PetscErrorCode DMPatchSolve(DM dm)
{
MPI_Comm comm;
MPI_Comm commz;
DM dmc;
PetscSF sfz, sfzr;
Vec XC;
MatStencil patchSize, commSize, gridRank, lower, upper;
PetscInt M, N, P, i, j, k, l, m, n, p = 0;
PetscMPIInt rank, size;
PetscInt debug = 0;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = DMPatchGetCoarse(dm, &dmc);CHKERRQ(ierr);
ierr = DMPatchGetPatchSize(dm, &patchSize);CHKERRQ(ierr);
ierr = DMPatchGetCommSize(dm, &commSize);CHKERRQ(ierr);
ierr = DMPatchGetCommSize(dm, &commSize);CHKERRQ(ierr);
ierr = DMGetGlobalVector(dmc, &XC);CHKERRQ(ierr);
ierr = DMDAGetInfo(dmc, 0, &M, &N, &P, &l, &m, &n, 0,0,0,0,0,0);CHKERRQ(ierr);
M = PetscMax(M, 1); l = PetscMax(l, 1);
N = PetscMax(N, 1); m = PetscMax(m, 1);
P = PetscMax(P, 1); n = PetscMax(n, 1);
gridRank.i = rank % l;
gridRank.j = rank/l % m;
gridRank.k = rank/(l*m) % n;
if (commSize.i*commSize.j*commSize.k == size || commSize.i*commSize.j*commSize.k == 0) {
commSize.i = l; commSize.j = m; commSize.k = n;
commz = comm;
} else if (commSize.i*commSize.j*commSize.k == 1) {
commz = PETSC_COMM_SELF;
} else {
const PetscMPIInt newComm = ((gridRank.k/commSize.k)*(m/commSize.j) + gridRank.j/commSize.j)*(l/commSize.i) + (gridRank.i/commSize.i);
const PetscMPIInt newRank = ((gridRank.k%commSize.k)*commSize.j + gridRank.j%commSize.j)*commSize.i + (gridRank.i%commSize.i);
ierr = MPI_Comm_split(comm, newComm, newRank, &commz);CHKERRQ(ierr);
if (debug) {ierr = PetscPrintf(PETSC_COMM_SELF, "Rank %d color %d key %d commz %d\n", rank, newComm, newRank, *((PetscMPIInt*) &commz));CHKERRQ(ierr);}
}
/*
Assumptions:
- patchSize divides gridSize
- commSize divides gridSize
- commSize divides l,m,n
Ignore multiple patches per rank for now
Multiple ranks per patch:
- l,m,n divides patchSize
- commSize divides patchSize
*/
for (k = 0; k < P; k += PetscMax(patchSize.k, 1)) {
for (j = 0; j < N; j += PetscMax(patchSize.j, 1)) {
for (i = 0; i < M; i += PetscMax(patchSize.i, 1), ++p) {
MPI_Comm commp = MPI_COMM_NULL;
DM dmz = NULL;
#if 0
DM dmf = NULL;
Mat interpz = NULL;
#endif
Vec XZ = NULL;
PetscScalar *xcarray = NULL;
PetscScalar *xzarray = NULL;
if ((gridRank.k/commSize.k == p/(l/commSize.i * m/commSize.j) % n/commSize.k) &&
(gridRank.j/commSize.j == p/(l/commSize.i) % m/commSize.j) &&
(gridRank.i/commSize.i == p % l/commSize.i)) {
if (debug) {ierr = PetscPrintf(PETSC_COMM_SELF, "Rank %d is accepting Patch %d\n", rank, p);CHKERRQ(ierr);}
commp = commz;
}
/* Zoom to coarse patch */
lower.i = i; lower.j = j; lower.k = k;
upper.i = i + patchSize.i; upper.j = j + patchSize.j; upper.k = k + patchSize.k;
ierr = DMPatchZoom(dmc, XC, lower, upper, commp, &dmz, &sfz, &sfzr);CHKERRQ(ierr);
lower.c = 0; /* initialize member, otherwise compiler issues warnings */
upper.c = 0; /* initialize member, otherwise compiler issues warnings */
/* Debug */
ierr = PetscPrintf(comm, "Patch %d: (%d, %d, %d)--(%d, %d, %d)\n", p, lower.i, lower.j, lower.k, upper.i, upper.j, upper.k);CHKERRQ(ierr);
if (dmz) {ierr = DMView(dmz, PETSC_VIEWER_STDOUT_(commz));CHKERRQ(ierr);}
ierr = PetscSFView(sfz, PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr);
ierr = PetscSFView(sfzr, PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr);
/* Scatter Xcoarse -> Xzoom */
if (dmz) {ierr = DMGetGlobalVector(dmz, &XZ);CHKERRQ(ierr);}
if (XZ) {ierr = VecGetArray(XZ, &xzarray);CHKERRQ(ierr);}
ierr = VecGetArray(XC, &xcarray);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sfz, MPIU_SCALAR, xcarray, xzarray);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sfz, MPIU_SCALAR, xcarray, xzarray);CHKERRQ(ierr);
ierr = VecRestoreArray(XC, &xcarray);CHKERRQ(ierr);
if (XZ) {ierr = VecRestoreArray(XZ, &xzarray);CHKERRQ(ierr);}
#if 0
/* Interpolate Xzoom -> Xfine, note that this may be on subcomms */
ierr = DMRefine(dmz, MPI_COMM_NULL, &dmf);CHKERRQ(ierr);
ierr = DMCreateInterpolation(dmz, dmf, &interpz, NULL);CHKERRQ(ierr);
ierr = DMInterpolate(dmz, interpz, dmf);CHKERRQ(ierr);
/* Smooth Xfine using two-step smoother, normal smoother plus Kaczmarz---moves back and forth from dmzoom to dmfine */
/* Compute residual Rfine */
/* Restrict Rfine to Rzoom_restricted */
#endif
/* Scatter Rzoom_restricted -> Rcoarse_restricted */
if (XZ) {ierr = VecGetArray(XZ, &xzarray);CHKERRQ(ierr);}
ierr = VecGetArray(XC, &xcarray);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sfzr, MPIU_SCALAR, xzarray, xcarray, MPIU_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sfzr, MPIU_SCALAR, xzarray, xcarray, MPIU_SUM);CHKERRQ(ierr);
ierr = VecRestoreArray(XC, &xcarray);CHKERRQ(ierr);
if (XZ) {ierr = VecRestoreArray(XZ, &xzarray);CHKERRQ(ierr);}
if (dmz) {ierr = DMRestoreGlobalVector(dmz, &XZ);CHKERRQ(ierr);}
/* Compute global residual Rcoarse */
/* TauCoarse = Rcoarse - Rcoarse_restricted */
ierr = PetscSFDestroy(&sfz);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfzr);CHKERRQ(ierr);
ierr = DMDestroy(&dmz);CHKERRQ(ierr);
}
}
}
ierr = DMRestoreGlobalVector(dmc, &XC);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
PetscSFGetMultiSF - gets the inner SF implemeting gathers and scatters
Collective
Input Argument:
. sf - star forest that may contain roots with 0 or with more than 1 vertex
Output Arguments:
. multi - star forest with split roots, such that each root has degree exactly 1
Level: developer
Notes:
In most cases, users should use PetscSFGatherBegin() and PetscSFScatterBegin() instead of manipulating multi
directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming
edge, it is a candidate for future optimization that might involve its removal.
.seealso: PetscSFSetGraph(), PetscSFGatherBegin(), PetscSFScatterBegin()
@*/
PetscErrorCode PetscSFGetMultiSF(PetscSF sf,PetscSF *multi)
{
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
PetscValidPointer(multi,2);
if (sf->nroots < 0) { /* Graph has not been set yet; why do we need this? */
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi);CHKERRQ(ierr);
*multi = sf->multi;
PetscFunctionReturn(0);
}
if (!sf->multi) {
const PetscInt *indegree;
PetscInt i,*inoffset,*outones,*outoffset,maxlocal;
PetscSFNode *remote;
ierr = PetscSFComputeDegreeBegin(sf,&indegree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sf,&indegree);CHKERRQ(ierr);
for (i=0,maxlocal=0; i<sf->nleaves; i++) maxlocal = PetscMax(maxlocal,(sf->mine ? sf->mine[i] : i)+1);
ierr = PetscMalloc3(sf->nroots+1,&inoffset,maxlocal,&outones,maxlocal,&outoffset);CHKERRQ(ierr);
inoffset[0] = 0;
for (i=0; i<sf->nroots; i++) inoffset[i+1] = inoffset[i] + indegree[i];
for (i=0; i<maxlocal; i++) outones[i] = 1;
ierr = PetscSFFetchAndOpBegin(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpEnd(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM);CHKERRQ(ierr);
for (i=0; i<sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */
#if 0
#if defined(PETSC_USE_DEBUG) /* Check that the expected number of increments occurred */
for (i=0; i<sf->nroots; i++) {
if (inoffset[i] + indegree[i] != inoffset[i+1]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect result after PetscSFFetchAndOp");
}
#endif
#endif
ierr = PetscMalloc1(sf->nleaves,&remote);CHKERRQ(ierr);
for (i=0; i<sf->nleaves; i++) {
remote[i].rank = sf->remote[i].rank;
remote[i].index = outoffset[sf->mine ? sf->mine[i] : i];
}
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
if (sf->rankorder) { /* Sort the ranks */
PetscMPIInt rank;
PetscInt *inranks,*newoffset,*outranks,*newoutoffset,*tmpoffset,maxdegree;
PetscSFNode *newremote;
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
for (i=0,maxdegree=0; i<sf->nroots; i++) maxdegree = PetscMax(maxdegree,indegree[i]);
ierr = PetscMalloc5(sf->multi->nroots,&inranks,sf->multi->nroots,&newoffset,maxlocal,&outranks,maxlocal,&newoutoffset,maxdegree,&tmpoffset);CHKERRQ(ierr);
for (i=0; i<maxlocal; i++) outranks[i] = rank;
ierr = PetscSFReduceBegin(sf->multi,MPIU_INT,outranks,inranks,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf->multi,MPIU_INT,outranks,inranks,MPIU_REPLACE);CHKERRQ(ierr);
/* Sort the incoming ranks at each vertex, build the inverse map */
for (i=0; i<sf->nroots; i++) {
PetscInt j;
for (j=0; j<indegree[i]; j++) tmpoffset[j] = j;
ierr = PetscSortIntWithArray(indegree[i],inranks+inoffset[i],tmpoffset);CHKERRQ(ierr);
for (j=0; j<indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j;
}
ierr = PetscSFBcastBegin(sf->multi,MPIU_INT,newoffset,newoutoffset);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf->multi,MPIU_INT,newoffset,newoutoffset);CHKERRQ(ierr);
ierr = PetscMalloc1(sf->nleaves,&newremote);CHKERRQ(ierr);
for (i=0; i<sf->nleaves; i++) {
newremote[i].rank = sf->remote[i].rank;
newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i];
}
ierr = PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,newremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscFree5(inranks,newoffset,outranks,newoutoffset,tmpoffset);CHKERRQ(ierr);
}
ierr = PetscFree3(inoffset,outones,outoffset);CHKERRQ(ierr);
}
*multi = sf->multi;
PetscFunctionReturn(0);
}
PetscErrorCode DMPlexPreallocateOperator(DM dm, PetscInt bs, PetscSection section, PetscSection sectionGlobal, PetscInt dnz[], PetscInt onz[], PetscInt dnzu[], PetscInt onzu[], Mat A, PetscBool fillMatrix)
{
MPI_Comm comm;
MatType mtype;
PetscSF sf, sfDof, sfAdj;
PetscSection leafSectionAdj, rootSectionAdj, sectionAdj, anchorSectionAdj;
PetscInt nroots, nleaves, l, p;
const PetscInt *leaves;
const PetscSFNode *remotes;
PetscInt dim, pStart, pEnd, numDof, globalOffStart, globalOffEnd, numCols;
PetscInt *tmpAdj = NULL, *adj, *rootAdj, *anchorAdj = NULL, *cols, *remoteOffsets;
PetscInt adjSize;
PetscLayout rLayout;
PetscInt locRows, rStart, rEnd, r;
PetscMPIInt size;
PetscBool doCommLocal, doComm, debug = PETSC_FALSE, isSymBlock, isSymSeqBlock, isSymMPIBlock;
PetscBool useAnchors;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
PetscValidHeaderSpecific(section, PETSC_SECTION_CLASSID, 3);
PetscValidHeaderSpecific(sectionGlobal, PETSC_SECTION_CLASSID, 4);
PetscValidHeaderSpecific(A, MAT_CLASSID, 9);
if (dnz) PetscValidPointer(dnz,5);
if (onz) PetscValidPointer(onz,6);
if (dnzu) PetscValidPointer(dnzu,7);
if (onzu) PetscValidPointer(onzu,8);
ierr = PetscLogEventBegin(DMPLEX_Preallocate,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(NULL, "-dm_view_preallocation", &debug, NULL);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMGetPointSF(dm, &sf);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL);CHKERRQ(ierr);
doCommLocal = (size > 1) && (nroots >= 0) ? PETSC_TRUE : PETSC_FALSE;
ierr = MPI_Allreduce(&doCommLocal, &doComm, 1, MPIU_BOOL, MPI_LAND, comm);CHKERRQ(ierr);
/* Create dof SF based on point SF */
if (debug) {
ierr = PetscPrintf(comm, "Input Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(section, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Input Global Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(sectionGlobal, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Input SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sf, NULL);CHKERRQ(ierr);
}
ierr = PetscSFCreateRemoteOffsets(sf, section, section, &remoteOffsets);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, section, remoteOffsets, section, &sfDof);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Dof SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sfDof, NULL);CHKERRQ(ierr);
}
/* Create section for dof adjacency (dof ==> # adj dof) */
ierr = PetscSectionGetChart(section, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(section, &numDof);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(leafSectionAdj, 0, numDof);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSectionAdj, 0, numDof);CHKERRQ(ierr);
/* Fill in the ghost dofs on the interface */
ierr = PetscSFGetGraph(sf, NULL, &nleaves, &leaves, &remotes);CHKERRQ(ierr);
/* use constraints in finding adjacency in this routine */
ierr = DMPlexGetAdjacencyUseAnchors(dm,&useAnchors);CHKERRQ(ierr);
ierr = DMPlexSetAdjacencyUseAnchors(dm,PETSC_TRUE);CHKERRQ(ierr);
/*
section - maps points to (# dofs, local dofs)
sectionGlobal - maps points to (# dofs, global dofs)
leafSectionAdj - maps unowned local dofs to # adj dofs
rootSectionAdj - maps owned local dofs to # adj dofs
adj - adj global dofs indexed by leafSectionAdj
rootAdj - adj global dofs indexed by rootSectionAdj
sf - describes shared points across procs
sfDof - describes shared dofs across procs
sfAdj - describes shared adjacent dofs across procs
** The bootstrapping process involves six rounds with similar structure of visiting neighbors of each point.
(0). If there are point-to-point constraints, add the adjacencies of constrained points to anchors in anchorAdj
(This is done in DMPlexComputeAnchorAdjacencies())
1. Visit unowned points on interface, count adjacencies placing in leafSectionAdj
Reduce those counts to rootSectionAdj (now redundantly counting some interface points)
2. Visit owned points on interface, count adjacencies placing in rootSectionAdj
Create sfAdj connecting rootSectionAdj and leafSectionAdj
3. Visit unowned points on interface, write adjacencies to adj
Gather adj to rootAdj (note that there is redundancy in rootAdj when multiple procs find the same adjacencies)
4. Visit owned points on interface, write adjacencies to rootAdj
Remove redundancy in rootAdj
** The last two traversals use transitive closure
5. Visit all owned points in the subdomain, count dofs for each point (sectionAdj)
Allocate memory addressed by sectionAdj (cols)
6. Visit all owned points in the subdomain, insert dof adjacencies into cols
** Knowing all the column adjacencies, check ownership and sum into dnz and onz
*/
ierr = DMPlexComputeAnchorAdjacencies(dm,section,sectionGlobal,&anchorSectionAdj,&anchorAdj);CHKERRQ(ierr);
for (l = 0; l < nleaves; ++l) {
PetscInt dof, off, d, q, anDof;
PetscInt p = leaves[l], numAdj = PETSC_DETERMINE;
if ((p < pStart) || (p >= pEnd)) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(leafSectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(leafSectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(leafSectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency Section for Preallocation on Leaves:\n");CHKERRQ(ierr);
ierr = PetscSectionView(leafSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Get maximum remote adjacency sizes for owned dofs on interface (roots) */
if (doComm) {
ierr = PetscSFReduceBegin(sfDof, MPIU_INT, leafSectionAdj->atlasDof, rootSectionAdj->atlasDof, MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sfDof, MPIU_INT, leafSectionAdj->atlasDof, rootSectionAdj->atlasDof, MPI_SUM);CHKERRQ(ierr);
}
if (debug) {
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Add in local adjacency sizes for owned dofs on interface (roots) */
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, adof, dof, off, d, q, anDof;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(rootSectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(rootSectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots after local additions:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Create adj SF based on dof SF */
ierr = PetscSFCreateRemoteOffsets(sfDof, rootSectionAdj, leafSectionAdj, &remoteOffsets);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sfDof, rootSectionAdj, remoteOffsets, leafSectionAdj, &sfAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sfAdj, NULL);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sfDof);CHKERRQ(ierr);
/* Create leaf adjacency */
ierr = PetscSectionSetUp(leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(leafSectionAdj, &adjSize);CHKERRQ(ierr);
ierr = PetscCalloc1(adjSize, &adj);CHKERRQ(ierr);
for (l = 0; l < nleaves; ++l) {
PetscInt dof, off, d, q, anDof, anOff;
PetscInt p = leaves[l], numAdj = PETSC_DETERMINE;
if ((p < pStart) || (p >= pEnd)) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
PetscInt aoff, i = 0;
ierr = PetscSectionGetOffset(leafSectionAdj, d, &aoff);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd) {
adj[aoff+i] = (ngoff < 0 ? -(ngoff+1) : ngoff) + nd;
++i;
}
}
for (q = 0; q < anDof; q++) {
adj[aoff+i] = anchorAdj[anOff+q];
++i;
}
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Leaf adjacency indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, adj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Gather adjacent indices to root */
ierr = PetscSectionGetStorageSize(rootSectionAdj, &adjSize);CHKERRQ(ierr);
ierr = PetscMalloc1(adjSize, &rootAdj);CHKERRQ(ierr);
for (r = 0; r < adjSize; ++r) rootAdj[r] = -1;
if (doComm) {
const PetscInt *indegree;
PetscInt *remoteadj, radjsize = 0;
ierr = PetscSFComputeDegreeBegin(sfAdj, &indegree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sfAdj, &indegree);CHKERRQ(ierr);
for (p = 0; p < adjSize; ++p) radjsize += indegree[p];
ierr = PetscMalloc1(radjsize, &remoteadj);CHKERRQ(ierr);
ierr = PetscSFGatherBegin(sfAdj, MPIU_INT, adj, remoteadj);CHKERRQ(ierr);
ierr = PetscSFGatherEnd(sfAdj, MPIU_INT, adj, remoteadj);CHKERRQ(ierr);
for (p = 0, l = 0, r = 0; p < adjSize; ++p, l = PetscMax(p, l + indegree[p-1])) {
PetscInt s;
for (s = 0; s < indegree[p]; ++s, ++r) rootAdj[l+s] = remoteadj[r];
}
if (r != radjsize) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Inconsistency in communication %d != %d", r, radjsize);
if (l != adjSize) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Inconsistency in communication %d != %d", l, adjSize);
ierr = PetscFree(remoteadj);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sfAdj);CHKERRQ(ierr);
ierr = PetscFree(adj);CHKERRQ(ierr);
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Root adjacency indices after gather\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Add in local adjacency indices for owned dofs on interface (roots) */
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, adof, dof, off, d, q, anDof, anOff;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
PetscInt adof, aoff, i;
ierr = PetscSectionGetDof(rootSectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, d, &aoff);CHKERRQ(ierr);
i = adof-1;
for (q = 0; q < anDof; q++) {
rootAdj[aoff+i] = anchorAdj[anOff+q];
--i;
}
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd) {
rootAdj[aoff+i] = ngoff < 0 ? -(ngoff+1)+nd : ngoff+nd;
--i;
}
}
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Root adjacency indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Compress indices */
ierr = PetscSectionSetUp(rootSectionAdj);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt dof, cdof, off, d;
PetscInt adof, aoff;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
for (d = off; d < off+dof-cdof; ++d) {
ierr = PetscSectionGetDof(rootSectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, d, &aoff);CHKERRQ(ierr);
ierr = PetscSortRemoveDupsInt(&adof, &rootAdj[aoff]);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSectionAdj, d, adof);CHKERRQ(ierr);
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots after compression:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Root adjacency indices after compression\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Build adjacency section: Maps global indices to sets of adjacent global indices */
ierr = PetscSectionGetOffsetRange(sectionGlobal, &globalOffStart, &globalOffEnd);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, §ionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(sectionAdj, globalOffStart, globalOffEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, dof, cdof, off, goff, d, q, anDof;
PetscBool found = PETSC_TRUE;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
for (d = 0; d < dof-cdof; ++d) {
PetscInt ldof, rdof;
ierr = PetscSectionGetDof(leafSectionAdj, off+d, &ldof);CHKERRQ(ierr);
ierr = PetscSectionGetDof(rootSectionAdj, off+d, &rdof);CHKERRQ(ierr);
if (ldof > 0) {
/* We do not own this point */
} else if (rdof > 0) {
ierr = PetscSectionSetDof(sectionAdj, goff+d, rdof);CHKERRQ(ierr);
} else {
found = PETSC_FALSE;
}
}
if (found) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, noff;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, padj, &noff);CHKERRQ(ierr);
for (d = goff; d < goff+dof-cdof; ++d) {
ierr = PetscSectionAddDof(sectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = goff; d < goff+dof-cdof; ++d) {
ierr = PetscSectionAddDof(sectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(sectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(sectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Get adjacent indices */
ierr = PetscSectionGetStorageSize(sectionAdj, &numCols);CHKERRQ(ierr);
ierr = PetscMalloc1(numCols, &cols);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, dof, cdof, off, goff, d, q, anDof, anOff;
PetscBool found = PETSC_TRUE;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
for (d = 0; d < dof-cdof; ++d) {
PetscInt ldof, rdof;
ierr = PetscSectionGetDof(leafSectionAdj, off+d, &ldof);CHKERRQ(ierr);
ierr = PetscSectionGetDof(rootSectionAdj, off+d, &rdof);CHKERRQ(ierr);
if (ldof > 0) {
/* We do not own this point */
} else if (rdof > 0) {
PetscInt aoff, roff;
ierr = PetscSectionGetOffset(sectionAdj, goff+d, &aoff);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, off+d, &roff);CHKERRQ(ierr);
ierr = PetscMemcpy(&cols[aoff], &rootAdj[roff], rdof * sizeof(PetscInt));CHKERRQ(ierr);
} else {
found = PETSC_FALSE;
}
}
if (found) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = goff; d < goff+dof-cdof; ++d) {
PetscInt adof, aoff, i = 0;
ierr = PetscSectionGetDof(sectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, d, &aoff);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
const PetscInt *ncind;
/* Adjacent points may not be in the section chart */
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintIndices(section, padj, &ncind);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd, ++i) {
cols[aoff+i] = ngoff < 0 ? -(ngoff+1)+nd : ngoff+nd;
}
}
for (q = 0; q < anDof; q++, i++) {
cols[aoff+i] = anchorAdj[anOff + q];
}
if (i != adof) SETERRQ4(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of entries %D != %D for dof %D (point %D)", i, adof, d, p);
}
}
ierr = PetscSectionDestroy(&anchorSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSectionAdj);CHKERRQ(ierr);
ierr = PetscFree(anchorAdj);CHKERRQ(ierr);
ierr = PetscFree(rootAdj);CHKERRQ(ierr);
ierr = PetscFree(tmpAdj);CHKERRQ(ierr);
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Column indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, numCols, cols, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Create allocation vectors from adjacency graph */
ierr = MatGetLocalSize(A, &locRows, NULL);CHKERRQ(ierr);
ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)A), &rLayout);CHKERRQ(ierr);
ierr = PetscLayoutSetLocalSize(rLayout, locRows);CHKERRQ(ierr);
ierr = PetscLayoutSetBlockSize(rLayout, 1);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(rLayout);CHKERRQ(ierr);
ierr = PetscLayoutGetRange(rLayout, &rStart, &rEnd);CHKERRQ(ierr);
ierr = PetscLayoutDestroy(&rLayout);CHKERRQ(ierr);
/* Only loop over blocks of rows */
if (rStart%bs || rEnd%bs) SETERRQ3(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid layout [%d, %d) for matrix, must be divisible by block size %d", rStart, rEnd, bs);
for (r = rStart/bs; r < rEnd/bs; ++r) {
const PetscInt row = r*bs;
PetscInt numCols, cStart, c;
ierr = PetscSectionGetDof(sectionAdj, row, &numCols);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, row, &cStart);CHKERRQ(ierr);
for (c = cStart; c < cStart+numCols; ++c) {
if ((cols[c] >= rStart*bs) && (cols[c] < rEnd*bs)) {
++dnz[r-rStart];
if (cols[c] >= row) ++dnzu[r-rStart];
} else {
++onz[r-rStart];
if (cols[c] >= row) ++onzu[r-rStart];
}
}
}
if (bs > 1) {
for (r = 0; r < locRows/bs; ++r) {
dnz[r] /= bs;
onz[r] /= bs;
dnzu[r] /= bs;
onzu[r] /= bs;
}
}
/* Set matrix pattern */
ierr = MatXAIJSetPreallocation(A, bs, dnz, onz, dnzu, onzu);CHKERRQ(ierr);
ierr = MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr);
/* Check for symmetric storage */
ierr = MatGetType(A, &mtype);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATSBAIJ, &isSymBlock);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATSEQSBAIJ, &isSymSeqBlock);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATMPISBAIJ, &isSymMPIBlock);CHKERRQ(ierr);
if (isSymBlock || isSymSeqBlock || isSymMPIBlock) {ierr = MatSetOption(A, MAT_IGNORE_LOWER_TRIANGULAR, PETSC_TRUE);CHKERRQ(ierr);}
/* Fill matrix with zeros */
if (fillMatrix) {
PetscScalar *values;
PetscInt maxRowLen = 0;
for (r = rStart; r < rEnd; ++r) {
PetscInt len;
ierr = PetscSectionGetDof(sectionAdj, r, &len);CHKERRQ(ierr);
maxRowLen = PetscMax(maxRowLen, len);
}
ierr = PetscCalloc1(maxRowLen, &values);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
PetscInt numCols, cStart;
ierr = PetscSectionGetDof(sectionAdj, r, &numCols);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, r, &cStart);CHKERRQ(ierr);
ierr = MatSetValues(A, 1, &r, numCols, &cols[cStart], values, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree(values);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
}
/* restore original useAnchors */
ierr = DMPlexSetAdjacencyUseAnchors(dm,useAnchors);CHKERRQ(ierr);
ierr = PetscSectionDestroy(§ionAdj);CHKERRQ(ierr);
ierr = PetscFree(cols);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Preallocate,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
PetscSF sf,sfDup,sfInv,sfEmbed,sfA,sfB,sfBA;
const PetscInt *degree;
PetscErrorCode ierr;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test setup */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = CheckRanksEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test setup then reset */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test view (no graph set, no type set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set graph then view (no type set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set type then view (no graph set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set type then graph then view */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set graph then type */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test Bcast */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPI_INT,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFBcastEnd (sf,MPI_INT,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test Reduce */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPI_INT,NULL,NULL,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPI_INT,NULL,NULL,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPI_INT,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPI_INT,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test FetchAndOp */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpBegin(sf,MPI_INT,NULL,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpEnd (sf,MPI_INT,NULL,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test ComputeDegree */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(sf,°ree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sf,°ree);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFDuplicate() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_GRAPH,&sfDup);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfDup);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfDup);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateInverseSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateInverseSF(sf,&sfInv);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfInv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfInv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateEmbeddedSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateEmbeddedSF(sf,0,NULL,&sfEmbed);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateEmbeddedLeafSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateEmbeddedLeafSF(sf,0,NULL,&sfEmbed);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCompose() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sfA);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfA,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sfB);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfB,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCompose(sfA,sfB,&sfBA);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfBA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfBA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfB);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);
ierr = (*mat->ops->increaseoverlap)(mat,1,is,ov);CHKERRQ(ierr);
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){
ierr = PetscCommDestroy(&scomm);CHKERRQ(ierr);
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){
ierr = PetscCommDestroy(&scomm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
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);
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];
}
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 to avoid using uninitialized variable */
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);
ierr = PetscFree(localsizes_sc);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots,&indices_recv);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,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);
ierr = PetscFree2(indices_ov_rd,sources_sc_rd);CHKERRQ(ierr);
ierr = ISCreateGeneral(scomm,nroots,indices_recv,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr);
ierr = MatGetSubMatricesMPI(mat,1,&is_sc,&is_sc,MAT_INITIAL_MATRIX,&smat);CHKERRQ(ierr);
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);
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);
ierr = ISDestroy(&is_sc);CHKERRQ(ierr);
ierr = ISDestroy(&partitioning);CHKERRQ(ierr);
ierr = PetscCommDestroy(&scomm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
