PetscErrorCode MatPartitioningHierarchical_ReassembleFineparts(Mat adj, IS fineparts, ISLocalToGlobalMapping mapping, IS *sfineparts)
{
PetscInt *local_indices, *global_indices,*owners,*sfineparts_indices,localsize,i;
const PetscInt *ranges,*fineparts_indices;
PetscMPIInt rank;
MPI_Comm comm;
PetscLayout rmap;
PetscSFNode *remote;
PetscSF sf;
PetscErrorCode ierr;
PetscFunctionBegin;
/*get communicator */
ierr = PetscObjectGetComm((PetscObject)adj,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetLayouts(adj,&rmap,PETSC_NULL);CHKERRQ(ierr);
ierr = ISGetLocalSize(fineparts,&localsize);CHKERRQ(ierr);
ierr = PetscCalloc2(localsize,&global_indices,localsize,&local_indices);CHKERRQ(ierr);
for(i=0; i<localsize; i++){
local_indices[i] = i;
}
/*global indices */
ierr = ISLocalToGlobalMappingApply(mapping,localsize,local_indices,global_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(localsize,&owners);CHKERRQ(ierr);
/*find owners for global indices */
for(i=0; i<localsize; i++){
ierr = PetscLayoutFindOwner(rmap,global_indices[i],&owners[i]);CHKERRQ(ierr);
}
/*ranges */
ierr = PetscLayoutGetRanges(rmap,&ranges);CHKERRQ(ierr);
ierr = PetscCalloc1(ranges[rank+1]-ranges[rank],&sfineparts_indices);CHKERRQ(ierr);
ierr = ISGetIndices(fineparts,&fineparts_indices);CHKERRQ(ierr);
/*create a SF to exchange data */
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscCalloc1(localsize,&remote);CHKERRQ(ierr);
for(i=0; i<localsize; i++){
remote[i].rank = owners[i];
remote[i].index = global_indices[i]-ranges[owners[i]];
}
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
/*not sure how to add prefix to sf*/
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,localsize,localsize,PETSC_NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPIU_INT,fineparts_indices,sfineparts_indices,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf,MPIU_INT,fineparts_indices,sfineparts_indices,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = ISRestoreIndices(fineparts,&fineparts_indices);CHKERRQ(ierr);
/* comm self */
ierr = ISCreateGeneral(comm,ranges[rank+1]-ranges[rank],sfineparts_indices,PETSC_OWN_POINTER,sfineparts);CHKERRQ(ierr);
ierr = PetscFree2(global_indices,local_indices);CHKERRQ(ierr);
ierr = PetscFree(owners);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatISSetPreallocation_IS(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
{
Mat_IS *matis = (Mat_IS*)(B->data);
PetscSF sf;
PetscInt bs,i,nroots,*rootdata,nleaves,*leafdata,nlocalcols;
const PetscInt *gidxs;
PetscErrorCode ierr;
PetscFunctionBegin;
if (!matis->A) {
SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_SUP,"You should first call MatSetLocalToGlobalMapping");
}
ierr = MatGetLocalSize(B,&nroots,NULL);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,&nleaves,&nlocalcols);CHKERRQ(ierr);
ierr = MatGetBlockSize(matis->A,&bs);CHKERRQ(ierr);
ierr = PetscCalloc2(nroots,&rootdata,nleaves,&leafdata);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)B),&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,&gidxs);CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(sf,B->rmap,nleaves,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,&gidxs);CHKERRQ(ierr);
if (!d_nnz) {
for (i=0;i<nroots;i++) rootdata[i] += d_nz;
} else {
for (i=0;i<nroots;i++) rootdata[i] += d_nnz[i];
}
if (!o_nnz) {
for (i=0;i<nroots;i++) rootdata[i] += o_nz;
} else {
for (i=0;i<nroots;i++) rootdata[i] += o_nnz[i];
}
ierr = PetscSFBcastBegin(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves;i++) {
leafdata[i] = PetscMin(leafdata[i],nlocalcols);
}
ierr = MatSeqAIJSetPreallocation(matis->A,0,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves/bs;i++) {
leafdata[i] = leafdata[i*bs]/bs;
}
ierr = MatSeqBAIJSetPreallocation(matis->A,bs,0,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves/bs;i++) {
leafdata[i] = leafdata[i]-i;
}
ierr = MatSeqSBAIJSetPreallocation(matis->A,bs,0,leafdata);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFree2(rootdata,leafdata);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
Vec B,Bout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscScalar *bufB;
PetscScalar *bufBout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 2) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for two MPI processes\n");
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 2;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
if (rank == 0) {
iremote[0].rank = 0;
iremote[0].index = 0;
iremote[1].rank = 1;
iremote[1].index = 0;
} else {
iremote[0].rank = 1;
iremote[0].index = 0;
iremote[1].rank = 0;
iremote[1].index = 0;
}
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,2,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&B);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Bout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArray(B,&bufB);CHKERRQ(ierr);
for (i=0; i<2; i++) {
bufA[i] = (PetscScalar)rank;
bufB[i] = (PetscScalar)(rank) + 10.0;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(B,&bufB);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecGetArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecRestoreArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecView(Bout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&B);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = VecDestroy(&Bout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
/*
* Increase overlap for the sub-matrix across sub communicator
* sub-matrix could be a graph or numerical matrix
* */
PetscErrorCode MatIncreaseOverlapSplit_Single(Mat mat,IS *is,PetscInt ov)
{
PetscInt i,nindx,*indices_sc,*indices_ov,localsize,*localsizes_sc,localsize_tmp;
PetscInt *indices_ov_rd,nroots,nleaves,*localoffsets,*indices_recv,*sources_sc,*sources_sc_rd;
const PetscInt *indices;
PetscMPIInt srank,ssize,issamecomm,k,grank;
IS is_sc,allis_sc,partitioning;
MPI_Comm gcomm,dcomm,scomm;
PetscSF sf;
PetscSFNode *remote;
Mat *smat;
MatPartitioning part;
PetscErrorCode ierr;
PetscFunctionBegin;
/* get a sub communicator before call individual MatIncreaseOverlap
* since the sub communicator may be changed.
* */
ierr = PetscObjectGetComm((PetscObject)(*is),&dcomm);CHKERRQ(ierr);
/*make a copy before the original one is deleted*/
ierr = PetscCommDuplicate(dcomm,&scomm,NULL);CHKERRQ(ierr);
/*get a global communicator, where mat should be a global matrix */
ierr = PetscObjectGetComm((PetscObject)mat,&gcomm);CHKERRQ(ierr);
/*increase overlap on each individual subdomain*/
ierr = (*mat->ops->increaseoverlap)(mat,1,is,ov);CHKERRQ(ierr);
/*compare communicators */
ierr = MPI_Comm_compare(gcomm,scomm,&issamecomm);CHKERRQ(ierr);
/* if the sub-communicator is the same as the global communicator,
* user does not want to use a sub-communicator
* */
if(issamecomm == MPI_IDENT || issamecomm == MPI_CONGRUENT) PetscFunctionReturn(0);
/* if the sub-communicator is petsc_comm_self,
* user also does not care the sub-communicator
* */
ierr = MPI_Comm_compare(scomm,PETSC_COMM_SELF,&issamecomm);CHKERRQ(ierr);
if(issamecomm == MPI_IDENT || issamecomm == MPI_CONGRUENT){PetscFunctionReturn(0);}
/*local rank, size in a sub-communicator */
ierr = MPI_Comm_rank(scomm,&srank);CHKERRQ(ierr);
ierr = MPI_Comm_size(scomm,&ssize);CHKERRQ(ierr);
ierr = MPI_Comm_rank(gcomm,&grank);CHKERRQ(ierr);
/*create a new IS based on sub-communicator
* since the old IS is often based on petsc_comm_self
* */
ierr = ISGetLocalSize(*is,&nindx);CHKERRQ(ierr);
ierr = PetscCalloc1(nindx,&indices_sc);CHKERRQ(ierr);
ierr = ISGetIndices(*is,&indices);CHKERRQ(ierr);
ierr = PetscMemcpy(indices_sc,indices,sizeof(PetscInt)*nindx);CHKERRQ(ierr);
ierr = ISRestoreIndices(*is,&indices);CHKERRQ(ierr);
/*we do not need any more*/
ierr = ISDestroy(is);CHKERRQ(ierr);
/*create a index set based on the sub communicator */
ierr = ISCreateGeneral(scomm,nindx,indices_sc,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr);
/*gather all indices within the sub communicator*/
ierr = ISAllGather(is_sc,&allis_sc);CHKERRQ(ierr);
ierr = ISDestroy(&is_sc);CHKERRQ(ierr);
/* gather local sizes */
ierr = PetscMalloc1(ssize,&localsizes_sc);CHKERRQ(ierr);
/*get individual local sizes for all index sets*/
ierr = MPI_Gather(&nindx,1,MPIU_INT,localsizes_sc,1,MPIU_INT,0,scomm);CHKERRQ(ierr);
/*only root does these computations */
if(!srank){
/*get local size for the big index set*/
ierr = ISGetLocalSize(allis_sc,&localsize);CHKERRQ(ierr);
ierr = PetscCalloc2(localsize,&indices_ov,localsize,&sources_sc);CHKERRQ(ierr);
ierr = PetscCalloc2(localsize,&indices_ov_rd,localsize,&sources_sc_rd);CHKERRQ(ierr);
ierr = ISGetIndices(allis_sc,&indices);CHKERRQ(ierr);
ierr = PetscMemcpy(indices_ov,indices,sizeof(PetscInt)*localsize);CHKERRQ(ierr);
ierr = ISRestoreIndices(allis_sc,&indices);CHKERRQ(ierr);
/*we do not need it any more */
ierr = ISDestroy(&allis_sc);CHKERRQ(ierr);
/*assign corresponding sources */
localsize_tmp = 0;
for(k=0; k<ssize; k++){
for(i=0; i<localsizes_sc[k]; i++){
sources_sc[localsize_tmp++] = k;
}
}
/*record where indices come from */
ierr = PetscSortIntWithArray(localsize,indices_ov,sources_sc);CHKERRQ(ierr);
/*count local sizes for reduced indices */
ierr = PetscMemzero(localsizes_sc,sizeof(PetscInt)*ssize);CHKERRQ(ierr);
/*initialize the first entity*/
if(localsize){
indices_ov_rd[0] = indices_ov[0];
sources_sc_rd[0] = sources_sc[0];
localsizes_sc[sources_sc[0]]++;
}
localsize_tmp = 1;
/*remove duplicate integers */
for(i=1; i<localsize; i++){
if(indices_ov[i] != indices_ov[i-1]){
indices_ov_rd[localsize_tmp] = indices_ov[i];
sources_sc_rd[localsize_tmp++] = sources_sc[i];
localsizes_sc[sources_sc[i]]++;
}
}
ierr = PetscFree2(indices_ov,sources_sc);CHKERRQ(ierr);
ierr = PetscCalloc1(ssize+1,&localoffsets);CHKERRQ(ierr);
for(k=0; k<ssize; k++){
localoffsets[k+1] = localoffsets[k] + localsizes_sc[k];
}
/*construct a star forest to send data back */
nleaves = localoffsets[ssize];
ierr = PetscMemzero(localoffsets,(ssize+1)*sizeof(PetscInt));CHKERRQ(ierr);
nroots = localsizes_sc[srank];
ierr = PetscCalloc1(nleaves,&remote);CHKERRQ(ierr);
for(i=0; i<nleaves; i++){
remote[i].rank = sources_sc_rd[i];
remote[i].index = localoffsets[sources_sc_rd[i]]++;
}
ierr = PetscFree(localoffsets);CHKERRQ(ierr);
}else{
ierr = ISDestroy(&allis_sc);CHKERRQ(ierr);
/*Allocate a 'zero' pointer */
ierr = PetscCalloc1(0,&remote);CHKERRQ(ierr);
nleaves = 0;
indices_ov_rd = 0;
sources_sc_rd = 0;
}
/*scatter sizes to everybody */
ierr = MPI_Scatter(localsizes_sc,1, MPIU_INT,&nroots,1, MPIU_INT,0,scomm);CHKERRQ(ierr);
/*free memory */
ierr = PetscFree(localsizes_sc);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots,&indices_recv);CHKERRQ(ierr);
/*ierr = MPI_Comm_dup(scomm,&dcomm);CHKERRQ(ierr);*/
/*set data back to every body */
ierr = PetscSFCreate(scomm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPIU_INT,indices_ov_rd,indices_recv,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf,MPIU_INT,indices_ov_rd,indices_recv,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* free memory */
ierr = PetscFree2(indices_ov_rd,sources_sc_rd);CHKERRQ(ierr);
/*create a index set*/
ierr = ISCreateGeneral(scomm,nroots,indices_recv,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr);
/*construct a parallel submatrix */
ierr = MatGetSubMatricesMPI(mat,1,&is_sc,&is_sc,MAT_INITIAL_MATRIX,&smat);CHKERRQ(ierr);
/* we do not need them any more */
ierr = ISDestroy(&allis_sc);CHKERRQ(ierr);
/*create a partitioner to repartition the sub-matrix*/
ierr = MatPartitioningCreate(scomm,&part);CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,smat[0]);CHKERRQ(ierr);
#if PETSC_HAVE_PARMETIS
/* if there exists a ParMETIS installation, we try to use ParMETIS
* because a repartition routine possibly work better
* */
ierr = MatPartitioningSetType(part,MATPARTITIONINGPARMETIS);CHKERRQ(ierr);
/*try to use reparition function, instead of partition function */
ierr = MatPartitioningParmetisSetRepartition(part);CHKERRQ(ierr);
#else
/*we at least provide a default partitioner to rebalance the computation */
ierr = MatPartitioningSetType(part,MATPARTITIONINGAVERAGE);CHKERRQ(ierr);
#endif
/*user can pick up any partitioner by using an option*/
ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr);
/* apply partition */
ierr = MatPartitioningApply(part,&partitioning);CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr);
ierr = MatDestroy(&(smat[0]));CHKERRQ(ierr);
ierr = PetscFree(smat);CHKERRQ(ierr);
/* get local rows including overlap */
ierr = ISBuildTwoSided(partitioning,is_sc,is);CHKERRQ(ierr);
/* destroy */
ierr = ISDestroy(&is_sc);CHKERRQ(ierr);
ierr = ISDestroy(&partitioning);CHKERRQ(ierr);
ierr = PetscCommDestroy(&scomm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
DMPlexDistribute - Distributes the mesh and any associated sections.
Not Collective
Input Parameter:
+ dm - The original DMPlex object
. partitioner - The partitioning package, or NULL for the default
- overlap - The overlap of partitions, 0 is the default
Output Parameter:
+ sf - The PetscSF used for point distribution
- parallelMesh - The distributed DMPlex object, or NULL
Note: If the mesh was not distributed, the return value is NULL.
The user can control the definition of adjacency for the mesh using DMPlexGetAdjacencyUseCone() and
DMPlexSetAdjacencyUseClosure(). They should choose the combination appropriate for the function
representation on the mesh.
Level: intermediate
.keywords: mesh, elements
.seealso: DMPlexCreate(), DMPlexDistributeByFace(), DMPlexSetAdjacencyUseCone(), DMPlexSetAdjacencyUseClosure()
@*/
PetscErrorCode DMPlexDistribute(DM dm, const char partitioner[], PetscInt overlap, PetscSF *sf, DM *dmParallel)
{
DM_Plex *mesh = (DM_Plex*) dm->data, *pmesh;
MPI_Comm comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS origCellPart, origPart, cellPart, part;
PetscSection origCellPartSection, origPartSection, cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF, pointSF, coneSF;
ISLocalToGlobalMapping renumbering;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets;
PetscInt *cones, *newCones, newConesSize;
PetscBool flg;
PetscMPIInt rank, numProcs, p;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
if (sf) PetscValidPointer(sf,4);
PetscValidPointer(dmParallel,5);
ierr = PetscLogEventBegin(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
*dmParallel = NULL;
if (numProcs == 1) PetscFunctionReturn(0);
ierr = DMPlexGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = PetscLogEventBegin(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
if (overlap > 1) SETERRQ(PetscObjectComm((PetscObject)dm), PETSC_ERR_SUP, "Overlap > 1 not yet implemented");
ierr = DMPlexCreatePartition(dm, partitioner, height, overlap > 0 ? PETSC_TRUE : PETSC_FALSE, &cellPartSection, &cellPart, &origCellPartSection, &origCellPart);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) numRemoteRanks = numProcs;
else numRemoteRanks = 0;
ierr = PetscMalloc1(numRemoteRanks, &remoteRanks);CHKERRQ(ierr);
for (p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-partition_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, NULL);CHKERRQ(ierr);
if (origCellPart) {
ierr = PetscPrintf(comm, "Original Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(origCellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(origCellPart, NULL);CHKERRQ(ierr);
}
ierr = PetscSFView(partSF, NULL);CHKERRQ(ierr);
}
/* Close the partition over the mesh */
ierr = DMPlexCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMPlexCreate(comm, dmParallel);CHKERRQ(ierr);
ierr = DMPlexSetDimension(*dmParallel, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *dmParallel, "Parallel Mesh");CHKERRQ(ierr);
pmesh = (DM_Plex*) (*dmParallel)->data;
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, &pointSF);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, NULL);CHKERRQ(ierr);
ierr = PetscSFView(pointSF, NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, NULL);CHKERRQ(ierr);
}
ierr = PetscLogEventEnd(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscLogEventBegin(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMPlexGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMPlexGetConeSection(*dmParallel, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMSetUp(*dmParallel);CHKERRQ(ierr);
{
PetscInt pStart, pEnd, p;
ierr = PetscSectionGetChart(newConeSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt coneSize;
ierr = PetscSectionGetDof(newConeSection, p, &coneSize);CHKERRQ(ierr);
pmesh->maxConeSize = PetscMax(pmesh->maxConeSize, coneSize);
}
}
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMPlexGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetCones(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyBlock(renumbering, IS_GTOLM_MASK, newConesSize, newCones, NULL, newCones);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-cones_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, NULL);CHKERRQ(ierr);
}
ierr = DMPlexGetConeOrientations(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetConeOrientations(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Create supports and stratify sieve */
{
PetscInt pStart, pEnd;
ierr = PetscSectionGetChart(pmesh->coneSection, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionSetChart(pmesh->supportSection, pStart, pEnd);CHKERRQ(ierr);
}
ierr = DMPlexSymmetrize(*dmParallel);CHKERRQ(ierr);
ierr = DMPlexStratify(*dmParallel);CHKERRQ(ierr);
/* Distribute Coordinates */
{
PetscSection originalCoordSection, newCoordSection;
Vec originalCoordinates, newCoordinates;
PetscInt bs;
const char *name;
ierr = DMGetCoordinateSection(dm, &originalCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinateSection(*dmParallel, &newCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinatesLocal(dm, &originalCoordinates);CHKERRQ(ierr);
ierr = VecCreate(comm, &newCoordinates);CHKERRQ(ierr);
ierr = PetscObjectGetName((PetscObject) originalCoordinates, &name);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) newCoordinates, name);CHKERRQ(ierr);
ierr = DMPlexDistributeField(dm, pointSF, originalCoordSection, originalCoordinates, newCoordSection, newCoordinates);CHKERRQ(ierr);
ierr = DMSetCoordinatesLocal(*dmParallel, newCoordinates);CHKERRQ(ierr);
ierr = VecGetBlockSize(originalCoordinates, &bs);CHKERRQ(ierr);
ierr = VecSetBlockSize(newCoordinates, bs);CHKERRQ(ierr);
ierr = VecDestroy(&newCoordinates);CHKERRQ(ierr);
}
/* Distribute labels */
ierr = PetscLogEventBegin(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
{
DMLabel next = mesh->labels, newNext = pmesh->labels;
PetscInt numLabels = 0, l;
/* Bcast number of labels */
while (next) {++numLabels; next = next->next;}
ierr = MPI_Bcast(&numLabels, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
next = mesh->labels;
for (l = 0; l < numLabels; ++l) {
DMLabel labelNew;
PetscBool isdepth;
/* Skip "depth" because it is recreated */
if (!rank) {ierr = PetscStrcmp(next->name, "depth", &isdepth);CHKERRQ(ierr);}
ierr = MPI_Bcast(&isdepth, 1, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
if (isdepth) {if (!rank) next = next->next; continue;}
ierr = DMLabelDistribute(next, partSection, part, renumbering, &labelNew);CHKERRQ(ierr);
/* Insert into list */
if (newNext) newNext->next = labelNew;
else pmesh->labels = labelNew;
newNext = labelNew;
if (!rank) next = next->next;
}
}
ierr = PetscLogEventEnd(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
/* Setup hybrid structure */
{
const PetscInt *gpoints;
PetscInt depth, n, d;
for (d = 0; d <= dim; ++d) {pmesh->hybridPointMax[d] = mesh->hybridPointMax[d];}
ierr = MPI_Bcast(pmesh->hybridPointMax, dim+1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetSize(renumbering, &n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(renumbering, &gpoints);CHKERRQ(ierr);
ierr = DMPlexGetDepth(dm, &depth);CHKERRQ(ierr);
for (d = 0; d <= dim; ++d) {
PetscInt pmax = pmesh->hybridPointMax[d], newmax = 0, pEnd, stratum[2], p;
if (pmax < 0) continue;
ierr = DMPlexGetDepthStratum(dm, d > depth ? depth : d, &stratum[0], &stratum[1]);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(*dmParallel, d, NULL, &pEnd);CHKERRQ(ierr);
ierr = MPI_Bcast(stratum, 2, MPIU_INT, 0, comm);CHKERRQ(ierr);
for (p = 0; p < n; ++p) {
const PetscInt point = gpoints[p];
if ((point >= stratum[0]) && (point < stratum[1]) && (point >= pmax)) ++newmax;
}
if (newmax > 0) pmesh->hybridPointMax[d] = pEnd - newmax;
else pmesh->hybridPointMax[d] = -1;
}
ierr = ISLocalToGlobalMappingRestoreIndices(renumbering, &gpoints);CHKERRQ(ierr);
}
/* Cleanup Partition */
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Create point SF for parallel mesh */
ierr = PetscLogEventBegin(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
{
const PetscInt *leaves;
PetscSFNode *remotePoints, *rowners, *lowners;
PetscInt numRoots, numLeaves, numGhostPoints = 0, p, gp, *ghostPoints;
PetscInt pStart, pEnd;
ierr = DMPlexGetChart(*dmParallel, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSFGetGraph(pointSF, &numRoots, &numLeaves, &leaves, NULL);CHKERRQ(ierr);
ierr = PetscMalloc2(numRoots,&rowners,numLeaves,&lowners);CHKERRQ(ierr);
for (p=0; p<numRoots; p++) {
rowners[p].rank = -1;
rowners[p].index = -1;
}
if (origCellPart) {
/* Make sure points in the original partition are not assigned to other procs */
const PetscInt *origPoints;
ierr = DMPlexCreatePartitionClosure(dm, origCellPartSection, origCellPart, &origPartSection, &origPart);CHKERRQ(ierr);
ierr = ISGetIndices(origPart, &origPoints);CHKERRQ(ierr);
for (p = 0; p < numProcs; ++p) {
PetscInt dof, off, d;
ierr = PetscSectionGetDof(origPartSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(origPartSection, p, &off);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
rowners[origPoints[d]].rank = p;
}
}
ierr = ISRestoreIndices(origPart, &origPoints);CHKERRQ(ierr);
ierr = ISDestroy(&origPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origPartSection);CHKERRQ(ierr);
}
ierr = ISDestroy(&origCellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origCellPartSection);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].rank == rank) { /* Either put in a bid or we know we own it */
lowners[p].rank = rank;
lowners[p].index = leaves ? leaves[p] : p;
} else if (lowners[p].rank >= 0) { /* Point already claimed so flag so that MAXLOC does not listen to us */
lowners[p].rank = -2;
lowners[p].index = -2;
}
}
for (p=0; p<numRoots; p++) { /* Root must not participate in the rediction, flag so that MAXLOC does not use */
rowners[p].rank = -3;
rowners[p].index = -3;
}
ierr = PetscSFReduceBegin(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].index < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cell partition corrupt: point not claimed");
if (lowners[p].rank != rank) ++numGhostPoints;
}
ierr = PetscMalloc1(numGhostPoints, &ghostPoints);CHKERRQ(ierr);
ierr = PetscMalloc1(numGhostPoints, &remotePoints);CHKERRQ(ierr);
for (p = 0, gp = 0; p < numLeaves; ++p) {
if (lowners[p].rank != rank) {
ghostPoints[gp] = leaves ? leaves[p] : p;
remotePoints[gp].rank = lowners[p].rank;
remotePoints[gp].index = lowners[p].index;
++gp;
}
}
ierr = PetscFree2(rowners,lowners);CHKERRQ(ierr);
ierr = PetscSFSetGraph((*dmParallel)->sf, pEnd - pStart, numGhostPoints, ghostPoints, PETSC_OWN_POINTER, remotePoints, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions((*dmParallel)->sf);CHKERRQ(ierr);
}
pmesh->useCone = mesh->useCone;
pmesh->useClosure = mesh->useClosure;
ierr = PetscLogEventEnd(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
/* Copy BC */
ierr = DMPlexCopyBoundary(dm, *dmParallel);CHKERRQ(ierr);
/* Cleanup */
if (sf) {*sf = pointSF;}
else {ierr = PetscSFDestroy(&pointSF);CHKERRQ(ierr);}
ierr = DMSetFromOptions(*dmParallel);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
PetscBool test_dupped_type;
MPI_Datatype contig;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for one MPI process");
ierr = PetscOptionsBegin(PETSC_COMM_WORLD,"","PetscSF type freeing options","none");CHKERRQ(ierr);
test_dupped_type = PETSC_FALSE;
ierr = PetscOptionsBool("-test_dupped_type", "Test dupped input type","",test_dupped_type,&test_dupped_type,NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 1;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
iremote[0].rank = 0;
iremote[0].index = 0;
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,4,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
for (i=0; i<4; i++) {
bufA[i] = (PetscScalar)i;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = MPI_Type_contiguous(4, MPIU_SCALAR, &contig);CHKERRQ(ierr);
ierr = MPI_Type_commit(&contig);CHKERRQ(ierr);
if (test_dupped_type) {
MPI_Datatype tmp;
ierr = MPI_Type_dup(contig, &tmp);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
contig = tmp;
}
for (i=0;i<10000;i++) {
ierr = PetscSFBcastBegin(sf,contig,bufA,bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,contig,bufA,bufAout);CHKERRQ(ierr);
}
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
/*@
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);
}
/*
* The interface should be easy to use for both MatGetSubMatrix (parallel sub-matrix) and MatGetSubMatrices (sequential sub-matrices)
* */
static PetscErrorCode MatGetSubMatrix_MPIAdj_data(Mat adj,IS irows, IS icols, PetscInt **sadj_xadj,PetscInt **sadj_adjncy,PetscInt **sadj_values)
{
PetscInt nlrows_is,icols_n,i,j,nroots,nleaves,owner,rlocalindex,*ncols_send,*ncols_recv;
PetscInt nlrows_mat,*adjncy_recv,Ncols_recv,Ncols_send,*xadj_recv,*values_recv;
PetscInt *ncols_recv_offsets,loc,rnclos,*sadjncy,*sxadj,*svalues,isvalue;
const PetscInt *irows_indices,*icols_indices,*xadj, *adjncy;
Mat_MPIAdj *a = (Mat_MPIAdj*)adj->data;
PetscLayout rmap;
MPI_Comm comm;
PetscSF sf;
PetscSFNode *iremote;
PetscBool done;
PetscErrorCode ierr;
PetscFunctionBegin;
/* communicator */
ierr = PetscObjectGetComm((PetscObject)adj,&comm);CHKERRQ(ierr);
/* Layouts */
ierr = MatGetLayouts(adj,&rmap,PETSC_NULL);CHKERRQ(ierr);
/* get rows information */
ierr = ISGetLocalSize(irows,&nlrows_is);CHKERRQ(ierr);
ierr = ISGetIndices(irows,&irows_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(nlrows_is,&iremote);CHKERRQ(ierr);
/* construct sf graph*/
nleaves = nlrows_is;
for(i=0; i<nlrows_is; i++){
owner = -1;
rlocalindex = -1;
ierr = PetscLayoutFindOwnerIndex(rmap,irows_indices[i],&owner,&rlocalindex);CHKERRQ(ierr);
iremote[i].rank = owner;
iremote[i].index = rlocalindex;
}
ierr = MatGetRowIJ(adj,0,PETSC_FALSE,PETSC_FALSE,&nlrows_mat,&xadj,&adjncy,&done);CHKERRQ(ierr);
ierr = PetscCalloc4(nlrows_mat,&ncols_send,nlrows_is,&xadj_recv,nlrows_is+1,&ncols_recv_offsets,nlrows_is,&ncols_recv);CHKERRQ(ierr);
nroots = nlrows_mat;
for(i=0; i<nlrows_mat; i++){
ncols_send[i] = xadj[i+1]-xadj[i];
}
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,ncols_send,ncols_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,ncols_send,ncols_recv);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,xadj,xadj_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,xadj,xadj_recv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
Ncols_recv =0;
for(i=0; i<nlrows_is; i++){
Ncols_recv += ncols_recv[i];
ncols_recv_offsets[i+1] = ncols_recv[i]+ncols_recv_offsets[i];
}
Ncols_send = 0;
for(i=0; i<nlrows_mat; i++){
Ncols_send += ncols_send[i];
}
ierr = PetscCalloc1(Ncols_recv,&iremote);CHKERRQ(ierr);
ierr = PetscCalloc1(Ncols_recv,&adjncy_recv);CHKERRQ(ierr);
nleaves = Ncols_recv;
Ncols_recv = 0;
for(i=0; i<nlrows_is; i++){
ierr = PetscLayoutFindOwner(rmap,irows_indices[i],&owner);CHKERRQ(ierr);
for(j=0; j<ncols_recv[i]; j++){
iremote[Ncols_recv].rank = owner;
iremote[Ncols_recv++].index = xadj_recv[i]+j;
}
}
ierr = ISRestoreIndices(irows,&irows_indices);CHKERRQ(ierr);
/*if we need to deal with edge weights ???*/
if(a->values){isvalue=1;}else{isvalue=0;}
/*involve a global communication */
/*ierr = MPI_Allreduce(&isvalue,&isvalue,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr);*/
if(isvalue){ierr = PetscCalloc1(Ncols_recv,&values_recv);CHKERRQ(ierr);}
nroots = Ncols_send;
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,adjncy,adjncy_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,adjncy,adjncy_recv);CHKERRQ(ierr);
if(isvalue){
ierr = PetscSFBcastBegin(sf,MPIU_INT,a->values,values_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,a->values,values_recv);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = MatRestoreRowIJ(adj,0,PETSC_FALSE,PETSC_FALSE,&nlrows_mat,&xadj,&adjncy,&done);CHKERRQ(ierr);
ierr = ISGetLocalSize(icols,&icols_n);CHKERRQ(ierr);
ierr = ISGetIndices(icols,&icols_indices);CHKERRQ(ierr);
rnclos = 0;
for(i=0; i<nlrows_is; i++){
for(j=ncols_recv_offsets[i]; j<ncols_recv_offsets[i+1]; j++){
ierr = PetscFindInt(adjncy_recv[j], icols_n, icols_indices, &loc);CHKERRQ(ierr);
if(loc<0){
adjncy_recv[j] = -1;
if(isvalue) values_recv[j] = -1;
ncols_recv[i]--;
}else{
rnclos++;
}
}
}
ierr = ISRestoreIndices(icols,&icols_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(rnclos,&sadjncy);CHKERRQ(ierr);
if(isvalue) {ierr = PetscCalloc1(rnclos,&svalues);CHKERRQ(ierr);}
ierr = PetscCalloc1(nlrows_is+1,&sxadj);CHKERRQ(ierr);
rnclos = 0;
for(i=0; i<nlrows_is; i++){
for(j=ncols_recv_offsets[i]; j<ncols_recv_offsets[i+1]; j++){
if(adjncy_recv[j]<0) continue;
sadjncy[rnclos] = adjncy_recv[j];
if(isvalue) svalues[rnclos] = values_recv[j];
rnclos++;
}
}
for(i=0; i<nlrows_is; i++){
sxadj[i+1] = sxadj[i]+ncols_recv[i];
}
if(sadj_xadj) { *sadj_xadj = sxadj;}else { ierr = PetscFree(sxadj);CHKERRQ(ierr);}
if(sadj_adjncy){ *sadj_adjncy = sadjncy;}else{ ierr = PetscFree(sadjncy);CHKERRQ(ierr);}
if(sadj_values){
if(isvalue) *sadj_values = svalues; else *sadj_values=0;
}else{
if(isvalue) {ierr = PetscFree(svalues);CHKERRQ(ierr);}
}
ierr = PetscFree4(ncols_send,xadj_recv,ncols_recv_offsets,ncols_recv);CHKERRQ(ierr);
ierr = PetscFree(adjncy_recv);CHKERRQ(ierr);
if(isvalue) {ierr = PetscFree(values_recv);CHKERRQ(ierr);}
PetscFunctionReturn(0);
}
PETSC_EXTERN PetscErrorCode MatColoringCreateBipartiteGraph(MatColoring mc,PetscSF *etoc,PetscSF *etor)
{
PetscErrorCode ierr;
PetscInt nentries,ncolentries,idx;
PetscInt i,j,rs,re,cs,ce,cn;
PetscInt *rowleaf,*colleaf,*rowdata;
PetscInt ncol;
const PetscScalar *vcol;
const PetscInt *icol;
const PetscInt *coldegrees,*rowdegrees;
Mat m = mc->mat;
PetscFunctionBegin;
ierr = MatGetOwnershipRange(m,&rs,&re);CHKERRQ(ierr);
ierr = MatGetOwnershipRangeColumn(m,&cs,&ce);CHKERRQ(ierr);
cn = ce-cs;
nentries=0;
for (i=rs;i<re;i++) {
ierr = MatGetRow(m,i,&ncol,NULL,&vcol);CHKERRQ(ierr);
for (j=0;j<ncol;j++) {
nentries++;
}
ierr = MatRestoreRow(m,i,&ncol,NULL,&vcol);CHKERRQ(ierr);
}
ierr = PetscMalloc(sizeof(PetscInt)*nentries,&rowleaf);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*nentries,&rowdata);CHKERRQ(ierr);
idx=0;
for (i=rs;i<re;i++) {
ierr = MatGetRow(m,i,&ncol,&icol,&vcol);CHKERRQ(ierr);
for (j=0;j<ncol;j++) {
rowleaf[idx] = icol[j];
rowdata[idx] = i;
idx++;
}
ierr = MatRestoreRow(m,i,&ncol,&icol,&vcol);CHKERRQ(ierr);
}
if (idx != nentries) SETERRQ2(PetscObjectComm((PetscObject)m),PETSC_ERR_NOT_CONVERGED,"Bad number of entries %d vs %d",idx,nentries);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)m),etoc);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)m),etor);CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(*etoc,m->cmap,nentries,NULL,PETSC_COPY_VALUES,rowleaf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(*etoc);CHKERRQ(ierr);
/* determine the number of entries in the column matrix */
ierr = PetscLogEventBegin(Mat_Coloring_Comm,*etoc,0,0,0);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(*etoc,&coldegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(*etoc,&coldegrees);CHKERRQ(ierr);
ierr = PetscLogEventEnd(Mat_Coloring_Comm,*etoc,0,0,0);CHKERRQ(ierr);
ncolentries=0;
for (i=0;i<cn;i++) {
ncolentries += coldegrees[i];
}
ierr = PetscMalloc(sizeof(PetscInt)*ncolentries,&colleaf);CHKERRQ(ierr);
/* create the one going the other way by building the leaf set */
ierr = PetscLogEventBegin(Mat_Coloring_Comm,*etoc,0,0,0);CHKERRQ(ierr);
ierr = PetscSFGatherBegin(*etoc,MPIU_INT,rowdata,colleaf);CHKERRQ(ierr);
ierr = PetscSFGatherEnd(*etoc,MPIU_INT,rowdata,colleaf);CHKERRQ(ierr);
ierr = PetscLogEventEnd(Mat_Coloring_Comm,*etoc,0,0,0);CHKERRQ(ierr);
/* this one takes mat entries in *columns* to rows -- you never have to actually be able to order the leaf entries. */
ierr = PetscSFSetGraphLayout(*etor,m->rmap,ncolentries,NULL,PETSC_COPY_VALUES,colleaf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(*etor);CHKERRQ(ierr);
ierr = PetscLogEventBegin(Mat_Coloring_Comm,*etor,0,0,0);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(*etor,&rowdegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(*etor,&rowdegrees);CHKERRQ(ierr);
ierr = PetscLogEventEnd(Mat_Coloring_Comm,*etor,0,0,0);CHKERRQ(ierr);
ierr = PetscFree(rowdata);CHKERRQ(ierr);
ierr = PetscFree(rowleaf);CHKERRQ(ierr);
ierr = PetscFree(colleaf);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
