static PetscErrorCode PetscCommBuildTwoSidedFReq_Reference(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscMPIInt nto,const PetscMPIInt *toranks,const void *todata,
PetscMPIInt *nfrom,PetscMPIInt **fromranks,void *fromdata,PetscMPIInt ntags,MPI_Request **toreqs,MPI_Request **fromreqs,
PetscErrorCode (*send)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,PetscMPIInt,void*,MPI_Request[],void*),
PetscErrorCode (*recv)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,void*,MPI_Request[],void*),void *ctx)
{
PetscErrorCode ierr;
PetscMPIInt i,*tag;
MPI_Aint lb,unitbytes;
MPI_Request *sendreq,*recvreq;
PetscFunctionBegin;
ierr = PetscMalloc1(ntags,&tag);CHKERRQ(ierr);
if (ntags > 0) {
ierr = PetscCommDuplicate(comm,&comm,&tag[0]);CHKERRQ(ierr);
}
for (i=1; i<ntags; i++) {
ierr = PetscCommGetNewTag(comm,&tag[i]);CHKERRQ(ierr);
}
/* Perform complete initial rendezvous */
ierr = PetscCommBuildTwoSided(comm,count,dtype,nto,toranks,todata,nfrom,fromranks,fromdata);CHKERRQ(ierr);
ierr = PetscMalloc1(nto*ntags,&sendreq);CHKERRQ(ierr);
ierr = PetscMalloc1(*nfrom*ntags,&recvreq);CHKERRQ(ierr);
ierr = MPI_Type_get_extent(dtype,&lb,&unitbytes);CHKERRQ(ierr);
if (lb != 0) SETERRQ1(comm,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld\n",(long)lb);
for (i=0; i<nto; i++) {
PetscMPIInt k;
for (k=0; k<ntags; k++) sendreq[i*ntags+k] = MPI_REQUEST_NULL;
ierr = (*send)(comm,tag,i,toranks[i],((char*)todata)+count*unitbytes*i,sendreq+i*ntags,ctx);CHKERRQ(ierr);
}
for (i=0; i<*nfrom; i++) {
void *header = (*(char**)fromdata) + count*unitbytes*i;
PetscMPIInt k;
for (k=0; k<ntags; k++) recvreq[i*ntags+k] = MPI_REQUEST_NULL;
ierr = (*recv)(comm,tag,(*fromranks)[i],header,recvreq+i*ntags,ctx);CHKERRQ(ierr);
}
ierr = PetscFree(tag);CHKERRQ(ierr);
ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
*toreqs = sendreq;
*fromreqs = recvreq;
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
PetscErrorCode ierr;
PetscMPIInt rank,size,*toranks,*fromranks,nto,nfrom;
PetscInt i,n;
PetscBool verbose,build_twosided_f;
Unit *todata,*fromdata;
MPI_Datatype dtype;
PetscInitialize(&argc,&argv,(char*)0,help);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
verbose = PETSC_FALSE;
ierr = PetscOptionsGetBool(NULL,"-verbose",&verbose,NULL);CHKERRQ(ierr);
build_twosided_f = PETSC_FALSE;
ierr = PetscOptionsGetBool(NULL,"-build_twosided_f",&build_twosided_f,NULL);CHKERRQ(ierr);
for (i=1,nto=0; i<size; i*=2) nto++;
ierr = PetscMalloc2(nto,&todata,nto,&toranks);CHKERRQ(ierr);
for (n=0,i=1; i<size; n++,i*=2) {
toranks[n] = (rank+i) % size;
todata[n].rank = (rank+i) % size;
todata[n].value = (PetscScalar)rank;
todata[n].ok[0] = 'o';
todata[n].ok[1] = 'k';
todata[n].ok[2] = 0;
}
if (verbose) {
for (i=0; i<nto; i++) {
ierr = PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] TO %d: {%D, %g, \"%s\"}\n",rank,toranks[i],todata[i].rank,(double)PetscRealPart(todata[i].value),todata[i].ok);CHKERRQ(ierr);
}
ierr = PetscSynchronizedFlush(PETSC_COMM_WORLD,PETSC_STDOUT);CHKERRQ(ierr);
}
ierr = MakeDatatype(&dtype);CHKERRQ(ierr);
if (build_twosided_f) {
struct FCtx fctx;
PetscMPIInt *todummy,*fromdummy;
fctx.rank = rank;
fctx.nto = nto;
fctx.toranks = toranks;
fctx.todata = todata;
ierr = PetscSegBufferCreate(sizeof(Unit),1,&fctx.seg);CHKERRQ(ierr);
ierr = PetscMalloc1(nto,&todummy);CHKERRQ(ierr);
for (i=0; i<nto; i++) todummy[i] = rank;
ierr = PetscCommBuildTwoSidedF(PETSC_COMM_WORLD,1,MPI_INT,nto,toranks,todummy,&nfrom,&fromranks,&fromdummy,2,FSend,FRecv,&fctx);CHKERRQ(ierr);
ierr = PetscFree(todummy);CHKERRQ(ierr);
ierr = PetscFree(fromdummy);CHKERRQ(ierr);
ierr = PetscSegBufferExtractAlloc(fctx.seg,&fromdata);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&fctx.seg);CHKERRQ(ierr);
} else {
ierr = PetscCommBuildTwoSided(PETSC_COMM_WORLD,1,dtype,nto,toranks,todata,&nfrom,&fromranks,&fromdata);CHKERRQ(ierr);
}
ierr = MPI_Type_free(&dtype);CHKERRQ(ierr);
if (verbose) {
PetscInt *iranks,*iperm;
ierr = PetscMalloc2(nfrom,&iranks,nfrom,&iperm);CHKERRQ(ierr);
for (i=0; i<nfrom; i++) {
iranks[i] = fromranks[i];
iperm[i] = i;
}
/* Receive ordering is non-deterministic in general, so sort to make verbose output deterministic. */
ierr = PetscSortIntWithPermutation(nfrom,iranks,iperm);CHKERRQ(ierr);
for (i=0; i<nfrom; i++) {
PetscInt ip = iperm[i];
ierr = PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] FROM %d: {%D, %g, \"%s\"}\n",rank,fromranks[ip],fromdata[ip].rank,(double)PetscRealPart(fromdata[ip].value),fromdata[ip].ok);CHKERRQ(ierr);
}
ierr = PetscSynchronizedFlush(PETSC_COMM_WORLD,PETSC_STDOUT);CHKERRQ(ierr);
ierr = PetscFree2(iranks,iperm);CHKERRQ(ierr);
}
if (nto != nfrom) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"[%d] From ranks %d does not match To ranks %d",rank,nto,nfrom);
for (i=1; i<size; i*=2) {
PetscMPIInt expected_rank = (rank-i+size)%size;
PetscBool flg;
for (n=0; n<nfrom; n++) {
if (expected_rank == fromranks[n]) goto found;
}
SETERRQ2(PETSC_COMM_WORLD,PETSC_ERR_PLIB,"[%d] Could not find expected from rank %d",rank,expected_rank);
found:
if (PetscRealPart(fromdata[n].value) != expected_rank) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"[%d] Got data %g from rank %d",rank,(double)PetscRealPart(fromdata[n].value),expected_rank);
ierr = PetscStrcmp(fromdata[n].ok,"ok",&flg);CHKERRQ(ierr);
if (!flg) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"[%d] Got string %s from rank %d",rank,fromdata[n].ok,expected_rank);
}
ierr = PetscFree2(todata,toranks);CHKERRQ(ierr);
ierr = PetscFree(fromdata);CHKERRQ(ierr);
ierr = PetscFree(fromranks);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
/*@
ISBuildTwoSided - Takes an IS that describes where we will go. Generates an IS that contains new numbers from remote or local
on the IS.
Collective on IS
Input Parameters
. to - an IS describes where we will go. Negative target rank will be ignored
. toindx - an IS describes what indices should send. NULL means sending natural numbering
Output Parameter:
. rows - contains new numbers from remote or local
Level: advanced
.seealso: MatPartitioningCreate(), ISPartitioningToNumbering(), ISPartitioningCount()
@*/
PetscErrorCode ISBuildTwoSided(IS ito,IS toindx, IS *rows)
{
const PetscInt *ito_indices,*toindx_indices;
PetscInt *send_indices,rstart,*recv_indices,nrecvs,nsends;
PetscInt *tosizes,*fromsizes,i,j,*tosizes_tmp,*tooffsets_tmp,ito_ln;
PetscMPIInt *toranks,*fromranks,size,target_rank,*fromperm_newtoold,nto,nfrom;
PetscLayout isrmap;
MPI_Comm comm;
PetscSF sf;
PetscSFNode *iremote;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)ito,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = ISGetLocalSize(ito,&ito_ln);CHKERRQ(ierr);
/* why we do not have ISGetLayout? */
isrmap = ito->map;
ierr = PetscLayoutGetRange(isrmap,&rstart,NULL);CHKERRQ(ierr);
ierr = ISGetIndices(ito,&ito_indices);CHKERRQ(ierr);
ierr = PetscCalloc2(size,&tosizes_tmp,size+1,&tooffsets_tmp);CHKERRQ(ierr);
for(i=0; i<ito_ln; i++){
if(ito_indices[i]<0) continue;
#if defined(PETSC_USE_DEBUG)
if(ito_indices[i]>=size) SETERRQ2(comm,PETSC_ERR_ARG_OUTOFRANGE,"target rank %d is larger than communicator size %d ",ito_indices[i],size);
#endif
tosizes_tmp[ito_indices[i]]++;
}
nto = 0;
for(i=0; i<size; i++){
tooffsets_tmp[i+1] = tooffsets_tmp[i]+tosizes_tmp[i];
if(tosizes_tmp[i]>0) nto++;
}
ierr = PetscCalloc2(nto,&toranks,2*nto,&tosizes);CHKERRQ(ierr);
nto = 0;
for(i=0; i<size; i++){
if(tosizes_tmp[i]>0){
toranks[nto] = i;
tosizes[2*nto] = tosizes_tmp[i];/* size */
tosizes[2*nto+1] = tooffsets_tmp[i];/* offset */
nto++;
}
}
nsends = tooffsets_tmp[size];
ierr = PetscCalloc1(nsends,&send_indices);CHKERRQ(ierr);
if(toindx){
ierr = ISGetIndices(toindx,&toindx_indices);CHKERRQ(ierr);
}
for(i=0; i<ito_ln; i++){
if(ito_indices[i]<0) continue;
target_rank = ito_indices[i];
send_indices[tooffsets_tmp[target_rank]] = toindx? toindx_indices[i]:(i+rstart);
tooffsets_tmp[target_rank]++;
}
if(toindx){
ierr = ISRestoreIndices(toindx,&toindx_indices);CHKERRQ(ierr);
}
ierr = ISRestoreIndices(ito,&ito_indices);CHKERRQ(ierr);
ierr = PetscFree2(tosizes_tmp,tooffsets_tmp);CHKERRQ(ierr);
ierr = PetscCommBuildTwoSided(comm,2,MPIU_INT,nto,toranks,tosizes,&nfrom,&fromranks,&fromsizes);CHKERRQ(ierr);
ierr = PetscFree2(toranks,tosizes);CHKERRQ(ierr);
ierr = PetscCalloc1(nfrom,&fromperm_newtoold);CHKERRQ(ierr);
for(i=0; i<nfrom; i++){
fromperm_newtoold[i] = i;
}
ierr = PetscSortMPIIntWithArray(nfrom,fromranks,fromperm_newtoold);CHKERRQ(ierr);
nrecvs = 0;
for(i=0; i<nfrom; i++){
nrecvs += fromsizes[i*2];
}
ierr = PetscCalloc1(nrecvs,&recv_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(nrecvs,&iremote);CHKERRQ(ierr);
nrecvs = 0;
for(i=0; i<nfrom; i++){
for(j=0; j<fromsizes[2*fromperm_newtoold[i]]; j++){
iremote[nrecvs].rank = fromranks[i];
iremote[nrecvs++].index = fromsizes[2*fromperm_newtoold[i]+1]+j;
}
}
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nsends,nrecvs,NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
/* how to put a prefix ? */
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,send_indices,recv_indices);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,send_indices,recv_indices);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFree(fromranks);CHKERRQ(ierr);
ierr = PetscFree(fromsizes);CHKERRQ(ierr);
ierr = PetscFree(fromperm_newtoold);CHKERRQ(ierr);
ierr = PetscFree(send_indices);CHKERRQ(ierr);
if(rows){
ierr = PetscSortInt(nrecvs,recv_indices);CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, nrecvs,recv_indices,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
}else{
ierr = PetscFree(recv_indices);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}