11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 6674ae819SStefano Zampini 7*669cc0f4SStefano Zampini 8*669cc0f4SStefano Zampini #undef __FUNCT__ 9*669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 10*669cc0f4SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PCBDDCGraph graph, MatNullSpace *nnsp) 11*669cc0f4SStefano Zampini { 12*669cc0f4SStefano Zampini PetscErrorCode ierr; 13*669cc0f4SStefano Zampini ISLocalToGlobalMapping rmap; 14*669cc0f4SStefano Zampini IS *faces,*edges; 15*669cc0f4SStefano Zampini Vec p,v,quad_vec,*quad_vecs; 16*669cc0f4SStefano Zampini PetscScalar *vals; 17*669cc0f4SStefano Zampini const PetscScalar *array; 18*669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 19*669cc0f4SStefano Zampini 20*669cc0f4SStefano Zampini PetscFunctionBegin; 21*669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 22*669cc0f4SStefano Zampini if (graph->twodim) { 23*669cc0f4SStefano Zampini lmaxneighs = 2; 24*669cc0f4SStefano Zampini } else { 25*669cc0f4SStefano Zampini lmaxneighs = 1; 26*669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 27*669cc0f4SStefano Zampini const PetscInt *idxs; 28*669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 29*669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 30*669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 31*669cc0f4SStefano Zampini } 32*669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 33*669cc0f4SStefano Zampini } 34*669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 35*669cc0f4SStefano Zampini maxsize = 0; 36*669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 37*669cc0f4SStefano Zampini PetscInt nn; 38*669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 39*669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 40*669cc0f4SStefano Zampini } 41*669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 42*669cc0f4SStefano Zampini PetscInt nn; 43*669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 44*669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 45*669cc0f4SStefano Zampini } 46*669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 47*669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 48*669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 49*669cc0f4SStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 50*669cc0f4SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&rmap,NULL);CHKERRQ(ierr); 51*669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 52*669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 53*669cc0f4SStefano Zampini PetscInt first,last; 54*669cc0f4SStefano Zampini 55*669cc0f4SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],rmap);CHKERRQ(ierr); 56*669cc0f4SStefano Zampini /* the near-null space fo BDDC carries information on quadrature weights, 57*669cc0f4SStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 58*669cc0f4SStefano Zampini and create a suitable set of basis vectors first */ 59*669cc0f4SStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 60*669cc0f4SStefano Zampini if (i>=first && i < last) { 61*669cc0f4SStefano Zampini PetscScalar *data; 62*669cc0f4SStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 63*669cc0f4SStefano Zampini data[i-first] = 1.; 64*669cc0f4SStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 65*669cc0f4SStefano Zampini } 66*669cc0f4SStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 67*669cc0f4SStefano Zampini } 68*669cc0f4SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 69*669cc0f4SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 70*669cc0f4SStefano Zampini /* compute local quad vec */ 71*669cc0f4SStefano Zampini ierr = MatCreateVecs(divudotp,&v,&p);CHKERRQ(ierr); 72*669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 73*669cc0f4SStefano Zampini ierr = MatMultTranspose(divudotp,p,v);CHKERRQ(ierr); 74*669cc0f4SStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 75*669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 76*669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 77*669cc0f4SStefano Zampini const PetscInt *idxs; 78*669cc0f4SStefano Zampini PetscInt idx,nn,j; 79*669cc0f4SStefano Zampini 80*669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 81*669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 82*669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 83*669cc0f4SStefano Zampini ierr = PetscFindInt(PetscGlobalRank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 84*669cc0f4SStefano Zampini idx = -(idx+1); 85*669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 86*669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 87*669cc0f4SStefano Zampini ierr = ISDestroy(&faces[i]);CHKERRQ(ierr); 88*669cc0f4SStefano Zampini } 89*669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 90*669cc0f4SStefano Zampini const PetscInt *idxs; 91*669cc0f4SStefano Zampini PetscInt idx,nn,j; 92*669cc0f4SStefano Zampini 93*669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 94*669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 95*669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 96*669cc0f4SStefano Zampini ierr = PetscFindInt(PetscGlobalRank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 97*669cc0f4SStefano Zampini idx = -(idx+1); 98*669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 99*669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 100*669cc0f4SStefano Zampini ierr = ISDestroy(&edges[i]);CHKERRQ(ierr); 101*669cc0f4SStefano Zampini } 102*669cc0f4SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 103*669cc0f4SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 104*669cc0f4SStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 105*669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 106*669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 107*669cc0f4SStefano Zampini 108*669cc0f4SStefano Zampini /* assemble near null space */ 109*669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 110*669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 111*669cc0f4SStefano Zampini } 112*669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 113*669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 114*669cc0f4SStefano Zampini ierr = VecNormalize(quad_vecs[i],NULL);CHKERRQ(ierr); 115*669cc0f4SStefano Zampini } 116*669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 117*669cc0f4SStefano Zampini PetscFunctionReturn(0); 118*669cc0f4SStefano Zampini } 119*669cc0f4SStefano Zampini 120*669cc0f4SStefano Zampini 121a3df083aSStefano Zampini #undef __FUNCT__ 1221f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 1231f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 1241f4df5f7SStefano Zampini { 1251f4df5f7SStefano Zampini PetscErrorCode ierr; 1261f4df5f7SStefano Zampini Vec local,global; 1271f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1281f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1291f4df5f7SStefano Zampini 1301f4df5f7SStefano Zampini PetscFunctionBegin; 1311f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 1321f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 1331f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 1341f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 1351f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 1361f4df5f7SStefano Zampini PetscInt i; 1371f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1381f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 1391f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1401f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 1411f4df5f7SStefano Zampini } 1421f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 1431f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 1441f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 1451f4df5f7SStefano Zampini } 1461f4df5f7SStefano Zampini } else { 147986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 1481f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 149986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 150986cdee1SStefano Zampini if (i > 1) { 151986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 1521f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1531f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 1541f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1551f4df5f7SStefano Zampini } 1561f4df5f7SStefano Zampini } 1571f4df5f7SStefano Zampini } 158986cdee1SStefano Zampini } 1591f4df5f7SStefano Zampini 1601f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 1611f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 1621f4df5f7SStefano Zampini } 1631f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 1641f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1651f4df5f7SStefano Zampini } 1661f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 1671f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1681f4df5f7SStefano Zampini } 1691f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 1701f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 1711f4df5f7SStefano Zampini PetscFunctionReturn(0); 1721f4df5f7SStefano Zampini } 1731f4df5f7SStefano Zampini 1741f4df5f7SStefano Zampini #undef __FUNCT__ 1753e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 1763e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 1773e589ea0SStefano Zampini { 1783e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 1793e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 1803e589ea0SStefano Zampini PetscErrorCode ierr; 1813e589ea0SStefano Zampini 1823e589ea0SStefano Zampini PetscFunctionBegin; 1833e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 1843e589ea0SStefano Zampini PetscFunctionReturn(0); 1853e589ea0SStefano Zampini } 1863e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1873e589ea0SStefano Zampini Vec swap; 1883e589ea0SStefano Zampini 1893e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 1903e589ea0SStefano Zampini swap = pcbddc->work_change; 1913e589ea0SStefano Zampini pcbddc->work_change = r; 1923e589ea0SStefano Zampini r = swap; 1933e589ea0SStefano Zampini } 1943e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1953e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1963e589ea0SStefano Zampini if (pcbddc->benign_n) { 1973e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 1983e589ea0SStefano Zampini } else { 1993e589ea0SStefano Zampini ierr = VecSet(pcis->vec2_D,0.);CHKERRQ(ierr); 2003e589ea0SStefano Zampini } 2013e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 2023e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2033e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2043e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2053e589ea0SStefano Zampini Vec swap; 2063e589ea0SStefano Zampini 2073e589ea0SStefano Zampini swap = r; 2083e589ea0SStefano Zampini r = pcbddc->work_change; 2093e589ea0SStefano Zampini pcbddc->work_change = swap; 2103e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 2113e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 2123e589ea0SStefano Zampini } 2133e589ea0SStefano Zampini PetscFunctionReturn(0); 2143e589ea0SStefano Zampini } 2153e589ea0SStefano Zampini 2163e589ea0SStefano Zampini #undef __FUNCT__ 217a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 218a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 219a3df083aSStefano Zampini { 220a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 221a3df083aSStefano Zampini PetscErrorCode ierr; 222a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 223a3df083aSStefano Zampini 224a3df083aSStefano Zampini PetscFunctionBegin; 225a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 226a3df083aSStefano Zampini if (transpose) { 227a3df083aSStefano Zampini apply_right = ctx->apply_left; 228a3df083aSStefano Zampini apply_left = ctx->apply_right; 229a3df083aSStefano Zampini } else { 230a3df083aSStefano Zampini apply_right = ctx->apply_right; 231a3df083aSStefano Zampini apply_left = ctx->apply_left; 232a3df083aSStefano Zampini } 233a3df083aSStefano Zampini reset_x = PETSC_FALSE; 234a3df083aSStefano Zampini if (apply_right) { 235a3df083aSStefano Zampini const PetscScalar *ax; 236a3df083aSStefano Zampini PetscInt nl,i; 237a3df083aSStefano Zampini 238a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 239a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 240a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 241a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 242a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 243a3df083aSStefano Zampini PetscScalar sum,val; 244a3df083aSStefano Zampini const PetscInt *idxs; 245a3df083aSStefano Zampini PetscInt nz,j; 246a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 247a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 248a3df083aSStefano Zampini sum = 0.; 249a3df083aSStefano Zampini if (ctx->apply_p0) { 250a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 251a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 252a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 253a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 254a3df083aSStefano Zampini } 255a3df083aSStefano Zampini } else { 256a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 257a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 258a3df083aSStefano Zampini } 259a3df083aSStefano Zampini } 260a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 261a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 262a3df083aSStefano Zampini } 263a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 264a3df083aSStefano Zampini reset_x = PETSC_TRUE; 265a3df083aSStefano Zampini } 266a3df083aSStefano Zampini if (transpose) { 267a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 268a3df083aSStefano Zampini } else { 269a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 270a3df083aSStefano Zampini } 271a3df083aSStefano Zampini if (reset_x) { 272a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 273a3df083aSStefano Zampini } 274a3df083aSStefano Zampini if (apply_left) { 275a3df083aSStefano Zampini PetscScalar *ay; 276a3df083aSStefano Zampini PetscInt i; 277a3df083aSStefano Zampini 278a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 279a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 280a3df083aSStefano Zampini PetscScalar sum,val; 281a3df083aSStefano Zampini const PetscInt *idxs; 282a3df083aSStefano Zampini PetscInt nz,j; 283a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 284a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 285a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 286a3df083aSStefano Zampini if (ctx->apply_p0) { 287a3df083aSStefano Zampini sum = 0.; 288a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 289a3df083aSStefano Zampini sum += ay[idxs[j]]; 290a3df083aSStefano Zampini ay[idxs[j]] += val; 291a3df083aSStefano Zampini } 292a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 293a3df083aSStefano Zampini } else { 294a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 295a3df083aSStefano Zampini ay[idxs[j]] += val; 296a3df083aSStefano Zampini } 297a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 298a3df083aSStefano Zampini } 299a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 300a3df083aSStefano Zampini } 301a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 302a3df083aSStefano Zampini } 303a3df083aSStefano Zampini PetscFunctionReturn(0); 304a3df083aSStefano Zampini } 305a3df083aSStefano Zampini 306a3df083aSStefano Zampini #undef __FUNCT__ 307a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 308a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 309a3df083aSStefano Zampini { 310a3df083aSStefano Zampini PetscErrorCode ierr; 311a3df083aSStefano Zampini 312a3df083aSStefano Zampini PetscFunctionBegin; 313a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 314a3df083aSStefano Zampini PetscFunctionReturn(0); 315a3df083aSStefano Zampini } 316a3df083aSStefano Zampini 317a3df083aSStefano Zampini #undef __FUNCT__ 318a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 319a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 320a3df083aSStefano Zampini { 321a3df083aSStefano Zampini PetscErrorCode ierr; 322a3df083aSStefano Zampini 323a3df083aSStefano Zampini PetscFunctionBegin; 324a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 325a3df083aSStefano Zampini PetscFunctionReturn(0); 326a3df083aSStefano Zampini } 327a3df083aSStefano Zampini 328a3df083aSStefano Zampini #undef __FUNCT__ 329a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 330a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 331a3df083aSStefano Zampini { 332a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 333a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 334a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 335a3df083aSStefano Zampini PetscErrorCode ierr; 336a3df083aSStefano Zampini 337a3df083aSStefano Zampini PetscFunctionBegin; 338a3df083aSStefano Zampini if (!restore) { 3391dd7afcfSStefano Zampini Mat A_IB,A_BI; 340a3df083aSStefano Zampini PetscScalar *work; 341b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 342a3df083aSStefano Zampini 3431dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 3441dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 3451dd7afcfSStefano Zampini } 3461dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 3471dd7afcfSStefano Zampini PetscFunctionReturn(0); 3481dd7afcfSStefano Zampini } 349a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 350a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 351a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 352a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 353a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 354a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 355a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 356a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 357a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 358a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 359a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 360a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 361059032f7SStefano Zampini if (reuse) { 362a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 3631dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 364059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 365059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 366059032f7SStefano Zampini PetscInt i; 367059032f7SStefano Zampini 368059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 369059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 370059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 371059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 372059032f7SStefano Zampini } 373059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 3741dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 375059032f7SStefano Zampini } 376a3df083aSStefano Zampini ctx->A = pcis->A_IB; 377a3df083aSStefano Zampini ctx->work = work; 378a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 379a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 380a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 381a3df083aSStefano Zampini pcis->A_IB = A_IB; 382a3df083aSStefano Zampini 383a3df083aSStefano Zampini /* A_BI as A_IB^T */ 384a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 385a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 386a3df083aSStefano Zampini pcis->A_BI = A_BI; 387a3df083aSStefano Zampini } else { 3881dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 3891dd7afcfSStefano Zampini PetscFunctionReturn(0); 3901dd7afcfSStefano Zampini } 391a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 392a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 393a3df083aSStefano Zampini pcis->A_IB = ctx->A; 3941dd7afcfSStefano Zampini ctx->A = NULL; 3951dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 3961dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 3971dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 3981dd7afcfSStefano Zampini if (ctx->free) { 399059032f7SStefano Zampini PetscInt i; 4001dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 401059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 402059032f7SStefano Zampini } 403059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 404059032f7SStefano Zampini } 405a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 406a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 407a3df083aSStefano Zampini } 408a3df083aSStefano Zampini PetscFunctionReturn(0); 409a3df083aSStefano Zampini } 410a3df083aSStefano Zampini 411a3df083aSStefano Zampini /* used just in bddc debug mode */ 412a3df083aSStefano Zampini #undef __FUNCT__ 413a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 414a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 415a3df083aSStefano Zampini { 416a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 417a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 418a3df083aSStefano Zampini Mat An; 419a3df083aSStefano Zampini PetscErrorCode ierr; 420a3df083aSStefano Zampini 421a3df083aSStefano Zampini PetscFunctionBegin; 422a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 423a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 424a3df083aSStefano Zampini if (is1) { 425a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 426a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 427a3df083aSStefano Zampini } else { 428a3df083aSStefano Zampini *B = An; 429a3df083aSStefano Zampini } 430a3df083aSStefano Zampini PetscFunctionReturn(0); 431a3df083aSStefano Zampini } 432a3df083aSStefano Zampini 4331cf9b237SStefano Zampini /* TODO: add reuse flag */ 4341cf9b237SStefano Zampini #undef __FUNCT__ 4351cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 4361cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 4371cf9b237SStefano Zampini { 4381cf9b237SStefano Zampini Mat Bt; 4391cf9b237SStefano Zampini PetscScalar *a,*bdata; 4401cf9b237SStefano Zampini const PetscInt *ii,*ij; 4411cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 4421cf9b237SStefano Zampini PetscBool flg_row; 4431cf9b237SStefano Zampini PetscErrorCode ierr; 4441cf9b237SStefano Zampini 4451cf9b237SStefano Zampini PetscFunctionBegin; 4461cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 4471cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4481cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 4491cf9b237SStefano Zampini nnz = n; 4501cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 4511cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 4521cf9b237SStefano Zampini } 4531cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 4541cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 4551cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 4561cf9b237SStefano Zampini nnz = 0; 4571cf9b237SStefano Zampini bii[0] = 0; 4581cf9b237SStefano Zampini for (i=0;i<n;i++) { 4591cf9b237SStefano Zampini PetscInt j; 4601cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 4611cf9b237SStefano Zampini PetscScalar entry = a[j]; 4621cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 4631cf9b237SStefano Zampini bij[nnz] = ij[j]; 4641cf9b237SStefano Zampini bdata[nnz] = entry; 4651cf9b237SStefano Zampini nnz++; 4661cf9b237SStefano Zampini } 4671cf9b237SStefano Zampini } 4681cf9b237SStefano Zampini bii[i+1] = nnz; 4691cf9b237SStefano Zampini } 4701cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 4711cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 4721cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4731cf9b237SStefano Zampini { 4741cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 4751cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 4761cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 4771cf9b237SStefano Zampini } 4781cf9b237SStefano Zampini *B = Bt; 4791cf9b237SStefano Zampini PetscFunctionReturn(0); 4801cf9b237SStefano Zampini } 4811cf9b237SStefano Zampini 482674ae819SStefano Zampini #undef __FUNCT__ 4834f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 4844f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 4854f1b2e48SStefano Zampini { 4864f1b2e48SStefano Zampini Mat B; 4874f1b2e48SStefano Zampini IS is_dummy,*cc_n; 4884f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 4894f1b2e48SStefano Zampini PCBDDCGraph graph; 4904f1b2e48SStefano Zampini PetscInt i,n; 4914f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 4924f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 4934f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 4944f1b2e48SStefano Zampini PetscErrorCode ierr; 4954f1b2e48SStefano Zampini 4964f1b2e48SStefano Zampini PetscFunctionBegin; 49763c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 49863c961adSStefano Zampini *ncc = 0; 49963c961adSStefano Zampini *cc = NULL; 50063c961adSStefano Zampini PetscFunctionReturn(0); 50163c961adSStefano Zampini } 5024f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 5034f1b2e48SStefano Zampini if (!isseqaij && filter) { 5041cf9b237SStefano Zampini PetscBool isseqdense; 5051cf9b237SStefano Zampini 5061cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 5071cf9b237SStefano Zampini if (!isseqdense) { 5084f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 5091cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 5101cf9b237SStefano Zampini PetscScalar *array; 5111cf9b237SStefano Zampini PetscReal chop=1.e-6; 5121cf9b237SStefano Zampini 5131cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 5141cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 5151cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 5161cf9b237SStefano Zampini for (i=0;i<n;i++) { 5171cf9b237SStefano Zampini PetscInt j; 5181cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 5191cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 5201cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 5211cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 5221cf9b237SStefano Zampini } 5231cf9b237SStefano Zampini } 5241cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 5259d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 5261cf9b237SStefano Zampini } 5274f1b2e48SStefano Zampini } else { 5284f1b2e48SStefano Zampini B = A; 5294f1b2e48SStefano Zampini } 5304f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5314f1b2e48SStefano Zampini 5324f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 5334f1b2e48SStefano Zampini if (filter) { 5344f1b2e48SStefano Zampini PetscScalar *data; 5354f1b2e48SStefano Zampini PetscInt j,cum; 5364f1b2e48SStefano Zampini 5374f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 5384f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 5394f1b2e48SStefano Zampini cum = 0; 5404f1b2e48SStefano Zampini for (i=0;i<n;i++) { 5414f1b2e48SStefano Zampini PetscInt t; 5424f1b2e48SStefano Zampini 5434f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 5444f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 5454f1b2e48SStefano Zampini continue; 5464f1b2e48SStefano Zampini } 5474f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 5484f1b2e48SStefano Zampini } 5494f1b2e48SStefano Zampini t = xadj_filtered[i]; 5504f1b2e48SStefano Zampini xadj_filtered[i] = cum; 5514f1b2e48SStefano Zampini cum += t; 5524f1b2e48SStefano Zampini } 5534f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 5544f1b2e48SStefano Zampini } else { 5554f1b2e48SStefano Zampini xadj_filtered = NULL; 5564f1b2e48SStefano Zampini adjncy_filtered = NULL; 5574f1b2e48SStefano Zampini } 5584f1b2e48SStefano Zampini 5594f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 5604f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 5614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 5634f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 5644f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 5654f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 5664f1b2e48SStefano Zampini if (xadj_filtered) { 5674f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 5684f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 5694f1b2e48SStefano Zampini } else { 5704f1b2e48SStefano Zampini graph->xadj = xadj; 5714f1b2e48SStefano Zampini graph->adjncy = adjncy; 5724f1b2e48SStefano Zampini } 5734f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5754f1b2e48SStefano Zampini /* partial clean up */ 5764f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 5774f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5781cf9b237SStefano Zampini if (A != B) { 5794f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini 5824f1b2e48SStefano Zampini /* get back data */ 5831cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 5841cf9b237SStefano Zampini if (cc) { 5854f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 5864f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 5874f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini } 5894f1b2e48SStefano Zampini *cc = cc_n; 5901cf9b237SStefano Zampini } 5914f1b2e48SStefano Zampini /* clean up graph */ 5924f1b2e48SStefano Zampini graph->xadj = 0; 5934f1b2e48SStefano Zampini graph->adjncy = 0; 5944f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5954f1b2e48SStefano Zampini PetscFunctionReturn(0); 5964f1b2e48SStefano Zampini } 5974f1b2e48SStefano Zampini 5984f1b2e48SStefano Zampini #undef __FUNCT__ 5995408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 6005408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 6015408967cSStefano Zampini { 6025408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6035408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 604dee84bffSStefano Zampini IS dirIS = NULL; 6054f1b2e48SStefano Zampini PetscInt i; 6065408967cSStefano Zampini PetscErrorCode ierr; 6075408967cSStefano Zampini 6085408967cSStefano Zampini PetscFunctionBegin; 609dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 6105408967cSStefano Zampini if (zerodiag) { 6115408967cSStefano Zampini Mat A; 6125408967cSStefano Zampini Vec vec3_N; 6135408967cSStefano Zampini PetscScalar *vals; 6145408967cSStefano Zampini const PetscInt *idxs; 615d12d3064SStefano Zampini PetscInt nz,*count; 6165408967cSStefano Zampini 6175408967cSStefano Zampini /* p0 */ 6185408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 6195408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 6205408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6215408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6224f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 6235408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6245408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6255408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6265408967cSStefano Zampini /* v_I */ 6275408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 6285408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 6295408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6305408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 6315408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6325408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 6335408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6345408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6355408967cSStefano Zampini if (dirIS) { 6365408967cSStefano Zampini PetscInt n; 6375408967cSStefano Zampini 6385408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 6395408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 6405408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 6415408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6425408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 6435408967cSStefano Zampini } 6445408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 6455408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 6465408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 6475408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 648*669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 6495408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 6505408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 651fbfcb133SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) { 652b9b0e38cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)",PetscAbsScalar(vals[0])); 6535408967cSStefano Zampini } 6545408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6555408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 656d12d3064SStefano Zampini 657d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 658d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 659d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 660d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 661d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 662d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 663d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 664d12d3064SStefano Zampini if (count[idxs[i]]) { 665d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 666d12d3064SStefano Zampini } 667d4d8cf7bSStefano Zampini } 668d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 669d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 6705408967cSStefano Zampini } 671dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 6725408967cSStefano Zampini 6735408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 6745408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 6754f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 6765408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 6774f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 6785408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 6794f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6804f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 6814f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 6824f1b2e48SStefano Zampini } 6835408967cSStefano Zampini } 6845408967cSStefano Zampini PetscFunctionReturn(0); 6855408967cSStefano Zampini } 6865408967cSStefano Zampini 6875408967cSStefano Zampini #undef __FUNCT__ 688339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 689339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 690339f8db1SStefano Zampini { 691339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6924f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 693b0f5fe93SStefano Zampini PetscInt nz,n; 6941f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 6954f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 696339f8db1SStefano Zampini PetscErrorCode ierr; 697339f8db1SStefano Zampini 698339f8db1SStefano Zampini PetscFunctionBegin; 6999f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 7009f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 701a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 702a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 703a3df083aSStefano Zampini } 704a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 705a3df083aSStefano Zampini pcbddc->benign_n = 0; 7064f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 7074f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 7084f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 7094f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 7104f1b2e48SStefano Zampini since the local Schur complements are SPD 7114f1b2e48SStefano Zampini */ 7124f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 7134f1b2e48SStefano Zampini have_null = PETSC_TRUE; 71440fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 7154f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 7164f1b2e48SStefano Zampini 7174f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 7184f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 7194f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 7204f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 721ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 72240fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 72340fa8d13SStefano Zampini if (!sorted) { 72440fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 72540fa8d13SStefano Zampini } 72640fa8d13SStefano Zampini } else { 72740fa8d13SStefano Zampini pressures = NULL; 72840fa8d13SStefano Zampini } 72997d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 73097d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 73127b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 73297d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 733339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 734339f8db1SStefano Zampini if (!sorted) { 735339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 736339f8db1SStefano Zampini } 737339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 7384f1b2e48SStefano Zampini if (!nz) { 7394f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 7404f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 74140fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 74240fa8d13SStefano Zampini } 7434f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 7444f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 7454f1b2e48SStefano Zampini zerodiag_subs = NULL; 7464f1b2e48SStefano Zampini pcbddc->benign_n = 0; 7471f4df5f7SStefano Zampini n_interior_dofs = 0; 7481f4df5f7SStefano Zampini interior_dofs = NULL; 7491f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 7501f4df5f7SStefano Zampini PetscInt n,i,j; 7511f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 7521f4df5f7SStefano Zampini PetscInt *iwork; 7531f4df5f7SStefano Zampini 7541f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 7551f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7561f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 7571f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 7581f4df5f7SStefano Zampini for (i=0;i<n_neigh;i++) 7591f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 7601f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 7611f4df5f7SStefano Zampini for (i=0;i<n;i++) 7621f4df5f7SStefano Zampini if (!iwork[i]) 7631f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 7641f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 7651f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7661f4df5f7SStefano Zampini } 7674f1b2e48SStefano Zampini if (has_null_pressures) { 7684f1b2e48SStefano Zampini IS *subs; 7691f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 7701f4df5f7SStefano Zampini const PetscInt *idxs; 7711f4df5f7SStefano Zampini PetscScalar *array; 7721f4df5f7SStefano Zampini Vec *work; 7731f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 7744f1b2e48SStefano Zampini 7754f1b2e48SStefano Zampini subs = pcbddc->local_subs; 7764f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 7771f4df5f7SStefano Zampini /* these vectors are needed to check if the constant on pressures is in the kernel of the local operator B (i.e. B(v_I,p0) should be zero) */ 7781f4df5f7SStefano Zampini if (pcbddc->current_level) { 7791f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 7801f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 7811f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 7821f4df5f7SStefano Zampini /* work[0] = 1_p */ 7831f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 7841f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 7851f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 7861f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 7871f4df5f7SStefano Zampini /* work[0] = 1_v */ 7881f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 7891f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 7901f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 7911f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 7921f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 7931f4df5f7SStefano Zampini } 7944f1b2e48SStefano Zampini if (nsubs > 1) { 7954f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 7964f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 7974f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 7984f1b2e48SStefano Zampini IS t_zerodiag_subs; 7994f1b2e48SStefano Zampini PetscInt nl; 8004f1b2e48SStefano Zampini 8014f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 8024f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 8034f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 8044f1b2e48SStefano Zampini if (nl) { 8054f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8064f1b2e48SStefano Zampini 8071f4df5f7SStefano Zampini if (pcbddc->current_level) { 8081f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 8091f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 8101f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 8111f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8121f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8131f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8141f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 8151f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 8161f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 8171f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 8181f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 8191f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8201f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8211f4df5f7SStefano Zampini valid = PETSC_FALSE; 8221f4df5f7SStefano Zampini break; 8231f4df5f7SStefano Zampini } 8241f4df5f7SStefano Zampini } 8251f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 8261f4df5f7SStefano Zampini } 8271f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 8281f4df5f7SStefano Zampini IS t_bc; 8291f4df5f7SStefano Zampini PetscInt nzb; 8301f4df5f7SStefano Zampini 8311f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 8321f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 8331f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 8341f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8351f4df5f7SStefano Zampini } 8361f4df5f7SStefano Zampini if (valid && pressures) { 8374f1b2e48SStefano Zampini IS t_pressure_subs; 8384f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 8394f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 8404f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 8414f1b2e48SStefano Zampini } 8424f1b2e48SStefano Zampini if (valid) { 8434f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 8444f1b2e48SStefano Zampini pcbddc->benign_n++; 8454f1b2e48SStefano Zampini } else { 8464f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 8474f1b2e48SStefano Zampini } 8484f1b2e48SStefano Zampini } 8494f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 8504f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 8514f1b2e48SStefano Zampini } 8524f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 8534f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8541f4df5f7SStefano Zampini 8551f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 8561f4df5f7SStefano Zampini PetscInt nzb; 8571f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 8581f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8591f4df5f7SStefano Zampini } 8601f4df5f7SStefano Zampini if (valid && pressures) { 8614f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 8624f1b2e48SStefano Zampini } 8631f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 8641f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 8651f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 8661f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8671f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8681f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8691f4df5f7SStefano Zampini valid = PETSC_FALSE; 8701f4df5f7SStefano Zampini break; 8711f4df5f7SStefano Zampini } 8721f4df5f7SStefano Zampini } 8731f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8741f4df5f7SStefano Zampini } 8754f1b2e48SStefano Zampini if (valid) { 8764f1b2e48SStefano Zampini pcbddc->benign_n = 1; 877ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 8784f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 8794f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 8804f1b2e48SStefano Zampini } 8814f1b2e48SStefano Zampini } 8821f4df5f7SStefano Zampini if (pcbddc->current_level) { 8831f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 8844f1b2e48SStefano Zampini } 8851f4df5f7SStefano Zampini } 8861f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 8874f1b2e48SStefano Zampini 8884f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 889b9b0e38cSStefano Zampini PetscInt n; 890b9b0e38cSStefano Zampini 8914f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 8924f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 893b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 894b9b0e38cSStefano Zampini if (n) { 8954f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 8964f1b2e48SStefano Zampini have_null = PETSC_FALSE; 8974f1b2e48SStefano Zampini } 898b9b0e38cSStefano Zampini } 8994f1b2e48SStefano Zampini 9004f1b2e48SStefano Zampini /* final check for null pressures */ 9014f1b2e48SStefano Zampini if (zerodiag && pressures) { 9024f1b2e48SStefano Zampini PetscInt nz,np; 9034f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 9044f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 9054f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 9064f1b2e48SStefano Zampini } 9074f1b2e48SStefano Zampini 9084f1b2e48SStefano Zampini if (recompute_zerodiag) { 9094f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9104f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 9114f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 9124f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 9134f1b2e48SStefano Zampini } else { 9144f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 9154f1b2e48SStefano Zampini 9164f1b2e48SStefano Zampini nzn = 0; 9174f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9184f1b2e48SStefano Zampini PetscInt ns; 9194f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9204f1b2e48SStefano Zampini nzn += ns; 9214f1b2e48SStefano Zampini } 9224f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 9234f1b2e48SStefano Zampini nzn = 0; 9244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9254f1b2e48SStefano Zampini PetscInt ns,*idxs; 9264f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9274f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9284f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 9294f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9304f1b2e48SStefano Zampini nzn += ns; 9314f1b2e48SStefano Zampini } 9324f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 9334f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 9344f1b2e48SStefano Zampini } 9354f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9364f1b2e48SStefano Zampini } 9374f1b2e48SStefano Zampini 938*669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 939*669cc0f4SStefano Zampini if (pcbddc->benign_compute_nonetflux && !pcbddc->divudotp) { 940*669cc0f4SStefano Zampini Mat A; 9411f4df5f7SStefano Zampini IS isused = NULL; 9421f4df5f7SStefano Zampini if (pressures) { 9431f4df5f7SStefano Zampini isused = pressures; 9441f4df5f7SStefano Zampini } else { 9451f4df5f7SStefano Zampini isused = zerodiag; 9461f4df5f7SStefano Zampini } 947*669cc0f4SStefano Zampini if (!isused) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Don't know how to extract div u dot p! Please provide the pressure field"); 948*669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 949*669cc0f4SStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&pcbddc->divudotp);CHKERRQ(ierr); 950b3afcdbeSStefano Zampini } 951b3afcdbeSStefano Zampini 952b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 9534f1b2e48SStefano Zampini if (has_null_pressures) { 9544f1b2e48SStefano Zampini IS zerodiagc; 9554f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 9564f1b2e48SStefano Zampini PetscInt i,s,*nnz; 9574f1b2e48SStefano Zampini 9584f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 959339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 960339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 961339f8db1SStefano Zampini /* local change of basis for pressures */ 962339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 96397d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 964339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 965339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 966339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 9674f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 9684f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9694f1b2e48SStefano Zampini PetscInt nzs,j; 9704f1b2e48SStefano Zampini 9714f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 9724f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 9734f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 9744f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 9754f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 9764f1b2e48SStefano Zampini } 977339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 978339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 979339f8db1SStefano Zampini /* set identity on velocities */ 980339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 981339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 982339f8db1SStefano Zampini } 9834f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 9844f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 9859f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 9864f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 987339f8db1SStefano Zampini /* set change on pressures */ 9884f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 9894f1b2e48SStefano Zampini PetscScalar *array; 9904f1b2e48SStefano Zampini PetscInt nzs; 9914f1b2e48SStefano Zampini 9924f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 9934f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 9944f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 995339f8db1SStefano Zampini PetscScalar vals[2]; 996339f8db1SStefano Zampini PetscInt cols[2]; 997339f8db1SStefano Zampini 998339f8db1SStefano Zampini cols[0] = idxs[i]; 9994f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 1000339f8db1SStefano Zampini vals[0] = 1.; 1001b0f5fe93SStefano Zampini vals[1] = 1.; 10024f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1003339f8db1SStefano Zampini } 10044f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 10054f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 10064f1b2e48SStefano Zampini array[nzs-1] = 1.; 10074f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 10084f1b2e48SStefano Zampini /* store local idxs for p0 */ 10094f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 10104f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 1011339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 10124f1b2e48SStefano Zampini } 1013339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1014339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1015a3df083aSStefano Zampini /* project if needed */ 1016a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10171dd7afcfSStefano Zampini Mat M; 10181dd7afcfSStefano Zampini 10191dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 1020339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 10211dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 10221dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 1023a3df083aSStefano Zampini } 10244f1b2e48SStefano Zampini /* store global idxs for p0 */ 10254f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1026339f8db1SStefano Zampini } 1027ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 10284f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 1029b0f5fe93SStefano Zampini 1030b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 1031b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 103227b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 103327b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1034339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 1035339f8db1SStefano Zampini PetscFunctionReturn(0); 1036339f8db1SStefano Zampini } 1037339f8db1SStefano Zampini 1038339f8db1SStefano Zampini #undef __FUNCT__ 1039015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 1040015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 1041efc2fbd9SStefano Zampini { 1042efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1043de9d7bd0SStefano Zampini PetscScalar *array; 1044efc2fbd9SStefano Zampini PetscErrorCode ierr; 1045efc2fbd9SStefano Zampini 1046efc2fbd9SStefano Zampini PetscFunctionBegin; 1047efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 1048efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 10494f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1050efc2fbd9SStefano Zampini } 1051de9d7bd0SStefano Zampini if (get) { 1052efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 10534f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 10544f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 1055efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 1056de9d7bd0SStefano Zampini } else { 1057de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 1058de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1059de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1060de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 1061efc2fbd9SStefano Zampini } 1062efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1063efc2fbd9SStefano Zampini } 1064efc2fbd9SStefano Zampini 1065efc2fbd9SStefano Zampini #undef __FUNCT__ 1066c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 1067c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 1068c263805aSStefano Zampini { 1069c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1070c263805aSStefano Zampini PetscErrorCode ierr; 1071c263805aSStefano Zampini 1072c263805aSStefano Zampini PetscFunctionBegin; 1073c263805aSStefano Zampini /* TODO: add error checking 1074c263805aSStefano Zampini - avoid nested pop (or push) calls. 1075c263805aSStefano Zampini - cannot push before pop. 10761c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 1077c263805aSStefano Zampini */ 10784f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1079efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1080efc2fbd9SStefano Zampini } 1081c263805aSStefano Zampini if (pop) { 1082a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10834f1b2e48SStefano Zampini IS is_p0; 10844f1b2e48SStefano Zampini MatReuse reuse; 1085c263805aSStefano Zampini 1086c263805aSStefano Zampini /* extract B_0 */ 10874f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 10884f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 10894f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 10904f1b2e48SStefano Zampini } 10914f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 10924f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 1093c263805aSStefano Zampini /* remove rows and cols from local problem */ 1094c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 109597d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 10964f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 10974f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 1098a3df083aSStefano Zampini } else { 1099a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1100a3df083aSStefano Zampini PetscScalar *vals; 1101a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 1102a3df083aSStefano Zampini 1103a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 1104a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 1105a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 11060b5adadeSStefano Zampini PetscInt *nnz; 1107a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 1108a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 1109a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1110331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 1111331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1112331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 1113331e053bSStefano Zampini nnz[i] = n - nnz[i]; 1114331e053bSStefano Zampini } 1115331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 1116331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1117331e053bSStefano Zampini } 1118a3df083aSStefano Zampini 1119a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1120a3df083aSStefano Zampini PetscScalar *array; 1121a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 1122a3df083aSStefano Zampini 1123a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 1124a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1125a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1126a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 1127a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1128a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 1129a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 1130a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 1131a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 1132a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 1133a3df083aSStefano Zampini cum = 0; 1134a3df083aSStefano Zampini for (j=0;j<n;j++) { 113522db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 1136a3df083aSStefano Zampini vals[cum] = array[j]; 1137a3df083aSStefano Zampini idxs_ins[cum] = j; 1138a3df083aSStefano Zampini cum++; 1139a3df083aSStefano Zampini } 1140a3df083aSStefano Zampini } 1141a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 1142a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 1143a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1144a3df083aSStefano Zampini } 1145a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1146a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1147a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 1148a3df083aSStefano Zampini } 1149c263805aSStefano Zampini } else { /* push */ 1150a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 11514f1b2e48SStefano Zampini PetscInt i; 11524f1b2e48SStefano Zampini 11534f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 11544f1b2e48SStefano Zampini PetscScalar *B0_vals; 11554f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 11564f1b2e48SStefano Zampini 11574f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 11584f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 11597b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 11604f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 11614f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 11624f1b2e48SStefano Zampini } 1163c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1164c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1165a3df083aSStefano Zampini } else { 1166a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 1167a3df083aSStefano Zampini } 1168c263805aSStefano Zampini } 1169c263805aSStefano Zampini PetscFunctionReturn(0); 1170c263805aSStefano Zampini } 1171c263805aSStefano Zampini 1172c263805aSStefano Zampini #undef __FUNCT__ 1173b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 117408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 1175b1b3d7a2SStefano Zampini { 1176b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 117708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 117808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 117908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 118008122e43SStefano Zampini PetscScalar *work,lwork; 118108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 118208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 118308122e43SStefano Zampini PetscReal *eigs,thresh; 11841b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 1185f6f667cfSStefano Zampini PetscBool allocated_S_St; 118608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 118708122e43SStefano Zampini PetscReal *rwork; 118808122e43SStefano Zampini #endif 1189b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1190b1b3d7a2SStefano Zampini 1191b1b3d7a2SStefano Zampini PetscFunctionBegin; 1192b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 1193af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 1194af25d912SStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 119506a4e24aSStefano Zampini 1196fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1197fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1198fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1199fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 12001575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1201fd14bc51SStefano Zampini } 1202fd14bc51SStefano Zampini 1203e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 1204e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 1205e496cd5dSStefano Zampini } 1206e496cd5dSStefano Zampini 120708122e43SStefano Zampini /* max size of subsets */ 120808122e43SStefano Zampini mss = 0; 120908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 121008122e43SStefano Zampini PetscInt subset_size; 1211862806e4SStefano Zampini 121208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 121308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 121408122e43SStefano Zampini } 121508122e43SStefano Zampini 121608122e43SStefano Zampini /* min/max and threshold */ 121708122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 1218f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 121908122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 1220f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 1221f6f667cfSStefano Zampini if (nmin) { 1222f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 1223f6f667cfSStefano Zampini } 122408122e43SStefano Zampini 122508122e43SStefano Zampini /* allocate lapack workspace */ 122608122e43SStefano Zampini cum = cum2 = 0; 122708122e43SStefano Zampini maxneigs = 0; 122808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 122908122e43SStefano Zampini PetscInt n,subset_size; 1230f6f667cfSStefano Zampini 123108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 123208122e43SStefano Zampini n = PetscMin(subset_size,nmax); 12339162d606SStefano Zampini cum += subset_size; 12349162d606SStefano Zampini cum2 += subset_size*n; 123508122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 123608122e43SStefano Zampini } 123708122e43SStefano Zampini if (mss) { 12389ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 123908122e43SStefano Zampini PetscBLASInt B_itype = 1; 124008122e43SStefano Zampini PetscBLASInt B_N = mss; 12414c6709b3SStefano Zampini PetscReal zero = 0.0; 12424c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 124308122e43SStefano Zampini 124408122e43SStefano Zampini B_lwork = -1; 124508122e43SStefano Zampini S = NULL; 124608122e43SStefano Zampini St = NULL; 1247a58a30b4SStefano Zampini eigs = NULL; 1248a58a30b4SStefano Zampini eigv = NULL; 1249a58a30b4SStefano Zampini B_iwork = NULL; 1250a58a30b4SStefano Zampini B_ifail = NULL; 1251d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1252d1710679SStefano Zampini rwork = NULL; 1253d1710679SStefano Zampini #endif 12548bec7fa6SStefano Zampini thresh = 1.0; 125508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 125608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 125708122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 125808122e43SStefano Zampini #else 125908122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 126008122e43SStefano Zampini #endif 126108122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 126208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 126308122e43SStefano Zampini } else { 126408122e43SStefano Zampini /* TODO */ 126508122e43SStefano Zampini } 126608122e43SStefano Zampini } else { 126708122e43SStefano Zampini lwork = 0; 126808122e43SStefano Zampini } 126908122e43SStefano Zampini 127008122e43SStefano Zampini nv = 0; 1271d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 1272d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 127308122e43SStefano Zampini } 12744c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1275f6f667cfSStefano Zampini if (allocated_S_St) { 1276f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1277f6f667cfSStefano Zampini } 1278f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 127908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 128008122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 128108122e43SStefano Zampini #endif 12829162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 12839162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 12849162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 128508122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 12869162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 128708122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 128808122e43SStefano Zampini 128908122e43SStefano Zampini maxneigs = 0; 129072b8c272SStefano Zampini cum = cumarray = 0; 12919162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 12929162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1293d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 129408122e43SStefano Zampini const PetscInt *idxs; 129508122e43SStefano Zampini 1296d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 129708122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 129808122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 129908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 130008122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 13019162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 13029162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 130308122e43SStefano Zampini } 1304d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 130508122e43SStefano Zampini } 130608122e43SStefano Zampini 130708122e43SStefano Zampini if (mss) { /* multilevel */ 130808122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 130908122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 131008122e43SStefano Zampini } 131108122e43SStefano Zampini 1312ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 131308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 131408122e43SStefano Zampini const PetscInt *idxs; 13159d54b7f4SStefano Zampini PetscReal upper,lower; 1316862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 131708122e43SStefano Zampini PetscBLASInt B_N; 1318aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 131908122e43SStefano Zampini 13209d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 13219d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 13229d54b7f4SStefano Zampini lower = thresh; 13239d54b7f4SStefano Zampini } else { 13249d54b7f4SStefano Zampini upper = 1./thresh; 13259d54b7f4SStefano Zampini lower = 0.; 13269d54b7f4SStefano Zampini } 1327862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1328ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1329f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1330f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 13319ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1332aff50787SStefano Zampini PetscInt j,k; 1333aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1334aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1335aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 133608122e43SStefano Zampini } 133708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1338aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1339aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1340aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1341aff50787SStefano Zampini } 134208122e43SStefano Zampini } 134308122e43SStefano Zampini } else { 134408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 134508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 134608122e43SStefano Zampini } 13478bec7fa6SStefano Zampini } else { 1348f6f667cfSStefano Zampini S = Sarray + cumarray; 1349f6f667cfSStefano Zampini St = Starray + cumarray; 13508bec7fa6SStefano Zampini } 1351aff50787SStefano Zampini /* see if we can save some work */ 1352b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1353aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1354aff50787SStefano Zampini } 1355aff50787SStefano Zampini 1356b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1357aff50787SStefano Zampini B_neigs = 0; 1358aff50787SStefano Zampini } else { 13599ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 136008122e43SStefano Zampini PetscBLASInt B_itype = 1; 1361f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 13624c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 13639552c7c7SStefano Zampini PetscInt nmin_s; 1364b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 136508122e43SStefano Zampini 1366fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 13678bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 1368fd14bc51SStefano Zampini } 1369d16cbb6bSStefano Zampini 1370b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1371b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1372b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1373b7ab4a40SStefano Zampini } 1374b7ab4a40SStefano Zampini 137508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1376b7ab4a40SStefano Zampini if (compute_range) { 1377d16cbb6bSStefano Zampini 1378d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 137908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 13809d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 138108122e43SStefano Zampini #else 13829d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 138308122e43SStefano Zampini #endif 1384b7ab4a40SStefano Zampini } else if (!same_data) { 1385d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1386d16cbb6bSStefano Zampini B_IL = 1; 1387d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 13889d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1389d16cbb6bSStefano Zampini #else 13909d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1391d16cbb6bSStefano Zampini #endif 1392b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1393b7ab4a40SStefano Zampini PetscInt k; 1394b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1395b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1396b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1397b7ab4a40SStefano Zampini nmin = nmax; 1398b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1399b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1400b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1401b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1402b7ab4a40SStefano Zampini } 1403d16cbb6bSStefano Zampini } 140408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 140508122e43SStefano Zampini if (B_ierr) { 14066c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 14076c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 14086c4ed002SBarry Smith else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 140908122e43SStefano Zampini } 141008122e43SStefano Zampini 141108122e43SStefano Zampini if (B_neigs > nmax) { 1412fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1413fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1414fd14bc51SStefano Zampini } 14159d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 141608122e43SStefano Zampini B_neigs = nmax; 141708122e43SStefano Zampini } 141808122e43SStefano Zampini 14199552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 14209552c7c7SStefano Zampini if (B_neigs < nmin_s) { 142108122e43SStefano Zampini PetscBLASInt B_neigs2; 142208122e43SStefano Zampini 14239d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1424f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 14259d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 14269d54b7f4SStefano Zampini } else { 14279d54b7f4SStefano Zampini B_IL = B_neigs + 1; 14289d54b7f4SStefano Zampini B_IU = nmin_s; 14299d54b7f4SStefano Zampini } 1430fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1431fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 1432fd14bc51SStefano Zampini } 14339ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 143408122e43SStefano Zampini PetscInt j; 143508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 143608122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 143708122e43SStefano Zampini } 143808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 143908122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 144008122e43SStefano Zampini } 144108122e43SStefano Zampini } else { 144208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 144308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 144408122e43SStefano Zampini } 144508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 144608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 14479d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 144808122e43SStefano Zampini #else 14499d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 145008122e43SStefano Zampini #endif 145108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 145208122e43SStefano Zampini B_neigs += B_neigs2; 145308122e43SStefano Zampini } 145408122e43SStefano Zampini if (B_ierr) { 14556c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 14566c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 14576c4ed002SBarry Smith else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 145808122e43SStefano Zampini } 1459fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1460ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 146108122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 146208122e43SStefano Zampini if (eigs[j] == 0.0) { 1463ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 146408122e43SStefano Zampini } else { 14659d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1466ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 14679d54b7f4SStefano Zampini } else { 14689d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 14699d54b7f4SStefano Zampini } 1470fd14bc51SStefano Zampini } 147108122e43SStefano Zampini } 147208122e43SStefano Zampini } 147308122e43SStefano Zampini } else { 147408122e43SStefano Zampini /* TODO */ 147508122e43SStefano Zampini } 1476aff50787SStefano Zampini } 14776c3e6151SStefano Zampini /* change the basis back to the original one */ 14786c3e6151SStefano Zampini if (sub_schurs->change) { 147972b8c272SStefano Zampini Mat change,phi,phit; 14806c3e6151SStefano Zampini 14816c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 14826c3e6151SStefano Zampini PetscInt ii; 14836c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 14846c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 14856c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 14866c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 14876c3e6151SStefano Zampini } 14886c3e6151SStefano Zampini } 14896c3e6151SStefano Zampini } 149072b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 14916c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 149272b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 14936c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 14946c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 14956c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 14966c3e6151SStefano Zampini } 14978bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 14988bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 14999162d606SStefano Zampini if (B_neigs) { 15009162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1501fd14bc51SStefano Zampini 1502fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 15039552c7c7SStefano Zampini PetscInt ii; 15049552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1505ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15069552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1507ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1508ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1509ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1510ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1511ac47001eSStefano Zampini #else 1512ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 1513ac47001eSStefano Zampini #endif 15149552c7c7SStefano Zampini } 15159552c7c7SStefano Zampini } 1516fd14bc51SStefano Zampini } 15179162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 15189162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 15199162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 15209162d606SStefano Zampini cum++; 152108122e43SStefano Zampini } 152208122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 152308122e43SStefano Zampini /* shift for next computation */ 152408122e43SStefano Zampini cumarray += subset_size*subset_size; 152508122e43SStefano Zampini } 1526fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1527fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1528fd14bc51SStefano Zampini } 152908122e43SStefano Zampini 153008122e43SStefano Zampini if (mss) { 153108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 153208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1533f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1534f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1535f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 153608122e43SStefano Zampini } 1537f6f667cfSStefano Zampini if (allocated_S_St) { 1538f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1539f6f667cfSStefano Zampini } 1540f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 154108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 154208122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 154308122e43SStefano Zampini #endif 154408122e43SStefano Zampini if (pcbddc->dbg_flag) { 15451b968477SStefano Zampini PetscInt maxneigs_r; 1546b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 15479b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 154808122e43SStefano Zampini } 154908122e43SStefano Zampini PetscFunctionReturn(0); 155008122e43SStefano Zampini } 1551b1b3d7a2SStefano Zampini 1552674ae819SStefano Zampini #undef __FUNCT__ 1553c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1554c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1555c8587f34SStefano Zampini { 15568629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1557c8587f34SStefano Zampini PetscErrorCode ierr; 1558c8587f34SStefano Zampini 1559c8587f34SStefano Zampini PetscFunctionBegin; 1560f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 15615e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1562c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1563c8587f34SStefano Zampini 1564684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 15650fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1566684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1567c8587f34SStefano Zampini 15688629588bSStefano Zampini /* 15698629588bSStefano Zampini Setup local correction and local part of coarse basis. 15708629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 15718629588bSStefano Zampini */ 157247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 15738629588bSStefano Zampini 15748629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 15758629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 15768629588bSStefano Zampini 15778629588bSStefano Zampini /* free */ 15788629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1579c8587f34SStefano Zampini PetscFunctionReturn(0); 1580c8587f34SStefano Zampini } 1581c8587f34SStefano Zampini 1582c8587f34SStefano Zampini #undef __FUNCT__ 1583674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1584674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1585674ae819SStefano Zampini { 1586674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1587674ae819SStefano Zampini PetscErrorCode ierr; 1588674ae819SStefano Zampini 1589674ae819SStefano Zampini PetscFunctionBegin; 1590674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1591674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 159230368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1593674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1594785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1595674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1596f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1597f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1598785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 159963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 160063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1601674ae819SStefano Zampini PetscFunctionReturn(0); 1602674ae819SStefano Zampini } 1603674ae819SStefano Zampini 1604674ae819SStefano Zampini #undef __FUNCT__ 1605674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1606674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1607674ae819SStefano Zampini { 1608674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16094f1b2e48SStefano Zampini PetscInt i; 1610674ae819SStefano Zampini PetscErrorCode ierr; 1611674ae819SStefano Zampini 1612674ae819SStefano Zampini PetscFunctionBegin; 1613b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1614674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 161516909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 16161dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1617674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1618*669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 1619674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 16204f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 16214f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 16224f1b2e48SStefano Zampini } 16234f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1624b334f244SStefano Zampini if (pcbddc->sub_schurs) { 1625b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1626b334f244SStefano Zampini } 1627674ae819SStefano Zampini PetscFunctionReturn(0); 1628674ae819SStefano Zampini } 1629674ae819SStefano Zampini 1630674ae819SStefano Zampini #undef __FUNCT__ 1631674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1632674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1633674ae819SStefano Zampini { 1634674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1635674ae819SStefano Zampini PetscErrorCode ierr; 1636674ae819SStefano Zampini 1637674ae819SStefano Zampini PetscFunctionBegin; 1638674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 163958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1640ca92afb2SStefano Zampini PetscScalar *array; 164106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 164206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 164358da7f69SStefano Zampini } 1644674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1645674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 164615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 164715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1648674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1649674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1650674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 165106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1652674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1653674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 16548ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1655674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1656674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1657674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1658f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1659f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1660f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1661f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1662727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 16630e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1664f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 166570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 166681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 16670369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 16681dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 16694f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 16708b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1671ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1672ca92afb2SStefano Zampini PetscInt i; 1673ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1674ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1675ca92afb2SStefano Zampini } 1676ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1677ca92afb2SStefano Zampini } 16784f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1679674ae819SStefano Zampini PetscFunctionReturn(0); 1680674ae819SStefano Zampini } 1681674ae819SStefano Zampini 1682674ae819SStefano Zampini #undef __FUNCT__ 1683f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1684f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 16856bfb1811SStefano Zampini { 16866bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16876bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 16886bfb1811SStefano Zampini VecType impVecType; 16894f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 16906bfb1811SStefano Zampini PetscErrorCode ierr; 16916bfb1811SStefano Zampini 16926bfb1811SStefano Zampini PetscFunctionBegin; 16936c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1694e7b262bdSStefano Zampini /* get sizes */ 16954f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1696b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 16976bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1698e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1699e7b262bdSStefano Zampini /* R nodes */ 1700e7b262bdSStefano Zampini old_size = -1; 1701e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1702e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1703e7b262bdSStefano Zampini } 1704e7b262bdSStefano Zampini if (n_R != old_size) { 1705e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1706e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17076bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 17086bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 17096bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 17106bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1711e7b262bdSStefano Zampini } 1712e7b262bdSStefano Zampini /* local primal dofs */ 1713e7b262bdSStefano Zampini old_size = -1; 1714e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1715e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1716e7b262bdSStefano Zampini } 1717e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1718e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 171983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1720e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 17216bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1722e7b262bdSStefano Zampini } 1723e7b262bdSStefano Zampini /* local explicit constraints */ 1724e7b262bdSStefano Zampini old_size = -1; 1725e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1726e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1727e7b262bdSStefano Zampini } 1728e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1729e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 173083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 173183b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 173283b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 173383b7ccabSStefano Zampini } 17346bfb1811SStefano Zampini PetscFunctionReturn(0); 17356bfb1811SStefano Zampini } 17366bfb1811SStefano Zampini 17376bfb1811SStefano Zampini #undef __FUNCT__ 173847f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 173947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 174088ebb749SStefano Zampini { 174125084f0cSStefano Zampini PetscErrorCode ierr; 174225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 174388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 174488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1745d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 174625084f0cSStefano Zampini /* submatrices of local problem */ 174780677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 174806656605SStefano Zampini /* submatrices of local coarse problem */ 174906656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 175025084f0cSStefano Zampini /* working matrices */ 175106656605SStefano Zampini Mat C_CR; 175225084f0cSStefano Zampini /* additional working stuff */ 175306656605SStefano Zampini PC pc_R; 17544f1b2e48SStefano Zampini Mat F; 17555cbda25cSStefano Zampini Vec dummy_vec; 1756a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 175725084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 175806656605SStefano Zampini PetscScalar *work; 175906656605SStefano Zampini PetscInt *idx_V_B; 1760ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 176106656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1762ffd830a3SStefano Zampini 176325084f0cSStefano Zampini /* some shortcuts to scalars */ 176406656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 176588ebb749SStefano Zampini 176688ebb749SStefano Zampini PetscFunctionBegin; 1767ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1768ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1769ffd830a3SStefano Zampini } 1770ffd830a3SStefano Zampini 1771ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1772b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 17734f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1774b371cd4fSStefano Zampini n_B = pcis->n_B; 1775b371cd4fSStefano Zampini n_D = pcis->n - n_B; 177688ebb749SStefano Zampini n_R = pcis->n - n_vertices; 177788ebb749SStefano Zampini 177888ebb749SStefano Zampini /* vertices in boundary numbering */ 1779785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 17800e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 17816c4ed002SBarry Smith if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D\n",n_vertices,i); 178288ebb749SStefano Zampini 178306656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1784019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 178506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 178606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 178706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 178806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 178906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 179006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 179106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 179206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 179306656605SStefano Zampini 179406656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 179506656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 179606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 179706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 179806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1799ffd830a3SStefano Zampini lda_rhs = n_R; 1800a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 180106656605SStefano Zampini if (isLU || isILU || isCHOL) { 180206656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1803b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 1804df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1805d62866d3SStefano Zampini MatFactorType type; 1806d62866d3SStefano Zampini 1807df4d28bfSStefano Zampini F = reuse_solver->F; 18086816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1809d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1810ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 181122db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 181206656605SStefano Zampini } else { 181306656605SStefano Zampini F = NULL; 181406656605SStefano Zampini } 181506656605SStefano Zampini 1816ffd830a3SStefano Zampini /* allocate workspace */ 1817ffd830a3SStefano Zampini n = 0; 1818ffd830a3SStefano Zampini if (n_constraints) { 1819ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1820ffd830a3SStefano Zampini } 1821ffd830a3SStefano Zampini if (n_vertices) { 1822ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1823ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1824ffd830a3SStefano Zampini } 1825ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1826ffd830a3SStefano Zampini 18275cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 18285cbda25cSStefano Zampini dummy_vec = NULL; 18295cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 18305cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 18315cbda25cSStefano Zampini } 18325cbda25cSStefano Zampini 183388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 183488ebb749SStefano Zampini if (n_constraints) { 183572b8c272SStefano Zampini Mat M1,M2,M3,C_B; 183606656605SStefano Zampini IS is_aux; 183780677318SStefano Zampini PetscScalar *array,*array2; 183806656605SStefano Zampini 1839f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 184080677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 184188ebb749SStefano Zampini 184225084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 184325084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 18448ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 184572b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 184688ebb749SStefano Zampini 184780677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 184880677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1849ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 185088ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 185106656605SStefano Zampini const PetscScalar *row_cmat_values; 185206656605SStefano Zampini const PetscInt *row_cmat_indices; 185306656605SStefano Zampini PetscInt size_of_constraint,j; 185488ebb749SStefano Zampini 185506656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 185606656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1857ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 185806656605SStefano Zampini } 185906656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 186006656605SStefano Zampini } 1861ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 186206656605SStefano Zampini if (F) { 186306656605SStefano Zampini Mat B; 186406656605SStefano Zampini 1865ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1866a3df083aSStefano Zampini if (need_benign_correction) { 1867df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1868a3df083aSStefano Zampini 186972b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 187072b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1871a3df083aSStefano Zampini } 187280677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1873a3df083aSStefano Zampini if (need_benign_correction) { 1874a3df083aSStefano Zampini PetscScalar *marr; 1875df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1876a3df083aSStefano Zampini 1877a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 18785cbda25cSStefano Zampini if (lda_rhs != n_R) { 18795cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 18805cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18815cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 18825cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18835cbda25cSStefano Zampini } 18845cbda25cSStefano Zampini } else { 1885a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1886a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18875cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1888a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1889a3df083aSStefano Zampini } 18905cbda25cSStefano Zampini } 1891a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1892a3df083aSStefano Zampini } 189306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 189406656605SStefano Zampini } else { 189580677318SStefano Zampini PetscScalar *marr; 189680677318SStefano Zampini 189780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 189806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1899ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1900ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 190106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 190206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 190306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 190406656605SStefano Zampini } 190580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 190606656605SStefano Zampini } 190780677318SStefano Zampini if (!pcbddc->switch_static) { 190880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 190980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 191080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 191180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1912ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 191380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 191480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 191780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 191880677318SStefano Zampini } 191980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 192080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 192172b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 192280677318SStefano Zampini } else { 1923ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1924ffd830a3SStefano Zampini IS dummy; 1925ffd830a3SStefano Zampini 1926ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 192772b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1928ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1929ffd830a3SStefano Zampini } else { 193080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 193180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1932ffd830a3SStefano Zampini } 193325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 193480677318SStefano Zampini } 193580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 193680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 193780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 193806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 193906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 194080677318SStefano Zampini if (isCHOL) { 194180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 194280677318SStefano Zampini } else { 194325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 194480677318SStefano Zampini } 194580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 194606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 194725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 194825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 194925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 195080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 195172b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 195272b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 195306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 195406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1955f4ddd8eeSStefano Zampini } 1956fc227af8SStefano Zampini 1957fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 195888ebb749SStefano Zampini if (n_vertices) { 195906656605SStefano Zampini IS is_aux; 19603a50541eSStefano Zampini 1961b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 19626816873aSStefano Zampini IS tis; 19636816873aSStefano Zampini 19646816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 19656816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 19666816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 19676816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 19686816873aSStefano Zampini } else { 19693a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 19706816873aSStefano Zampini } 19719577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 19729577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 197304708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 197425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 197588ebb749SStefano Zampini } 197688ebb749SStefano Zampini 197788ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1978f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 197906656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 198006656605SStefano Zampini if (pcbddc->coarse_phi_D) { 198106656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 198206656605SStefano Zampini } 1983f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 198406656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 198506656605SStefano Zampini PetscScalar *marray; 198606656605SStefano Zampini 198706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 198806656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1989f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1990f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1991f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1992f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1993f4ddd8eeSStefano Zampini } 1994f4ddd8eeSStefano Zampini } 199506656605SStefano Zampini 1996f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 199706656605SStefano Zampini PetscScalar *marray; 199888ebb749SStefano Zampini 199906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20008eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 200106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 200288ebb749SStefano Zampini } 20033301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 200406656605SStefano Zampini n *= 2; 200588ebb749SStefano Zampini } 200606656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 200706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 200806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20098eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 201006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 201106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 201288ebb749SStefano Zampini } 20133301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 201406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 20158eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 201606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 201706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 201888ebb749SStefano Zampini } 201988ebb749SStefano Zampini } else { 2020c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 2021c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 20221b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2023c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 2024c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 2025c0553b1fSStefano Zampini } 202688ebb749SStefano Zampini } 202706656605SStefano Zampini } 2028019a44ceSStefano Zampini 202906656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 20304f1b2e48SStefano Zampini p0_lidx_I = NULL; 20314f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 2032d12edf2fSStefano Zampini const PetscInt *idxs; 2033d12edf2fSStefano Zampini 2034d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 20354f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 20364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20374f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 20384f1b2e48SStefano Zampini } 2039d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 2040d12edf2fSStefano Zampini } 2041d16cbb6bSStefano Zampini 204206656605SStefano Zampini /* vertices */ 204306656605SStefano Zampini if (n_vertices) { 204416f15bc4SStefano Zampini 2045af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 204604708bb6SStefano Zampini 204716f15bc4SStefano Zampini if (n_R) { 204814393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 204906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 205016f15bc4SStefano Zampini PetscScalar *x,*y; 205104708bb6SStefano Zampini PetscBool isseqaij; 205206656605SStefano Zampini 205321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 205414393ed6SStefano Zampini if (need_benign_correction) { 205514393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 205614393ed6SStefano Zampini IS is_p0; 205714393ed6SStefano Zampini PetscInt *idxs_p0,n; 205814393ed6SStefano Zampini 205914393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 206014393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 206114393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2062af25d912SStefano Zampini if (n != pcbddc->benign_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 206314393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 206414393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 206514393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 206614393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 206714393ed6SStefano Zampini } 206814393ed6SStefano Zampini 2069ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2070af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2071ffd830a3SStefano Zampini } else { 2072ca92afb2SStefano Zampini PetscScalar *av,*array; 2073ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2074ca92afb2SStefano Zampini PetscInt n; 2075ca92afb2SStefano Zampini PetscBool flg_row; 2076ffd830a3SStefano Zampini 2077ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2078ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 20799d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2080ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2081ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2082ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2083ca92afb2SStefano Zampini PetscInt j; 2084ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2085ffd830a3SStefano Zampini } 2086ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2087ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2088ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2089ffd830a3SStefano Zampini } 2090ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2091a3df083aSStefano Zampini if (need_benign_correction) { 2092df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2093a3df083aSStefano Zampini PetscScalar *marr; 2094a3df083aSStefano Zampini 2095a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 209614393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 209714393ed6SStefano Zampini 209814393ed6SStefano Zampini | 0 0 0 | (V) 209914393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 210014393ed6SStefano Zampini | 0 0 -1 | (p0) 210114393ed6SStefano Zampini 210214393ed6SStefano Zampini */ 2103df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 210414393ed6SStefano Zampini const PetscScalar *vals; 210514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 210614393ed6SStefano Zampini PetscInt n,j,nz; 210714393ed6SStefano Zampini 2108df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2109df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 211014393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 211114393ed6SStefano Zampini for (j=0;j<n;j++) { 211214393ed6SStefano Zampini PetscScalar val = vals[j]; 211314393ed6SStefano Zampini PetscInt k,col = idxs[j]; 211414393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 211514393ed6SStefano Zampini } 211614393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2117df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 211814393ed6SStefano Zampini } 211972b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 212072b8c272SStefano Zampini } 212172b8c272SStefano Zampini if (F) { 212214393ed6SStefano Zampini /* need to correct the rhs */ 212372b8c272SStefano Zampini if (need_benign_correction) { 212472b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 212572b8c272SStefano Zampini PetscScalar *marr; 212672b8c272SStefano Zampini 212772b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 21285cbda25cSStefano Zampini if (lda_rhs != n_R) { 21295cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21305cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21315cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 21325cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21335cbda25cSStefano Zampini } 21345cbda25cSStefano Zampini } else { 2135a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2136a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21375cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2138a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2139a3df083aSStefano Zampini } 21405cbda25cSStefano Zampini } 2141a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2142a3df083aSStefano Zampini } 214306656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 214414393ed6SStefano Zampini /* need to correct the solution */ 2145a3df083aSStefano Zampini if (need_benign_correction) { 2146df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2147a3df083aSStefano Zampini PetscScalar *marr; 2148a3df083aSStefano Zampini 2149a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 21505cbda25cSStefano Zampini if (lda_rhs != n_R) { 21515cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21525cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21535cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 21545cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21555cbda25cSStefano Zampini } 21565cbda25cSStefano Zampini } else { 2157a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2158a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21595cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2160a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2161a3df083aSStefano Zampini } 21625cbda25cSStefano Zampini } 2163a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2164a3df083aSStefano Zampini } 216506656605SStefano Zampini } else { 216606656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 216706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2168ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2169ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 217106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217306656605SStefano Zampini } 217406656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 217506656605SStefano Zampini } 217680677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2177ffd830a3SStefano Zampini /* S_VV and S_CV */ 217806656605SStefano Zampini if (n_constraints) { 217906656605SStefano Zampini Mat B; 218080677318SStefano Zampini 2181ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 218280677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2183ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2184ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 218580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 218880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 218980677318SStefano Zampini } 2190ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 219180677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 219280677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2193ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 219480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 219506656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2196ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2197ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 219806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 219906656605SStefano Zampini } 220004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 220104708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2202511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 220304708bb6SStefano Zampini } 2204ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2205ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2206ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2207ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2208ffd830a3SStefano Zampini } 220906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 221014393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 221114393ed6SStefano Zampini if (need_benign_correction) { 2212df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 221314393ed6SStefano Zampini PetscScalar *marr,*sums; 221414393ed6SStefano Zampini 221514393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 221614393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2217df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 221814393ed6SStefano Zampini const PetscScalar *vals; 221914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 222014393ed6SStefano Zampini PetscInt n,j,nz; 222114393ed6SStefano Zampini 2222df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2223df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 222414393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 222514393ed6SStefano Zampini PetscInt k; 222614393ed6SStefano Zampini sums[j] = 0.; 222714393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 222814393ed6SStefano Zampini } 222914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 223014393ed6SStefano Zampini for (j=0;j<n;j++) { 223114393ed6SStefano Zampini PetscScalar val = vals[j]; 223214393ed6SStefano Zampini PetscInt k; 223314393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 223414393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 223514393ed6SStefano Zampini } 223614393ed6SStefano Zampini } 223714393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2238df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 223914393ed6SStefano Zampini } 224014393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 224114393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 224214393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 224314393ed6SStefano Zampini } 224480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 224506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 224606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 224706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 224806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 224906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 225006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 225106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2252d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2253019a44ceSStefano Zampini } else { 2254d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2255d16cbb6bSStefano Zampini } 225621eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2257d16cbb6bSStefano Zampini 225806656605SStefano Zampini /* coarse basis functions */ 225906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 226016f15bc4SStefano Zampini PetscScalar *y; 226116f15bc4SStefano Zampini 2262ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 226306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 226406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 226506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 226606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 226706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 226806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 226906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 227006656605SStefano Zampini 227106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22724f1b2e48SStefano Zampini PetscInt j; 22734f1b2e48SStefano Zampini 227406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 227506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 227606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 227706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 227806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 22794f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 228006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 228106656605SStefano Zampini } 228206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 228306656605SStefano Zampini } 228404708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 228504708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 228606656605SStefano Zampini } 22875cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 228806656605SStefano Zampini 228906656605SStefano Zampini if (n_constraints) { 229006656605SStefano Zampini Mat B; 229106656605SStefano Zampini 2292ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 229306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 229480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 229506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 229606656605SStefano Zampini if (n_vertices) { 229780677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 229880677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 229980677318SStefano Zampini } else { 230080677318SStefano Zampini Mat S_VCt; 230180677318SStefano Zampini 2302ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2303ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 230472b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2305ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2306ffd830a3SStefano Zampini } 230780677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 230880677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 230980677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 231080677318SStefano Zampini } 231106656605SStefano Zampini } 231206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 231306656605SStefano Zampini /* coarse basis functions */ 231406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 231506656605SStefano Zampini PetscScalar *y; 231606656605SStefano Zampini 2317ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 231806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 231906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 232006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 232306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 232406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23254f1b2e48SStefano Zampini PetscInt j; 23264f1b2e48SStefano Zampini 232706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 232806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 232906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23324f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 233306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 233406656605SStefano Zampini } 233506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 233606656605SStefano Zampini } 233706656605SStefano Zampini } 233880677318SStefano Zampini if (n_constraints) { 233980677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 234080677318SStefano Zampini } 23414f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 234272b8c272SStefano Zampini 234372b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 234472b8c272SStefano Zampini if (pcbddc->benign_n) { 234572b8c272SStefano Zampini Mat B0_B,B0_BPHI; 234672b8c272SStefano Zampini IS is_dummy; 234772b8c272SStefano Zampini PetscScalar *data; 234872b8c272SStefano Zampini PetscInt j; 234972b8c272SStefano Zampini 235072b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 235172b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 235272b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 235372b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 235486c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 235572b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 235672b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 235772b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 235872b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 235972b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 236072b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 236172b8c272SStefano Zampini } 236272b8c272SStefano Zampini } 236372b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 236472b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 236572b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 236672b8c272SStefano Zampini } 2367019a44ceSStefano Zampini 236806656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 23693301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2370ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2371ffd830a3SStefano Zampini PetscScalar *marray; 237206656605SStefano Zampini 237306656605SStefano Zampini if (n_constraints) { 2374ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 237506656605SStefano Zampini 2376af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 237706656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2378ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 237916f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 238006656605SStefano Zampini if (n_vertices) { 2381ffd830a3SStefano Zampini Mat S_VCT; 238206656605SStefano Zampini 238306656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2384ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 238516f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 238606656605SStefano Zampini } 2387ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 23885b782168SStefano Zampini } else { 23895b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 239006656605SStefano Zampini } 239116f15bc4SStefano Zampini if (n_vertices && n_R) { 2392ffd830a3SStefano Zampini PetscScalar *av,*marray; 2393ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2394ffd830a3SStefano Zampini PetscInt n; 2395ffd830a3SStefano Zampini PetscBool flg_row; 239606656605SStefano Zampini 2397ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2398af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2399ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2400ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2401ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2402ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2403ffd830a3SStefano Zampini PetscInt j; 2404ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2405ffd830a3SStefano Zampini } 2406ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2407ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2408ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 240906656605SStefano Zampini } 241006656605SStefano Zampini 2411ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2412ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2413ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2414ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2415ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 241606656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 241706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 241806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 241906656605SStefano Zampini } 2420ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 24215b782168SStefano Zampini if (B_C) { 2422ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2423ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2424ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2425ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2426ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2427ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2428ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 242906656605SStefano Zampini } 2430ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 24315b782168SStefano Zampini } 24325b782168SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 24335b782168SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 24345b782168SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 24355b782168SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 24365b782168SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 24375b782168SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 24385b782168SStefano Zampini } 24395b782168SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 244006656605SStefano Zampini /* coarse basis functions */ 244106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 244206656605SStefano Zampini PetscScalar *y; 244306656605SStefano Zampini 2444ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 244506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 244606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 244706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 244806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 244906656605SStefano Zampini if (i<n_vertices) { 245006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 245106656605SStefano Zampini } 245206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 245306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 245406656605SStefano Zampini 245506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 245606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 245706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 245806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 245906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 246106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 246206656605SStefano Zampini } 246306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 246406656605SStefano Zampini } 2465ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2466ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 246706656605SStefano Zampini } 2468d62866d3SStefano Zampini /* free memory */ 246988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 247006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 247106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 247206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 247306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2474d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2475d62866d3SStefano Zampini if (n_vertices) { 2476d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2477d62866d3SStefano Zampini } 2478d62866d3SStefano Zampini if (n_constraints) { 2479d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2480d62866d3SStefano Zampini } 248188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 248288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 248388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2484d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 248588ebb749SStefano Zampini Mat coarse_sub_mat; 248625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 248788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 248888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 248988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 24908bec7fa6SStefano Zampini Mat C_B,CPHI; 24918bec7fa6SStefano Zampini IS is_dummy; 24928bec7fa6SStefano Zampini Vec mones; 249388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 249488ebb749SStefano Zampini PetscReal real_value; 249588ebb749SStefano Zampini 2496a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2497a3df083aSStefano Zampini Mat A; 2498a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2499a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2500a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2501a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2502a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2503a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2504a3df083aSStefano Zampini } else { 250588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 250688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 250788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 250888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2509a3df083aSStefano Zampini } 251088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 251188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2512ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 251388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 251488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 251588ebb749SStefano Zampini } 251688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 251788ebb749SStefano Zampini 251825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 25193301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 252025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2521ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 252288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 252388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 252488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 252588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 252688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 252788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 252888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 252988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 253088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 253188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 253288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 253388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 253488ebb749SStefano Zampini } else { 253588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 253688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 253788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 253888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 253988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 254088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 254188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 254288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 254388ebb749SStefano Zampini } 254488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 254588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 254688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2547511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 25484f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2549fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2550d12edf2fSStefano Zampini PetscScalar *data,*data2; 25514f1b2e48SStefano Zampini PetscInt j; 2552d12edf2fSStefano Zampini 25534f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2554fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2555d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 255686c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2557d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2558d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 25594f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 25604f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2561d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 25624f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 25634f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 25644f1b2e48SStefano Zampini } 2565d12edf2fSStefano Zampini } 2566d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2567d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2568d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2569d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2570d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2571d12edf2fSStefano Zampini } 2572d12edf2fSStefano Zampini #if 0 2573d12edf2fSStefano Zampini { 2574d12edf2fSStefano Zampini PetscViewer viewer; 2575d12edf2fSStefano Zampini char filename[256]; 2576ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2577d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2578d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2579ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2580ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2581ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2582d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 258372b8c272SStefano Zampini if (save_change) { 258472b8c272SStefano Zampini Mat phi_B; 258572b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 258672b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 258772b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 258872b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 258972b8c272SStefano Zampini } else { 2590ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2591ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 259272b8c272SStefano Zampini } 2593ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2594ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2595ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2596ffd830a3SStefano Zampini } 2597ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2598ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2599ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2600ffd830a3SStefano Zampini } 260172b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2602ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2603ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2604ffd830a3SStefano Zampini } 2605d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2606d12edf2fSStefano Zampini } 2607d12edf2fSStefano Zampini #endif 260881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 26098bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 26101575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 261106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 26128bec7fa6SStefano Zampini 26138bec7fa6SStefano Zampini /* check constraints */ 2614a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2615a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 26164f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 26178bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2618a00504b5SStefano Zampini } else { 2619a00504b5SStefano Zampini PetscScalar *data; 2620a00504b5SStefano Zampini Mat tmat; 2621a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2622a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2623a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2624a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2625a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2626a00504b5SStefano Zampini } 26278bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 26288bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 26298bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 26308bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2631bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2632ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2633bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2634bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2635bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2636bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2637bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 263888ebb749SStefano Zampini } 26398bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 26408bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 26418bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 26428bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 264325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 264488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 264588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 264688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 264788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 264888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 264988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 265088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 265188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 265288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 265388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2654ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 265588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 265688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 265788ebb749SStefano Zampini } 265888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 265988ebb749SStefano Zampini } 26608629588bSStefano Zampini /* get back data */ 26618629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 266288ebb749SStefano Zampini PetscFunctionReturn(0); 266388ebb749SStefano Zampini } 266488ebb749SStefano Zampini 266588ebb749SStefano Zampini #undef __FUNCT__ 2666d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2667d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2668aa0d41d4SStefano Zampini { 2669d65f70fdSStefano Zampini Mat *work_mat; 2670d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2671d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2672c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2673aa0d41d4SStefano Zampini PetscErrorCode ierr; 2674aa0d41d4SStefano Zampini 2675aa0d41d4SStefano Zampini PetscFunctionBegin; 2676d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2677d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2678d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2679d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2680aa0d41d4SStefano Zampini 2681d65f70fdSStefano Zampini if (!rsorted) { 2682906d46d4SStefano Zampini const PetscInt *idxs; 2683906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2684aa0d41d4SStefano Zampini 2685d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2686d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2687d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2688d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2689aa0d41d4SStefano Zampini } 2690d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2691d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2692d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2693d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2694aa0d41d4SStefano Zampini } 2695d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2696d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2697d65f70fdSStefano Zampini } else { 2698d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2699d65f70fdSStefano Zampini isrow_s = isrow; 2700aa0d41d4SStefano Zampini } 2701906d46d4SStefano Zampini 2702d65f70fdSStefano Zampini if (!csorted) { 2703d65f70fdSStefano Zampini if (isrow == iscol) { 2704d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2705d65f70fdSStefano Zampini iscol_s = isrow_s; 2706d65f70fdSStefano Zampini } else { 2707d65f70fdSStefano Zampini const PetscInt *idxs; 2708d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2709906d46d4SStefano Zampini 2710d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2711d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2712d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2713d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2714d65f70fdSStefano Zampini } 2715d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2716d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2717d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2718d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2719d65f70fdSStefano Zampini } 2720d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2721d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2722d65f70fdSStefano Zampini } 2723d65f70fdSStefano Zampini } else { 2724d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2725d65f70fdSStefano Zampini iscol_s = iscol; 2726d65f70fdSStefano Zampini } 2727d65f70fdSStefano Zampini 2728d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2729d65f70fdSStefano Zampini 2730d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2731906d46d4SStefano Zampini Mat new_mat; 2732d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2733906d46d4SStefano Zampini 2734d65f70fdSStefano Zampini if (!rsorted) { 2735d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2736d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2737d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2738d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2739906d46d4SStefano Zampini } 2740d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2741d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2742d65f70fdSStefano Zampini } else { 2743d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2744906d46d4SStefano Zampini } 2745d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2746d65f70fdSStefano Zampini 2747d65f70fdSStefano Zampini if (!csorted) { 2748d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2749d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2750d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2751d65f70fdSStefano Zampini } else { 2752d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2753d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2754d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2755d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2756d65f70fdSStefano Zampini } 2757d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2758d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2759d65f70fdSStefano Zampini } 2760d65f70fdSStefano Zampini } else { 2761d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2762d65f70fdSStefano Zampini } 2763d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2764d65f70fdSStefano Zampini 2765d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2766d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2767d65f70fdSStefano Zampini work_mat[0] = new_mat; 2768d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2769d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2770d65f70fdSStefano Zampini } 2771d65f70fdSStefano Zampini 2772d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2773d65f70fdSStefano Zampini *B = work_mat[0]; 2774d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2775d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2776d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2777d65f70fdSStefano Zampini PetscFunctionReturn(0); 2778d65f70fdSStefano Zampini } 2779d65f70fdSStefano Zampini 2780d65f70fdSStefano Zampini #undef __FUNCT__ 27815e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 27825e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2783aa0d41d4SStefano Zampini { 2784aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 27855e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2786d65f70fdSStefano Zampini Mat new_mat; 27875e8657edSStefano Zampini IS is_local,is_global; 2788d65f70fdSStefano Zampini PetscInt local_size; 2789d65f70fdSStefano Zampini PetscBool isseqaij; 2790aa0d41d4SStefano Zampini PetscErrorCode ierr; 2791aa0d41d4SStefano Zampini 2792aa0d41d4SStefano Zampini PetscFunctionBegin; 2793aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 27945e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 27955e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2796b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2797aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2798d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2799aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2800906d46d4SStefano Zampini 2801906d46d4SStefano Zampini /* check */ 2802906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2803906d46d4SStefano Zampini Vec x,x_change; 2804906d46d4SStefano Zampini PetscReal error; 2805906d46d4SStefano Zampini 28065e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2807906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 28085e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2809e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2810e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2811d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2812e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2813e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2814906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2815906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2816906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2817906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2818906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2819906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2820906d46d4SStefano Zampini } 2821906d46d4SStefano Zampini 282222d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 28239b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 282422d5777bSStefano Zampini if (isseqaij) { 2825a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2826a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2827aa0d41d4SStefano Zampini } else { 2828a00504b5SStefano Zampini Mat work_mat; 28291cf9b237SStefano Zampini 2830a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2831aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2832a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 28331d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2834aa0d41d4SStefano Zampini } 28353301b35fSStefano Zampini if (matis->A->symmetric_set) { 28363301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2837e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 28383301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2839e496cd5dSStefano Zampini #endif 28403301b35fSStefano Zampini } 2841d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2842aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2843aa0d41d4SStefano Zampini } 2844aa0d41d4SStefano Zampini 2845aa0d41d4SStefano Zampini #undef __FUNCT__ 2846a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 28478ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2848a64d13efSStefano Zampini { 2849a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2850a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2851d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 285253892102SStefano Zampini PetscInt *idx_R_local=NULL; 28533a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 28543a50541eSStefano Zampini PetscInt vbs,bs; 28556816873aSStefano Zampini PetscBT bitmask=NULL; 2856a64d13efSStefano Zampini PetscErrorCode ierr; 2857a64d13efSStefano Zampini 2858a64d13efSStefano Zampini PetscFunctionBegin; 2859b23d619eSStefano Zampini /* 2860b23d619eSStefano Zampini No need to setup local scatters if 2861b23d619eSStefano Zampini - primal space is unchanged 2862b23d619eSStefano Zampini AND 2863b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2864b23d619eSStefano Zampini AND 2865b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2866b23d619eSStefano Zampini */ 2867b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2868f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2869f4ddd8eeSStefano Zampini } 2870f4ddd8eeSStefano Zampini /* destroy old objects */ 2871f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2872f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2873f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2874a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2875b371cd4fSStefano Zampini n_B = pcis->n_B; 2876b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2877b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 28783a50541eSStefano Zampini 2879a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 28806816873aSStefano Zampini 288153892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2882b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2883854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2884a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2885a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 28860e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2887a64d13efSStefano Zampini } 2888a64d13efSStefano Zampini 2889a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 28904641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 28916816873aSStefano Zampini idx_R_local[n_R++] = i; 2892a64d13efSStefano Zampini } 2893a64d13efSStefano Zampini } 2894df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2895df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 28966816873aSStefano Zampini 2897df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2898df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 28996816873aSStefano Zampini } 29003a50541eSStefano Zampini 29013a50541eSStefano Zampini /* Block code */ 29023a50541eSStefano Zampini vbs = 1; 29033a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 29043a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 29053a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 29063a50541eSStefano Zampini PetscInt *vary; 2907b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2908785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 29093a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2910d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2911d3df7717SStefano Zampini /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */ 29120e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2913d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 29143a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 29153a50541eSStefano Zampini is_blocked = PETSC_FALSE; 29163a50541eSStefano Zampini break; 29173a50541eSStefano Zampini } 29183a50541eSStefano Zampini } 2919d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2920d3df7717SStefano Zampini } else { 2921d3df7717SStefano Zampini /* Verify directly the R set */ 2922d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2923d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2924d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2925d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2926d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2927d3df7717SStefano Zampini break; 2928d3df7717SStefano Zampini } 2929d3df7717SStefano Zampini } 2930d3df7717SStefano Zampini } 2931d3df7717SStefano Zampini } 29323a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 29333a50541eSStefano Zampini vbs = bs; 29343a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 29353a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 29363a50541eSStefano Zampini } 29373a50541eSStefano Zampini } 29383a50541eSStefano Zampini } 29393a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2940b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 2941df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 294253892102SStefano Zampini 2943df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2944df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 294553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2946df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 294753892102SStefano Zampini } else { 29483a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 294953892102SStefano Zampini } 2950a64d13efSStefano Zampini 2951a64d13efSStefano Zampini /* print some info if requested */ 2952a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2953a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2954a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29551575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2956a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2957a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 29584f1b2e48SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"r_size = %d, v_size = %d, constraints = %d, local_primal_size = %d\n",n_R,n_vertices,pcbddc->local_primal_size-n_vertices-pcbddc->benign_n,pcbddc->local_primal_size);CHKERRQ(ierr); 2959a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2960a64d13efSStefano Zampini } 2961a64d13efSStefano Zampini 2962a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2963b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 29646816873aSStefano Zampini IS is_aux1,is_aux2; 29656816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 29666816873aSStefano Zampini 29673a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2968854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2969854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2970a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 29714641a718SStefano Zampini for (i=0; i<n_D; i++) { 29724641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 29734641a718SStefano Zampini } 2974a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2975a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 29764641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 29774641a718SStefano Zampini aux_array1[j++] = i; 2978a64d13efSStefano Zampini } 2979a64d13efSStefano Zampini } 2980a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2981a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2982a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 29834641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 29844641a718SStefano Zampini aux_array2[j++] = i; 2985a64d13efSStefano Zampini } 2986a64d13efSStefano Zampini } 2987a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2988a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2989a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2990a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2991a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2992a64d13efSStefano Zampini 29938eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2994785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2995a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 29964641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 29974641a718SStefano Zampini aux_array1[j++] = i; 2998a64d13efSStefano Zampini } 2999a64d13efSStefano Zampini } 3000a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3001a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 3002a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3003a64d13efSStefano Zampini } 30044641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 30053a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3006d62866d3SStefano Zampini } else { 3007df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 30086816873aSStefano Zampini IS tis; 30096816873aSStefano Zampini PetscInt schur_size; 30106816873aSStefano Zampini 3011df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 30126816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 3013df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 30146816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 30156816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 30166816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 30176816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 30186816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 3019d62866d3SStefano Zampini } 3020d62866d3SStefano Zampini } 3021a64d13efSStefano Zampini PetscFunctionReturn(0); 3022a64d13efSStefano Zampini } 3023a64d13efSStefano Zampini 3024304d26faSStefano Zampini 3025304d26faSStefano Zampini #undef __FUNCT__ 3026304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 3027684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 3028304d26faSStefano Zampini { 3029304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3030304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3031304d26faSStefano Zampini PC pc_temp; 3032304d26faSStefano Zampini Mat A_RR; 3033f4ddd8eeSStefano Zampini MatReuse reuse; 3034304d26faSStefano Zampini PetscScalar m_one = -1.0; 3035304d26faSStefano Zampini PetscReal value; 303604708bb6SStefano Zampini PetscInt n_D,n_R; 3037c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 3038304d26faSStefano Zampini PetscErrorCode ierr; 3039e604994aSStefano Zampini /* prefixes stuff */ 3040312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 3041e604994aSStefano Zampini size_t len; 3042304d26faSStefano Zampini 3043304d26faSStefano Zampini PetscFunctionBegin; 3044304d26faSStefano Zampini 3045e604994aSStefano Zampini /* compute prefixes */ 3046e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 3047e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 3048e604994aSStefano Zampini if (!pcbddc->current_level) { 3049e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3050e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3051e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3052e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3053e604994aSStefano Zampini } else { 3054e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3055312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3056e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3057e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 3058312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3059312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 306034d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 306134d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3062e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3063e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3064e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 3065e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 3066e604994aSStefano Zampini } 3067e604994aSStefano Zampini 3068304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3069684f6988SStefano Zampini if (dirichlet) { 3070d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3071450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 3072b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 3073450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3074450f8f5eSStefano Zampini } 3075450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3076a3df083aSStefano Zampini Mat A_IIn; 3077a3df083aSStefano Zampini 3078a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3079a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3080a3df083aSStefano Zampini pcis->A_II = A_IIn; 3081a3df083aSStefano Zampini } 3082450f8f5eSStefano Zampini } 30833301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 30843301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3085964fefecSStefano Zampini } 3086ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3087964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3088304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3089304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3090304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3091304d26faSStefano Zampini /* default */ 3092304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3093e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 30949577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3095304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 30969577ea80SStefano Zampini if (issbaij) { 30979577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 30989577ea80SStefano Zampini } else { 3099304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31009577ea80SStefano Zampini } 3101304d26faSStefano Zampini /* Allow user's customization */ 3102304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3103304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3104304d26faSStefano Zampini } 3105d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3106b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3107df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3108d62866d3SStefano Zampini 3109df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3110d5574798SStefano Zampini } 3111304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3112304d26faSStefano Zampini if (!n_D) { 3113304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3114304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3115304d26faSStefano Zampini } 3116304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3117304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3118304d26faSStefano Zampini /* set ksp_D into pcis data */ 3119304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3120304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3121304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3122684f6988SStefano Zampini } 3123304d26faSStefano Zampini 3124304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3125684f6988SStefano Zampini A_RR = 0; 3126684f6988SStefano Zampini if (neumann) { 3127d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 312804708bb6SStefano Zampini PetscInt ibs,mbs; 312904708bb6SStefano Zampini PetscBool issbaij; 313004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3131f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 31328ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3133f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3134f4ddd8eeSStefano Zampini PetscInt nn_R; 313581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3136f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3137f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3138f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3139f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3140f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3141f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3142f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3143727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3144f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3145f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3146f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3147f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3148f4ddd8eeSStefano Zampini } 3149f4ddd8eeSStefano Zampini } 3150f4ddd8eeSStefano Zampini /* last check */ 3151d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3152f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3153f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3154f4ddd8eeSStefano Zampini } 3155f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3156f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3157f4ddd8eeSStefano Zampini } 3158a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3159af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3160af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 316104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 316204708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 316304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 316404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 316504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3166af732b37SStefano Zampini } else { 3167511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 31686816873aSStefano Zampini } 316904708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 317004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 317104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 317204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 317304708bb6SStefano Zampini } else { 3174511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 317504708bb6SStefano Zampini } 317604708bb6SStefano Zampini } 3177a00504b5SStefano Zampini /* extract A_RR */ 3178b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3179a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3180a00504b5SStefano Zampini 3181a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 318216e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3183a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 318416e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 318516e386b8SStefano Zampini } else { 3186a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3187a00504b5SStefano Zampini } 3188a00504b5SStefano Zampini } else { 3189a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3190a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3191a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3192a00504b5SStefano Zampini } 3193a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3194f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 319516e386b8SStefano Zampini } 31963301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31973301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 31986816873aSStefano Zampini } 3199f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3200304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3201304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3202304d26faSStefano Zampini /* default */ 3203304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3204e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3205304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 32069577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 32079577ea80SStefano Zampini if (issbaij) { 32089577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 32099577ea80SStefano Zampini } else { 3210304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 32119577ea80SStefano Zampini } 3212304d26faSStefano Zampini /* Allow user's customization */ 3213304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3214304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3215304d26faSStefano Zampini } 3216304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3217304d26faSStefano Zampini if (!n_R) { 3218304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3219304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3220304d26faSStefano Zampini } 32215cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3222df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3223b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3224df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3225d62866d3SStefano Zampini 3226df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3227d62866d3SStefano Zampini } 3228304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3229304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3230684f6988SStefano Zampini } 3231304d26faSStefano Zampini 3232684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3233684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32341575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3235684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3236684f6988SStefano Zampini } 3237c7017625SStefano Zampini 3238c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3239c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3240c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3241c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3242c7017625SStefano Zampini } 3243c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3244c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3245c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3246c7017625SStefano Zampini } 3247c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3248c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3249c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3250c7017625SStefano Zampini } 3251c7017625SStefano Zampini 3252c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3253c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3254684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 32550fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 32560fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 32570fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 32580fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 32590fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3260e604994aSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Dirichlet solve (%s) = % 1.14e \n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_D))->prefix,value);CHKERRQ(ierr); 3261c7017625SStefano Zampini if (check_corr[0]) { 3262c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3263c7017625SStefano Zampini } 3264304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3265304d26faSStefano Zampini } 3266684f6988SStefano Zampini if (neumann) { /* Neumann */ 32670fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 32680fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 32690fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 32700fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 32710fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3272e604994aSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Neumann solve (%s) = % 1.14e\n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_R))->prefix,value);CHKERRQ(ierr); 3273c7017625SStefano Zampini if (check_corr[1]) { 3274c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3275c7017625SStefano Zampini } 3276304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3277304d26faSStefano Zampini } 3278684f6988SStefano Zampini } 32795cbda25cSStefano Zampini /* free Neumann problem's matrix */ 32805cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3281304d26faSStefano Zampini PetscFunctionReturn(0); 3282304d26faSStefano Zampini } 3283304d26faSStefano Zampini 3284304d26faSStefano Zampini #undef __FUNCT__ 3285ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 328680677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3287674ae819SStefano Zampini { 3288674ae819SStefano Zampini PetscErrorCode ierr; 3289674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3290be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3291b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 3292674ae819SStefano Zampini 3293674ae819SStefano Zampini PetscFunctionBegin; 3294b334f244SStefano Zampini if (!reuse_solver) { 329580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 329620c7b377SStefano Zampini } 329780677318SStefano Zampini if (!pcbddc->switch_static) { 329880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 329980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 330080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 330120c7b377SStefano Zampini } 3302b334f244SStefano Zampini if (!reuse_solver) { 330380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 330480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 330520c7b377SStefano Zampini } else { 3306df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3307be83ff47SStefano Zampini 3308df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3309df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 331020c7b377SStefano Zampini } 3311be83ff47SStefano Zampini } else { 331280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 331380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 331480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 331580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 331680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 331780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 331880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 331980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 332080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3321674ae819SStefano Zampini } 3322674ae819SStefano Zampini } 3323b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 332480677318SStefano Zampini if (applytranspose) { 332580677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 332680677318SStefano Zampini } else { 332780677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 332880677318SStefano Zampini } 3329be83ff47SStefano Zampini } else { 3330df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3331be83ff47SStefano Zampini 3332be83ff47SStefano Zampini if (applytranspose) { 3333df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3334be83ff47SStefano Zampini } else { 3335df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3336be83ff47SStefano Zampini } 3337be83ff47SStefano Zampini } 333880677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 333980677318SStefano Zampini if (!pcbddc->switch_static) { 3340b334f244SStefano Zampini if (!reuse_solver) { 334180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 334280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3343be83ff47SStefano Zampini } else { 3344df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3345be83ff47SStefano Zampini 3346df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3347df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3348be83ff47SStefano Zampini } 334980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 335080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 335180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 335280677318SStefano Zampini } 335380677318SStefano Zampini } else { 335480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 335980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 336080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 336180677318SStefano Zampini } 336280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 336380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 336480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 336580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3366674ae819SStefano Zampini } 3367674ae819SStefano Zampini PetscFunctionReturn(0); 3368674ae819SStefano Zampini } 3369674ae819SStefano Zampini 3370dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3371674ae819SStefano Zampini #undef __FUNCT__ 3372674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3373dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3374674ae819SStefano Zampini { 3375674ae819SStefano Zampini PetscErrorCode ierr; 3376674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3377674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3378674ae819SStefano Zampini const PetscScalar zero = 0.0; 3379674ae819SStefano Zampini 3380674ae819SStefano Zampini PetscFunctionBegin; 3381dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 33824fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3383dc359a40SStefano Zampini if (applytranspose) { 3384674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 33858eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3386dc359a40SStefano Zampini } else { 3387674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3388674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 338915aaf578SStefano Zampini } 33904fee134fSStefano Zampini } else { 33914fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 33924fee134fSStefano Zampini } 3393efc2fbd9SStefano Zampini 3394efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 33954f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3396efc2fbd9SStefano Zampini PetscScalar *array; 33974f1b2e48SStefano Zampini PetscInt j; 3398efc2fbd9SStefano Zampini 3399efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34004f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3401efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3402efc2fbd9SStefano Zampini } 3403efc2fbd9SStefano Zampini 340412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 340512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 340612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340812edc857SStefano Zampini 34099f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 341012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 341151694757SStefano Zampini Mat coarse_mat; 3412964fefecSStefano Zampini Vec rhs,sol; 341351694757SStefano Zampini MatNullSpace nullsp; 341427b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3415964fefecSStefano Zampini 341627b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 341727b6a85dSStefano Zampini PC coarse_pc; 341827b6a85dSStefano Zampini 341927b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 342027b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 342127b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 342227b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 342327b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 342427b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 34253bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 342627b6a85dSStefano Zampini } 342727b6a85dSStefano Zampini } 3428964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3429964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 343051694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 343151694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 343251694757SStefano Zampini if (nullsp) { 343351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 343451694757SStefano Zampini } 343512edc857SStefano Zampini if (applytranspose) { 34361f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only) { 34371f4df5f7SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 34382701bc32SStefano Zampini } else { 3439964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 34402701bc32SStefano Zampini } 34412701bc32SStefano Zampini } else { 34421f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 34432701bc32SStefano Zampini PC coarse_pc; 34442701bc32SStefano Zampini 34452701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 34462701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 34473e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 34482701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 344912edc857SStefano Zampini } else { 3450964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 345112edc857SStefano Zampini } 34522701bc32SStefano Zampini } 34531d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 345427b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 345527b6a85dSStefano Zampini PC coarse_pc; 345627b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 345727b6a85dSStefano Zampini 345827b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 345927b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 346027b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 34613bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 346227b6a85dSStefano Zampini } 346351694757SStefano Zampini if (nullsp) { 346451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 346551694757SStefano Zampini } 346612edc857SStefano Zampini } 3467674ae819SStefano Zampini 3468674ae819SStefano Zampini /* Local solution on R nodes */ 34694fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 347080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 34719f00e9b4SStefano Zampini } 34729f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 34739f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 347412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3475674ae819SStefano Zampini 34764fee134fSStefano Zampini /* Sum contributions from the two levels */ 34774fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3478dc359a40SStefano Zampini if (applytranspose) { 3479dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3480dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3481dc359a40SStefano Zampini } else { 3482674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 34838eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3484dc359a40SStefano Zampini } 3485efc2fbd9SStefano Zampini /* store p0 */ 34864f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3487efc2fbd9SStefano Zampini PetscScalar *array; 34884f1b2e48SStefano Zampini PetscInt j; 3489efc2fbd9SStefano Zampini 3490efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34914f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3492efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3493efc2fbd9SStefano Zampini } 34944fee134fSStefano Zampini } else { /* expand the coarse solution */ 34954fee134fSStefano Zampini if (applytranspose) { 34964fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 34974fee134fSStefano Zampini } else { 34984fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 34994fee134fSStefano Zampini } 35004fee134fSStefano Zampini } 3501674ae819SStefano Zampini PetscFunctionReturn(0); 3502674ae819SStefano Zampini } 3503674ae819SStefano Zampini 3504674ae819SStefano Zampini #undef __FUNCT__ 3505674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 350612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3507674ae819SStefano Zampini { 3508674ae819SStefano Zampini PetscErrorCode ierr; 3509674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 351058da7f69SStefano Zampini PetscScalar *array; 351112edc857SStefano Zampini Vec from,to; 3512674ae819SStefano Zampini 3513674ae819SStefano Zampini PetscFunctionBegin; 351412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 351512edc857SStefano Zampini from = pcbddc->coarse_vec; 351612edc857SStefano Zampini to = pcbddc->vec1_P; 351712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 351812edc857SStefano Zampini Vec tvec; 351958da7f69SStefano Zampini 352058da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 352158da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 352212edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 352358da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 352458da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 352558da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 352612edc857SStefano Zampini } 352712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 352812edc857SStefano Zampini from = pcbddc->vec1_P; 352912edc857SStefano Zampini to = pcbddc->coarse_vec; 353012edc857SStefano Zampini } 353112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3532674ae819SStefano Zampini PetscFunctionReturn(0); 3533674ae819SStefano Zampini } 3534674ae819SStefano Zampini 3535674ae819SStefano Zampini #undef __FUNCT__ 3536674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 353712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3538674ae819SStefano Zampini { 3539674ae819SStefano Zampini PetscErrorCode ierr; 3540674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 354158da7f69SStefano Zampini PetscScalar *array; 354212edc857SStefano Zampini Vec from,to; 3543674ae819SStefano Zampini 3544674ae819SStefano Zampini PetscFunctionBegin; 354512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 354612edc857SStefano Zampini from = pcbddc->coarse_vec; 354712edc857SStefano Zampini to = pcbddc->vec1_P; 354812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 354912edc857SStefano Zampini from = pcbddc->vec1_P; 355012edc857SStefano Zampini to = pcbddc->coarse_vec; 355112edc857SStefano Zampini } 355212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 355312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 355412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 355512edc857SStefano Zampini Vec tvec; 355658da7f69SStefano Zampini 355712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 355858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 355958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 356058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 356158da7f69SStefano Zampini } 356258da7f69SStefano Zampini } else { 356358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 356458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 356512edc857SStefano Zampini } 356612edc857SStefano Zampini } 3567674ae819SStefano Zampini PetscFunctionReturn(0); 3568674ae819SStefano Zampini } 3569674ae819SStefano Zampini 3570984c4197SStefano Zampini /* uncomment for testing purposes */ 3571984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3572674ae819SStefano Zampini #undef __FUNCT__ 3573674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3574674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3575674ae819SStefano Zampini { 3576674ae819SStefano Zampini PetscErrorCode ierr; 3577674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3578674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3579674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3580984c4197SStefano Zampini /* one and zero */ 3581984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3582984c4197SStefano Zampini /* space to store constraints and their local indices */ 35839162d606SStefano Zampini PetscScalar *constraints_data; 35849162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 35859162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 35869162d606SStefano Zampini PetscInt *constraints_n; 3587984c4197SStefano Zampini /* iterators */ 3588b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3589984c4197SStefano Zampini /* BLAS integers */ 3590e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3591e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3592c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3593727cdba6SStefano Zampini /* reuse */ 35940e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 35950e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3596984c4197SStefano Zampini /* change of basis */ 3597b3d85658SStefano Zampini PetscBool qr_needed; 35989162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3599984c4197SStefano Zampini /* auxiliary stuff */ 360064efe560SStefano Zampini PetscInt *nnz,*is_indices; 36018a0068c3SStefano Zampini PetscInt ncc; 3602984c4197SStefano Zampini /* some quantities */ 360345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3604a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3605984c4197SStefano Zampini 3606674ae819SStefano Zampini PetscFunctionBegin; 36078e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 36088e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 36098e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 361016909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3611088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3612088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 36130e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 36140e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 36150e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36160e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36170e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3618088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3619cf5a6209SStefano Zampini 3620cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 36219162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3622cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3623cf5a6209SStefano Zampini const Vec *nearnullvecs; 3624cf5a6209SStefano Zampini Vec *localnearnullsp; 3625cf5a6209SStefano Zampini PetscScalar *array; 3626cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3627cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3628674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3629b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3630674ae819SStefano Zampini PetscScalar *work; 3631674ae819SStefano Zampini PetscReal *singular_vals; 3632674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3633674ae819SStefano Zampini PetscReal *rwork; 3634674ae819SStefano Zampini #endif 3635674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3636674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3637674ae819SStefano Zampini #else 3638964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3639964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3640674ae819SStefano Zampini #endif 3641674ae819SStefano Zampini 3642674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3643d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3644e4d548c7SStefano Zampini /* print some info */ 36451f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3646e4d548c7SStefano Zampini PetscInt nv; 3647e4d548c7SStefano Zampini 3648e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 3649e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3650e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3651e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3652e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3653e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3654e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3655e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3656e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3657e4d548c7SStefano Zampini } 3658e4d548c7SStefano Zampini 3659d06fc5fdSStefano Zampini /* free unneeded index sets */ 3660d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3661d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3662674ae819SStefano Zampini } 3663d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3664d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3665d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3666d06fc5fdSStefano Zampini } 3667d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3668d06fc5fdSStefano Zampini n_ISForEdges = 0; 3669d06fc5fdSStefano Zampini } 3670d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3671d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3672d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3673d06fc5fdSStefano Zampini } 3674d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3675d06fc5fdSStefano Zampini n_ISForFaces = 0; 3676d06fc5fdSStefano Zampini } 367770022509SStefano Zampini 3678674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3679674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3680674ae819SStefano Zampini if (nearnullsp) { 3681674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3682f4ddd8eeSStefano Zampini /* remove any stored info */ 3683f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3684f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3685f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3686f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3687f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3688473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3689f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3690f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3691f4ddd8eeSStefano Zampini } 3692984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3693984c4197SStefano Zampini nnsp_size = 0; 3694674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3695674ae819SStefano Zampini } 3696984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3697984c4197SStefano Zampini max_constraints = nnsp_size; 3698984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3699984c4197SStefano Zampini 3700674ae819SStefano Zampini /* 3701674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 37029162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 37039162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 37049162d606SStefano Zampini There can be multiple constraints per connected component 3705674ae819SStefano Zampini */ 3706674ae819SStefano Zampini n_vertices = 0; 3707674ae819SStefano Zampini if (ISForVertices) { 3708674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3709674ae819SStefano Zampini } 37109162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 37119162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 37129162d606SStefano Zampini 37139162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 37149162d606SStefano Zampini total_counts *= max_constraints; 3715674ae819SStefano Zampini total_counts += n_vertices; 37164641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 37179162d606SStefano Zampini 3718674ae819SStefano Zampini total_counts = 0; 3719674ae819SStefano Zampini max_size_of_constraint = 0; 3720674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 37219162d606SStefano Zampini IS used_is; 3722674ae819SStefano Zampini if (i<n_ISForEdges) { 37239162d606SStefano Zampini used_is = ISForEdges[i]; 3724674ae819SStefano Zampini } else { 37259162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3726674ae819SStefano Zampini } 37279162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3728674ae819SStefano Zampini total_counts += j; 3729674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3730674ae819SStefano Zampini } 37319162d606SStefano Zampini ierr = PetscMalloc3(total_counts*max_constraints+n_vertices,&constraints_data,total_counts+n_vertices,&constraints_idxs,total_counts+n_vertices,&constraints_idxs_B);CHKERRQ(ierr); 37329162d606SStefano Zampini 3733984c4197SStefano Zampini /* get local part of global near null space vectors */ 3734785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3735984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3736984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3737e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3738e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3739984c4197SStefano Zampini } 3740674ae819SStefano Zampini 3741242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3742242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3743a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3744242a89d7SStefano Zampini 3745984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3746a773dcb8SStefano Zampini if (!skip_lapack) { 3747674ae819SStefano Zampini PetscScalar temp_work; 3748911cabfeSStefano Zampini 3749674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3750984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3751785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3752785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3753785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3754674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3755785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3756674ae819SStefano Zampini #endif 3757674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3758c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3759c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3760674ae819SStefano Zampini lwork = -1; 3761674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3762674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3763c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3764674ae819SStefano Zampini #else 3765c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3766674ae819SStefano Zampini #endif 3767674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3768984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3769674ae819SStefano Zampini #else /* on missing GESVD */ 3770674ae819SStefano Zampini /* SVD */ 3771674ae819SStefano Zampini PetscInt max_n,min_n; 3772674ae819SStefano Zampini max_n = max_size_of_constraint; 3773984c4197SStefano Zampini min_n = max_constraints; 3774984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3775674ae819SStefano Zampini min_n = max_size_of_constraint; 3776984c4197SStefano Zampini max_n = max_constraints; 3777674ae819SStefano Zampini } 3778785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3779674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3780785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3781674ae819SStefano Zampini #endif 3782674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3783674ae819SStefano Zampini lwork = -1; 3784e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3785e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3786b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3787674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3788674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 37899162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr)); 3790674ae819SStefano Zampini #else 37919162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr)); 3792674ae819SStefano Zampini #endif 3793674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3794984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3795984c4197SStefano Zampini #endif /* on missing GESVD */ 3796674ae819SStefano Zampini /* Allocate optimal workspace */ 3797674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3798854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3799674ae819SStefano Zampini } 3800674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3801674ae819SStefano Zampini total_counts = 0; 38029162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 38039162d606SStefano Zampini constraints_data_ptr[0] = 0; 3804674ae819SStefano Zampini /* vertices */ 38059162d606SStefano Zampini if (n_vertices) { 3806674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 38079162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3808674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 38099162d606SStefano Zampini constraints_n[total_counts] = 1; 38109162d606SStefano Zampini constraints_data[total_counts] = 1.0; 38119162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 38129162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3813674ae819SStefano Zampini total_counts++; 3814674ae819SStefano Zampini } 3815674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3816674ae819SStefano Zampini n_vertices = total_counts; 3817674ae819SStefano Zampini } 3818984c4197SStefano Zampini 3819674ae819SStefano Zampini /* edges and faces */ 38209162d606SStefano Zampini total_counts_cc = total_counts; 3821911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 38229162d606SStefano Zampini IS used_is; 38239162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 38249162d606SStefano Zampini 3825911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 38269162d606SStefano Zampini used_is = ISForEdges[ncc]; 3827984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3828674ae819SStefano Zampini } else { 38299162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3830984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3831674ae819SStefano Zampini } 3832674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 38339162d606SStefano Zampini 38349162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 38359162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3836984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3837984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3838674ae819SStefano Zampini if (nnsp_has_cnst) { 38395b08dc53SStefano Zampini PetscScalar quad_value; 38409162d606SStefano Zampini 38419162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38429162d606SStefano Zampini idxs_copied = PETSC_TRUE; 38439162d606SStefano Zampini 3844a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3845674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3846a773dcb8SStefano Zampini } else { 3847a773dcb8SStefano Zampini quad_value = 1.0; 3848a773dcb8SStefano Zampini } 3849674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38509162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3851674ae819SStefano Zampini } 38529162d606SStefano Zampini temp_constraints++; 3853674ae819SStefano Zampini total_counts++; 3854674ae819SStefano Zampini } 3855674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3856984c4197SStefano Zampini PetscReal real_value; 38579162d606SStefano Zampini PetscScalar *ptr_to_data; 38589162d606SStefano Zampini 3859984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 38609162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3861674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38629162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3863674ae819SStefano Zampini } 3864984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3865984c4197SStefano Zampini /* check if array is null on the connected component */ 3866e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 38679162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 38685b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3869674ae819SStefano Zampini temp_constraints++; 3870674ae819SStefano Zampini total_counts++; 38719162d606SStefano Zampini if (!idxs_copied) { 38729162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38739162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3874674ae819SStefano Zampini } 3875674ae819SStefano Zampini } 38769162d606SStefano Zampini } 38779162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 387845a1bb75SStefano Zampini valid_constraints = temp_constraints; 3879eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3880a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 38819162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 38829162d606SStefano Zampini 38839162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3884a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 38859162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3886a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 38879162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3888a773dcb8SStefano Zampini } else { /* perform SVD */ 3889984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 38909162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3891674ae819SStefano Zampini 3892674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3893984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3894984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3895984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3896984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3897984c4197SStefano Zampini from that computed using LAPACKgesvd 3898984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3899984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3900984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3901674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3902e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3903984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3904674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3905674ae819SStefano Zampini for (k=0;k<j+1;k++) { 39069162d606SStefano Zampini PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k] = BLASdot_(&Blas_N,ptr_to_data+k*size_of_constraint,&Blas_one,ptr_to_data+j*size_of_constraint,&Blas_one)); 3907674ae819SStefano Zampini } 3908674ae819SStefano Zampini } 3909e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3910e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3911e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3912674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3913c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3914674ae819SStefano Zampini #else 3915c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3916674ae819SStefano Zampini #endif 3917674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3918984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3919984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3920674ae819SStefano Zampini j = 0; 3921984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3922674ae819SStefano Zampini total_counts = total_counts-j; 392345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3924e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3925c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3926c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3927c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3928c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3929c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3930c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3931674ae819SStefano Zampini if (j<temp_constraints) { 3932984c4197SStefano Zampini PetscInt ii; 3933984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3934674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39359162d606SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,ptr_to_data,&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC)); 3936674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3937984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3938674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 39399162d606SStefano Zampini ptr_to_data[k*size_of_constraint+ii] = singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii]; 3940674ae819SStefano Zampini } 3941674ae819SStefano Zampini } 3942674ae819SStefano Zampini } 3943674ae819SStefano Zampini #else /* on missing GESVD */ 3944e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3945e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3946b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3947674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3948674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 39499162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr)); 3950674ae819SStefano Zampini #else 39519162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr)); 3952674ae819SStefano Zampini #endif 3953984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3954674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3955984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3956e310c8b4SStefano Zampini k = temp_constraints; 3957e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3958674ae819SStefano Zampini j = 0; 3959e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 396045a1bb75SStefano Zampini valid_constraints = k-j; 3961911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3962984c4197SStefano Zampini #endif /* on missing GESVD */ 3963674ae819SStefano Zampini } 3964a773dcb8SStefano Zampini } 39659162d606SStefano Zampini /* update pointers information */ 39669162d606SStefano Zampini if (valid_constraints) { 39679162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 39689162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 39699162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 39709162d606SStefano Zampini /* set change_of_basis flag */ 397145a1bb75SStefano Zampini if (boolforchange) { 3972b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 39739162d606SStefano Zampini } 3974b3d85658SStefano Zampini total_counts_cc++; 397545a1bb75SStefano Zampini } 397645a1bb75SStefano Zampini } 3977984c4197SStefano Zampini /* free workspace */ 39788f1c130eSStefano Zampini if (!skip_lapack) { 3979984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3980984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3981984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3982984c4197SStefano Zampini #endif 3983984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3984984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3985984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3986984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3987984c4197SStefano Zampini #endif 3988984c4197SStefano Zampini } 3989984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3990984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3991984c4197SStefano Zampini } 3992984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3993cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3994cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3995cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3996cf5a6209SStefano Zampini } 3997cf5a6209SStefano Zampini if (n_ISForFaces) { 3998cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3999cf5a6209SStefano Zampini } 4000cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4001cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4002cf5a6209SStefano Zampini } 4003cf5a6209SStefano Zampini if (n_ISForEdges) { 4004cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4005cf5a6209SStefano Zampini } 4006cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 400708122e43SStefano Zampini } else { 400808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4009984c4197SStefano Zampini 401008122e43SStefano Zampini total_counts = 0; 401108122e43SStefano Zampini n_vertices = 0; 4012d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 4013d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 401408122e43SStefano Zampini } 401508122e43SStefano Zampini max_constraints = 0; 40169162d606SStefano Zampini total_counts_cc = 0; 401708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 401808122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 40199162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 402008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 402108122e43SStefano Zampini } 40229162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 40239162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 40249162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 40259162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 402674d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 40279162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 40289162d606SStefano Zampini total_counts_cc = 0; 40299162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 40309162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 40319162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 403208122e43SStefano Zampini } 403308122e43SStefano Zampini } 40349162d606SStefano Zampini #if 0 40359162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 40369162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40379162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 40389162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 40399162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 40409162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 40419162d606SStefano Zampini } 40429162d606SStefano Zampini printf("\n"); 40439162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 40449162d606SStefano Zampini } 40451b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 40468bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 40471b968477SStefano Zampini } 40481b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 40498bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] sub %d, edge %d, n %d\n",PetscGlobalRank,i,(PetscBool)PetscBTLookup(sub_schurs->is_edge,i),pcbddc->adaptive_constraints_n[i+n_vertices]); 40501b968477SStefano Zampini } 405108122e43SStefano Zampini #endif 405208122e43SStefano Zampini 40538bec7fa6SStefano Zampini max_size_of_constraint = 0; 40549162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) max_size_of_constraint = PetscMax(max_size_of_constraint,constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]); 40559162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 405608122e43SStefano Zampini /* Change of basis */ 4057b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 405808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 405908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 406008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 4061b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 406208122e43SStefano Zampini } 406308122e43SStefano Zampini } 406408122e43SStefano Zampini } 406508122e43SStefano Zampini } 4066984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 40674f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 406808122e43SStefano Zampini 40699162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 40709162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 40716c4ed002SBarry Smith if (i != constraints_idxs_ptr[total_counts_cc]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %D != %D\n",constraints_idxs_ptr[total_counts_cc],i); 4072674ae819SStefano Zampini 4073674ae819SStefano Zampini /* Create constraint matrix */ 4074674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 407516f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4076984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4077984c4197SStefano Zampini 4078984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4079a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4080a717540cSStefano Zampini qr_needed = PETSC_FALSE; 408174d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4082984c4197SStefano Zampini total_primal_vertices=0; 4083b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 40849162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40859162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 408672b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 40879162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4088b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 408964efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 40909162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 40919162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4092a717540cSStefano Zampini } 4093b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 409491af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4095a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4096a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4097a717540cSStefano Zampini } 4098fa434743SStefano Zampini } else { 4099b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4100fa434743SStefano Zampini } 4101a717540cSStefano Zampini } 4102b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4103b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4104674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 410570022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4106b3d85658SStefano Zampini 41074f1b2e48SStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 41080e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 41090e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4110984c4197SStefano Zampini 4111984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 411274d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4113785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4114984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 411574d5cdf7SStefano Zampini 4116984c4197SStefano Zampini j = total_primal_vertices; 411774d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4118b3d85658SStefano Zampini cum = total_primal_vertices; 41199162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41204641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4121b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4122b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4123b3d85658SStefano Zampini cum++; 41249162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 412574d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 412674d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 412774d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 412874d5cdf7SStefano Zampini } 41299162d606SStefano Zampini j += constraints_n[i]; 4130674ae819SStefano Zampini } 4131674ae819SStefano Zampini } 4132674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4133674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4134088faed8SStefano Zampini 4135674ae819SStefano Zampini /* set values in constraint matrix */ 4136984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 41370e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4138674ae819SStefano Zampini } 4139984c4197SStefano Zampini total_counts = total_primal_vertices; 41409162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41414641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 41429162d606SStefano Zampini PetscInt *cols; 41439162d606SStefano Zampini 41449162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 41459162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 41469162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41479162d606SStefano Zampini PetscInt row = total_counts+k; 41489162d606SStefano Zampini PetscScalar *vals; 41499162d606SStefano Zampini 41509162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 41519162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 41529162d606SStefano Zampini } 41539162d606SStefano Zampini total_counts += constraints_n[i]; 4154674ae819SStefano Zampini } 4155674ae819SStefano Zampini } 4156674ae819SStefano Zampini /* assembling */ 4157674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4158674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4159088faed8SStefano Zampini 4160984c4197SStefano Zampini /* 41616a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4162984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4163f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4164984c4197SStefano Zampini */ 4165674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4166674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4167026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4168984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4169984c4197SStefano Zampini /* working stuff for GEQRF */ 417081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4171984c4197SStefano Zampini PetscBLASInt lqr_work; 4172984c4197SStefano Zampini /* working stuff for UNGQR */ 4173984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4174984c4197SStefano Zampini PetscBLASInt lgqr_work; 4175984c4197SStefano Zampini /* working stuff for TRTRS */ 4176984c4197SStefano Zampini PetscScalar *trs_rhs; 41773f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4178984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4179984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4180984c4197SStefano Zampini PetscScalar *start_vals; 4181984c4197SStefano Zampini /* working stuff for values insertion */ 41824641a718SStefano Zampini PetscBT is_primal; 418364efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4184906d46d4SStefano Zampini /* matrix sizes */ 4185906d46d4SStefano Zampini PetscInt global_size,local_size; 4186906d46d4SStefano Zampini /* temporary change of basis */ 4187906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4188cf5a6209SStefano Zampini /* extra space for debugging */ 4189cf5a6209SStefano Zampini PetscScalar *dbg_work; 4190984c4197SStefano Zampini 4191906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4192906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 419316f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4194bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4195906d46d4SStefano Zampini /* nonzeros for local mat */ 4196bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 41971dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4198bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 41991dd7afcfSStefano Zampini } else { 42001dd7afcfSStefano Zampini const PetscInt *ii; 42011dd7afcfSStefano Zampini PetscInt n; 42021dd7afcfSStefano Zampini PetscBool flg_row; 42031dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42041dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 42051dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42061dd7afcfSStefano Zampini } 42079162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4208a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 42099162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4210a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 42119162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4212a717540cSStefano Zampini } else { 42139162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 42149162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4215a717540cSStefano Zampini } 4216a717540cSStefano Zampini } 4217a717540cSStefano Zampini } 4218906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4219bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 42201dd7afcfSStefano Zampini /* Set interior change in the matrix */ 42211dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4222bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4223906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4224a717540cSStefano Zampini } 42251dd7afcfSStefano Zampini } else { 42261dd7afcfSStefano Zampini const PetscInt *ii,*jj; 42271dd7afcfSStefano Zampini PetscScalar *aa; 42281dd7afcfSStefano Zampini PetscInt n; 42291dd7afcfSStefano Zampini PetscBool flg_row; 42301dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42311dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42321dd7afcfSStefano Zampini for (i=0;i<n;i++) { 42331dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 42341dd7afcfSStefano Zampini } 42351dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42361dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42371dd7afcfSStefano Zampini } 4238a717540cSStefano Zampini 4239a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4240a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4241a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4242a717540cSStefano Zampini } 4243a717540cSStefano Zampini 4244a717540cSStefano Zampini 4245a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4246a717540cSStefano Zampini /* 4247a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4248a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4249a717540cSStefano Zampini 4250a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4251a717540cSStefano Zampini 4252a6b551f4SStefano Zampini - Using the following block transformation if there is only a primal dof on the cc (and -pc_bddc_use_qr_single is not specified) 4253a6b551f4SStefano Zampini 4254a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4255a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4256a717540cSStefano Zampini | ... | 4257a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4258a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4259a717540cSStefano Zampini 4260a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4261a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4262a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4263a6b551f4SStefano Zampini 4264a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4265a717540cSStefano Zampini */ 4266a717540cSStefano Zampini if (qr_needed) { 4267984c4197SStefano Zampini /* space to store Q */ 4268854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4269984c4197SStefano Zampini /* first we issue queries for optimal work */ 42703f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 42713f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 42723f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4273984c4197SStefano Zampini lqr_work = -1; 42743f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4275984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4276984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4277785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4278984c4197SStefano Zampini lgqr_work = -1; 42793f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 42803f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 42813f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 42823f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 42833f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 42843f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4285984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4286984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4287785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4288984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4289785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4290984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4291785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4292a717540cSStefano Zampini /* allocating workspace for check */ 4293a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4294cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4295a717540cSStefano Zampini } 4296a717540cSStefano Zampini } 4297984c4197SStefano Zampini /* array to store whether a node is primal or not */ 42984641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4299473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 43000e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 43016c4ed002SBarry Smith if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %D != %D\n",total_primal_vertices,i); 430239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 430339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 430439e2fb2aSStefano Zampini } 430539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4306984c4197SStefano Zampini 4307a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 43089162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 43099162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 43104641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4311984c4197SStefano Zampini /* get constraint info */ 43129162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4313984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4314984c4197SStefano Zampini 4315984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43169162d606SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %d: %d need a change of basis (size %d)\n",total_counts,primal_dofs,size_of_constraint);CHKERRQ(ierr); 4317674ae819SStefano Zampini } 4318984c4197SStefano Zampini 4319fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4320a717540cSStefano Zampini 4321a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4322a717540cSStefano Zampini if (pcbddc->dbg_flag) { 43239162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4324a717540cSStefano Zampini } 4325984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 43269162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4327984c4197SStefano Zampini 4328984c4197SStefano Zampini /* compute QR decomposition of constraints */ 43293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43303f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4332674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43333f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4334984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4335674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4336984c4197SStefano Zampini 4337984c4197SStefano Zampini /* explictly compute R^-T */ 4338984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4339984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 43403f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43413f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 43423f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4344984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43453f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4346984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4347984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4348984c4197SStefano Zampini 4349a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 43503f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43513f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 43523f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 43533f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4354984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43553f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4356984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4357984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4358984c4197SStefano Zampini 4359984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4360984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4361984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 43623f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43633f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43643f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 43653f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43663f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 43673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4368984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43699162d606SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,constraints_data+constraints_data_ptr[total_counts],&Blas_LDC)); 4370984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 43719162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4372984c4197SStefano Zampini 4373984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 43749162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4375984c4197SStefano Zampini /* insert cols for primal dofs */ 4376984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4377984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 43789162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4379906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4380984c4197SStefano Zampini } 4381984c4197SStefano Zampini /* insert cols for dual dofs */ 4382984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 43839162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4384984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 43859162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4386906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4387984c4197SStefano Zampini j++; 4388674ae819SStefano Zampini } 4389674ae819SStefano Zampini } 4390984c4197SStefano Zampini 4391984c4197SStefano Zampini /* check change of basis */ 4392984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4393984c4197SStefano Zampini PetscInt ii,jj; 4394984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4395c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4396c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4397c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4398c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4399c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4400c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4401984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4402cf5a6209SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("T","N",&Blas_M,&Blas_N,&Blas_K,&one,dbg_work,&Blas_LDA,qr_basis,&Blas_LDB,&zero,&dbg_work[size_of_constraint*primal_dofs],&Blas_LDC)); 4403984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4404984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4405984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4406cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4407cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) valid_qr = PETSC_FALSE; 4408674ae819SStefano Zampini } 4409674ae819SStefano Zampini } 4410984c4197SStefano Zampini if (!valid_qr) { 441122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4412984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4413984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4414cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4415cf5a6209SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not orthogonal to constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii])); 4416674ae819SStefano Zampini } 4417cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4418cf5a6209SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not unitary w.r.t constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii])); 4419984c4197SStefano Zampini } 4420984c4197SStefano Zampini } 4421984c4197SStefano Zampini } 4422674ae819SStefano Zampini } else { 442322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4424674ae819SStefano Zampini } 4425674ae819SStefano Zampini } 4426a717540cSStefano Zampini } else { /* simple transformation block */ 4427a717540cSStefano Zampini PetscInt row,col; 4428a6b551f4SStefano Zampini PetscScalar val,norm; 4429a6b551f4SStefano Zampini 4430a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44319162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,constraints_data+constraints_data_ptr[total_counts],&Blas_one,constraints_data+constraints_data_ptr[total_counts],&Blas_one)); 4432a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 44339162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 44349162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4435bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 44369162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4437906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 44389162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4439a717540cSStefano Zampini } else { 4440a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 44419162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4442a717540cSStefano Zampini if (row != col) { 44439162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4444a717540cSStefano Zampini } else { 44459162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4446a717540cSStefano Zampini } 4447906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4448a717540cSStefano Zampini } 4449a717540cSStefano Zampini } 4450a717540cSStefano Zampini } 445198a51de6SStefano Zampini if (pcbddc->dbg_flag) { 445222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4453a717540cSStefano Zampini } 4454674ae819SStefano Zampini } 4455984c4197SStefano Zampini } else { 4456984c4197SStefano Zampini if (pcbddc->dbg_flag) { 44579162d606SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d does not need a change of basis (size %d)\n",total_counts,size_of_constraint);CHKERRQ(ierr); 4458674ae819SStefano Zampini } 4459674ae819SStefano Zampini } 4460674ae819SStefano Zampini } 4461a717540cSStefano Zampini 4462a717540cSStefano Zampini /* free workspace */ 4463a717540cSStefano Zampini if (qr_needed) { 4464984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4465cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4466984c4197SStefano Zampini } 4467984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4468984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4469984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4470984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4471984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4472674ae819SStefano Zampini } 4473a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4474906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4475906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4476906d46d4SStefano Zampini 4477906d46d4SStefano Zampini /* assembling of global change of variable */ 447888c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4479bbb9e6c6SStefano Zampini Mat tmat; 448016f15bc4SStefano Zampini PetscInt bs; 448116f15bc4SStefano Zampini 4482906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4483906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4484bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4485bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4486bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4487bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 448816f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 448916f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4490906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4491bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4492bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4493bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4494bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4495bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4496e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4497e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4498bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4499bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 450088c03ad3SStefano Zampini 4501906d46d4SStefano Zampini /* check */ 4502906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4503906d46d4SStefano Zampini PetscReal error; 4504906d46d4SStefano Zampini Vec x,x_change; 4505906d46d4SStefano Zampini 4506906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4507906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4508906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4509906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4510e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4511e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4512bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4513e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4514e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4515906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4516906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4517906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4518906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4519bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4520906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4521906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4522906d46d4SStefano Zampini } 4523b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4524b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4525b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4526bf3a8328SStefano Zampini 4527bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4528bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4529bf3a8328SStefano Zampini } 4530b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 4531ac632422SStefano Zampini Mat S_new,tmat; 4532bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4533bbb9e6c6SStefano Zampini 4534bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 45356816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4536bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4537bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4538bf3a8328SStefano Zampini IS is_V; 4539b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4540b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4541b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4542b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4543b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4544bf3a8328SStefano Zampini } 4545bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4546ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4547b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4548ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4549bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4550bf3a8328SStefano Zampini const PetscScalar *array; 4551bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4552bf3a8328SStefano Zampini PetscInt i,n_V; 4553bf3a8328SStefano Zampini 4554b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4555b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4556b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4557b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4558b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4559b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4560b087196eSStefano Zampini PetscScalar val; 4561b087196eSStefano Zampini PetscInt idx; 4562b087196eSStefano Zampini 4563b087196eSStefano Zampini idx = idxs_V[i]; 4564b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4565b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4566b087196eSStefano Zampini } 4567b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4568b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4569bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4570bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4571bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4572bf3a8328SStefano Zampini } 4573ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4574ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4575ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4576ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4577b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4578ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4579bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4580b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4581bf3a8328SStefano Zampini } 4582ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4583ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4584ac632422SStefano Zampini } 4585b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 458688c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4587b96c3477SStefano Zampini } 4588c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4589b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 4590c9db6a07SStefano Zampini PetscInt i; 4591c9db6a07SStefano Zampini 4592c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4593c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4594c9db6a07SStefano Zampini } 4595c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4596c9db6a07SStefano Zampini } 4597b96c3477SStefano Zampini } 459816909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 459916909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 460016909a7fSStefano Zampini } else { 4601906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 460216909a7fSStefano Zampini } 46031dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 460427b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 460572b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 460672b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 460772b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 460872b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 460972b8c272SStefano Zampini } 46101dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 461127b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4612b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4613b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4614906d46d4SStefano Zampini } else { 46151dd7afcfSStefano Zampini Mat benign_global = NULL; 461627b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 46171dd7afcfSStefano Zampini Mat tmat; 46181dd7afcfSStefano Zampini 46191dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 46201dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 46211dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 46221dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46231dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46241dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 46251dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46261dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46271dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 46281dd7afcfSStefano Zampini if (pcbddc->benign_change) { 46291dd7afcfSStefano Zampini Mat M; 46301dd7afcfSStefano Zampini 46311dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 46321dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 46331dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 46341dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4635906d46d4SStefano Zampini } else { 46361dd7afcfSStefano Zampini Mat eye; 46371dd7afcfSStefano Zampini PetscScalar *array; 46381dd7afcfSStefano Zampini 46391dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46401dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 46411dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 46421dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4643906d46d4SStefano Zampini } 46441dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46451dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46461dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46471dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 46481dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 46491dd7afcfSStefano Zampini } 46501dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 46511dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 46521dd7afcfSStefano Zampini } 46531dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 46541dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 46551dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 465627b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 46571dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 46581dd7afcfSStefano Zampini } 46591dd7afcfSStefano Zampini } 466016909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 466116909a7fSStefano Zampini IS is_global; 466216909a7fSStefano Zampini const PetscInt *gidxs; 466316909a7fSStefano Zampini 466416909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 466516909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 466616909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 466716909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 466816909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 466916909a7fSStefano Zampini } 46701dd7afcfSStefano Zampini } 46711dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 46721dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4673b9b85e73SStefano Zampini } 4674a717540cSStefano Zampini 467572b8c272SStefano Zampini if (!pcbddc->fake_change) { 46764f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 46774f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 46784f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 46794f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4680019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4681019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4682019a44ceSStefano Zampini pcbddc->local_primal_size++; 4683019a44ceSStefano Zampini } 4684019a44ceSStefano Zampini 4685019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4686727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4687727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 46889f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 4689c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 46900e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 46919f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 4692727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4693727cdba6SStefano Zampini } 46940e6343abSStefano Zampini } 4695727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4696b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 469772b8c272SStefano Zampini } 469872b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4699727cdba6SStefano Zampini 4700a717540cSStefano Zampini /* flush dbg viewer */ 4701b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4702b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4703b8ffe317SStefano Zampini } 4704a717540cSStefano Zampini 4705e310c8b4SStefano Zampini /* free workspace */ 4706a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 47074641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 470808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 47099162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 47109162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 471108122e43SStefano Zampini } else { 47129162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 47139162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 47149162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 471508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 471608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 47179162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 47189162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 471908122e43SStefano Zampini } 4720674ae819SStefano Zampini PetscFunctionReturn(0); 4721674ae819SStefano Zampini } 4722674ae819SStefano Zampini 4723674ae819SStefano Zampini #undef __FUNCT__ 4724674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4725674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4726674ae819SStefano Zampini { 4727674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4728674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4729674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 473014f95afaSStefano Zampini PetscInt ierr,i,N; 4731674ae819SStefano Zampini 4732674ae819SStefano Zampini PetscFunctionBegin; 47338e61c736SStefano Zampini /* Reset previously computed graph */ 47348e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4735674ae819SStefano Zampini /* Init local Graph struct */ 47367fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 47373bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4738674ae819SStefano Zampini 4739575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 47405099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 47415099eff2SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %d, expected %d\n",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs); 4742575ad6abSStefano Zampini } 47439577ea80SStefano Zampini 4744674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4745d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 47464d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 47474d379d7bSStefano Zampini PetscInt nvtxs; 4748e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4749674ae819SStefano Zampini 47502fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 47512fffb893SStefano Zampini if (flg_row) { 47524d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4753b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 47542fffb893SStefano Zampini } 47552fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4756674ae819SStefano Zampini } 47579b28b941SStefano Zampini if (pcbddc->dbg_flag) { 47589b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4759674ae819SStefano Zampini } 4760674ae819SStefano Zampini 4761674ae819SStefano Zampini /* Setup of Graph */ 476214f95afaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,pcbddc->vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4763674ae819SStefano Zampini 47644f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 47654f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 47664f1b2e48SStefano Zampini PetscInt *local_subs; 47674f1b2e48SStefano Zampini 47684f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 47694f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 47704f1b2e48SStefano Zampini const PetscInt *idxs; 47714f1b2e48SStefano Zampini PetscInt nl,j; 47724f1b2e48SStefano Zampini 47734f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 47744f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 47754f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 47764f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 47774f1b2e48SStefano Zampini } 47784f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 47794f1b2e48SStefano Zampini } 47804f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 47814f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 47824f1b2e48SStefano Zampini } 47834f1b2e48SStefano Zampini 4784674ae819SStefano Zampini /* Graph's connected components analysis */ 4785674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4786674ae819SStefano Zampini PetscFunctionReturn(0); 4787674ae819SStefano Zampini } 4788674ae819SStefano Zampini 4789dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4790674ae819SStefano Zampini #undef __FUNCT__ 4791674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4792dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4793674ae819SStefano Zampini { 4794dc456d91SStefano Zampini PetscSF sf; 4795dc456d91SStefano Zampini PetscLayout map; 4796dc456d91SStefano Zampini const PetscInt *idxs; 4797dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4798dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4799dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4800dc456d91SStefano Zampini PetscMPIInt commsize; 4801674ae819SStefano Zampini PetscBool first_found; 4802674ae819SStefano Zampini PetscErrorCode ierr; 4803674ae819SStefano Zampini 4804674ae819SStefano Zampini PetscFunctionBegin; 4805dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4806dc456d91SStefano Zampini if (subset_mult) { 4807dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4808dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4809dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4810674ae819SStefano Zampini } 4811dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4812dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4813dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4814dc456d91SStefano Zampini for (i=0;i<n;i++) { 4815dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4816dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4817674ae819SStefano Zampini } 4818dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4819b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4820dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4821dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4822dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4823dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4824dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4825dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4826dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4827dc456d91SStefano Zampini 4828dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4829dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4830dc456d91SStefano Zampini if (subset_mult) { 4831dc456d91SStefano Zampini const PetscInt* idxs_mult; 4832dc456d91SStefano Zampini 4833dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4834dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4835dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4836674ae819SStefano Zampini } else { 4837dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4838674ae819SStefano Zampini } 4839dc456d91SStefano Zampini /* local size of new subset */ 4840dc456d91SStefano Zampini n_n = 0; 4841dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4842dc456d91SStefano Zampini 4843dc456d91SStefano Zampini /* global indexes in layout */ 4844dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4845dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4846dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4847dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4848dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4849dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4850dc456d91SStefano Zampini 4851dc456d91SStefano Zampini /* reduce from leaves to roots */ 4852dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 485364a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 485464a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4855dc456d91SStefano Zampini 4856dc456d91SStefano Zampini /* count indexes in local part of layout */ 4857674ae819SStefano Zampini nlocals = 0; 4858674ae819SStefano Zampini first_index = -1; 4859674ae819SStefano Zampini first_found = PETSC_FALSE; 4860dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4861dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4862674ae819SStefano Zampini first_found = PETSC_TRUE; 4863674ae819SStefano Zampini first_index = i; 4864674ae819SStefano Zampini } 4865dc456d91SStefano Zampini nlocals += root_data[i]; 4866674ae819SStefano Zampini } 4867dc456d91SStefano Zampini 4868dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 48695fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4870dc456d91SStefano Zampini start = 0; 487164a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 48725fa240b1SStefano Zampini #else 487364a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 48745fa240b1SStefano Zampini start = start-nlocals; 48755fa240b1SStefano Zampini #endif 48765fa240b1SStefano Zampini 4877dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4878dc456d91SStefano Zampini *N_n = start + nlocals; 4879dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4880dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4881674ae819SStefano Zampini } 48825fa240b1SStefano Zampini 48835fa240b1SStefano Zampini /* adapt root data with cumulative */ 4884674ae819SStefano Zampini if (first_found) { 4885dc456d91SStefano Zampini PetscInt old_index; 4886dc456d91SStefano Zampini 4887dc456d91SStefano Zampini root_data[first_index] += start; 4888674ae819SStefano Zampini old_index = first_index; 4889dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4890dc456d91SStefano Zampini if (root_data[i]) { 4891dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4892674ae819SStefano Zampini old_index = i; 4893674ae819SStefano Zampini } 4894674ae819SStefano Zampini } 4895674ae819SStefano Zampini } 4896dc456d91SStefano Zampini 4897dc456d91SStefano Zampini /* from roots to leaves */ 4898dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4899dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4900dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4901dc456d91SStefano Zampini 4902dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4903dc456d91SStefano Zampini if (subset_mult) { 4904dc456d91SStefano Zampini const PetscInt* idxs_mult; 4905dc456d91SStefano Zampini PetscInt cum; 4906dc456d91SStefano Zampini 4907dc456d91SStefano Zampini cum = 0; 4908dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4909dc456d91SStefano Zampini for (i=0;i<n;i++) { 4910dc456d91SStefano Zampini PetscInt j; 4911dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4912674ae819SStefano Zampini } 4913dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4914674ae819SStefano Zampini } else { 4915dc456d91SStefano Zampini for (i=0;i<n;i++) { 4916dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4917674ae819SStefano Zampini } 4918674ae819SStefano Zampini } 4919dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4920dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4921674ae819SStefano Zampini PetscFunctionReturn(0); 4922674ae819SStefano Zampini } 49239a7d3425SStefano Zampini 4924*669cc0f4SStefano Zampini /* this implements stabilized Gram-Schmidt */ 49259a7d3425SStefano Zampini #undef __FUNCT__ 49269a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 49279a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 49289a7d3425SStefano Zampini { 49299a7d3425SStefano Zampini PetscInt i,j; 49309a7d3425SStefano Zampini PetscScalar *alphas; 49319a7d3425SStefano Zampini PetscErrorCode ierr; 49329a7d3425SStefano Zampini 49339a7d3425SStefano Zampini PetscFunctionBegin; 4934785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 49359a7d3425SStefano Zampini for (i=0;i<n;i++) { 49369a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 4937*669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 4938*669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 4939*669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 49409a7d3425SStefano Zampini } 49419a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 49429a7d3425SStefano Zampini PetscFunctionReturn(0); 49439a7d3425SStefano Zampini } 49449a7d3425SStefano Zampini 4945e7931f94SStefano Zampini #undef __FUNCT__ 494670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 494757de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4948e7931f94SStefano Zampini { 494957de7509SStefano Zampini Mat A; 4950e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4951e7931f94SStefano Zampini PetscMPIInt size,rank,color; 495252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 495352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 495427b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 495557de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 495627b6a85dSStefano Zampini PetscInt xadj_count, *count; 495727b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 495827b6a85dSStefano Zampini PetscSubcomm psubcomm; 495927b6a85dSStefano Zampini MPI_Comm subcomm; 496052e5ac9dSStefano Zampini PetscErrorCode ierr; 4961a57a6d2fSStefano Zampini 4962e7931f94SStefano Zampini PetscFunctionBegin; 496357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 496457de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 496557de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 496657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 496757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 496857de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 496957de7509SStefano Zampini 497057de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 497157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 497257de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 497357de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 497457de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 497557de7509SStefano Zampini im_active = !!(n); 497657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 497757de7509SStefano Zampini void_procs = size - active_procs; 497857de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 497957de7509SStefano Zampini if (void_procs) { 498057de7509SStefano Zampini PetscInt ncand; 498157de7509SStefano Zampini 498257de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 498357de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 498457de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 498557de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 498657de7509SStefano Zampini if (!procs_candidates[i]) { 498757de7509SStefano Zampini procs_candidates[ncand++] = i; 498857de7509SStefano Zampini } 498957de7509SStefano Zampini } 499057de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 499157de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 499257de7509SStefano Zampini } 499357de7509SStefano Zampini 499457de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 499557de7509SStefano Zampini if (active_procs < *n_subdomains) { 499657de7509SStefano Zampini PetscInt issize,isidx; 499757de7509SStefano Zampini if (im_active) { 499857de7509SStefano Zampini issize = 1; 499957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 500057de7509SStefano Zampini isidx = procs_candidates[rank]; 500157de7509SStefano Zampini } else { 500257de7509SStefano Zampini isidx = rank; 500357de7509SStefano Zampini } 500457de7509SStefano Zampini } else { 500557de7509SStefano Zampini issize = 0; 500657de7509SStefano Zampini isidx = -1; 500757de7509SStefano Zampini } 500857de7509SStefano Zampini *n_subdomains = active_procs; 500957de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 5010daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 501157de7509SStefano Zampini PetscFunctionReturn(0); 501257de7509SStefano Zampini } 5013c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 5014c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 501527b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 5016e7931f94SStefano Zampini 5017e7931f94SStefano Zampini /* Get info on mapping */ 50183bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 50193bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 5020e7931f94SStefano Zampini 5021e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 5022785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 5023e7931f94SStefano Zampini xadj[0] = 0; 5024e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 5025785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 5026785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 502727b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 502827b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 502927b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 503027b6a85dSStefano Zampini count[shared[i][j]] += 1; 5031e7931f94SStefano Zampini 503227b6a85dSStefano Zampini xadj_count = 0; 50332b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 503427b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 503527b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 5036d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 5037d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 5038d023bfaeSStefano Zampini xadj_count++; 503927b6a85dSStefano Zampini break; 504027b6a85dSStefano Zampini } 5041e7931f94SStefano Zampini } 5042e7931f94SStefano Zampini } 5043d023bfaeSStefano Zampini xadj[1] = xadj_count; 504427b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 50453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 5046e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 5047e7931f94SStefano Zampini 50483837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 5049e7931f94SStefano Zampini 505027b6a85dSStefano Zampini /* Restrict work on active processes only */ 505127b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 505227b6a85dSStefano Zampini if (void_procs) { 505327b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 505427b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 505527b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 505627b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 505727b6a85dSStefano Zampini } else { 505827b6a85dSStefano Zampini psubcomm = NULL; 505927b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 506027b6a85dSStefano Zampini } 506127b6a85dSStefano Zampini 506227b6a85dSStefano Zampini v_wgt = NULL; 506327b6a85dSStefano Zampini if (!color) { 5064e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 5065e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 5066e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5067c8587f34SStefano Zampini } else { 506852e5ac9dSStefano Zampini Mat subdomain_adj; 506952e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 507052e5ac9dSStefano Zampini MatPartitioning partitioner; 507127b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 507252e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 507357de7509SStefano Zampini PetscMPIInt size; 5074b0c7d250SStefano Zampini PetscBool aggregate; 5075b0c7d250SStefano Zampini 507627b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 507727b6a85dSStefano Zampini if (void_procs) { 507827b6a85dSStefano Zampini PetscInt prank = rank; 5079785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 508027b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 5081e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 5082e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 5083c8587f34SStefano Zampini } 5084e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 508527b6a85dSStefano Zampini } else { 508627b6a85dSStefano Zampini oldranks = NULL; 508727b6a85dSStefano Zampini } 5088b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 508927b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 5090b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 5091b0c7d250SStefano Zampini PetscMPIInt nrank; 5092b0c7d250SStefano Zampini PetscScalar *vals; 5093b0c7d250SStefano Zampini 509427b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5095b0c7d250SStefano Zampini lrows = 0; 5096b0c7d250SStefano Zampini if (nrank<redprocs) { 5097b0c7d250SStefano Zampini lrows = size/redprocs; 5098b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5099b0c7d250SStefano Zampini } 510027b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5101b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5102b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5103b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5104b0c7d250SStefano Zampini row = nrank; 5105b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5106b0c7d250SStefano Zampini cols = adjncy; 5107b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5108b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5109b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5110b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5111b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 511252e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 511352e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 511452e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5115b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 511627b6a85dSStefano Zampini if (use_vwgt) { 511727b6a85dSStefano Zampini Vec v; 511827b6a85dSStefano Zampini const PetscScalar *array; 511927b6a85dSStefano Zampini PetscInt nl; 512027b6a85dSStefano Zampini 512127b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 512227b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 512327b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 512427b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 512527b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 512627b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 512727b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 512822db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 512927b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 513027b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 513127b6a85dSStefano Zampini } 5132b0c7d250SStefano Zampini } else { 513327b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 513427b6a85dSStefano Zampini if (use_vwgt) { 513527b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 513627b6a85dSStefano Zampini v_wgt[0] = local_size; 513727b6a85dSStefano Zampini } 5138b0c7d250SStefano Zampini } 513922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5140e7931f94SStefano Zampini 5141e7931f94SStefano Zampini /* Partition */ 514227b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5143e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 514427b6a85dSStefano Zampini if (v_wgt) { 5145e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5146c8587f34SStefano Zampini } 514757de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 514857de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5149e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5150e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 515122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5152e7931f94SStefano Zampini 515352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 515452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 515552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 515652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 515757de7509SStefano Zampini if (!aggregate) { 515857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 515927b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 516027b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 516127b6a85dSStefano Zampini #endif 516257de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 516327b6a85dSStefano Zampini } else if (oldranks) { 5164b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 516527b6a85dSStefano Zampini } else { 516627b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 516757de7509SStefano Zampini } 516828143c3dSStefano Zampini } else { 5169b0c7d250SStefano Zampini PetscInt idxs[1]; 5170b0c7d250SStefano Zampini PetscMPIInt tag; 5171b0c7d250SStefano Zampini MPI_Request *reqs; 5172b0c7d250SStefano Zampini 5173b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5174b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5175b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 517627b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 517728143c3dSStefano Zampini } 517827b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5179b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5180b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 518157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 518227b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 518327b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 518427b6a85dSStefano Zampini #endif 518557de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 518627b6a85dSStefano Zampini } else if (oldranks) { 5187b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 518827b6a85dSStefano Zampini } else { 518927b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5190e7931f94SStefano Zampini } 519157de7509SStefano Zampini } 519252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5193e7931f94SStefano Zampini /* clean up */ 5194e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 519552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5196e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5197e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5198e7931f94SStefano Zampini } 519927b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 520057de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5201e7931f94SStefano Zampini 5202e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5203e7931f94SStefano Zampini i = 1; 520427b6a85dSStefano Zampini if (!color) i=0; 520557de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5206e7931f94SStefano Zampini PetscFunctionReturn(0); 5207e7931f94SStefano Zampini } 5208e7931f94SStefano Zampini 5209e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5210e7931f94SStefano Zampini 5211e7931f94SStefano Zampini #undef __FUNCT__ 5212e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 521357de7509SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, PetscBool reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 5214e7931f94SStefano Zampini { 521570cf5478SStefano Zampini Mat local_mat; 5216e7931f94SStefano Zampini IS is_sends_internal; 52179d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 521828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 52199d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5220e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5221e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5222e7931f94SStefano Zampini const PetscInt* is_indices; 5223e7931f94SStefano Zampini MatType new_local_type; 5224e7931f94SStefano Zampini /* buffers */ 5225e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 522628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 52279d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5228e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 5229e7931f94SStefano Zampini /* MPI */ 523028143c3dSStefano Zampini MPI_Comm comm,comm_n; 523128143c3dSStefano Zampini PetscSubcomm subcomm; 5232e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 523328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 523428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 523528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 523628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 523728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 5238e7931f94SStefano Zampini PetscErrorCode ierr; 5239e7931f94SStefano Zampini 5240e7931f94SStefano Zampini PetscFunctionBegin; 524157de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5242e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 524328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 524457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 524557de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 524657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 524757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 524857de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 524957de7509SStefano Zampini 525057de7509SStefano Zampini /* further checks */ 5251e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5252e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5253e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5254e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5255e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 525657de7509SStefano Zampini if (reuse && *mat_n) { 525770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 525857de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 525970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 526028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 526170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 526270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 526370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 526470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 526570cf5478SStefano Zampini } 5266e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5267e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 526857de7509SStefano Zampini 5269e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5270e7931f94SStefano Zampini if (!is_sends) { 527128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 527257de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5273c8587f34SStefano Zampini } else { 5274e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5275e7931f94SStefano Zampini is_sends_internal = is_sends; 5276c8587f34SStefano Zampini } 5277e7931f94SStefano Zampini 5278e7931f94SStefano Zampini /* get comm */ 5279a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5280e7931f94SStefano Zampini 5281e7931f94SStefano Zampini /* compute number of sends */ 5282e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5283e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5284e7931f94SStefano Zampini 5285e7931f94SStefano Zampini /* compute number of receives */ 5286e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5287785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5288e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5289e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5290e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5291e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5292e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5293e7931f94SStefano Zampini 529428143c3dSStefano Zampini /* restrict comm if requested */ 529528143c3dSStefano Zampini subcomm = 0; 529628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 529728143c3dSStefano Zampini if (restrict_comm) { 5298779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5299779c1cceSStefano Zampini 530028143c3dSStefano Zampini color = 0; 530153a05cb3SStefano Zampini if (restrict_full) { 530253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 530353a05cb3SStefano Zampini } else { 530453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 530553a05cb3SStefano Zampini } 5306b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 530728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 530828143c3dSStefano Zampini /* check if reuse has been requested */ 530957de7509SStefano Zampini if (reuse) { 531028143c3dSStefano Zampini if (*mat_n) { 531128143c3dSStefano Zampini PetscMPIInt subcommsize2; 531228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 531328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 531428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 531528143c3dSStefano Zampini } else { 531628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 531728143c3dSStefano Zampini } 531828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5319779c1cceSStefano Zampini PetscMPIInt rank; 5320779c1cceSStefano Zampini 5321779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 532228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 532328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 532428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5325306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 532628143c3dSStefano Zampini } 532728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 532828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 532928143c3dSStefano Zampini } else { 533028143c3dSStefano Zampini comm_n = comm; 533128143c3dSStefano Zampini } 533228143c3dSStefano Zampini 5333e7931f94SStefano Zampini /* prepare send/receive buffers */ 5334785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5335e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5336785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5337e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 533828143c3dSStefano Zampini if (nis) { 5339854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 534028143c3dSStefano Zampini } 5341e7931f94SStefano Zampini 534228143c3dSStefano Zampini /* Get data from local matrices */ 53436c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5344e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5345e7931f94SStefano Zampini /* 5346e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5347e7931f94SStefano Zampini send_buffer_idxs should contain: 5348e7931f94SStefano Zampini - MatType_PRIVATE type 5349e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5350e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5351e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5352e7931f94SStefano Zampini */ 53536c4ed002SBarry Smith else { 5354e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 53553bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5356854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5357e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5358e7931f94SStefano Zampini send_buffer_idxs[1] = i; 53593bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5360e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 53613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5362e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5363e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5364e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5365e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5366c8587f34SStefano Zampini } 5367c8587f34SStefano Zampini } 5368e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 536928143c3dSStefano Zampini /* additional is (if any) */ 537028143c3dSStefano Zampini if (nis) { 537128143c3dSStefano Zampini PetscMPIInt psum; 537228143c3dSStefano Zampini PetscInt j; 537328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 537428143c3dSStefano Zampini PetscInt plen; 537528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 537628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 537728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 537828143c3dSStefano Zampini } 5379854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 538028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 538128143c3dSStefano Zampini PetscInt plen; 538228143c3dSStefano Zampini const PetscInt *is_array_idxs; 538328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 538428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 538528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 538628143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 538728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 538828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 538928143c3dSStefano Zampini } 539028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 539128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 539228143c3dSStefano Zampini } 539328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 539428143c3dSStefano Zampini } 539528143c3dSStefano Zampini 5396e7931f94SStefano Zampini buf_size_idxs = 0; 5397e7931f94SStefano Zampini buf_size_vals = 0; 539828143c3dSStefano Zampini buf_size_idxs_is = 0; 5399e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5400e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5401e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 540228143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5403e7931f94SStefano Zampini } 5404785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5405785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 540695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5407e7931f94SStefano Zampini 5408e7931f94SStefano Zampini /* get new tags for clean communications */ 5409e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5410e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 541128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5412e7931f94SStefano Zampini 5413e7931f94SStefano Zampini /* allocate for requests */ 5414785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5415785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 541695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5417785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5418785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 541995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5420e7931f94SStefano Zampini 5421e7931f94SStefano Zampini /* communications */ 5422e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5423e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 542428143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5425e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5426e7931f94SStefano Zampini source_dest = onodes[i]; 5427e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5428e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5429e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5430e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 543128143c3dSStefano Zampini if (nis) { 543257de7509SStefano Zampini source_dest = onodes_is[i]; 543328143c3dSStefano Zampini ierr = MPI_Irecv(ptr_idxs_is,olengths_idxs_is[i],MPIU_INT,source_dest,tag_idxs_is,comm,&recv_req_idxs_is[i]);CHKERRQ(ierr); 543428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 543528143c3dSStefano Zampini } 5436e7931f94SStefano Zampini } 5437e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5438e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5439e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5440e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 544128143c3dSStefano Zampini if (nis) { 544228143c3dSStefano Zampini ierr = MPI_Isend(send_buffer_idxs_is,ilengths_idxs_is[source_dest],MPIU_INT,source_dest,tag_idxs_is,comm,&send_req_idxs_is[i]);CHKERRQ(ierr); 544328143c3dSStefano Zampini } 5444e7931f94SStefano Zampini } 5445e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5446e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5447e7931f94SStefano Zampini 5448e7931f94SStefano Zampini /* assemble new l2g map */ 5449e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5450e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 54519d30be91SStefano Zampini new_local_rows = 0; 5452e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 54539d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5454e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5455e7931f94SStefano Zampini } 54569d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5457e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 54589d30be91SStefano Zampini new_local_rows = 0; 5459e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 54609d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 54619d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5462e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5463e7931f94SStefano Zampini } 54649d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 54659d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5466e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5467e7931f94SStefano Zampini 5468e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5469e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5470e7931f94SStefano Zampini /* it also assumes that if the block size is set, than it is the same among all local matrices (see checks at the beginning of the function) */ 5471e7931f94SStefano Zampini if (n_recvs) { 547228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5473e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5474e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5475e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5476e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5477e7931f94SStefano Zampini break; 5478e7931f94SStefano Zampini } 5479e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5480e7931f94SStefano Zampini } 5481e7931f94SStefano Zampini switch (new_local_type_private) { 548228143c3dSStefano Zampini case MATDENSE_PRIVATE: 548328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5484e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5485e7931f94SStefano Zampini bs = 1; 548628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 548728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 548828143c3dSStefano Zampini bs = 1; 548928143c3dSStefano Zampini } 5490e7931f94SStefano Zampini break; 5491e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5492e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5493e7931f94SStefano Zampini bs = 1; 5494e7931f94SStefano Zampini break; 5495e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5496e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5497e7931f94SStefano Zampini break; 5498e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5499e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5500e7931f94SStefano Zampini break; 5501e7931f94SStefano Zampini default: 55029d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5503e7931f94SStefano Zampini break; 5504e7931f94SStefano Zampini } 550528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 550628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 550728143c3dSStefano Zampini bs = 1; 5508e7931f94SStefano Zampini } 5509e7931f94SStefano Zampini 551070cf5478SStefano Zampini /* create MATIS object if needed */ 551157de7509SStefano Zampini if (!reuse) { 5512e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5513e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 551470cf5478SStefano Zampini } else { 551570cf5478SStefano Zampini /* it also destroys the local matrices */ 551657de7509SStefano Zampini if (*mat_n) { 551770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 551857de7509SStefano Zampini } else { /* this is a fake object */ 551957de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 552057de7509SStefano Zampini } 552170cf5478SStefano Zampini } 552270cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5523e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 55249d30be91SStefano Zampini 55259d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 55269d30be91SStefano Zampini 55279d30be91SStefano Zampini /* Global to local map of received indices */ 55289d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 55299d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 55309d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 55319d30be91SStefano Zampini 55329d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 55339d30be91SStefano Zampini buf_size_idxs = 0; 55349d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 55359d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 55369d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 55379d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 55389d30be91SStefano Zampini } 55399d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 55409d30be91SStefano Zampini 55419d30be91SStefano Zampini /* set preallocation */ 55429d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 55439d30be91SStefano Zampini if (!newisdense) { 55449d30be91SStefano Zampini PetscInt *new_local_nnz=0; 55459d30be91SStefano Zampini 55469d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 55479d30be91SStefano Zampini if (n_recvs) { 55489d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 55499d30be91SStefano Zampini } 55509d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 55519d30be91SStefano Zampini PetscInt j; 55529d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 55539d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 55549d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 55559d30be91SStefano Zampini } 55569d30be91SStefano Zampini } else { 55579d30be91SStefano Zampini /* TODO */ 55589d30be91SStefano Zampini } 55599d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 55609d30be91SStefano Zampini } 55619d30be91SStefano Zampini if (new_local_nnz) { 55629d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 55639d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 55649d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 55659d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 55669d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 55679d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 55689d30be91SStefano Zampini } else { 55699d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55709d30be91SStefano Zampini } 55719d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 55729d30be91SStefano Zampini } else { 55739d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55749d30be91SStefano Zampini } 5575e7931f94SStefano Zampini 5576e7931f94SStefano Zampini /* set values */ 5577e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 55789d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5579e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5580e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5581e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 55829d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5583e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5584e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5585e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 558628143c3dSStefano Zampini } else { 558728143c3dSStefano Zampini /* TODO */ 5588e7931f94SStefano Zampini } 5589e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5590e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5591e7931f94SStefano Zampini } 5592e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5593e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 559470cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 559570cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55969d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 55979d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5598e7931f94SStefano Zampini 5599dfd14d43SStefano Zampini #if 0 560028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5601e7931f94SStefano Zampini Vec lvec,rvec; 5602e7931f94SStefano Zampini PetscReal infty_error; 5603e7931f94SStefano Zampini 56042a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5605e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5606e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5607e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 560870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5609e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5610e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5611e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5612e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5613e7931f94SStefano Zampini } 561428143c3dSStefano Zampini #endif 5615e7931f94SStefano Zampini 561628143c3dSStefano Zampini /* assemble new additional is (if any) */ 561728143c3dSStefano Zampini if (nis) { 561828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 561928143c3dSStefano Zampini 562028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5621854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 562228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 562328143c3dSStefano Zampini psum = 0; 562428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 562528143c3dSStefano Zampini for (j=0;j<nis;j++) { 562628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 562728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 562828143c3dSStefano Zampini psum += plen; 562928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 563028143c3dSStefano Zampini } 563128143c3dSStefano Zampini } 5632854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5633854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 563428143c3dSStefano Zampini for (i=1;i<nis;i++) { 563528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 563628143c3dSStefano Zampini } 563728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 563828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 563928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 564028143c3dSStefano Zampini for (j=0;j<nis;j++) { 564128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 564228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 564328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 564428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 564528143c3dSStefano Zampini } 564628143c3dSStefano Zampini } 564728143c3dSStefano Zampini for (i=0;i<nis;i++) { 564828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 564928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 565028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 565128143c3dSStefano Zampini } 565228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 565328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 565428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 565528143c3dSStefano Zampini } 5656e7931f94SStefano Zampini /* free workspace */ 565728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5658e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5659e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5660e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5661e7931f94SStefano Zampini if (isdense) { 5662e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5663e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5664e7931f94SStefano Zampini } else { 5665e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5666e7931f94SStefano Zampini } 566728143c3dSStefano Zampini if (nis) { 566828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 566928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 567028143c3dSStefano Zampini } 5671e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5672e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 567328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5674e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5675e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 567628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5677e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5678e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5679e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5680e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5681e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 568228143c3dSStefano Zampini if (nis) { 568328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 568428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 568528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 568628143c3dSStefano Zampini } 568728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 568828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 568928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 569028143c3dSStefano Zampini for (i=0;i<nis;i++) { 569128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 569228143c3dSStefano Zampini } 569353a05cb3SStefano Zampini *mat_n = NULL; 569428143c3dSStefano Zampini } 5695e7931f94SStefano Zampini PetscFunctionReturn(0); 5696e7931f94SStefano Zampini } 5697a57a6d2fSStefano Zampini 569812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5699af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 570012edc857SStefano Zampini 5701c8587f34SStefano Zampini #undef __FUNCT__ 5702c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5703c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5704c8587f34SStefano Zampini { 5705c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5706c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 570720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 57089881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 570920a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 57106e683305SStefano Zampini IS coarse_is,*isarray; 57116e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 571230368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5713f9eb5b7dSStefano Zampini PC pc_temp; 5714c8587f34SStefano Zampini PCType coarse_pc_type; 5715c8587f34SStefano Zampini KSPType coarse_ksp_type; 5716f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 57174f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 57186e683305SStefano Zampini Mat t_coarse_mat_is; 571957de7509SStefano Zampini PetscInt ncoarse; 572068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 572122bc73bbSStefano Zampini PetscScalar *array; 572257de7509SStefano Zampini MatReuse coarse_mat_reuse; 572357de7509SStefano Zampini PetscBool restr, full_restr, have_void; 57249881197aSStefano Zampini PetscErrorCode ierr; 5725fdc09c96SStefano Zampini 5726c8587f34SStefano Zampini PetscFunctionBegin; 5727c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 572868457ee5SStefano Zampini if (pcbddc->new_primal_space || pcbddc->coarse_size == -1) { /* a new primal space is present or it is the first initialization, so recompute global numbering */ 5729fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 57305a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5731fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5732f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5733f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5734f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5735fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 573651bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 573751bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5738dc4bcba2SStefano Zampini PC pc; 5739dc4bcba2SStefano Zampini PetscBool isbddc; 5740dc4bcba2SStefano Zampini 5741dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5742dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5743dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5744dc4bcba2SStefano Zampini if (isbddc) { 574563c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 574663c961adSStefano Zampini } else { 5747727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 574863c961adSStefano Zampini } 5749fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5750fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5751fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5752f4ddd8eeSStefano Zampini } 5753fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5754fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5755f4ddd8eeSStefano Zampini } 575670cf5478SStefano Zampini /* reset any subassembling information */ 575757de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 575870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 575957de7509SStefano Zampini } 57606e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5761fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5762f4ddd8eeSStefano Zampini } 576357de7509SStefano Zampini /* assemble coarse matrix */ 576457de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 576557de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 576657de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 576757de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 576818a45a71SStefano Zampini } else { 576957de7509SStefano Zampini coarse_mat = NULL; 577057de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 57716e683305SStefano Zampini } 5772e7931f94SStefano Zampini 5773abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5774abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5775abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5776abbbba34SStefano Zampini 5777abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 577822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 577922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 578022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 578122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5782e176bc59SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,NULL,&t_coarse_mat_is);CHKERRQ(ierr); 57836e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 57846e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57856e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5786abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5787abbbba34SStefano Zampini 578857de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 578957de7509SStefano Zampini im_active = !!(pcis->n); 579057de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 579157de7509SStefano Zampini 579257de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 579357de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 579457de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 579557de7509SStefano Zampini coarse_mat_is = NULL; 579657de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 579757de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 579857de7509SStefano Zampini full_restr = PETSC_TRUE; 579957de7509SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(pcbddc->coarse_size,pcbddc->coarse_eqs_per_proc); 580057de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 580157de7509SStefano Zampini if (multilevel_requested) { 580257de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 580357de7509SStefano Zampini restr = PETSC_FALSE; 580457de7509SStefano Zampini full_restr = PETSC_FALSE; 580557de7509SStefano Zampini } else { 580657de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 580757de7509SStefano Zampini restr = PETSC_TRUE; 580857de7509SStefano Zampini full_restr = PETSC_TRUE; 580957de7509SStefano Zampini } 581057de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 581157de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 581257de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 581357de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 581457de7509SStefano Zampini PetscInt psum; 581557de7509SStefano Zampini PetscMPIInt size; 581657de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 581757de7509SStefano Zampini else psum = 0; 581857de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 581957de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 582057de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 582157de7509SStefano Zampini } 582257de7509SStefano Zampini /* determine if we can go multilevel */ 582357de7509SStefano Zampini if (multilevel_requested) { 582457de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 582557de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 582657de7509SStefano Zampini } 582757de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 582857de7509SStefano Zampini 5829e4d548c7SStefano Zampini /* dump subassembling pattern */ 5830e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5831e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5832e4d548c7SStefano Zampini } 5833e4d548c7SStefano Zampini 58346e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 583527b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 58366e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 58376e683305SStefano Zampini const PetscInt *idxs; 58386e683305SStefano Zampini ISLocalToGlobalMapping tmap; 58396e683305SStefano Zampini 58406e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 58410be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 58426e683305SStefano Zampini /* allocate space for temporary storage */ 5843854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5844854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 58456e683305SStefano Zampini /* allocate for IS array */ 58466e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 58476e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 584827b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 584930368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5850854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 58516e683305SStefano Zampini /* dofs splitting */ 58526e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 58536e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 58546e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 58556e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58566e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58576e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58586e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 585930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 58606e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 58616e683305SStefano Zampini } 58626e683305SStefano Zampini /* neumann boundaries */ 58636e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 58646e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 58656e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 58666e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58676e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58686e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58696e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 587030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 58716e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 58726e683305SStefano Zampini } 58736e683305SStefano Zampini /* free memory */ 58746e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 58756e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 58766e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 58776e683305SStefano Zampini } else { 58786e683305SStefano Zampini nis = 0; 58796e683305SStefano Zampini nisdofs = 0; 58806e683305SStefano Zampini nisneu = 0; 588130368db7SStefano Zampini nisvert = 0; 58826e683305SStefano Zampini isarray = NULL; 58836e683305SStefano Zampini } 58846e683305SStefano Zampini /* destroy no longer needed map */ 58856e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 58866e683305SStefano Zampini 588757de7509SStefano Zampini /* subassemble */ 588857de7509SStefano Zampini if (multilevel_allowed) { 588957de7509SStefano Zampini PetscBool reuse,reuser; 589057de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 589157de7509SStefano Zampini else reuse = PETSC_FALSE; 589257de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 589357de7509SStefano Zampini if (reuser) { 589457de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray);CHKERRQ(ierr); 589574e2c79eSStefano Zampini } else { 589657de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 589774e2c79eSStefano Zampini } 58981f4df5f7SStefano Zampini /* TODO: if (pcbddc->benign_have_null) -> give a hint to the coarser levels if they have to locally apply the benign trick or not */ 589974e2c79eSStefano Zampini } else { 590057de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 59016e683305SStefano Zampini } 590257de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 590357de7509SStefano Zampini PetscMPIInt size; 590457de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 590557de7509SStefano Zampini if (!multilevel_allowed) { 590657de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 59076e683305SStefano Zampini } else { 590857de7509SStefano Zampini Mat A; 5909779c1cceSStefano Zampini 591057de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 591157de7509SStefano Zampini if (coarse_mat_is) { 591257de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 591357de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 591457de7509SStefano Zampini coarse_mat = coarse_mat_is; 591557de7509SStefano Zampini } 591657de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 591757de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 591857de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5919779c1cceSStefano Zampini } 5920779c1cceSStefano Zampini } 592157de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 592257de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 59236e683305SStefano Zampini 59246e683305SStefano Zampini /* create local to global scatters for coarse problem */ 592568457ee5SStefano Zampini if (compute_vecs) { 59266e683305SStefano Zampini PetscInt lrows; 59276e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 592857de7509SStefano Zampini if (coarse_mat) { 592957de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 59306e683305SStefano Zampini } else { 59316e683305SStefano Zampini lrows = 0; 59326e683305SStefano Zampini } 59336e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 59346e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 59356e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 59366e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59376e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59386e683305SStefano Zampini } 59396e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5940c8587f34SStefano Zampini 5941f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5942f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5943f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5944f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5945f9eb5b7dSStefano Zampini } else { 5946f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5947f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5948c8587f34SStefano Zampini } 5949c8587f34SStefano Zampini 59506e683305SStefano Zampini /* print some info if requested */ 59516e683305SStefano Zampini if (pcbddc->dbg_flag) { 59526e683305SStefano Zampini if (!multilevel_allowed) { 59536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 59546e683305SStefano Zampini if (multilevel_requested) { 59556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Not enough active processes on level %d (active processes %d, coarsening ratio %d)\n",pcbddc->current_level,active_procs,pcbddc->coarsening_ratio);CHKERRQ(ierr); 59566e683305SStefano Zampini } else if (pcbddc->max_levels) { 59576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 59586e683305SStefano Zampini } 59596e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59606e683305SStefano Zampini } 59616e683305SStefano Zampini } 59626e683305SStefano Zampini 5963f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 596457de7509SStefano Zampini if (coarse_mat) { 59656a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 59666e683305SStefano Zampini if (pcbddc->dbg_flag) { 596757de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 59686e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 59696e683305SStefano Zampini } 5970f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5971312be037SStefano Zampini char prefix[256],str_level[16]; 5972e604994aSStefano Zampini size_t len; 597357de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5974422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5975c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5976f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 597757de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5978c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 59796e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5980c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5981c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5982e604994aSStefano Zampini /* prefix */ 5983e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5984e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5985e604994aSStefano Zampini if (!pcbddc->current_level) { 5986e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5987e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5988c8587f34SStefano Zampini } else { 5989e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5990312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5991312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 599234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5993312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5994e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5995e604994aSStefano Zampini } 5996e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 59973e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 59983e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 59993e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 60003e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 6001f9eb5b7dSStefano Zampini /* allow user customization */ 6002f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 60033e3c6dadSStefano Zampini } 60043e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 600551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 60063e3c6dadSStefano Zampini if (nisdofs) { 60073e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 60083e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 60093e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 60103e3c6dadSStefano Zampini } 60113e3c6dadSStefano Zampini } 60123e3c6dadSStefano Zampini if (nisneu) { 60133e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 60143e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 6015312be037SStefano Zampini } 601630368db7SStefano Zampini if (nisvert) { 601730368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 601830368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 601930368db7SStefano Zampini } 6020f9eb5b7dSStefano Zampini 6021f9eb5b7dSStefano Zampini /* get some info after set from options */ 6022f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 6023f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 60244f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 60256e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 6026f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6027f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 6028f9eb5b7dSStefano Zampini } 602939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 60304f3a063dSStefano Zampini if (isredundant) { 60314f3a063dSStefano Zampini KSP inner_ksp; 60324f3a063dSStefano Zampini PC inner_pc; 60334f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 60344f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 60354f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 60364f3a063dSStefano Zampini } 6037f9eb5b7dSStefano Zampini 603857de7509SStefano Zampini /* parameters which miss an API */ 603957de7509SStefano Zampini if (isbddc) { 6040720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 6041720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 604257de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 604327b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 604427b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 604527b6a85dSStefano Zampini pcbddc_coarse->benign_compute_nonetflux = PETSC_TRUE; 6046720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 604759e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 6048720d30f9SStefano Zampini } 6049d4d8cf7bSStefano Zampini } 60509881197aSStefano Zampini 60513301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 60525a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 60533301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 60543301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 60553301b35fSStefano Zampini } 60563301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 60573301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 60583301b35fSStefano Zampini } 60593301b35fSStefano Zampini if (pc->pmat->spd_set) { 60603301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 60613301b35fSStefano Zampini } 606227b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 606327b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 606427b6a85dSStefano Zampini } 60656e683305SStefano Zampini /* set operators */ 60665f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 60676e683305SStefano Zampini if (pcbddc->dbg_flag) { 60686e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 60696e683305SStefano Zampini } 60706e683305SStefano Zampini } 60716e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6072b1ecc7b1SStefano Zampini #if 0 6073b9b85e73SStefano Zampini { 6074b9b85e73SStefano Zampini PetscViewer viewer; 6075b9b85e73SStefano Zampini char filename[256]; 6076b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6077b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 60786a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6079b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6080f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6081b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6082b9b85e73SStefano Zampini } 6083b9b85e73SStefano Zampini #endif 6084f9eb5b7dSStefano Zampini 608598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 608698a51de6SStefano Zampini Vec crhs,csol; 608704708bb6SStefano Zampini 6088f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6089f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6090f347579bSStefano Zampini if (!csol) { 60912a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6092f9eb5b7dSStefano Zampini } 6093f347579bSStefano Zampini if (!crhs) { 60942a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6095f347579bSStefano Zampini } 6096b0f5fe93SStefano Zampini } 6097b0f5fe93SStefano Zampini 6098b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6099b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6100b0f5fe93SStefano Zampini 6101b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 61024f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 61034f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 61044f1b2e48SStefano Zampini } 6105b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6106b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6107b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6108b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6109b0f5fe93SStefano Zampini if (coarse_mat) { 6110b0f5fe93SStefano Zampini Vec nullv; 6111b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6112b0f5fe93SStefano Zampini PetscInt nl; 6113b0f5fe93SStefano Zampini 6114b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6115b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6116b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6117b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6118b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6119b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6120b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6121b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6122b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6123b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6124b0f5fe93SStefano Zampini } 6125b0f5fe93SStefano Zampini } 6126b0f5fe93SStefano Zampini 6127b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6128b0f5fe93SStefano Zampini PetscBool ispreonly; 6129b0f5fe93SStefano Zampini 6130b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6131b0f5fe93SStefano Zampini PetscBool isnull; 6132b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6133bef83e63SStefano Zampini if (isnull) { 6134b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6135b0f5fe93SStefano Zampini } 6136bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6137b0f5fe93SStefano Zampini } 6138b0f5fe93SStefano Zampini /* setup coarse ksp */ 6139b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6140cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6141cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 61426e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6143c8587f34SStefano Zampini KSP check_ksp; 61442b510759SStefano Zampini KSPType check_ksp_type; 6145c8587f34SStefano Zampini PC check_pc; 61466e683305SStefano Zampini Vec check_vec,coarse_vec; 61476a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 61482b510759SStefano Zampini PetscInt its; 61496e683305SStefano Zampini PetscBool compute_eigs; 61506e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 61516e683305SStefano Zampini PetscInt neigs; 61528e185a42SStefano Zampini const char *prefix; 6153c8587f34SStefano Zampini 61542b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 61556e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6156422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 615723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6158f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6159e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6160e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6161e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 61622b510759SStefano Zampini if (ispreonly) { 61632b510759SStefano Zampini check_ksp_type = KSPPREONLY; 61646e683305SStefano Zampini compute_eigs = PETSC_FALSE; 61652b510759SStefano Zampini } else { 6166cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 61676e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6168c8587f34SStefano Zampini } 6169c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 61706e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 61716e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 61726e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6173a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6174a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6175a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6176a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6177c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6178c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6179c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6180c8587f34SStefano Zampini /* create random vec */ 61812701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6182c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 61836e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6184c8587f34SStefano Zampini /* solve coarse problem */ 61856e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6186cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 61876e683305SStefano Zampini if (compute_eigs) { 6188854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6189854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 61906e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 61916e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 61926e683305SStefano Zampini lambda_min = eigs_r[0]; 61936e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 61942701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 61952701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6196cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6197cbcc2c2aSStefano Zampini } 6198c8587f34SStefano Zampini } 6199c8587f34SStefano Zampini } 6200cbcc2c2aSStefano Zampini 6201c8587f34SStefano Zampini /* check coarse problem residual error */ 62026e683305SStefano Zampini if (pcbddc->dbg_flag) { 62036e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 62046e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62056e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6206c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 62076e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 62086e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6209779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 62106e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 62116e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 62126e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 62136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6214b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6215b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6216b0f5fe93SStefano Zampini } 62176e683305SStefano Zampini if (compute_eigs) { 62186e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6219deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6220c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 62216e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 62226e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem eigenvalues (estimated with %d iterations of %s): %1.6e %1.6e (%1.6e %1.6e)\n",its,check_ksp_type,lambda_min,lambda_max,lambda_min_s,lambda_max_s);CHKERRQ(ierr); 62236e683305SStefano Zampini for (i=0;i<neigs;i++) { 62246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6225c8587f34SStefano Zampini } 62266e683305SStefano Zampini } 62276e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 62286e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62296e683305SStefano Zampini } 6230e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 62312701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6232c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 62336e683305SStefano Zampini if (compute_eigs) { 62346e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 62356e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6236c8587f34SStefano Zampini } 62376e683305SStefano Zampini } 62386e683305SStefano Zampini } 6239bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6240cbcc2c2aSStefano Zampini /* print additional info */ 6241cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 62426e683305SStefano Zampini /* waits until all processes reaches this point */ 62436e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6244cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6245cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6246cbcc2c2aSStefano Zampini } 6247cbcc2c2aSStefano Zampini 62482b510759SStefano Zampini /* free memory */ 6249fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6250c8587f34SStefano Zampini PetscFunctionReturn(0); 6251c8587f34SStefano Zampini } 6252674ae819SStefano Zampini 6253f34684f1SStefano Zampini #undef __FUNCT__ 6254f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6255f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6256f34684f1SStefano Zampini { 6257f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6258f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6259f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6260dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6261dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 626273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6263dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6264f34684f1SStefano Zampini PetscErrorCode ierr; 6265f34684f1SStefano Zampini 6266f34684f1SStefano Zampini PetscFunctionBegin; 6267f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 62686c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6269dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 62703bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6271dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6272dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6273dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6274dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6275dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6276dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 62776c4ed002SBarry Smith if (local_size != pcbddc->local_primal_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %D != %D",local_size,pcbddc->local_primal_size); 6278dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6279dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6280dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6281dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6282dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6283f34684f1SStefano Zampini 6284f34684f1SStefano Zampini /* check numbering */ 6285f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6286019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6287dc456d91SStefano Zampini PetscInt i; 6288b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6289f34684f1SStefano Zampini 6290f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6291f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6292f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 62931575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6294019a44ceSStefano Zampini /* counter */ 6295019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6296019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6297019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6298019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6299019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6300019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6301f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6302f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6303727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6304f34684f1SStefano Zampini } 6305f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6306f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6307f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6308e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6309e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6310e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6311e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6312f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6313019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6314f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6315019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 63162c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 631775c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6318b9b85e73SStefano Zampini set_error = PETSC_TRUE; 63192c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 63202c66d082SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d (gid %d) owned by %d processes instead of %d!\n",PetscGlobalRank,i,gi,owned,neigh);CHKERRQ(ierr); 6321f34684f1SStefano Zampini } 6322f34684f1SStefano Zampini } 6323019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6324b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6325f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6326f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6327f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6328f34684f1SStefano Zampini } 6329f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6330f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6331e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6332e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6333f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6334f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6335b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6336ca8b9ea9SStefano Zampini PetscInt *gidxs; 6337ca8b9ea9SStefano Zampini 6338ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 63393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6340f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6341f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6342f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6343f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 63444bc2dc4bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_primal_indices[%d]=%d (%d,%d)\n",i,local_primal_indices[i],pcbddc->primal_indices_local_idxs[i],gidxs[i]);CHKERRQ(ierr); 6345f34684f1SStefano Zampini } 6346f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6347ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6348f34684f1SStefano Zampini } 6349f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 63501575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6351302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6352f34684f1SStefano Zampini } 63538bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6354f34684f1SStefano Zampini /* get back data */ 6355f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6356f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6357674ae819SStefano Zampini PetscFunctionReturn(0); 6358674ae819SStefano Zampini } 6359674ae819SStefano Zampini 6360e456f2a8SStefano Zampini #undef __FUNCT__ 6361e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6362a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6363e456f2a8SStefano Zampini { 6364e456f2a8SStefano Zampini IS localis_t; 6365a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6366e456f2a8SStefano Zampini PetscScalar *vals; 6367e456f2a8SStefano Zampini PetscErrorCode ierr; 6368e456f2a8SStefano Zampini 6369e456f2a8SStefano Zampini PetscFunctionBegin; 6370a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6371e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6372854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6373e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6374e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6375a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6376a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 63771035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6378a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 63791035eff8SStefano Zampini } 6380a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6381e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6382e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6383a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6384a7dc3881SStefano Zampini /* now compute set in local ordering */ 6385a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6386a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6387a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6388a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6389a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6390ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6391e456f2a8SStefano Zampini lsize++; 6392e456f2a8SStefano Zampini } 6393e456f2a8SStefano Zampini } 6394854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6395a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6396ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6397e456f2a8SStefano Zampini idxs[lsize++] = i; 6398e456f2a8SStefano Zampini } 6399e456f2a8SStefano Zampini } 6400a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6401a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6402e456f2a8SStefano Zampini *localis = localis_t; 6403e456f2a8SStefano Zampini PetscFunctionReturn(0); 6404e456f2a8SStefano Zampini } 6405906d46d4SStefano Zampini 6406b96c3477SStefano Zampini #undef __FUNCT__ 6407b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 640808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6409b96c3477SStefano Zampini { 6410a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6411b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6412b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6413a64f4aa4SStefano Zampini Mat S_j; 6414b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6415b96c3477SStefano Zampini PetscBool free_used_adj; 6416b96c3477SStefano Zampini PetscErrorCode ierr; 6417b96c3477SStefano Zampini 6418b96c3477SStefano Zampini PetscFunctionBegin; 6419b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6420b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 642108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6422b96c3477SStefano Zampini used_xadj = NULL; 6423b96c3477SStefano Zampini used_adjncy = NULL; 6424b96c3477SStefano Zampini } else { 642508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 642608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 642708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 642808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6429b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6430b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6431b96c3477SStefano Zampini } else { 64322fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6433b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6434b96c3477SStefano Zampini PetscInt nvtxs; 6435b96c3477SStefano Zampini 64362fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 64372fffb893SStefano Zampini if (flg_row) { 6438b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6439b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6440b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6441b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 64422fffb893SStefano Zampini } else { 64432fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 64442fffb893SStefano Zampini used_xadj = NULL; 64452fffb893SStefano Zampini used_adjncy = NULL; 64462fffb893SStefano Zampini } 64472fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6448b96c3477SStefano Zampini } 6449b96c3477SStefano Zampini } 6450d5574798SStefano Zampini 6451d5574798SStefano Zampini /* setup sub_schurs data */ 6452a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6453df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6454df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6455a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 645691af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6457a64f4aa4SStefano Zampini } else { 64586816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6459b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6460a3df083aSStefano Zampini PetscInt benign_n; 646172b8c272SStefano Zampini Mat change = NULL; 64629d54b7f4SStefano Zampini Vec scaling = NULL; 646372b8c272SStefano Zampini IS change_primal = NULL; 6464a3df083aSStefano Zampini 64655feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 64665feab87aSStefano Zampini PetscInt n_vertices; 64675feab87aSStefano Zampini 64685feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 64692034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 64705feab87aSStefano Zampini } 647104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 647204708bb6SStefano Zampini if (!isseqaij) { 647304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 647404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 647504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 647604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 647704708bb6SStefano Zampini } else { 6478511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 647904708bb6SStefano Zampini } 648004708bb6SStefano Zampini } 6481a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6482a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6483ca92afb2SStefano Zampini } else { 6484a3df083aSStefano Zampini benign_n = 0; 6485ca92afb2SStefano Zampini } 6486b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6487b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6488b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 648972b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 649022db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6491b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 649222db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6493b7ab4a40SStefano Zampini } 6494b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6495b7ab4a40SStefano Zampini We need to compute the change of basis according to the quadrature weights attached to pmat via MatSetNearNullSpace, and this could not be done (at the moment) without some hacking */ 6496b7ab4a40SStefano Zampini if (need_change) { 649788c03ad3SStefano Zampini PC_IS *pcisf; 649888c03ad3SStefano Zampini PC_BDDC *pcbddcf; 649988c03ad3SStefano Zampini PC pcf; 650088c03ad3SStefano Zampini 6501e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 650288c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 650388c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 650488c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 650588c03ad3SStefano Zampini /* hacks */ 650688c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 650772b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 650872b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 650972b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 651072b8c272SStefano Zampini pcisf->n = pcis->n; 651172b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 651288c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 651388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 651488c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 651588c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 651688c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 651788c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 651872b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 651988c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 652088c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 652172b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 652272b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 652372b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 652472b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 652572b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 652688c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 652772b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 652888c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 652988c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 653088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 653188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 653288c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 653388c03ad3SStefano Zampini pcf->ops->destroy = NULL; 653488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 653588c03ad3SStefano Zampini } 65369d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 653791af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 653872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 653972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6540ca92afb2SStefano Zampini } 6541d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6542b96c3477SStefano Zampini 6543b96c3477SStefano Zampini /* free adjacency */ 6544b96c3477SStefano Zampini if (free_used_adj) { 6545b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6546b96c3477SStefano Zampini } 6547b96c3477SStefano Zampini PetscFunctionReturn(0); 6548b96c3477SStefano Zampini } 6549b96c3477SStefano Zampini 6550b96c3477SStefano Zampini #undef __FUNCT__ 6551b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 655208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6553b96c3477SStefano Zampini { 6554b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6555b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6556b96c3477SStefano Zampini PCBDDCGraph graph; 6557b96c3477SStefano Zampini PetscErrorCode ierr; 6558b96c3477SStefano Zampini 6559b96c3477SStefano Zampini PetscFunctionBegin; 6560b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 656108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 65623301b35fSStefano Zampini IS verticesIS,verticescomm; 65633301b35fSStefano Zampini PetscInt vsize,*idxs; 6564b96c3477SStefano Zampini 6565b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 65663301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 65673301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 65683301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 65693301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 65703301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6571b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 65727fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6573441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 65743301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6575b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6576b96c3477SStefano Zampini } else { 6577b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6578b96c3477SStefano Zampini } 6579e4d548c7SStefano Zampini /* print some info */ 6580e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6581e4d548c7SStefano Zampini IS vertices; 6582e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 6583e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6584e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6585e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6586e4d548c7SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 6587e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6588e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6589e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6590e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6591e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6592e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6593e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6594e4d548c7SStefano Zampini } 6595b96c3477SStefano Zampini 6596b96c3477SStefano Zampini /* sub_schurs init */ 6597b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 6598b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 6599b334f244SStefano Zampini } 6600b334f244SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6601a64f4aa4SStefano Zampini 6602b96c3477SStefano Zampini /* free graph struct */ 660308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6604b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6605b96c3477SStefano Zampini } 6606b96c3477SStefano Zampini PetscFunctionReturn(0); 6607b96c3477SStefano Zampini } 6608fa34dd3eSStefano Zampini 6609fa34dd3eSStefano Zampini #undef __FUNCT__ 6610fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6611fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6612fa34dd3eSStefano Zampini { 6613fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6614fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6615fa34dd3eSStefano Zampini PetscErrorCode ierr; 6616fa34dd3eSStefano Zampini 6617fa34dd3eSStefano Zampini PetscFunctionBegin; 6618fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6619fa34dd3eSStefano Zampini IS zerodiag = NULL; 66204f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6621fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 66224f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 662375c01103SStefano Zampini PetscReal norm; 6624fa34dd3eSStefano Zampini PetscInt i; 6625fa34dd3eSStefano Zampini 6626fa34dd3eSStefano Zampini /* B0 and B0_B */ 6627fa34dd3eSStefano Zampini if (zerodiag) { 6628fa34dd3eSStefano Zampini IS dummy; 6629fa34dd3eSStefano Zampini 66304f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 66314f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6632fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6633fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6634fa34dd3eSStefano Zampini } 6635fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6636fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6637fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6638fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6639fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6640fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6641fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6642fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6643fa34dd3eSStefano Zampini /* S_j */ 6644fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6645fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6646fa34dd3eSStefano Zampini 6647fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6648fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6649fa34dd3eSStefano Zampini /* continuous in primal space */ 6650fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6651fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6652fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6653fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 66544f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 66554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6656fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6657fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6658fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6659fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6660fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6661fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6662fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6663fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6664fa34dd3eSStefano Zampini 6665fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6666fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6667fa34dd3eSStefano Zampini /* local with Schur */ 6668fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6669fa34dd3eSStefano Zampini if (zerodiag) { 6670fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 66714f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6672fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6673fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6674fa34dd3eSStefano Zampini } 6675fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6676fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6677fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6678fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6679fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6680fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6681fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6682fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6683fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6684fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6685fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6686fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6687fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6688fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6689fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6690fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6691fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6692fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6693fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6694fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6695fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6696fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6697fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6698fa34dd3eSStefano Zampini if (zerodiag) { 6699fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6700fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67014f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6702fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6703fa34dd3eSStefano Zampini } 6704fa34dd3eSStefano Zampini /* BDDC */ 6705fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6706fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6707fa34dd3eSStefano Zampini 6708fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6709fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6710fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6711fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 67124f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 67134f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6714fa34dd3eSStefano Zampini } 67154f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6716fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6717fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6718fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6719fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6720fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6721fa34dd3eSStefano Zampini } 6722fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6723fa34dd3eSStefano Zampini } 6724