PetscErrorCode KSPSetUp_AGMRES(KSP ksp)
{
PetscErrorCode ierr;
PetscInt hes;
PetscInt nloc;
KSP_AGMRES *agmres = (KSP_AGMRES*)ksp->data;
PetscInt neig = agmres->neig;
PetscInt max_k = agmres->max_k;
PetscInt N = MAXKSPSIZE;
PetscInt lwork = PetscMax(8 * N + 16, 4 * neig * (N - neig));
PetscFunctionBegin;
if (ksp->pc_side == PC_SYMMETRIC) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"no symmetric preconditioning for KSPAGMRES");
max_k = agmres->max_k;
N = MAXKSPSIZE;
/* Preallocate space during the call to KSPSetup_GMRES for the Krylov basis */
agmres->q_preallocate = PETSC_TRUE; /* No allocation on the fly */
/* Preallocate space to compute later the eigenvalues in GMRES */
ksp->calc_sings = PETSC_TRUE;
agmres->max_k = N; /* Set the augmented size to be allocated in KSPSetup_GMRES */
ierr = KSPSetUp_DGMRES(ksp);CHKERRQ(ierr);
agmres->max_k = max_k;
hes = (N + 1) * (N + 1);
/* Data for the Newton basis GMRES */
ierr = PetscCalloc4(max_k,&agmres->Rshift,max_k,&agmres->Ishift,hes,&agmres->Rloc,((N+1)*4),&agmres->wbufptr);CHKERRQ(ierr);
ierr = PetscMalloc7((N+1),&agmres->Scale,(N+1),&agmres->sgn,(N+1),&agmres->tloc,(N+1),&agmres->temp,(N+1),&agmres->tau,lwork,&agmres->work,(N+1),&agmres->nrs);CHKERRQ(ierr);
ierr = PetscMemzero(agmres->Scale, (N+1)*sizeof(PetscScalar));CHKERRQ(ierr);
ierr = PetscMemzero(agmres->sgn, (N+1)*sizeof(PetscScalar));CHKERRQ(ierr);
ierr = PetscMemzero(agmres->tloc, (N+1)*sizeof(PetscScalar));CHKERRQ(ierr);
ierr = PetscMemzero(agmres->temp, (N+1)*sizeof(PetscScalar));CHKERRQ(ierr);
/* Allocate space for the vectors in the orthogonalized basis*/
ierr = VecGetLocalSize(agmres->vecs[0], &nloc);CHKERRQ(ierr);
ierr = PetscMalloc1(nloc*(N+1), &agmres->Qloc);CHKERRQ(ierr);
/* Init the ring of processors for the roddec orthogonalization */
ierr = KSPAGMRESRoddecInitNeighboor(ksp);CHKERRQ(ierr);
if (agmres->neig < 1) PetscFunctionReturn(0);
/* Allocate space for the deflation */
ierr = PetscMalloc1(N, &agmres->select);CHKERRQ(ierr);
ierr = VecDuplicateVecs(VEC_V(0), N, &agmres->TmpU);CHKERRQ(ierr);
ierr = PetscMalloc2(N*N, &agmres->MatEigL, N*N, &agmres->MatEigR);CHKERRQ(ierr);
/* ierr = PetscMalloc6(N*N, &agmres->Q, N*N, &agmres->Z, N, &agmres->wr, N, &agmres->wi, N, &agmres->beta, N, &agmres->modul);CHKERRQ(ierr); */
ierr = PetscMalloc3(N*N, &agmres->Q, N*N, &agmres->Z, N, &agmres->beta);CHKERRQ(ierr);
ierr = PetscMalloc2((N+1),&agmres->perm,(2*neig*N),&agmres->iwork);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PetscFEGeomCreate(PetscQuadrature quad, PetscInt numCells, PetscInt dimEmbed, PetscBool faceData, PetscFEGeom **geom)
{
PetscFEGeom *g;
PetscInt dim, Nq, N;
const PetscReal *p;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscQuadratureGetData(quad,&dim,NULL,&Nq,&p,NULL);CHKERRQ(ierr);
ierr = PetscNew(&g);CHKERRQ(ierr);
g->xi = p;
g->numCells = numCells;
g->numPoints = Nq;
g->dim = dim;
g->dimEmbed = dimEmbed;
N = numCells * Nq;
ierr = PetscCalloc3(N * dimEmbed, &g->v, N * dimEmbed * dimEmbed, &g->J, N, &g->detJ);CHKERRQ(ierr);
if (faceData) {
ierr = PetscCalloc4(numCells, &g->face, N * dimEmbed, &g->n, N * dimEmbed * dimEmbed, &(g->suppInvJ[0]), N * dimEmbed * dimEmbed, &(g->suppInvJ[1]));CHKERRQ(ierr);
}
ierr = PetscCalloc1(N * dimEmbed * dimEmbed, &g->invJ);CHKERRQ(ierr);
*geom = g;
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);
}
/*@
DMPlexOrient - Give a consistent orientation to the input mesh
Input Parameters:
. dm - The DM
Note: The orientation data for the DM are change in-place.
$ This routine will fail for non-orientable surfaces, such as the Moebius strip.
Level: advanced
.seealso: DMCreate(), DMPLEX
@*/
PetscErrorCode DMPlexOrient(DM dm)
{
MPI_Comm comm;
PetscSF sf;
const PetscInt *lpoints;
const PetscSFNode *rpoints;
PetscSFNode *rorntComp = NULL, *lorntComp = NULL;
PetscInt *numNeighbors, **neighbors;
PetscSFNode *nrankComp;
PetscBool *match, *flipped;
PetscBT seenCells, flippedCells, seenFaces;
PetscInt *faceFIFO, fTop, fBottom, *cellComp, *faceComp;
PetscInt numLeaves, numRoots, dim, h, cStart, cEnd, c, cell, fStart, fEnd, face, off, totNeighbors = 0;
PetscMPIInt rank, size, numComponents, comp = 0;
PetscBool flg, flg2;
PetscViewer viewer = NULL, selfviewer = NULL;
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 = PetscOptionsHasName(((PetscObject) dm)->options,((PetscObject) dm)->prefix, "-orientation_view", &flg);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->options,((PetscObject) dm)->prefix, "-orientation_view_synchronized", &flg2);CHKERRQ(ierr);
ierr = DMGetPointSF(dm, &sf);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf, &numRoots, &numLeaves, &lpoints, &rpoints);CHKERRQ(ierr);
/* Truth Table
mismatch flips do action mismatch flipA ^ flipB action
F 0 flips no F F F
F 1 flip yes F T T
F 2 flips no T F T
T 0 flips yes T T F
T 1 flip no
T 2 flips yes
*/
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMPlexGetVTKCellHeight(dm, &h);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, h, &cStart, &cEnd);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, h+1, &fStart, &fEnd);CHKERRQ(ierr);
ierr = PetscBTCreate(cEnd - cStart, &seenCells);CHKERRQ(ierr);
ierr = PetscBTMemzero(cEnd - cStart, seenCells);CHKERRQ(ierr);
ierr = PetscBTCreate(cEnd - cStart, &flippedCells);CHKERRQ(ierr);
ierr = PetscBTMemzero(cEnd - cStart, flippedCells);CHKERRQ(ierr);
ierr = PetscBTCreate(fEnd - fStart, &seenFaces);CHKERRQ(ierr);
ierr = PetscBTMemzero(fEnd - fStart, seenFaces);CHKERRQ(ierr);
ierr = PetscCalloc3(fEnd - fStart, &faceFIFO, cEnd-cStart, &cellComp, fEnd-fStart, &faceComp);CHKERRQ(ierr);
/*
OLD STYLE
- Add an integer array over cells and faces (component) for connected component number
Foreach component
- Mark the initial cell as seen
- Process component as usual
- Set component for all seenCells
- Wipe seenCells and seenFaces (flippedCells can stay)
- Generate parallel adjacency for component using SF and seenFaces
- Collect numComponents adj data from each proc to 0
- Build same serial graph
- Use same solver
- Use Scatterv to to send back flipped flags for each component
- Negate flippedCells by component
NEW STYLE
- Create the adj on each process
- Bootstrap to complete graph on proc 0
*/
/* Loop over components */
for (cell = cStart; cell < cEnd; ++cell) cellComp[cell-cStart] = -1;
do {
/* Look for first unmarked cell */
for (cell = cStart; cell < cEnd; ++cell) if (cellComp[cell-cStart] < 0) break;
if (cell >= cEnd) break;
/* Initialize FIFO with first cell in component */
{
const PetscInt *cone;
PetscInt coneSize;
fTop = fBottom = 0;
ierr = DMPlexGetConeSize(dm, cell, &coneSize);CHKERRQ(ierr);
ierr = DMPlexGetCone(dm, cell, &cone);CHKERRQ(ierr);
for (c = 0; c < coneSize; ++c) {
faceFIFO[fBottom++] = cone[c];
ierr = PetscBTSet(seenFaces, cone[c]-fStart);CHKERRQ(ierr);
}
ierr = PetscBTSet(seenCells, cell-cStart);CHKERRQ(ierr);
}
/* Consider each face in FIFO */
while (fTop < fBottom) {
ierr = DMPlexCheckFace_Internal(dm, faceFIFO, &fTop, &fBottom, cStart, fStart, fEnd, seenCells, flippedCells, seenFaces);CHKERRQ(ierr);
}
/* Set component for cells and faces */
for (cell = 0; cell < cEnd-cStart; ++cell) {
if (PetscBTLookup(seenCells, cell)) cellComp[cell] = comp;
}
for (face = 0; face < fEnd-fStart; ++face) {
if (PetscBTLookup(seenFaces, face)) faceComp[face] = comp;
}
/* Wipe seenCells and seenFaces for next component */
ierr = PetscBTMemzero(fEnd - fStart, seenFaces);CHKERRQ(ierr);
ierr = PetscBTMemzero(cEnd - cStart, seenCells);CHKERRQ(ierr);
++comp;
} while (1);
numComponents = comp;
if (flg) {
PetscViewer v;
ierr = PetscViewerASCIIGetStdout(comm, &v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPushSynchronized(v);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(v, "[%d]BT for serial flipped cells:\n", rank);CHKERRQ(ierr);
ierr = PetscBTView(cEnd-cStart, flippedCells, v);CHKERRQ(ierr);
ierr = PetscViewerFlush(v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPopSynchronized(v);CHKERRQ(ierr);
}
/* Now all subdomains are oriented, but we need a consistent parallel orientation */
if (numLeaves >= 0) {
/* Store orientations of boundary faces*/
ierr = PetscCalloc2(numRoots,&rorntComp,numRoots,&lorntComp);CHKERRQ(ierr);
for (face = fStart; face < fEnd; ++face) {
const PetscInt *cone, *support, *ornt;
PetscInt coneSize, supportSize;
ierr = DMPlexGetSupportSize(dm, face, &supportSize);CHKERRQ(ierr);
if (supportSize != 1) continue;
ierr = DMPlexGetSupport(dm, face, &support);CHKERRQ(ierr);
ierr = DMPlexGetCone(dm, support[0], &cone);CHKERRQ(ierr);
ierr = DMPlexGetConeSize(dm, support[0], &coneSize);CHKERRQ(ierr);
ierr = DMPlexGetConeOrientation(dm, support[0], &ornt);CHKERRQ(ierr);
for (c = 0; c < coneSize; ++c) if (cone[c] == face) break;
if (dim == 1) {
/* Use cone position instead, shifted to -1 or 1 */
if (PetscBTLookup(flippedCells, support[0]-cStart)) rorntComp[face].rank = 1-c*2;
else rorntComp[face].rank = c*2-1;
} else {
if (PetscBTLookup(flippedCells, support[0]-cStart)) rorntComp[face].rank = ornt[c] < 0 ? -1 : 1;
else rorntComp[face].rank = ornt[c] < 0 ? 1 : -1;
}
rorntComp[face].index = faceComp[face-fStart];
}
/* Communicate boundary edge orientations */
ierr = PetscSFBcastBegin(sf, MPIU_2INT, rorntComp, lorntComp);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_2INT, rorntComp, lorntComp);CHKERRQ(ierr);
}
/* Get process adjacency */
ierr = PetscMalloc2(numComponents, &numNeighbors, numComponents, &neighbors);CHKERRQ(ierr);
viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)dm));
if (flg2) {ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr);}
ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&selfviewer);CHKERRQ(ierr);
for (comp = 0; comp < numComponents; ++comp) {
PetscInt l, n;
numNeighbors[comp] = 0;
ierr = PetscMalloc1(PetscMax(numLeaves, 0), &neighbors[comp]);CHKERRQ(ierr);
/* I know this is p^2 time in general, but for bounded degree its alright */
for (l = 0; l < numLeaves; ++l) {
const PetscInt face = lpoints[l];
/* Find a representative face (edge) separating pairs of procs */
if ((face >= fStart) && (face < fEnd) && (faceComp[face-fStart] == comp)) {
const PetscInt rrank = rpoints[l].rank;
const PetscInt rcomp = lorntComp[face].index;
for (n = 0; n < numNeighbors[comp]; ++n) if ((rrank == rpoints[neighbors[comp][n]].rank) && (rcomp == lorntComp[lpoints[neighbors[comp][n]]].index)) break;
if (n >= numNeighbors[comp]) {
PetscInt supportSize;
ierr = DMPlexGetSupportSize(dm, face, &supportSize);CHKERRQ(ierr);
if (supportSize != 1) SETERRQ1(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Boundary faces should see one cell, not %d", supportSize);
if (flg) {ierr = PetscViewerASCIIPrintf(selfviewer, "[%d]: component %d, Found representative leaf %d (face %d) connecting to face %d on (%d, %d) with orientation %d\n", rank, comp, l, face, rpoints[l].index, rrank, rcomp, lorntComp[face].rank);CHKERRQ(ierr);}
neighbors[comp][numNeighbors[comp]++] = l;
}
}
}
totNeighbors += numNeighbors[comp];
}
ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&selfviewer);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
if (flg2) {ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr);}
ierr = PetscMalloc2(totNeighbors, &nrankComp, totNeighbors, &match);CHKERRQ(ierr);
for (comp = 0, off = 0; comp < numComponents; ++comp) {
PetscInt n;
for (n = 0; n < numNeighbors[comp]; ++n, ++off) {
const PetscInt face = lpoints[neighbors[comp][n]];
const PetscInt o = rorntComp[face].rank*lorntComp[face].rank;
if (o < 0) match[off] = PETSC_TRUE;
else if (o > 0) match[off] = PETSC_FALSE;
else SETERRQ5(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid face %d (%d, %d) neighbor: %d comp: %d", face, rorntComp[face], lorntComp[face], neighbors[comp][n], comp);
nrankComp[off].rank = rpoints[neighbors[comp][n]].rank;
nrankComp[off].index = lorntComp[lpoints[neighbors[comp][n]]].index;
}
ierr = PetscFree(neighbors[comp]);CHKERRQ(ierr);
}
/* Collect the graph on 0 */
if (numLeaves >= 0) {
Mat G;
PetscBT seenProcs, flippedProcs;
PetscInt *procFIFO, pTop, pBottom;
PetscInt *N = NULL, *Noff;
PetscSFNode *adj = NULL;
PetscBool *val = NULL;
PetscMPIInt *recvcounts = NULL, *displs = NULL, *Nc, p, o;
PetscMPIInt size = 0;
ierr = PetscCalloc1(numComponents, &flipped);CHKERRQ(ierr);
if (!rank) {ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);}
ierr = PetscCalloc4(size, &recvcounts, size+1, &displs, size, &Nc, size+1, &Noff);CHKERRQ(ierr);
ierr = MPI_Gather(&numComponents, 1, MPI_INT, Nc, 1, MPI_INT, 0, comm);CHKERRQ(ierr);
for (p = 0; p < size; ++p) {
displs[p+1] = displs[p] + Nc[p];
}
if (!rank) {ierr = PetscMalloc1(displs[size],&N);CHKERRQ(ierr);}
ierr = MPI_Gatherv(numNeighbors, numComponents, MPIU_INT, N, Nc, displs, MPIU_INT, 0, comm);CHKERRQ(ierr);
for (p = 0, o = 0; p < size; ++p) {
recvcounts[p] = 0;
for (c = 0; c < Nc[p]; ++c, ++o) recvcounts[p] += N[o];
displs[p+1] = displs[p] + recvcounts[p];
}
if (!rank) {ierr = PetscMalloc2(displs[size], &adj, displs[size], &val);CHKERRQ(ierr);}
ierr = MPI_Gatherv(nrankComp, totNeighbors, MPIU_2INT, adj, recvcounts, displs, MPIU_2INT, 0, comm);CHKERRQ(ierr);
ierr = MPI_Gatherv(match, totNeighbors, MPIU_BOOL, val, recvcounts, displs, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
ierr = PetscFree2(numNeighbors, neighbors);CHKERRQ(ierr);
if (!rank) {
for (p = 1; p <= size; ++p) {Noff[p] = Noff[p-1] + Nc[p-1];}
if (flg) {
PetscInt n;
for (p = 0, off = 0; p < size; ++p) {
for (c = 0; c < Nc[p]; ++c) {
ierr = PetscPrintf(PETSC_COMM_SELF, "Proc %d Comp %d:\n", p, c);CHKERRQ(ierr);
for (n = 0; n < N[Noff[p]+c]; ++n, ++off) {
ierr = PetscPrintf(PETSC_COMM_SELF, " edge (%d, %d) (%d):\n", adj[off].rank, adj[off].index, val[off]);CHKERRQ(ierr);
}
}
}
}
/* Symmetrize the graph */
ierr = MatCreate(PETSC_COMM_SELF, &G);CHKERRQ(ierr);
ierr = MatSetSizes(G, Noff[size], Noff[size], Noff[size], Noff[size]);CHKERRQ(ierr);
ierr = MatSetUp(G);CHKERRQ(ierr);
for (p = 0, off = 0; p < size; ++p) {
for (c = 0; c < Nc[p]; ++c) {
const PetscInt r = Noff[p]+c;
PetscInt n;
for (n = 0; n < N[r]; ++n, ++off) {
const PetscInt q = Noff[adj[off].rank] + adj[off].index;
const PetscScalar o = val[off] ? 1.0 : 0.0;
ierr = MatSetValues(G, 1, &r, 1, &q, &o, INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(G, 1, &q, 1, &r, &o, INSERT_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatAssemblyBegin(G, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(G, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscBTCreate(Noff[size], &seenProcs);CHKERRQ(ierr);
ierr = PetscBTMemzero(Noff[size], seenProcs);CHKERRQ(ierr);
ierr = PetscBTCreate(Noff[size], &flippedProcs);CHKERRQ(ierr);
ierr = PetscBTMemzero(Noff[size], flippedProcs);CHKERRQ(ierr);
ierr = PetscMalloc1(Noff[size], &procFIFO);CHKERRQ(ierr);
pTop = pBottom = 0;
for (p = 0; p < Noff[size]; ++p) {
if (PetscBTLookup(seenProcs, p)) continue;
/* Initialize FIFO with next proc */
procFIFO[pBottom++] = p;
ierr = PetscBTSet(seenProcs, p);CHKERRQ(ierr);
/* Consider each proc in FIFO */
while (pTop < pBottom) {
const PetscScalar *ornt;
const PetscInt *neighbors;
PetscInt proc, nproc, seen, flippedA, flippedB, mismatch, numNeighbors, n;
proc = procFIFO[pTop++];
flippedA = PetscBTLookup(flippedProcs, proc) ? 1 : 0;
ierr = MatGetRow(G, proc, &numNeighbors, &neighbors, &ornt);CHKERRQ(ierr);
/* Loop over neighboring procs */
for (n = 0; n < numNeighbors; ++n) {
nproc = neighbors[n];
mismatch = PetscRealPart(ornt[n]) > 0.5 ? 0 : 1;
seen = PetscBTLookup(seenProcs, nproc);
flippedB = PetscBTLookup(flippedProcs, nproc) ? 1 : 0;
if (mismatch ^ (flippedA ^ flippedB)) {
if (seen) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Previously seen procs %d and %d do not match: Fault mesh is non-orientable", proc, nproc);
if (!flippedB) {
ierr = PetscBTSet(flippedProcs, nproc);CHKERRQ(ierr);
} else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Inconsistent mesh orientation: Fault mesh is non-orientable");
} else if (mismatch && flippedA && flippedB) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Attempt to flip already flipped cell: Fault mesh is non-orientable");
if (!seen) {
procFIFO[pBottom++] = nproc;
ierr = PetscBTSet(seenProcs, nproc);CHKERRQ(ierr);
}
}
}
}
ierr = PetscFree(procFIFO);CHKERRQ(ierr);
ierr = MatDestroy(&G);CHKERRQ(ierr);
ierr = PetscFree2(adj, val);CHKERRQ(ierr);
ierr = PetscBTDestroy(&seenProcs);CHKERRQ(ierr);
}
/* Scatter flip flags */
{
PetscBool *flips = NULL;
if (!rank) {
ierr = PetscMalloc1(Noff[size], &flips);CHKERRQ(ierr);
for (p = 0; p < Noff[size]; ++p) {
flips[p] = PetscBTLookup(flippedProcs, p) ? PETSC_TRUE : PETSC_FALSE;
if (flg && flips[p]) {ierr = PetscPrintf(comm, "Flipping Proc+Comp %d:\n", p);CHKERRQ(ierr);}
}
for (p = 0; p < size; ++p) {
displs[p+1] = displs[p] + Nc[p];
}
}
ierr = MPI_Scatterv(flips, Nc, displs, MPIU_BOOL, flipped, numComponents, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
ierr = PetscFree(flips);CHKERRQ(ierr);
}
if (!rank) {ierr = PetscBTDestroy(&flippedProcs);CHKERRQ(ierr);}
ierr = PetscFree(N);CHKERRQ(ierr);
ierr = PetscFree4(recvcounts, displs, Nc, Noff);CHKERRQ(ierr);
ierr = PetscFree2(nrankComp, match);CHKERRQ(ierr);
/* Decide whether to flip cells in each component */
for (c = 0; c < cEnd-cStart; ++c) {if (flipped[cellComp[c]]) {ierr = PetscBTNegate(flippedCells, c);CHKERRQ(ierr);}}
ierr = PetscFree(flipped);CHKERRQ(ierr);
}
if (flg) {
PetscViewer v;
ierr = PetscViewerASCIIGetStdout(comm, &v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPushSynchronized(v);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(v, "[%d]BT for parallel flipped cells:\n", rank);CHKERRQ(ierr);
ierr = PetscBTView(cEnd-cStart, flippedCells, v);CHKERRQ(ierr);
ierr = PetscViewerFlush(v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPopSynchronized(v);CHKERRQ(ierr);
}
/* Reverse flipped cells in the mesh */
for (c = cStart; c < cEnd; ++c) {
if (PetscBTLookup(flippedCells, c-cStart)) {
ierr = DMPlexReverseCell(dm, c);CHKERRQ(ierr);
}
}
ierr = PetscBTDestroy(&seenCells);CHKERRQ(ierr);
ierr = PetscBTDestroy(&flippedCells);CHKERRQ(ierr);
ierr = PetscBTDestroy(&seenFaces);CHKERRQ(ierr);
ierr = PetscFree2(numNeighbors, neighbors);CHKERRQ(ierr);
ierr = PetscFree2(rorntComp, lorntComp);CHKERRQ(ierr);
ierr = PetscFree3(faceFIFO, cellComp, faceComp);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PipesView(Vec X,DM networkdm,Wash wash)
{
PetscErrorCode ierr;
Pipe pipe;
DMNetworkComponentGenericDataType *nwarr;
PetscInt pipeOffset,key,Start,End;
PetscMPIInt rank;
PetscInt nx,nnodes,nidx,*idx1,*idx2,*idx1_h,*idx2_h,idx_start,i,k,k1,xstart,j1;
Vec Xq,Xh,localX;
IS is1_q,is2_q,is1_h,is2_h;
VecScatter ctx_q,ctx_h;
PetscFunctionBegin;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
/* get num of local and global total nnodes */
nidx = wash->nnodes_loc;
ierr = MPIU_Allreduce(&nidx,&nx,1,MPIU_INT,MPI_SUM,PETSC_COMM_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&Xq);CHKERRQ(ierr);
if (rank == 0) { /* all entries of Xq are in proc[0] */
ierr = VecSetSizes(Xq,nx,PETSC_DECIDE);CHKERRQ(ierr);
} else {
ierr = VecSetSizes(Xq,0,PETSC_DECIDE);CHKERRQ(ierr);
}
ierr = VecSetFromOptions(Xq);CHKERRQ(ierr);
ierr = VecSet(Xq,0.0);CHKERRQ(ierr);
ierr = VecDuplicate(Xq,&Xh);CHKERRQ(ierr);
ierr = DMGetLocalVector(networkdm,&localX);CHKERRQ(ierr);
/* set idx1 and idx2 */
ierr = PetscCalloc4(nidx,&idx1,nidx,&idx2,nidx,&idx1_h,nidx,&idx2_h);CHKERRQ(ierr);
ierr = DMNetworkGetComponentDataArray(networkdm,&nwarr);CHKERRQ(ierr);
ierr = DMNetworkGetEdgeRange(networkdm,&Start, &End);CHKERRQ(ierr);
ierr = VecGetOwnershipRange(X,&xstart,NULL);CHKERRQ(ierr);
k1 = 0;
j1 = 0;
for (i = Start; i < End; i++) {
ierr = DMNetworkGetComponentTypeOffset(networkdm,i,0,&key,&pipeOffset);CHKERRQ(ierr);
pipe = (Pipe)(nwarr+pipeOffset);
nnodes = pipe->nnodes;
idx_start = pipe->id*nnodes;
for (k=0; k<nnodes; k++) {
idx1[k1] = xstart + j1*2*nnodes + 2*k;
idx2[k1] = idx_start + k;
idx1_h[k1] = xstart + j1*2*nnodes + 2*k + 1;
idx2_h[k1] = idx_start + k;
k1++;
}
j1++;
}
ierr = ISCreateGeneral(PETSC_COMM_SELF,nidx,idx1,PETSC_COPY_VALUES,&is1_q);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,nidx,idx2,PETSC_COPY_VALUES,&is2_q);CHKERRQ(ierr);
ierr = VecScatterCreate(X,is1_q,Xq,is2_q,&ctx_q);CHKERRQ(ierr);
ierr = VecScatterBegin(ctx_q,X,Xq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(ctx_q,X,Xq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,nidx,idx1_h,PETSC_COPY_VALUES,&is1_h);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,nidx,idx2_h,PETSC_COPY_VALUES,&is2_h);CHKERRQ(ierr);
ierr = VecScatterCreate(X,is1_h,Xh,is2_h,&ctx_h);CHKERRQ(ierr);
ierr = VecScatterBegin(ctx_h,X,Xh,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(ctx_h,X,Xh,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
if (!rank) printf("Xq: \n");
ierr = VecView(Xq,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
if (!rank) printf("Xh: \n");
ierr = VecView(Xh,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecScatterDestroy(&ctx_q);CHKERRQ(ierr);
ierr = PetscFree4(idx1,idx2,idx1_h,idx2_h);CHKERRQ(ierr);
ierr = ISDestroy(&is1_q);CHKERRQ(ierr);
ierr = ISDestroy(&is2_q);CHKERRQ(ierr);
ierr = VecScatterDestroy(&ctx_h);CHKERRQ(ierr);
ierr = ISDestroy(&is1_h);CHKERRQ(ierr);
ierr = ISDestroy(&is2_h);CHKERRQ(ierr);
ierr = VecDestroy(&Xq);CHKERRQ(ierr);
ierr = VecDestroy(&Xh);CHKERRQ(ierr);
ierr = DMRestoreLocalVector(networkdm,&localX);CHKERRQ(ierr);
PetscFunctionReturn(0);
}