) {
-  int system_tag = 999;  // used negative int but smpi_create_request() declares this illegal (to be checked)
+static void tree_antibcast(void *buf, int count, MPI_Datatype datatype,
+                           int root, MPI_Comm comm, proc_tree_t tree)
+{
+  int system_tag = 999;         // used negative int but smpi_create_request() declares this illegal (to be checked)
   int rank, i;
-  MPI_Request* requests;
+  MPI_Request *requests;
 
   rank = smpi_comm_rank(comm);
   // everyone sends to its parent, except root.
-  if(!tree->isRoot) {
-    DEBUG3("<%d> tree_antibcast(): i am not root: send to %d, tag=%d)",
+  if (!tree->isRoot) {
+    XBT_DEBUG("<%d> tree_antibcast(): i am not root: send to %d, tag=%d)",
            rank, tree->parent, system_tag + rank);
-    smpi_mpi_send(buf, count, datatype, tree->parent, system_tag + rank, comm);
+    smpi_mpi_send(buf, count, datatype, tree->parent, system_tag + rank,
+                  comm);
   }
   //every one receives as many messages as it has children
   requests = xbt_new(MPI_Request, tree->numChildren);
-  DEBUG2("<%d> creates %d requests (1 per child)\n", rank, tree->numChildren);
-  for(i = 0; i < tree->numChildren; i++) {
-    if(tree->child[i] == -1) {
+  XBT_DEBUG("<%d> creates %d requests (1 per child)", rank,
+         tree->numChildren);
+  for (i = 0; i < tree->numChildren; i++) {
+    if (tree->child[i] == -1) {
       requests[i] = MPI_REQUEST_NULL;
     } else {
-      DEBUG3("<%d> recv from <%d>, tag=%d", rank, tree->child[i], system_tag + tree->child[i]);
-      requests[i] = smpi_mpi_irecv(buf, count, datatype, tree->child[i], system_tag + tree->child[i], comm);
+      XBT_DEBUG("<%d> recv from <%d>, tag=%d", rank, tree->child[i],
+             system_tag + tree->child[i]);
+      requests[i] =
+          smpi_irecv_init(buf, count, datatype, tree->child[i],
+                          system_tag + tree->child[i], comm);
     }
   }
+  smpi_mpi_startall(tree->numChildren, requests);
   smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE);
   xbt_free(requests);
-} 
+}
 
 /**
- * bcast with a binary, ternary, or whatever tree .. 
+ * bcast with a binary, ternary, or whatever tree ..
  **/
-void nary_tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, int arity) {
-  proc_tree_t tree = alloc_tree(arity); 
+void nary_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
+                     MPI_Comm comm, int arity)
+{
+  proc_tree_t tree = alloc_tree(arity);
   int rank, size;
 
   rank = smpi_comm_rank(comm);
@@ -161,12 +179,13 @@ void nary_tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_
 }
 
 /**
- * barrier with a binary, ternary, or whatever tree .. 
+ * barrier with a binary, ternary, or whatever tree ..
  **/
-void nary_tree_barrier(MPI_Comm comm, int arity) {
-  proc_tree_t tree = alloc_tree( arity ); 
+void nary_tree_barrier(MPI_Comm comm, int arity)
+{
+  proc_tree_t tree = alloc_tree(arity);
   int rank, size;
-  char dummy='$';
+  char dummy = '$';
 
   rank = smpi_comm_rank(comm);
   size = smpi_comm_size(comm);
@@ -177,60 +196,125 @@ void nary_tree_barrier(MPI_Comm comm, int arity) {
 }
 
 /**
- * Alltoall Bruck 
+ * Alltoall Bruck
  *
  * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12
  **/
-int smpi_coll_tuned_alltoall_bruck(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
-  DEBUG0("coll:tuned:alltoall_intra_bruck ** NOT IMPLEMENTED YET**");
+int smpi_coll_tuned_alltoall_bruck(void *sendbuf, int sendcount,
+                                   MPI_Datatype sendtype, void *recvbuf,
+                                   int recvcount, MPI_Datatype recvtype,
+                                   MPI_Comm comm)
+{
+  int system_tag = 777;
+  int i, rank, size, err, count;
+  MPI_Aint lb;
+  MPI_Aint sendextent = 0;
+  MPI_Aint recvextent = 0;
+  MPI_Request *requests;
+
+  // FIXME: check implementation
+  rank = smpi_comm_rank(comm);
+  size = smpi_comm_size(comm);
+  XBT_DEBUG("<%d> algorithm alltoall_bruck() called.", rank);
+  err = smpi_datatype_extent(sendtype, &lb, &sendextent);
+  err = smpi_datatype_extent(recvtype, &lb, &recvextent);
+  /* Local copy from self */
+  err =
+      smpi_datatype_copy(&((char *) sendbuf)[rank * sendextent], sendcount,
+                         sendtype, &((char *) recvbuf)[rank * recvextent],
+                         recvcount, recvtype);
+  if (err == MPI_SUCCESS && size > 1) {
+    /* Initiate all send/recv to/from others. */
+    requests = xbt_new(MPI_Request, 2 * (size - 1));
+    count = 0;
+    /* Create all receives that will be posted first */
+    for (i = 0; i < size; ++i) {
+      if (i == rank) {
+        XBT_DEBUG("<%d> skip request creation [src = %d, recvcount = %d]",
+               rank, i, recvcount);
+        continue;
+      }
+      requests[count] =
+          smpi_irecv_init(&((char *) recvbuf)[i * recvextent], recvcount,
+                          recvtype, i, system_tag, comm);
+      count++;
+    }
+    /* Now create all sends  */
+    for (i = 0; i < size; ++i) {
+      if (i == rank) {
+        XBT_DEBUG("<%d> skip request creation [dst = %d, sendcount = %d]",
+               rank, i, sendcount);
+        continue;
+      }
+      requests[count] =
+          smpi_isend_init(&((char *) sendbuf)[i * sendextent], sendcount,
+                          sendtype, i, system_tag, comm);
+      count++;
+    }
+    /* Wait for them all. */
+    smpi_mpi_startall(count, requests);
+    XBT_DEBUG("<%d> wait for %d requests", rank, count);
+    smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
+    xbt_free(requests);
+  }
   return MPI_SUCCESS;
 }
 
 /**
  * Alltoall basic_linear
  **/
-int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
+int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount,
+                                          MPI_Datatype sendtype,
+                                          void *recvbuf, int recvcount,
+                                          MPI_Datatype recvtype,
+                                          MPI_Comm comm)
+{
   int system_tag = 888;
   int i, rank, size, err, count;
-  MPI_Aint lb, sendinc, recvinc;
+  MPI_Aint lb;
+  MPI_Aint sendinc = 0;
+  MPI_Aint recvinc = 0;
   MPI_Request *requests;
 
   /* Initialize. */
   rank = smpi_comm_rank(comm);
   size = smpi_comm_size(comm);
-  DEBUG1("<%d> algorithm alltoall_basic_linear() called.", rank);
+  XBT_DEBUG("<%d> algorithm alltoall_basic_linear() called.", rank);
   err = smpi_datatype_extent(sendtype, &lb, &sendinc);
   err = smpi_datatype_extent(recvtype, &lb, &recvinc);
   sendinc *= sendcount;
   recvinc *= recvcount;
   /* simple optimization */
-  err = smpi_datatype_copy(&((char*)sendbuf)[rank * sendinc], sendcount, sendtype, &((char*)recvbuf)[rank * recvinc], recvcount, recvtype);
-  if(err == MPI_SUCCESS && size > 1) {
+  err =
+      smpi_datatype_copy(&((char *) sendbuf)[rank * sendinc], sendcount,
+                         sendtype, &((char *) recvbuf)[rank * recvinc],
+                         recvcount, recvtype);
+  if (err == MPI_SUCCESS && size > 1) {
     /* Initiate all send/recv to/from others. */
     requests = xbt_new(MPI_Request, 2 * (size - 1));
     /* Post all receives first -- a simple optimization */
     count = 0;
-    for(i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
-      requests[count] = smpi_mpi_irecv(&((char*)recvbuf)[i * recvinc], recvcount, recvtype, i, system_tag, comm);
+    for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
+      requests[count] =
+          smpi_irecv_init(&((char *) recvbuf)[i * recvinc], recvcount,
+                          recvtype, i, system_tag, comm);
       count++;
     }
-    /* Now post all sends in reverse order 
+    /* Now post all sends in reverse order
      *   - We would like to minimize the search time through message queue
      *     when messages actually arrive in the order in which they were posted.
      * TODO: check the previous assertion
      */
-    for(i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size ) {
-      requests[count] = smpi_mpi_isend(&((char*)sendbuf)[i * sendinc], sendcount, sendtype, i, system_tag, comm);
+    for (i = (rank + size - 1) % size; i != rank;
+         i = (i + size - 1) % size) {
+      requests[count] =
+          smpi_isend_init(&((char *) sendbuf)[i * sendinc], sendcount,
+                          sendtype, i, system_tag, comm);
       count++;
     }
-    /* Wait for them all.  If there's an error, note that we don't
-     * care what the error was -- just that there *was* an error.  The
-     * PML will finish all requests, even if one or more of them fail.
-     * i.e., by the end of this call, all the requests are free-able.
-     * So free them anyway -- even if there was an error, and return
-     * the error after we free everything.
-     */
-    DEBUG2("<%d> wait for %d requests", rank, count);
+    /* Wait for them all. */
+    smpi_mpi_startall(count, requests);
+    XBT_DEBUG("<%d> wait for %d requests", rank, count);
     smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
     xbt_free(requests);
   }
@@ -242,75 +326,97 @@ int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount, MPI_Data
  *
  * this algorithm performs size steps (1<=s<=size) and
  * at each step s, a process p sends iand receive to.from a unique distinct remote process
- * size=5 : s=1:  4->0->1, 0->1->2, 1->2->3, ... 
+ * size=5 : s=1:  4->0->1, 0->1->2, 1->2->3, ...
  *          s=2:  3->0->2, 4->1->3, 0->2->4, 1->3->0 , 2->4->1
- *          .... 
+ *          ....
  * Openmpi calls this routine when the message size sent to each rank is greater than 3000 bytes
  **/
-int smpi_coll_tuned_alltoall_pairwise(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
+int smpi_coll_tuned_alltoall_pairwise(void *sendbuf, int sendcount,
+                                      MPI_Datatype sendtype, void *recvbuf,
+                                      int recvcount, MPI_Datatype recvtype,
+                                      MPI_Comm comm)
+{
   int system_tag = 999;
   int rank, size, step, sendto, recvfrom, sendsize, recvsize;
 
   rank = smpi_comm_rank(comm);
   size = smpi_comm_size(comm);
-  DEBUG1("<%d> algorithm alltoall_pairwise() called.", rank);
+  XBT_DEBUG("<%d> algorithm alltoall_pairwise() called.", rank);
   sendsize = smpi_datatype_size(sendtype);
   recvsize = smpi_datatype_size(recvtype);
   /* Perform pairwise exchange - starting from 1 so the local copy is last */
-  for(step = 1; step < size + 1; step++) {
+  for (step = 1; step < size + 1; step++) {
     /* who do we talk to in this step? */
-    sendto  = (rank+step)%size;
-    recvfrom = (rank+size-step)%size;
+    sendto = (rank + step) % size;
+    recvfrom = (rank + size - step) % size;
     /* send and receive */
-    smpi_mpi_sendrecv(&((char*)sendbuf)[sendto * sendsize * sendcount], sendcount, sendtype, sendto, system_tag, &((char*)recvbuf)[recvfrom * recvsize * recvcount], recvcount, recvtype, recvfrom, system_tag, comm, MPI_STATUS_IGNORE);
+    smpi_mpi_sendrecv(&((char *) sendbuf)[sendto * sendsize * sendcount],
+                      sendcount, sendtype, sendto, system_tag,
+                      &((char *) recvbuf)[recvfrom * recvsize * recvcount],
+                      recvcount, recvtype, recvfrom, system_tag, comm,
+                      MPI_STATUS_IGNORE);
   }
   return MPI_SUCCESS;
 }
 
-int smpi_coll_basic_alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm) {
+int smpi_coll_basic_alltoallv(void *sendbuf, int *sendcounts,
+                              int *senddisps, MPI_Datatype sendtype,
+                              void *recvbuf, int *recvcounts,
+                              int *recvdisps, MPI_Datatype recvtype,
+                              MPI_Comm comm)
+{
   int system_tag = 889;
   int i, rank, size, err, count;
-  MPI_Aint lb, sendextent, recvextent;
-  MPI_Request* requests;
+  MPI_Aint lb;
+  MPI_Aint sendextent = 0;
+  MPI_Aint recvextent = 0;
+  MPI_Request *requests;
 
   /* Initialize. */
   rank = smpi_comm_rank(comm);
   size = smpi_comm_size(comm);
-  DEBUG1("<%d> algorithm basic_alltoallv() called.", rank);
+  XBT_DEBUG("<%d> algorithm basic_alltoallv() called.", rank);
   err = smpi_datatype_extent(sendtype, &lb, &sendextent);
   err = smpi_datatype_extent(recvtype, &lb, &recvextent);
   /* Local copy from self */
-  err = smpi_datatype_copy(&((char*)sendbuf)[senddisps[rank] * sendextent], sendcounts[rank], sendtype, &((char*)recvbuf)[recvdisps[rank] * recvextent], recvcounts[rank], recvtype);
-  if(err == MPI_SUCCESS && size > 1) {
+  err =
+      smpi_datatype_copy(&((char *) sendbuf)[senddisps[rank] * sendextent],
+                         sendcounts[rank], sendtype,
+                         &((char *) recvbuf)[recvdisps[rank] * recvextent],
+                         recvcounts[rank], recvtype);
+  if (err == MPI_SUCCESS && size > 1) {
     /* Initiate all send/recv to/from others. */
     requests = xbt_new(MPI_Request, 2 * (size - 1));
     count = 0;
     /* Create all receives that will be posted first */
-    for(i = 0; i < size; ++i) {
-      if(i == rank || recvcounts[i] == 0) {
-        DEBUG3("<%d> skip request creation [src = %d, recvcounts[src] = %d]", rank, i, recvcounts[i]);
+    for (i = 0; i < size; ++i) {
+      if (i == rank || recvcounts[i] == 0) {
+        XBT_DEBUG
+            ("<%d> skip request creation [src = %d, recvcounts[src] = %d]",
+             rank, i, recvcounts[i]);
         continue;
       }
-      requests[count] = smpi_mpi_irecv(&((char*)recvbuf)[recvdisps[i] * recvextent], recvcounts[i], recvtype, i, system_tag, comm);
+      requests[count] =
+          smpi_irecv_init(&((char *) recvbuf)[recvdisps[i] * recvextent],
+                          recvcounts[i], recvtype, i, system_tag, comm);
       count++;
     }
     /* Now create all sends  */
-    for(i = 0; i < size; ++i) {
-      if(i == rank || sendcounts[i] == 0) {
-        DEBUG3("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]", rank, i, sendcounts[i]);
+    for (i = 0; i < size; ++i) {
+      if (i == rank || sendcounts[i] == 0) {
+        XBT_DEBUG
+            ("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]",
+             rank, i, sendcounts[i]);
         continue;
       }
-      requests[count] = smpi_mpi_isend(&((char*)sendbuf)[senddisps[i] * sendextent], sendcounts[i], sendtype, i, system_tag, comm);
+      requests[count] =
+          smpi_isend_init(&((char *) sendbuf)[senddisps[i] * sendextent],
+                          sendcounts[i], sendtype, i, system_tag, comm);
       count++;
     }
-    /* Wait for them all.  If there's an error, note that we don't
-     * care what the error was -- just that there *was* an error.  The
-     * PML will finish all requests, even if one or more of them fail.
-     * i.e., by the end of this call, all the requests are free-able.
-     * So free them anyway -- even if there was an error, and return
-     * the error after we free everything.
-     */
-    DEBUG2("<%d> wait for %d requests", rank, count);
+    /* Wait for them all. */
+    smpi_mpi_startall(count, requests);
+    XBT_DEBUG("<%d> wait for %d requests", rank, count);
     smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
     xbt_free(requests);
   }