#include <stdio.h>
#include <stdlib.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)");
/* proc_tree taken and translated from P2P-MPI */
/* wait for data from my parent in the tree */
if (!tree->isRoot) {
-#ifdef DEBUG_STEPH
- printf("[%d] recv(%d from %d, tag=%d)\n",rank,rank, tree->parent, system_tag+rank);
-#endif
+ DEBUG3("<%d> tree_bcast(): i am not root: recv from %d, tag=%d)",rank,tree->parent,system_tag+rank);
retval = smpi_create_request(buf, count, datatype,
tree->parent, rank,
system_tag + rank,
rank,retval,__FILE__,__LINE__);
}
smpi_mpi_irecv(request);
-#ifdef DEBUG_STEPH
- printf("[%d] waiting on irecv from %d\n",rank , tree->parent);
-#endif
+ DEBUG2("<%d> tree_bcast(): waiting on irecv from %d",rank, tree->parent);
smpi_mpi_wait(request, MPI_STATUS_IGNORE);
xbt_mallocator_release(smpi_global->request_mallocator, request);
}
requests = xbt_malloc( tree->numChildren * sizeof(smpi_mpi_request_t));
-#ifdef DEBUG_STEPH
- printf("[%d] creates %d requests\n",rank,tree->numChildren);
-#endif
+ DEBUG2("<%d> creates %d requests (1 per child)\n",rank,tree->numChildren);
/* iniates sends to ranks lower in the tree */
for (i=0; i < tree->numChildren; i++) {
if (tree->child[i] != -1) {
-#ifdef DEBUG_STEPH
- printf("[%d] send(%d->%d, tag=%d)\n",rank,rank, tree->child[i], system_tag+tree->child[i]);
-#endif
+ DEBUG3("<%d> send to <%d>,tag=%d",rank,tree->child[i], system_tag+tree->child[i]);
retval = smpi_create_request(buf, count, datatype,
rank, tree->child[i],
system_tag + tree->child[i],
comm, &(requests[i]));
-#ifdef DEBUG_STEPH
- printf("[%d] after create req[%d]=%p req->(src=%d,dst=%d)\n",rank , i, requests[i],requests[i]->src,requests[i]->dst );
-#endif
if (MPI_SUCCESS != retval) {
printf("** internal error: smpi_create_request() rank=%d returned retval=%d, %s:%d\n",
rank,retval,__FILE__,__LINE__);
int retval;
rank = smpi_mpi_comm_rank( comm );
+ DEBUG2("<%d> entered nary_tree_bcast(), arity=%d",rank,arity);
// arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
proc_tree_t tree = alloc_tree( arity );
build_tree( rank, comm->size, &tree );
/**
* Barrier
**/
-
int nary_tree_barrier( MPI_Comm comm , int arity)
{
int rank;
}
+/**
+ * Alltoall pairwise
+ *
+ * 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, ...
+ * 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 datatype,
void* recvbuf, int recvcount, MPI_Datatype recvdatatype, MPI_Comm comm)
void * tmpsend, *tmprecv;
rank = smpi_mpi_comm_rank(comm);
+ INFO1("<%d> algorithm alltoall_pairwise() called.\n",rank);
+
+
/* Perform pairwise exchange - starting from 1 so the local copy is last */
for (step = 1; step < size+1; step++) {
comm, MPI_STATUS_IGNORE);
}
return(retval);
+}
+/**
+ * helper: copy a datatype into another (in the simple case dt1=dt2)
+**/
+int copy_dt( void *sbuf, int scount, const MPI_Datatype sdtype, void *rbuf, int rcount, const MPI_Datatype rdtype);
+int copy_dt( void *sbuf, int scount, const MPI_Datatype sdtype,
+ void *rbuf, int rcount, const MPI_Datatype rdtype)
+{
+ /* First check if we really have something to do */
+ if (0 == rcount) {
+ return ((0 == scount) ? MPI_SUCCESS : MPI_ERR_TRUNCATE);
+ }
+ /* If same datatypes used, just copy. */
+ if (sdtype == rdtype) {
+ int count = ( scount < rcount ? scount : rcount );
+ memcpy( rbuf, sbuf, sdtype->size*count);
+ return ((scount > rcount) ? MPI_ERR_TRUNCATE : MPI_SUCCESS);
+ }
+ /* FIXME: cases
+ * - If receive packed.
+ * - If send packed
+ * to be treated once we have the MPI_Pack things ...
+ **/
+ return( MPI_SUCCESS );
}
+/**
+ *
+ **/
+int smpi_coll_tuned_alltoall_basic_linear(void *sbuf, int scount, MPI_Datatype sdtype,
+ void* rbuf, int rcount, MPI_Datatype rdtype, MPI_Comm comm)
+{
+ int i;
+ int system_alltoall_tag = 888;
+ int rank;
+ int size = comm->size;
+ int err;
+ char *psnd;
+ char *prcv;
+ int nreq = 0;
+ MPI_Aint lb;
+ MPI_Aint sndinc;
+ MPI_Aint rcvinc;
+ MPI_Request *reqs;
+
+ /* Initialize. */
+ rank = smpi_mpi_comm_rank(comm);
+ DEBUG1("<%d> algorithm alltoall_basic_linear() called.",rank);
+
+ err = smpi_mpi_type_get_extent(sdtype, &lb, &sndinc);
+ err = smpi_mpi_type_get_extent(rdtype, &lb, &rcvinc);
+ sndinc *= scount;
+ rcvinc *= rcount;
+ /* simple optimization */
+ psnd = ((char *) sbuf) + (rank * sndinc);
+ prcv = ((char *) rbuf) + (rank * rcvinc);
+
+ err = copy_dt( psnd, scount, sdtype, prcv, rcount, rdtype );
+ if (MPI_SUCCESS != err) {
+ return err;
+ }
+
+ /* If only one process, we're done. */
+ if (1 == size) {
+ return MPI_SUCCESS;
+ }
+
+ /* Initiate all send/recv to/from others. */
+ reqs = xbt_malloc(2*(size-1) * sizeof(smpi_mpi_request_t));
+
+ prcv = (char *) rbuf;
+ psnd = (char *) sbuf;
+
+ /* Post all receives first -- a simple optimization */
+ for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
+ err = smpi_create_request( prcv + (i * rcvinc), rcount, rdtype,
+ i, rank,
+ system_alltoall_tag,
+ comm, &(reqs[nreq]));
+ if (MPI_SUCCESS != err) {
+ DEBUG2("<%d> failed to create request for rank %d",rank,i);
+ for (i=0;i< nreq;i++)
+ xbt_mallocator_release(smpi_global->request_mallocator, reqs[i]);
+ return err;
+ }
+ nreq++;
+ }
+ /* 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.
+ * */
+ for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size ) {
+ err = smpi_create_request (psnd + (i * sndinc), scount, sdtype,
+ rank, i,
+ system_alltoall_tag,
+ comm, &(reqs[nreq]));
+ if (MPI_SUCCESS != err) {
+ DEBUG2("<%d> failed to create request for rank %d\n",rank,i);
+ for (i=0;i< nreq;i++)
+ xbt_mallocator_release(smpi_global->request_mallocator, reqs[i]);
+ return err;
+ }
+ nreq++;
+ }
+
+ /* Start your engines. This will never return an error. */
+ for ( i=0; i< nreq/2; i++ ) {
+ DEBUG3("<%d> issued irecv request reqs[%d]=%p",rank,i,reqs[i]);
+ smpi_mpi_irecv( reqs[i] );
+ }
+ for ( i= nreq/2; i<nreq; i++ ) {
+ DEBUG3("<%d> issued isend request reqs[%d]=%p",rank,i,reqs[i]);
+ smpi_mpi_isend( reqs[i] );
+ }
+
+
+ /* 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,nreq);
+ // waitall is buggy: use a loop instead for the moment
+ // err = smpi_mpi_waitall(nreq, reqs, MPI_STATUS_IGNORE);
+ for (i=0;i<nreq;i++) {
+ err = smpi_mpi_wait( reqs[i], MPI_STATUS_IGNORE);
+ }
+
+ /* Free the reqs */
+ assert( nreq == 2*(size-1) );
+ for (i=0;i< nreq;i++) {
+ xbt_mallocator_release(smpi_global->request_mallocator, reqs[i]);
+ }
+ xbt_free( reqs );
+ return err;
+}
+
+
+/**
+ * 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 sdtype,
void* recvbuf, int recvcount, MPI_Datatype rdtype, MPI_Comm comm)
{
- /*
+/* int size = comm->size;
int i, k, line = -1;
- int rank, size;
int sendto, recvfrom, distance, *displs=NULL, *blen=NULL;
int maxpacksize, packsize, position;
char * tmpbuf=NULL, *packbuf=NULL;
int weallocated = 0;
MPI_Datatype iddt;
- size = ompi_comm_size(comm);
- rank = ompi_comm_rank(comm);
-
- OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:alltoall_intra_bruck rank %d", rank));
-
- err = ompi_ddt_get_extent (sdtype, &lb, &sext);
- if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
-
- err = ompi_ddt_get_extent (rdtype, &lb, &rext);
- if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
-
+ rank = smpi_mpi_comm_rank(comm);
+*/
+ INFO0("coll:tuned:alltoall_intra_bruck ** NOT IMPLEMENTED YET**");
+/*
+ displs = xbt_malloc(size*sizeof(int));
+ blen = xbt_malloc(size*sizeof(int));
- displs = (int *) malloc(size*sizeof(int));
- if (displs == NULL) { line = __LINE__; err = -1; goto err_hndl; }
- blen = (int *) malloc(size*sizeof(int));
- if (blen == NULL) { line = __LINE__; err = -1; goto err_hndl; }
weallocated = 1;
*/
/* Prepare for packing data */
- /*err = MPI_Pack_size( scount*size, sdtype, comm, &maxpacksize );
- if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
+/*
+ err = MPI_Pack_size( scount*size, sdtype, comm, &maxpacksize );
+ if (err != MPI_SUCCESS) { }
*/
/* pack buffer allocation */
/* packbuf = (char*) malloc((unsigned) maxpacksize);
- if (packbuf == NULL) { line = __LINE__; err = -1; goto err_hndl; }
+ if (packbuf == NULL) { }
*/
/* tmp buffer allocation for message data */
-/* tmpbuf = (char *) malloc(scount*size*sext);
- if (tmpbuf == NULL) { line = __LINE__; err = -1; goto err_hndl; }
+/* tmpbuf = xbt_malloc(scount*size*sext);
+ if (tmpbuf == NULL) { }
*/
/* Step 1 - local rotation - shift up by rank */
}
return err;
*/
- int NOTYET=1;
- return NOTYET;
+ return -1; /* FIXME: to be changed*/
}
