#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <assert.h>
#include "private.h"
#include "smpi_coll_private.h"
/* 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__);
system_tag + rank,
comm, &request);
if (MPI_SUCCESS != retval) {
- printf("** internal error: smpi_create_request() rank=%d returned retval=%d, %s:%d\n",
+ ERROR4("** internal error: smpi_create_request() rank=%d returned retval=%d, %s:%d\n",
rank,retval,__FILE__,__LINE__);
}
smpi_mpi_isend(request);
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 );
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++) {
}
/**
- *
+ * Alltoall basic_linear
**/
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_tag = 999;
+ int system_alltoall_tag = 888;
int rank;
int size = comm->size;
int err;
/* 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);
for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
err = smpi_create_request( prcv + (i * rcvinc), rcount, rdtype,
i, rank,
- system_tag + i,
+ system_alltoall_tag,
comm, &(reqs[nreq]));
if (MPI_SUCCESS != err) {
- DEBUG2("[%d] failed to create request for rank %d\n",rank,i);
+ 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;
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_tag + i,
+ system_alltoall_tag,
comm, &(reqs[nreq]));
if (MPI_SUCCESS != err) {
- DEBUG2("[%d] failed to create request for rank %d\n",rank,i);
+ 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;
/* 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] );
}
* So free them anyway -- even if there was an error, and return
* the error after we free everything. */
- err = smpi_mpi_waitall(nreq, reqs, MPI_STATUS_IGNORE);
+ 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 */
- for (i=0;i< 2*(size-1);i++) {
+ 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;
}
return -1; /* FIXME: to be changed*/
}
+static void print_buffer_int(void *buf, int len, char *msg, int rank);
+static void print_buffer_int(void *buf, int len, char *msg, int rank)
+{
+ int tmp, *v;
+ fprintf(stderr,"**<%d> %s (#%d): ", rank, msg,len);
+ for (tmp = 0; tmp < len; tmp++) {
+ v = buf;
+ fprintf(stderr,"[%d (%p)]", v[tmp],v+tmp);
+ }
+ fprintf(stderr,"\n");
+ free(msg);
+}
+
+
+
+/**
+ * alltoallv basic
+ **/
+
+int smpi_coll_basic_alltoallv(void *sbuf, int *scounts, int *sdisps, MPI_Datatype sdtype,
+ void *rbuf, int *rcounts, int *rdisps, MPI_Datatype rdtype,
+ MPI_Comm comm) {
+
+ int i;
+ int system_alltoallv_tag = 889;
+ int rank;
+ int size = comm->size;
+ int err;
+ char *psnd;
+ char *prcv;
+ //int nreq = 0;
+ int rreq = 0;
+ int sreq = 0;
+ MPI_Aint lb;
+ MPI_Aint sndextent;
+ MPI_Aint rcvextent;
+ MPI_Request *reqs;
+
+ /* Initialize. */
+ rank = smpi_mpi_comm_rank(comm);
+ DEBUG1("<%d> algorithm basic_alltoallv() called.",rank);
+
+ err = smpi_mpi_type_get_extent(sdtype, &lb, &sndextent);
+ err = smpi_mpi_type_get_extent(rdtype, &lb, &rcvextent);
+
+ psnd = (char *)sbuf;
+ //print_buffer_int(psnd,size*size,xbt_strdup("sbuff"),rank);
+
+ /* copy the local sbuf to rbuf when it's me */
+ psnd = ((char *) sbuf) + (sdisps[rank] * sndextent);
+ prcv = ((char *) rbuf) + (rdisps[rank] * rcvextent);
+
+ if (0 != scounts[rank]) {
+ err = copy_dt( psnd, scounts[rank], sdtype, prcv, rcounts[rank], 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));
+
+
+ /* Create all receives that will be posted first */
+ for (i = 0; i < size; ++i) {
+ if (i == rank || 0 == rcounts[i]) {
+ DEBUG3("<%d> skip req creation i=%d,rcounts[i]=%d",rank,i, rcounts[i]);
+ continue;
+ }
+ prcv = ((char *) rbuf) + (rdisps[i] * rcvextent);
+
+ err = smpi_create_request( prcv, rcounts[i], rdtype,
+ i, rank,
+ system_alltoallv_tag,
+ comm, &(reqs[rreq]));
+ if (MPI_SUCCESS != err) {
+ DEBUG2("<%d> failed to create request for rank %d",rank,i);
+ for (i=0;i< rreq;i++)
+ xbt_mallocator_release(smpi_global->request_mallocator, reqs[i]);
+ return err;
+ }
+ rreq++;
+ }
+ DEBUG2("<%d> %d irecv reqs created",rank,rreq);
+ /* Now create all sends */
+ for (i = 0; i < size; ++i) {
+ if (i == rank || 0 == scounts[i]) {
+ DEBUG3("<%d> skip req creation i=%d,scounts[i]=%d",rank,i, scounts[i]);
+ continue;
+ }
+ psnd = ((char *) sbuf) + (sdisps[i] * sndextent);
+
+ //fprintf(stderr,"<%d> send %d elems to <%d>\n",rank,scounts[i],i);
+ //print_buffer_int(psnd,scounts[i],xbt_strdup("sbuff part"),rank);
+ err = smpi_create_request (psnd, scounts[i], sdtype,
+ rank, i,
+ system_alltoallv_tag,
+ comm, &(reqs[rreq+sreq]));
+ if (MPI_SUCCESS != err) {
+ DEBUG2("<%d> failed to create request for rank %d\n",rank,i);
+ for (i=0;i< rreq+sreq;i++)
+ xbt_mallocator_release(smpi_global->request_mallocator, reqs[i]);
+ return err;
+ }
+ sreq++;
+ }
+ DEBUG2("<%d> %d isend reqs created",rank,sreq);
+
+ /* Start your engines. This will never return an error. */
+ for ( i=0; i< rreq; i++ ) {
+ DEBUG3("<%d> issued irecv request reqs[%d]=%p",rank,i,reqs[i]);
+ smpi_mpi_irecv( reqs[i] );
+ }
+ DEBUG3("<%d> for (i=%d;i<%d)",rank,rreq,sreq);
+ for ( i=rreq; i< rreq+sreq; 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,rreq+sreq);
+ // waitall is buggy: use a loop instead for the moment
+ // err = smpi_mpi_waitall(nreq, reqs, MPI_STATUS_IGNORE);
+ for (i=0;i< rreq+sreq;i++) {
+ err = smpi_mpi_wait( reqs[i], MPI_STATUS_IGNORE);
+ }
+
+ /* Free the reqs */
+ /* nreq might be < 2*(size-1) since some request creations are skipped */
+ for (i=0;i< rreq+sreq;i++) {
+ xbt_mallocator_release(smpi_global->request_mallocator, reqs[i]);
+ }
+ xbt_free( reqs );
+ return err;
+}
+
+
+
/**
* -----------------------------------------------------------------------------------------------------