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 7669cc0f4SStefano Zampini 8669cc0f4SStefano Zampini #undef __FUNCT__ 9669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 10669cc0f4SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PCBDDCGraph graph, MatNullSpace *nnsp) 11669cc0f4SStefano Zampini { 12669cc0f4SStefano Zampini PetscErrorCode ierr; 13669cc0f4SStefano Zampini ISLocalToGlobalMapping rmap; 14669cc0f4SStefano Zampini IS *faces,*edges; 15669cc0f4SStefano Zampini Vec p,v,quad_vec,*quad_vecs; 16669cc0f4SStefano Zampini PetscScalar *vals; 17669cc0f4SStefano Zampini const PetscScalar *array; 18669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 191ae86dd6SStefano Zampini PetscMPIInt rank; 20669cc0f4SStefano Zampini 21669cc0f4SStefano Zampini PetscFunctionBegin; 22669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 23669cc0f4SStefano Zampini if (graph->twodim) { 24669cc0f4SStefano Zampini lmaxneighs = 2; 25669cc0f4SStefano Zampini } else { 26669cc0f4SStefano Zampini lmaxneighs = 1; 27669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 28669cc0f4SStefano Zampini const PetscInt *idxs; 29669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 30669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 31669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 32669cc0f4SStefano Zampini } 33669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 34669cc0f4SStefano Zampini } 35669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 36669cc0f4SStefano Zampini maxsize = 0; 37669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 38669cc0f4SStefano Zampini PetscInt nn; 39669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 40669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 41669cc0f4SStefano Zampini } 42669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 43669cc0f4SStefano Zampini PetscInt nn; 44669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 45669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 46669cc0f4SStefano Zampini } 47669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 48669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 49669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 50669cc0f4SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&rmap,NULL);CHKERRQ(ierr); 51669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 521ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 53669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 54669cc0f4SStefano Zampini PetscInt first,last; 55669cc0f4SStefano Zampini 56669cc0f4SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],rmap);CHKERRQ(ierr); 57669cc0f4SStefano Zampini /* the near-null space fo BDDC carries information on quadrature weights, 58669cc0f4SStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 59669cc0f4SStefano Zampini and create a suitable set of basis vectors first */ 60669cc0f4SStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 61669cc0f4SStefano Zampini if (i>=first && i < last) { 62669cc0f4SStefano Zampini PetscScalar *data; 63669cc0f4SStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 64669cc0f4SStefano Zampini data[i-first] = 1.; 65669cc0f4SStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 66669cc0f4SStefano Zampini } 67669cc0f4SStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 68669cc0f4SStefano Zampini } 69669cc0f4SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 701ae86dd6SStefano Zampini for (i=0;i<maxneighs;i++) { /* reset vectors */ 711ae86dd6SStefano Zampini PetscInt first,last; 721ae86dd6SStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 731ae86dd6SStefano Zampini if (i>=first && i < last) { 741ae86dd6SStefano Zampini PetscScalar *data; 751ae86dd6SStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 761ae86dd6SStefano Zampini data[i-first] = 0.; 771ae86dd6SStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 781ae86dd6SStefano Zampini } 791ae86dd6SStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 801ae86dd6SStefano Zampini } 81669cc0f4SStefano Zampini /* compute local quad vec */ 82669cc0f4SStefano Zampini ierr = MatCreateVecs(divudotp,&v,&p);CHKERRQ(ierr); 83669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 84669cc0f4SStefano Zampini ierr = MatMultTranspose(divudotp,p,v);CHKERRQ(ierr); 85669cc0f4SStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 86669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 871ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 881ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 89669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 90669cc0f4SStefano Zampini const PetscInt *idxs; 91669cc0f4SStefano Zampini PetscInt idx,nn,j; 92669cc0f4SStefano Zampini 93669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 94669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 95669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 961ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 97669cc0f4SStefano Zampini idx = -(idx+1); 98669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 99669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 100669cc0f4SStefano Zampini ierr = ISDestroy(&faces[i]);CHKERRQ(ierr); 101669cc0f4SStefano Zampini } 102669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 103669cc0f4SStefano Zampini const PetscInt *idxs; 104669cc0f4SStefano Zampini PetscInt idx,nn,j; 105669cc0f4SStefano Zampini 106669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 107669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 108669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 1091ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 110669cc0f4SStefano Zampini idx = -(idx+1); 111669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 112669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 113669cc0f4SStefano Zampini ierr = ISDestroy(&edges[i]);CHKERRQ(ierr); 114669cc0f4SStefano Zampini } 115669cc0f4SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 116669cc0f4SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 117669cc0f4SStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 118669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 119669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 120669cc0f4SStefano Zampini 121669cc0f4SStefano Zampini /* assemble near null space */ 122669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 123669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 124669cc0f4SStefano Zampini } 125669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 126669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 127669cc0f4SStefano Zampini } 128669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 129669cc0f4SStefano Zampini PetscFunctionReturn(0); 130669cc0f4SStefano Zampini } 131669cc0f4SStefano Zampini 132669cc0f4SStefano Zampini 133a3df083aSStefano Zampini #undef __FUNCT__ 1341f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 1351f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 1361f4df5f7SStefano Zampini { 1371f4df5f7SStefano Zampini PetscErrorCode ierr; 1381f4df5f7SStefano Zampini Vec local,global; 1391f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1401f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1411f4df5f7SStefano Zampini 1421f4df5f7SStefano Zampini PetscFunctionBegin; 1431f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 1441f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 1451f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 1461f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 1471f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 1481f4df5f7SStefano Zampini PetscInt i; 1491f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1501f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 1511f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1521f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 1531f4df5f7SStefano Zampini } 1541f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 1551f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 1561f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 1571f4df5f7SStefano Zampini } 1581f4df5f7SStefano Zampini } else { 159986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 1601f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 161986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 162986cdee1SStefano Zampini if (i > 1) { 163986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 1641f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1651f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 1661f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1671f4df5f7SStefano Zampini } 1681f4df5f7SStefano Zampini } 1691f4df5f7SStefano Zampini } 170986cdee1SStefano Zampini } 1711f4df5f7SStefano Zampini 1721f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 1731f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 1741f4df5f7SStefano Zampini } 1751f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 1761f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1771f4df5f7SStefano Zampini } 1781f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 1791f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1801f4df5f7SStefano Zampini } 1811f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 1821f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 1831f4df5f7SStefano Zampini PetscFunctionReturn(0); 1841f4df5f7SStefano Zampini } 1851f4df5f7SStefano Zampini 1861f4df5f7SStefano Zampini #undef __FUNCT__ 1873e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 1883e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 1893e589ea0SStefano Zampini { 1903e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 1913e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 1923e589ea0SStefano Zampini PetscErrorCode ierr; 1933e589ea0SStefano Zampini 1943e589ea0SStefano Zampini PetscFunctionBegin; 1953e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 1963e589ea0SStefano Zampini PetscFunctionReturn(0); 1973e589ea0SStefano Zampini } 1983e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1993e589ea0SStefano Zampini Vec swap; 2003e589ea0SStefano Zampini 2013e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 2023e589ea0SStefano Zampini swap = pcbddc->work_change; 2033e589ea0SStefano Zampini pcbddc->work_change = r; 2043e589ea0SStefano Zampini r = swap; 2053e589ea0SStefano Zampini } 2063e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2073e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2083e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 2093e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 2103e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2113e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2123e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2133e589ea0SStefano Zampini Vec swap; 2143e589ea0SStefano Zampini 2153e589ea0SStefano Zampini swap = r; 2163e589ea0SStefano Zampini r = pcbddc->work_change; 2173e589ea0SStefano Zampini pcbddc->work_change = swap; 2183e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 2193e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 2203e589ea0SStefano Zampini } 2213e589ea0SStefano Zampini PetscFunctionReturn(0); 2223e589ea0SStefano Zampini } 2233e589ea0SStefano Zampini 2243e589ea0SStefano Zampini #undef __FUNCT__ 225a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 226a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 227a3df083aSStefano Zampini { 228a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 229a3df083aSStefano Zampini PetscErrorCode ierr; 230a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 231a3df083aSStefano Zampini 232a3df083aSStefano Zampini PetscFunctionBegin; 233a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 234a3df083aSStefano Zampini if (transpose) { 235a3df083aSStefano Zampini apply_right = ctx->apply_left; 236a3df083aSStefano Zampini apply_left = ctx->apply_right; 237a3df083aSStefano Zampini } else { 238a3df083aSStefano Zampini apply_right = ctx->apply_right; 239a3df083aSStefano Zampini apply_left = ctx->apply_left; 240a3df083aSStefano Zampini } 241a3df083aSStefano Zampini reset_x = PETSC_FALSE; 242a3df083aSStefano Zampini if (apply_right) { 243a3df083aSStefano Zampini const PetscScalar *ax; 244a3df083aSStefano Zampini PetscInt nl,i; 245a3df083aSStefano Zampini 246a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 247a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 248a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 249a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 250a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 251a3df083aSStefano Zampini PetscScalar sum,val; 252a3df083aSStefano Zampini const PetscInt *idxs; 253a3df083aSStefano Zampini PetscInt nz,j; 254a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 255a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 256a3df083aSStefano Zampini sum = 0.; 257a3df083aSStefano Zampini if (ctx->apply_p0) { 258a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 259a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 260a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 261a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 262a3df083aSStefano Zampini } 263a3df083aSStefano Zampini } else { 264a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 265a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 266a3df083aSStefano Zampini } 267a3df083aSStefano Zampini } 268a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 269a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 270a3df083aSStefano Zampini } 271a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 272a3df083aSStefano Zampini reset_x = PETSC_TRUE; 273a3df083aSStefano Zampini } 274a3df083aSStefano Zampini if (transpose) { 275a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 276a3df083aSStefano Zampini } else { 277a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 278a3df083aSStefano Zampini } 279a3df083aSStefano Zampini if (reset_x) { 280a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 281a3df083aSStefano Zampini } 282a3df083aSStefano Zampini if (apply_left) { 283a3df083aSStefano Zampini PetscScalar *ay; 284a3df083aSStefano Zampini PetscInt i; 285a3df083aSStefano Zampini 286a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 287a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 288a3df083aSStefano Zampini PetscScalar sum,val; 289a3df083aSStefano Zampini const PetscInt *idxs; 290a3df083aSStefano Zampini PetscInt nz,j; 291a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 292a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 293a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 294a3df083aSStefano Zampini if (ctx->apply_p0) { 295a3df083aSStefano Zampini sum = 0.; 296a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 297a3df083aSStefano Zampini sum += ay[idxs[j]]; 298a3df083aSStefano Zampini ay[idxs[j]] += val; 299a3df083aSStefano Zampini } 300a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 301a3df083aSStefano Zampini } else { 302a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 303a3df083aSStefano Zampini ay[idxs[j]] += val; 304a3df083aSStefano Zampini } 305a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 306a3df083aSStefano Zampini } 307a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 308a3df083aSStefano Zampini } 309a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 310a3df083aSStefano Zampini } 311a3df083aSStefano Zampini PetscFunctionReturn(0); 312a3df083aSStefano Zampini } 313a3df083aSStefano Zampini 314a3df083aSStefano Zampini #undef __FUNCT__ 315a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 316a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 317a3df083aSStefano Zampini { 318a3df083aSStefano Zampini PetscErrorCode ierr; 319a3df083aSStefano Zampini 320a3df083aSStefano Zampini PetscFunctionBegin; 321a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 322a3df083aSStefano Zampini PetscFunctionReturn(0); 323a3df083aSStefano Zampini } 324a3df083aSStefano Zampini 325a3df083aSStefano Zampini #undef __FUNCT__ 326a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 327a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 328a3df083aSStefano Zampini { 329a3df083aSStefano Zampini PetscErrorCode ierr; 330a3df083aSStefano Zampini 331a3df083aSStefano Zampini PetscFunctionBegin; 332a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 333a3df083aSStefano Zampini PetscFunctionReturn(0); 334a3df083aSStefano Zampini } 335a3df083aSStefano Zampini 336a3df083aSStefano Zampini #undef __FUNCT__ 337a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 338a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 339a3df083aSStefano Zampini { 340a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 341a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 342a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 343a3df083aSStefano Zampini PetscErrorCode ierr; 344a3df083aSStefano Zampini 345a3df083aSStefano Zampini PetscFunctionBegin; 346a3df083aSStefano Zampini if (!restore) { 3471dd7afcfSStefano Zampini Mat A_IB,A_BI; 348a3df083aSStefano Zampini PetscScalar *work; 349b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 350a3df083aSStefano Zampini 3511dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 3521dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 3531dd7afcfSStefano Zampini } 3541dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 3551dd7afcfSStefano Zampini PetscFunctionReturn(0); 3561dd7afcfSStefano Zampini } 357a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 358a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 359a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 360a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 361a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 362a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 363a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 364a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 365a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 366a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 367a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 368a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 369059032f7SStefano Zampini if (reuse) { 370a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 3711dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 372059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 373059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 374059032f7SStefano Zampini PetscInt i; 375059032f7SStefano Zampini 376059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 377059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 378059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 379059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 380059032f7SStefano Zampini } 381059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 3821dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 383059032f7SStefano Zampini } 384a3df083aSStefano Zampini ctx->A = pcis->A_IB; 385a3df083aSStefano Zampini ctx->work = work; 386a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 387a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 388a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 389a3df083aSStefano Zampini pcis->A_IB = A_IB; 390a3df083aSStefano Zampini 391a3df083aSStefano Zampini /* A_BI as A_IB^T */ 392a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 393a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 394a3df083aSStefano Zampini pcis->A_BI = A_BI; 395a3df083aSStefano Zampini } else { 3961dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 3971dd7afcfSStefano Zampini PetscFunctionReturn(0); 3981dd7afcfSStefano Zampini } 399a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 400a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 401a3df083aSStefano Zampini pcis->A_IB = ctx->A; 4021dd7afcfSStefano Zampini ctx->A = NULL; 4031dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 4041dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 4051dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 4061dd7afcfSStefano Zampini if (ctx->free) { 407059032f7SStefano Zampini PetscInt i; 4081dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 409059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 410059032f7SStefano Zampini } 411059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 412059032f7SStefano Zampini } 413a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 414a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 415a3df083aSStefano Zampini } 416a3df083aSStefano Zampini PetscFunctionReturn(0); 417a3df083aSStefano Zampini } 418a3df083aSStefano Zampini 419a3df083aSStefano Zampini /* used just in bddc debug mode */ 420a3df083aSStefano Zampini #undef __FUNCT__ 421a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 422a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 423a3df083aSStefano Zampini { 424a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 425a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 426a3df083aSStefano Zampini Mat An; 427a3df083aSStefano Zampini PetscErrorCode ierr; 428a3df083aSStefano Zampini 429a3df083aSStefano Zampini PetscFunctionBegin; 430a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 431a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 432a3df083aSStefano Zampini if (is1) { 433a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 434a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 435a3df083aSStefano Zampini } else { 436a3df083aSStefano Zampini *B = An; 437a3df083aSStefano Zampini } 438a3df083aSStefano Zampini PetscFunctionReturn(0); 439a3df083aSStefano Zampini } 440a3df083aSStefano Zampini 4411cf9b237SStefano Zampini /* TODO: add reuse flag */ 4421cf9b237SStefano Zampini #undef __FUNCT__ 4431cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 4441cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 4451cf9b237SStefano Zampini { 4461cf9b237SStefano Zampini Mat Bt; 4471cf9b237SStefano Zampini PetscScalar *a,*bdata; 4481cf9b237SStefano Zampini const PetscInt *ii,*ij; 4491cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 4501cf9b237SStefano Zampini PetscBool flg_row; 4511cf9b237SStefano Zampini PetscErrorCode ierr; 4521cf9b237SStefano Zampini 4531cf9b237SStefano Zampini PetscFunctionBegin; 4541cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 4551cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4561cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 4571cf9b237SStefano Zampini nnz = n; 4581cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 4591cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 4601cf9b237SStefano Zampini } 4611cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 4621cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 4631cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 4641cf9b237SStefano Zampini nnz = 0; 4651cf9b237SStefano Zampini bii[0] = 0; 4661cf9b237SStefano Zampini for (i=0;i<n;i++) { 4671cf9b237SStefano Zampini PetscInt j; 4681cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 4691cf9b237SStefano Zampini PetscScalar entry = a[j]; 4701cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 4711cf9b237SStefano Zampini bij[nnz] = ij[j]; 4721cf9b237SStefano Zampini bdata[nnz] = entry; 4731cf9b237SStefano Zampini nnz++; 4741cf9b237SStefano Zampini } 4751cf9b237SStefano Zampini } 4761cf9b237SStefano Zampini bii[i+1] = nnz; 4771cf9b237SStefano Zampini } 4781cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 4791cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 4801cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4811cf9b237SStefano Zampini { 4821cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 4831cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 4841cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 4851cf9b237SStefano Zampini } 4861cf9b237SStefano Zampini *B = Bt; 4871cf9b237SStefano Zampini PetscFunctionReturn(0); 4881cf9b237SStefano Zampini } 4891cf9b237SStefano Zampini 490674ae819SStefano Zampini #undef __FUNCT__ 4914f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 4924f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 4934f1b2e48SStefano Zampini { 4944f1b2e48SStefano Zampini Mat B; 4954f1b2e48SStefano Zampini IS is_dummy,*cc_n; 4964f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 4974f1b2e48SStefano Zampini PCBDDCGraph graph; 4984f1b2e48SStefano Zampini PetscInt i,n; 4994f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 5004f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 5014f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 5024f1b2e48SStefano Zampini PetscErrorCode ierr; 5034f1b2e48SStefano Zampini 5044f1b2e48SStefano Zampini PetscFunctionBegin; 50563c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 50663c961adSStefano Zampini *ncc = 0; 50763c961adSStefano Zampini *cc = NULL; 50863c961adSStefano Zampini PetscFunctionReturn(0); 50963c961adSStefano Zampini } 5104f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 5114f1b2e48SStefano Zampini if (!isseqaij && filter) { 5121cf9b237SStefano Zampini PetscBool isseqdense; 5131cf9b237SStefano Zampini 5141cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 5151cf9b237SStefano Zampini if (!isseqdense) { 5164f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 5171cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 5181cf9b237SStefano Zampini PetscScalar *array; 5191cf9b237SStefano Zampini PetscReal chop=1.e-6; 5201cf9b237SStefano Zampini 5211cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 5221cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 5231cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 5241cf9b237SStefano Zampini for (i=0;i<n;i++) { 5251cf9b237SStefano Zampini PetscInt j; 5261cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 5271cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 5281cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 5291cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 5301cf9b237SStefano Zampini } 5311cf9b237SStefano Zampini } 5321cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 5339d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 5341cf9b237SStefano Zampini } 5354f1b2e48SStefano Zampini } else { 5364f1b2e48SStefano Zampini B = A; 5374f1b2e48SStefano Zampini } 5384f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5394f1b2e48SStefano Zampini 5404f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 5414f1b2e48SStefano Zampini if (filter) { 5424f1b2e48SStefano Zampini PetscScalar *data; 5434f1b2e48SStefano Zampini PetscInt j,cum; 5444f1b2e48SStefano Zampini 5454f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 5464f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 5474f1b2e48SStefano Zampini cum = 0; 5484f1b2e48SStefano Zampini for (i=0;i<n;i++) { 5494f1b2e48SStefano Zampini PetscInt t; 5504f1b2e48SStefano Zampini 5514f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 5524f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 5534f1b2e48SStefano Zampini continue; 5544f1b2e48SStefano Zampini } 5554f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 5564f1b2e48SStefano Zampini } 5574f1b2e48SStefano Zampini t = xadj_filtered[i]; 5584f1b2e48SStefano Zampini xadj_filtered[i] = cum; 5594f1b2e48SStefano Zampini cum += t; 5604f1b2e48SStefano Zampini } 5614f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini } else { 5634f1b2e48SStefano Zampini xadj_filtered = NULL; 5644f1b2e48SStefano Zampini adjncy_filtered = NULL; 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini 5674f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 5684f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 5704f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 5724f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 5734f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini if (xadj_filtered) { 5754f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 5764f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 5774f1b2e48SStefano Zampini } else { 5784f1b2e48SStefano Zampini graph->xadj = xadj; 5794f1b2e48SStefano Zampini graph->adjncy = adjncy; 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 5824f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5834f1b2e48SStefano Zampini /* partial clean up */ 5844f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 5854f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5861cf9b237SStefano Zampini if (A != B) { 5874f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini } 5894f1b2e48SStefano Zampini 5904f1b2e48SStefano Zampini /* get back data */ 5911cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 5921cf9b237SStefano Zampini if (cc) { 5934f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 5944f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 5954f1b2e48SStefano 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); 5964f1b2e48SStefano Zampini } 5974f1b2e48SStefano Zampini *cc = cc_n; 5981cf9b237SStefano Zampini } 5994f1b2e48SStefano Zampini /* clean up graph */ 6004f1b2e48SStefano Zampini graph->xadj = 0; 6014f1b2e48SStefano Zampini graph->adjncy = 0; 6024f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini PetscFunctionReturn(0); 6044f1b2e48SStefano Zampini } 6054f1b2e48SStefano Zampini 6064f1b2e48SStefano Zampini #undef __FUNCT__ 6075408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 6085408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 6095408967cSStefano Zampini { 6105408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6115408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 612dee84bffSStefano Zampini IS dirIS = NULL; 6134f1b2e48SStefano Zampini PetscInt i; 6145408967cSStefano Zampini PetscErrorCode ierr; 6155408967cSStefano Zampini 6165408967cSStefano Zampini PetscFunctionBegin; 617dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 6185408967cSStefano Zampini if (zerodiag) { 6195408967cSStefano Zampini Mat A; 6205408967cSStefano Zampini Vec vec3_N; 6215408967cSStefano Zampini PetscScalar *vals; 6225408967cSStefano Zampini const PetscInt *idxs; 623d12d3064SStefano Zampini PetscInt nz,*count; 6245408967cSStefano Zampini 6255408967cSStefano Zampini /* p0 */ 6265408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 6275408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 6285408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6295408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6304f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 6315408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6325408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6335408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6345408967cSStefano Zampini /* v_I */ 6355408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 6365408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 6375408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6385408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 6395408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6405408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 6415408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6425408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6435408967cSStefano Zampini if (dirIS) { 6445408967cSStefano Zampini PetscInt n; 6455408967cSStefano Zampini 6465408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 6475408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 6485408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 6495408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6505408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 6515408967cSStefano Zampini } 6525408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 6535408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 6545408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 6555408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 656669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 6575408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 6585408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 659fbfcb133SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) { 660b9b0e38cSStefano 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])); 6615408967cSStefano Zampini } 6625408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6635408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 664d12d3064SStefano Zampini 665d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 666d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 667d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 668d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 669d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 670d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 671d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 672d12d3064SStefano Zampini if (count[idxs[i]]) { 673d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 674d12d3064SStefano Zampini } 675d4d8cf7bSStefano Zampini } 676d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 677d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 6785408967cSStefano Zampini } 679dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 6805408967cSStefano Zampini 6815408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 6825408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 6834f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 6845408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 6854f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 6865408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 6874f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6884f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 6894f1b2e48SStefano 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); 6904f1b2e48SStefano Zampini } 6915408967cSStefano Zampini } 6925408967cSStefano Zampini PetscFunctionReturn(0); 6935408967cSStefano Zampini } 6945408967cSStefano Zampini 6955408967cSStefano Zampini #undef __FUNCT__ 696339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 697339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 698339f8db1SStefano Zampini { 699339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7004f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 701b0f5fe93SStefano Zampini PetscInt nz,n; 7021f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 7034f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 704339f8db1SStefano Zampini PetscErrorCode ierr; 705339f8db1SStefano Zampini 706339f8db1SStefano Zampini PetscFunctionBegin; 7079f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 7089f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 709a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 710a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 711a3df083aSStefano Zampini } 712a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 713a3df083aSStefano Zampini pcbddc->benign_n = 0; 7141ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 7154f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 7164f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 7174f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 7181ae86dd6SStefano Zampini since the local Schur complements are already SPD 7194f1b2e48SStefano Zampini */ 7204f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 7214f1b2e48SStefano Zampini have_null = PETSC_TRUE; 72240fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 7234f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 7244f1b2e48SStefano Zampini 7254f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 7264f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 7274f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 7284f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 729ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 73040fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 73140fa8d13SStefano Zampini if (!sorted) { 73240fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 73340fa8d13SStefano Zampini } 73440fa8d13SStefano Zampini } else { 73540fa8d13SStefano Zampini pressures = NULL; 73640fa8d13SStefano Zampini } 73797d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 73897d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 73927b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 74097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 741339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 742339f8db1SStefano Zampini if (!sorted) { 743339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 744339f8db1SStefano Zampini } 745339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 7464f1b2e48SStefano Zampini if (!nz) { 7474f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 7484f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 74940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 75040fa8d13SStefano Zampini } 7514f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 7524f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 7534f1b2e48SStefano Zampini zerodiag_subs = NULL; 7544f1b2e48SStefano Zampini pcbddc->benign_n = 0; 7551f4df5f7SStefano Zampini n_interior_dofs = 0; 7561f4df5f7SStefano Zampini interior_dofs = NULL; 7571f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 7581f4df5f7SStefano Zampini PetscInt n,i,j; 7591f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 7601f4df5f7SStefano Zampini PetscInt *iwork; 7611f4df5f7SStefano Zampini 7621f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 7631f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7641f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 7651f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 7661f4df5f7SStefano Zampini for (i=0;i<n_neigh;i++) 7671f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 7681f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 7691f4df5f7SStefano Zampini for (i=0;i<n;i++) 7701f4df5f7SStefano Zampini if (!iwork[i]) 7711f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 7721f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 7731f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7741f4df5f7SStefano Zampini } 7754f1b2e48SStefano Zampini if (has_null_pressures) { 7764f1b2e48SStefano Zampini IS *subs; 7771f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 7781f4df5f7SStefano Zampini const PetscInt *idxs; 7791f4df5f7SStefano Zampini PetscScalar *array; 7801f4df5f7SStefano Zampini Vec *work; 7811f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 7824f1b2e48SStefano Zampini 7834f1b2e48SStefano Zampini subs = pcbddc->local_subs; 7844f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 7851f4df5f7SStefano 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) */ 7861f4df5f7SStefano Zampini if (pcbddc->current_level) { 7871f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 7881f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 7891f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 7901f4df5f7SStefano Zampini /* work[0] = 1_p */ 7911f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 7921f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 7931f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 7941f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 7951f4df5f7SStefano Zampini /* work[0] = 1_v */ 7961f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 7971f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 7981f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 7991f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 8001f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 8011f4df5f7SStefano Zampini } 8024f1b2e48SStefano Zampini if (nsubs > 1) { 8034f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 8044f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 8054f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 8064f1b2e48SStefano Zampini IS t_zerodiag_subs; 8074f1b2e48SStefano Zampini PetscInt nl; 8084f1b2e48SStefano Zampini 8094f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 8104f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 8114f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 8124f1b2e48SStefano Zampini if (nl) { 8134f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8144f1b2e48SStefano Zampini 8151f4df5f7SStefano Zampini if (pcbddc->current_level) { 8161f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 8171f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 8181f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 8191f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8201f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8211f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8221f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 8231f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 8241f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 8251f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 8261f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 8271f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8281f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8291f4df5f7SStefano Zampini valid = PETSC_FALSE; 8301f4df5f7SStefano Zampini break; 8311f4df5f7SStefano Zampini } 8321f4df5f7SStefano Zampini } 8331f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 8341f4df5f7SStefano Zampini } 8351f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 8361f4df5f7SStefano Zampini IS t_bc; 8371f4df5f7SStefano Zampini PetscInt nzb; 8381f4df5f7SStefano Zampini 8391f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 8401f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 8411f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 8421f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8431f4df5f7SStefano Zampini } 8441f4df5f7SStefano Zampini if (valid && pressures) { 8454f1b2e48SStefano Zampini IS t_pressure_subs; 8464f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 8474f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 8484f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 8494f1b2e48SStefano Zampini } 8504f1b2e48SStefano Zampini if (valid) { 8514f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 8524f1b2e48SStefano Zampini pcbddc->benign_n++; 8534f1b2e48SStefano Zampini } else { 8544f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 8554f1b2e48SStefano Zampini } 8564f1b2e48SStefano Zampini } 8574f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 8584f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 8594f1b2e48SStefano Zampini } 8604f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 8614f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8621f4df5f7SStefano Zampini 8631f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 8641f4df5f7SStefano Zampini PetscInt nzb; 8651f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 8661f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8671f4df5f7SStefano Zampini } 8681f4df5f7SStefano Zampini if (valid && pressures) { 8694f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 8704f1b2e48SStefano Zampini } 8711f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 8721f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 8731f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 8741f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8751f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8761f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8771f4df5f7SStefano Zampini valid = PETSC_FALSE; 8781f4df5f7SStefano Zampini break; 8791f4df5f7SStefano Zampini } 8801f4df5f7SStefano Zampini } 8811f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8821f4df5f7SStefano Zampini } 8834f1b2e48SStefano Zampini if (valid) { 8844f1b2e48SStefano Zampini pcbddc->benign_n = 1; 885ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 8864f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 8874f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 8884f1b2e48SStefano Zampini } 8894f1b2e48SStefano Zampini } 8901f4df5f7SStefano Zampini if (pcbddc->current_level) { 8911f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 8924f1b2e48SStefano Zampini } 8931f4df5f7SStefano Zampini } 8941f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 8954f1b2e48SStefano Zampini 8964f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 897b9b0e38cSStefano Zampini PetscInt n; 898b9b0e38cSStefano Zampini 8994f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9004f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 901b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 902b9b0e38cSStefano Zampini if (n) { 9034f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 9044f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9054f1b2e48SStefano Zampini } 906b9b0e38cSStefano Zampini } 9074f1b2e48SStefano Zampini 9084f1b2e48SStefano Zampini /* final check for null pressures */ 9094f1b2e48SStefano Zampini if (zerodiag && pressures) { 9104f1b2e48SStefano Zampini PetscInt nz,np; 9114f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 9124f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 9134f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 9144f1b2e48SStefano Zampini } 9154f1b2e48SStefano Zampini 9164f1b2e48SStefano Zampini if (recompute_zerodiag) { 9174f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9184f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 9194f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 9204f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 9214f1b2e48SStefano Zampini } else { 9224f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 9234f1b2e48SStefano Zampini 9244f1b2e48SStefano Zampini nzn = 0; 9254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9264f1b2e48SStefano Zampini PetscInt ns; 9274f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9284f1b2e48SStefano Zampini nzn += ns; 9294f1b2e48SStefano Zampini } 9304f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 9314f1b2e48SStefano Zampini nzn = 0; 9324f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9334f1b2e48SStefano Zampini PetscInt ns,*idxs; 9344f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9354f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9364f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 9374f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9384f1b2e48SStefano Zampini nzn += ns; 9394f1b2e48SStefano Zampini } 9404f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 9414f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 9424f1b2e48SStefano Zampini } 9434f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9444f1b2e48SStefano Zampini } 9454f1b2e48SStefano Zampini 946669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 947669cc0f4SStefano Zampini if (pcbddc->benign_compute_nonetflux && !pcbddc->divudotp) { 948669cc0f4SStefano Zampini Mat A; 9491f4df5f7SStefano Zampini IS isused = NULL; 9501ae86dd6SStefano Zampini PetscInt n; 9511f4df5f7SStefano Zampini if (pressures) { 9521f4df5f7SStefano Zampini isused = pressures; 9531f4df5f7SStefano Zampini } else { 9541f4df5f7SStefano Zampini isused = zerodiag; 9551f4df5f7SStefano Zampini } 956669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 9571ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 9581ae86dd6SStefano Zampini if (!isused && n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Don't know how to extract div u dot p! Please provide the pressure field"); 9591ae86dd6SStefano Zampini if (n) { 960669cc0f4SStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&pcbddc->divudotp);CHKERRQ(ierr); 9611ae86dd6SStefano Zampini } else { 9621ae86dd6SStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&pcbddc->divudotp);CHKERRQ(ierr); 9631ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9641ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 9651ae86dd6SStefano Zampini } 966b3afcdbeSStefano Zampini } 967b3afcdbeSStefano Zampini 968b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 9694f1b2e48SStefano Zampini if (has_null_pressures) { 9704f1b2e48SStefano Zampini IS zerodiagc; 9714f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 9724f1b2e48SStefano Zampini PetscInt i,s,*nnz; 9734f1b2e48SStefano Zampini 9744f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 975339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 976339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 977339f8db1SStefano Zampini /* local change of basis for pressures */ 978339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 97997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 980339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 981339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 982339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 9834f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 9844f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9854f1b2e48SStefano Zampini PetscInt nzs,j; 9864f1b2e48SStefano Zampini 9874f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 9884f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 9894f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 9904f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 9914f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 9924f1b2e48SStefano Zampini } 993339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 994339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 995339f8db1SStefano Zampini /* set identity on velocities */ 996339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 997339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 998339f8db1SStefano Zampini } 9994f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 10004f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 10019f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 10024f1b2e48SStefano 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); 1003339f8db1SStefano Zampini /* set change on pressures */ 10044f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 10054f1b2e48SStefano Zampini PetscScalar *array; 10064f1b2e48SStefano Zampini PetscInt nzs; 10074f1b2e48SStefano Zampini 10084f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 10094f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 10104f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 1011339f8db1SStefano Zampini PetscScalar vals[2]; 1012339f8db1SStefano Zampini PetscInt cols[2]; 1013339f8db1SStefano Zampini 1014339f8db1SStefano Zampini cols[0] = idxs[i]; 10154f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 1016339f8db1SStefano Zampini vals[0] = 1.; 1017b0f5fe93SStefano Zampini vals[1] = 1.; 10184f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1019339f8db1SStefano Zampini } 10204f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 10214f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 10224f1b2e48SStefano Zampini array[nzs-1] = 1.; 10234f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 10244f1b2e48SStefano Zampini /* store local idxs for p0 */ 10254f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 10264f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 1027339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 10284f1b2e48SStefano Zampini } 1029339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1030339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1031a3df083aSStefano Zampini /* project if needed */ 1032a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10331dd7afcfSStefano Zampini Mat M; 10341dd7afcfSStefano Zampini 10351dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 1036339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 10371dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 10381dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 1039a3df083aSStefano Zampini } 10404f1b2e48SStefano Zampini /* store global idxs for p0 */ 10414f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1042339f8db1SStefano Zampini } 1043ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 10444f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 1045b0f5fe93SStefano Zampini 1046b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 1047b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 104827b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 104927b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1050339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 1051339f8db1SStefano Zampini PetscFunctionReturn(0); 1052339f8db1SStefano Zampini } 1053339f8db1SStefano Zampini 1054339f8db1SStefano Zampini #undef __FUNCT__ 1055015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 1056015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 1057efc2fbd9SStefano Zampini { 1058efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1059de9d7bd0SStefano Zampini PetscScalar *array; 1060efc2fbd9SStefano Zampini PetscErrorCode ierr; 1061efc2fbd9SStefano Zampini 1062efc2fbd9SStefano Zampini PetscFunctionBegin; 1063efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 1064efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 10654f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1066efc2fbd9SStefano Zampini } 1067de9d7bd0SStefano Zampini if (get) { 1068efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 10694f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 10704f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 1071efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 1072de9d7bd0SStefano Zampini } else { 1073de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 1074de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1075de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1076de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 1077efc2fbd9SStefano Zampini } 1078efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1079efc2fbd9SStefano Zampini } 1080efc2fbd9SStefano Zampini 1081efc2fbd9SStefano Zampini #undef __FUNCT__ 1082c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 1083c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 1084c263805aSStefano Zampini { 1085c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1086c263805aSStefano Zampini PetscErrorCode ierr; 1087c263805aSStefano Zampini 1088c263805aSStefano Zampini PetscFunctionBegin; 1089c263805aSStefano Zampini /* TODO: add error checking 1090c263805aSStefano Zampini - avoid nested pop (or push) calls. 1091c263805aSStefano Zampini - cannot push before pop. 10921c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 1093c263805aSStefano Zampini */ 10944f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1095efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1096efc2fbd9SStefano Zampini } 1097c263805aSStefano Zampini if (pop) { 1098a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10994f1b2e48SStefano Zampini IS is_p0; 11004f1b2e48SStefano Zampini MatReuse reuse; 1101c263805aSStefano Zampini 1102c263805aSStefano Zampini /* extract B_0 */ 11034f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 11044f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 11054f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 11064f1b2e48SStefano Zampini } 11074f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 11084f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 1109c263805aSStefano Zampini /* remove rows and cols from local problem */ 1110c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 111197d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 11124f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 11134f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 1114a3df083aSStefano Zampini } else { 1115a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1116a3df083aSStefano Zampini PetscScalar *vals; 1117a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 1118a3df083aSStefano Zampini 1119a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 1120a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 1121a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 11220b5adadeSStefano Zampini PetscInt *nnz; 1123a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 1124a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 1125a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1126331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 1127331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1128331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 1129331e053bSStefano Zampini nnz[i] = n - nnz[i]; 1130331e053bSStefano Zampini } 1131331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 1132331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1133331e053bSStefano Zampini } 1134a3df083aSStefano Zampini 1135a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1136a3df083aSStefano Zampini PetscScalar *array; 1137a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 1138a3df083aSStefano Zampini 1139a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 1140a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1141a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1142a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 1143a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1144a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 1145a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 1146a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 1147a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 1148a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 1149a3df083aSStefano Zampini cum = 0; 1150a3df083aSStefano Zampini for (j=0;j<n;j++) { 115122db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 1152a3df083aSStefano Zampini vals[cum] = array[j]; 1153a3df083aSStefano Zampini idxs_ins[cum] = j; 1154a3df083aSStefano Zampini cum++; 1155a3df083aSStefano Zampini } 1156a3df083aSStefano Zampini } 1157a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 1158a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 1159a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1160a3df083aSStefano Zampini } 1161a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1162a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1163a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 1164a3df083aSStefano Zampini } 1165c263805aSStefano Zampini } else { /* push */ 1166a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 11674f1b2e48SStefano Zampini PetscInt i; 11684f1b2e48SStefano Zampini 11694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 11704f1b2e48SStefano Zampini PetscScalar *B0_vals; 11714f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 11724f1b2e48SStefano Zampini 11734f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 11744f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 11757b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 11764f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 11774f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 11784f1b2e48SStefano Zampini } 1179c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1180c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1181a3df083aSStefano Zampini } else { 1182a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 1183a3df083aSStefano Zampini } 1184c263805aSStefano Zampini } 1185c263805aSStefano Zampini PetscFunctionReturn(0); 1186c263805aSStefano Zampini } 1187c263805aSStefano Zampini 1188c263805aSStefano Zampini #undef __FUNCT__ 1189b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 119008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 1191b1b3d7a2SStefano Zampini { 1192b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 119308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 119408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 119508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 119608122e43SStefano Zampini PetscScalar *work,lwork; 119708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 119808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 119908122e43SStefano Zampini PetscReal *eigs,thresh; 12001b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 1201f6f667cfSStefano Zampini PetscBool allocated_S_St; 120208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 120308122e43SStefano Zampini PetscReal *rwork; 120408122e43SStefano Zampini #endif 1205b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1206b1b3d7a2SStefano Zampini 1207b1b3d7a2SStefano Zampini PetscFunctionBegin; 1208b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 1209af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 1210af25d912SStefano 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); 121106a4e24aSStefano Zampini 1212fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1213fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1214fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1215fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 12161575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1217fd14bc51SStefano Zampini } 1218fd14bc51SStefano Zampini 1219e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 1220e496cd5dSStefano 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); 1221e496cd5dSStefano Zampini } 1222e496cd5dSStefano Zampini 122308122e43SStefano Zampini /* max size of subsets */ 122408122e43SStefano Zampini mss = 0; 122508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 122608122e43SStefano Zampini PetscInt subset_size; 1227862806e4SStefano Zampini 122808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 122908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 123008122e43SStefano Zampini } 123108122e43SStefano Zampini 123208122e43SStefano Zampini /* min/max and threshold */ 123308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 1234f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 123508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 1236f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 1237f6f667cfSStefano Zampini if (nmin) { 1238f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 1239f6f667cfSStefano Zampini } 124008122e43SStefano Zampini 124108122e43SStefano Zampini /* allocate lapack workspace */ 124208122e43SStefano Zampini cum = cum2 = 0; 124308122e43SStefano Zampini maxneigs = 0; 124408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 124508122e43SStefano Zampini PetscInt n,subset_size; 1246f6f667cfSStefano Zampini 124708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 124808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 12499162d606SStefano Zampini cum += subset_size; 12509162d606SStefano Zampini cum2 += subset_size*n; 125108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 125208122e43SStefano Zampini } 125308122e43SStefano Zampini if (mss) { 12549ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 125508122e43SStefano Zampini PetscBLASInt B_itype = 1; 125608122e43SStefano Zampini PetscBLASInt B_N = mss; 12574c6709b3SStefano Zampini PetscReal zero = 0.0; 12584c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 125908122e43SStefano Zampini 126008122e43SStefano Zampini B_lwork = -1; 126108122e43SStefano Zampini S = NULL; 126208122e43SStefano Zampini St = NULL; 1263a58a30b4SStefano Zampini eigs = NULL; 1264a58a30b4SStefano Zampini eigv = NULL; 1265a58a30b4SStefano Zampini B_iwork = NULL; 1266a58a30b4SStefano Zampini B_ifail = NULL; 1267d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1268d1710679SStefano Zampini rwork = NULL; 1269d1710679SStefano Zampini #endif 12708bec7fa6SStefano Zampini thresh = 1.0; 127108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 127208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 127308122e43SStefano 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)); 127408122e43SStefano Zampini #else 127508122e43SStefano 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)); 127608122e43SStefano Zampini #endif 127708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 127808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 127908122e43SStefano Zampini } else { 128008122e43SStefano Zampini /* TODO */ 128108122e43SStefano Zampini } 128208122e43SStefano Zampini } else { 128308122e43SStefano Zampini lwork = 0; 128408122e43SStefano Zampini } 128508122e43SStefano Zampini 128608122e43SStefano Zampini nv = 0; 1287d62866d3SStefano 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) */ 1288d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 128908122e43SStefano Zampini } 12904c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1291f6f667cfSStefano Zampini if (allocated_S_St) { 1292f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1293f6f667cfSStefano Zampini } 1294f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 129508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 129608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 129708122e43SStefano Zampini #endif 12989162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 12999162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 13009162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 130108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 13029162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 130308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 130408122e43SStefano Zampini 130508122e43SStefano Zampini maxneigs = 0; 130672b8c272SStefano Zampini cum = cumarray = 0; 13079162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 13089162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1309d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 131008122e43SStefano Zampini const PetscInt *idxs; 131108122e43SStefano Zampini 1312d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 131308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 131408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 131508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 131608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 13179162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 13189162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 131908122e43SStefano Zampini } 1320d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 132108122e43SStefano Zampini } 132208122e43SStefano Zampini 132308122e43SStefano Zampini if (mss) { /* multilevel */ 132408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 132508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 132608122e43SStefano Zampini } 132708122e43SStefano Zampini 1328ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 132908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 133008122e43SStefano Zampini const PetscInt *idxs; 13319d54b7f4SStefano Zampini PetscReal upper,lower; 1332862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 133308122e43SStefano Zampini PetscBLASInt B_N; 1334aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 133508122e43SStefano Zampini 13369d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 13379d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 13389d54b7f4SStefano Zampini lower = thresh; 13399d54b7f4SStefano Zampini } else { 13409d54b7f4SStefano Zampini upper = 1./thresh; 13419d54b7f4SStefano Zampini lower = 0.; 13429d54b7f4SStefano Zampini } 1343862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1344ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1345f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1346f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 13479ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1348aff50787SStefano Zampini PetscInt j,k; 1349aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1350aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1351aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 135208122e43SStefano Zampini } 135308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1354aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1355aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1356aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1357aff50787SStefano Zampini } 135808122e43SStefano Zampini } 135908122e43SStefano Zampini } else { 136008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 136108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 136208122e43SStefano Zampini } 13638bec7fa6SStefano Zampini } else { 1364f6f667cfSStefano Zampini S = Sarray + cumarray; 1365f6f667cfSStefano Zampini St = Starray + cumarray; 13668bec7fa6SStefano Zampini } 1367aff50787SStefano Zampini /* see if we can save some work */ 1368b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1369aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1370aff50787SStefano Zampini } 1371aff50787SStefano Zampini 1372b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1373aff50787SStefano Zampini B_neigs = 0; 1374aff50787SStefano Zampini } else { 13759ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 137608122e43SStefano Zampini PetscBLASInt B_itype = 1; 1377f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 13784c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 13799552c7c7SStefano Zampini PetscInt nmin_s; 1380b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 138108122e43SStefano Zampini 1382fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 13838bec7fa6SStefano 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]]); 1384fd14bc51SStefano Zampini } 1385d16cbb6bSStefano Zampini 1386b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1387b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1388b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1389b7ab4a40SStefano Zampini } 1390b7ab4a40SStefano Zampini 139108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1392b7ab4a40SStefano Zampini if (compute_range) { 1393d16cbb6bSStefano Zampini 1394d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 139508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 13969d54b7f4SStefano 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)); 139708122e43SStefano Zampini #else 13989d54b7f4SStefano 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)); 139908122e43SStefano Zampini #endif 1400b7ab4a40SStefano Zampini } else if (!same_data) { 1401d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1402d16cbb6bSStefano Zampini B_IL = 1; 1403d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 14049d54b7f4SStefano 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)); 1405d16cbb6bSStefano Zampini #else 14069d54b7f4SStefano 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)); 1407d16cbb6bSStefano Zampini #endif 1408b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1409b7ab4a40SStefano Zampini PetscInt k; 1410b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1411b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1412b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1413b7ab4a40SStefano Zampini nmin = nmax; 1414b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1415b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1416b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1417b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1418b7ab4a40SStefano Zampini } 1419d16cbb6bSStefano Zampini } 142008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 142108122e43SStefano Zampini if (B_ierr) { 14226c4ed002SBarry 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); 14236c4ed002SBarry 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); 14246c4ed002SBarry 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); 142508122e43SStefano Zampini } 142608122e43SStefano Zampini 142708122e43SStefano Zampini if (B_neigs > nmax) { 1428fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1429fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1430fd14bc51SStefano Zampini } 14319d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 143208122e43SStefano Zampini B_neigs = nmax; 143308122e43SStefano Zampini } 143408122e43SStefano Zampini 14359552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 14369552c7c7SStefano Zampini if (B_neigs < nmin_s) { 143708122e43SStefano Zampini PetscBLASInt B_neigs2; 143808122e43SStefano Zampini 14399d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1440f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 14419d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 14429d54b7f4SStefano Zampini } else { 14439d54b7f4SStefano Zampini B_IL = B_neigs + 1; 14449d54b7f4SStefano Zampini B_IU = nmin_s; 14459d54b7f4SStefano Zampini } 1446fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1447fd14bc51SStefano 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); 1448fd14bc51SStefano Zampini } 14499ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 14501ae86dd6SStefano Zampini PetscInt j,k; 145108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 14521ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 14531ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 14541ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 145508122e43SStefano Zampini } 145608122e43SStefano Zampini } 145708122e43SStefano Zampini } else { 145808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 145908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 146008122e43SStefano Zampini } 146108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 146208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 14639d54b7f4SStefano 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)); 146408122e43SStefano Zampini #else 14659d54b7f4SStefano 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)); 146608122e43SStefano Zampini #endif 146708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 146808122e43SStefano Zampini B_neigs += B_neigs2; 146908122e43SStefano Zampini } 147008122e43SStefano Zampini if (B_ierr) { 14716c4ed002SBarry 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); 14726c4ed002SBarry 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); 14736c4ed002SBarry 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); 147408122e43SStefano Zampini } 1475fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1476ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 147708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 147808122e43SStefano Zampini if (eigs[j] == 0.0) { 1479ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 148008122e43SStefano Zampini } else { 14819d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1482ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 14839d54b7f4SStefano Zampini } else { 14849d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 14859d54b7f4SStefano Zampini } 1486fd14bc51SStefano Zampini } 148708122e43SStefano Zampini } 148808122e43SStefano Zampini } 148908122e43SStefano Zampini } else { 149008122e43SStefano Zampini /* TODO */ 149108122e43SStefano Zampini } 1492aff50787SStefano Zampini } 14936c3e6151SStefano Zampini /* change the basis back to the original one */ 14946c3e6151SStefano Zampini if (sub_schurs->change) { 149572b8c272SStefano Zampini Mat change,phi,phit; 14966c3e6151SStefano Zampini 14976c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 14986c3e6151SStefano Zampini PetscInt ii; 14996c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 15006c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15016c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 15026c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 15036c3e6151SStefano Zampini } 15046c3e6151SStefano Zampini } 15056c3e6151SStefano Zampini } 150672b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 15076c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 150872b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 15096c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 15106c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 15116c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 15126c3e6151SStefano Zampini } 15138bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 15148bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 15159162d606SStefano Zampini if (B_neigs) { 15169162d606SStefano 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); 1517fd14bc51SStefano Zampini 1518fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 15199552c7c7SStefano Zampini PetscInt ii; 15209552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1521ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15229552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1523ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1524ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1525ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1526ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1527ac47001eSStefano Zampini #else 1528ac47001eSStefano 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); 1529ac47001eSStefano Zampini #endif 15309552c7c7SStefano Zampini } 15319552c7c7SStefano Zampini } 1532fd14bc51SStefano Zampini } 15339162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 15349162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 15359162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 15369162d606SStefano Zampini cum++; 153708122e43SStefano Zampini } 153808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 153908122e43SStefano Zampini /* shift for next computation */ 154008122e43SStefano Zampini cumarray += subset_size*subset_size; 154108122e43SStefano Zampini } 1542fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1543fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1544fd14bc51SStefano Zampini } 154508122e43SStefano Zampini 154608122e43SStefano Zampini if (mss) { 154708122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 154808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1549f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1550f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1551f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 155208122e43SStefano Zampini } 1553f6f667cfSStefano Zampini if (allocated_S_St) { 1554f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1555f6f667cfSStefano Zampini } 1556f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 155708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 155808122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 155908122e43SStefano Zampini #endif 156008122e43SStefano Zampini if (pcbddc->dbg_flag) { 15611b968477SStefano Zampini PetscInt maxneigs_r; 1562b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 15639b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 156408122e43SStefano Zampini } 156508122e43SStefano Zampini PetscFunctionReturn(0); 156608122e43SStefano Zampini } 1567b1b3d7a2SStefano Zampini 1568674ae819SStefano Zampini #undef __FUNCT__ 1569c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1570c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1571c8587f34SStefano Zampini { 15728629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1573c8587f34SStefano Zampini PetscErrorCode ierr; 1574c8587f34SStefano Zampini 1575c8587f34SStefano Zampini PetscFunctionBegin; 1576f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 15775e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1578c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1579c8587f34SStefano Zampini 1580684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 15810fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1582684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1583c8587f34SStefano Zampini 15848629588bSStefano Zampini /* 15858629588bSStefano Zampini Setup local correction and local part of coarse basis. 15868629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 15878629588bSStefano Zampini */ 158847f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 15898629588bSStefano Zampini 15908629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 15918629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 15928629588bSStefano Zampini 15938629588bSStefano Zampini /* free */ 15948629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1595c8587f34SStefano Zampini PetscFunctionReturn(0); 1596c8587f34SStefano Zampini } 1597c8587f34SStefano Zampini 1598c8587f34SStefano Zampini #undef __FUNCT__ 1599674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1600674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1601674ae819SStefano Zampini { 1602674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1603674ae819SStefano Zampini PetscErrorCode ierr; 1604674ae819SStefano Zampini 1605674ae819SStefano Zampini PetscFunctionBegin; 1606674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1607674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 160830368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1609674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1610785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1611674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1612f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1613f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1614785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 161563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 161663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1617674ae819SStefano Zampini PetscFunctionReturn(0); 1618674ae819SStefano Zampini } 1619674ae819SStefano Zampini 1620674ae819SStefano Zampini #undef __FUNCT__ 1621674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1622674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1623674ae819SStefano Zampini { 1624674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16254f1b2e48SStefano Zampini PetscInt i; 1626674ae819SStefano Zampini PetscErrorCode ierr; 1627674ae819SStefano Zampini 1628674ae819SStefano Zampini PetscFunctionBegin; 1629b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1630674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 163116909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 16321dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1633674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1634669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 1635674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 16364f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 16374f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 16384f1b2e48SStefano Zampini } 16394f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1640b334f244SStefano Zampini if (pcbddc->sub_schurs) { 1641b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1642b334f244SStefano Zampini } 1643674ae819SStefano Zampini PetscFunctionReturn(0); 1644674ae819SStefano Zampini } 1645674ae819SStefano Zampini 1646674ae819SStefano Zampini #undef __FUNCT__ 1647674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1648674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1649674ae819SStefano Zampini { 1650674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1651674ae819SStefano Zampini PetscErrorCode ierr; 1652674ae819SStefano Zampini 1653674ae819SStefano Zampini PetscFunctionBegin; 1654674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 165558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1656ca92afb2SStefano Zampini PetscScalar *array; 165706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 165806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 165958da7f69SStefano Zampini } 1660674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1661674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 166215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 166315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1664674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1665674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1666674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 166706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1668674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1669674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 16708ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1671674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1672674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1673674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1674f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1675f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1676f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1677f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1678727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 16790e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1680f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 168170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 168281d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 16830369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 16841dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 16854f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 16868b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1687ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1688ca92afb2SStefano Zampini PetscInt i; 1689ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1690ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1691ca92afb2SStefano Zampini } 1692ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1693ca92afb2SStefano Zampini } 16944f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1695674ae819SStefano Zampini PetscFunctionReturn(0); 1696674ae819SStefano Zampini } 1697674ae819SStefano Zampini 1698674ae819SStefano Zampini #undef __FUNCT__ 1699f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1700f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 17016bfb1811SStefano Zampini { 17026bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 17036bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 17046bfb1811SStefano Zampini VecType impVecType; 17054f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 17066bfb1811SStefano Zampini PetscErrorCode ierr; 17076bfb1811SStefano Zampini 17086bfb1811SStefano Zampini PetscFunctionBegin; 17096c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1710e7b262bdSStefano Zampini /* get sizes */ 17114f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1712b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 17136bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1714e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1715e7b262bdSStefano Zampini /* R nodes */ 1716e7b262bdSStefano Zampini old_size = -1; 1717e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1718e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1719e7b262bdSStefano Zampini } 1720e7b262bdSStefano Zampini if (n_R != old_size) { 1721e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1722e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17236bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 17246bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 17256bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 17266bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1727e7b262bdSStefano Zampini } 1728e7b262bdSStefano Zampini /* local primal dofs */ 1729e7b262bdSStefano Zampini old_size = -1; 1730e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1731e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1732e7b262bdSStefano Zampini } 1733e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1734e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 173583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1736e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 17376bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1738e7b262bdSStefano Zampini } 1739e7b262bdSStefano Zampini /* local explicit constraints */ 1740e7b262bdSStefano Zampini old_size = -1; 1741e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1742e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1743e7b262bdSStefano Zampini } 1744e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1745e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 174683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 174783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 174883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 174983b7ccabSStefano Zampini } 17506bfb1811SStefano Zampini PetscFunctionReturn(0); 17516bfb1811SStefano Zampini } 17526bfb1811SStefano Zampini 17536bfb1811SStefano Zampini #undef __FUNCT__ 175447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 175547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 175688ebb749SStefano Zampini { 175725084f0cSStefano Zampini PetscErrorCode ierr; 175825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 175988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 176088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1761d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 176225084f0cSStefano Zampini /* submatrices of local problem */ 176380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 176406656605SStefano Zampini /* submatrices of local coarse problem */ 176506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 176625084f0cSStefano Zampini /* working matrices */ 176706656605SStefano Zampini Mat C_CR; 176825084f0cSStefano Zampini /* additional working stuff */ 176906656605SStefano Zampini PC pc_R; 17704f1b2e48SStefano Zampini Mat F; 17715cbda25cSStefano Zampini Vec dummy_vec; 1772a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 177325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 177406656605SStefano Zampini PetscScalar *work; 177506656605SStefano Zampini PetscInt *idx_V_B; 1776ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 177706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1778ffd830a3SStefano Zampini 177925084f0cSStefano Zampini /* some shortcuts to scalars */ 178006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 178188ebb749SStefano Zampini 178288ebb749SStefano Zampini PetscFunctionBegin; 1783ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1784ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1785ffd830a3SStefano Zampini } 1786ffd830a3SStefano Zampini 1787ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1788b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 17894f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1790b371cd4fSStefano Zampini n_B = pcis->n_B; 1791b371cd4fSStefano Zampini n_D = pcis->n - n_B; 179288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 179388ebb749SStefano Zampini 179488ebb749SStefano Zampini /* vertices in boundary numbering */ 1795785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 17960e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 17976c4ed002SBarry 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); 179888ebb749SStefano Zampini 179906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1800019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 180106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 180206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 180306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 180406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 180506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 180606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 180706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 180806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 180906656605SStefano Zampini 181006656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 181106656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 181206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 181306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 181406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1815ffd830a3SStefano Zampini lda_rhs = n_R; 1816a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 181706656605SStefano Zampini if (isLU || isILU || isCHOL) { 181806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1819b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 1820df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1821d62866d3SStefano Zampini MatFactorType type; 1822d62866d3SStefano Zampini 1823df4d28bfSStefano Zampini F = reuse_solver->F; 18246816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1825d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1826ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 182722db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 182806656605SStefano Zampini } else { 182906656605SStefano Zampini F = NULL; 183006656605SStefano Zampini } 183106656605SStefano Zampini 1832ffd830a3SStefano Zampini /* allocate workspace */ 1833ffd830a3SStefano Zampini n = 0; 1834ffd830a3SStefano Zampini if (n_constraints) { 1835ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1836ffd830a3SStefano Zampini } 1837ffd830a3SStefano Zampini if (n_vertices) { 1838ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1839ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1840ffd830a3SStefano Zampini } 1841ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1842ffd830a3SStefano Zampini 18435cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 18445cbda25cSStefano Zampini dummy_vec = NULL; 18455cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 18465cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 18475cbda25cSStefano Zampini } 18485cbda25cSStefano Zampini 184988ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 185088ebb749SStefano Zampini if (n_constraints) { 185172b8c272SStefano Zampini Mat M1,M2,M3,C_B; 185206656605SStefano Zampini IS is_aux; 185380677318SStefano Zampini PetscScalar *array,*array2; 185406656605SStefano Zampini 1855f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 185680677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 185788ebb749SStefano Zampini 185825084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 185925084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 18608ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 186172b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 186288ebb749SStefano Zampini 186380677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 186480677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1865ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 186688ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 186706656605SStefano Zampini const PetscScalar *row_cmat_values; 186806656605SStefano Zampini const PetscInt *row_cmat_indices; 186906656605SStefano Zampini PetscInt size_of_constraint,j; 187088ebb749SStefano Zampini 187106656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 187206656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1873ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 187406656605SStefano Zampini } 187506656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 187606656605SStefano Zampini } 1877ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 187806656605SStefano Zampini if (F) { 187906656605SStefano Zampini Mat B; 188006656605SStefano Zampini 1881ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1882a3df083aSStefano Zampini if (need_benign_correction) { 1883df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1884a3df083aSStefano Zampini 188572b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 188672b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1887a3df083aSStefano Zampini } 188880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1889a3df083aSStefano Zampini if (need_benign_correction) { 1890a3df083aSStefano Zampini PetscScalar *marr; 1891df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1892a3df083aSStefano Zampini 1893a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 18945cbda25cSStefano Zampini if (lda_rhs != n_R) { 18955cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 18965cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18975cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 18985cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18995cbda25cSStefano Zampini } 19005cbda25cSStefano Zampini } else { 1901a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1902a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19035cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1904a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1905a3df083aSStefano Zampini } 19065cbda25cSStefano Zampini } 1907a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1908a3df083aSStefano Zampini } 190906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 191006656605SStefano Zampini } else { 191180677318SStefano Zampini PetscScalar *marr; 191280677318SStefano Zampini 191380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 191406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1915ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1916ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 191706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 191806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 191906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 192006656605SStefano Zampini } 192180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 192206656605SStefano Zampini } 192380677318SStefano Zampini if (!pcbddc->switch_static) { 192480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 192580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 192680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 192780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1928ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 192980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 193080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 193380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 193480677318SStefano Zampini } 193580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 193680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 193772b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 193880677318SStefano Zampini } else { 1939ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1940ffd830a3SStefano Zampini IS dummy; 1941ffd830a3SStefano Zampini 1942ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 194372b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1944ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1945ffd830a3SStefano Zampini } else { 194680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 194780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1948ffd830a3SStefano Zampini } 194925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 195080677318SStefano Zampini } 195180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 195280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 195380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 195406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 195506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 195680677318SStefano Zampini if (isCHOL) { 195780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 195880677318SStefano Zampini } else { 195925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 196080677318SStefano Zampini } 196180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 196206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 196325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 196425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 196525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 196680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 196772b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 196872b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 196906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 197006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1971f4ddd8eeSStefano Zampini } 1972fc227af8SStefano Zampini 1973fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 197488ebb749SStefano Zampini if (n_vertices) { 197506656605SStefano Zampini IS is_aux; 19763a50541eSStefano Zampini 1977b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 19786816873aSStefano Zampini IS tis; 19796816873aSStefano Zampini 19806816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 19816816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 19826816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 19836816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 19846816873aSStefano Zampini } else { 19853a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 19866816873aSStefano Zampini } 19879577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 19889577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 198904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 199025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 199188ebb749SStefano Zampini } 199288ebb749SStefano Zampini 199388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1994f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 199506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 199606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 199706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 199806656605SStefano Zampini } 1999f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 200006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 200106656605SStefano Zampini PetscScalar *marray; 200206656605SStefano Zampini 200306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 200406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 2005f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2006f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 2007f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 2008f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2009f4ddd8eeSStefano Zampini } 2010f4ddd8eeSStefano Zampini } 201106656605SStefano Zampini 2012f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 201306656605SStefano Zampini PetscScalar *marray; 201488ebb749SStefano Zampini 201506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20168eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 201706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 201888ebb749SStefano Zampini } 20193301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 202006656605SStefano Zampini n *= 2; 202188ebb749SStefano Zampini } 202206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 202306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 202406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20258eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 202606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 202706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 202888ebb749SStefano Zampini } 20293301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 203006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 20318eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 203206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 203306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 203488ebb749SStefano Zampini } 203588ebb749SStefano Zampini } else { 2036c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 2037c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 20381b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2039c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 2040c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 2041c0553b1fSStefano Zampini } 204288ebb749SStefano Zampini } 204306656605SStefano Zampini } 2044019a44ceSStefano Zampini 204506656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 20464f1b2e48SStefano Zampini p0_lidx_I = NULL; 20474f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 2048d12edf2fSStefano Zampini const PetscInt *idxs; 2049d12edf2fSStefano Zampini 2050d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 20514f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 20524f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20534f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 20544f1b2e48SStefano Zampini } 2055d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 2056d12edf2fSStefano Zampini } 2057d16cbb6bSStefano Zampini 205806656605SStefano Zampini /* vertices */ 205906656605SStefano Zampini if (n_vertices) { 206016f15bc4SStefano Zampini 2061af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 206204708bb6SStefano Zampini 206316f15bc4SStefano Zampini if (n_R) { 206414393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 206506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 206616f15bc4SStefano Zampini PetscScalar *x,*y; 206704708bb6SStefano Zampini PetscBool isseqaij; 206806656605SStefano Zampini 206921eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 207014393ed6SStefano Zampini if (need_benign_correction) { 207114393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 207214393ed6SStefano Zampini IS is_p0; 207314393ed6SStefano Zampini PetscInt *idxs_p0,n; 207414393ed6SStefano Zampini 207514393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 207614393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 207714393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2078af25d912SStefano 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); 207914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 208014393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 208114393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 208214393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 208314393ed6SStefano Zampini } 208414393ed6SStefano Zampini 2085ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2086af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2087ffd830a3SStefano Zampini } else { 2088ca92afb2SStefano Zampini PetscScalar *av,*array; 2089ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2090ca92afb2SStefano Zampini PetscInt n; 2091ca92afb2SStefano Zampini PetscBool flg_row; 2092ffd830a3SStefano Zampini 2093ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2094ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 20959d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2096ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2097ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2098ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2099ca92afb2SStefano Zampini PetscInt j; 2100ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2101ffd830a3SStefano Zampini } 2102ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2103ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2104ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2105ffd830a3SStefano Zampini } 2106ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2107a3df083aSStefano Zampini if (need_benign_correction) { 2108df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2109a3df083aSStefano Zampini PetscScalar *marr; 2110a3df083aSStefano Zampini 2111a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 211214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 211314393ed6SStefano Zampini 211414393ed6SStefano Zampini | 0 0 0 | (V) 211514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 211614393ed6SStefano Zampini | 0 0 -1 | (p0) 211714393ed6SStefano Zampini 211814393ed6SStefano Zampini */ 2119df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 212014393ed6SStefano Zampini const PetscScalar *vals; 212114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 212214393ed6SStefano Zampini PetscInt n,j,nz; 212314393ed6SStefano Zampini 2124df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2125df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 212614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 212714393ed6SStefano Zampini for (j=0;j<n;j++) { 212814393ed6SStefano Zampini PetscScalar val = vals[j]; 212914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 213014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 213114393ed6SStefano Zampini } 213214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2133df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 213414393ed6SStefano Zampini } 213572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 213672b8c272SStefano Zampini } 213772b8c272SStefano Zampini if (F) { 213814393ed6SStefano Zampini /* need to correct the rhs */ 213972b8c272SStefano Zampini if (need_benign_correction) { 214072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 214172b8c272SStefano Zampini PetscScalar *marr; 214272b8c272SStefano Zampini 214372b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 21445cbda25cSStefano Zampini if (lda_rhs != n_R) { 21455cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21465cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21475cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 21485cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21495cbda25cSStefano Zampini } 21505cbda25cSStefano Zampini } else { 2151a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2152a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21535cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2154a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2155a3df083aSStefano Zampini } 21565cbda25cSStefano Zampini } 2157a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2158a3df083aSStefano Zampini } 215906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 216014393ed6SStefano Zampini /* need to correct the solution */ 2161a3df083aSStefano Zampini if (need_benign_correction) { 2162df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2163a3df083aSStefano Zampini PetscScalar *marr; 2164a3df083aSStefano Zampini 2165a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 21665cbda25cSStefano Zampini if (lda_rhs != n_R) { 21675cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21685cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21695cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 21705cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21715cbda25cSStefano Zampini } 21725cbda25cSStefano Zampini } else { 2173a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2174a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21755cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2176a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2177a3df083aSStefano Zampini } 21785cbda25cSStefano Zampini } 2179a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2180a3df083aSStefano Zampini } 218106656605SStefano Zampini } else { 218206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 218306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2184ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2185ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 218606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 218706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 218806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218906656605SStefano Zampini } 219006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 219106656605SStefano Zampini } 219280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2193ffd830a3SStefano Zampini /* S_VV and S_CV */ 219406656605SStefano Zampini if (n_constraints) { 219506656605SStefano Zampini Mat B; 219680677318SStefano Zampini 2197ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 219880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2199ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2200ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 220180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 220480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 220580677318SStefano Zampini } 2206ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 220780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 220880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2209ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 221080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 221106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2212ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2213ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 221406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 221506656605SStefano Zampini } 221604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 221704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2218511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 221904708bb6SStefano Zampini } 2220ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2221ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2222ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2223ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2224ffd830a3SStefano Zampini } 222506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 222614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 222714393ed6SStefano Zampini if (need_benign_correction) { 2228df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 222914393ed6SStefano Zampini PetscScalar *marr,*sums; 223014393ed6SStefano Zampini 223114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 223214393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2233df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 223414393ed6SStefano Zampini const PetscScalar *vals; 223514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 223614393ed6SStefano Zampini PetscInt n,j,nz; 223714393ed6SStefano Zampini 2238df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2239df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 224014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 224114393ed6SStefano Zampini PetscInt k; 224214393ed6SStefano Zampini sums[j] = 0.; 224314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 224414393ed6SStefano Zampini } 224514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 224614393ed6SStefano Zampini for (j=0;j<n;j++) { 224714393ed6SStefano Zampini PetscScalar val = vals[j]; 224814393ed6SStefano Zampini PetscInt k; 224914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 225014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 225114393ed6SStefano Zampini } 225214393ed6SStefano Zampini } 225314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2254df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 225514393ed6SStefano Zampini } 225614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 225714393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 225814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 225914393ed6SStefano Zampini } 226080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 226106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 226206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 226306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 226406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 226506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 226606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 226706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2268d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2269019a44ceSStefano Zampini } else { 2270d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2271d16cbb6bSStefano Zampini } 227221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2273d16cbb6bSStefano Zampini 227406656605SStefano Zampini /* coarse basis functions */ 227506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 227616f15bc4SStefano Zampini PetscScalar *y; 227716f15bc4SStefano Zampini 2278ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 227906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 228006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 228106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 228206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 228306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 228406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 228506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 228606656605SStefano Zampini 228706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 22884f1b2e48SStefano Zampini PetscInt j; 22894f1b2e48SStefano Zampini 229006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 229106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 229206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 229306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 229406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 22954f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 229606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 229706656605SStefano Zampini } 229806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 229906656605SStefano Zampini } 230004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 230104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 230206656605SStefano Zampini } 23035cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 230406656605SStefano Zampini 230506656605SStefano Zampini if (n_constraints) { 230606656605SStefano Zampini Mat B; 230706656605SStefano Zampini 2308ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 230906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 231080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 231106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 231206656605SStefano Zampini if (n_vertices) { 231380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 231480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 231580677318SStefano Zampini } else { 231680677318SStefano Zampini Mat S_VCt; 231780677318SStefano Zampini 2318ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2319ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 232072b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2321ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2322ffd830a3SStefano Zampini } 232380677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 232480677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 232580677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 232680677318SStefano Zampini } 232706656605SStefano Zampini } 232806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 232906656605SStefano Zampini /* coarse basis functions */ 233006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 233106656605SStefano Zampini PetscScalar *y; 233206656605SStefano Zampini 2333ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 233406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 233506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 233606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 233906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 234006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23414f1b2e48SStefano Zampini PetscInt j; 23424f1b2e48SStefano Zampini 234306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 234406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 234506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 234606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 234706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23484f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 234906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 235006656605SStefano Zampini } 235106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 235206656605SStefano Zampini } 235306656605SStefano Zampini } 235480677318SStefano Zampini if (n_constraints) { 235580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 235680677318SStefano Zampini } 23574f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 235872b8c272SStefano Zampini 235972b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 236072b8c272SStefano Zampini if (pcbddc->benign_n) { 236172b8c272SStefano Zampini Mat B0_B,B0_BPHI; 236272b8c272SStefano Zampini IS is_dummy; 236372b8c272SStefano Zampini PetscScalar *data; 236472b8c272SStefano Zampini PetscInt j; 236572b8c272SStefano Zampini 236672b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 236772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 236872b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 236972b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 237086c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 237172b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 237272b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 237372b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 237472b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 237572b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 237672b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 237772b8c272SStefano Zampini } 237872b8c272SStefano Zampini } 237972b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 238072b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 238172b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 238272b8c272SStefano Zampini } 2383019a44ceSStefano Zampini 238406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 23853301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2386ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2387ffd830a3SStefano Zampini PetscScalar *marray; 238806656605SStefano Zampini 238906656605SStefano Zampini if (n_constraints) { 2390ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 239106656605SStefano Zampini 2392af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 239306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2394ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 239516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 239606656605SStefano Zampini if (n_vertices) { 2397ffd830a3SStefano Zampini Mat S_VCT; 239806656605SStefano Zampini 239906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2400ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 240116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 240206656605SStefano Zampini } 2403ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 24045b782168SStefano Zampini } else { 24055b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 240606656605SStefano Zampini } 240716f15bc4SStefano Zampini if (n_vertices && n_R) { 2408ffd830a3SStefano Zampini PetscScalar *av,*marray; 2409ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2410ffd830a3SStefano Zampini PetscInt n; 2411ffd830a3SStefano Zampini PetscBool flg_row; 241206656605SStefano Zampini 2413ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2414af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2415ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2416ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2417ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2418ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2419ffd830a3SStefano Zampini PetscInt j; 2420ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2421ffd830a3SStefano Zampini } 2422ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2423ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2424ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 242506656605SStefano Zampini } 242606656605SStefano Zampini 2427ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2428ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2429ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2430ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2431ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 243206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 243306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 243406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 243506656605SStefano Zampini } 2436ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 24375b782168SStefano Zampini if (B_C) { 2438ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2439ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2440ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2441ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2442ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2443ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2444ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 244506656605SStefano Zampini } 2446ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 24475b782168SStefano Zampini } 24485b782168SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 24495b782168SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 24505b782168SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 24515b782168SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 24525b782168SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 24535b782168SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 24545b782168SStefano Zampini } 24555b782168SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 245606656605SStefano Zampini /* coarse basis functions */ 245706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 245806656605SStefano Zampini PetscScalar *y; 245906656605SStefano Zampini 2460ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 246106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 246206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 246306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 246506656605SStefano Zampini if (i<n_vertices) { 246606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 246706656605SStefano Zampini } 246806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 246906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 247006656605SStefano Zampini 247106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 247206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 247306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 247406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 247506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 247606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 247706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 247806656605SStefano Zampini } 247906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 248006656605SStefano Zampini } 2481ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2482ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 248306656605SStefano Zampini } 2484d62866d3SStefano Zampini /* free memory */ 248588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 248606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 248706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 248806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 248906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2490d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2491d62866d3SStefano Zampini if (n_vertices) { 2492d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2493d62866d3SStefano Zampini } 2494d62866d3SStefano Zampini if (n_constraints) { 2495d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2496d62866d3SStefano Zampini } 249788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 249888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 249988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2500d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 250188ebb749SStefano Zampini Mat coarse_sub_mat; 250225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 250388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 250488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 250588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 25068bec7fa6SStefano Zampini Mat C_B,CPHI; 25078bec7fa6SStefano Zampini IS is_dummy; 25088bec7fa6SStefano Zampini Vec mones; 250988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 251088ebb749SStefano Zampini PetscReal real_value; 251188ebb749SStefano Zampini 2512a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2513a3df083aSStefano Zampini Mat A; 2514a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2515a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2516a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2517a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2518a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2519a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2520a3df083aSStefano Zampini } else { 252188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 252288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 252388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 252488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2525a3df083aSStefano Zampini } 252688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 252788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2528ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 252988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 253088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 253188ebb749SStefano Zampini } 253288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 253388ebb749SStefano Zampini 253425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 25353301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 253625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2537ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 253888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 253988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 254088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 254188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 254288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 254388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 254488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 254588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 254688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 254788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 254888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 254988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 255088ebb749SStefano Zampini } else { 255188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 255288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 255388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 255488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 255588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 255688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 255788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 255888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 255988ebb749SStefano Zampini } 256088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 256188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 256288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2563511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 25644f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2565fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2566d12edf2fSStefano Zampini PetscScalar *data,*data2; 25674f1b2e48SStefano Zampini PetscInt j; 2568d12edf2fSStefano Zampini 25694f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2570fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2571d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 257286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2573d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2574d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 25754f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 25764f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2577d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 25784f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 25794f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 25804f1b2e48SStefano Zampini } 2581d12edf2fSStefano Zampini } 2582d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2583d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2584d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2585d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2586d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2587d12edf2fSStefano Zampini } 2588d12edf2fSStefano Zampini #if 0 2589d12edf2fSStefano Zampini { 2590d12edf2fSStefano Zampini PetscViewer viewer; 2591d12edf2fSStefano Zampini char filename[256]; 2592ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2593d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2594d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2595ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2596ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2597ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2598d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 259972b8c272SStefano Zampini if (save_change) { 260072b8c272SStefano Zampini Mat phi_B; 260172b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 260272b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 260372b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 260472b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 260572b8c272SStefano Zampini } else { 2606ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2607ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 260872b8c272SStefano Zampini } 2609ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2610ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2611ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2612ffd830a3SStefano Zampini } 2613ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2614ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2615ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2616ffd830a3SStefano Zampini } 261772b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2618ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2619ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2620ffd830a3SStefano Zampini } 2621d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2622d12edf2fSStefano Zampini } 2623d12edf2fSStefano Zampini #endif 262481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 26258bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 26261575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 262706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 26288bec7fa6SStefano Zampini 26298bec7fa6SStefano Zampini /* check constraints */ 2630a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2631a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 26324f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 26338bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2634a00504b5SStefano Zampini } else { 2635a00504b5SStefano Zampini PetscScalar *data; 2636a00504b5SStefano Zampini Mat tmat; 2637a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2638a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2639a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2640a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2641a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2642a00504b5SStefano Zampini } 26438bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 26448bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 26458bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 26468bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2647bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2648ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2649bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2650bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2651bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2652bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2653bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 265488ebb749SStefano Zampini } 26558bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 26568bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 26578bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 26588bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 265925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 266088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 266188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 266288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 266388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 266488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 266588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 266688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 266788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 266888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 266988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2670ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 267188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 267288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 267388ebb749SStefano Zampini } 267488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 267588ebb749SStefano Zampini } 26768629588bSStefano Zampini /* get back data */ 26778629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 267888ebb749SStefano Zampini PetscFunctionReturn(0); 267988ebb749SStefano Zampini } 268088ebb749SStefano Zampini 268188ebb749SStefano Zampini #undef __FUNCT__ 2682d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2683d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2684aa0d41d4SStefano Zampini { 2685d65f70fdSStefano Zampini Mat *work_mat; 2686d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2687d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2688c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2689aa0d41d4SStefano Zampini PetscErrorCode ierr; 2690aa0d41d4SStefano Zampini 2691aa0d41d4SStefano Zampini PetscFunctionBegin; 2692d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2693d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2694d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2695d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2696aa0d41d4SStefano Zampini 2697d65f70fdSStefano Zampini if (!rsorted) { 2698906d46d4SStefano Zampini const PetscInt *idxs; 2699906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2700aa0d41d4SStefano Zampini 2701d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2702d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2703d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2704d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2705aa0d41d4SStefano Zampini } 2706d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2707d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2708d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2709d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2710aa0d41d4SStefano Zampini } 2711d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2712d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2713d65f70fdSStefano Zampini } else { 2714d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2715d65f70fdSStefano Zampini isrow_s = isrow; 2716aa0d41d4SStefano Zampini } 2717906d46d4SStefano Zampini 2718d65f70fdSStefano Zampini if (!csorted) { 2719d65f70fdSStefano Zampini if (isrow == iscol) { 2720d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2721d65f70fdSStefano Zampini iscol_s = isrow_s; 2722d65f70fdSStefano Zampini } else { 2723d65f70fdSStefano Zampini const PetscInt *idxs; 2724d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2725906d46d4SStefano Zampini 2726d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2727d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2728d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2729d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2730d65f70fdSStefano Zampini } 2731d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2732d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2733d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2734d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2735d65f70fdSStefano Zampini } 2736d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2737d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2738d65f70fdSStefano Zampini } 2739d65f70fdSStefano Zampini } else { 2740d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2741d65f70fdSStefano Zampini iscol_s = iscol; 2742d65f70fdSStefano Zampini } 2743d65f70fdSStefano Zampini 2744d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2745d65f70fdSStefano Zampini 2746d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2747906d46d4SStefano Zampini Mat new_mat; 2748d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2749906d46d4SStefano Zampini 2750d65f70fdSStefano Zampini if (!rsorted) { 2751d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2752d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2753d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2754d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2755906d46d4SStefano Zampini } 2756d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2757d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2758d65f70fdSStefano Zampini } else { 2759d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2760906d46d4SStefano Zampini } 2761d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2762d65f70fdSStefano Zampini 2763d65f70fdSStefano Zampini if (!csorted) { 2764d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2765d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2766d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2767d65f70fdSStefano Zampini } else { 2768d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2769d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2770d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2771d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2772d65f70fdSStefano Zampini } 2773d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2774d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2775d65f70fdSStefano Zampini } 2776d65f70fdSStefano Zampini } else { 2777d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2778d65f70fdSStefano Zampini } 2779d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2780d65f70fdSStefano Zampini 2781d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2782d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2783d65f70fdSStefano Zampini work_mat[0] = new_mat; 2784d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2785d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2786d65f70fdSStefano Zampini } 2787d65f70fdSStefano Zampini 2788d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2789d65f70fdSStefano Zampini *B = work_mat[0]; 2790d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2791d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2792d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2793d65f70fdSStefano Zampini PetscFunctionReturn(0); 2794d65f70fdSStefano Zampini } 2795d65f70fdSStefano Zampini 2796d65f70fdSStefano Zampini #undef __FUNCT__ 27975e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 27985e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2799aa0d41d4SStefano Zampini { 2800aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 28015e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2802d65f70fdSStefano Zampini Mat new_mat; 28035e8657edSStefano Zampini IS is_local,is_global; 2804d65f70fdSStefano Zampini PetscInt local_size; 2805d65f70fdSStefano Zampini PetscBool isseqaij; 2806aa0d41d4SStefano Zampini PetscErrorCode ierr; 2807aa0d41d4SStefano Zampini 2808aa0d41d4SStefano Zampini PetscFunctionBegin; 2809aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 28105e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 28115e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2812b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2813aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2814d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2815aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2816906d46d4SStefano Zampini 2817906d46d4SStefano Zampini /* check */ 2818906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2819906d46d4SStefano Zampini Vec x,x_change; 2820906d46d4SStefano Zampini PetscReal error; 2821906d46d4SStefano Zampini 28225e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2823906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 28245e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2825e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2826e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2827d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2828e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2829e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2830906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2831906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2832906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2833906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2834906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2835906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2836906d46d4SStefano Zampini } 2837906d46d4SStefano Zampini 283822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 28399b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 284022d5777bSStefano Zampini if (isseqaij) { 2841a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2842a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2843aa0d41d4SStefano Zampini } else { 2844a00504b5SStefano Zampini Mat work_mat; 28451cf9b237SStefano Zampini 2846a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2847aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2848a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 28491d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2850aa0d41d4SStefano Zampini } 28513301b35fSStefano Zampini if (matis->A->symmetric_set) { 28523301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2853e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 28543301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2855e496cd5dSStefano Zampini #endif 28563301b35fSStefano Zampini } 2857d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2858aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2859aa0d41d4SStefano Zampini } 2860aa0d41d4SStefano Zampini 2861aa0d41d4SStefano Zampini #undef __FUNCT__ 2862a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 28638ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2864a64d13efSStefano Zampini { 2865a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2866a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2867d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 286853892102SStefano Zampini PetscInt *idx_R_local=NULL; 28693a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 28703a50541eSStefano Zampini PetscInt vbs,bs; 28716816873aSStefano Zampini PetscBT bitmask=NULL; 2872a64d13efSStefano Zampini PetscErrorCode ierr; 2873a64d13efSStefano Zampini 2874a64d13efSStefano Zampini PetscFunctionBegin; 2875b23d619eSStefano Zampini /* 2876b23d619eSStefano Zampini No need to setup local scatters if 2877b23d619eSStefano Zampini - primal space is unchanged 2878b23d619eSStefano Zampini AND 2879b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2880b23d619eSStefano Zampini AND 2881b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2882b23d619eSStefano Zampini */ 2883b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2884f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2885f4ddd8eeSStefano Zampini } 2886f4ddd8eeSStefano Zampini /* destroy old objects */ 2887f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2888f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2889f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2890a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2891b371cd4fSStefano Zampini n_B = pcis->n_B; 2892b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2893b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 28943a50541eSStefano Zampini 2895a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 28966816873aSStefano Zampini 289753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2898b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2899854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2900a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2901a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 29020e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2903a64d13efSStefano Zampini } 2904a64d13efSStefano Zampini 2905a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 29064641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 29076816873aSStefano Zampini idx_R_local[n_R++] = i; 2908a64d13efSStefano Zampini } 2909a64d13efSStefano Zampini } 2910df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2911df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 29126816873aSStefano Zampini 2913df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2914df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 29156816873aSStefano Zampini } 29163a50541eSStefano Zampini 29173a50541eSStefano Zampini /* Block code */ 29183a50541eSStefano Zampini vbs = 1; 29193a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 29203a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 29213a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 29223a50541eSStefano Zampini PetscInt *vary; 2923b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2924785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 29253a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2926d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2927d3df7717SStefano 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 */ 29280e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2929d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 29303a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 29313a50541eSStefano Zampini is_blocked = PETSC_FALSE; 29323a50541eSStefano Zampini break; 29333a50541eSStefano Zampini } 29343a50541eSStefano Zampini } 2935d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2936d3df7717SStefano Zampini } else { 2937d3df7717SStefano Zampini /* Verify directly the R set */ 2938d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2939d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2940d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2941d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2942d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2943d3df7717SStefano Zampini break; 2944d3df7717SStefano Zampini } 2945d3df7717SStefano Zampini } 2946d3df7717SStefano Zampini } 2947d3df7717SStefano Zampini } 29483a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 29493a50541eSStefano Zampini vbs = bs; 29503a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 29513a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 29523a50541eSStefano Zampini } 29533a50541eSStefano Zampini } 29543a50541eSStefano Zampini } 29553a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2956b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 2957df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 295853892102SStefano Zampini 2959df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2960df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 296153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2962df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 296353892102SStefano Zampini } else { 29643a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 296553892102SStefano Zampini } 2966a64d13efSStefano Zampini 2967a64d13efSStefano Zampini /* print some info if requested */ 2968a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2969a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2970a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29711575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2972a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2973a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 29744f1b2e48SStefano 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); 2975a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2976a64d13efSStefano Zampini } 2977a64d13efSStefano Zampini 2978a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2979b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 29806816873aSStefano Zampini IS is_aux1,is_aux2; 29816816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 29826816873aSStefano Zampini 29833a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2984854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2985854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2986a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 29874641a718SStefano Zampini for (i=0; i<n_D; i++) { 29884641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 29894641a718SStefano Zampini } 2990a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2991a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 29924641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 29934641a718SStefano Zampini aux_array1[j++] = i; 2994a64d13efSStefano Zampini } 2995a64d13efSStefano Zampini } 2996a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2997a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2998a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 29994641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 30004641a718SStefano Zampini aux_array2[j++] = i; 3001a64d13efSStefano Zampini } 3002a64d13efSStefano Zampini } 3003a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3004a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 3005a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 3006a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3007a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 3008a64d13efSStefano Zampini 30098eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3010785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 3011a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30124641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 30134641a718SStefano Zampini aux_array1[j++] = i; 3014a64d13efSStefano Zampini } 3015a64d13efSStefano Zampini } 3016a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3017a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 3018a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3019a64d13efSStefano Zampini } 30204641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 30213a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3022d62866d3SStefano Zampini } else { 3023df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 30246816873aSStefano Zampini IS tis; 30256816873aSStefano Zampini PetscInt schur_size; 30266816873aSStefano Zampini 3027df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 30286816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 3029df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 30306816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 30316816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 30326816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 30336816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 30346816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 3035d62866d3SStefano Zampini } 3036d62866d3SStefano Zampini } 3037a64d13efSStefano Zampini PetscFunctionReturn(0); 3038a64d13efSStefano Zampini } 3039a64d13efSStefano Zampini 3040304d26faSStefano Zampini 3041304d26faSStefano Zampini #undef __FUNCT__ 3042304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 3043684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 3044304d26faSStefano Zampini { 3045304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3046304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3047304d26faSStefano Zampini PC pc_temp; 3048304d26faSStefano Zampini Mat A_RR; 3049f4ddd8eeSStefano Zampini MatReuse reuse; 3050304d26faSStefano Zampini PetscScalar m_one = -1.0; 3051304d26faSStefano Zampini PetscReal value; 305204708bb6SStefano Zampini PetscInt n_D,n_R; 3053c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 3054304d26faSStefano Zampini PetscErrorCode ierr; 3055e604994aSStefano Zampini /* prefixes stuff */ 3056312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 3057e604994aSStefano Zampini size_t len; 3058304d26faSStefano Zampini 3059304d26faSStefano Zampini PetscFunctionBegin; 3060304d26faSStefano Zampini 3061e604994aSStefano Zampini /* compute prefixes */ 3062e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 3063e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 3064e604994aSStefano Zampini if (!pcbddc->current_level) { 3065e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3066e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3067e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3068e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3069e604994aSStefano Zampini } else { 3070e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3071312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3072e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3073e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 3074312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3075312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 307634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 307734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3078e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3079e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3080e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 3081e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 3082e604994aSStefano Zampini } 3083e604994aSStefano Zampini 3084304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3085684f6988SStefano Zampini if (dirichlet) { 3086d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3087450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 3088b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 3089450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3090450f8f5eSStefano Zampini } 3091450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3092a3df083aSStefano Zampini Mat A_IIn; 3093a3df083aSStefano Zampini 3094a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3095a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3096a3df083aSStefano Zampini pcis->A_II = A_IIn; 3097a3df083aSStefano Zampini } 3098450f8f5eSStefano Zampini } 30993301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31003301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3101964fefecSStefano Zampini } 3102ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3103964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3104304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3105304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3106304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3107304d26faSStefano Zampini /* default */ 3108304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3109e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 31109577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3111304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 31129577ea80SStefano Zampini if (issbaij) { 31139577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 31149577ea80SStefano Zampini } else { 3115304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31169577ea80SStefano Zampini } 3117304d26faSStefano Zampini /* Allow user's customization */ 3118304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3119304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3120304d26faSStefano Zampini } 3121d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3122b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3123df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3124d62866d3SStefano Zampini 3125df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3126d5574798SStefano Zampini } 3127304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3128304d26faSStefano Zampini if (!n_D) { 3129304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3130304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3131304d26faSStefano Zampini } 3132304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3133304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3134304d26faSStefano Zampini /* set ksp_D into pcis data */ 3135304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3136304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3137304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3138684f6988SStefano Zampini } 3139304d26faSStefano Zampini 3140304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3141684f6988SStefano Zampini A_RR = 0; 3142684f6988SStefano Zampini if (neumann) { 3143d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 314404708bb6SStefano Zampini PetscInt ibs,mbs; 314504708bb6SStefano Zampini PetscBool issbaij; 314604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3147f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 31488ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3149f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3150f4ddd8eeSStefano Zampini PetscInt nn_R; 315181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3152f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3153f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3154f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3155f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3156f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3157f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3158f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3159727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3160f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3161f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3162f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3163f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3164f4ddd8eeSStefano Zampini } 3165f4ddd8eeSStefano Zampini } 3166f4ddd8eeSStefano Zampini /* last check */ 3167d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3168f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3169f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3170f4ddd8eeSStefano Zampini } 3171f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3172f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3173f4ddd8eeSStefano Zampini } 3174a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3175af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3176af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 317704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 317804708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 317904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 318004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 318104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3182af732b37SStefano Zampini } else { 3183511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 31846816873aSStefano Zampini } 318504708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 318604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 318704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 318804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 318904708bb6SStefano Zampini } else { 3190511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 319104708bb6SStefano Zampini } 319204708bb6SStefano Zampini } 3193a00504b5SStefano Zampini /* extract A_RR */ 3194b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3195a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3196a00504b5SStefano Zampini 3197a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 319816e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3199a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 320016e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 320116e386b8SStefano Zampini } else { 3202a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3203a00504b5SStefano Zampini } 3204a00504b5SStefano Zampini } else { 3205a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3206a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3207a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3208a00504b5SStefano Zampini } 3209a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3210f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 321116e386b8SStefano Zampini } 32123301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 32133301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 32146816873aSStefano Zampini } 3215f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3216304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3217304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3218304d26faSStefano Zampini /* default */ 3219304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3220e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3221304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 32229577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 32239577ea80SStefano Zampini if (issbaij) { 32249577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 32259577ea80SStefano Zampini } else { 3226304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 32279577ea80SStefano Zampini } 3228304d26faSStefano Zampini /* Allow user's customization */ 3229304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3230304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3231304d26faSStefano Zampini } 3232304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3233304d26faSStefano Zampini if (!n_R) { 3234304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3235304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3236304d26faSStefano Zampini } 32375cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3238df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3239b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3240df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3241d62866d3SStefano Zampini 3242df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3243d62866d3SStefano Zampini } 3244304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3245304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3246684f6988SStefano Zampini } 3247304d26faSStefano Zampini 3248684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3249684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32501575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3251684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3252684f6988SStefano Zampini } 3253c7017625SStefano Zampini 3254c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3255c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3256c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3257c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3258c7017625SStefano Zampini } 3259c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3260c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3261c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3262c7017625SStefano Zampini } 3263c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3264c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3265c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3266c7017625SStefano Zampini } 3267c7017625SStefano Zampini 3268c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3269c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3270684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 32710fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 32720fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 32730fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 32740fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 32750fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3276e604994aSStefano 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); 3277c7017625SStefano Zampini if (check_corr[0]) { 3278c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3279c7017625SStefano Zampini } 3280304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3281304d26faSStefano Zampini } 3282684f6988SStefano Zampini if (neumann) { /* Neumann */ 32830fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 32840fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 32850fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 32860fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 32870fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3288e604994aSStefano 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); 3289c7017625SStefano Zampini if (check_corr[1]) { 3290c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3291c7017625SStefano Zampini } 3292304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3293304d26faSStefano Zampini } 3294684f6988SStefano Zampini } 32955cbda25cSStefano Zampini /* free Neumann problem's matrix */ 32965cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3297304d26faSStefano Zampini PetscFunctionReturn(0); 3298304d26faSStefano Zampini } 3299304d26faSStefano Zampini 3300304d26faSStefano Zampini #undef __FUNCT__ 3301ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 330280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3303674ae819SStefano Zampini { 3304674ae819SStefano Zampini PetscErrorCode ierr; 3305674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3306be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3307b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 3308674ae819SStefano Zampini 3309674ae819SStefano Zampini PetscFunctionBegin; 3310b334f244SStefano Zampini if (!reuse_solver) { 331180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 331220c7b377SStefano Zampini } 331380677318SStefano Zampini if (!pcbddc->switch_static) { 331480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 331580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 331680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 331720c7b377SStefano Zampini } 3318b334f244SStefano Zampini if (!reuse_solver) { 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); 332120c7b377SStefano Zampini } else { 3322df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3323be83ff47SStefano Zampini 3324df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3325df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 332620c7b377SStefano Zampini } 3327be83ff47SStefano Zampini } else { 332880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 332980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 333080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 333180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 333280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 333380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 333480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 333580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 333680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3337674ae819SStefano Zampini } 3338674ae819SStefano Zampini } 3339b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 334080677318SStefano Zampini if (applytranspose) { 334180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 334280677318SStefano Zampini } else { 334380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 334480677318SStefano Zampini } 3345be83ff47SStefano Zampini } else { 3346df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3347be83ff47SStefano Zampini 3348be83ff47SStefano Zampini if (applytranspose) { 3349df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3350be83ff47SStefano Zampini } else { 3351df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3352be83ff47SStefano Zampini } 3353be83ff47SStefano Zampini } 335480677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 335580677318SStefano Zampini if (!pcbddc->switch_static) { 3356b334f244SStefano Zampini if (!reuse_solver) { 335780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3359be83ff47SStefano Zampini } else { 3360df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3361be83ff47SStefano Zampini 3362df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3363df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3364be83ff47SStefano Zampini } 336580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 336680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 336780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 336880677318SStefano Zampini } 336980677318SStefano Zampini } else { 337080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 337180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 337280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 337380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 337480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 337580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 337680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 337780677318SStefano Zampini } 337880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 337980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 338080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 338180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3382674ae819SStefano Zampini } 3383674ae819SStefano Zampini PetscFunctionReturn(0); 3384674ae819SStefano Zampini } 3385674ae819SStefano Zampini 3386dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3387674ae819SStefano Zampini #undef __FUNCT__ 3388674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3389dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3390674ae819SStefano Zampini { 3391674ae819SStefano Zampini PetscErrorCode ierr; 3392674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3393674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3394674ae819SStefano Zampini const PetscScalar zero = 0.0; 3395674ae819SStefano Zampini 3396674ae819SStefano Zampini PetscFunctionBegin; 3397dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 33984fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3399dc359a40SStefano Zampini if (applytranspose) { 3400674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 34018eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3402dc359a40SStefano Zampini } else { 3403674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3404674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 340515aaf578SStefano Zampini } 34064fee134fSStefano Zampini } else { 34074fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 34084fee134fSStefano Zampini } 3409efc2fbd9SStefano Zampini 3410efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 34114f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3412efc2fbd9SStefano Zampini PetscScalar *array; 34134f1b2e48SStefano Zampini PetscInt j; 3414efc2fbd9SStefano Zampini 3415efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34164f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3417efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3418efc2fbd9SStefano Zampini } 3419efc2fbd9SStefano Zampini 342012edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 342112edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 342212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342412edc857SStefano Zampini 34259f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 342612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 342751694757SStefano Zampini Mat coarse_mat; 3428964fefecSStefano Zampini Vec rhs,sol; 342951694757SStefano Zampini MatNullSpace nullsp; 343027b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3431964fefecSStefano Zampini 343227b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 343327b6a85dSStefano Zampini PC coarse_pc; 343427b6a85dSStefano Zampini 343527b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 343627b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 343727b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 343827b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 343927b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 344027b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 34413bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 344227b6a85dSStefano Zampini } 344327b6a85dSStefano Zampini } 3444964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3445964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 344651694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 344751694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 344851694757SStefano Zampini if (nullsp) { 344951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 345051694757SStefano Zampini } 345112edc857SStefano Zampini if (applytranspose) { 34521f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only) { 34531f4df5f7SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 34542701bc32SStefano Zampini } else { 3455964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 34562701bc32SStefano Zampini } 34572701bc32SStefano Zampini } else { 34581f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 34592701bc32SStefano Zampini PC coarse_pc; 34602701bc32SStefano Zampini 34612701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 34622701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 34633e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 34642701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 346512edc857SStefano Zampini } else { 3466964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 346712edc857SStefano Zampini } 34682701bc32SStefano Zampini } 34691d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 347027b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 347127b6a85dSStefano Zampini PC coarse_pc; 347227b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 347327b6a85dSStefano Zampini 347427b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 347527b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 347627b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 34773bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 347827b6a85dSStefano Zampini } 347951694757SStefano Zampini if (nullsp) { 348051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 348151694757SStefano Zampini } 348212edc857SStefano Zampini } 3483674ae819SStefano Zampini 3484674ae819SStefano Zampini /* Local solution on R nodes */ 34854fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 348680677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 34879f00e9b4SStefano Zampini } 34889f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 34899f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 349012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3491674ae819SStefano Zampini 34924fee134fSStefano Zampini /* Sum contributions from the two levels */ 34934fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3494dc359a40SStefano Zampini if (applytranspose) { 3495dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3496dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3497dc359a40SStefano Zampini } else { 3498674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 34998eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3500dc359a40SStefano Zampini } 3501efc2fbd9SStefano Zampini /* store p0 */ 35024f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3503efc2fbd9SStefano Zampini PetscScalar *array; 35044f1b2e48SStefano Zampini PetscInt j; 3505efc2fbd9SStefano Zampini 3506efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 35074f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3508efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3509efc2fbd9SStefano Zampini } 35104fee134fSStefano Zampini } else { /* expand the coarse solution */ 35114fee134fSStefano Zampini if (applytranspose) { 35124fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35134fee134fSStefano Zampini } else { 35144fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35154fee134fSStefano Zampini } 35164fee134fSStefano Zampini } 3517674ae819SStefano Zampini PetscFunctionReturn(0); 3518674ae819SStefano Zampini } 3519674ae819SStefano Zampini 3520674ae819SStefano Zampini #undef __FUNCT__ 3521674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 352212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3523674ae819SStefano Zampini { 3524674ae819SStefano Zampini PetscErrorCode ierr; 3525674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 352658da7f69SStefano Zampini PetscScalar *array; 352712edc857SStefano Zampini Vec from,to; 3528674ae819SStefano Zampini 3529674ae819SStefano Zampini PetscFunctionBegin; 353012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 353112edc857SStefano Zampini from = pcbddc->coarse_vec; 353212edc857SStefano Zampini to = pcbddc->vec1_P; 353312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 353412edc857SStefano Zampini Vec tvec; 353558da7f69SStefano Zampini 353658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 353758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 353812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 353958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 354058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 354158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 354212edc857SStefano Zampini } 354312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 354412edc857SStefano Zampini from = pcbddc->vec1_P; 354512edc857SStefano Zampini to = pcbddc->coarse_vec; 354612edc857SStefano Zampini } 354712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3548674ae819SStefano Zampini PetscFunctionReturn(0); 3549674ae819SStefano Zampini } 3550674ae819SStefano Zampini 3551674ae819SStefano Zampini #undef __FUNCT__ 3552674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 355312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3554674ae819SStefano Zampini { 3555674ae819SStefano Zampini PetscErrorCode ierr; 3556674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 355758da7f69SStefano Zampini PetscScalar *array; 355812edc857SStefano Zampini Vec from,to; 3559674ae819SStefano Zampini 3560674ae819SStefano Zampini PetscFunctionBegin; 356112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 356212edc857SStefano Zampini from = pcbddc->coarse_vec; 356312edc857SStefano Zampini to = pcbddc->vec1_P; 356412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 356512edc857SStefano Zampini from = pcbddc->vec1_P; 356612edc857SStefano Zampini to = pcbddc->coarse_vec; 356712edc857SStefano Zampini } 356812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 356912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 357012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 357112edc857SStefano Zampini Vec tvec; 357258da7f69SStefano Zampini 357312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 357458da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 357558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 357658da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 357758da7f69SStefano Zampini } 357858da7f69SStefano Zampini } else { 357958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 358058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 358112edc857SStefano Zampini } 358212edc857SStefano Zampini } 3583674ae819SStefano Zampini PetscFunctionReturn(0); 3584674ae819SStefano Zampini } 3585674ae819SStefano Zampini 3586984c4197SStefano Zampini /* uncomment for testing purposes */ 3587984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3588674ae819SStefano Zampini #undef __FUNCT__ 3589674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3590674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3591674ae819SStefano Zampini { 3592674ae819SStefano Zampini PetscErrorCode ierr; 3593674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3594674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3595674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3596984c4197SStefano Zampini /* one and zero */ 3597984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3598984c4197SStefano Zampini /* space to store constraints and their local indices */ 35999162d606SStefano Zampini PetscScalar *constraints_data; 36009162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 36019162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 36029162d606SStefano Zampini PetscInt *constraints_n; 3603984c4197SStefano Zampini /* iterators */ 3604b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3605984c4197SStefano Zampini /* BLAS integers */ 3606e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3607e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3608c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3609727cdba6SStefano Zampini /* reuse */ 36100e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 36110e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3612984c4197SStefano Zampini /* change of basis */ 3613b3d85658SStefano Zampini PetscBool qr_needed; 36149162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3615984c4197SStefano Zampini /* auxiliary stuff */ 361664efe560SStefano Zampini PetscInt *nnz,*is_indices; 36178a0068c3SStefano Zampini PetscInt ncc; 3618984c4197SStefano Zampini /* some quantities */ 361945a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3620a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3621984c4197SStefano Zampini 3622674ae819SStefano Zampini PetscFunctionBegin; 36238e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 36248e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 36258e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 362616909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3627088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3628088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 36290e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 36300e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 36310e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36320e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36330e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3634088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3635cf5a6209SStefano Zampini 3636cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 36379162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3638cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3639cf5a6209SStefano Zampini const Vec *nearnullvecs; 3640cf5a6209SStefano Zampini Vec *localnearnullsp; 3641cf5a6209SStefano Zampini PetscScalar *array; 3642cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3643cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3644674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3645b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3646674ae819SStefano Zampini PetscScalar *work; 3647674ae819SStefano Zampini PetscReal *singular_vals; 3648674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3649674ae819SStefano Zampini PetscReal *rwork; 3650674ae819SStefano Zampini #endif 3651674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3652674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3653674ae819SStefano Zampini #else 3654964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3655964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3656674ae819SStefano Zampini #endif 3657674ae819SStefano Zampini 3658674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3659d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3660e4d548c7SStefano Zampini /* print some info */ 36611f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3662e4d548c7SStefano Zampini PetscInt nv; 3663e4d548c7SStefano Zampini 3664e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 3665e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3666e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3667e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3668e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3669e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3670e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3671e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3672e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3673e4d548c7SStefano Zampini } 3674e4d548c7SStefano Zampini 3675d06fc5fdSStefano Zampini /* free unneeded index sets */ 3676d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3677d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3678674ae819SStefano Zampini } 3679d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3680d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3681d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3682d06fc5fdSStefano Zampini } 3683d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3684d06fc5fdSStefano Zampini n_ISForEdges = 0; 3685d06fc5fdSStefano Zampini } 3686d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3687d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3688d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3689d06fc5fdSStefano Zampini } 3690d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3691d06fc5fdSStefano Zampini n_ISForFaces = 0; 3692d06fc5fdSStefano Zampini } 369370022509SStefano Zampini 3694674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3695674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3696674ae819SStefano Zampini if (nearnullsp) { 3697674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3698f4ddd8eeSStefano Zampini /* remove any stored info */ 3699f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3700f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3701f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3702f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3703f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3704473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3705f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3706f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3707f4ddd8eeSStefano Zampini } 3708984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3709984c4197SStefano Zampini nnsp_size = 0; 3710674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3711674ae819SStefano Zampini } 3712984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3713984c4197SStefano Zampini max_constraints = nnsp_size; 3714984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3715984c4197SStefano Zampini 3716674ae819SStefano Zampini /* 3717674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 37189162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 37199162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 37209162d606SStefano Zampini There can be multiple constraints per connected component 3721674ae819SStefano Zampini */ 3722674ae819SStefano Zampini n_vertices = 0; 3723674ae819SStefano Zampini if (ISForVertices) { 3724674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3725674ae819SStefano Zampini } 37269162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 37279162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 37289162d606SStefano Zampini 37299162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 37309162d606SStefano Zampini total_counts *= max_constraints; 3731674ae819SStefano Zampini total_counts += n_vertices; 37324641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 37339162d606SStefano Zampini 3734674ae819SStefano Zampini total_counts = 0; 3735674ae819SStefano Zampini max_size_of_constraint = 0; 3736674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 37379162d606SStefano Zampini IS used_is; 3738674ae819SStefano Zampini if (i<n_ISForEdges) { 37399162d606SStefano Zampini used_is = ISForEdges[i]; 3740674ae819SStefano Zampini } else { 37419162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3742674ae819SStefano Zampini } 37439162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3744674ae819SStefano Zampini total_counts += j; 3745674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3746674ae819SStefano Zampini } 37479162d606SStefano 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); 37489162d606SStefano Zampini 3749984c4197SStefano Zampini /* get local part of global near null space vectors */ 3750785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3751984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3752984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3753e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3754e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3755984c4197SStefano Zampini } 3756674ae819SStefano Zampini 3757242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3758242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3759a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3760242a89d7SStefano Zampini 3761984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3762a773dcb8SStefano Zampini if (!skip_lapack) { 3763674ae819SStefano Zampini PetscScalar temp_work; 3764911cabfeSStefano Zampini 3765674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3766984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3767785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3768785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3769785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3770674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3771785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3772674ae819SStefano Zampini #endif 3773674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3774c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3775c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3776674ae819SStefano Zampini lwork = -1; 3777674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3778674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3779c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3780674ae819SStefano Zampini #else 3781c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3782674ae819SStefano Zampini #endif 3783674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3784984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3785674ae819SStefano Zampini #else /* on missing GESVD */ 3786674ae819SStefano Zampini /* SVD */ 3787674ae819SStefano Zampini PetscInt max_n,min_n; 3788674ae819SStefano Zampini max_n = max_size_of_constraint; 3789984c4197SStefano Zampini min_n = max_constraints; 3790984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3791674ae819SStefano Zampini min_n = max_size_of_constraint; 3792984c4197SStefano Zampini max_n = max_constraints; 3793674ae819SStefano Zampini } 3794785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3795674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3796785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3797674ae819SStefano Zampini #endif 3798674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3799674ae819SStefano Zampini lwork = -1; 3800e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3801e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3802b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3803674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3804674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 38059162d606SStefano 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)); 3806674ae819SStefano Zampini #else 38079162d606SStefano 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)); 3808674ae819SStefano Zampini #endif 3809674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3810984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3811984c4197SStefano Zampini #endif /* on missing GESVD */ 3812674ae819SStefano Zampini /* Allocate optimal workspace */ 3813674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3814854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3815674ae819SStefano Zampini } 3816674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3817674ae819SStefano Zampini total_counts = 0; 38189162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 38199162d606SStefano Zampini constraints_data_ptr[0] = 0; 3820674ae819SStefano Zampini /* vertices */ 38219162d606SStefano Zampini if (n_vertices) { 3822674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 38239162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3824674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 38259162d606SStefano Zampini constraints_n[total_counts] = 1; 38269162d606SStefano Zampini constraints_data[total_counts] = 1.0; 38279162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 38289162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3829674ae819SStefano Zampini total_counts++; 3830674ae819SStefano Zampini } 3831674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3832674ae819SStefano Zampini n_vertices = total_counts; 3833674ae819SStefano Zampini } 3834984c4197SStefano Zampini 3835674ae819SStefano Zampini /* edges and faces */ 38369162d606SStefano Zampini total_counts_cc = total_counts; 3837911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 38389162d606SStefano Zampini IS used_is; 38399162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 38409162d606SStefano Zampini 3841911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 38429162d606SStefano Zampini used_is = ISForEdges[ncc]; 3843984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3844674ae819SStefano Zampini } else { 38459162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3846984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3847674ae819SStefano Zampini } 3848674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 38499162d606SStefano Zampini 38509162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 38519162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3852984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3853984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3854674ae819SStefano Zampini if (nnsp_has_cnst) { 38555b08dc53SStefano Zampini PetscScalar quad_value; 38569162d606SStefano Zampini 38579162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38589162d606SStefano Zampini idxs_copied = PETSC_TRUE; 38599162d606SStefano Zampini 3860a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3861674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3862a773dcb8SStefano Zampini } else { 3863a773dcb8SStefano Zampini quad_value = 1.0; 3864a773dcb8SStefano Zampini } 3865674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38669162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3867674ae819SStefano Zampini } 38689162d606SStefano Zampini temp_constraints++; 3869674ae819SStefano Zampini total_counts++; 3870674ae819SStefano Zampini } 3871674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3872984c4197SStefano Zampini PetscReal real_value; 38739162d606SStefano Zampini PetscScalar *ptr_to_data; 38749162d606SStefano Zampini 3875984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 38769162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3877674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38789162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3879674ae819SStefano Zampini } 3880984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3881984c4197SStefano Zampini /* check if array is null on the connected component */ 3882e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 38839162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 38845b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3885674ae819SStefano Zampini temp_constraints++; 3886674ae819SStefano Zampini total_counts++; 38879162d606SStefano Zampini if (!idxs_copied) { 38889162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38899162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3890674ae819SStefano Zampini } 3891674ae819SStefano Zampini } 38929162d606SStefano Zampini } 38939162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 389445a1bb75SStefano Zampini valid_constraints = temp_constraints; 3895eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3896a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 38979162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 38989162d606SStefano Zampini 38999162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3900a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39019162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3902a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 39039162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3904a773dcb8SStefano Zampini } else { /* perform SVD */ 3905984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 39069162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3907674ae819SStefano Zampini 3908674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3909984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3910984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3911984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3912984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3913984c4197SStefano Zampini from that computed using LAPACKgesvd 3914984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3915984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3916984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3917674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3918e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3919984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3920674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3921674ae819SStefano Zampini for (k=0;k<j+1;k++) { 39229162d606SStefano 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)); 3923674ae819SStefano Zampini } 3924674ae819SStefano Zampini } 3925e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3926e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3927e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3928674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3929c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3930674ae819SStefano Zampini #else 3931c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3932674ae819SStefano Zampini #endif 3933674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3934984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3935984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3936674ae819SStefano Zampini j = 0; 3937984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3938674ae819SStefano Zampini total_counts = total_counts-j; 393945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3940e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3941c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3942c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3943c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3944c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3945c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3946c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3947674ae819SStefano Zampini if (j<temp_constraints) { 3948984c4197SStefano Zampini PetscInt ii; 3949984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3950674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39519162d606SStefano 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)); 3952674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3953984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3954674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 39559162d606SStefano 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]; 3956674ae819SStefano Zampini } 3957674ae819SStefano Zampini } 3958674ae819SStefano Zampini } 3959674ae819SStefano Zampini #else /* on missing GESVD */ 3960e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3961e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3962b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3963674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3964674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 39659162d606SStefano 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)); 3966674ae819SStefano Zampini #else 39679162d606SStefano 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)); 3968674ae819SStefano Zampini #endif 3969984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3970674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3971984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3972e310c8b4SStefano Zampini k = temp_constraints; 3973e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3974674ae819SStefano Zampini j = 0; 3975e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 397645a1bb75SStefano Zampini valid_constraints = k-j; 3977911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3978984c4197SStefano Zampini #endif /* on missing GESVD */ 3979674ae819SStefano Zampini } 3980a773dcb8SStefano Zampini } 39819162d606SStefano Zampini /* update pointers information */ 39829162d606SStefano Zampini if (valid_constraints) { 39839162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 39849162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 39859162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 39869162d606SStefano Zampini /* set change_of_basis flag */ 398745a1bb75SStefano Zampini if (boolforchange) { 3988b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 39899162d606SStefano Zampini } 3990b3d85658SStefano Zampini total_counts_cc++; 399145a1bb75SStefano Zampini } 399245a1bb75SStefano Zampini } 3993984c4197SStefano Zampini /* free workspace */ 39948f1c130eSStefano Zampini if (!skip_lapack) { 3995984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3996984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3997984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3998984c4197SStefano Zampini #endif 3999984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 4000984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4001984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 4002984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 4003984c4197SStefano Zampini #endif 4004984c4197SStefano Zampini } 4005984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4006984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 4007984c4197SStefano Zampini } 4008984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 4009cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 4010cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4011cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4012cf5a6209SStefano Zampini } 4013cf5a6209SStefano Zampini if (n_ISForFaces) { 4014cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4015cf5a6209SStefano Zampini } 4016cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4017cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4018cf5a6209SStefano Zampini } 4019cf5a6209SStefano Zampini if (n_ISForEdges) { 4020cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4021cf5a6209SStefano Zampini } 4022cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 402308122e43SStefano Zampini } else { 402408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4025984c4197SStefano Zampini 402608122e43SStefano Zampini total_counts = 0; 402708122e43SStefano Zampini n_vertices = 0; 4028d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 4029d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 403008122e43SStefano Zampini } 403108122e43SStefano Zampini max_constraints = 0; 40329162d606SStefano Zampini total_counts_cc = 0; 403308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 403408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 40359162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 403608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 403708122e43SStefano Zampini } 40389162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 40399162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 40409162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 40419162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 404274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 40439162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 40449162d606SStefano Zampini total_counts_cc = 0; 40459162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 40469162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 40479162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 404808122e43SStefano Zampini } 404908122e43SStefano Zampini } 40509162d606SStefano Zampini #if 0 40519162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 40529162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40539162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 40549162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 40559162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 40569162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 40579162d606SStefano Zampini } 40589162d606SStefano Zampini printf("\n"); 40599162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 40609162d606SStefano Zampini } 40611b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 40628bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 40631b968477SStefano Zampini } 40641b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 40658bec7fa6SStefano 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]); 40661b968477SStefano Zampini } 406708122e43SStefano Zampini #endif 406808122e43SStefano Zampini 40698bec7fa6SStefano Zampini max_size_of_constraint = 0; 40709162d606SStefano 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]); 40719162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 407208122e43SStefano Zampini /* Change of basis */ 4073b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 407408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 407508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 407608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 4077b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 407808122e43SStefano Zampini } 407908122e43SStefano Zampini } 408008122e43SStefano Zampini } 408108122e43SStefano Zampini } 4082984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 40834f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 408408122e43SStefano Zampini 40859162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 40869162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 40876c4ed002SBarry 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); 4088674ae819SStefano Zampini 4089674ae819SStefano Zampini /* Create constraint matrix */ 4090674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 409116f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4092984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4093984c4197SStefano Zampini 4094984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4095a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4096a717540cSStefano Zampini qr_needed = PETSC_FALSE; 409774d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4098984c4197SStefano Zampini total_primal_vertices=0; 4099b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 41009162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 41019162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 410272b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 41039162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4104b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 410564efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 41069162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41079162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4108a717540cSStefano Zampini } 4109b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 411091af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4111a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4112a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4113a717540cSStefano Zampini } 4114fa434743SStefano Zampini } else { 4115b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4116fa434743SStefano Zampini } 4117a717540cSStefano Zampini } 4118b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4119b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4120674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 412170022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4122b3d85658SStefano Zampini 41234f1b2e48SStefano 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); 41240e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 41250e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4126984c4197SStefano Zampini 4127984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 412874d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4129785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4130984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 413174d5cdf7SStefano Zampini 4132984c4197SStefano Zampini j = total_primal_vertices; 413374d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4134b3d85658SStefano Zampini cum = total_primal_vertices; 41359162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41364641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4137b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4138b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4139b3d85658SStefano Zampini cum++; 41409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 414174d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 414274d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 414374d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 414474d5cdf7SStefano Zampini } 41459162d606SStefano Zampini j += constraints_n[i]; 4146674ae819SStefano Zampini } 4147674ae819SStefano Zampini } 4148674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4149674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4150088faed8SStefano Zampini 4151674ae819SStefano Zampini /* set values in constraint matrix */ 4152984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 41530e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4154674ae819SStefano Zampini } 4155984c4197SStefano Zampini total_counts = total_primal_vertices; 41569162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41574641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 41589162d606SStefano Zampini PetscInt *cols; 41599162d606SStefano Zampini 41609162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 41619162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 41629162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41639162d606SStefano Zampini PetscInt row = total_counts+k; 41649162d606SStefano Zampini PetscScalar *vals; 41659162d606SStefano Zampini 41669162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 41679162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 41689162d606SStefano Zampini } 41699162d606SStefano Zampini total_counts += constraints_n[i]; 4170674ae819SStefano Zampini } 4171674ae819SStefano Zampini } 4172674ae819SStefano Zampini /* assembling */ 4173674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4174674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4175088faed8SStefano Zampini 4176984c4197SStefano Zampini /* 41776a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4178984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4179f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4180984c4197SStefano Zampini */ 4181674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4182674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4183026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4184984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4185984c4197SStefano Zampini /* working stuff for GEQRF */ 418681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4187984c4197SStefano Zampini PetscBLASInt lqr_work; 4188984c4197SStefano Zampini /* working stuff for UNGQR */ 4189984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4190984c4197SStefano Zampini PetscBLASInt lgqr_work; 4191984c4197SStefano Zampini /* working stuff for TRTRS */ 4192984c4197SStefano Zampini PetscScalar *trs_rhs; 41933f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4194984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4195984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4196984c4197SStefano Zampini PetscScalar *start_vals; 4197984c4197SStefano Zampini /* working stuff for values insertion */ 41984641a718SStefano Zampini PetscBT is_primal; 419964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4200906d46d4SStefano Zampini /* matrix sizes */ 4201906d46d4SStefano Zampini PetscInt global_size,local_size; 4202906d46d4SStefano Zampini /* temporary change of basis */ 4203906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4204cf5a6209SStefano Zampini /* extra space for debugging */ 4205cf5a6209SStefano Zampini PetscScalar *dbg_work; 4206984c4197SStefano Zampini 4207906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4208906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 420916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4210bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4211906d46d4SStefano Zampini /* nonzeros for local mat */ 4212bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 42131dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4214bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 42151dd7afcfSStefano Zampini } else { 42161dd7afcfSStefano Zampini const PetscInt *ii; 42171dd7afcfSStefano Zampini PetscInt n; 42181dd7afcfSStefano Zampini PetscBool flg_row; 42191dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42201dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 42211dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42221dd7afcfSStefano Zampini } 42239162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4224a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 42259162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4226a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 42279162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4228a717540cSStefano Zampini } else { 42299162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 42309162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4231a717540cSStefano Zampini } 4232a717540cSStefano Zampini } 4233a717540cSStefano Zampini } 4234906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4235bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 42361dd7afcfSStefano Zampini /* Set interior change in the matrix */ 42371dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4238bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4239906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4240a717540cSStefano Zampini } 42411dd7afcfSStefano Zampini } else { 42421dd7afcfSStefano Zampini const PetscInt *ii,*jj; 42431dd7afcfSStefano Zampini PetscScalar *aa; 42441dd7afcfSStefano Zampini PetscInt n; 42451dd7afcfSStefano Zampini PetscBool flg_row; 42461dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42471dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42481dd7afcfSStefano Zampini for (i=0;i<n;i++) { 42491dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 42501dd7afcfSStefano Zampini } 42511dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42521dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42531dd7afcfSStefano Zampini } 4254a717540cSStefano Zampini 4255a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4256a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4257a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4258a717540cSStefano Zampini } 4259a717540cSStefano Zampini 4260a717540cSStefano Zampini 4261a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4262a717540cSStefano Zampini /* 4263a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4264a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4265a717540cSStefano Zampini 4266a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4267a717540cSStefano Zampini 4268a6b551f4SStefano 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) 4269a6b551f4SStefano Zampini 4270a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4271a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4272a717540cSStefano Zampini | ... | 4273a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4274a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4275a717540cSStefano Zampini 4276a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4277a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4278a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4279a6b551f4SStefano Zampini 4280a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4281a717540cSStefano Zampini */ 4282a717540cSStefano Zampini if (qr_needed) { 4283984c4197SStefano Zampini /* space to store Q */ 4284854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4285984c4197SStefano Zampini /* first we issue queries for optimal work */ 42863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 42873f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 42883f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4289984c4197SStefano Zampini lqr_work = -1; 42903f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4291984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4292984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4293785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4294984c4197SStefano Zampini lgqr_work = -1; 42953f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 42963f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 42973f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 42983f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 42993f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 43003f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4301984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4302984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4303785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4304984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4305785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4306984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4307785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4308a717540cSStefano Zampini /* allocating workspace for check */ 4309a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4310cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4311a717540cSStefano Zampini } 4312a717540cSStefano Zampini } 4313984c4197SStefano Zampini /* array to store whether a node is primal or not */ 43144641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4315473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 43160e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 43176c4ed002SBarry 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); 431839e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 431939e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 432039e2fb2aSStefano Zampini } 432139e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4322984c4197SStefano Zampini 4323a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 43249162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 43259162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 43264641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4327984c4197SStefano Zampini /* get constraint info */ 43289162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4329984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4330984c4197SStefano Zampini 4331984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43329162d606SStefano 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); 4333674ae819SStefano Zampini } 4334984c4197SStefano Zampini 4335fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4336a717540cSStefano Zampini 4337a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4338a717540cSStefano Zampini if (pcbddc->dbg_flag) { 43399162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4340a717540cSStefano Zampini } 4341984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 43429162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4343984c4197SStefano Zampini 4344984c4197SStefano Zampini /* compute QR decomposition of constraints */ 43453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43473f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4348674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43493f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4350984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4351674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4352984c4197SStefano Zampini 4353984c4197SStefano Zampini /* explictly compute R^-T */ 4354984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4355984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 43563f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43573f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 43583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4360984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43613f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4362984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4363984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4364984c4197SStefano Zampini 4365a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 43663f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 43683f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 43693f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4370984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43713f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4372984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4373984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4374984c4197SStefano Zampini 4375984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4376984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4377984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 43783f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43793f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43803f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 43813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43823f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 43833f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4384984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43859162d606SStefano 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)); 4386984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 43879162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4388984c4197SStefano Zampini 4389984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 43909162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4391984c4197SStefano Zampini /* insert cols for primal dofs */ 4392984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4393984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 43949162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4395906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4396984c4197SStefano Zampini } 4397984c4197SStefano Zampini /* insert cols for dual dofs */ 4398984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 43999162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4400984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 44019162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4402906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4403984c4197SStefano Zampini j++; 4404674ae819SStefano Zampini } 4405674ae819SStefano Zampini } 4406984c4197SStefano Zampini 4407984c4197SStefano Zampini /* check change of basis */ 4408984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4409984c4197SStefano Zampini PetscInt ii,jj; 4410984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4411c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4412c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4413c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4414c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4415c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4416c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4417984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4418cf5a6209SStefano 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)); 4419984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4420984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4421984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4422cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4423cf5a6209SStefano 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; 4424674ae819SStefano Zampini } 4425674ae819SStefano Zampini } 4426984c4197SStefano Zampini if (!valid_qr) { 442722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4428984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4429984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4430cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4431cf5a6209SStefano 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])); 4432674ae819SStefano Zampini } 4433cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4434cf5a6209SStefano 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])); 4435984c4197SStefano Zampini } 4436984c4197SStefano Zampini } 4437984c4197SStefano Zampini } 4438674ae819SStefano Zampini } else { 443922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4440674ae819SStefano Zampini } 4441674ae819SStefano Zampini } 4442a717540cSStefano Zampini } else { /* simple transformation block */ 4443a717540cSStefano Zampini PetscInt row,col; 4444a6b551f4SStefano Zampini PetscScalar val,norm; 4445a6b551f4SStefano Zampini 4446a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44479162d606SStefano 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)); 4448a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 44499162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 44509162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4451bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 44529162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4453906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 44549162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4455a717540cSStefano Zampini } else { 4456a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 44579162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4458a717540cSStefano Zampini if (row != col) { 44599162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4460a717540cSStefano Zampini } else { 44619162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4462a717540cSStefano Zampini } 4463906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4464a717540cSStefano Zampini } 4465a717540cSStefano Zampini } 4466a717540cSStefano Zampini } 446798a51de6SStefano Zampini if (pcbddc->dbg_flag) { 446822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4469a717540cSStefano Zampini } 4470674ae819SStefano Zampini } 4471984c4197SStefano Zampini } else { 4472984c4197SStefano Zampini if (pcbddc->dbg_flag) { 44739162d606SStefano 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); 4474674ae819SStefano Zampini } 4475674ae819SStefano Zampini } 4476674ae819SStefano Zampini } 4477a717540cSStefano Zampini 4478a717540cSStefano Zampini /* free workspace */ 4479a717540cSStefano Zampini if (qr_needed) { 4480984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4481cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4482984c4197SStefano Zampini } 4483984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4484984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4485984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4486984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4487984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4488674ae819SStefano Zampini } 4489a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4490906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4491906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4492906d46d4SStefano Zampini 4493906d46d4SStefano Zampini /* assembling of global change of variable */ 449488c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4495bbb9e6c6SStefano Zampini Mat tmat; 449616f15bc4SStefano Zampini PetscInt bs; 449716f15bc4SStefano Zampini 4498906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4499906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4500bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4501bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4502bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4503bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 450416f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 450516f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4506906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4507bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4508bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4509bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4510bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4511bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4512e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4513e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4514bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4515bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 451688c03ad3SStefano Zampini 4517906d46d4SStefano Zampini /* check */ 4518906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4519906d46d4SStefano Zampini PetscReal error; 4520906d46d4SStefano Zampini Vec x,x_change; 4521906d46d4SStefano Zampini 4522906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4523906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4524906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4525906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4526e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4527e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4528bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4529e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4530e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4531906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4532906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4533906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4534906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4535bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4536906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4537906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4538906d46d4SStefano Zampini } 4539b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4540b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4541b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4542bf3a8328SStefano Zampini 4543bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4544bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4545bf3a8328SStefano Zampini } 4546b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 4547ac632422SStefano Zampini Mat S_new,tmat; 4548bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4549bbb9e6c6SStefano Zampini 4550bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 45516816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4552bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4553bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4554bf3a8328SStefano Zampini IS is_V; 4555b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4556b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4557b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4558b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4559b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4560bf3a8328SStefano Zampini } 4561bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4562ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4563b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4564ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4565bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4566bf3a8328SStefano Zampini const PetscScalar *array; 4567bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4568bf3a8328SStefano Zampini PetscInt i,n_V; 4569bf3a8328SStefano Zampini 4570b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4571b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4572b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4573b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4574b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4575b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4576b087196eSStefano Zampini PetscScalar val; 4577b087196eSStefano Zampini PetscInt idx; 4578b087196eSStefano Zampini 4579b087196eSStefano Zampini idx = idxs_V[i]; 4580b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4581b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4582b087196eSStefano Zampini } 4583b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4584b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4585bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4586bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4587bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4588bf3a8328SStefano Zampini } 4589ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4590ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4591ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4592ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4593b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4594ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4595bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4596b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4597bf3a8328SStefano Zampini } 4598ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4599ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4600ac632422SStefano Zampini } 4601b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 460288c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4603b96c3477SStefano Zampini } 4604c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4605b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 4606c9db6a07SStefano Zampini PetscInt i; 4607c9db6a07SStefano Zampini 4608c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4609c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4610c9db6a07SStefano Zampini } 4611c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4612c9db6a07SStefano Zampini } 4613b96c3477SStefano Zampini } 461416909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 461516909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 461616909a7fSStefano Zampini } else { 4617906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 461816909a7fSStefano Zampini } 46191dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 462027b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 462172b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 462272b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 462372b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 462472b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 462572b8c272SStefano Zampini } 46261dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 462727b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4628b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4629b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4630906d46d4SStefano Zampini } else { 46311dd7afcfSStefano Zampini Mat benign_global = NULL; 463227b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 46331dd7afcfSStefano Zampini Mat tmat; 46341dd7afcfSStefano Zampini 46351dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 46361dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 46371dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 46381dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46391dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46401dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 46411dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46421dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46431dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 46441dd7afcfSStefano Zampini if (pcbddc->benign_change) { 46451dd7afcfSStefano Zampini Mat M; 46461dd7afcfSStefano Zampini 46471dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 46481dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 46491dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 46501dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4651906d46d4SStefano Zampini } else { 46521dd7afcfSStefano Zampini Mat eye; 46531dd7afcfSStefano Zampini PetscScalar *array; 46541dd7afcfSStefano Zampini 46551dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46561dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 46571dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 46581dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4659906d46d4SStefano Zampini } 46601dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46611dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46621dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46631dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 46641dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 46651dd7afcfSStefano Zampini } 46661dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 46671dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 46681dd7afcfSStefano Zampini } 46691dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 46701dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 46711dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 467227b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 46731dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 46741dd7afcfSStefano Zampini } 46751dd7afcfSStefano Zampini } 467616909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 467716909a7fSStefano Zampini IS is_global; 467816909a7fSStefano Zampini const PetscInt *gidxs; 467916909a7fSStefano Zampini 468016909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 468116909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 468216909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 468316909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 468416909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 468516909a7fSStefano Zampini } 46861dd7afcfSStefano Zampini } 46871dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 46881dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4689b9b85e73SStefano Zampini } 4690a717540cSStefano Zampini 469172b8c272SStefano Zampini if (!pcbddc->fake_change) { 46924f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 46934f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 46944f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 46954f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4696019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4697019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4698019a44ceSStefano Zampini pcbddc->local_primal_size++; 4699019a44ceSStefano Zampini } 4700019a44ceSStefano Zampini 4701019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4702727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4703727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 47049f47a83aSStefano 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); 4705c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 47060e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 47079f47a83aSStefano 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); 4708727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4709727cdba6SStefano Zampini } 47100e6343abSStefano Zampini } 4711727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4712b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 471372b8c272SStefano Zampini } 471472b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4715727cdba6SStefano Zampini 4716a717540cSStefano Zampini /* flush dbg viewer */ 4717b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4718b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4719b8ffe317SStefano Zampini } 4720a717540cSStefano Zampini 4721e310c8b4SStefano Zampini /* free workspace */ 4722a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 47234641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 472408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 47259162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 47269162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 472708122e43SStefano Zampini } else { 47289162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 47299162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 47309162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 473108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 473208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 47339162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 47349162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 473508122e43SStefano Zampini } 4736674ae819SStefano Zampini PetscFunctionReturn(0); 4737674ae819SStefano Zampini } 4738674ae819SStefano Zampini 4739674ae819SStefano Zampini #undef __FUNCT__ 4740674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4741674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4742674ae819SStefano Zampini { 4743674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4744674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4745674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 474614f95afaSStefano Zampini PetscInt ierr,i,N; 4747674ae819SStefano Zampini 4748674ae819SStefano Zampini PetscFunctionBegin; 47498e61c736SStefano Zampini /* Reset previously computed graph */ 47508e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4751674ae819SStefano Zampini /* Init local Graph struct */ 47527fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 47533bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4754674ae819SStefano Zampini 4755575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 47565099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 47575099eff2SStefano 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); 4758575ad6abSStefano Zampini } 47599577ea80SStefano Zampini 4760674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4761d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 47624d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 47634d379d7bSStefano Zampini PetscInt nvtxs; 4764e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4765674ae819SStefano Zampini 47662fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 47672fffb893SStefano Zampini if (flg_row) { 47684d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4769b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 47702fffb893SStefano Zampini } 47712fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4772674ae819SStefano Zampini } 47739b28b941SStefano Zampini if (pcbddc->dbg_flag) { 47749b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4775674ae819SStefano Zampini } 4776674ae819SStefano Zampini 4777674ae819SStefano Zampini /* Setup of Graph */ 4778*4b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 477914f95afaSStefano 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); 4780674ae819SStefano Zampini 47814f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 47824f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 47834f1b2e48SStefano Zampini PetscInt *local_subs; 47844f1b2e48SStefano Zampini 47854f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 47864f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 47874f1b2e48SStefano Zampini const PetscInt *idxs; 47884f1b2e48SStefano Zampini PetscInt nl,j; 47894f1b2e48SStefano Zampini 47904f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 47914f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 47924f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 47934f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 47944f1b2e48SStefano Zampini } 47954f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 47964f1b2e48SStefano Zampini } 47974f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 47984f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 47994f1b2e48SStefano Zampini } 48004f1b2e48SStefano Zampini 4801674ae819SStefano Zampini /* Graph's connected components analysis */ 4802674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4803674ae819SStefano Zampini PetscFunctionReturn(0); 4804674ae819SStefano Zampini } 4805674ae819SStefano Zampini 4806dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4807674ae819SStefano Zampini #undef __FUNCT__ 4808674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4809dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4810674ae819SStefano Zampini { 4811dc456d91SStefano Zampini PetscSF sf; 4812dc456d91SStefano Zampini PetscLayout map; 4813dc456d91SStefano Zampini const PetscInt *idxs; 4814dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4815dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4816dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4817dc456d91SStefano Zampini PetscMPIInt commsize; 4818674ae819SStefano Zampini PetscBool first_found; 4819674ae819SStefano Zampini PetscErrorCode ierr; 4820674ae819SStefano Zampini 4821674ae819SStefano Zampini PetscFunctionBegin; 4822dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4823dc456d91SStefano Zampini if (subset_mult) { 4824dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4825dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4826dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4827674ae819SStefano Zampini } 4828dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4829dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4830dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4831dc456d91SStefano Zampini for (i=0;i<n;i++) { 4832dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4833dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4834674ae819SStefano Zampini } 4835dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4836b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4837dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4838dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4839dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4840dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4841dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4842dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4843dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4844dc456d91SStefano Zampini 4845dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4846dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4847dc456d91SStefano Zampini if (subset_mult) { 4848dc456d91SStefano Zampini const PetscInt* idxs_mult; 4849dc456d91SStefano Zampini 4850dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4851dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4852dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4853674ae819SStefano Zampini } else { 4854dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4855674ae819SStefano Zampini } 4856dc456d91SStefano Zampini /* local size of new subset */ 4857dc456d91SStefano Zampini n_n = 0; 4858dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4859dc456d91SStefano Zampini 4860dc456d91SStefano Zampini /* global indexes in layout */ 4861dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4862dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4863dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4864dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4865dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4866dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4867dc456d91SStefano Zampini 4868dc456d91SStefano Zampini /* reduce from leaves to roots */ 4869dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 487064a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 487164a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4872dc456d91SStefano Zampini 4873dc456d91SStefano Zampini /* count indexes in local part of layout */ 4874674ae819SStefano Zampini nlocals = 0; 4875674ae819SStefano Zampini first_index = -1; 4876674ae819SStefano Zampini first_found = PETSC_FALSE; 4877dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4878dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4879674ae819SStefano Zampini first_found = PETSC_TRUE; 4880674ae819SStefano Zampini first_index = i; 4881674ae819SStefano Zampini } 4882dc456d91SStefano Zampini nlocals += root_data[i]; 4883674ae819SStefano Zampini } 4884dc456d91SStefano Zampini 4885dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 48865fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4887dc456d91SStefano Zampini start = 0; 488864a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 48895fa240b1SStefano Zampini #else 489064a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 48915fa240b1SStefano Zampini start = start-nlocals; 48925fa240b1SStefano Zampini #endif 48935fa240b1SStefano Zampini 4894dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4895dc456d91SStefano Zampini *N_n = start + nlocals; 4896dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4897dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4898674ae819SStefano Zampini } 48995fa240b1SStefano Zampini 49005fa240b1SStefano Zampini /* adapt root data with cumulative */ 4901674ae819SStefano Zampini if (first_found) { 4902dc456d91SStefano Zampini PetscInt old_index; 4903dc456d91SStefano Zampini 4904dc456d91SStefano Zampini root_data[first_index] += start; 4905674ae819SStefano Zampini old_index = first_index; 4906dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4907dc456d91SStefano Zampini if (root_data[i]) { 4908dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4909674ae819SStefano Zampini old_index = i; 4910674ae819SStefano Zampini } 4911674ae819SStefano Zampini } 4912674ae819SStefano Zampini } 4913dc456d91SStefano Zampini 4914dc456d91SStefano Zampini /* from roots to leaves */ 4915dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4916dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4917dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4918dc456d91SStefano Zampini 4919dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4920dc456d91SStefano Zampini if (subset_mult) { 4921dc456d91SStefano Zampini const PetscInt* idxs_mult; 4922dc456d91SStefano Zampini PetscInt cum; 4923dc456d91SStefano Zampini 4924dc456d91SStefano Zampini cum = 0; 4925dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4926dc456d91SStefano Zampini for (i=0;i<n;i++) { 4927dc456d91SStefano Zampini PetscInt j; 4928dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4929674ae819SStefano Zampini } 4930dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4931674ae819SStefano Zampini } else { 4932dc456d91SStefano Zampini for (i=0;i<n;i++) { 4933dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4934674ae819SStefano Zampini } 4935674ae819SStefano Zampini } 4936dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4937dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4938674ae819SStefano Zampini PetscFunctionReturn(0); 4939674ae819SStefano Zampini } 49409a7d3425SStefano Zampini 4941669cc0f4SStefano Zampini /* this implements stabilized Gram-Schmidt */ 49429a7d3425SStefano Zampini #undef __FUNCT__ 49439a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 49449a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 49459a7d3425SStefano Zampini { 49469a7d3425SStefano Zampini PetscInt i,j; 49479a7d3425SStefano Zampini PetscScalar *alphas; 49489a7d3425SStefano Zampini PetscErrorCode ierr; 49499a7d3425SStefano Zampini 49509a7d3425SStefano Zampini PetscFunctionBegin; 4951785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 49529a7d3425SStefano Zampini for (i=0;i<n;i++) { 49539a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 4954669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 4955669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 4956669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 49579a7d3425SStefano Zampini } 49589a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 49599a7d3425SStefano Zampini PetscFunctionReturn(0); 49609a7d3425SStefano Zampini } 49619a7d3425SStefano Zampini 4962e7931f94SStefano Zampini #undef __FUNCT__ 496370cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 496457de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4965e7931f94SStefano Zampini { 496657de7509SStefano Zampini Mat A; 4967e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4968e7931f94SStefano Zampini PetscMPIInt size,rank,color; 496952e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 497052e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 497127b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 497257de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 497327b6a85dSStefano Zampini PetscInt xadj_count, *count; 497427b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 497527b6a85dSStefano Zampini PetscSubcomm psubcomm; 497627b6a85dSStefano Zampini MPI_Comm subcomm; 497752e5ac9dSStefano Zampini PetscErrorCode ierr; 4978a57a6d2fSStefano Zampini 4979e7931f94SStefano Zampini PetscFunctionBegin; 498057de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 498157de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 498257de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 498357de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 498457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 498557de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 498657de7509SStefano Zampini 498757de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 498857de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 498957de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 499057de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 499157de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 499257de7509SStefano Zampini im_active = !!(n); 499357de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 499457de7509SStefano Zampini void_procs = size - active_procs; 499557de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 499657de7509SStefano Zampini if (void_procs) { 499757de7509SStefano Zampini PetscInt ncand; 499857de7509SStefano Zampini 499957de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 500057de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 500157de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 500257de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 500357de7509SStefano Zampini if (!procs_candidates[i]) { 500457de7509SStefano Zampini procs_candidates[ncand++] = i; 500557de7509SStefano Zampini } 500657de7509SStefano Zampini } 500757de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 500857de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 500957de7509SStefano Zampini } 501057de7509SStefano Zampini 501157de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 501257de7509SStefano Zampini if (active_procs < *n_subdomains) { 501357de7509SStefano Zampini PetscInt issize,isidx; 501457de7509SStefano Zampini if (im_active) { 501557de7509SStefano Zampini issize = 1; 501657de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 501757de7509SStefano Zampini isidx = procs_candidates[rank]; 501857de7509SStefano Zampini } else { 501957de7509SStefano Zampini isidx = rank; 502057de7509SStefano Zampini } 502157de7509SStefano Zampini } else { 502257de7509SStefano Zampini issize = 0; 502357de7509SStefano Zampini isidx = -1; 502457de7509SStefano Zampini } 502557de7509SStefano Zampini *n_subdomains = active_procs; 502657de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 5027daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 502857de7509SStefano Zampini PetscFunctionReturn(0); 502957de7509SStefano Zampini } 5030c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 5031c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 503227b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 5033e7931f94SStefano Zampini 5034e7931f94SStefano Zampini /* Get info on mapping */ 50353bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 50363bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 5037e7931f94SStefano Zampini 5038e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 5039785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 5040e7931f94SStefano Zampini xadj[0] = 0; 5041e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 5042785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 5043785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 504427b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 504527b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 504627b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 504727b6a85dSStefano Zampini count[shared[i][j]] += 1; 5048e7931f94SStefano Zampini 504927b6a85dSStefano Zampini xadj_count = 0; 50502b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 505127b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 505227b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 5053d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 5054d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 5055d023bfaeSStefano Zampini xadj_count++; 505627b6a85dSStefano Zampini break; 505727b6a85dSStefano Zampini } 5058e7931f94SStefano Zampini } 5059e7931f94SStefano Zampini } 5060d023bfaeSStefano Zampini xadj[1] = xadj_count; 506127b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 50623bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 5063e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 5064e7931f94SStefano Zampini 50653837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 5066e7931f94SStefano Zampini 506727b6a85dSStefano Zampini /* Restrict work on active processes only */ 506827b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 506927b6a85dSStefano Zampini if (void_procs) { 507027b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 507127b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 507227b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 507327b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 507427b6a85dSStefano Zampini } else { 507527b6a85dSStefano Zampini psubcomm = NULL; 507627b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 507727b6a85dSStefano Zampini } 507827b6a85dSStefano Zampini 507927b6a85dSStefano Zampini v_wgt = NULL; 508027b6a85dSStefano Zampini if (!color) { 5081e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 5082e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 5083e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5084c8587f34SStefano Zampini } else { 508552e5ac9dSStefano Zampini Mat subdomain_adj; 508652e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 508752e5ac9dSStefano Zampini MatPartitioning partitioner; 508827b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 508952e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 509057de7509SStefano Zampini PetscMPIInt size; 5091b0c7d250SStefano Zampini PetscBool aggregate; 5092b0c7d250SStefano Zampini 509327b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 509427b6a85dSStefano Zampini if (void_procs) { 509527b6a85dSStefano Zampini PetscInt prank = rank; 5096785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 509727b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 5098e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 5099e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 5100c8587f34SStefano Zampini } 5101e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 510227b6a85dSStefano Zampini } else { 510327b6a85dSStefano Zampini oldranks = NULL; 510427b6a85dSStefano Zampini } 5105b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 510627b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 5107b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 5108b0c7d250SStefano Zampini PetscMPIInt nrank; 5109b0c7d250SStefano Zampini PetscScalar *vals; 5110b0c7d250SStefano Zampini 511127b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5112b0c7d250SStefano Zampini lrows = 0; 5113b0c7d250SStefano Zampini if (nrank<redprocs) { 5114b0c7d250SStefano Zampini lrows = size/redprocs; 5115b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5116b0c7d250SStefano Zampini } 511727b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5118b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5119b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5120b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5121b0c7d250SStefano Zampini row = nrank; 5122b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5123b0c7d250SStefano Zampini cols = adjncy; 5124b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5125b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5126b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5127b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5128b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 512952e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 513052e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 513152e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5132b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 513327b6a85dSStefano Zampini if (use_vwgt) { 513427b6a85dSStefano Zampini Vec v; 513527b6a85dSStefano Zampini const PetscScalar *array; 513627b6a85dSStefano Zampini PetscInt nl; 513727b6a85dSStefano Zampini 513827b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 513927b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 514027b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 514127b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 514227b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 514327b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 514427b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 514522db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 514627b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 514727b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 514827b6a85dSStefano Zampini } 5149b0c7d250SStefano Zampini } else { 515027b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 515127b6a85dSStefano Zampini if (use_vwgt) { 515227b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 515327b6a85dSStefano Zampini v_wgt[0] = local_size; 515427b6a85dSStefano Zampini } 5155b0c7d250SStefano Zampini } 515622b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5157e7931f94SStefano Zampini 5158e7931f94SStefano Zampini /* Partition */ 515927b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5160e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 516127b6a85dSStefano Zampini if (v_wgt) { 5162e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5163c8587f34SStefano Zampini } 516457de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 516557de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5166e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5167e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 516822b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5169e7931f94SStefano Zampini 517052e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 517152e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 517252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 517352e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 517457de7509SStefano Zampini if (!aggregate) { 517557de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 517627b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 517727b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 517827b6a85dSStefano Zampini #endif 517957de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 518027b6a85dSStefano Zampini } else if (oldranks) { 5181b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 518227b6a85dSStefano Zampini } else { 518327b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 518457de7509SStefano Zampini } 518528143c3dSStefano Zampini } else { 5186b0c7d250SStefano Zampini PetscInt idxs[1]; 5187b0c7d250SStefano Zampini PetscMPIInt tag; 5188b0c7d250SStefano Zampini MPI_Request *reqs; 5189b0c7d250SStefano Zampini 5190b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5191b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5192b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 519327b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 519428143c3dSStefano Zampini } 519527b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5196b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5197b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 519857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 519927b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 520027b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 520127b6a85dSStefano Zampini #endif 520257de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 520327b6a85dSStefano Zampini } else if (oldranks) { 5204b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 520527b6a85dSStefano Zampini } else { 520627b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5207e7931f94SStefano Zampini } 520857de7509SStefano Zampini } 520952e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5210e7931f94SStefano Zampini /* clean up */ 5211e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 521252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5213e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5214e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5215e7931f94SStefano Zampini } 521627b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 521757de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5218e7931f94SStefano Zampini 5219e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5220e7931f94SStefano Zampini i = 1; 522127b6a85dSStefano Zampini if (!color) i=0; 522257de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5223e7931f94SStefano Zampini PetscFunctionReturn(0); 5224e7931f94SStefano Zampini } 5225e7931f94SStefano Zampini 5226e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5227e7931f94SStefano Zampini 5228e7931f94SStefano Zampini #undef __FUNCT__ 5229e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 52301ae86dd6SStefano 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[], PetscInt nvecs, Vec nnsp_vec[]) 5231e7931f94SStefano Zampini { 523270cf5478SStefano Zampini Mat local_mat; 5233e7931f94SStefano Zampini IS is_sends_internal; 52349d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 52351ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 52369d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5237e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5238e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5239e7931f94SStefano Zampini const PetscInt* is_indices; 5240e7931f94SStefano Zampini MatType new_local_type; 5241e7931f94SStefano Zampini /* buffers */ 5242e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 524328143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 52449d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5245e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 52461ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 5247e7931f94SStefano Zampini /* MPI */ 524828143c3dSStefano Zampini MPI_Comm comm,comm_n; 524928143c3dSStefano Zampini PetscSubcomm subcomm; 5250e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 525128143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 525228143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 52531ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 52541ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 52551ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 5256e7931f94SStefano Zampini PetscErrorCode ierr; 5257e7931f94SStefano Zampini 5258e7931f94SStefano Zampini PetscFunctionBegin; 525957de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5260e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 526128143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 526257de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 526357de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 526457de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 526557de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 526657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 52671ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 52681ae86dd6SStefano Zampini if (nvecs) { 52691ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 52701ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 52711ae86dd6SStefano Zampini } 527257de7509SStefano Zampini /* further checks */ 5273e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5274e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5275e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5276e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5277e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 527857de7509SStefano Zampini if (reuse && *mat_n) { 527970cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 528057de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 528170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 528228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 528370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 528470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 528570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 528670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 528770cf5478SStefano Zampini } 5288e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5289e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 529057de7509SStefano Zampini 5291e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5292e7931f94SStefano Zampini if (!is_sends) { 529328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 529457de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5295c8587f34SStefano Zampini } else { 5296e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5297e7931f94SStefano Zampini is_sends_internal = is_sends; 5298c8587f34SStefano Zampini } 5299e7931f94SStefano Zampini 5300e7931f94SStefano Zampini /* get comm */ 5301a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5302e7931f94SStefano Zampini 5303e7931f94SStefano Zampini /* compute number of sends */ 5304e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5305e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5306e7931f94SStefano Zampini 5307e7931f94SStefano Zampini /* compute number of receives */ 5308e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5309785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5310e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5311e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5312e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5313e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5314e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5315e7931f94SStefano Zampini 531628143c3dSStefano Zampini /* restrict comm if requested */ 531728143c3dSStefano Zampini subcomm = 0; 531828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 531928143c3dSStefano Zampini if (restrict_comm) { 5320779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5321779c1cceSStefano Zampini 532228143c3dSStefano Zampini color = 0; 532353a05cb3SStefano Zampini if (restrict_full) { 532453a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 532553a05cb3SStefano Zampini } else { 532653a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 532753a05cb3SStefano Zampini } 5328b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 532928143c3dSStefano Zampini subcommsize = commsize - subcommsize; 533028143c3dSStefano Zampini /* check if reuse has been requested */ 533157de7509SStefano Zampini if (reuse) { 533228143c3dSStefano Zampini if (*mat_n) { 533328143c3dSStefano Zampini PetscMPIInt subcommsize2; 533428143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 533528143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 533628143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 533728143c3dSStefano Zampini } else { 533828143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 533928143c3dSStefano Zampini } 534028143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5341779c1cceSStefano Zampini PetscMPIInt rank; 5342779c1cceSStefano Zampini 5343779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 534428143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 534528143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 534628143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5347306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 534828143c3dSStefano Zampini } 534928143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 535028143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 535128143c3dSStefano Zampini } else { 535228143c3dSStefano Zampini comm_n = comm; 535328143c3dSStefano Zampini } 535428143c3dSStefano Zampini 5355e7931f94SStefano Zampini /* prepare send/receive buffers */ 5356785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5357e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5358785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5359e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 536028143c3dSStefano Zampini if (nis) { 5361854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 536228143c3dSStefano Zampini } 5363e7931f94SStefano Zampini 536428143c3dSStefano Zampini /* Get data from local matrices */ 53656c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5366e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5367e7931f94SStefano Zampini /* 5368e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5369e7931f94SStefano Zampini send_buffer_idxs should contain: 5370e7931f94SStefano Zampini - MatType_PRIVATE type 5371e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5372e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5373e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5374e7931f94SStefano Zampini */ 53756c4ed002SBarry Smith else { 5376e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 53773bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5378854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5379e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5380e7931f94SStefano Zampini send_buffer_idxs[1] = i; 53813bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5382e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 53833bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5384e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5385e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5386e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5387e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5388c8587f34SStefano Zampini } 5389c8587f34SStefano Zampini } 5390e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 539128143c3dSStefano Zampini /* additional is (if any) */ 539228143c3dSStefano Zampini if (nis) { 539328143c3dSStefano Zampini PetscMPIInt psum; 539428143c3dSStefano Zampini PetscInt j; 539528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 539628143c3dSStefano Zampini PetscInt plen; 539728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 539828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 539928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 540028143c3dSStefano Zampini } 5401854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 540228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 540328143c3dSStefano Zampini PetscInt plen; 540428143c3dSStefano Zampini const PetscInt *is_array_idxs; 540528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 540628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 540728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 540828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 540928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 541028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 541128143c3dSStefano Zampini } 541228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 541328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 541428143c3dSStefano Zampini } 541528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 541628143c3dSStefano Zampini } 541728143c3dSStefano Zampini 5418e7931f94SStefano Zampini buf_size_idxs = 0; 5419e7931f94SStefano Zampini buf_size_vals = 0; 542028143c3dSStefano Zampini buf_size_idxs_is = 0; 54211ae86dd6SStefano Zampini buf_size_vecs = 0; 5422e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5423e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5424e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 542528143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 54261ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 5427e7931f94SStefano Zampini } 5428785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5429785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 543095ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 54311ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 5432e7931f94SStefano Zampini 5433e7931f94SStefano Zampini /* get new tags for clean communications */ 5434e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5435e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 543628143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 54371ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 5438e7931f94SStefano Zampini 5439e7931f94SStefano Zampini /* allocate for requests */ 5440785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5441785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 544295ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 54431ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 5444785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5445785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 544695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 54471ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 5448e7931f94SStefano Zampini 5449e7931f94SStefano Zampini /* communications */ 5450e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5451e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 545228143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 54531ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 5454e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5455e7931f94SStefano Zampini source_dest = onodes[i]; 5456e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5457e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5458e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5459e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 546028143c3dSStefano Zampini if (nis) { 546157de7509SStefano Zampini source_dest = onodes_is[i]; 546228143c3dSStefano 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); 546328143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 546428143c3dSStefano Zampini } 54651ae86dd6SStefano Zampini if (nvecs) { 54661ae86dd6SStefano Zampini source_dest = onodes[i]; 54671ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 54681ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 54691ae86dd6SStefano Zampini } 5470e7931f94SStefano Zampini } 5471e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5472e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5473e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5474e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 547528143c3dSStefano Zampini if (nis) { 547628143c3dSStefano 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); 547728143c3dSStefano Zampini } 54781ae86dd6SStefano Zampini if (nvecs) { 54791ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 54801ae86dd6SStefano Zampini ierr = MPI_Isend(send_buffer_vecs,ilengths_idxs[source_dest]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&send_req_vecs[i]);CHKERRQ(ierr); 54811ae86dd6SStefano Zampini } 5482e7931f94SStefano Zampini } 5483e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5484e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5485e7931f94SStefano Zampini 5486e7931f94SStefano Zampini /* assemble new l2g map */ 5487e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5488e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 54899d30be91SStefano Zampini new_local_rows = 0; 5490e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 54919d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5492e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5493e7931f94SStefano Zampini } 54949d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5495e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 54969d30be91SStefano Zampini new_local_rows = 0; 5497e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 54989d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 54999d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5500e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5501e7931f94SStefano Zampini } 55029d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 55039d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5504e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5505e7931f94SStefano Zampini 5506e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5507e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5508e7931f94SStefano 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) */ 5509e7931f94SStefano Zampini if (n_recvs) { 551028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5511e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5512e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5513e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5514e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5515e7931f94SStefano Zampini break; 5516e7931f94SStefano Zampini } 5517e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5518e7931f94SStefano Zampini } 5519e7931f94SStefano Zampini switch (new_local_type_private) { 552028143c3dSStefano Zampini case MATDENSE_PRIVATE: 552128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5522e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5523e7931f94SStefano Zampini bs = 1; 552428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 552528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 552628143c3dSStefano Zampini bs = 1; 552728143c3dSStefano Zampini } 5528e7931f94SStefano Zampini break; 5529e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5530e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5531e7931f94SStefano Zampini bs = 1; 5532e7931f94SStefano Zampini break; 5533e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5534e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5535e7931f94SStefano Zampini break; 5536e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5537e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5538e7931f94SStefano Zampini break; 5539e7931f94SStefano Zampini default: 55409d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5541e7931f94SStefano Zampini break; 5542e7931f94SStefano Zampini } 554328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 554428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 554528143c3dSStefano Zampini bs = 1; 5546e7931f94SStefano Zampini } 5547e7931f94SStefano Zampini 554870cf5478SStefano Zampini /* create MATIS object if needed */ 554957de7509SStefano Zampini if (!reuse) { 5550e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5551e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 555270cf5478SStefano Zampini } else { 555370cf5478SStefano Zampini /* it also destroys the local matrices */ 555457de7509SStefano Zampini if (*mat_n) { 555570cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 555657de7509SStefano Zampini } else { /* this is a fake object */ 555757de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 555857de7509SStefano Zampini } 555970cf5478SStefano Zampini } 556070cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5561e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 55629d30be91SStefano Zampini 55639d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 55649d30be91SStefano Zampini 55659d30be91SStefano Zampini /* Global to local map of received indices */ 55669d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 55679d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 55689d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 55699d30be91SStefano Zampini 55709d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 55719d30be91SStefano Zampini buf_size_idxs = 0; 55729d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 55739d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 55749d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 55759d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 55769d30be91SStefano Zampini } 55779d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 55789d30be91SStefano Zampini 55799d30be91SStefano Zampini /* set preallocation */ 55809d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 55819d30be91SStefano Zampini if (!newisdense) { 55829d30be91SStefano Zampini PetscInt *new_local_nnz=0; 55839d30be91SStefano Zampini 55849d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 55859d30be91SStefano Zampini if (n_recvs) { 55869d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 55879d30be91SStefano Zampini } 55889d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 55899d30be91SStefano Zampini PetscInt j; 55909d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 55919d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 55929d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 55939d30be91SStefano Zampini } 55949d30be91SStefano Zampini } else { 55959d30be91SStefano Zampini /* TODO */ 55969d30be91SStefano Zampini } 55979d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 55989d30be91SStefano Zampini } 55999d30be91SStefano Zampini if (new_local_nnz) { 56009d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 56019d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 56029d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 56039d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 56049d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 56059d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 56069d30be91SStefano Zampini } else { 56079d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 56089d30be91SStefano Zampini } 56099d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 56109d30be91SStefano Zampini } else { 56119d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 56129d30be91SStefano Zampini } 5613e7931f94SStefano Zampini 5614e7931f94SStefano Zampini /* set values */ 5615e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 56169d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5617e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5618e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5619e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 56209d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5621e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5622e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5623e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 562428143c3dSStefano Zampini } else { 562528143c3dSStefano Zampini /* TODO */ 5626e7931f94SStefano Zampini } 5627e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5628e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5629e7931f94SStefano Zampini } 5630e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5631e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 563270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 563370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 56349d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 5635e7931f94SStefano Zampini 5636dfd14d43SStefano Zampini #if 0 563728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5638e7931f94SStefano Zampini Vec lvec,rvec; 5639e7931f94SStefano Zampini PetscReal infty_error; 5640e7931f94SStefano Zampini 56412a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5642e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5643e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5644e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 564570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5646e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5647e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5648e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5649e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5650e7931f94SStefano Zampini } 565128143c3dSStefano Zampini #endif 5652e7931f94SStefano Zampini 565328143c3dSStefano Zampini /* assemble new additional is (if any) */ 565428143c3dSStefano Zampini if (nis) { 565528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 565628143c3dSStefano Zampini 565728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5658854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 565928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 566028143c3dSStefano Zampini psum = 0; 566128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 566228143c3dSStefano Zampini for (j=0;j<nis;j++) { 566328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 566428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 566528143c3dSStefano Zampini psum += plen; 566628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 566728143c3dSStefano Zampini } 566828143c3dSStefano Zampini } 5669854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5670854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 567128143c3dSStefano Zampini for (i=1;i<nis;i++) { 567228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 567328143c3dSStefano Zampini } 567428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 567528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 567628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 567728143c3dSStefano Zampini for (j=0;j<nis;j++) { 567828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 567928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 568028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 568128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 568228143c3dSStefano Zampini } 568328143c3dSStefano Zampini } 568428143c3dSStefano Zampini for (i=0;i<nis;i++) { 568528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 568628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 568728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 568828143c3dSStefano Zampini } 568928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 569028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 569128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 569228143c3dSStefano Zampini } 5693e7931f94SStefano Zampini /* free workspace */ 569428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5695e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5696e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5697e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5698e7931f94SStefano Zampini if (isdense) { 5699e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5700e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5701e7931f94SStefano Zampini } else { 5702e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5703e7931f94SStefano Zampini } 570428143c3dSStefano Zampini if (nis) { 570528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 570628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 570728143c3dSStefano Zampini } 57081ae86dd6SStefano Zampini 57091ae86dd6SStefano Zampini if (nvecs) { 57101ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 57111ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 57121ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 57131ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 57141ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 57151ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 57161ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 57171ae86dd6SStefano Zampini /* set values */ 57181ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 57191ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 57201ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 57211ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 57221ae86dd6SStefano Zampini PetscInt j; 57231ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 57241ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 57251ae86dd6SStefano Zampini } 57261ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 57271ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 57281ae86dd6SStefano Zampini } 57291ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 57301ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 57311ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 57321ae86dd6SStefano Zampini } 57331ae86dd6SStefano Zampini 57341ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 57351ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5736e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5737e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 57381ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 573928143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5740e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5741e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 57421ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 574328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5744e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5745e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5746e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5747e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5748e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 574928143c3dSStefano Zampini if (nis) { 575028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 575128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 575228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 575328143c3dSStefano Zampini } 575428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 575528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 575628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 575728143c3dSStefano Zampini for (i=0;i<nis;i++) { 575828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 575928143c3dSStefano Zampini } 57601ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 57611ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 57621ae86dd6SStefano Zampini } 576353a05cb3SStefano Zampini *mat_n = NULL; 576428143c3dSStefano Zampini } 5765e7931f94SStefano Zampini PetscFunctionReturn(0); 5766e7931f94SStefano Zampini } 5767a57a6d2fSStefano Zampini 576812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5769af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 577012edc857SStefano Zampini 5771c8587f34SStefano Zampini #undef __FUNCT__ 5772c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5773c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5774c8587f34SStefano Zampini { 5775c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5776c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 577720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 57781ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 57799881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 578020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 57816e683305SStefano Zampini IS coarse_is,*isarray; 57826e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 578330368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5784f9eb5b7dSStefano Zampini PC pc_temp; 5785c8587f34SStefano Zampini PCType coarse_pc_type; 5786c8587f34SStefano Zampini KSPType coarse_ksp_type; 5787f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 57884f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 57896e683305SStefano Zampini Mat t_coarse_mat_is; 579057de7509SStefano Zampini PetscInt ncoarse; 579168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 579222bc73bbSStefano Zampini PetscScalar *array; 579357de7509SStefano Zampini MatReuse coarse_mat_reuse; 579457de7509SStefano Zampini PetscBool restr, full_restr, have_void; 57959881197aSStefano Zampini PetscErrorCode ierr; 5796fdc09c96SStefano Zampini 5797c8587f34SStefano Zampini PetscFunctionBegin; 5798c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 579968457ee5SStefano 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 */ 5800fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 58015a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5802fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5803f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5804f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5805f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5806fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 580751bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 580851bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5809dc4bcba2SStefano Zampini PC pc; 5810dc4bcba2SStefano Zampini PetscBool isbddc; 5811dc4bcba2SStefano Zampini 5812dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5813dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5814dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5815dc4bcba2SStefano Zampini if (isbddc) { 581663c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 581763c961adSStefano Zampini } else { 5818727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 581963c961adSStefano Zampini } 5820fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5821fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5822fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5823f4ddd8eeSStefano Zampini } 5824fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5825fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5826f4ddd8eeSStefano Zampini } 582770cf5478SStefano Zampini /* reset any subassembling information */ 582857de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 582970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 583057de7509SStefano Zampini } 58316e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5832fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5833f4ddd8eeSStefano Zampini } 583457de7509SStefano Zampini /* assemble coarse matrix */ 583557de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 583657de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 583757de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 583857de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 583918a45a71SStefano Zampini } else { 584057de7509SStefano Zampini coarse_mat = NULL; 584157de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 58426e683305SStefano Zampini } 5843e7931f94SStefano Zampini 5844abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5845abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5846abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5847abbbba34SStefano Zampini 5848abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 584922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 585022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 585122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 585222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5853e176bc59SStefano 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); 58546e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 58556e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 58566e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5857abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5858abbbba34SStefano Zampini 585957de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 586057de7509SStefano Zampini im_active = !!(pcis->n); 586157de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 586257de7509SStefano Zampini 586357de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 586457de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 586557de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 586657de7509SStefano Zampini coarse_mat_is = NULL; 586757de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 586857de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 586957de7509SStefano Zampini full_restr = PETSC_TRUE; 58701ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 587157de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 587257de7509SStefano Zampini if (multilevel_requested) { 587357de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 587457de7509SStefano Zampini restr = PETSC_FALSE; 587557de7509SStefano Zampini full_restr = PETSC_FALSE; 587657de7509SStefano Zampini } else { 587757de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 587857de7509SStefano Zampini restr = PETSC_TRUE; 587957de7509SStefano Zampini full_restr = PETSC_TRUE; 588057de7509SStefano Zampini } 5881*4b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 588257de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 588357de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 588457de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 588557de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 588657de7509SStefano Zampini PetscInt psum; 588757de7509SStefano Zampini PetscMPIInt size; 588857de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 588957de7509SStefano Zampini else psum = 0; 589057de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 589157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 589257de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 589357de7509SStefano Zampini } 589457de7509SStefano Zampini /* determine if we can go multilevel */ 589557de7509SStefano Zampini if (multilevel_requested) { 589657de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 589757de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 589857de7509SStefano Zampini } 589957de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 590057de7509SStefano Zampini 5901e4d548c7SStefano Zampini /* dump subassembling pattern */ 5902e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5903e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5904e4d548c7SStefano Zampini } 5905e4d548c7SStefano Zampini 59066e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 590727b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 59086e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 59096e683305SStefano Zampini const PetscInt *idxs; 59106e683305SStefano Zampini ISLocalToGlobalMapping tmap; 59116e683305SStefano Zampini 59126e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 59130be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 59146e683305SStefano Zampini /* allocate space for temporary storage */ 5915854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5916854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 59176e683305SStefano Zampini /* allocate for IS array */ 59186e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 59196e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 592027b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 592130368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5922854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 59236e683305SStefano Zampini /* dofs splitting */ 59246e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 59256e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 59266e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 59276e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 59286e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 59296e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 59306e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 593130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 59326e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 59336e683305SStefano Zampini } 59346e683305SStefano Zampini /* neumann boundaries */ 59356e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 59366e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 59376e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 59386e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 59396e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 59406e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 59416e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 594230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 59436e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 59446e683305SStefano Zampini } 59456e683305SStefano Zampini /* free memory */ 59466e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 59476e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 59486e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 59496e683305SStefano Zampini } else { 59506e683305SStefano Zampini nis = 0; 59516e683305SStefano Zampini nisdofs = 0; 59526e683305SStefano Zampini nisneu = 0; 595330368db7SStefano Zampini nisvert = 0; 59546e683305SStefano Zampini isarray = NULL; 59556e683305SStefano Zampini } 59566e683305SStefano Zampini /* destroy no longer needed map */ 59576e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 59586e683305SStefano Zampini 595957de7509SStefano Zampini /* subassemble */ 596057de7509SStefano Zampini if (multilevel_allowed) { 59611ae86dd6SStefano Zampini Vec vp[1]; 59621ae86dd6SStefano Zampini PetscInt nvecs = 0; 596357de7509SStefano Zampini PetscBool reuse,reuser; 59641ae86dd6SStefano Zampini 596557de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 596657de7509SStefano Zampini else reuse = PETSC_FALSE; 596757de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 59681ae86dd6SStefano Zampini vp[0] = NULL; 59691ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 59701ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 59711ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 59721ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 59731ae86dd6SStefano Zampini nvecs = 1; 59741ae86dd6SStefano Zampini 59751ae86dd6SStefano Zampini if (pcbddc->divudotp) { 59761ae86dd6SStefano Zampini Mat B; 59771ae86dd6SStefano Zampini Vec v,p; 59781ae86dd6SStefano Zampini IS dummy; 59791ae86dd6SStefano Zampini PetscInt np; 59801ae86dd6SStefano Zampini 59811ae86dd6SStefano Zampini ierr = MatGetSize(pcbddc->divudotp,&np,NULL);CHKERRQ(ierr); 59821ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 59831ae86dd6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 59841ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 59851ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 59861ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 59871ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 59881ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 59891ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 59901ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 59911ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 59921ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 59931ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 59941ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 59951ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 599674e2c79eSStefano Zampini } 59971ae86dd6SStefano Zampini } 59981ae86dd6SStefano Zampini if (reuser) { 59991ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 600074e2c79eSStefano Zampini } else { 60011ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray,nvecs,vp);CHKERRQ(ierr); 60021ae86dd6SStefano Zampini } 60031ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 60041ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 60051ae86dd6SStefano Zampini PetscInt nl; 60061ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 60071ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 60081ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 60091ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 60101ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 60111ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 60121ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 60131ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 60141ae86dd6SStefano Zampini } 60151ae86dd6SStefano Zampini } else { 60161ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray,0,NULL);CHKERRQ(ierr); 60176e683305SStefano Zampini } 601857de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 601957de7509SStefano Zampini PetscMPIInt size; 602057de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 602157de7509SStefano Zampini if (!multilevel_allowed) { 602257de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 60236e683305SStefano Zampini } else { 602457de7509SStefano Zampini Mat A; 6025779c1cceSStefano Zampini 602657de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 602757de7509SStefano Zampini if (coarse_mat_is) { 602857de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 602957de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 603057de7509SStefano Zampini coarse_mat = coarse_mat_is; 603157de7509SStefano Zampini } 603257de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 603357de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 603457de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 6035779c1cceSStefano Zampini } 6036779c1cceSStefano Zampini } 603757de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 603857de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 60396e683305SStefano Zampini 60406e683305SStefano Zampini /* create local to global scatters for coarse problem */ 604168457ee5SStefano Zampini if (compute_vecs) { 60426e683305SStefano Zampini PetscInt lrows; 60436e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 604457de7509SStefano Zampini if (coarse_mat) { 604557de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 60466e683305SStefano Zampini } else { 60476e683305SStefano Zampini lrows = 0; 60486e683305SStefano Zampini } 60496e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 60506e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 60516e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 60526e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 60536e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 60546e683305SStefano Zampini } 60556e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 6056c8587f34SStefano Zampini 6057f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 6058f9eb5b7dSStefano Zampini if (multilevel_allowed) { 6059f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 6060f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 6061f9eb5b7dSStefano Zampini } else { 6062f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 6063f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 6064c8587f34SStefano Zampini } 6065c8587f34SStefano Zampini 60666e683305SStefano Zampini /* print some info if requested */ 60676e683305SStefano Zampini if (pcbddc->dbg_flag) { 60686e683305SStefano Zampini if (!multilevel_allowed) { 60696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 60706e683305SStefano Zampini if (multilevel_requested) { 60716e683305SStefano 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); 60726e683305SStefano Zampini } else if (pcbddc->max_levels) { 60736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 60746e683305SStefano Zampini } 60756e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 60766e683305SStefano Zampini } 60776e683305SStefano Zampini } 60786e683305SStefano Zampini 6079f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 608057de7509SStefano Zampini if (coarse_mat) { 60816a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 60826e683305SStefano Zampini if (pcbddc->dbg_flag) { 608357de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 60846e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 60856e683305SStefano Zampini } 6086f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 6087312be037SStefano Zampini char prefix[256],str_level[16]; 6088e604994aSStefano Zampini size_t len; 608957de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 6090422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 6091c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 6092f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 609357de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6094c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 60956e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 6096c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 6097c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6098e604994aSStefano Zampini /* prefix */ 6099e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 6100e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 6101e604994aSStefano Zampini if (!pcbddc->current_level) { 6102e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 6103e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 6104c8587f34SStefano Zampini } else { 6105e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 6106312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 6107312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 610834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 6109312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 6110e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 6111e604994aSStefano Zampini } 6112e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 61133e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 61143e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 61153e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 61163e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 6117f9eb5b7dSStefano Zampini /* allow user customization */ 6118f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 61193e3c6dadSStefano Zampini } 61203e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 612151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 61223e3c6dadSStefano Zampini if (nisdofs) { 61233e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 61243e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 61253e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 61263e3c6dadSStefano Zampini } 61273e3c6dadSStefano Zampini } 61283e3c6dadSStefano Zampini if (nisneu) { 61293e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 61303e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 6131312be037SStefano Zampini } 613230368db7SStefano Zampini if (nisvert) { 613330368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 613430368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 613530368db7SStefano Zampini } 6136f9eb5b7dSStefano Zampini 6137f9eb5b7dSStefano Zampini /* get some info after set from options */ 6138f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 6139f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 61404f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 61416e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 6142f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6143f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 6144f9eb5b7dSStefano Zampini } 614539f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 61464f3a063dSStefano Zampini if (isredundant) { 61474f3a063dSStefano Zampini KSP inner_ksp; 61484f3a063dSStefano Zampini PC inner_pc; 61494f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 61504f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 61514f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 61524f3a063dSStefano Zampini } 6153f9eb5b7dSStefano Zampini 615457de7509SStefano Zampini /* parameters which miss an API */ 615557de7509SStefano Zampini if (isbddc) { 6156720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 6157720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 615857de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 615927b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 616027b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 61611ae86dd6SStefano Zampini if (coarsedivudotp) { 61621ae86dd6SStefano Zampini ierr = MatDestroy(&pcbddc_coarse->divudotp);CHKERRQ(ierr); 61631ae86dd6SStefano Zampini pcbddc_coarse->divudotp = coarsedivudotp; 61641ae86dd6SStefano Zampini coarsedivudotp = NULL; 61651ae86dd6SStefano Zampini } 616627b6a85dSStefano Zampini pcbddc_coarse->benign_compute_nonetflux = PETSC_TRUE; 6167720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 616859e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 6169720d30f9SStefano Zampini } 6170d4d8cf7bSStefano Zampini } 61719881197aSStefano Zampini 61723301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 61735a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 61743301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 61753301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 61763301b35fSStefano Zampini } 61773301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 61783301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 61793301b35fSStefano Zampini } 61803301b35fSStefano Zampini if (pc->pmat->spd_set) { 61813301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 61823301b35fSStefano Zampini } 618327b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 618427b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 618527b6a85dSStefano Zampini } 61866e683305SStefano Zampini /* set operators */ 61875f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 61886e683305SStefano Zampini if (pcbddc->dbg_flag) { 61896e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 61906e683305SStefano Zampini } 61916e683305SStefano Zampini } 61926e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6193b1ecc7b1SStefano Zampini #if 0 6194b9b85e73SStefano Zampini { 6195b9b85e73SStefano Zampini PetscViewer viewer; 6196b9b85e73SStefano Zampini char filename[256]; 6197b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6198b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 61996a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6200b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6201f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6202b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6203b9b85e73SStefano Zampini } 6204b9b85e73SStefano Zampini #endif 6205f9eb5b7dSStefano Zampini 620698a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 620798a51de6SStefano Zampini Vec crhs,csol; 620804708bb6SStefano Zampini 6209f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6210f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6211f347579bSStefano Zampini if (!csol) { 62122a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6213f9eb5b7dSStefano Zampini } 6214f347579bSStefano Zampini if (!crhs) { 62152a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6216f347579bSStefano Zampini } 6217b0f5fe93SStefano Zampini } 62181ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 6219b0f5fe93SStefano Zampini 6220b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6221b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6222b0f5fe93SStefano Zampini 6223b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 62244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 62254f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 62264f1b2e48SStefano Zampini } 6227b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6228b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6229b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6230b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6231b0f5fe93SStefano Zampini if (coarse_mat) { 6232b0f5fe93SStefano Zampini Vec nullv; 6233b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6234b0f5fe93SStefano Zampini PetscInt nl; 6235b0f5fe93SStefano Zampini 6236b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6237b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6238b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6239b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6240b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6241b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6242b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6243b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6244b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6245b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6246b0f5fe93SStefano Zampini } 6247b0f5fe93SStefano Zampini } 6248b0f5fe93SStefano Zampini 6249b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6250b0f5fe93SStefano Zampini PetscBool ispreonly; 6251b0f5fe93SStefano Zampini 6252b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6253b0f5fe93SStefano Zampini PetscBool isnull; 6254b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6255bef83e63SStefano Zampini if (isnull) { 6256b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6257b0f5fe93SStefano Zampini } 6258bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6259b0f5fe93SStefano Zampini } 6260b0f5fe93SStefano Zampini /* setup coarse ksp */ 6261b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6262cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6263cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 62646e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6265c8587f34SStefano Zampini KSP check_ksp; 62662b510759SStefano Zampini KSPType check_ksp_type; 6267c8587f34SStefano Zampini PC check_pc; 62686e683305SStefano Zampini Vec check_vec,coarse_vec; 62696a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 62702b510759SStefano Zampini PetscInt its; 62716e683305SStefano Zampini PetscBool compute_eigs; 62726e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 62736e683305SStefano Zampini PetscInt neigs; 62748e185a42SStefano Zampini const char *prefix; 6275c8587f34SStefano Zampini 62762b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 62776e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6278422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 627923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6280f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6281e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6282e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6283e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 62842b510759SStefano Zampini if (ispreonly) { 62852b510759SStefano Zampini check_ksp_type = KSPPREONLY; 62866e683305SStefano Zampini compute_eigs = PETSC_FALSE; 62872b510759SStefano Zampini } else { 6288cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 62896e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6290c8587f34SStefano Zampini } 6291c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 62926e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 62936e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 62946e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6295a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6296a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6297a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6298a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6299c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6300c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6301c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6302c8587f34SStefano Zampini /* create random vec */ 63032701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6304c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 63056e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6306c8587f34SStefano Zampini /* solve coarse problem */ 63076e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6308cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 63096e683305SStefano Zampini if (compute_eigs) { 6310854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6311854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 63126e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 63131ae86dd6SStefano Zampini if (neigs) { 63146e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 63156e683305SStefano Zampini lambda_min = eigs_r[0]; 63166e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 63172701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 63182701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6319cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6320cbcc2c2aSStefano Zampini } 6321c8587f34SStefano Zampini } 6322c8587f34SStefano Zampini } 63231ae86dd6SStefano Zampini } 6324cbcc2c2aSStefano Zampini 6325c8587f34SStefano Zampini /* check coarse problem residual error */ 63266e683305SStefano Zampini if (pcbddc->dbg_flag) { 63276e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 63286e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 63296e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6330c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 63316e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 63326e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6333779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 63346e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 63356e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 63366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 63376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6338b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6339b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6340b0f5fe93SStefano Zampini } 63416e683305SStefano Zampini if (compute_eigs) { 63426e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6343deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6344c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 63456e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 63466e683305SStefano 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); 63476e683305SStefano Zampini for (i=0;i<neigs;i++) { 63486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6349c8587f34SStefano Zampini } 63506e683305SStefano Zampini } 63516e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 63526e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 63536e683305SStefano Zampini } 6354e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 63552701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6356c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 63576e683305SStefano Zampini if (compute_eigs) { 63586e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 63596e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6360c8587f34SStefano Zampini } 63616e683305SStefano Zampini } 63626e683305SStefano Zampini } 6363bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6364cbcc2c2aSStefano Zampini /* print additional info */ 6365cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 63666e683305SStefano Zampini /* waits until all processes reaches this point */ 63676e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6368cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6369cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6370cbcc2c2aSStefano Zampini } 6371cbcc2c2aSStefano Zampini 63722b510759SStefano Zampini /* free memory */ 6373fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6374c8587f34SStefano Zampini PetscFunctionReturn(0); 6375c8587f34SStefano Zampini } 6376674ae819SStefano Zampini 6377f34684f1SStefano Zampini #undef __FUNCT__ 6378f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6379f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6380f34684f1SStefano Zampini { 6381f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6382f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6383f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6384dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6385dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 638673be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6387dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6388f34684f1SStefano Zampini PetscErrorCode ierr; 6389f34684f1SStefano Zampini 6390f34684f1SStefano Zampini PetscFunctionBegin; 6391f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 63926c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6393dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 63943bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6395dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6396dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6397dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6398dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6399dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6400dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 64016c4ed002SBarry 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); 6402dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6403dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6404dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6405dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6406dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6407f34684f1SStefano Zampini 6408f34684f1SStefano Zampini /* check numbering */ 6409f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6410019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6411dc456d91SStefano Zampini PetscInt i; 6412b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6413f34684f1SStefano Zampini 6414f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6415f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6416f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 64171575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6418019a44ceSStefano Zampini /* counter */ 6419019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6420019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6421019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6422019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6423019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6424019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6425f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6426f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6427727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6428f34684f1SStefano Zampini } 6429f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6430f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6431f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6432e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6433e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6434e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6435e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6436f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6437019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6438f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6439019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 64402c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 644175c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6442b9b85e73SStefano Zampini set_error = PETSC_TRUE; 64432c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 64442c66d082SStefano 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); 6445f34684f1SStefano Zampini } 6446f34684f1SStefano Zampini } 6447019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6448b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6449f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6450f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6451f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6452f34684f1SStefano Zampini } 6453f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6454f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6455e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6456e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6457f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6458f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6459b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6460ca8b9ea9SStefano Zampini PetscInt *gidxs; 6461ca8b9ea9SStefano Zampini 6462ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 64633bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6464f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6465f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6466f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6467f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 64684bc2dc4bSStefano 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); 6469f34684f1SStefano Zampini } 6470f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6471ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6472f34684f1SStefano Zampini } 6473f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 64741575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6475302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6476f34684f1SStefano Zampini } 64778bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6478f34684f1SStefano Zampini /* get back data */ 6479f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6480f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6481674ae819SStefano Zampini PetscFunctionReturn(0); 6482674ae819SStefano Zampini } 6483674ae819SStefano Zampini 6484e456f2a8SStefano Zampini #undef __FUNCT__ 6485e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6486a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6487e456f2a8SStefano Zampini { 6488e456f2a8SStefano Zampini IS localis_t; 6489a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6490e456f2a8SStefano Zampini PetscScalar *vals; 6491e456f2a8SStefano Zampini PetscErrorCode ierr; 6492e456f2a8SStefano Zampini 6493e456f2a8SStefano Zampini PetscFunctionBegin; 6494a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6495e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6496854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6497e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6498e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6499a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6500a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 65011035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6502a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 65031035eff8SStefano Zampini } 6504a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6505e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6506e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6507a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6508a7dc3881SStefano Zampini /* now compute set in local ordering */ 6509a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6510a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6511a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6512a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6513a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6514ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6515e456f2a8SStefano Zampini lsize++; 6516e456f2a8SStefano Zampini } 6517e456f2a8SStefano Zampini } 6518854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6519a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6520ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6521e456f2a8SStefano Zampini idxs[lsize++] = i; 6522e456f2a8SStefano Zampini } 6523e456f2a8SStefano Zampini } 6524a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6525a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6526e456f2a8SStefano Zampini *localis = localis_t; 6527e456f2a8SStefano Zampini PetscFunctionReturn(0); 6528e456f2a8SStefano Zampini } 6529906d46d4SStefano Zampini 6530b96c3477SStefano Zampini #undef __FUNCT__ 6531b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 653208122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6533b96c3477SStefano Zampini { 6534a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6535b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6536b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6537a64f4aa4SStefano Zampini Mat S_j; 6538b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6539b96c3477SStefano Zampini PetscBool free_used_adj; 6540b96c3477SStefano Zampini PetscErrorCode ierr; 6541b96c3477SStefano Zampini 6542b96c3477SStefano Zampini PetscFunctionBegin; 6543b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6544b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 654508122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6546b96c3477SStefano Zampini used_xadj = NULL; 6547b96c3477SStefano Zampini used_adjncy = NULL; 6548b96c3477SStefano Zampini } else { 654908122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 655008122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 655108122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 655208122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6553b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6554b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6555b96c3477SStefano Zampini } else { 65562fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6557b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6558b96c3477SStefano Zampini PetscInt nvtxs; 6559b96c3477SStefano Zampini 65602fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 65612fffb893SStefano Zampini if (flg_row) { 6562b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6563b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6564b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6565b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 65662fffb893SStefano Zampini } else { 65672fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 65682fffb893SStefano Zampini used_xadj = NULL; 65692fffb893SStefano Zampini used_adjncy = NULL; 65702fffb893SStefano Zampini } 65712fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6572b96c3477SStefano Zampini } 6573b96c3477SStefano Zampini } 6574d5574798SStefano Zampini 6575d5574798SStefano Zampini /* setup sub_schurs data */ 6576a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6577df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6578df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6579a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 658091af6908SStefano 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); 6581a64f4aa4SStefano Zampini } else { 65826816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6583b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6584a3df083aSStefano Zampini PetscInt benign_n; 658572b8c272SStefano Zampini Mat change = NULL; 65869d54b7f4SStefano Zampini Vec scaling = NULL; 658772b8c272SStefano Zampini IS change_primal = NULL; 6588a3df083aSStefano Zampini 65895feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 65905feab87aSStefano Zampini PetscInt n_vertices; 65915feab87aSStefano Zampini 65925feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 65932034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 65945feab87aSStefano Zampini } 659504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 659604708bb6SStefano Zampini if (!isseqaij) { 659704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 659804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 659904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 660004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 660104708bb6SStefano Zampini } else { 6602511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 660304708bb6SStefano Zampini } 660404708bb6SStefano Zampini } 6605a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6606a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6607ca92afb2SStefano Zampini } else { 6608a3df083aSStefano Zampini benign_n = 0; 6609ca92afb2SStefano Zampini } 6610b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6611b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6612b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 661372b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 661422db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6615b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 661622db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6617b7ab4a40SStefano Zampini } 6618b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6619b7ab4a40SStefano 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 */ 6620b7ab4a40SStefano Zampini if (need_change) { 662188c03ad3SStefano Zampini PC_IS *pcisf; 662288c03ad3SStefano Zampini PC_BDDC *pcbddcf; 662388c03ad3SStefano Zampini PC pcf; 662488c03ad3SStefano Zampini 6625e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 662688c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 662788c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 662888c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 662988c03ad3SStefano Zampini /* hacks */ 663088c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 663172b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 663272b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 663372b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 663472b8c272SStefano Zampini pcisf->n = pcis->n; 663572b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 663688c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 663788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 663888c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 663988c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 664088c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 664188c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 664272b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 664388c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 664488c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 664572b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 664672b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 664772b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 664872b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 664972b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 665088c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 665172b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 665288c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 665388c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 665488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 665588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 665688c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 665788c03ad3SStefano Zampini pcf->ops->destroy = NULL; 665888c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 665988c03ad3SStefano Zampini } 66609d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 666191af6908SStefano 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); 666272b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 666372b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6664ca92afb2SStefano Zampini } 6665d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6666b96c3477SStefano Zampini 6667b96c3477SStefano Zampini /* free adjacency */ 6668b96c3477SStefano Zampini if (free_used_adj) { 6669b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6670b96c3477SStefano Zampini } 6671b96c3477SStefano Zampini PetscFunctionReturn(0); 6672b96c3477SStefano Zampini } 6673b96c3477SStefano Zampini 6674b96c3477SStefano Zampini #undef __FUNCT__ 6675b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 667608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6677b96c3477SStefano Zampini { 6678b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6679b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6680b96c3477SStefano Zampini PCBDDCGraph graph; 6681b96c3477SStefano Zampini PetscErrorCode ierr; 6682b96c3477SStefano Zampini 6683b96c3477SStefano Zampini PetscFunctionBegin; 6684b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 668508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 66863301b35fSStefano Zampini IS verticesIS,verticescomm; 66873301b35fSStefano Zampini PetscInt vsize,*idxs; 6688b96c3477SStefano Zampini 6689b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 66903301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 66913301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66923301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 66933301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66943301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6695b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 66967fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6697441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 66983301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6699b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6700b96c3477SStefano Zampini } else { 6701b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6702b96c3477SStefano Zampini } 6703e4d548c7SStefano Zampini /* print some info */ 6704e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6705e4d548c7SStefano Zampini IS vertices; 6706e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 6707e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6708e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6709e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6710e4d548c7SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 6711e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6712e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6713e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6714e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6715e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6716e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6717e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6718e4d548c7SStefano Zampini } 6719b96c3477SStefano Zampini 6720b96c3477SStefano Zampini /* sub_schurs init */ 6721b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 6722b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 6723b334f244SStefano Zampini } 6724b334f244SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6725a64f4aa4SStefano Zampini 6726b96c3477SStefano Zampini /* free graph struct */ 672708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6728b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6729b96c3477SStefano Zampini } 6730b96c3477SStefano Zampini PetscFunctionReturn(0); 6731b96c3477SStefano Zampini } 6732fa34dd3eSStefano Zampini 6733fa34dd3eSStefano Zampini #undef __FUNCT__ 6734fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6735fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6736fa34dd3eSStefano Zampini { 6737fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6738fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6739fa34dd3eSStefano Zampini PetscErrorCode ierr; 6740fa34dd3eSStefano Zampini 6741fa34dd3eSStefano Zampini PetscFunctionBegin; 6742fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6743fa34dd3eSStefano Zampini IS zerodiag = NULL; 67444f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6745fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 67464f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 674775c01103SStefano Zampini PetscReal norm; 6748fa34dd3eSStefano Zampini PetscInt i; 6749fa34dd3eSStefano Zampini 6750fa34dd3eSStefano Zampini /* B0 and B0_B */ 6751fa34dd3eSStefano Zampini if (zerodiag) { 6752fa34dd3eSStefano Zampini IS dummy; 6753fa34dd3eSStefano Zampini 67544f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 67554f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6756fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6757fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6758fa34dd3eSStefano Zampini } 6759fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6760fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6761fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6762fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6763fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6764fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6765fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6766fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6767fa34dd3eSStefano Zampini /* S_j */ 6768fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6769fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6770fa34dd3eSStefano Zampini 6771fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6772fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6773fa34dd3eSStefano Zampini /* continuous in primal space */ 6774fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6775fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6776fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6777fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 67784f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 67794f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6780fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6781fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6782fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6783fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6784fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6785fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6786fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6787fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6788fa34dd3eSStefano Zampini 6789fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6790fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6791fa34dd3eSStefano Zampini /* local with Schur */ 6792fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6793fa34dd3eSStefano Zampini if (zerodiag) { 6794fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6796fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6797fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6798fa34dd3eSStefano Zampini } 6799fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6800fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6801fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6802fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6803fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6804fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6805fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6806fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6807fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6808fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6809fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6810fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6811fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6812fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6813fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6814fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6815fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6816fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6817fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6818fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6819fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6820fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6821fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6822fa34dd3eSStefano Zampini if (zerodiag) { 6823fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6824fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 68254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6826fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6827fa34dd3eSStefano Zampini } 6828fa34dd3eSStefano Zampini /* BDDC */ 6829fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6830fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6831fa34dd3eSStefano Zampini 6832fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6833fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6834fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6835fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 68364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 68374f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6838fa34dd3eSStefano Zampini } 68394f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6840fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6841fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6842fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6843fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6844fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6845fa34dd3eSStefano Zampini } 6846fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6847fa34dd3eSStefano Zampini } 6848