PetscErrorCode MatColoringGetDegrees(Mat G,PetscInt distance,PetscInt *degrees)
{
PetscInt j,i,s,e,n,ln,lm,degree,bidx,idx,dist;
Mat lG,*lGs;
IS ris;
PetscErrorCode ierr;
PetscInt *seen;
const PetscInt *gidx;
PetscInt *idxbuf;
PetscInt *distbuf;
PetscInt ncols;
const PetscInt *cols;
PetscBool isSEQAIJ;
Mat_SeqAIJ *aij;
PetscInt *Gi,*Gj;
PetscFunctionBegin;
ierr = MatGetOwnershipRange(G,&s,&e);CHKERRQ(ierr);
n=e-s;
ierr = ISCreateStride(PetscObjectComm((PetscObject)G),n,s,1,&ris);CHKERRQ(ierr);
ierr = MatIncreaseOverlap(G,1,&ris,distance);CHKERRQ(ierr);
ierr = ISSort(ris);CHKERRQ(ierr);
ierr = MatGetSubMatrices(G,1,&ris,&ris,MAT_INITIAL_MATRIX,&lGs);CHKERRQ(ierr);
lG = lGs[0];
ierr = PetscObjectTypeCompare((PetscObject)lG,MATSEQAIJ,&isSEQAIJ);CHKERRQ(ierr);
if (!isSEQAIJ) {
SETERRQ(PetscObjectComm((PetscObject)G),PETSC_ERR_ARG_WRONGSTATE,"MatColoringDegrees requires an MPI/SEQAIJ Matrix");
}
ierr = MatGetSize(lG,&ln,&lm);CHKERRQ(ierr);
aij = (Mat_SeqAIJ*)lG->data;
Gi = aij->i;
Gj = aij->j;
ierr = PetscMalloc3(lm,&seen,lm,&idxbuf,lm,&distbuf);CHKERRQ(ierr);
for (i=0;i<ln;i++) {
seen[i]=-1;
}
ierr = ISGetIndices(ris,&gidx);CHKERRQ(ierr);
for (i=0;i<ln;i++) {
if (gidx[i] >= e || gidx[i] < s) continue;
bidx=-1;
ncols = Gi[i+1]-Gi[i];
cols = &(Gj[Gi[i]]);
degree = 0;
/* place the distance-one neighbors on the queue */
for (j=0;j<ncols;j++) {
bidx++;
seen[cols[j]] = i;
distbuf[bidx] = 1;
idxbuf[bidx] = cols[j];
}
while (bidx >= 0) {
/* pop */
idx = idxbuf[bidx];
dist = distbuf[bidx];
bidx--;
degree++;
if (dist < distance) {
ncols = Gi[idx+1]-Gi[idx];
cols = &(Gj[Gi[idx]]);
for (j=0;j<ncols;j++) {
if (seen[cols[j]] != i) {
bidx++;
seen[cols[j]] = i;
idxbuf[bidx] = cols[j];
distbuf[bidx] = dist+1;
}
}
}
}
degrees[gidx[i]-s] = degree;
}
ierr = ISRestoreIndices(ris,&gidx);CHKERRQ(ierr);
ierr = ISDestroy(&ris);CHKERRQ(ierr);
ierr = PetscFree3(seen,idxbuf,distbuf);CHKERRQ(ierr);
ierr = MatDestroyMatrices(1,&lGs);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat S, IS is_A_I, IS is_A_B, PetscInt ncc, IS is_cc[], PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers)
{
Mat A_II,A_IB,A_BI,A_BB;
ISLocalToGlobalMapping BtoNmap,ItoNmap;
PetscBT touched;
PetscInt i,n_I,n_B,n_local,*local_numbering;
PetscBool is_sorted;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = ISSorted(is_A_I,&is_sorted);CHKERRQ(ierr);
if (!is_sorted) {
SETERRQ(PetscObjectComm((PetscObject)is_A_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted");
}
ierr = ISSorted(is_A_B,&is_sorted);CHKERRQ(ierr);
if (!is_sorted) {
SETERRQ(PetscObjectComm((PetscObject)is_A_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted");
}
/* get sizes */
ierr = ISGetLocalSize(is_A_I,&n_I);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_A_B,&n_B);CHKERRQ(ierr);
n_local = n_I+n_B;
/* maps */
ierr = ISLocalToGlobalMappingCreateIS(is_A_B,&BtoNmap);CHKERRQ(ierr);
if (nlayers >= 0 && xadj != NULL && adjncy != NULL) { /* I problems have a different size of the original ones */
ierr = ISLocalToGlobalMappingCreateIS(is_A_I,&ItoNmap);CHKERRQ(ierr);
/* allocate some auxiliary space */
ierr = PetscMalloc1(n_local,&local_numbering);CHKERRQ(ierr);
ierr = PetscBTCreate(n_local,&touched);CHKERRQ(ierr);
} else {
ItoNmap = 0;
local_numbering = 0;
touched = 0;
}
/* get Schur complement matrices */
ierr = MatSchurComplementGetSubMatrices(S,&A_II,NULL,&A_IB,&A_BI,&A_BB);CHKERRQ(ierr);
/* allocate space for schur complements */
ierr = PetscMalloc5(ncc,&sub_schurs->is_AEj_I,ncc,&sub_schurs->is_AEj_B,ncc,&sub_schurs->S_Ej,ncc,&sub_schurs->work1,ncc,&sub_schurs->work2);CHKERRQ(ierr);
sub_schurs->n_subs = ncc;
/* cycle on subsets and extract schur complements */
for (i=0;i<sub_schurs->n_subs;i++) {
Mat AE_II,AE_IE,AE_EI,AE_EE;
IS is_I,is_subset_B;
/* get IS for subsets in B numbering */
ierr = ISDuplicate(is_cc[i],&sub_schurs->is_AEj_B[i]);CHKERRQ(ierr);
ierr = ISSort(sub_schurs->is_AEj_B[i]);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyIS(BtoNmap,IS_GTOLM_DROP,sub_schurs->is_AEj_B[i],&is_subset_B);CHKERRQ(ierr);
/* BB block on subset */
ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr);
if (ItoNmap) { /* is ItoNmap has been computed, extracts only a part of I dofs */
const PetscInt* idx_B;
PetscInt n_local_dofs,n_prev_added,j,layer,subset_size;
/* all boundary dofs must be skipped when adding layers */
ierr = PetscBTMemzero(n_local,touched);CHKERRQ(ierr);
ierr = ISGetIndices(is_A_B,&idx_B);CHKERRQ(ierr);
for (j=0;j<n_B;j++) {
ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr);
}
ierr = ISRestoreIndices(is_A_B,&idx_B);CHKERRQ(ierr);
/* add next layers of dofs */
ierr = ISGetLocalSize(is_cc[i],&subset_size);CHKERRQ(ierr);
ierr = ISGetIndices(is_cc[i],&idx_B);CHKERRQ(ierr);
ierr = PetscMemcpy(local_numbering,idx_B,subset_size*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISRestoreIndices(is_cc[i],&idx_B);CHKERRQ(ierr);
n_local_dofs = subset_size;
n_prev_added = subset_size;
for (layer=0;layer<nlayers;layer++) {
PetscInt n_added;
if (n_local_dofs == n_I+subset_size) break;
if (n_local_dofs > n_I+subset_size) {
SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %d. Out of bound access (%d > %d)",layer,n_local_dofs,n_I+subset_size);
}
ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr);
n_prev_added = n_added;
n_local_dofs += n_added;
if (!n_added) break;
}
/* IS for I dofs in original numbering and in I numbering */
ierr = ISCreateGeneral(PetscObjectComm((PetscObject)ItoNmap),n_local_dofs-subset_size,local_numbering+subset_size,PETSC_COPY_VALUES,&sub_schurs->is_AEj_I[i]);CHKERRQ(ierr);
ierr = ISSort(sub_schurs->is_AEj_I[i]);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,sub_schurs->is_AEj_I[i],&is_I);CHKERRQ(ierr);
/* II block */
ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr);
} else { /* in this case we can take references of already existing IS and matrices for I dofs */
/* IS for I dofs in original numbering */
ierr = PetscObjectReference((PetscObject)is_A_I);CHKERRQ(ierr);
sub_schurs->is_AEj_I[i] = is_A_I;
/* IS for I dofs in I numbering TODO: "first" argument of ISCreateStride is not general */
ierr = ISCreateStride(PetscObjectComm((PetscObject)is_A_I),n_I,0,1,&is_I);CHKERRQ(ierr);
/* II block is the same */
ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr);
AE_II = A_II;
}
/* IE block */
ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr);
/* EI block */
ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr);
/* setup Schur complements on subset */
ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&sub_schurs->S_Ej[i]);CHKERRQ(ierr);
ierr = MatGetVecs(sub_schurs->S_Ej[i],&sub_schurs->work1[i],&sub_schurs->work2[i]);CHKERRQ(ierr);
if (AE_II == A_II) { /* we can reuse the same ksp */
KSP ksp;
ierr = MatSchurComplementGetKSP(S,&ksp);CHKERRQ(ierr);
ierr = MatSchurComplementSetKSP(sub_schurs->S_Ej[i],ksp);CHKERRQ(ierr);
} else { /* build new ksp object which inherits ksp and pc types from the original one */
KSP origksp,schurksp;
PC origpc,schurpc;
KSPType ksp_type;
PCType pc_type;
PetscInt n_internal;
ierr = MatSchurComplementGetKSP(S,&origksp);CHKERRQ(ierr);
ierr = MatSchurComplementGetKSP(sub_schurs->S_Ej[i],&schurksp);CHKERRQ(ierr);
ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr);
ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr);
ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr);
ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr);
ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr);
ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr);
ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr);
if (n_internal) { /* UMFPACK gives error with 0 sized problems */
MatSolverPackage solver=NULL;
ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr);
if (solver) {
ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr);
}
}
ierr = KSPSetUp(schurksp);CHKERRQ(ierr);
}
/* free */
ierr = MatDestroy(&AE_II);CHKERRQ(ierr);
ierr = MatDestroy(&AE_EE);CHKERRQ(ierr);
ierr = MatDestroy(&AE_IE);CHKERRQ(ierr);
ierr = MatDestroy(&AE_EI);CHKERRQ(ierr);
ierr = ISDestroy(&is_I);CHKERRQ(ierr);
ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr);
}
/* free */
ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&BtoNmap);CHKERRQ(ierr);
ierr = PetscFree(local_numbering);CHKERRQ(ierr);
ierr = PetscBTDestroy(&touched);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatColoringCreateSmallestLastWeights(MatColoring mc,PetscReal *weights)
{
PetscInt *degrees,*degb,*llprev,*llnext;
PetscInt j,i,s,e,n,nin,ln,lm,degree,maxdegree=0,bidx,idx,dist,distance=mc->dist;
Mat lG,*lGs;
IS ris;
PetscErrorCode ierr;
PetscInt *seen;
const PetscInt *gidx;
PetscInt *idxbuf;
PetscInt *distbuf;
PetscInt ncols,nxt,prv,cur;
const PetscInt *cols;
PetscBool isSEQAIJ;
Mat_SeqAIJ *aij;
PetscInt *Gi,*Gj,*rperm;
Mat G = mc->mat;
PetscReal *lweights,r;
PetscRandom rand;
PetscFunctionBegin;
ierr = MatGetOwnershipRange(G,&s,&e);CHKERRQ(ierr);
n=e-s;
ierr = ISCreateStride(PetscObjectComm((PetscObject)G),n,s,1,&ris);CHKERRQ(ierr);
ierr = MatIncreaseOverlap(G,1,&ris,distance+1);CHKERRQ(ierr);
ierr = ISSort(ris);CHKERRQ(ierr);
ierr = MatGetSubMatrices(G,1,&ris,&ris,MAT_INITIAL_MATRIX,&lGs);CHKERRQ(ierr);
lG = lGs[0];
ierr = PetscObjectTypeCompare((PetscObject)lG,MATSEQAIJ,&isSEQAIJ);CHKERRQ(ierr);
if (!isSEQAIJ) {
SETERRQ(PetscObjectComm((PetscObject)G),PETSC_ERR_ARG_WRONGSTATE,"MatColoringDegrees requires an MPI/SEQAIJ Matrix");
}
ierr = MatGetSize(lG,&ln,&lm);CHKERRQ(ierr);
aij = (Mat_SeqAIJ*)lG->data;
Gi = aij->i;
Gj = aij->j;
ierr = PetscMalloc3(lm,&seen,lm,&idxbuf,lm,&distbuf);CHKERRQ(ierr);
ierr = PetscMalloc1(lm,°rees);CHKERRQ(ierr);
ierr = PetscMalloc1(lm,&lweights);CHKERRQ(ierr);
for (i=0;i<ln;i++) {
seen[i]=-1;
lweights[i] = 1.;
}
ierr = ISGetIndices(ris,&gidx);CHKERRQ(ierr);
for (i=0;i<ln;i++) {
bidx=-1;
ncols = Gi[i+1]-Gi[i];
cols = &(Gj[Gi[i]]);
degree = 0;
/* place the distance-one neighbors on the queue */
for (j=0;j<ncols;j++) {
bidx++;
seen[cols[j]] = i;
distbuf[bidx] = 1;
idxbuf[bidx] = cols[j];
}
while (bidx >= 0) {
/* pop */
idx = idxbuf[bidx];
dist = distbuf[bidx];
bidx--;
degree++;
if (dist < distance) {
ncols = Gi[idx+1]-Gi[idx];
cols = &(Gj[Gi[idx]]);
for (j=0;j<ncols;j++) {
if (seen[cols[j]] != i) {
bidx++;
seen[cols[j]] = i;
idxbuf[bidx] = cols[j];
distbuf[bidx] = dist+1;
}
}
}
}
degrees[i] = degree;
if (degree > maxdegree) maxdegree = degree;
}
/* bucket by degree by some random permutation */
ierr = PetscRandomCreate(PetscObjectComm((PetscObject)mc),&rand);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(rand);CHKERRQ(ierr);
ierr = PetscMalloc1(ln,&rperm);CHKERRQ(ierr);
for (i=0;i<ln;i++) {
ierr = PetscRandomGetValueReal(rand,&r);CHKERRQ(ierr);
lweights[i] = r;
rperm[i]=i;
}
ierr = PetscSortRealWithPermutation(lm,lweights,rperm);CHKERRQ(ierr);
ierr = PetscMalloc1(maxdegree+1,°b);CHKERRQ(ierr);
ierr = PetscMalloc2(ln,&llnext,ln,&llprev);CHKERRQ(ierr);
for (i=0;i<maxdegree+1;i++) {
degb[i] = -1;
}
for (i=0;i<ln;i++) {
llnext[i] = -1;
llprev[i] = -1;
seen[i] = -1;
}
for (i=0;i<ln;i++) {
idx = rperm[i];
llnext[idx] = degb[degrees[idx]];
if (degb[degrees[idx]] > 0) llprev[degb[degrees[idx]]] = idx;
degb[degrees[idx]] = idx;
}
ierr = PetscFree(rperm);CHKERRQ(ierr);
/* remove the lowest degree one */
i=0;
nin=0;
while (i != maxdegree+1) {
for (i=1;i<maxdegree+1; i++) {
if (degb[i] > 0) {
cur = degb[i];
nin++;
degrees[cur] = 0;
degb[i] = llnext[cur];
bidx=-1;
ncols = Gi[cur+1]-Gi[cur];
cols = &(Gj[Gi[cur]]);
/* place the distance-one neighbors on the queue */
for (j=0;j<ncols;j++) {
if (cols[j] != cur) {
bidx++;
seen[cols[j]] = i;
distbuf[bidx] = 1;
idxbuf[bidx] = cols[j];
}
}
while (bidx >= 0) {
/* pop */
idx = idxbuf[bidx];
dist = distbuf[bidx];
bidx--;
nxt=llnext[idx];
prv=llprev[idx];
if (degrees[idx] > 0) {
/* change up the degree of the neighbors still in the graph */
if (lweights[idx] <= lweights[cur]) lweights[idx] = lweights[cur]+1;
if (nxt > 0) {
llprev[nxt] = prv;
}
if (prv > 0) {
llnext[prv] = nxt;
} else {
degb[degrees[idx]] = nxt;
}
degrees[idx]--;
llnext[idx] = degb[degrees[idx]];
llprev[idx] = -1;
if (degb[degrees[idx]] >= 0) {
llprev[degb[degrees[idx]]] = idx;
}
degb[degrees[idx]] = idx;
if (dist < distance) {
ncols = Gi[idx+1]-Gi[idx];
cols = &(Gj[Gi[idx]]);
for (j=0;j<ncols;j++) {
if (seen[cols[j]] != i) {
bidx++;
seen[cols[j]] = i;
idxbuf[bidx] = cols[j];
distbuf[bidx] = dist+1;
}
}
}
}
}
break;
}
}
}
for (i=0;i<lm;i++) {
if (gidx[i] >= s && gidx[i] < e) {
weights[gidx[i]-s] = lweights[i];
}
}
ierr = PetscRandomDestroy(&rand);CHKERRQ(ierr);
ierr = PetscFree(degb);CHKERRQ(ierr);
ierr = PetscFree2(llnext,llprev);CHKERRQ(ierr);
ierr = PetscFree(degrees);CHKERRQ(ierr);
ierr = PetscFree(lweights);CHKERRQ(ierr);
ierr = ISRestoreIndices(ris,&gidx);CHKERRQ(ierr);
ierr = ISDestroy(&ris);CHKERRQ(ierr);
ierr = PetscFree3(seen,idxbuf,distbuf);CHKERRQ(ierr);
ierr = MatDestroyMatrices(1,&lGs);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat mat)
{
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
Mat_SeqAIJ *B = (Mat_SeqAIJ*)(aij->B->data);
PetscErrorCode ierr;
PetscInt i,j,*aj = B->j,ec = 0,*garray;
IS from,to;
Vec gvec;
PetscBool useblockis;
#if defined (PETSC_USE_CTABLE)
PetscTable gid1_lid1;
PetscTablePosition tpos;
PetscInt gid,lid;
#else
PetscInt N = mat->cmap->N,*indices;
#endif
PetscFunctionBegin;
#if defined (PETSC_USE_CTABLE)
/* use a table */
ierr = PetscTableCreate(aij->B->rmap->n,mat->cmap->N+1,&gid1_lid1);CHKERRQ(ierr);
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt data,gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr);
if (!data) {
/* one based table */
ierr = PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);CHKERRQ(ierr);
}
}
}
/* form array of columns we need */
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr);
ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
while (tpos) {
ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr);
gid--;
lid--;
garray[lid] = gid;
}
ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */
ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
for (i=0; i<ec; i++) {
ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr);
}
/* compact out the extra columns in B */
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr);
lid --;
aj[B->i[i] + j] = lid;
}
}
aij->B->cmap->n = aij->B->cmap->N = ec;
ierr = PetscLayoutSetUp((aij->B->cmap));CHKERRQ(ierr);
ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr);
#else
/* Make an array as long as the number of columns */
/* mark those columns that are in aij->B */
ierr = PetscMalloc((N+1)*sizeof(PetscInt),&indices);CHKERRQ(ierr);
ierr = PetscMemzero(indices,N*sizeof(PetscInt));CHKERRQ(ierr);
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
if (!indices[aj[B->i[i] + j] ]) ec++;
indices[aj[B->i[i] + j] ] = 1;
}
}
/* form array of columns we need */
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr);
ec = 0;
for (i=0; i<N; i++) {
if (indices[i]) garray[ec++] = i;
}
/* make indices now point into garray */
for (i=0; i<ec; i++) {
indices[garray[i]] = i;
}
/* compact out the extra columns in B */
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
}
}
aij->B->cmap->n = aij->B->cmap->N = ec;
ierr = PetscLayoutSetUp((aij->B->cmap));CHKERRQ(ierr);
ierr = PetscFree(indices);CHKERRQ(ierr);
#endif
/* create local vector that is used to scatter into */
ierr = VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);CHKERRQ(ierr);
/* create two temporary Index sets for build scatter gather */
/* check for the special case where blocks are communicated for faster VecScatterXXX */
useblockis = PETSC_FALSE;
if (mat->cmap->bs > 1) {
PetscInt bs = mat->cmap->bs,ibs,ga;
if (!(ec % bs)) {
useblockis = PETSC_TRUE;
for (i=0; i<ec/bs; i++) {
if ((ga = garray[ibs = i*bs]) % bs) {
useblockis = PETSC_FALSE;
break;
}
for (j=1; j<bs; j++) {
if (garray[ibs+j] != ga+j) {
useblockis = PETSC_FALSE;
break;
}
}
if (!useblockis) break;
}
}
}
#if defined(PETSC_USE_DEBUG)
i = (PetscInt)useblockis;
ierr = MPI_Allreduce(&i,&j,1,MPIU_INT,MPI_MIN,((PetscObject)mat)->comm); CHKERRQ(ierr);
if(j!=i) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Use of blocked not consistant (I am usning blocked)");
#endif
if (useblockis) {
PetscInt *ga,bs = mat->cmap->bs,iec = ec/bs;
if(ec%bs)SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"ec=%D bs=%D",ec,bs);
ierr = PetscInfo(mat,"Using block index set to define scatter\n");
ierr = PetscMalloc(iec*sizeof(PetscInt),&ga);CHKERRQ(ierr);
for (i=0; i<iec; i++) ga[i] = garray[i*bs]/bs;
ierr = ISCreateBlock(((PetscObject)mat)->comm,bs,iec,ga,PETSC_OWN_POINTER,&from);CHKERRQ(ierr);
} else {
ierr = ISCreateGeneral(((PetscObject)mat)->comm,ec,garray,PETSC_COPY_VALUES,&from);CHKERRQ(ierr);
}
ierr = ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);CHKERRQ(ierr);
/* create temporary global vector to generate scatter context */
/* This does not allocate the array's memory so is efficient */
ierr = VecCreateMPIWithArray(((PetscObject)mat)->comm,1,mat->cmap->n,mat->cmap->N,PETSC_NULL,&gvec);CHKERRQ(ierr);
/* generate the scatter context */
ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,aij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,aij->lvec);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,from);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,to);CHKERRQ(ierr);
aij->garray = garray;
ierr = PetscLogObjectMemory(mat,(ec+1)*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = VecDestroy(&gvec);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
PetscErrorCode ierr;
PetscInt n = 5,N,i;
PetscMPIInt size,rank;
PetscScalar value,zero = 0.0;
Vec x,y;
IS is1,is2;
VecScatter ctx = 0;
ierr = PetscInitialize(&argc,&argv,(char*)0,help);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
/* create two vectors */
N = size*n;
ierr = VecCreate(PETSC_COMM_WORLD,&y);CHKERRQ(ierr);
ierr = VecSetSizes(y,n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetFromOptions(y);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&x);CHKERRQ(ierr);
ierr = VecSetSizes(x,n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetFromOptions(x);CHKERRQ(ierr);
/* create two index sets */
ierr = ISCreateStride(PETSC_COMM_WORLD,n,n*rank,1,&is1);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_WORLD,n,(n*(rank+1))%N,1,&is2);CHKERRQ(ierr);
/* fill local part of parallel vector x */
value = (PetscScalar)(rank+1);
for (i=n*rank; i<n*(rank+1); i++) {
ierr = VecSetValues(x,1,&i,&value,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(x);CHKERRQ(ierr);
ierr = VecAssemblyEnd(x);CHKERRQ(ierr);
ierr = VecSet(y,zero);CHKERRQ(ierr);
ierr = VecScatterCreate(x,is1,y,is2,&ctx);CHKERRQ(ierr);
for (i=0; i<100; i++) {
PetscReal ynorm;
PetscInt j;
ierr = VecNormBegin(y,NORM_2,&ynorm);CHKERRQ(ierr);
ierr = PetscCommSplitReductionBegin(((PetscObject)y)->comm);CHKERRQ(ierr);
for (j=0; j<3; j++) {
ierr = VecScatterBegin(ctx,x,y,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(ctx,x,y,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
}
ierr = VecNormEnd(y,NORM_2,&ynorm);CHKERRQ(ierr);
/* ierr = PetscPrintf(PETSC_COMM_WORLD,"ynorm = %8.2G\n",ynorm);CHKERRQ(ierr); */
}
ierr = VecScatterDestroy(&ctx);CHKERRQ(ierr);
ierr = VecView(y,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&y);CHKERRQ(ierr);
ierr = ISDestroy(&is1);CHKERRQ(ierr);
ierr = ISDestroy(&is2);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
PetscErrorCode PCBDDCNullSpaceAssembleCorrection(PC pc,IS local_dofs)
{
PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
PC_IS *pcis = (PC_IS*)pc->data;
Mat_IS* matis = (Mat_IS*)pc->pmat->data;
KSP *local_ksp;
PC newpc;
NullSpaceCorrection_ctx shell_ctx;
Mat local_mat,local_pmat,small_mat,inv_small_mat;
MatStructure local_mat_struct;
Vec work1,work2;
const Vec *nullvecs;
VecScatter scatter_ctx;
IS is_aux;
MatFactorInfo matinfo;
PetscScalar *basis_mat,*Kbasis_mat,*array,*array_mat;
PetscScalar one = 1.0,zero = 0.0, m_one = -1.0;
PetscInt basis_dofs,basis_size,nnsp_size,i,k,n_I,n_R;
PetscBool nnsp_has_cnst;
PetscErrorCode ierr;
PetscFunctionBegin;
/* Infer the local solver */
ierr = ISGetSize(local_dofs,&basis_dofs);CHKERRQ(ierr);
ierr = VecGetSize(pcis->vec1_D,&n_I);CHKERRQ(ierr);
ierr = VecGetSize(pcbddc->vec1_R,&n_R);CHKERRQ(ierr);
if (basis_dofs == n_I) {
/* Dirichlet solver */
local_ksp = &pcbddc->ksp_D;
} else if (basis_dofs == n_R) {
/* Neumann solver */
local_ksp = &pcbddc->ksp_R;
} else {
SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in %s: unknown local IS size %d. n_I=%d, n_R=%d)\n",__FUNCT__,basis_dofs,n_I,n_R);
}
ierr = KSPGetOperators(*local_ksp,&local_mat,&local_pmat,&local_mat_struct);CHKERRQ(ierr);
/* Get null space vecs */
ierr = MatNullSpaceGetVecs(pcbddc->NullSpace,&nnsp_has_cnst,&nnsp_size,&nullvecs);CHKERRQ(ierr);
basis_size = nnsp_size;
if (nnsp_has_cnst) {
basis_size++;
}
if (basis_dofs) {
/* Create shell ctx */
ierr = PetscMalloc(sizeof(*shell_ctx),&shell_ctx);CHKERRQ(ierr);
/* Create work vectors in shell context */
ierr = VecCreate(PETSC_COMM_SELF,&shell_ctx->work_small_1);CHKERRQ(ierr);
ierr = VecSetSizes(shell_ctx->work_small_1,basis_size,basis_size);CHKERRQ(ierr);
ierr = VecSetType(shell_ctx->work_small_1,VECSEQ);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_small_1,&shell_ctx->work_small_2);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_SELF,&shell_ctx->work_full_1);CHKERRQ(ierr);
ierr = VecSetSizes(shell_ctx->work_full_1,basis_dofs,basis_dofs);CHKERRQ(ierr);
ierr = VecSetType(shell_ctx->work_full_1,VECSEQ);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&shell_ctx->work_full_2);CHKERRQ(ierr);
/* Allocate workspace */
ierr = MatCreateSeqDense(PETSC_COMM_SELF,basis_dofs,basis_size,NULL,&shell_ctx->basis_mat );CHKERRQ(ierr);
ierr = MatCreateSeqDense(PETSC_COMM_SELF,basis_dofs,basis_size,NULL,&shell_ctx->Kbasis_mat);CHKERRQ(ierr);
ierr = MatDenseGetArray(shell_ctx->basis_mat,&basis_mat);CHKERRQ(ierr);
ierr = MatDenseGetArray(shell_ctx->Kbasis_mat,&Kbasis_mat);CHKERRQ(ierr);
/* Restrict local null space on selected dofs (Dirichlet or Neumann)
and compute matrices N and K*N */
ierr = VecDuplicate(shell_ctx->work_full_1,&work1);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work2);CHKERRQ(ierr);
ierr = VecScatterCreate(pcis->vec1_N,local_dofs,work1,(IS)0,&scatter_ctx);CHKERRQ(ierr);
}
for (k=0;k<nnsp_size;k++) {
ierr = VecScatterBegin(matis->ctx,nullvecs[k],pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(matis->ctx,nullvecs[k],pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
if (basis_dofs) {
ierr = VecPlaceArray(work1,(const PetscScalar*)&basis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = VecScatterBegin(scatter_ctx,pcis->vec1_N,work1,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(scatter_ctx,pcis->vec1_N,work1,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecPlaceArray(work2,(const PetscScalar*)&Kbasis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work2);CHKERRQ(ierr);
ierr = VecResetArray(work1);CHKERRQ(ierr);
ierr = VecResetArray(work2);CHKERRQ(ierr);
}
}
if (basis_dofs) {
if (nnsp_has_cnst) {
ierr = VecPlaceArray(work1,(const PetscScalar*)&basis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = VecSet(work1,one);CHKERRQ(ierr);
ierr = VecPlaceArray(work2,(const PetscScalar*)&Kbasis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work2);CHKERRQ(ierr);
ierr = VecResetArray(work1);CHKERRQ(ierr);
ierr = VecResetArray(work2);CHKERRQ(ierr);
}
ierr = VecDestroy(&work1);CHKERRQ(ierr);
ierr = VecDestroy(&work2);CHKERRQ(ierr);
ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(shell_ctx->basis_mat,&basis_mat);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(shell_ctx->Kbasis_mat,&Kbasis_mat);CHKERRQ(ierr);
/* Assemble another Mat object in shell context */
ierr = MatTransposeMatMult(shell_ctx->basis_mat,shell_ctx->Kbasis_mat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&small_mat);CHKERRQ(ierr);
ierr = MatFactorInfoInitialize(&matinfo);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,basis_size,0,1,&is_aux);CHKERRQ(ierr);
ierr = MatLUFactor(small_mat,is_aux,is_aux,&matinfo);CHKERRQ(ierr);
ierr = ISDestroy(&is_aux);CHKERRQ(ierr);
ierr = PetscMalloc(basis_size*basis_size*sizeof(PetscScalar),&array_mat);CHKERRQ(ierr);
for (k=0;k<basis_size;k++) {
ierr = VecSet(shell_ctx->work_small_1,zero);CHKERRQ(ierr);
ierr = VecSetValue(shell_ctx->work_small_1,k,one,INSERT_VALUES);CHKERRQ(ierr);
ierr = VecAssemblyBegin(shell_ctx->work_small_1);CHKERRQ(ierr);
ierr = VecAssemblyEnd(shell_ctx->work_small_1);CHKERRQ(ierr);
ierr = MatSolve(small_mat,shell_ctx->work_small_1,shell_ctx->work_small_2);CHKERRQ(ierr);
ierr = VecGetArrayRead(shell_ctx->work_small_2,(const PetscScalar**)&array);CHKERRQ(ierr);
for (i=0;i<basis_size;i++) {
array_mat[i*basis_size+k]=array[i];
}
ierr = VecRestoreArrayRead(shell_ctx->work_small_2,(const PetscScalar**)&array);CHKERRQ(ierr);
}
ierr = MatCreateSeqDense(PETSC_COMM_SELF,basis_size,basis_size,array_mat,&inv_small_mat);CHKERRQ(ierr);
ierr = MatMatMult(shell_ctx->basis_mat,inv_small_mat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&shell_ctx->Lbasis_mat);CHKERRQ(ierr);
ierr = PetscFree(array_mat);CHKERRQ(ierr);
ierr = MatDestroy(&inv_small_mat);CHKERRQ(ierr);
ierr = MatDestroy(&small_mat);CHKERRQ(ierr);
ierr = MatScale(shell_ctx->Kbasis_mat,m_one);CHKERRQ(ierr);
/* Rebuild local PC */
ierr = KSPGetPC(*local_ksp,&shell_ctx->local_pc);CHKERRQ(ierr);
ierr = PetscObjectReference((PetscObject)shell_ctx->local_pc);CHKERRQ(ierr);
ierr = PCCreate(PETSC_COMM_SELF,&newpc);CHKERRQ(ierr);
ierr = PCSetOperators(newpc,local_mat,local_mat,SAME_PRECONDITIONER);CHKERRQ(ierr);
ierr = PCSetType(newpc,PCSHELL);CHKERRQ(ierr);
ierr = PCShellSetContext(newpc,shell_ctx);CHKERRQ(ierr);
ierr = PCShellSetApply(newpc,PCBDDCApplyNullSpaceCorrectionPC);CHKERRQ(ierr);
ierr = PCShellSetDestroy(newpc,PCBDDCDestroyNullSpaceCorrectionPC);CHKERRQ(ierr);
ierr = PCSetUp(newpc);CHKERRQ(ierr);
ierr = KSPSetPC(*local_ksp,newpc);CHKERRQ(ierr);
ierr = PCDestroy(&newpc);CHKERRQ(ierr);
ierr = KSPSetUp(*local_ksp);CHKERRQ(ierr);
}
/* test */
if (pcbddc->dbg_flag && basis_dofs) {
KSP check_ksp;
PC check_pc;
Mat test_mat;
Vec work3;
PetscReal test_err,lambda_min,lambda_max;
PetscBool setsym,issym=PETSC_FALSE;
PetscInt tabs;
ierr = PetscViewerASCIIGetTab(pcbddc->dbg_viewer,&tabs);CHKERRQ(ierr);
ierr = KSPGetPC(*local_ksp,&check_pc);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work1);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work2);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work3);CHKERRQ(ierr);
ierr = VecSetRandom(shell_ctx->work_small_1,NULL);CHKERRQ(ierr);
ierr = MatMult(shell_ctx->basis_mat,shell_ctx->work_small_1,work1);CHKERRQ(ierr);
ierr = VecCopy(work1,work2);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work3);CHKERRQ(ierr);
ierr = PCApply(check_pc,work3,work1);CHKERRQ(ierr);
ierr = VecAXPY(work1,m_one,work2);CHKERRQ(ierr);
ierr = VecNorm(work1,NORM_INFINITY,&test_err);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d error for nullspace correction for ",PetscGlobalRank);
ierr = PetscViewerASCIIUseTabs(pcbddc->dbg_viewer,PETSC_FALSE);CHKERRQ(ierr);
if (basis_dofs == n_I) {
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Dirichlet ");
} else {
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Neumann ");
}
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"solver is :%1.14e\n",test_err);
ierr = PetscViewerASCIISetTab(pcbddc->dbg_viewer,tabs);CHKERRQ(ierr);
ierr = PetscViewerASCIIUseTabs(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
ierr = MatTransposeMatMult(shell_ctx->Lbasis_mat,shell_ctx->Kbasis_mat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&test_mat);CHKERRQ(ierr);
ierr = MatShift(test_mat,one);CHKERRQ(ierr);
ierr = MatNorm(test_mat,NORM_INFINITY,&test_err);CHKERRQ(ierr);
ierr = MatDestroy(&test_mat);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d error for nullspace matrices is :%1.14e\n",PetscGlobalRank,test_err);
/* Create ksp object suitable for extreme eigenvalues' estimation */
ierr = KSPCreate(PETSC_COMM_SELF,&check_ksp);CHKERRQ(ierr);
ierr = KSPSetOperators(check_ksp,local_mat,local_mat,SAME_PRECONDITIONER);CHKERRQ(ierr);
ierr = KSPSetTolerances(check_ksp,1.e-8,1.e-8,PETSC_DEFAULT,basis_dofs);CHKERRQ(ierr);
ierr = KSPSetComputeSingularValues(check_ksp,PETSC_TRUE);CHKERRQ(ierr);
ierr = MatIsSymmetricKnown(pc->pmat,&setsym,&issym);CHKERRQ(ierr);
if (issym) {
ierr = KSPSetType(check_ksp,KSPCG);CHKERRQ(ierr);
}
ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr);
ierr = KSPSetUp(check_ksp);CHKERRQ(ierr);
ierr = VecSetRandom(work1,NULL);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work2);CHKERRQ(ierr);
ierr = KSPSolve(check_ksp,work2,work2);CHKERRQ(ierr);
ierr = VecAXPY(work2,m_one,work1);CHKERRQ(ierr);
ierr = VecNorm(work2,NORM_INFINITY,&test_err);CHKERRQ(ierr);
ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max,&lambda_min);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(check_ksp,&k);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d error for adapted KSP %1.14e (it %d, eigs %1.6e %1.6e)\n",PetscGlobalRank,test_err,k,lambda_min,lambda_max);
ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr);
ierr = VecDestroy(&work1);CHKERRQ(ierr);
ierr = VecDestroy(&work2);CHKERRQ(ierr);
ierr = VecDestroy(&work3);CHKERRQ(ierr);
}
/* all processes shoud call this, even the void ones */
if (pcbddc->dbg_flag) {
ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
int main(int argc,char **args)
{
Mat A,Atrans,sA,*submatA,*submatsA;
PetscErrorCode ierr;
PetscMPIInt size,rank;
PetscInt bs=1,mbs=10,ov=1,i,j,k,*rows,*cols,nd=2,*idx,rstart,rend,sz,M,N,Mbs;
PetscScalar *vals,rval,one=1.0;
IS *is1,*is2;
PetscRandom rand;
PetscBool flg,TestOverlap,TestSubMat,TestAllcols;
PetscLogStage stages[2];
PetscInt vid = -1;
PetscInitialize(&argc,&args,(char *)0,help);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(PETSC_NULL,"-mat_block_size",&bs,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(PETSC_NULL,"-mat_mbs",&mbs,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(PETSC_NULL,"-ov",&ov,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(PETSC_NULL,"-nd",&nd,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(PETSC_NULL,"-view_id",&vid,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsHasName(PETSC_NULL, "-test_overlap", &TestOverlap);CHKERRQ(ierr);
ierr = PetscOptionsHasName(PETSC_NULL, "-test_submat", &TestSubMat);CHKERRQ(ierr);
ierr = PetscOptionsHasName(PETSC_NULL, "-test_allcols", &TestAllcols);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,mbs*bs,mbs*bs,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
ierr = MatSetType(A,MATBAIJ);CHKERRQ(ierr);
ierr = MatSeqBAIJSetPreallocation(A,bs,PETSC_DEFAULT,PETSC_NULL);
ierr = MatMPIBAIJSetPreallocation(A,bs,PETSC_DEFAULT,PETSC_NULL,PETSC_DEFAULT,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscRandomCreate(PETSC_COMM_WORLD,&rand);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(rand);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(A,&rstart,&rend);CHKERRQ(ierr);
ierr = MatGetSize(A,&M,&N);
Mbs = M/bs;
ierr = PetscMalloc(bs*sizeof(PetscInt),&rows);CHKERRQ(ierr);
ierr = PetscMalloc(bs*sizeof(PetscInt),&cols);CHKERRQ(ierr);
ierr = PetscMalloc(bs*bs*sizeof(PetscScalar),&vals);CHKERRQ(ierr);
ierr = PetscMalloc(M*sizeof(PetscScalar),&idx);CHKERRQ(ierr);
/* Now set blocks of values */
for (j=0; j<bs*bs; j++) vals[j] = 0.0;
for (i=0; i<Mbs; i++){
cols[0] = i*bs; rows[0] = i*bs;
for (j=1; j<bs; j++) {
rows[j] = rows[j-1]+1;
cols[j] = cols[j-1]+1;
}
ierr = MatSetValues(A,bs,rows,bs,cols,vals,ADD_VALUES);CHKERRQ(ierr);
}
/* second, add random blocks */
for (i=0; i<20*bs; i++) {
ierr = PetscRandomGetValue(rand,&rval);CHKERRQ(ierr);
cols[0] = bs*(PetscInt)(PetscRealPart(rval)*Mbs);
ierr = PetscRandomGetValue(rand,&rval);CHKERRQ(ierr);
rows[0] = rstart + bs*(PetscInt)(PetscRealPart(rval)*mbs);
for (j=1; j<bs; j++) {
rows[j] = rows[j-1]+1;
cols[j] = cols[j-1]+1;
}
for (j=0; j<bs*bs; j++) {
ierr = PetscRandomGetValue(rand,&rval);CHKERRQ(ierr);
vals[j] = rval;
}
ierr = MatSetValues(A,bs,rows,bs,cols,vals,ADD_VALUES);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* make A a symmetric matrix: A <- A^T + A */
ierr = MatTranspose(A,MAT_INITIAL_MATRIX, &Atrans);CHKERRQ(ierr);
ierr = MatAXPY(A,one,Atrans,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
ierr = MatDestroy(&Atrans);CHKERRQ(ierr);
ierr = MatTranspose(A,MAT_INITIAL_MATRIX, &Atrans);
ierr = MatEqual(A, Atrans, &flg);
if (flg) {
ierr = MatSetOption(A,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr);
} else {
SETERRQ(PETSC_COMM_SELF,1,"A+A^T is non-symmetric");
}
ierr = MatDestroy(&Atrans);CHKERRQ(ierr);
/* create a SeqSBAIJ matrix sA (= A) */
ierr = MatConvert(A,MATSBAIJ,MAT_INITIAL_MATRIX,&sA);CHKERRQ(ierr);
if (vid >= 0 && vid < size){
if (!rank) printf("A: \n");
ierr = MatView(A,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
if (!rank) printf("sA: \n");
ierr = MatView(sA,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Test sA==A through MatMult() */
ierr = MatMultEqual(A,sA,10,&flg);CHKERRQ(ierr);
if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG ,"Error in MatConvert(): A != sA");
/* Test MatIncreaseOverlap() */
ierr = PetscMalloc(nd*sizeof(IS **),&is1);CHKERRQ(ierr);
ierr = PetscMalloc(nd*sizeof(IS **),&is2);CHKERRQ(ierr);
for (i=0; i<nd; i++) {
if (!TestAllcols){
ierr = PetscRandomGetValue(rand,&rval);CHKERRQ(ierr);
sz = (PetscInt)((0.5+0.2*PetscRealPart(rval))*mbs); /* 0.5*mbs < sz < 0.7*mbs */
for (j=0; j<sz; j++) {
ierr = PetscRandomGetValue(rand,&rval);CHKERRQ(ierr);
idx[j*bs] = bs*(PetscInt)(PetscRealPart(rval)*Mbs);
for (k=1; k<bs; k++) idx[j*bs+k] = idx[j*bs]+k;
}
ierr = ISCreateGeneral(PETSC_COMM_SELF,sz*bs,idx,PETSC_COPY_VALUES,is1+i);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,sz*bs,idx,PETSC_COPY_VALUES,is2+i);CHKERRQ(ierr);
if (rank == vid){
ierr = PetscPrintf(PETSC_COMM_SELF," [%d] IS sz[%d]: %d\n",rank,i,sz);CHKERRQ(ierr);
ierr = ISView(is2[i],PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr);
}
} else { /* Test all rows and colums */
sz = M;
ierr = ISCreateStride(PETSC_COMM_SELF,sz,0,1,is1+i);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,sz,0,1,is2+i);CHKERRQ(ierr);
if (rank == vid){
PetscBool colflag;
ierr = ISIdentity(is2[i],&colflag);CHKERRQ(ierr);
printf("[%d] is2[%d], colflag %d\n",rank,i,colflag);
ierr = ISView(is2[i],PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr);
}
}
}
ierr = PetscLogStageRegister("MatOv_SBAIJ",&stages[0]);
ierr = PetscLogStageRegister("MatOv_BAIJ",&stages[1]);
/* Test MatIncreaseOverlap */
if (TestOverlap){
ierr = PetscLogStagePush(stages[0]);CHKERRQ(ierr);
ierr = MatIncreaseOverlap(sA,nd,is2,ov);CHKERRQ(ierr);
ierr = PetscLogStagePop();CHKERRQ(ierr);
ierr = PetscLogStagePush(stages[1]);CHKERRQ(ierr);
ierr = MatIncreaseOverlap(A,nd,is1,ov);CHKERRQ(ierr);
ierr = PetscLogStagePop();CHKERRQ(ierr);
if (rank == vid){
printf("\n[%d] IS from BAIJ:\n",rank);
ierr = ISView(is1[0],PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr);
printf("\n[%d] IS from SBAIJ:\n",rank);
ierr = ISView(is2[0],PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr);
}
for (i=0; i<nd; ++i) {
ierr = ISEqual(is1[i],is2[i],&flg);CHKERRQ(ierr);
if (!flg ){
if (rank == 0){
ierr = ISSort(is1[i]);CHKERRQ(ierr);
/* ISView(is1[i],PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); */
ierr = ISSort(is2[i]);CHKERRQ(ierr);
/* ISView(is2[i],PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); */
}
SETERRQ1(PETSC_COMM_SELF,1,"i=%D, is1 != is2",i);
}
}
}
/* Test MatGetSubmatrices */
if (TestSubMat){
for(i = 0; i < nd; ++i) {
ierr = ISSort(is1[i]); CHKERRQ(ierr);
ierr = ISSort(is2[i]); CHKERRQ(ierr);
}
ierr = MatGetSubMatrices(A,nd,is1,is1,MAT_INITIAL_MATRIX,&submatA);CHKERRQ(ierr);
ierr = MatGetSubMatrices(sA,nd,is2,is2,MAT_INITIAL_MATRIX,&submatsA);CHKERRQ(ierr);
ierr = MatMultEqual(A,sA,10,&flg);CHKERRQ(ierr);
if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A != sA");
/* Now test MatGetSubmatrices with MAT_REUSE_MATRIX option */
ierr = MatGetSubMatrices(A,nd,is1,is1,MAT_REUSE_MATRIX,&submatA);CHKERRQ(ierr);
ierr = MatGetSubMatrices(sA,nd,is2,is2,MAT_REUSE_MATRIX,&submatsA);CHKERRQ(ierr);
ierr = MatMultEqual(A,sA,10,&flg);CHKERRQ(ierr);
if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetSubmatrices(): A != sA");
for (i=0; i<nd; ++i) {
ierr = MatDestroy(&submatA[i]);CHKERRQ(ierr);
ierr = MatDestroy(&submatsA[i]);CHKERRQ(ierr);
}
ierr = PetscFree(submatA);CHKERRQ(ierr);
ierr = PetscFree(submatsA);CHKERRQ(ierr);
}
/* Free allocated memory */
for (i=0; i<nd; ++i) {
ierr = ISDestroy(&is1[i]);CHKERRQ(ierr);
ierr = ISDestroy(&is2[i]);CHKERRQ(ierr);
}
ierr = PetscFree(is1);CHKERRQ(ierr);
ierr = PetscFree(is2);CHKERRQ(ierr);
ierr = PetscFree(idx);CHKERRQ(ierr);
ierr = PetscFree(rows);CHKERRQ(ierr);
ierr = PetscFree(cols);CHKERRQ(ierr);
ierr = PetscFree(vals);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&sA);CHKERRQ(ierr);
ierr = PetscRandomDestroy(&rand);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
PetscErrorCode MatIncreaseOverlap_MPISBAIJ(Mat C,PetscInt is_max,IS is[],PetscInt ov)
{
PetscErrorCode ierr;
PetscInt i,N=C->cmap->N, bs=C->rmap->bs,M=C->rmap->N,Mbs=M/bs,*nidx,isz,iov;
IS *is_new,*is_row;
Mat *submats;
Mat_MPISBAIJ *c=(Mat_MPISBAIJ*)C->data;
Mat_SeqSBAIJ *asub_i;
PetscBT table;
PetscInt *ai,brow,nz,nis,l,nmax,nstages_local,nstages,max_no,pos;
const PetscInt *idx;
PetscBool flg,*allcolumns,*allrows;
PetscFunctionBegin;
ierr = PetscMalloc(is_max*sizeof(IS),&is_new);CHKERRQ(ierr);
/* Convert the indices into block format */
ierr = ISCompressIndicesGeneral(N,C->rmap->n,bs,is_max,is,is_new);CHKERRQ(ierr);
if (ov < 0){ SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");}
/* ----- previous non-scalable implementation ----- */
flg=PETSC_FALSE;
ierr = PetscOptionsHasName(PETSC_NULL, "-IncreaseOverlap_old", &flg);CHKERRQ(ierr);
if (flg){ /* previous non-scalable implementation */
printf("use previous non-scalable implementation...\n");
for (i=0; i<ov; ++i) {
ierr = MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr);
}
} else { /* implementation using modified BAIJ routines */
ierr = PetscMalloc((Mbs+1)*sizeof(PetscInt),&nidx);CHKERRQ(ierr);
ierr = PetscBTCreate(Mbs,&table);CHKERRQ(ierr); /* for column search */
ierr = PetscMalloc2(is_max+1,PetscBool,&allcolumns,is_max+1,PetscBool,&allrows);CHKERRQ(ierr);
/* Create is_row */
ierr = PetscMalloc(is_max*sizeof(IS **),&is_row);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,Mbs,0,1,&is_row[0]);CHKERRQ(ierr);
allrows[0] = PETSC_TRUE;
for (i=1; i<is_max; i++) {
is_row[i] = is_row[0]; /* reuse is_row[0] */
allrows[i] = PETSC_TRUE;
}
/* Allocate memory to hold all the submatrices - Modified from MatGetSubMatrices_MPIBAIJ() */
ierr = PetscMalloc((is_max+1)*sizeof(Mat),&submats);CHKERRQ(ierr);
/* Check for special case: each processor gets entire matrix columns */
for (i=0; i<is_max; i++) {
ierr = ISIdentity(is_new[i],&flg);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_new[i],&isz);CHKERRQ(ierr);
if (flg && isz == Mbs){
allcolumns[i] = PETSC_TRUE;
} else {
allcolumns[i] = PETSC_FALSE;
}
}
/* Determine the number of stages through which submatrices are done */
nmax = 20*1000000 / (c->Nbs * sizeof(PetscInt));
if (!nmax) nmax = 1;
nstages_local = is_max/nmax + ((is_max % nmax)?1:0);
/* Make sure every processor loops through the nstages */
ierr = MPI_Allreduce(&nstages_local,&nstages,1,MPIU_INT,MPI_MAX,((PetscObject)C)->comm);CHKERRQ(ierr);
for (iov=0; iov<ov; ++iov) {
/* 1) Get submats for column search */
for (i=0,pos=0; i<nstages; i++) {
if (pos+nmax <= is_max) max_no = nmax;
else if (pos == is_max) max_no = 0;
else max_no = is_max-pos;
c->ijonly = PETSC_TRUE;
ierr = MatGetSubMatrices_MPIBAIJ_local(C,max_no,is_row+pos,is_new+pos,MAT_INITIAL_MATRIX,allrows+pos,allcolumns+pos,submats+pos);CHKERRQ(ierr);
pos += max_no;
}
/* 2) Row search */
ierr = MatIncreaseOverlap_MPIBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr);
/* 3) Column search */
for (i=0; i<is_max; i++){
asub_i = (Mat_SeqSBAIJ*)submats[i]->data;
ai=asub_i->i;;
/* put is_new obtained from MatIncreaseOverlap_MPIBAIJ() to table */
ierr = PetscBTMemzero(Mbs,table);CHKERRQ(ierr);
ierr = ISGetIndices(is_new[i],&idx);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_new[i],&nis);CHKERRQ(ierr);
for (l=0; l<nis; l++) {
ierr = PetscBTSet(table,idx[l]);CHKERRQ(ierr);
nidx[l] = idx[l];
}
isz = nis;
/* add column entries to table */
for (brow=0; brow<Mbs; brow++){
nz = ai[brow+1] - ai[brow];
if (nz) {
if (!PetscBTLookupSet(table,brow)) nidx[isz++] = brow;
}
}
ierr = ISRestoreIndices(is_new[i],&idx);CHKERRQ(ierr);
ierr = ISDestroy(&is_new[i]);CHKERRQ(ierr);
/* create updated is_new */
ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,is_new+i);CHKERRQ(ierr);
}
/* Free tmp spaces */
for (i=0; i<is_max; i++){
ierr = MatDestroy(&submats[i]);CHKERRQ(ierr);
}
}
ierr = PetscFree2(allcolumns,allrows);CHKERRQ(ierr);
ierr = PetscBTDestroy(&table);CHKERRQ(ierr);
ierr = PetscFree(submats);CHKERRQ(ierr);
ierr = ISDestroy(&is_row[0]);CHKERRQ(ierr);
ierr = PetscFree(is_row);CHKERRQ(ierr);
ierr = PetscFree(nidx);CHKERRQ(ierr);
}
for (i=0; i<is_max; i++) {ierr = ISDestroy(&is[i]);CHKERRQ(ierr);}
ierr = ISExpandIndicesGeneral(N,N,bs,is_max,is_new,is);CHKERRQ(ierr);
for (i=0; i<is_max; i++) {ierr = ISDestroy(&is_new[i]);CHKERRQ(ierr);}
ierr = PetscFree(is_new);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc,char **args)
{
Mat A,F;
PetscViewer fd; /* viewer */
char file[PETSC_MAX_PATH_LEN]; /* input file name */
PetscErrorCode ierr;
PetscBool flg;
Vec x,y,w;
MatFactorInfo iluinfo;
IS perm;
PetscInt m;
PetscReal norm;
PetscInitialize(&argc,&args,(char *)0,help);
/*
Determine file from which we read the matrix
*/
ierr = PetscOptionsGetString(PETSC_NULL,"-f",file,PETSC_MAX_PATH_LEN,&flg);CHKERRQ(ierr);
if (!flg) SETERRQ(PETSC_COMM_WORLD,1,"Must indicate binary file with the -f option");
/*
Open binary file. Note that we use FILE_MODE_READ to indicate
reading from this file.
*/
ierr = PetscViewerBinaryOpen(PETSC_COMM_WORLD,file,FILE_MODE_READ,&fd);CHKERRQ(ierr);
/*
Load the matrix; then destroy the viewer.
*/
ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = MatSetType(A,MATSEQBAIJ);CHKERRQ(ierr);
ierr = MatLoad(A,fd);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&x);CHKERRQ(ierr);
ierr = VecLoad(x,fd);CHKERRQ(ierr);
ierr = PetscViewerDestroy(&fd);CHKERRQ(ierr);
ierr = VecDuplicate(x,&y);CHKERRQ(ierr);
ierr = VecDuplicate(x,&w);CHKERRQ(ierr);
ierr = MatGetFactor(A,"petsc",MAT_FACTOR_ILU,&F);CHKERRQ(ierr);
iluinfo.fill = 1.0;
ierr = MatGetSize(A,&m,0);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_WORLD,m,0,1,&perm);CHKERRQ(ierr);
ierr = MatLUFactorSymbolic(F,A,perm,perm,&iluinfo);CHKERRQ(ierr);
ierr = MatLUFactorNumeric(F,A,&iluinfo);CHKERRQ(ierr);
ierr = MatSolveTranspose(F,x,y);CHKERRQ(ierr);
F->ops->solvetranspose = MatSolveTranspose_SeqBAIJ_N;
ierr = MatSolveTranspose(F,x,w);CHKERRQ(ierr);
// VecView(w,0);VecView(y,0);
ierr = VecAXPY(w,-1.0,y);CHKERRQ(ierr);
ierr = VecNorm(w,NORM_2,&norm);CHKERRQ(ierr);
if (norm) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Norm of difference is nonzero %g\n",norm);CHKERRQ(ierr);
}
ierr = ISDestroy(&perm);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&F);CHKERRQ(ierr);
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&y);CHKERRQ(ierr);
ierr = VecDestroy(&w);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
static PetscErrorCode permutematrix(Mat Ain, Mat Bin, Mat *Aout, Mat *Bout, int **permIndices)
{
PetscErrorCode ierr;
MatPartitioning part;
IS isn, is, iscols;
PetscInt *nlocal,localCols,m,n;
PetscMPIInt size, rank;
MPI_Comm comm;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Ain,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetSize(Ain,&m,&n);CHKERRQ(ierr);
ierr = MatPartitioningCreate(comm,&part);CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,Ain);CHKERRQ(ierr);
ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr);
/* get new processor owner number of each vertex */
ierr = MatPartitioningApply(part,&is);CHKERRQ(ierr);
/* get new global number of each old global number */
ierr = ISPartitioningToNumbering(is,&isn);CHKERRQ(ierr);
ierr = PetscMalloc(size*sizeof(int),&nlocal);CHKERRQ(ierr);
/* get number of new vertices for each processor */
ierr = ISPartitioningCount(is,size,nlocal);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
/* get old global number of each new global number */
ierr = ISInvertPermutation(isn,nlocal[rank],&is);CHKERRQ(ierr);
ierr = ISDestroy(&isn);CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr);
ierr = ISSort(is);CHKERRQ(ierr);
/* If matrix is square, the permutation is applied to rows and columns;
otherwise it is only applied to rows. */
if (m == n) {
iscols = is;
localCols = nlocal[rank];
} else {
PetscInt lowj, highj;
ierr = MatGetOwnershipRangeColumn(Ain,&lowj,&highj);CHKERRQ(ierr);
localCols = highj-lowj;
ierr = ISCreateStride(comm,localCols, lowj, 1, &iscols);CHKERRQ(ierr);
}
/* copy permutation */
if (permIndices) {
const PetscInt *indices;
PetscInt i;
*permIndices = malloc(sizeof(int)*(nlocal[rank]+localCols));
ierr = ISGetIndices(is, &indices);CHKERRQ(ierr);
for (i=0; i<nlocal[rank]; i++) (*permIndices)[i] = indices[i];
ierr = ISRestoreIndices(is, &indices);CHKERRQ(ierr);
ierr = ISGetIndices(iscols, &indices);CHKERRQ(ierr);
for (i=0; i<localCols; i++) (*permIndices)[i+nlocal[rank]] = indices[i];
ierr = ISRestoreIndices(iscols, &indices);CHKERRQ(ierr);
}
ierr = PetscFree(nlocal);CHKERRQ(ierr);
ierr = MatGetSubMatrix(Ain,is,iscols,MAT_INITIAL_MATRIX,Aout);CHKERRQ(ierr);
if (Bin && Bout) {
ierr = MatGetSubMatrix(Bin,is,iscols,MAT_INITIAL_MATRIX,Bout);CHKERRQ(ierr);
}
ierr = ISDestroy(&is);CHKERRQ(ierr);
if (m != n) {
ierr = ISDestroy(&iscols);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
