/*@C
PCGAMGFilterGraph - filter (remove zero and possibly small values from the) graph and make it symmetric if requested
Collective on Mat
Input Parameter:
+ a_Gmat - the graph
. vfilter - threshold paramter [0,1)
- symm - make the result symmetric
Level: developer
Notes:
This is called before graph coarsers are called.
.seealso: PCGAMGSetThreshold()
@*/
PetscErrorCode PCGAMGFilterGraph(Mat *a_Gmat,PetscReal vfilter,PetscBool symm)
{
PetscErrorCode ierr;
PetscInt Istart,Iend,Ii,jj,ncols,nnz0,nnz1, NN, MM, nloc;
PetscMPIInt rank;
Mat Gmat = *a_Gmat, tGmat, matTrans;
MPI_Comm comm;
const PetscScalar *vals;
const PetscInt *idx;
PetscInt *d_nnz, *o_nnz;
Vec diag;
MatType mtype;
PetscFunctionBegin;
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
/* scale Gmat for all values between -1 and 1 */
ierr = MatCreateVecs(Gmat, &diag, 0);CHKERRQ(ierr);
ierr = MatGetDiagonal(Gmat, diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecSqrtAbs(diag);CHKERRQ(ierr);
ierr = MatDiagonalScale(Gmat, diag, diag);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
if (vfilter < 0.0 && !symm) {
/* Just use the provided matrix as the graph but make all values positive */
MatInfo info;
PetscScalar *avals;
PetscBool isaij,ismpiaij;
ierr = PetscObjectBaseTypeCompare((PetscObject)Gmat,MATSEQAIJ,&isaij);CHKERRQ(ierr);
ierr = PetscObjectBaseTypeCompare((PetscObject)Gmat,MATMPIAIJ,&ismpiaij);CHKERRQ(ierr);
if (!isaij && !ismpiaij) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_USER,"Require (MPI)AIJ matrix type");
if (isaij) {
ierr = MatGetInfo(Gmat,MAT_LOCAL,&info);CHKERRQ(ierr);
ierr = MatSeqAIJGetArray(Gmat,&avals);CHKERRQ(ierr);
for (jj = 0; jj<info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
ierr = MatSeqAIJRestoreArray(Gmat,&avals);CHKERRQ(ierr);
} else {
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)Gmat->data;
ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr);
ierr = MatSeqAIJGetArray(aij->A,&avals);CHKERRQ(ierr);
for (jj = 0; jj<info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
ierr = MatSeqAIJRestoreArray(aij->A,&avals);CHKERRQ(ierr);
ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr);
ierr = MatSeqAIJGetArray(aij->B,&avals);CHKERRQ(ierr);
for (jj = 0; jj<info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
ierr = MatSeqAIJRestoreArray(aij->B,&avals);CHKERRQ(ierr);
}
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
ierr = PetscObjectGetComm((PetscObject)Gmat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Gmat, &Istart, &Iend);CHKERRQ(ierr);
nloc = Iend - Istart;
ierr = MatGetSize(Gmat, &MM, &NN);CHKERRQ(ierr);
if (symm) {
ierr = MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans);CHKERRQ(ierr);
}
/* Determine upper bound on nonzeros needed in new filtered matrix */
ierr = PetscMalloc2(nloc, &d_nnz,nloc, &o_nnz);CHKERRQ(ierr);
for (Ii = Istart, jj = 0; Ii < Iend; Ii++, jj++) {
ierr = MatGetRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] = ncols;
o_nnz[jj] = ncols;
ierr = MatRestoreRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
if (symm) {
ierr = MatGetRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] += ncols;
o_nnz[jj] += ncols;
ierr = MatRestoreRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
}
if (d_nnz[jj] > nloc) d_nnz[jj] = nloc;
if (o_nnz[jj] > (MM-nloc)) o_nnz[jj] = MM - nloc;
}
ierr = MatGetType(Gmat,&mtype);CHKERRQ(ierr);
ierr = MatCreate(comm, &tGmat);CHKERRQ(ierr);
ierr = MatSetSizes(tGmat,nloc,nloc,MM,MM);CHKERRQ(ierr);
ierr = MatSetBlockSizes(tGmat, 1, 1);CHKERRQ(ierr);
ierr = MatSetType(tGmat, mtype);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(tGmat,0,d_nnz);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(tGmat,0,d_nnz,0,o_nnz);CHKERRQ(ierr);
ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr);
if (symm) {
ierr = MatDestroy(&matTrans);CHKERRQ(ierr);
} else {
/* all entries are generated locally so MatAssembly will be slightly faster for large process counts */
ierr = MatSetOption(tGmat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr);
}
for (Ii = Istart, nnz0 = nnz1 = 0; Ii < Iend; Ii++) {
ierr = MatGetRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
for (jj=0; jj<ncols; jj++,nnz0++) {
PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
if (PetscRealPart(sv) > vfilter) {
nnz1++;
if (symm) {
sv *= 0.5;
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(tGmat,1,&idx[jj],1,&Ii,&sv,ADD_VALUES);CHKERRQ(ierr);
} else {
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatRestoreRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
#if defined(PETSC_USE_INFO)
{
double t1 = (!nnz0) ? 1. : 100.*(double)nnz1/(double)nnz0, t2 = (!nloc) ? 1. : (double)nnz0/(double)nloc;
ierr = PetscInfo4(*a_Gmat,"\t %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%D)\n",t1,vfilter,t2,MM);CHKERRQ(ierr);
}
#endif
ierr = MatDestroy(&Gmat);CHKERRQ(ierr);
*a_Gmat = tGmat;
PetscFunctionReturn(0);
}
PetscErrorCode PCGAMGFilterGraph(Mat *a_Gmat,const PetscReal vfilter,const PetscBool symm,const PetscInt verbose)
{
PetscErrorCode ierr;
PetscInt Istart,Iend,Ii,jj,ncols,nnz0,nnz1, NN, MM, nloc;
PetscMPIInt rank, size;
Mat Gmat = *a_Gmat, tGmat, matTrans;
MPI_Comm comm;
const PetscScalar *vals;
const PetscInt *idx;
PetscInt *d_nnz, *o_nnz;
Vec diag;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Gmat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Gmat, &Istart, &Iend);CHKERRQ(ierr);
nloc = Iend - Istart;
ierr = MatGetSize(Gmat, &MM, &NN);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
/* scale Gmat so filter works */
ierr = MatGetVecs(Gmat, &diag, 0);CHKERRQ(ierr);
ierr = MatGetDiagonal(Gmat, diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecSqrtAbs(diag);CHKERRQ(ierr);
ierr = MatDiagonalScale(Gmat, diag, diag);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
if (symm) {
ierr = MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans);CHKERRQ(ierr);
}
/* filter - dup zeros out matrix */
ierr = PetscMalloc1(nloc, &d_nnz);CHKERRQ(ierr);
ierr = PetscMalloc1(nloc, &o_nnz);CHKERRQ(ierr);
for (Ii = Istart, jj = 0; Ii < Iend; Ii++, jj++) {
ierr = MatGetRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] = ncols;
o_nnz[jj] = ncols;
ierr = MatRestoreRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
if (symm) {
ierr = MatGetRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] += ncols;
o_nnz[jj] += ncols;
ierr = MatRestoreRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
}
if (d_nnz[jj] > nloc) d_nnz[jj] = nloc;
if (o_nnz[jj] > (MM-nloc)) o_nnz[jj] = MM - nloc;
}
ierr = MatCreateAIJ(comm, nloc, nloc, MM, MM, 0, d_nnz, 0, o_nnz, &tGmat);CHKERRQ(ierr);
ierr = PetscFree(d_nnz);CHKERRQ(ierr);
ierr = PetscFree(o_nnz);CHKERRQ(ierr);
if (symm) {
ierr = MatDestroy(&matTrans);CHKERRQ(ierr);
}
for (Ii = Istart, nnz0 = nnz1 = 0; Ii < Iend; Ii++) {
ierr = MatGetRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
for (jj=0; jj<ncols; jj++,nnz0++) {
PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
if (PetscRealPart(sv) > vfilter) {
nnz1++;
if (symm) {
sv *= 0.5;
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(tGmat,1,&idx[jj],1,&Ii,&sv,ADD_VALUES);CHKERRQ(ierr);
} else {
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatRestoreRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
if (verbose) {
if (verbose == 1) {
ierr = PetscPrintf(comm,"\t[%d]%s %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%d)\n",rank,__FUNCT__,
100.*(double)nnz1/(double)nnz0,vfilter,(double)nnz0/(double)nloc,MM);CHKERRQ(ierr);
} else {
PetscInt nnz[2],out[2];
nnz[0] = nnz0; nnz[1] = nnz1;
ierr = MPI_Allreduce(nnz, out, 2, MPIU_INT, MPI_SUM, comm);CHKERRQ(ierr);
ierr = PetscPrintf(comm,"\t[%d]%s %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%d)\n",rank,__FUNCT__,
100.*(double)out[1]/(double)out[0],vfilter,(double)out[0]/(double)MM,MM);CHKERRQ(ierr);
}
}
ierr = MatDestroy(&Gmat);CHKERRQ(ierr);
*a_Gmat = tGmat;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_LSQR(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,size1,size2;
PetscScalar rho,rhobar,phi,phibar,theta,c,s,tmp,tau;
PetscReal beta,alpha,rnorm;
Vec X,B,V,V1,U,U1,TMP,W,W2,SE,Z = NULL;
Mat Amat,Pmat;
MatStructure pflag;
KSP_LSQR *lsqr = (KSP_LSQR*)ksp->data;
PetscBool diagonalscale,nopreconditioner;
PetscFunctionBegin;
ierr = PCGetDiagonalScale(ksp->pc,&diagonalscale);CHKERRQ(ierr);
if (diagonalscale) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Krylov method %s does not support diagonal scaling",((PetscObject)ksp)->type_name);
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat,&pflag);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)ksp->pc,PCNONE,&nopreconditioner);CHKERRQ(ierr);
/* nopreconditioner =PETSC_FALSE; */
/* Calculate norm of right hand side */
ierr = VecNorm(ksp->vec_rhs,NORM_2,&lsqr->rhs_norm);CHKERRQ(ierr);
/* mark norm of matrix with negative number to indicate it has not yet been computed */
lsqr->anorm = -1.0;
/* vectors of length m, where system size is mxn */
B = ksp->vec_rhs;
U = lsqr->vwork_m[0];
U1 = lsqr->vwork_m[1];
/* vectors of length n */
X = ksp->vec_sol;
W = lsqr->vwork_n[0];
V = lsqr->vwork_n[1];
V1 = lsqr->vwork_n[2];
W2 = lsqr->vwork_n[3];
if (!nopreconditioner) Z = lsqr->vwork_n[4];
/* standard error vector */
SE = lsqr->se;
if (SE) {
ierr = VecGetSize(SE,&size1);CHKERRQ(ierr);
ierr = VecGetSize(X,&size2);CHKERRQ(ierr);
if (size1 != size2) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Standard error vector (size %d) does not match solution vector (size %d)",size1,size2);
ierr = VecSet(SE,0.0);CHKERRQ(ierr);
}
/* Compute initial residual, temporarily use work vector u */
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,U);CHKERRQ(ierr); /* u <- b - Ax */
ierr = VecAYPX(U,-1.0,B);CHKERRQ(ierr);
} else {
ierr = VecCopy(B,U);CHKERRQ(ierr); /* u <- b (x is 0) */
}
/* Test for nothing to do */
ierr = VecNorm(U,NORM_2,&rnorm);CHKERRQ(ierr);
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = rnorm;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
beta = rnorm;
ierr = VecScale(U,1.0/beta);CHKERRQ(ierr);
ierr = KSP_MatMultTranspose(ksp,Amat,U,V);CHKERRQ(ierr);
if (nopreconditioner) {
ierr = VecNorm(V,NORM_2,&alpha);CHKERRQ(ierr);
} else {
ierr = PCApply(ksp->pc,V,Z);CHKERRQ(ierr);
ierr = VecDotRealPart(V,Z,&alpha);CHKERRQ(ierr);
if (alpha <= 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
PetscFunctionReturn(0);
}
alpha = PetscSqrtReal(alpha);
ierr = VecScale(Z,1.0/alpha);CHKERRQ(ierr);
}
ierr = VecScale(V,1.0/alpha);CHKERRQ(ierr);
if (nopreconditioner) {
ierr = VecCopy(V,W);CHKERRQ(ierr);
} else {
ierr = VecCopy(Z,W);CHKERRQ(ierr);
}
lsqr->arnorm = alpha * beta;
phibar = beta;
rhobar = alpha;
i = 0;
do {
if (nopreconditioner) {
ierr = KSP_MatMult(ksp,Amat,V,U1);CHKERRQ(ierr);
} else {
ierr = KSP_MatMult(ksp,Amat,Z,U1);CHKERRQ(ierr);
}
ierr = VecAXPY(U1,-alpha,U);CHKERRQ(ierr);
ierr = VecNorm(U1,NORM_2,&beta);CHKERRQ(ierr);
if (beta == 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
ierr = VecScale(U1,1.0/beta);CHKERRQ(ierr);
ierr = KSP_MatMultTranspose(ksp,Amat,U1,V1);CHKERRQ(ierr);
ierr = VecAXPY(V1,-beta,V);CHKERRQ(ierr);
if (nopreconditioner) {
ierr = VecNorm(V1,NORM_2,&alpha);CHKERRQ(ierr);
} else {
ierr = PCApply(ksp->pc,V1,Z);CHKERRQ(ierr);
ierr = VecDotRealPart(V1,Z,&alpha);CHKERRQ(ierr);
if (alpha <= 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
alpha = PetscSqrtReal(alpha);
ierr = VecScale(Z,1.0/alpha);CHKERRQ(ierr);
}
ierr = VecScale(V1,1.0/alpha);CHKERRQ(ierr);
rho = PetscSqrtScalar(rhobar*rhobar + beta*beta);
c = rhobar / rho;
s = beta / rho;
theta = s * alpha;
rhobar = -c * alpha;
phi = c * phibar;
phibar = s * phibar;
tau = s * phi;
ierr = VecAXPY(X,phi/rho,W);CHKERRQ(ierr); /* x <- x + (phi/rho) w */
if (SE) {
ierr = VecCopy(W,W2);CHKERRQ(ierr);
ierr = VecSquare(W2);CHKERRQ(ierr);
ierr = VecScale(W2,1.0/(rho*rho));CHKERRQ(ierr);
ierr = VecAXPY(SE, 1.0, W2);CHKERRQ(ierr); /* SE <- SE + (w^2/rho^2) */
}
if (nopreconditioner) {
ierr = VecAYPX(W,-theta/rho,V1);CHKERRQ(ierr); /* w <- v - (theta/rho) w */
} else {
ierr = VecAYPX(W,-theta/rho,Z);CHKERRQ(ierr); /* w <- z - (theta/rho) w */
}
lsqr->arnorm = alpha*PetscAbsScalar(tau);
rnorm = PetscRealPart(phibar);
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = rnorm;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
SWAP(U1,U,TMP);
SWAP(V1,V,TMP);
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it && !ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
/* Finish off the standard error estimates */
if (SE) {
tmp = 1.0;
ierr = MatGetSize(Amat,&size1,&size2);CHKERRQ(ierr);
if (size1 > size2) tmp = size1 - size2;
tmp = rnorm / PetscSqrtScalar(tmp);
ierr = VecSqrtAbs(SE);CHKERRQ(ierr);
ierr = VecScale(SE,tmp);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PetscErrorCode heavyEdgeMatchAgg(IS perm,Mat a_Gmat,PetscInt verbose,PetscCoarsenData **a_locals_llist)
{
PetscErrorCode ierr;
PetscBool isMPI;
MPI_Comm wcomm = ((PetscObject)a_Gmat)->comm;
PetscInt sub_it,kk,n,ix,*idx,*ii,iter,Iend,my0;
PetscMPIInt rank,size;
const PetscInt nloc = a_Gmat->rmap->n,n_iter=6; /* need to figure out how to stop this */
PetscInt *lid_cprowID,*lid_gid;
PetscBool *lid_matched;
Mat_SeqAIJ *matA, *matB=0;
Mat_MPIAIJ *mpimat=0;
PetscScalar one=1.;
PetscCoarsenData *agg_llists = PETSC_NULL,*deleted_list = PETSC_NULL;
Mat cMat,tMat,P;
MatScalar *ap;
PetscMPIInt tag1,tag2;
PetscFunctionBegin;
ierr = MPI_Comm_rank(wcomm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(wcomm, &size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(a_Gmat, &my0, &Iend);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(wcomm, &tag1);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(wcomm, &tag2);CHKERRQ(ierr);
ierr = PetscMalloc(nloc*sizeof(PetscInt), &lid_gid);CHKERRQ(ierr); /* explicit array needed */
ierr = PetscMalloc(nloc*sizeof(PetscInt), &lid_cprowID);CHKERRQ(ierr);
ierr = PetscMalloc(nloc*sizeof(PetscBool), &lid_matched);CHKERRQ(ierr);
ierr = PetscCDCreate(nloc, &agg_llists);CHKERRQ(ierr);
/* ierr = PetscCDSetChuckSize(agg_llists, nloc+1);CHKERRQ(ierr); */
*a_locals_llist = agg_llists;
ierr = PetscCDCreate(size, &deleted_list);CHKERRQ(ierr);
ierr = PetscCDSetChuckSize(deleted_list, 100);CHKERRQ(ierr);
/* setup 'lid_gid' for scatters and add self to all lists */
for (kk=0;kk<nloc;kk++) {
lid_gid[kk] = kk + my0;
ierr = PetscCDAppendID(agg_llists, kk, my0+kk);CHKERRQ(ierr);
}
/* make a copy of the graph, this gets destroyed in iterates */
ierr = MatDuplicate(a_Gmat,MAT_COPY_VALUES,&cMat);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)a_Gmat, MATMPIAIJ, &isMPI);CHKERRQ(ierr);
iter = 0;
while(iter++ < n_iter) {
PetscScalar *cpcol_gid,*cpcol_max_ew,*cpcol_max_pe,*lid_max_ew;
PetscBool *cpcol_matched;
PetscMPIInt *cpcol_pe,proc;
Vec locMaxEdge,locMaxPE,ghostMaxEdge,ghostMaxPE;
PetscInt nEdges,n_nz_row,jj;
Edge *Edges;
PetscInt gid;
const PetscInt *perm_ix, n_sub_its = 120;
/* get submatrices of cMat */
if (isMPI) {
mpimat = (Mat_MPIAIJ*)cMat->data;
matA = (Mat_SeqAIJ*)mpimat->A->data;
matB = (Mat_SeqAIJ*)mpimat->B->data;
/* force compressed storage of B */
matB->compressedrow.check = PETSC_TRUE;
ierr = MatCheckCompressedRow(mpimat->B,&matB->compressedrow,matB->i,cMat->rmap->n,-1.0);CHKERRQ(ierr);
assert(matB->compressedrow.use);
} else {
matA = (Mat_SeqAIJ*)cMat->data;
}
assert(matA && !matA->compressedrow.use);
assert(matB==0 || matB->compressedrow.use);
/* set max edge on nodes */
ierr = MatGetVecs(cMat, &locMaxEdge, 0);CHKERRQ(ierr);
ierr = MatGetVecs(cMat, &locMaxPE, 0);CHKERRQ(ierr);
/* get 'cpcol_pe' & 'cpcol_gid' & init. 'cpcol_matched' using 'mpimat->lvec' */
if (mpimat) {
Vec vec;
PetscScalar vval;
ierr = MatGetVecs(cMat, &vec, 0);CHKERRQ(ierr);
/* cpcol_pe */
vval = (PetscScalar)(rank);
for (kk=0,gid=my0;kk<nloc;kk++,gid++) {
ierr = VecSetValues(vec, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(vec);CHKERRQ(ierr);
ierr = VecAssemblyEnd(vec);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr); /* get proc ID in 'cpcol_gid' */
ierr = VecGetLocalSize(mpimat->lvec, &n);CHKERRQ(ierr);
ierr = PetscMalloc(n*sizeof(PetscInt), &cpcol_pe);CHKERRQ(ierr);
for (kk=0;kk<n;kk++) cpcol_pe[kk] = (PetscMPIInt)PetscRealPart(cpcol_gid[kk]);
ierr = VecRestoreArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr);
/* cpcol_gid */
for (kk=0,gid=my0;kk<nloc;kk++,gid++) {
vval = (PetscScalar)(gid);
ierr = VecSetValues(vec, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(vec);CHKERRQ(ierr);
ierr = VecAssemblyEnd(vec);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecDestroy(&vec);CHKERRQ(ierr);
ierr = VecGetArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr); /* get proc ID in 'cpcol_gid' */
/* cpcol_matched */
ierr = VecGetLocalSize(mpimat->lvec, &n);CHKERRQ(ierr);
ierr = PetscMalloc(n*sizeof(PetscBool), &cpcol_matched);CHKERRQ(ierr);
for (kk=0;kk<n;kk++) cpcol_matched[kk] = PETSC_FALSE;
}
/* need an inverse map - locals */
for (kk=0;kk<nloc;kk++) lid_cprowID[kk] = -1;
/* set index into compressed row 'lid_cprowID' */
if (matB) {
ii = matB->compressedrow.i;
for (ix=0; ix<matB->compressedrow.nrows; ix++) {
lid_cprowID[matB->compressedrow.rindex[ix]] = ix;
}
}
/* get removed IS, use '' */
/* if (iter==1) { */
/* PetscInt *lid_rem,idx; */
/* ierr = PetscMalloc(nloc*sizeof(PetscInt), &lid_rem);CHKERRQ(ierr); */
/* for (kk=idx=0;kk<nloc;kk++){ */
/* PetscInt nn,lid=kk; */
/* ii = matA->i; nn = ii[lid+1] - ii[lid]; */
/* if ((ix=lid_cprowID[lid]) != -1) { /\* if I have any ghost neighbors *\/ */
/* ii = matB->compressedrow.i; */
/* nn += ii[ix+1] - ii[ix]; */
/* } */
/* if (nn < 2) { */
/* lid_rem[idx++] = kk + my0; */
/* } */
/* } */
/* ierr = PetscCDSetRemovedIS(agg_llists, wcomm, idx, lid_rem);CHKERRQ(ierr); */
/* ierr = PetscFree(lid_rem);CHKERRQ(ierr); */
/* } */
/* compute 'locMaxEdge' & 'locMaxPE', and create list of edges, count edges' */
for (nEdges=0,kk=0,gid=my0;kk<nloc;kk++,gid++){
PetscReal max_e = 0., tt;
PetscScalar vval;
PetscInt lid = kk;
PetscMPIInt max_pe=rank,pe;
ii = matA->i; n = ii[lid+1] - ii[lid]; idx = matA->j + ii[lid];
ap = matA->a + ii[lid];
for (jj=0; jj<n; jj++) {
PetscInt lidj = idx[jj];
if (lidj != lid && PetscRealPart(ap[jj]) > max_e) max_e = PetscRealPart(ap[jj]);
if (lidj > lid) nEdges++;
}
if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n ; jj++) {
if ((tt=PetscRealPart(ap[jj])) > max_e) max_e = tt;
nEdges++;
if ((pe=cpcol_pe[idx[jj]]) > max_pe) max_pe = pe;
}
}
vval = max_e;
ierr = VecSetValues(locMaxEdge, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr);
vval = (PetscScalar)max_pe;
ierr = VecSetValues(locMaxPE, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(locMaxEdge);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxEdge);CHKERRQ(ierr);
ierr = VecAssemblyBegin(locMaxPE);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxPE);CHKERRQ(ierr);
/* get 'cpcol_max_ew' & 'cpcol_max_pe' */
if (mpimat) {
ierr = VecDuplicate(mpimat->lvec, &ghostMaxEdge);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecDuplicate(mpimat->lvec, &ghostMaxPE);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
}
/* setup sorted list of edges */
ierr = PetscMalloc(nEdges*sizeof(Edge), &Edges);CHKERRQ(ierr);
ierr = ISGetIndices(perm, &perm_ix);CHKERRQ(ierr);
for (nEdges=n_nz_row=kk=0;kk<nloc;kk++){
PetscInt nn, lid = perm_ix[kk];
ii = matA->i; nn = n = ii[lid+1] - ii[lid]; idx = matA->j + ii[lid];
ap = matA->a + ii[lid];
for (jj=0; jj<n; jj++) {
PetscInt lidj = idx[jj]; assert(PetscRealPart(ap[jj])>0.);
if (lidj > lid) {
Edges[nEdges].lid0 = lid;
Edges[nEdges].gid1 = lidj + my0;
Edges[nEdges].cpid1 = -1;
Edges[nEdges].weight = PetscRealPart(ap[jj]);
nEdges++;
}
}
if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
nn += n;
for (jj=0 ; jj<n ; jj++) {
assert(PetscRealPart(ap[jj])>0.);
Edges[nEdges].lid0 = lid;
Edges[nEdges].gid1 = (PetscInt)PetscRealPart(cpcol_gid[idx[jj]]);
Edges[nEdges].cpid1 = idx[jj];
Edges[nEdges].weight = PetscRealPart(ap[jj]);
nEdges++;
}
}
if (nn > 1) n_nz_row++;
else if (iter == 1){
/* should select this because it is technically in the MIS but lets not */
ierr = PetscCDRemoveAll(agg_llists, lid);CHKERRQ(ierr);
}
}
ierr = ISRestoreIndices(perm,&perm_ix);CHKERRQ(ierr);
qsort(Edges, nEdges, sizeof(Edge), gamg_hem_compare);
/* projection matrix */
ierr = MatCreateAIJ(wcomm, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE, 1, 0, 1, 0, &P);CHKERRQ(ierr);
/* clear matched flags */
for (kk=0;kk<nloc;kk++) lid_matched[kk] = PETSC_FALSE;
/* process - communicate - process */
for (sub_it=0;sub_it<n_sub_its;sub_it++){
PetscInt nactive_edges;
ierr = VecGetArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
for (kk=nactive_edges=0;kk<nEdges;kk++){
/* HEM */
const Edge *e = &Edges[kk];
const PetscInt lid0=e->lid0,gid1=e->gid1,cpid1=e->cpid1,gid0=lid0+my0,lid1=gid1-my0;
PetscBool isOK = PETSC_TRUE;
/* skip if either (local) vertex is done already */
if (lid_matched[lid0] || (gid1>=my0 && gid1<Iend && lid_matched[gid1-my0])) {
continue;
}
/* skip if ghost vertex is done */
if (cpid1 != -1 && cpcol_matched[cpid1]) {
continue;
}
nactive_edges++;
/* skip if I have a bigger edge someplace (lid_max_ew gets updated) */
if (PetscRealPart(lid_max_ew[lid0]) > e->weight + 1.e-12) {
continue;
}
if (cpid1 == -1) {
if (PetscRealPart(lid_max_ew[lid1]) > e->weight + 1.e-12) {
continue;
}
} else {
/* see if edge might get matched on other proc */
PetscReal g_max_e = PetscRealPart(cpcol_max_ew[cpid1]);
if (g_max_e > e->weight + 1.e-12) {
continue;
} else if (e->weight > g_max_e - 1.e-12 && (PetscMPIInt)PetscRealPart(cpcol_max_pe[cpid1]) > rank) {
/* check for max_e == to this edge and larger processor that will deal with this */
continue;
}
}
/* check ghost for v0 */
if (isOK){
PetscReal max_e,ew;
if ((ix=lid_cprowID[lid0]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n && isOK; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
ew = PetscRealPart(ap[jj]); max_e = PetscRealPart(cpcol_max_ew[lidj]);
/* check for max_e == to this edge and larger processor that will deal with this */
if (ew > max_e - 1.e-12 && ew > PetscRealPart(lid_max_ew[lid0]) - 1.e-12 && (PetscMPIInt)PetscRealPart(cpcol_max_pe[lidj]) > rank){
isOK = PETSC_FALSE;
}
}
}
/* for v1 */
if (cpid1 == -1 && isOK){
if ((ix=lid_cprowID[lid1]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n && isOK ; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
ew = PetscRealPart(ap[jj]); max_e = PetscRealPart(cpcol_max_ew[lidj]);
/* check for max_e == to this edge and larger processor that will deal with this */
if (ew > max_e - 1.e-12 && ew > PetscRealPart(lid_max_ew[lid1]) - 1.e-12 && (PetscMPIInt)PetscRealPart(cpcol_max_pe[lidj]) > rank) {
isOK = PETSC_FALSE;
}
}
}
}
}
/* do it */
if (isOK){
if (cpid1 == -1) {
lid_matched[lid1] = PETSC_TRUE; /* keep track of what we've done this round */
ierr = PetscCDAppendRemove(agg_llists, lid0, lid1);CHKERRQ(ierr);
} else if (sub_it != n_sub_its-1) {
/* add gid1 to list of ghost deleted by me -- I need their children */
proc = cpcol_pe[cpid1];
cpcol_matched[cpid1] = PETSC_TRUE; /* messing with VecGetArray array -- needed??? */
ierr = PetscCDAppendID(deleted_list, proc, cpid1);CHKERRQ(ierr); /* cache to send messages */
ierr = PetscCDAppendID(deleted_list, proc, lid0);CHKERRQ(ierr);
} else {
continue;
}
lid_matched[lid0] = PETSC_TRUE; /* keep track of what we've done this round */
/* set projection */
ierr = MatSetValues(P,1,&gid0,1,&gid0,&one,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(P,1,&gid1,1,&gid0,&one,INSERT_VALUES);CHKERRQ(ierr);
} /* matched */
} /* edge loop */
/* deal with deleted ghost on first pass */
if (size>1 && sub_it != n_sub_its-1){
PetscCDPos pos; PetscBool ise = PETSC_FALSE;
PetscInt nSend1, **sbuffs1,nSend2;
#define REQ_BF_SIZE 100
MPI_Request *sreqs2[REQ_BF_SIZE],*rreqs2[REQ_BF_SIZE];
MPI_Status status;
/* send request */
for (proc=0,nSend1=0;proc<size;proc++){
ierr = PetscCDEmptyAt(deleted_list,proc,&ise);CHKERRQ(ierr);
if (!ise) nSend1++;
}
ierr = PetscMalloc(nSend1*sizeof(PetscInt*), &sbuffs1);CHKERRQ(ierr);
/* ierr = PetscMalloc4(nSend1, PetscInt*, sbuffs1, nSend1, PetscInt*, rbuffs1, nSend1, MPI_Request*, sreqs1, nSend1, MPI_Request*, rreqs1);CHKERRQ(ierr); */
/* PetscFree4(sbuffs1,rbuffs1,sreqs1,rreqs1); */
for (proc=0,nSend1=0;proc<size;proc++){
/* count ghosts */
ierr = PetscCDSizeAt(deleted_list,proc,&n);CHKERRQ(ierr);
if (n>0){
#define CHUNCK_SIZE 100
PetscInt *sbuff,*pt;
MPI_Request *request;
assert(n%2==0);
n /= 2;
ierr = PetscMalloc((2 + 2*n + n*CHUNCK_SIZE)*sizeof(PetscInt) + 2*sizeof(MPI_Request), &sbuff);CHKERRQ(ierr);
/* PetscMalloc4(2+2*n,PetscInt,sbuffs1[nSend1],n*CHUNCK_SIZE,PetscInt,rbuffs1[nSend1],1,MPI_Request,rreqs2[nSend1],1,MPI_Request,sreqs2[nSend1]); */
/* save requests */
sbuffs1[nSend1] = sbuff;
request = (MPI_Request*)sbuff;
sbuff = pt = (PetscInt*)(request+1);
*pt++ = n; *pt++ = rank;
ierr = PetscCDGetHeadPos(deleted_list,proc,&pos);CHKERRQ(ierr);
while(pos){
PetscInt lid0, cpid, gid;
ierr = PetscLLNGetID(pos, &cpid);CHKERRQ(ierr);
gid = (PetscInt)PetscRealPart(cpcol_gid[cpid]);
ierr = PetscCDGetNextPos(deleted_list,proc,&pos);CHKERRQ(ierr);
ierr = PetscLLNGetID(pos, &lid0);CHKERRQ(ierr);
ierr = PetscCDGetNextPos(deleted_list,proc,&pos);CHKERRQ(ierr);
*pt++ = gid; *pt++ = lid0;
}
/* send request tag1 [n, proc, n*[gid1,lid0] ] */
ierr = MPI_Isend(sbuff, 2*n+2, MPIU_INT, proc, tag1, wcomm, request);CHKERRQ(ierr);
/* post recieve */
request = (MPI_Request*)pt;
rreqs2[nSend1] = request; /* cache recv request */
pt = (PetscInt*)(request+1);
ierr = MPI_Irecv(pt, n*CHUNCK_SIZE, MPIU_INT, proc, tag2, wcomm, request);CHKERRQ(ierr);
/* clear list */
ierr = PetscCDRemoveAll(deleted_list, proc);CHKERRQ(ierr);
nSend1++;
}
}
/* recieve requests, send response, clear lists */
kk = nactive_edges;
ierr = MPI_Allreduce(&kk,&nactive_edges,1,MPIU_INT,MPI_SUM,wcomm);CHKERRQ(ierr); /* not correct syncronization and global */
nSend2 = 0;
while(1){
#define BF_SZ 10000
PetscMPIInt flag,count;
PetscInt rbuff[BF_SZ],*pt,*pt2,*pt3,count2,*sbuff,count3;
MPI_Request *request;
ierr = MPI_Iprobe(MPI_ANY_SOURCE, tag1, wcomm, &flag, &status);CHKERRQ(ierr);
if (!flag) break;
ierr = MPI_Get_count(&status, MPIU_INT, &count);CHKERRQ(ierr);
if (count > BF_SZ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"buffer too small for recieve: %d",count);
proc = status.MPI_SOURCE;
/* recieve request tag1 [n, proc, n*[gid1,lid0] ] */
ierr = MPI_Recv(rbuff, count, MPIU_INT, proc, tag1, wcomm, &status);CHKERRQ(ierr);
/* count sends */
pt = rbuff; count3 = count2 = 0;
n = *pt++; kk = *pt++; assert(kk==proc);
while(n--){
PetscInt gid1=*pt++, lid1=gid1-my0; kk=*pt++; assert(lid1>=0 && lid1<nloc);
if (lid_matched[lid1]){
PetscPrintf(PETSC_COMM_SELF,"\t *** [%d]%s %d) ERROR recieved deleted gid %d, deleted by (lid) %d from proc %d\n",rank,__FUNCT__,sub_it,gid1,kk);
PetscSleep(1);
}
assert(!lid_matched[lid1]);
lid_matched[lid1] = PETSC_TRUE; /* keep track of what we've done this round */
ierr = PetscCDSizeAt(agg_llists, lid1, &kk);CHKERRQ(ierr);
count2 += kk + 2;
count3++; /* number of verts requested (n) */
}
assert(pt-rbuff==count);
if (count2 > count3*CHUNCK_SIZE) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Irecv will be too small: %d",count2);
/* send tag2 *[lid0, n, n*[gid] ] */
ierr = PetscMalloc(count2*sizeof(PetscInt) + sizeof(MPI_Request), &sbuff);CHKERRQ(ierr);
request = (MPI_Request*)sbuff;
sreqs2[nSend2++] = request; /* cache request */
if (nSend2==REQ_BF_SIZE) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"buffer too small for requests: %d",nSend2);
pt2 = sbuff = (PetscInt*)(request+1);
pt = rbuff;
n = *pt++; kk = *pt++; assert(kk==proc);
while(n--){
/* read [n, proc, n*[gid1,lid0] */
PetscInt gid1=*pt++, lid1=gid1-my0, lid0=*pt++; assert(lid1>=0 && lid1<nloc);
/* write [lid0, n, n*[gid] ] */
*pt2++ = lid0;
pt3 = pt2++; /* save pointer for later */
/* for (pos=PetscCDGetHeadPos(agg_llists,lid1) ; pos ; pos=PetscCDGetNextPos(agg_llists,lid1,pos)){ */
ierr = PetscCDGetHeadPos(agg_llists,lid1,&pos);CHKERRQ(ierr);
while(pos){
PetscInt gid;
ierr = PetscLLNGetID(pos, &gid);CHKERRQ(ierr);
ierr = PetscCDGetNextPos(agg_llists,lid1,&pos);CHKERRQ(ierr);
*pt2++ = gid;
}
*pt3 = (pt2-pt3)-1;
/* clear list */
ierr = PetscCDRemoveAll(agg_llists, lid1);CHKERRQ(ierr);
}
assert(pt2-sbuff==count2); assert(pt-rbuff==count);
/* send requested data tag2 *[lid0, n, n*[gid1] ] */
ierr = MPI_Isend(sbuff, count2, MPIU_INT, proc, tag2, wcomm, request);CHKERRQ(ierr);
}
/* recieve tag2 *[lid0, n, n*[gid] ] */
for (kk=0;kk<nSend1;kk++){
PetscMPIInt count;
MPI_Request *request;
PetscInt *pt, *pt2;
request = rreqs2[kk]; /* no need to free -- buffer is in 'sbuffs1' */
ierr = MPI_Wait(request, &status);CHKERRQ(ierr);
ierr = MPI_Get_count(&status, MPIU_INT, &count);CHKERRQ(ierr);
pt = pt2 = (PetscInt*)(request+1);
while(pt-pt2 < count){
PetscInt lid0 = *pt++, n = *pt++; assert(lid0>=0 && lid0<nloc);
while(n--){
PetscInt gid1 = *pt++;
ierr = PetscCDAppendID(agg_llists, lid0, gid1);CHKERRQ(ierr);
}
}
assert(pt-pt2==count);
}
/* wait for tag1 isends */
while(nSend1--){
MPI_Request *request;
request = (MPI_Request*)sbuffs1[nSend1];
ierr = MPI_Wait(request, &status);CHKERRQ(ierr);
ierr = PetscFree(request);CHKERRQ(ierr);
}
ierr = PetscFree(sbuffs1);CHKERRQ(ierr);
/* wait for tag2 isends */
while(nSend2--){
MPI_Request *request = sreqs2[nSend2];
ierr = MPI_Wait(request, &status);CHKERRQ(ierr);
ierr = PetscFree(request);CHKERRQ(ierr);
}
ierr = VecRestoreArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecRestoreArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
/* get 'cpcol_matched' - use locMaxPE, ghostMaxEdge, cpcol_max_ew */
for (kk=0,gid=my0;kk<nloc;kk++,gid++) {
PetscScalar vval = lid_matched[kk] ? 1.0 : 0.0;
ierr = VecSetValues(locMaxPE, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(locMaxPE);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxPE);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxPE,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxPE,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecGetLocalSize(mpimat->lvec, &n);CHKERRQ(ierr);
for (kk=0;kk<n;kk++) {
cpcol_matched[kk] = (PetscBool)(PetscRealPart(cpcol_max_ew[kk]) != 0.0);
}
ierr = VecRestoreArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
} /* size > 1 */
/* compute 'locMaxEdge' */
ierr = VecRestoreArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
PetscReal max_e = 0.,tt;
PetscScalar vval;
PetscInt lid = kk;
if (lid_matched[lid]) vval = 0.;
else {
ii = matA->i; n = ii[lid+1] - ii[lid]; idx = matA->j + ii[lid];
ap = matA->a + ii[lid];
for (jj=0; jj<n; jj++) {
PetscInt lidj = idx[jj];
if (lid_matched[lidj]) continue; /* this is new - can change local max */
if (lidj != lid && PetscRealPart(ap[jj]) > max_e) max_e = PetscRealPart(ap[jj]);
}
if (lid_cprowID && (ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n ; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
if ((tt=PetscRealPart(ap[jj])) > max_e) max_e = tt;
}
}
}
vval = (PetscScalar)max_e;
ierr = VecSetValues(locMaxEdge, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(locMaxEdge);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxEdge);CHKERRQ(ierr);
if (size>1 && sub_it != n_sub_its-1){
/* compute 'cpcol_max_ew' */
ierr = VecScatterBegin(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecGetArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
/* compute 'cpcol_max_pe' */
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
PetscInt lid = kk;
PetscReal ew,v1_max_e,v0_max_e=PetscRealPart(lid_max_ew[lid]);
PetscScalar vval;
PetscMPIInt max_pe=rank,pe;
if (lid_matched[lid]) vval = (PetscScalar)rank;
else if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n ; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
ew = PetscRealPart(ap[jj]); v1_max_e = PetscRealPart(cpcol_max_ew[lidj]);
/* get max pe that has a max_e == to this edge w */
if ((pe=cpcol_pe[idx[jj]]) > max_pe && ew > v1_max_e - 1.e-12 && ew > v0_max_e - 1.e-12) max_pe = pe;
assert(ew < v0_max_e + 1.e-12 && ew < v1_max_e + 1.e-12);
}
vval = (PetscScalar)max_pe;
}
ierr = VecSetValues(locMaxPE, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(locMaxPE);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxPE);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
ierr = VecRestoreArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
} /* deal with deleted ghost */
if (verbose>2) PetscPrintf(wcomm,"\t[%d]%s %d.%d: %d active edges.\n",
rank,__FUNCT__,iter,sub_it,nactive_edges);
if (!nactive_edges) break;
} /* sub_it loop */
/* clean up iteration */
ierr = PetscFree(Edges);CHKERRQ(ierr);
if (mpimat){
ierr = VecRestoreArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecDestroy(&ghostMaxEdge);CHKERRQ(ierr);
ierr = VecRestoreArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
ierr = VecDestroy(&ghostMaxPE);CHKERRQ(ierr);
ierr = PetscFree(cpcol_pe);CHKERRQ(ierr);
ierr = PetscFree(cpcol_matched);CHKERRQ(ierr);
}
ierr = VecDestroy(&locMaxEdge);CHKERRQ(ierr);
ierr = VecDestroy(&locMaxPE);CHKERRQ(ierr);
if (mpimat){
ierr = VecRestoreArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr);
}
/* create next G if needed */
if (iter == n_iter) { /* hard wired test - need to look at full surrounded nodes or something */
ierr = MatDestroy(&P);CHKERRQ(ierr);
ierr = MatDestroy(&cMat);CHKERRQ(ierr);
break;
} else {
Vec diag;
/* add identity for unmatched vertices so they stay alive */
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
if (!lid_matched[kk]) {
gid = kk+my0;
ierr = MatGetRow(cMat,gid,&n,0,0);CHKERRQ(ierr);
if (n>1){
ierr = MatSetValues(P,1,&gid,1,&gid,&one,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatRestoreRow(cMat,gid,&n,0,0);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(P,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(P,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* project to make new graph with colapsed edges */
ierr = MatPtAP(cMat,P,MAT_INITIAL_MATRIX,1.0,&tMat);CHKERRQ(ierr);
ierr = MatDestroy(&P);CHKERRQ(ierr);
ierr = MatDestroy(&cMat);CHKERRQ(ierr);
cMat = tMat;
ierr = MatGetVecs(cMat, &diag, 0);CHKERRQ(ierr);
ierr = MatGetDiagonal(cMat, diag);CHKERRQ(ierr); /* effectively PCJACOBI */
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecSqrtAbs(diag);CHKERRQ(ierr);
ierr = MatDiagonalScale(cMat, diag, diag);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
}
} /* coarsen iterator */
/* make fake matrix */
if (size>1){
Mat mat;
PetscCDPos pos;
PetscInt gid, NN, MM, jj = 0, mxsz = 0;
for (kk=0;kk<nloc;kk++){
ierr = PetscCDSizeAt(agg_llists, kk, &jj);CHKERRQ(ierr);
if (jj > mxsz) mxsz = jj;
}
ierr = MatGetSize(a_Gmat, &MM, &NN);CHKERRQ(ierr);
if (mxsz > MM-nloc) mxsz = MM-nloc;
ierr = MatCreateAIJ(wcomm, nloc, nloc,PETSC_DETERMINE, PETSC_DETERMINE,0, 0, mxsz, 0, &mat);CHKERRQ(ierr);
/* */
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
/* for (pos=PetscCDGetHeadPos(agg_llists,kk) ; pos ; pos=PetscCDGetNextPos(agg_llists,kk,pos)){ */
ierr = PetscCDGetHeadPos(agg_llists,kk,&pos);CHKERRQ(ierr);
while(pos){
PetscInt gid1;
ierr = PetscLLNGetID(pos, &gid1);CHKERRQ(ierr);
ierr = PetscCDGetNextPos(agg_llists,kk,&pos);CHKERRQ(ierr);
if (gid1 < my0 || gid1 >= my0+nloc) {
ierr = MatSetValues(mat,1,&gid,1,&gid1,&one,ADD_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscCDSetMat(agg_llists, mat);CHKERRQ(ierr);
}
ierr = PetscFree(lid_cprowID);CHKERRQ(ierr);
ierr = PetscFree(lid_gid);CHKERRQ(ierr);
ierr = PetscFree(lid_matched);CHKERRQ(ierr);
ierr = PetscCDDestroy(deleted_list);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
