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 7d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 8d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 9d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 10d8203eabSStefano Zampini #undef __FUNCT__ 11d8203eabSStefano Zampini #define __FUNCT__ "PCBDDCNullSpaceCreate" 12d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 13d8203eabSStefano Zampini { 14d8203eabSStefano Zampini PetscErrorCode ierr; 15d8203eabSStefano Zampini PetscInt i; 16d8203eabSStefano Zampini 17d8203eabSStefano Zampini PetscFunctionBegin; 18d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 19d8203eabSStefano Zampini PetscInt first,last; 20d8203eabSStefano Zampini 21d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 22*86fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 23d8203eabSStefano Zampini if (i>=first && i < last) { 24d8203eabSStefano Zampini PetscScalar *data; 25d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 26d8203eabSStefano Zampini if (!has_const) { 27d8203eabSStefano Zampini data[i-first] = 1.; 28d8203eabSStefano Zampini } else { 29*86fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 30*86fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 31d8203eabSStefano Zampini } 32d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 33d8203eabSStefano Zampini } 34d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 35d8203eabSStefano Zampini } 36d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 37d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 38d8203eabSStefano Zampini PetscInt first,last; 39d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 40d8203eabSStefano Zampini if (i>=first && i < last) { 41d8203eabSStefano Zampini PetscScalar *data; 42d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 43d8203eabSStefano Zampini if (!has_const) { 44d8203eabSStefano Zampini data[i-first] = 0.; 45d8203eabSStefano Zampini } else { 46*86fa73c5SStefano Zampini data[2*i-first] = 0.; 47*86fa73c5SStefano Zampini data[2*i-first+1] = 0.; 48d8203eabSStefano Zampini } 49d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 50d8203eabSStefano Zampini } 51d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 52d8203eabSStefano Zampini } 53d8203eabSStefano Zampini PetscFunctionReturn(0); 54d8203eabSStefano Zampini } 55d8203eabSStefano Zampini 56669cc0f4SStefano Zampini #undef __FUNCT__ 57669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 588ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 59669cc0f4SStefano Zampini { 60a198735bSStefano Zampini Mat loc_divudotp; 61fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 628ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 63669cc0f4SStefano Zampini IS *faces,*edges; 64669cc0f4SStefano Zampini PetscScalar *vals; 65669cc0f4SStefano Zampini const PetscScalar *array; 66669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 671ae86dd6SStefano Zampini PetscMPIInt rank; 68a198735bSStefano Zampini PetscErrorCode ierr; 69669cc0f4SStefano Zampini 70669cc0f4SStefano Zampini PetscFunctionBegin; 71669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 72669cc0f4SStefano Zampini if (graph->twodim) { 73669cc0f4SStefano Zampini lmaxneighs = 2; 74669cc0f4SStefano Zampini } else { 75669cc0f4SStefano Zampini lmaxneighs = 1; 76669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 77669cc0f4SStefano Zampini const PetscInt *idxs; 78669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 79669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 80669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 81669cc0f4SStefano Zampini } 82669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 83669cc0f4SStefano Zampini } 84669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 85669cc0f4SStefano Zampini maxsize = 0; 86669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 87669cc0f4SStefano Zampini PetscInt nn; 88669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 89669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 90669cc0f4SStefano Zampini } 91669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 92669cc0f4SStefano Zampini PetscInt nn; 93669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 94669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 95669cc0f4SStefano Zampini } 96669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 97669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 98669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 998ae0ca82SStefano Zampini if (!transpose) { 1008ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 1018ae0ca82SStefano Zampini } else { 1028ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 1038ae0ca82SStefano Zampini } 104669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 1051ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 106d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 107669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1088ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 109669cc0f4SStefano Zampini } 110d8203eabSStefano Zampini 111669cc0f4SStefano Zampini /* compute local quad vec */ 112a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 1138ae0ca82SStefano Zampini if (!transpose) { 114a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 1158ae0ca82SStefano Zampini } else { 1168ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 1178ae0ca82SStefano Zampini } 118669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 1198ae0ca82SStefano Zampini if (!transpose) { 120a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 1218ae0ca82SStefano Zampini } else { 1228ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 1238ae0ca82SStefano Zampini } 124fa23a32eSStefano Zampini if (vl2l) { 125fa23a32eSStefano Zampini ierr = VecGetSubVector(v,vl2l,&vins);CHKERRQ(ierr); 126fa23a32eSStefano Zampini } else { 127fa23a32eSStefano Zampini vins = v; 128fa23a32eSStefano Zampini } 129fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 130669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 1319a962809SStefano Zampini 1321ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 1331ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 134669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 135669cc0f4SStefano Zampini const PetscInt *idxs; 136669cc0f4SStefano Zampini PetscInt idx,nn,j; 137669cc0f4SStefano Zampini 138669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 139669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 140669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 1411ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 142669cc0f4SStefano Zampini idx = -(idx+1); 143669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 144669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 145669cc0f4SStefano Zampini } 146669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 147669cc0f4SStefano Zampini const PetscInt *idxs; 148669cc0f4SStefano Zampini PetscInt idx,nn,j; 149669cc0f4SStefano Zampini 150669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 151669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 152669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 1531ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 154669cc0f4SStefano Zampini idx = -(idx+1); 155669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 156669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 157669cc0f4SStefano Zampini } 158c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 159fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 160fa23a32eSStefano Zampini if (vl2l) { 161fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 162fa23a32eSStefano Zampini } 163669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 164669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 165669cc0f4SStefano Zampini 166669cc0f4SStefano Zampini /* assemble near null space */ 167669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 168669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 169669cc0f4SStefano Zampini } 170669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 171669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 172669cc0f4SStefano Zampini } 173669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 174669cc0f4SStefano Zampini PetscFunctionReturn(0); 175669cc0f4SStefano Zampini } 176669cc0f4SStefano Zampini 177669cc0f4SStefano Zampini 178a3df083aSStefano Zampini #undef __FUNCT__ 1791f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 1801f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 1811f4df5f7SStefano Zampini { 1821f4df5f7SStefano Zampini PetscErrorCode ierr; 1831f4df5f7SStefano Zampini Vec local,global; 1841f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1851f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1861f4df5f7SStefano Zampini 1871f4df5f7SStefano Zampini PetscFunctionBegin; 1881f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 1891f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 1901f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 1911f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 1921f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 1931f4df5f7SStefano Zampini PetscInt i; 1941f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1951f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 1961f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1971f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 1981f4df5f7SStefano Zampini } 1991f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 2001f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 2011f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 2021f4df5f7SStefano Zampini } 2031f4df5f7SStefano Zampini } else { 204986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 2051f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 206986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 207986cdee1SStefano Zampini if (i > 1) { 208986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 2091f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 2101f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 2111f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 2121f4df5f7SStefano Zampini } 2131f4df5f7SStefano Zampini } 2141f4df5f7SStefano Zampini } 215986cdee1SStefano Zampini } 2161f4df5f7SStefano Zampini 2171f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 2181f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 2191f4df5f7SStefano Zampini } 2201f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 2211f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2221f4df5f7SStefano Zampini } 2231f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 2241f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 2251f4df5f7SStefano Zampini } 2261f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 2271f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 2281f4df5f7SStefano Zampini PetscFunctionReturn(0); 2291f4df5f7SStefano Zampini } 2301f4df5f7SStefano Zampini 2311f4df5f7SStefano Zampini #undef __FUNCT__ 2323e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 2333e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 2343e589ea0SStefano Zampini { 2353e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 2363e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2373e589ea0SStefano Zampini PetscErrorCode ierr; 2383e589ea0SStefano Zampini 2393e589ea0SStefano Zampini PetscFunctionBegin; 2403e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 2413e589ea0SStefano Zampini PetscFunctionReturn(0); 2423e589ea0SStefano Zampini } 2433e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2443e589ea0SStefano Zampini Vec swap; 2453e589ea0SStefano Zampini 2463e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 2473e589ea0SStefano Zampini swap = pcbddc->work_change; 2483e589ea0SStefano Zampini pcbddc->work_change = r; 2493e589ea0SStefano Zampini r = swap; 2503e589ea0SStefano Zampini } 2513e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2523e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2533e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 2543e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 2553e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2563e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2573e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2583e589ea0SStefano Zampini Vec swap; 2593e589ea0SStefano Zampini 2603e589ea0SStefano Zampini swap = r; 2613e589ea0SStefano Zampini r = pcbddc->work_change; 2623e589ea0SStefano Zampini pcbddc->work_change = swap; 2633e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 2643e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 2653e589ea0SStefano Zampini } 2663e589ea0SStefano Zampini PetscFunctionReturn(0); 2673e589ea0SStefano Zampini } 2683e589ea0SStefano Zampini 2693e589ea0SStefano Zampini #undef __FUNCT__ 270a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 271a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 272a3df083aSStefano Zampini { 273a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 274a3df083aSStefano Zampini PetscErrorCode ierr; 275a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 276a3df083aSStefano Zampini 277a3df083aSStefano Zampini PetscFunctionBegin; 278a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 279a3df083aSStefano Zampini if (transpose) { 280a3df083aSStefano Zampini apply_right = ctx->apply_left; 281a3df083aSStefano Zampini apply_left = ctx->apply_right; 282a3df083aSStefano Zampini } else { 283a3df083aSStefano Zampini apply_right = ctx->apply_right; 284a3df083aSStefano Zampini apply_left = ctx->apply_left; 285a3df083aSStefano Zampini } 286a3df083aSStefano Zampini reset_x = PETSC_FALSE; 287a3df083aSStefano Zampini if (apply_right) { 288a3df083aSStefano Zampini const PetscScalar *ax; 289a3df083aSStefano Zampini PetscInt nl,i; 290a3df083aSStefano Zampini 291a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 292a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 293a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 294a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 295a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 296a3df083aSStefano Zampini PetscScalar sum,val; 297a3df083aSStefano Zampini const PetscInt *idxs; 298a3df083aSStefano Zampini PetscInt nz,j; 299a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 300a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 301a3df083aSStefano Zampini sum = 0.; 302a3df083aSStefano Zampini if (ctx->apply_p0) { 303a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 304a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 305a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 306a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 307a3df083aSStefano Zampini } 308a3df083aSStefano Zampini } else { 309a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 310a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 311a3df083aSStefano Zampini } 312a3df083aSStefano Zampini } 313a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 314a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 315a3df083aSStefano Zampini } 316a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 317a3df083aSStefano Zampini reset_x = PETSC_TRUE; 318a3df083aSStefano Zampini } 319a3df083aSStefano Zampini if (transpose) { 320a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 321a3df083aSStefano Zampini } else { 322a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 323a3df083aSStefano Zampini } 324a3df083aSStefano Zampini if (reset_x) { 325a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 326a3df083aSStefano Zampini } 327a3df083aSStefano Zampini if (apply_left) { 328a3df083aSStefano Zampini PetscScalar *ay; 329a3df083aSStefano Zampini PetscInt i; 330a3df083aSStefano Zampini 331a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 332a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 333a3df083aSStefano Zampini PetscScalar sum,val; 334a3df083aSStefano Zampini const PetscInt *idxs; 335a3df083aSStefano Zampini PetscInt nz,j; 336a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 337a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 338a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 339a3df083aSStefano Zampini if (ctx->apply_p0) { 340a3df083aSStefano Zampini sum = 0.; 341a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 342a3df083aSStefano Zampini sum += ay[idxs[j]]; 343a3df083aSStefano Zampini ay[idxs[j]] += val; 344a3df083aSStefano Zampini } 345a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 346a3df083aSStefano Zampini } else { 347a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 348a3df083aSStefano Zampini ay[idxs[j]] += val; 349a3df083aSStefano Zampini } 350a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 351a3df083aSStefano Zampini } 352a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 353a3df083aSStefano Zampini } 354a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 355a3df083aSStefano Zampini } 356a3df083aSStefano Zampini PetscFunctionReturn(0); 357a3df083aSStefano Zampini } 358a3df083aSStefano Zampini 359a3df083aSStefano Zampini #undef __FUNCT__ 360a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 361a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 362a3df083aSStefano Zampini { 363a3df083aSStefano Zampini PetscErrorCode ierr; 364a3df083aSStefano Zampini 365a3df083aSStefano Zampini PetscFunctionBegin; 366a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 367a3df083aSStefano Zampini PetscFunctionReturn(0); 368a3df083aSStefano Zampini } 369a3df083aSStefano Zampini 370a3df083aSStefano Zampini #undef __FUNCT__ 371a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 372a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 373a3df083aSStefano Zampini { 374a3df083aSStefano Zampini PetscErrorCode ierr; 375a3df083aSStefano Zampini 376a3df083aSStefano Zampini PetscFunctionBegin; 377a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 378a3df083aSStefano Zampini PetscFunctionReturn(0); 379a3df083aSStefano Zampini } 380a3df083aSStefano Zampini 381a3df083aSStefano Zampini #undef __FUNCT__ 382a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 383a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 384a3df083aSStefano Zampini { 385a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 386a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 387a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 388a3df083aSStefano Zampini PetscErrorCode ierr; 389a3df083aSStefano Zampini 390a3df083aSStefano Zampini PetscFunctionBegin; 391a3df083aSStefano Zampini if (!restore) { 3921dd7afcfSStefano Zampini Mat A_IB,A_BI; 393a3df083aSStefano Zampini PetscScalar *work; 394b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 395a3df083aSStefano Zampini 3969a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 3979a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 398a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 399a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 400a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 401a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 402a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 403a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 404a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 405a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 406a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 407a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 408a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 409a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 410059032f7SStefano Zampini if (reuse) { 411a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 4121dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 413059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 414059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 415059032f7SStefano Zampini PetscInt i; 416059032f7SStefano Zampini 417059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 418059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 419059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 420059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 421059032f7SStefano Zampini } 422059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 4231dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 424059032f7SStefano Zampini } 425a3df083aSStefano Zampini ctx->A = pcis->A_IB; 426a3df083aSStefano Zampini ctx->work = work; 427a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 428a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 429a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 430a3df083aSStefano Zampini pcis->A_IB = A_IB; 431a3df083aSStefano Zampini 432a3df083aSStefano Zampini /* A_BI as A_IB^T */ 433a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 434a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 435a3df083aSStefano Zampini pcis->A_BI = A_BI; 436a3df083aSStefano Zampini } else { 4371dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 4381dd7afcfSStefano Zampini PetscFunctionReturn(0); 4391dd7afcfSStefano Zampini } 440a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 441a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 442a3df083aSStefano Zampini pcis->A_IB = ctx->A; 4431dd7afcfSStefano Zampini ctx->A = NULL; 4441dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 4451dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 4461dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 4471dd7afcfSStefano Zampini if (ctx->free) { 448059032f7SStefano Zampini PetscInt i; 4491dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 450059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 451059032f7SStefano Zampini } 452059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 453059032f7SStefano Zampini } 454a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 455a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 456a3df083aSStefano Zampini } 457a3df083aSStefano Zampini PetscFunctionReturn(0); 458a3df083aSStefano Zampini } 459a3df083aSStefano Zampini 460a3df083aSStefano Zampini /* used just in bddc debug mode */ 461a3df083aSStefano Zampini #undef __FUNCT__ 462a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 463a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 464a3df083aSStefano Zampini { 465a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 466a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 467a3df083aSStefano Zampini Mat An; 468a3df083aSStefano Zampini PetscErrorCode ierr; 469a3df083aSStefano Zampini 470a3df083aSStefano Zampini PetscFunctionBegin; 471a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 472a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 473a3df083aSStefano Zampini if (is1) { 474a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 475a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 476a3df083aSStefano Zampini } else { 477a3df083aSStefano Zampini *B = An; 478a3df083aSStefano Zampini } 479a3df083aSStefano Zampini PetscFunctionReturn(0); 480a3df083aSStefano Zampini } 481a3df083aSStefano Zampini 4821cf9b237SStefano Zampini /* TODO: add reuse flag */ 4831cf9b237SStefano Zampini #undef __FUNCT__ 4841cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 4851cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 4861cf9b237SStefano Zampini { 4871cf9b237SStefano Zampini Mat Bt; 4881cf9b237SStefano Zampini PetscScalar *a,*bdata; 4891cf9b237SStefano Zampini const PetscInt *ii,*ij; 4901cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 4911cf9b237SStefano Zampini PetscBool flg_row; 4921cf9b237SStefano Zampini PetscErrorCode ierr; 4931cf9b237SStefano Zampini 4941cf9b237SStefano Zampini PetscFunctionBegin; 4951cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 4961cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4971cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 4981cf9b237SStefano Zampini nnz = n; 4991cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 5001cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 5011cf9b237SStefano Zampini } 5021cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 5031cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 5041cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 5051cf9b237SStefano Zampini nnz = 0; 5061cf9b237SStefano Zampini bii[0] = 0; 5071cf9b237SStefano Zampini for (i=0;i<n;i++) { 5081cf9b237SStefano Zampini PetscInt j; 5091cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 5101cf9b237SStefano Zampini PetscScalar entry = a[j]; 5111cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 5121cf9b237SStefano Zampini bij[nnz] = ij[j]; 5131cf9b237SStefano Zampini bdata[nnz] = entry; 5141cf9b237SStefano Zampini nnz++; 5151cf9b237SStefano Zampini } 5161cf9b237SStefano Zampini } 5171cf9b237SStefano Zampini bii[i+1] = nnz; 5181cf9b237SStefano Zampini } 5191cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 5201cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 5211cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 5221cf9b237SStefano Zampini { 5231cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 5241cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 5251cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 5261cf9b237SStefano Zampini } 5271cf9b237SStefano Zampini *B = Bt; 5281cf9b237SStefano Zampini PetscFunctionReturn(0); 5291cf9b237SStefano Zampini } 5301cf9b237SStefano Zampini 531674ae819SStefano Zampini #undef __FUNCT__ 5324f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 5334f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 5344f1b2e48SStefano Zampini { 5354f1b2e48SStefano Zampini Mat B; 5364f1b2e48SStefano Zampini IS is_dummy,*cc_n; 5374f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 5384f1b2e48SStefano Zampini PCBDDCGraph graph; 5394f1b2e48SStefano Zampini PetscInt i,n; 5404f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 5414f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 5424f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 5434f1b2e48SStefano Zampini PetscErrorCode ierr; 5444f1b2e48SStefano Zampini 5454f1b2e48SStefano Zampini PetscFunctionBegin; 54663c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 54763c961adSStefano Zampini *ncc = 0; 54863c961adSStefano Zampini *cc = NULL; 54963c961adSStefano Zampini PetscFunctionReturn(0); 55063c961adSStefano Zampini } 5514f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 5524f1b2e48SStefano Zampini if (!isseqaij && filter) { 5531cf9b237SStefano Zampini PetscBool isseqdense; 5541cf9b237SStefano Zampini 5551cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 5561cf9b237SStefano Zampini if (!isseqdense) { 5574f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 5581cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 5591cf9b237SStefano Zampini PetscScalar *array; 5601cf9b237SStefano Zampini PetscReal chop=1.e-6; 5611cf9b237SStefano Zampini 5621cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 5631cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 5641cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 5651cf9b237SStefano Zampini for (i=0;i<n;i++) { 5661cf9b237SStefano Zampini PetscInt j; 5671cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 5681cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 5691cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 5701cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 5711cf9b237SStefano Zampini } 5721cf9b237SStefano Zampini } 5731cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 5749d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 5751cf9b237SStefano Zampini } 5764f1b2e48SStefano Zampini } else { 5774f1b2e48SStefano Zampini B = A; 5784f1b2e48SStefano Zampini } 5794f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5804f1b2e48SStefano Zampini 5814f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 5824f1b2e48SStefano Zampini if (filter) { 5834f1b2e48SStefano Zampini PetscScalar *data; 5844f1b2e48SStefano Zampini PetscInt j,cum; 5854f1b2e48SStefano Zampini 5864f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 5874f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini cum = 0; 5894f1b2e48SStefano Zampini for (i=0;i<n;i++) { 5904f1b2e48SStefano Zampini PetscInt t; 5914f1b2e48SStefano Zampini 5924f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 5934f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 5944f1b2e48SStefano Zampini continue; 5954f1b2e48SStefano Zampini } 5964f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 5974f1b2e48SStefano Zampini } 5984f1b2e48SStefano Zampini t = xadj_filtered[i]; 5994f1b2e48SStefano Zampini xadj_filtered[i] = cum; 6004f1b2e48SStefano Zampini cum += t; 6014f1b2e48SStefano Zampini } 6024f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini } else { 6044f1b2e48SStefano Zampini xadj_filtered = NULL; 6054f1b2e48SStefano Zampini adjncy_filtered = NULL; 6064f1b2e48SStefano Zampini } 6074f1b2e48SStefano Zampini 6084f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 6094f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 6104f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 6114f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 6124f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 6134f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 6144f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 6154f1b2e48SStefano Zampini if (xadj_filtered) { 6164f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 6174f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 6184f1b2e48SStefano Zampini } else { 6194f1b2e48SStefano Zampini graph->xadj = xadj; 6204f1b2e48SStefano Zampini graph->adjncy = adjncy; 6214f1b2e48SStefano Zampini } 6224f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 6234f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6244f1b2e48SStefano Zampini /* partial clean up */ 6254f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 6264f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 6271cf9b237SStefano Zampini if (A != B) { 6284f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 6294f1b2e48SStefano Zampini } 6304f1b2e48SStefano Zampini 6314f1b2e48SStefano Zampini /* get back data */ 6321cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 6331cf9b237SStefano Zampini if (cc) { 6344f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 6354f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 6364f1b2e48SStefano 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); 6374f1b2e48SStefano Zampini } 6384f1b2e48SStefano Zampini *cc = cc_n; 6391cf9b237SStefano Zampini } 6404f1b2e48SStefano Zampini /* clean up graph */ 6414f1b2e48SStefano Zampini graph->xadj = 0; 6424f1b2e48SStefano Zampini graph->adjncy = 0; 6434f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6444f1b2e48SStefano Zampini PetscFunctionReturn(0); 6454f1b2e48SStefano Zampini } 6464f1b2e48SStefano Zampini 6474f1b2e48SStefano Zampini #undef __FUNCT__ 6485408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 6495408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 6505408967cSStefano Zampini { 6515408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6525408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 653dee84bffSStefano Zampini IS dirIS = NULL; 6544f1b2e48SStefano Zampini PetscInt i; 6555408967cSStefano Zampini PetscErrorCode ierr; 6565408967cSStefano Zampini 6575408967cSStefano Zampini PetscFunctionBegin; 658dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 6595408967cSStefano Zampini if (zerodiag) { 6605408967cSStefano Zampini Mat A; 6615408967cSStefano Zampini Vec vec3_N; 6625408967cSStefano Zampini PetscScalar *vals; 6635408967cSStefano Zampini const PetscInt *idxs; 664d12d3064SStefano Zampini PetscInt nz,*count; 6655408967cSStefano Zampini 6665408967cSStefano Zampini /* p0 */ 6675408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 6685408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 6695408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6705408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6714f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 6725408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6735408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6745408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6755408967cSStefano Zampini /* v_I */ 6765408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 6775408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 6785408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6795408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 6805408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6815408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 6825408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6835408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6845408967cSStefano Zampini if (dirIS) { 6855408967cSStefano Zampini PetscInt n; 6865408967cSStefano Zampini 6875408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 6885408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 6895408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 6905408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6915408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 6925408967cSStefano Zampini } 6935408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 6945408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 6955408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 6965408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 697669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 6985408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 6995408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 7009a962809SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) 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])); 7015408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 7025408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 703d12d3064SStefano Zampini 704d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 705d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 706d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 707d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 708d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 709d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 7109a962809SStefano Zampini for (i=0;i<nz;i++) if (count[idxs[i]]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 711d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 712d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 7135408967cSStefano Zampini } 714dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 7155408967cSStefano Zampini 7165408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 7175408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 7184f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 7195408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 7204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 7215408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 722f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 723f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 724f2a566d8SStefano Zampini if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %d instead of %g\n",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i);CHKERRQ(ierr); 725f2a566d8SStefano Zampini } 7265408967cSStefano Zampini PetscFunctionReturn(0); 7275408967cSStefano Zampini } 7285408967cSStefano Zampini 7295408967cSStefano Zampini #undef __FUNCT__ 730339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 731339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 732339f8db1SStefano Zampini { 733339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7344f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 735b0f5fe93SStefano Zampini PetscInt nz,n; 7361f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 7374f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 738339f8db1SStefano Zampini PetscErrorCode ierr; 739339f8db1SStefano Zampini 740339f8db1SStefano Zampini PetscFunctionBegin; 7419f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 7429f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 743a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 744a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 745a3df083aSStefano Zampini } 746a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 747a3df083aSStefano Zampini pcbddc->benign_n = 0; 7481ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 7494f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 7504f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 7514f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 7521ae86dd6SStefano Zampini since the local Schur complements are already SPD 7534f1b2e48SStefano Zampini */ 7544f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 7554f1b2e48SStefano Zampini have_null = PETSC_TRUE; 75640fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 7574f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 7584f1b2e48SStefano Zampini 7594f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 7604f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 7614f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 7624f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 763ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 76440fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 76540fa8d13SStefano Zampini if (!sorted) { 76640fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 76740fa8d13SStefano Zampini } 76840fa8d13SStefano Zampini } else { 76940fa8d13SStefano Zampini pressures = NULL; 77040fa8d13SStefano Zampini } 77197d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 77297d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 77327b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 77497d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 775339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 776339f8db1SStefano Zampini if (!sorted) { 777339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 778339f8db1SStefano Zampini } 779339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 7804f1b2e48SStefano Zampini if (!nz) { 7814f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 7824f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 78340fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 78440fa8d13SStefano Zampini } 7854f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 7864f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 7874f1b2e48SStefano Zampini zerodiag_subs = NULL; 7884f1b2e48SStefano Zampini pcbddc->benign_n = 0; 7891f4df5f7SStefano Zampini n_interior_dofs = 0; 7901f4df5f7SStefano Zampini interior_dofs = NULL; 7911f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 7921f4df5f7SStefano Zampini PetscInt n,i,j; 7931f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 7941f4df5f7SStefano Zampini PetscInt *iwork; 7951f4df5f7SStefano Zampini 7961f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 7971f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7981f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 7991f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 80090648384SStefano Zampini for (i=1;i<n_neigh;i++) 8011f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 8021f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 8031f4df5f7SStefano Zampini for (i=0;i<n;i++) 8041f4df5f7SStefano Zampini if (!iwork[i]) 8051f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 8061f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 8071f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 8081f4df5f7SStefano Zampini } 8094f1b2e48SStefano Zampini if (has_null_pressures) { 8104f1b2e48SStefano Zampini IS *subs; 8111f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 8121f4df5f7SStefano Zampini const PetscInt *idxs; 8131f4df5f7SStefano Zampini PetscScalar *array; 8141f4df5f7SStefano Zampini Vec *work; 8151f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 8164f1b2e48SStefano Zampini 8174f1b2e48SStefano Zampini subs = pcbddc->local_subs; 8184f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 8191f4df5f7SStefano 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) */ 8201f4df5f7SStefano Zampini if (pcbddc->current_level) { 8211f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 8221f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 8231f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 8241f4df5f7SStefano Zampini /* work[0] = 1_p */ 8251f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 8261f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 8271f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8281f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 8291f4df5f7SStefano Zampini /* work[0] = 1_v */ 8301f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 8311f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 8321f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 8331f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 8341f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 8351f4df5f7SStefano Zampini } 8364f1b2e48SStefano Zampini if (nsubs > 1) { 8374f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 8384f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 8394f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 8404f1b2e48SStefano Zampini IS t_zerodiag_subs; 8414f1b2e48SStefano Zampini PetscInt nl; 8424f1b2e48SStefano Zampini 8434f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 8444f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 8454f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 8464f1b2e48SStefano Zampini if (nl) { 8474f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8484f1b2e48SStefano Zampini 8491f4df5f7SStefano Zampini if (pcbddc->current_level) { 8501f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 8511f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 8521f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 8531f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8541f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8551f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8561f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 8571f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 8581f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 8591f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 8601f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 8611f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8621f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8631f4df5f7SStefano Zampini valid = PETSC_FALSE; 8641f4df5f7SStefano Zampini break; 8651f4df5f7SStefano Zampini } 8661f4df5f7SStefano Zampini } 8671f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 8681f4df5f7SStefano Zampini } 8691f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 8701f4df5f7SStefano Zampini IS t_bc; 8711f4df5f7SStefano Zampini PetscInt nzb; 8721f4df5f7SStefano Zampini 8731f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 8741f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 8751f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 8761f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8771f4df5f7SStefano Zampini } 8781f4df5f7SStefano Zampini if (valid && pressures) { 8794f1b2e48SStefano Zampini IS t_pressure_subs; 8804f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 8814f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 8824f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 8834f1b2e48SStefano Zampini } 8844f1b2e48SStefano Zampini if (valid) { 8854f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 8864f1b2e48SStefano Zampini pcbddc->benign_n++; 8874f1b2e48SStefano Zampini } else { 8884f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 8894f1b2e48SStefano Zampini } 8904f1b2e48SStefano Zampini } 8914f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 8924f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 8934f1b2e48SStefano Zampini } 8944f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 8954f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8961f4df5f7SStefano Zampini 8971f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 8981f4df5f7SStefano Zampini PetscInt nzb; 8991f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 9001f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 9011f4df5f7SStefano Zampini } 9021f4df5f7SStefano Zampini if (valid && pressures) { 9034f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 9044f1b2e48SStefano Zampini } 9051f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 9061f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 9071f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 9081f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 9091f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 9101f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 9111f4df5f7SStefano Zampini valid = PETSC_FALSE; 9121f4df5f7SStefano Zampini break; 9131f4df5f7SStefano Zampini } 9141f4df5f7SStefano Zampini } 9151f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 9161f4df5f7SStefano Zampini } 9174f1b2e48SStefano Zampini if (valid) { 9184f1b2e48SStefano Zampini pcbddc->benign_n = 1; 919ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 9204f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 9214f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 9224f1b2e48SStefano Zampini } 9234f1b2e48SStefano Zampini } 9241f4df5f7SStefano Zampini if (pcbddc->current_level) { 9251f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 9264f1b2e48SStefano Zampini } 9271f4df5f7SStefano Zampini } 9281f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 9294f1b2e48SStefano Zampini 9304f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 931b9b0e38cSStefano Zampini PetscInt n; 932b9b0e38cSStefano Zampini 9334f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9344f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 935b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 936b9b0e38cSStefano Zampini if (n) { 9374f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 9384f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9394f1b2e48SStefano Zampini } 940b9b0e38cSStefano Zampini } 9414f1b2e48SStefano Zampini 9424f1b2e48SStefano Zampini /* final check for null pressures */ 9434f1b2e48SStefano Zampini if (zerodiag && pressures) { 9444f1b2e48SStefano Zampini PetscInt nz,np; 9454f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 9464f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 9474f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 9484f1b2e48SStefano Zampini } 9494f1b2e48SStefano Zampini 9504f1b2e48SStefano Zampini if (recompute_zerodiag) { 9514f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9524f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 9534f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 9544f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 9554f1b2e48SStefano Zampini } else { 9564f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 9574f1b2e48SStefano Zampini 9584f1b2e48SStefano Zampini nzn = 0; 9594f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9604f1b2e48SStefano Zampini PetscInt ns; 9614f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9624f1b2e48SStefano Zampini nzn += ns; 9634f1b2e48SStefano Zampini } 9644f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 9654f1b2e48SStefano Zampini nzn = 0; 9664f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9674f1b2e48SStefano Zampini PetscInt ns,*idxs; 9684f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9694f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9704f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 9714f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9724f1b2e48SStefano Zampini nzn += ns; 9734f1b2e48SStefano Zampini } 9744f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 9754f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 9764f1b2e48SStefano Zampini } 9774f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9784f1b2e48SStefano Zampini } 9794f1b2e48SStefano Zampini 980669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 981a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 982a198735bSStefano Zampini Mat A,loc_divudotp; 983a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 984a198735bSStefano Zampini IS row,col,isused = NULL; 985a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 986a198735bSStefano Zampini 9871f4df5f7SStefano Zampini if (pressures) { 9881f4df5f7SStefano Zampini isused = pressures; 9891f4df5f7SStefano Zampini } else { 9901f4df5f7SStefano Zampini isused = zerodiag; 9911f4df5f7SStefano Zampini } 992a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 993669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 9941ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 9951ae86dd6SStefano 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"); 996a198735bSStefano Zampini n_isused = 0; 997a198735bSStefano Zampini if (isused) { 998a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 999a198735bSStefano Zampini } 1000a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1001a198735bSStefano Zampini st = st-n_isused; 10021ae86dd6SStefano Zampini if (n) { 1003a198735bSStefano Zampini const PetscInt *gidxs; 1004a198735bSStefano Zampini 1005a198735bSStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 1006a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 1007a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 1008a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 1009a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 1010a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 10111ae86dd6SStefano Zampini } else { 1012a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 1013a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 1014a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 1015a198735bSStefano Zampini } 1016a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 1017a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 1018a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 1019a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 1020a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 1021a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 1022a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 1023a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 1024a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 1025a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 1026a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 1027a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 1028a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 1029a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 10301ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10311ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10321ae86dd6SStefano Zampini } 1033b3afcdbeSStefano Zampini 1034b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 10354f1b2e48SStefano Zampini if (has_null_pressures) { 10364f1b2e48SStefano Zampini IS zerodiagc; 10374f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 10384f1b2e48SStefano Zampini PetscInt i,s,*nnz; 10394f1b2e48SStefano Zampini 10404f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 1041339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 1042339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 1043339f8db1SStefano Zampini /* local change of basis for pressures */ 1044339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 104597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 1046339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 1047339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1048339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 10494f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 10504f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 10514f1b2e48SStefano Zampini PetscInt nzs,j; 10524f1b2e48SStefano Zampini 10534f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 10544f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 10554f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 10564f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 10574f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 10584f1b2e48SStefano Zampini } 1059339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 1060339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1061339f8db1SStefano Zampini /* set identity on velocities */ 1062339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 1063339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 1064339f8db1SStefano Zampini } 10654f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 10664f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 10679f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 10684f1b2e48SStefano 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); 1069339f8db1SStefano Zampini /* set change on pressures */ 10704f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 10714f1b2e48SStefano Zampini PetscScalar *array; 10724f1b2e48SStefano Zampini PetscInt nzs; 10734f1b2e48SStefano Zampini 10744f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 10754f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 10764f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 1077339f8db1SStefano Zampini PetscScalar vals[2]; 1078339f8db1SStefano Zampini PetscInt cols[2]; 1079339f8db1SStefano Zampini 1080339f8db1SStefano Zampini cols[0] = idxs[i]; 10814f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 1082339f8db1SStefano Zampini vals[0] = 1.; 1083b0f5fe93SStefano Zampini vals[1] = 1.; 10844f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1085339f8db1SStefano Zampini } 10864f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 10874f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 10884f1b2e48SStefano Zampini array[nzs-1] = 1.; 10894f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 10904f1b2e48SStefano Zampini /* store local idxs for p0 */ 10914f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 10924f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 1093339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 10944f1b2e48SStefano Zampini } 1095339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1096339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1097a3df083aSStefano Zampini /* project if needed */ 1098a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10991dd7afcfSStefano Zampini Mat M; 11001dd7afcfSStefano Zampini 11011dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 1102339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 11031dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 11041dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 1105a3df083aSStefano Zampini } 11064f1b2e48SStefano Zampini /* store global idxs for p0 */ 11074f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1108339f8db1SStefano Zampini } 1109ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 11104f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 1111b0f5fe93SStefano Zampini 1112b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 1113b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 111427b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 111527b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1116339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 1117339f8db1SStefano Zampini PetscFunctionReturn(0); 1118339f8db1SStefano Zampini } 1119339f8db1SStefano Zampini 1120339f8db1SStefano Zampini #undef __FUNCT__ 1121015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 1122015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 1123efc2fbd9SStefano Zampini { 1124efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1125de9d7bd0SStefano Zampini PetscScalar *array; 1126efc2fbd9SStefano Zampini PetscErrorCode ierr; 1127efc2fbd9SStefano Zampini 1128efc2fbd9SStefano Zampini PetscFunctionBegin; 1129efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 1130efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 11314f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1132efc2fbd9SStefano Zampini } 1133de9d7bd0SStefano Zampini if (get) { 1134efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 11354f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 11364f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 1137efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 1138de9d7bd0SStefano Zampini } else { 1139de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 1140de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1141de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1142de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 1143efc2fbd9SStefano Zampini } 1144efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1145efc2fbd9SStefano Zampini } 1146efc2fbd9SStefano Zampini 1147efc2fbd9SStefano Zampini #undef __FUNCT__ 1148c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 1149c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 1150c263805aSStefano Zampini { 1151c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1152c263805aSStefano Zampini PetscErrorCode ierr; 1153c263805aSStefano Zampini 1154c263805aSStefano Zampini PetscFunctionBegin; 1155c263805aSStefano Zampini /* TODO: add error checking 1156c263805aSStefano Zampini - avoid nested pop (or push) calls. 1157c263805aSStefano Zampini - cannot push before pop. 11581c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 1159c263805aSStefano Zampini */ 11604f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1161efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1162efc2fbd9SStefano Zampini } 1163c263805aSStefano Zampini if (pop) { 1164a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 11654f1b2e48SStefano Zampini IS is_p0; 11664f1b2e48SStefano Zampini MatReuse reuse; 1167c263805aSStefano Zampini 1168c263805aSStefano Zampini /* extract B_0 */ 11694f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 11704f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 11714f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 11724f1b2e48SStefano Zampini } 11734f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 11744f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 1175c263805aSStefano Zampini /* remove rows and cols from local problem */ 1176c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 117797d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 11784f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 11794f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 1180a3df083aSStefano Zampini } else { 1181a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1182a3df083aSStefano Zampini PetscScalar *vals; 1183a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 1184a3df083aSStefano Zampini 1185a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 1186a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 1187a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 11880b5adadeSStefano Zampini PetscInt *nnz; 1189a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 1190a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 1191a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1192331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 1193331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1194331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 1195331e053bSStefano Zampini nnz[i] = n - nnz[i]; 1196331e053bSStefano Zampini } 1197331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 1198331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1199331e053bSStefano Zampini } 1200a3df083aSStefano Zampini 1201a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1202a3df083aSStefano Zampini PetscScalar *array; 1203a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 1204a3df083aSStefano Zampini 1205a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 1206a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1207a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1208a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 1209a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1210a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 1211a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 1212a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 1213a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 1214a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 1215a3df083aSStefano Zampini cum = 0; 1216a3df083aSStefano Zampini for (j=0;j<n;j++) { 121722db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 1218a3df083aSStefano Zampini vals[cum] = array[j]; 1219a3df083aSStefano Zampini idxs_ins[cum] = j; 1220a3df083aSStefano Zampini cum++; 1221a3df083aSStefano Zampini } 1222a3df083aSStefano Zampini } 1223a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 1224a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 1225a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1226a3df083aSStefano Zampini } 1227a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1228a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1229a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 1230a3df083aSStefano Zampini } 1231c263805aSStefano Zampini } else { /* push */ 1232a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 12334f1b2e48SStefano Zampini PetscInt i; 12344f1b2e48SStefano Zampini 12354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 12364f1b2e48SStefano Zampini PetscScalar *B0_vals; 12374f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 12384f1b2e48SStefano Zampini 12394f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 12404f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 12417b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 12424f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 12434f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 12444f1b2e48SStefano Zampini } 1245c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1246c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1247a3df083aSStefano Zampini } else { 1248a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 1249a3df083aSStefano Zampini } 1250c263805aSStefano Zampini } 1251c263805aSStefano Zampini PetscFunctionReturn(0); 1252c263805aSStefano Zampini } 1253c263805aSStefano Zampini 1254c263805aSStefano Zampini #undef __FUNCT__ 1255b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 125608122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 1257b1b3d7a2SStefano Zampini { 1258b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 125908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 126008122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 126108122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 126208122e43SStefano Zampini PetscScalar *work,lwork; 126308122e43SStefano Zampini PetscScalar *St,*S,*eigv; 126408122e43SStefano Zampini PetscScalar *Sarray,*Starray; 126508122e43SStefano Zampini PetscReal *eigs,thresh; 12661b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 1267f6f667cfSStefano Zampini PetscBool allocated_S_St; 126808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 126908122e43SStefano Zampini PetscReal *rwork; 127008122e43SStefano Zampini #endif 1271b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1272b1b3d7a2SStefano Zampini 1273b1b3d7a2SStefano Zampini PetscFunctionBegin; 1274b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 1275af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 1276af25d912SStefano 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); 127706a4e24aSStefano Zampini 1278fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1279fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1280fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1281fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 12821575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1283fd14bc51SStefano Zampini } 1284fd14bc51SStefano Zampini 1285e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 1286e496cd5dSStefano 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); 1287e496cd5dSStefano Zampini } 1288e496cd5dSStefano Zampini 128908122e43SStefano Zampini /* max size of subsets */ 129008122e43SStefano Zampini mss = 0; 129108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 129208122e43SStefano Zampini PetscInt subset_size; 1293862806e4SStefano Zampini 129408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 129508122e43SStefano Zampini mss = PetscMax(mss,subset_size); 129608122e43SStefano Zampini } 129708122e43SStefano Zampini 129808122e43SStefano Zampini /* min/max and threshold */ 129908122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 1300f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 130108122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 1302f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 1303f6f667cfSStefano Zampini if (nmin) { 1304f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 1305f6f667cfSStefano Zampini } 130608122e43SStefano Zampini 130708122e43SStefano Zampini /* allocate lapack workspace */ 130808122e43SStefano Zampini cum = cum2 = 0; 130908122e43SStefano Zampini maxneigs = 0; 131008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 131108122e43SStefano Zampini PetscInt n,subset_size; 1312f6f667cfSStefano Zampini 131308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 131408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 13159162d606SStefano Zampini cum += subset_size; 13169162d606SStefano Zampini cum2 += subset_size*n; 131708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 131808122e43SStefano Zampini } 131908122e43SStefano Zampini if (mss) { 13209ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 132108122e43SStefano Zampini PetscBLASInt B_itype = 1; 132208122e43SStefano Zampini PetscBLASInt B_N = mss; 13234c6709b3SStefano Zampini PetscReal zero = 0.0; 13244c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 132508122e43SStefano Zampini 132608122e43SStefano Zampini B_lwork = -1; 132708122e43SStefano Zampini S = NULL; 132808122e43SStefano Zampini St = NULL; 1329a58a30b4SStefano Zampini eigs = NULL; 1330a58a30b4SStefano Zampini eigv = NULL; 1331a58a30b4SStefano Zampini B_iwork = NULL; 1332a58a30b4SStefano Zampini B_ifail = NULL; 1333d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1334d1710679SStefano Zampini rwork = NULL; 1335d1710679SStefano Zampini #endif 13368bec7fa6SStefano Zampini thresh = 1.0; 133708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 133808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 133908122e43SStefano 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)); 134008122e43SStefano Zampini #else 134108122e43SStefano 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)); 134208122e43SStefano Zampini #endif 134308122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 134408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 134508122e43SStefano Zampini } else { 134608122e43SStefano Zampini /* TODO */ 134708122e43SStefano Zampini } 134808122e43SStefano Zampini } else { 134908122e43SStefano Zampini lwork = 0; 135008122e43SStefano Zampini } 135108122e43SStefano Zampini 135208122e43SStefano Zampini nv = 0; 1353d62866d3SStefano 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) */ 1354d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 135508122e43SStefano Zampini } 13564c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1357f6f667cfSStefano Zampini if (allocated_S_St) { 1358f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1359f6f667cfSStefano Zampini } 1360f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 136108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 136208122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 136308122e43SStefano Zampini #endif 13649162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 13659162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 13669162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 136708122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 13689162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 136908122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 137008122e43SStefano Zampini 137108122e43SStefano Zampini maxneigs = 0; 137272b8c272SStefano Zampini cum = cumarray = 0; 13739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 13749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1375d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 137608122e43SStefano Zampini const PetscInt *idxs; 137708122e43SStefano Zampini 1378d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 137908122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 138008122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 138108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 138208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 13839162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 13849162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 138508122e43SStefano Zampini } 1386d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 138708122e43SStefano Zampini } 138808122e43SStefano Zampini 138908122e43SStefano Zampini if (mss) { /* multilevel */ 139008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 139108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 139208122e43SStefano Zampini } 139308122e43SStefano Zampini 1394ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 139508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 139608122e43SStefano Zampini const PetscInt *idxs; 13979d54b7f4SStefano Zampini PetscReal upper,lower; 1398862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 139908122e43SStefano Zampini PetscBLASInt B_N; 1400aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 140108122e43SStefano Zampini 14029d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 14039d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 14049d54b7f4SStefano Zampini lower = thresh; 14059d54b7f4SStefano Zampini } else { 14069d54b7f4SStefano Zampini upper = 1./thresh; 14079d54b7f4SStefano Zampini lower = 0.; 14089d54b7f4SStefano Zampini } 1409862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1410ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1411f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1412f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 14139ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1414aff50787SStefano Zampini PetscInt j,k; 1415aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1416aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1417aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 141808122e43SStefano Zampini } 141908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1420aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1421aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1422aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1423aff50787SStefano Zampini } 142408122e43SStefano Zampini } 142508122e43SStefano Zampini } else { 142608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 142708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 142808122e43SStefano Zampini } 14298bec7fa6SStefano Zampini } else { 1430f6f667cfSStefano Zampini S = Sarray + cumarray; 1431f6f667cfSStefano Zampini St = Starray + cumarray; 14328bec7fa6SStefano Zampini } 1433aff50787SStefano Zampini /* see if we can save some work */ 1434b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1435aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1436aff50787SStefano Zampini } 1437aff50787SStefano Zampini 1438b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1439aff50787SStefano Zampini B_neigs = 0; 1440aff50787SStefano Zampini } else { 14419ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 144208122e43SStefano Zampini PetscBLASInt B_itype = 1; 1443f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 14444c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 14459552c7c7SStefano Zampini PetscInt nmin_s; 1446b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 144708122e43SStefano Zampini 1448fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 14498bec7fa6SStefano 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]]); 1450fd14bc51SStefano Zampini } 1451d16cbb6bSStefano Zampini 1452b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1453b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1454b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1455b7ab4a40SStefano Zampini } 1456b7ab4a40SStefano Zampini 145708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1458b7ab4a40SStefano Zampini if (compute_range) { 1459d16cbb6bSStefano Zampini 1460d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 146108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 14629d54b7f4SStefano 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)); 146308122e43SStefano Zampini #else 14649d54b7f4SStefano 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)); 146508122e43SStefano Zampini #endif 1466b7ab4a40SStefano Zampini } else if (!same_data) { 1467d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1468d16cbb6bSStefano Zampini B_IL = 1; 1469d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 14709d54b7f4SStefano 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)); 1471d16cbb6bSStefano Zampini #else 14729d54b7f4SStefano 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)); 1473d16cbb6bSStefano Zampini #endif 1474b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1475b7ab4a40SStefano Zampini PetscInt k; 1476b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1477b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1478b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1479b7ab4a40SStefano Zampini nmin = nmax; 1480b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1481b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1482b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1483b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1484b7ab4a40SStefano Zampini } 1485d16cbb6bSStefano Zampini } 148608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 148708122e43SStefano Zampini if (B_ierr) { 14886c4ed002SBarry 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); 14896c4ed002SBarry 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); 14906c4ed002SBarry 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); 149108122e43SStefano Zampini } 149208122e43SStefano Zampini 149308122e43SStefano Zampini if (B_neigs > nmax) { 1494fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1495fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1496fd14bc51SStefano Zampini } 14979d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 149808122e43SStefano Zampini B_neigs = nmax; 149908122e43SStefano Zampini } 150008122e43SStefano Zampini 15019552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 15029552c7c7SStefano Zampini if (B_neigs < nmin_s) { 150308122e43SStefano Zampini PetscBLASInt B_neigs2; 150408122e43SStefano Zampini 15059d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1506f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 15079d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 15089d54b7f4SStefano Zampini } else { 15099d54b7f4SStefano Zampini B_IL = B_neigs + 1; 15109d54b7f4SStefano Zampini B_IU = nmin_s; 15119d54b7f4SStefano Zampini } 1512fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1513fd14bc51SStefano 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); 1514fd14bc51SStefano Zampini } 15159ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 15161ae86dd6SStefano Zampini PetscInt j,k; 151708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 15181ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 15191ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 15201ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 152108122e43SStefano Zampini } 152208122e43SStefano Zampini } 152308122e43SStefano Zampini } else { 152408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 152508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 152608122e43SStefano Zampini } 152708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 152808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 15299d54b7f4SStefano 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)); 153008122e43SStefano Zampini #else 15319d54b7f4SStefano 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)); 153208122e43SStefano Zampini #endif 153308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 153408122e43SStefano Zampini B_neigs += B_neigs2; 153508122e43SStefano Zampini } 153608122e43SStefano Zampini if (B_ierr) { 15376c4ed002SBarry 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); 15386c4ed002SBarry 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); 15396c4ed002SBarry 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); 154008122e43SStefano Zampini } 1541fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1542ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 154308122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 154408122e43SStefano Zampini if (eigs[j] == 0.0) { 1545ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 154608122e43SStefano Zampini } else { 15479d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1548ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 15499d54b7f4SStefano Zampini } else { 15509d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 15519d54b7f4SStefano Zampini } 1552fd14bc51SStefano Zampini } 155308122e43SStefano Zampini } 155408122e43SStefano Zampini } 155508122e43SStefano Zampini } else { 155608122e43SStefano Zampini /* TODO */ 155708122e43SStefano Zampini } 1558aff50787SStefano Zampini } 15596c3e6151SStefano Zampini /* change the basis back to the original one */ 15606c3e6151SStefano Zampini if (sub_schurs->change) { 156172b8c272SStefano Zampini Mat change,phi,phit; 15626c3e6151SStefano Zampini 15636c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 15646c3e6151SStefano Zampini PetscInt ii; 15656c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 15666c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15676c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 1568684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1569684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 1570684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 1571684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1572684229deSStefano Zampini #else 15736c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 1574684229deSStefano Zampini #endif 15756c3e6151SStefano Zampini } 15766c3e6151SStefano Zampini } 15776c3e6151SStefano Zampini } 157872b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 15796c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 158072b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 15816c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 15826c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 15836c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 15846c3e6151SStefano Zampini } 15858bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 15868bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 15879162d606SStefano Zampini if (B_neigs) { 15889162d606SStefano 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); 1589fd14bc51SStefano Zampini 1590fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 15919552c7c7SStefano Zampini PetscInt ii; 15929552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1593ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15949552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1595ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1596ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1597ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1598ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1599ac47001eSStefano Zampini #else 1600ac47001eSStefano 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); 1601ac47001eSStefano Zampini #endif 16029552c7c7SStefano Zampini } 16039552c7c7SStefano Zampini } 1604fd14bc51SStefano Zampini } 16059162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 16069162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 16079162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 16089162d606SStefano Zampini cum++; 160908122e43SStefano Zampini } 161008122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 161108122e43SStefano Zampini /* shift for next computation */ 161208122e43SStefano Zampini cumarray += subset_size*subset_size; 161308122e43SStefano Zampini } 1614fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1615fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1616fd14bc51SStefano Zampini } 161708122e43SStefano Zampini 161808122e43SStefano Zampini if (mss) { 161908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 162008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1621f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1622f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1623f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 162408122e43SStefano Zampini } 1625f6f667cfSStefano Zampini if (allocated_S_St) { 1626f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1627f6f667cfSStefano Zampini } 1628f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 162908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 163008122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 163108122e43SStefano Zampini #endif 163208122e43SStefano Zampini if (pcbddc->dbg_flag) { 16331b968477SStefano Zampini PetscInt maxneigs_r; 1634b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 16359b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 163608122e43SStefano Zampini } 163708122e43SStefano Zampini PetscFunctionReturn(0); 163808122e43SStefano Zampini } 1639b1b3d7a2SStefano Zampini 1640674ae819SStefano Zampini #undef __FUNCT__ 1641c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1642c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1643c8587f34SStefano Zampini { 16448629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1645c8587f34SStefano Zampini PetscErrorCode ierr; 1646c8587f34SStefano Zampini 1647c8587f34SStefano Zampini PetscFunctionBegin; 1648f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 16495e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1650c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1651c8587f34SStefano Zampini 1652684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 16530fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1654684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1655c8587f34SStefano Zampini 16568629588bSStefano Zampini /* 16578629588bSStefano Zampini Setup local correction and local part of coarse basis. 16588629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 16598629588bSStefano Zampini */ 166047f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 16618629588bSStefano Zampini 16628629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 16638629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 16648629588bSStefano Zampini 16658629588bSStefano Zampini /* free */ 16668629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1667c8587f34SStefano Zampini PetscFunctionReturn(0); 1668c8587f34SStefano Zampini } 1669c8587f34SStefano Zampini 1670c8587f34SStefano Zampini #undef __FUNCT__ 1671674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1672674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1673674ae819SStefano Zampini { 1674674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1675674ae819SStefano Zampini PetscErrorCode ierr; 1676674ae819SStefano Zampini 1677674ae819SStefano Zampini PetscFunctionBegin; 1678674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1679674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 168030368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1681674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1682785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1683674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1684f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1685f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1686785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 168763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 168863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1689674ae819SStefano Zampini PetscFunctionReturn(0); 1690674ae819SStefano Zampini } 1691674ae819SStefano Zampini 1692674ae819SStefano Zampini #undef __FUNCT__ 1693674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1694674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1695674ae819SStefano Zampini { 1696674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16974f1b2e48SStefano Zampini PetscInt i; 1698674ae819SStefano Zampini PetscErrorCode ierr; 1699674ae819SStefano Zampini 1700674ae819SStefano Zampini PetscFunctionBegin; 1701b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1702674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 170316909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 17041dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1705674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1706669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 1707fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 1708674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 170971582508SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 17104f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 17114f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 17124f1b2e48SStefano Zampini } 17134f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1714b334f244SStefano Zampini if (pcbddc->sub_schurs) { 1715b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1716b334f244SStefano Zampini } 1717674ae819SStefano Zampini PetscFunctionReturn(0); 1718674ae819SStefano Zampini } 1719674ae819SStefano Zampini 1720674ae819SStefano Zampini #undef __FUNCT__ 1721674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1722674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1723674ae819SStefano Zampini { 1724674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1725674ae819SStefano Zampini PetscErrorCode ierr; 1726674ae819SStefano Zampini 1727674ae819SStefano Zampini PetscFunctionBegin; 1728674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 172958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1730ca92afb2SStefano Zampini PetscScalar *array; 173106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 173206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 173358da7f69SStefano Zampini } 1734674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1735674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 173615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 173715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1738674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1739674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1740674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 174106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1742674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1743674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17448ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1745674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1746674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1747674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1748f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1749f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1750f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1751f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1752727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 17530e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1754f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 175570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 175681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 17570369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 17581dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 17594f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 17608b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1761ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1762ca92afb2SStefano Zampini PetscInt i; 1763ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1764ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1765ca92afb2SStefano Zampini } 1766ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1767ca92afb2SStefano Zampini } 17684f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1769674ae819SStefano Zampini PetscFunctionReturn(0); 1770674ae819SStefano Zampini } 1771674ae819SStefano Zampini 1772674ae819SStefano Zampini #undef __FUNCT__ 1773f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1774f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 17756bfb1811SStefano Zampini { 17766bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 17776bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 17786bfb1811SStefano Zampini VecType impVecType; 17794f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 17806bfb1811SStefano Zampini PetscErrorCode ierr; 17816bfb1811SStefano Zampini 17826bfb1811SStefano Zampini PetscFunctionBegin; 17836c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1784e7b262bdSStefano Zampini /* get sizes */ 17854f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1786b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 17876bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1788e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1789e7b262bdSStefano Zampini /* R nodes */ 1790e7b262bdSStefano Zampini old_size = -1; 1791e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1792e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1793e7b262bdSStefano Zampini } 1794e7b262bdSStefano Zampini if (n_R != old_size) { 1795e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1796e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17976bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 17986bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 17996bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 18006bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1801e7b262bdSStefano Zampini } 1802e7b262bdSStefano Zampini /* local primal dofs */ 1803e7b262bdSStefano Zampini old_size = -1; 1804e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1805e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1806e7b262bdSStefano Zampini } 1807e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1808e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 180983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1810e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 18116bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1812e7b262bdSStefano Zampini } 1813e7b262bdSStefano Zampini /* local explicit constraints */ 1814e7b262bdSStefano Zampini old_size = -1; 1815e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1816e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1817e7b262bdSStefano Zampini } 1818e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1819e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 182083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 182183b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 182283b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 182383b7ccabSStefano Zampini } 18246bfb1811SStefano Zampini PetscFunctionReturn(0); 18256bfb1811SStefano Zampini } 18266bfb1811SStefano Zampini 18276bfb1811SStefano Zampini #undef __FUNCT__ 182847f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 182947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 183088ebb749SStefano Zampini { 183125084f0cSStefano Zampini PetscErrorCode ierr; 183225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 183388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 183488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1835d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 183625084f0cSStefano Zampini /* submatrices of local problem */ 183780677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 183806656605SStefano Zampini /* submatrices of local coarse problem */ 183906656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 184025084f0cSStefano Zampini /* working matrices */ 184106656605SStefano Zampini Mat C_CR; 184225084f0cSStefano Zampini /* additional working stuff */ 184306656605SStefano Zampini PC pc_R; 18444f1b2e48SStefano Zampini Mat F; 18455cbda25cSStefano Zampini Vec dummy_vec; 1846a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 184725084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 184806656605SStefano Zampini PetscScalar *work; 184906656605SStefano Zampini PetscInt *idx_V_B; 1850ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 185106656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1852ffd830a3SStefano Zampini 185325084f0cSStefano Zampini /* some shortcuts to scalars */ 185406656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 185588ebb749SStefano Zampini 185688ebb749SStefano Zampini PetscFunctionBegin; 18579a962809SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1858ffd830a3SStefano Zampini 1859ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1860b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 18614f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1862b371cd4fSStefano Zampini n_B = pcis->n_B; 1863b371cd4fSStefano Zampini n_D = pcis->n - n_B; 186488ebb749SStefano Zampini n_R = pcis->n - n_vertices; 186588ebb749SStefano Zampini 186688ebb749SStefano Zampini /* vertices in boundary numbering */ 1867785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 18680e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 18696c4ed002SBarry 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); 187088ebb749SStefano Zampini 187106656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1872019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 187306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 187406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 187506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 187606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 187706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 187806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 187906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 188006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 188106656605SStefano Zampini 188206656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 188306656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 188406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 188506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 188606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1887ffd830a3SStefano Zampini lda_rhs = n_R; 1888a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 188906656605SStefano Zampini if (isLU || isILU || isCHOL) { 189006656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1891b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 1892df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1893d62866d3SStefano Zampini MatFactorType type; 1894d62866d3SStefano Zampini 1895df4d28bfSStefano Zampini F = reuse_solver->F; 18966816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1897d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1898ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 189922db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 190006656605SStefano Zampini } else { 190106656605SStefano Zampini F = NULL; 190206656605SStefano Zampini } 190306656605SStefano Zampini 1904ffd830a3SStefano Zampini /* allocate workspace */ 1905ffd830a3SStefano Zampini n = 0; 1906ffd830a3SStefano Zampini if (n_constraints) { 1907ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1908ffd830a3SStefano Zampini } 1909ffd830a3SStefano Zampini if (n_vertices) { 1910ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1911ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1912ffd830a3SStefano Zampini } 1913ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1914ffd830a3SStefano Zampini 19155cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 19165cbda25cSStefano Zampini dummy_vec = NULL; 19175cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 19185cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 19195cbda25cSStefano Zampini } 19205cbda25cSStefano Zampini 192188ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 192288ebb749SStefano Zampini if (n_constraints) { 192372b8c272SStefano Zampini Mat M1,M2,M3,C_B; 192406656605SStefano Zampini IS is_aux; 192580677318SStefano Zampini PetscScalar *array,*array2; 192606656605SStefano Zampini 1927f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 192880677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 192988ebb749SStefano Zampini 193025084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 193125084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 19328ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 193372b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 193488ebb749SStefano Zampini 193580677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 193680677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1937ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 193888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 193906656605SStefano Zampini const PetscScalar *row_cmat_values; 194006656605SStefano Zampini const PetscInt *row_cmat_indices; 194106656605SStefano Zampini PetscInt size_of_constraint,j; 194288ebb749SStefano Zampini 194306656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 194406656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1945ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 194606656605SStefano Zampini } 194706656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 194806656605SStefano Zampini } 1949ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 195006656605SStefano Zampini if (F) { 195106656605SStefano Zampini Mat B; 195206656605SStefano Zampini 1953ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1954a3df083aSStefano Zampini if (need_benign_correction) { 1955df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1956a3df083aSStefano Zampini 195772b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 195872b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1959a3df083aSStefano Zampini } 196080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1961a3df083aSStefano Zampini if (need_benign_correction) { 1962a3df083aSStefano Zampini PetscScalar *marr; 1963df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1964a3df083aSStefano Zampini 1965a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 19665cbda25cSStefano Zampini if (lda_rhs != n_R) { 19675cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 19685cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19695cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19705cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19715cbda25cSStefano Zampini } 19725cbda25cSStefano Zampini } else { 1973a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1974a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19755cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1976a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1977a3df083aSStefano Zampini } 19785cbda25cSStefano Zampini } 1979a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1980a3df083aSStefano Zampini } 198106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 198206656605SStefano Zampini } else { 198380677318SStefano Zampini PetscScalar *marr; 198480677318SStefano Zampini 198580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 198606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1987ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1988ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 198906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 199006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 199106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 199206656605SStefano Zampini } 199380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 199406656605SStefano Zampini } 199580677318SStefano Zampini if (!pcbddc->switch_static) { 199680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 199780677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 199880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 199980677318SStefano Zampini for (i=0;i<n_constraints;i++) { 2000ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 200180677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 200280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 200580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 200680677318SStefano Zampini } 200780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 200880677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 200972b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 201080677318SStefano Zampini } else { 2011ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2012ffd830a3SStefano Zampini IS dummy; 2013ffd830a3SStefano Zampini 2014ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 201572b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 2016ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 2017ffd830a3SStefano Zampini } else { 201880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 201980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 2020ffd830a3SStefano Zampini } 202125084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 202280677318SStefano Zampini } 202380677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 202480677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 202580677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 202606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 202706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 202880677318SStefano Zampini if (isCHOL) { 202980677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 203080677318SStefano Zampini } else { 203125084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 203280677318SStefano Zampini } 203380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 203406656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 203525084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 203625084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 203725084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 203880677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 203972b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 204072b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 204106656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 204206656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 2043f4ddd8eeSStefano Zampini } 2044fc227af8SStefano Zampini 2045fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 204688ebb749SStefano Zampini if (n_vertices) { 204706656605SStefano Zampini IS is_aux; 20483a50541eSStefano Zampini 2049b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 20506816873aSStefano Zampini IS tis; 20516816873aSStefano Zampini 20526816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 20536816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 20546816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 20556816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 20566816873aSStefano Zampini } else { 20573a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 20586816873aSStefano Zampini } 20599577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 20609577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 206104708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 206225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 206388ebb749SStefano Zampini } 206488ebb749SStefano Zampini 206588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 2066f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 206706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 206806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 206906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 207006656605SStefano Zampini } 2071f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 207206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 207306656605SStefano Zampini PetscScalar *marray; 207406656605SStefano Zampini 207506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 207606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 2077f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2078f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 2079f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 2080f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2081f4ddd8eeSStefano Zampini } 2082f4ddd8eeSStefano Zampini } 208306656605SStefano Zampini 2084f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 208506656605SStefano Zampini PetscScalar *marray; 208688ebb749SStefano Zampini 208706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20888eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 208906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 209088ebb749SStefano Zampini } 20913301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 209206656605SStefano Zampini n *= 2; 209388ebb749SStefano Zampini } 209406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 209506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 209606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20978eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 209806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 209906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 210088ebb749SStefano Zampini } 21013301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 210206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 21038eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 210406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 210506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 210688ebb749SStefano Zampini } 210788ebb749SStefano Zampini } else { 2108c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 2109c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 21101b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2111c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 2112c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 2113c0553b1fSStefano Zampini } 211488ebb749SStefano Zampini } 211506656605SStefano Zampini } 2116019a44ceSStefano Zampini 211706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 21184f1b2e48SStefano Zampini p0_lidx_I = NULL; 21194f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 2120d12edf2fSStefano Zampini const PetscInt *idxs; 2121d12edf2fSStefano Zampini 2122d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 21234f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 21244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 21254f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 21264f1b2e48SStefano Zampini } 2127d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 2128d12edf2fSStefano Zampini } 2129d16cbb6bSStefano Zampini 213006656605SStefano Zampini /* vertices */ 213106656605SStefano Zampini if (n_vertices) { 213216f15bc4SStefano Zampini 2133af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 213404708bb6SStefano Zampini 213516f15bc4SStefano Zampini if (n_R) { 213614393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 213706656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 213816f15bc4SStefano Zampini PetscScalar *x,*y; 213904708bb6SStefano Zampini PetscBool isseqaij; 214006656605SStefano Zampini 214121eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 214214393ed6SStefano Zampini if (need_benign_correction) { 214314393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 214414393ed6SStefano Zampini IS is_p0; 214514393ed6SStefano Zampini PetscInt *idxs_p0,n; 214614393ed6SStefano Zampini 214714393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 214814393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 214914393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2150af25d912SStefano 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); 215114393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 215214393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 215314393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 215414393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 215514393ed6SStefano Zampini } 215614393ed6SStefano Zampini 2157ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2158af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2159ffd830a3SStefano Zampini } else { 2160ca92afb2SStefano Zampini PetscScalar *av,*array; 2161ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2162ca92afb2SStefano Zampini PetscInt n; 2163ca92afb2SStefano Zampini PetscBool flg_row; 2164ffd830a3SStefano Zampini 2165ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2166ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 21679d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2168ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2169ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2170ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2171ca92afb2SStefano Zampini PetscInt j; 2172ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2173ffd830a3SStefano Zampini } 2174ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2175ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2176ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2177ffd830a3SStefano Zampini } 2178ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2179a3df083aSStefano Zampini if (need_benign_correction) { 2180df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2181a3df083aSStefano Zampini PetscScalar *marr; 2182a3df083aSStefano Zampini 2183a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 218414393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 218514393ed6SStefano Zampini 218614393ed6SStefano Zampini | 0 0 0 | (V) 218714393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 218814393ed6SStefano Zampini | 0 0 -1 | (p0) 218914393ed6SStefano Zampini 219014393ed6SStefano Zampini */ 2191df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 219214393ed6SStefano Zampini const PetscScalar *vals; 219314393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 219414393ed6SStefano Zampini PetscInt n,j,nz; 219514393ed6SStefano Zampini 2196df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2197df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 219814393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 219914393ed6SStefano Zampini for (j=0;j<n;j++) { 220014393ed6SStefano Zampini PetscScalar val = vals[j]; 220114393ed6SStefano Zampini PetscInt k,col = idxs[j]; 220214393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 220314393ed6SStefano Zampini } 220414393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2205df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 220614393ed6SStefano Zampini } 220772b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 220872b8c272SStefano Zampini } 220972b8c272SStefano Zampini if (F) { 221014393ed6SStefano Zampini /* need to correct the rhs */ 221172b8c272SStefano Zampini if (need_benign_correction) { 221272b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 221372b8c272SStefano Zampini PetscScalar *marr; 221472b8c272SStefano Zampini 221572b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 22165cbda25cSStefano Zampini if (lda_rhs != n_R) { 22175cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 22185cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 22195cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 22205cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 22215cbda25cSStefano Zampini } 22225cbda25cSStefano Zampini } else { 2223a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2224a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 22255cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2226a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2227a3df083aSStefano Zampini } 22285cbda25cSStefano Zampini } 2229a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2230a3df083aSStefano Zampini } 223106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 223214393ed6SStefano Zampini /* need to correct the solution */ 2233a3df083aSStefano Zampini if (need_benign_correction) { 2234df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2235a3df083aSStefano Zampini PetscScalar *marr; 2236a3df083aSStefano Zampini 2237a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 22385cbda25cSStefano Zampini if (lda_rhs != n_R) { 22395cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 22405cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 22415cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 22425cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 22435cbda25cSStefano Zampini } 22445cbda25cSStefano Zampini } else { 2245a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2246a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 22475cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2248a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2249a3df083aSStefano Zampini } 22505cbda25cSStefano Zampini } 2251a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2252a3df083aSStefano Zampini } 225306656605SStefano Zampini } else { 225406656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 225506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2256ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2257ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 225806656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 225906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 226006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 226106656605SStefano Zampini } 226206656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 226306656605SStefano Zampini } 226480677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2265ffd830a3SStefano Zampini /* S_VV and S_CV */ 226606656605SStefano Zampini if (n_constraints) { 226706656605SStefano Zampini Mat B; 226880677318SStefano Zampini 2269ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 227080677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2271ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2272ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 227380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 227480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 227580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 227680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 227780677318SStefano Zampini } 2278ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 227980677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 228080677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2281ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 228280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 228306656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2284ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2285ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 228606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 228706656605SStefano Zampini } 228804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 228904708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2290511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 229104708bb6SStefano Zampini } 2292ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2293ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2294ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2295ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2296ffd830a3SStefano Zampini } 229706656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 229814393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 229914393ed6SStefano Zampini if (need_benign_correction) { 2300df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 230114393ed6SStefano Zampini PetscScalar *marr,*sums; 230214393ed6SStefano Zampini 230314393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 230414393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2305df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 230614393ed6SStefano Zampini const PetscScalar *vals; 230714393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 230814393ed6SStefano Zampini PetscInt n,j,nz; 230914393ed6SStefano Zampini 2310df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2311df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 231214393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 231314393ed6SStefano Zampini PetscInt k; 231414393ed6SStefano Zampini sums[j] = 0.; 231514393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 231614393ed6SStefano Zampini } 231714393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 231814393ed6SStefano Zampini for (j=0;j<n;j++) { 231914393ed6SStefano Zampini PetscScalar val = vals[j]; 232014393ed6SStefano Zampini PetscInt k; 232114393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 232214393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 232314393ed6SStefano Zampini } 232414393ed6SStefano Zampini } 232514393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2326df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 232714393ed6SStefano Zampini } 232814393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 232914393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 233014393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 233114393ed6SStefano Zampini } 233280677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 233306656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 233406656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 233506656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 233606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 233706656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 233806656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 233906656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2340d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2341019a44ceSStefano Zampini } else { 2342d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2343d16cbb6bSStefano Zampini } 234421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2345d16cbb6bSStefano Zampini 234606656605SStefano Zampini /* coarse basis functions */ 234706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 234816f15bc4SStefano Zampini PetscScalar *y; 234916f15bc4SStefano Zampini 2350ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 235106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 235206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 235306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 235406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 235506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 235606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 235706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 235806656605SStefano Zampini 235906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23604f1b2e48SStefano Zampini PetscInt j; 23614f1b2e48SStefano Zampini 236206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 236306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 236406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 236506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 236606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23674f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 236806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 236906656605SStefano Zampini } 237006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 237106656605SStefano Zampini } 237204708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 237304708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 237406656605SStefano Zampini } 23755cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 237606656605SStefano Zampini 237706656605SStefano Zampini if (n_constraints) { 237806656605SStefano Zampini Mat B; 237906656605SStefano Zampini 2380ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 238106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 238280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 238306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 238406656605SStefano Zampini if (n_vertices) { 238580677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 238680677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 238780677318SStefano Zampini } else { 238880677318SStefano Zampini Mat S_VCt; 238980677318SStefano Zampini 2390ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2391ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 239272b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2393ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2394ffd830a3SStefano Zampini } 239580677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 239680677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 239780677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 239880677318SStefano Zampini } 239906656605SStefano Zampini } 240006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 240106656605SStefano Zampini /* coarse basis functions */ 240206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 240306656605SStefano Zampini PetscScalar *y; 240406656605SStefano Zampini 2405ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 240606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 240706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 240806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 240906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 241006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 241106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 241206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 24134f1b2e48SStefano Zampini PetscInt j; 24144f1b2e48SStefano Zampini 241506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 241606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 241706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 241806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 241906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 24204f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 242106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 242206656605SStefano Zampini } 242306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 242406656605SStefano Zampini } 242506656605SStefano Zampini } 242680677318SStefano Zampini if (n_constraints) { 242780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 242880677318SStefano Zampini } 24294f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 243072b8c272SStefano Zampini 243172b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 243272b8c272SStefano Zampini if (pcbddc->benign_n) { 243372b8c272SStefano Zampini Mat B0_B,B0_BPHI; 243472b8c272SStefano Zampini IS is_dummy; 243572b8c272SStefano Zampini PetscScalar *data; 243672b8c272SStefano Zampini PetscInt j; 243772b8c272SStefano Zampini 243872b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 243972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 244072b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 244172b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 244286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 244372b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 244472b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 244572b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 244672b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 244772b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 244872b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 244972b8c272SStefano Zampini } 245072b8c272SStefano Zampini } 245172b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 245272b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 245372b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 245472b8c272SStefano Zampini } 2455019a44ceSStefano Zampini 245606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 24573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2458ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2459ffd830a3SStefano Zampini PetscScalar *marray; 246006656605SStefano Zampini 246106656605SStefano Zampini if (n_constraints) { 2462ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 246306656605SStefano Zampini 2464af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 246506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2466ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 246716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 246806656605SStefano Zampini if (n_vertices) { 2469ffd830a3SStefano Zampini Mat S_VCT; 247006656605SStefano Zampini 247106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2472ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 247316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 247406656605SStefano Zampini } 2475ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 24765b782168SStefano Zampini } else { 24775b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 247806656605SStefano Zampini } 247916f15bc4SStefano Zampini if (n_vertices && n_R) { 2480ffd830a3SStefano Zampini PetscScalar *av,*marray; 2481ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2482ffd830a3SStefano Zampini PetscInt n; 2483ffd830a3SStefano Zampini PetscBool flg_row; 248406656605SStefano Zampini 2485ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2486af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2487ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2488ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2489ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2490ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2491ffd830a3SStefano Zampini PetscInt j; 2492ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2493ffd830a3SStefano Zampini } 2494ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2495ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2496ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 249706656605SStefano Zampini } 249806656605SStefano Zampini 2499ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2500ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2501ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2502ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2503ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 250406656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 250506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 250606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 250706656605SStefano Zampini } 2508ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 25095b782168SStefano Zampini if (B_C) { 2510ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2511ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2512ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2513ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2514ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2515ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2516ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 251706656605SStefano Zampini } 2518ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 25195b782168SStefano Zampini } 252006656605SStefano Zampini /* coarse basis functions */ 252106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 252206656605SStefano Zampini PetscScalar *y; 252306656605SStefano Zampini 2524ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 252506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 252606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 252706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 252906656605SStefano Zampini if (i<n_vertices) { 253006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 253106656605SStefano Zampini } 253206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 253306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 253406656605SStefano Zampini 253506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 253606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 253706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 253806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 253906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 254006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 254106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 254206656605SStefano Zampini } 254306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 254406656605SStefano Zampini } 2545ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2546ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 254706656605SStefano Zampini } 2548d62866d3SStefano Zampini /* free memory */ 254988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 255006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 255106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 255206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 255306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2554d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2555d62866d3SStefano Zampini if (n_vertices) { 2556d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2557d62866d3SStefano Zampini } 2558d62866d3SStefano Zampini if (n_constraints) { 2559d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2560d62866d3SStefano Zampini } 256188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 256288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 256388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2564d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 256588ebb749SStefano Zampini Mat coarse_sub_mat; 256625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 256788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 256888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 256988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 25708bec7fa6SStefano Zampini Mat C_B,CPHI; 25718bec7fa6SStefano Zampini IS is_dummy; 25728bec7fa6SStefano Zampini Vec mones; 257388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 257488ebb749SStefano Zampini PetscReal real_value; 257588ebb749SStefano Zampini 2576a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2577a3df083aSStefano Zampini Mat A; 2578a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2579a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2580a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2581a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2582a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2583a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2584a3df083aSStefano Zampini } else { 258588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 258688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 258788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 258888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2589a3df083aSStefano Zampini } 259088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 259188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2592ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 259388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 259488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 259588ebb749SStefano Zampini } 259688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 259788ebb749SStefano Zampini 259825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 25993301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 260025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2601ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 260288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 260388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 260488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 260588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 260688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 260788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 260888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 260988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 261088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 261188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 261288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 261388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 261488ebb749SStefano Zampini } else { 261588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 261688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 261788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 261888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 261988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 262088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 262188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 262288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 262388ebb749SStefano Zampini } 262488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 262588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 262688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2627511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 26284f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2629fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2630d12edf2fSStefano Zampini PetscScalar *data,*data2; 26314f1b2e48SStefano Zampini PetscInt j; 2632d12edf2fSStefano Zampini 26334f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2634fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2635d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 263686c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2637d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2638d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 26394f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 26404f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2641d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 26424f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 26434f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 26444f1b2e48SStefano Zampini } 2645d12edf2fSStefano Zampini } 2646d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2647d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2648d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2649d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2650d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2651d12edf2fSStefano Zampini } 2652d12edf2fSStefano Zampini #if 0 2653d12edf2fSStefano Zampini { 2654d12edf2fSStefano Zampini PetscViewer viewer; 2655d12edf2fSStefano Zampini char filename[256]; 2656ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2657d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2658d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2659ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2660ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2661ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2662d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 266372b8c272SStefano Zampini if (save_change) { 266472b8c272SStefano Zampini Mat phi_B; 266572b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 266672b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 266772b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 266872b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 266972b8c272SStefano Zampini } else { 2670ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2671ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 267272b8c272SStefano Zampini } 2673ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2674ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2675ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2676ffd830a3SStefano Zampini } 2677ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2678ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2679ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2680ffd830a3SStefano Zampini } 268172b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2682ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2683ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2684ffd830a3SStefano Zampini } 2685d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2686d12edf2fSStefano Zampini } 2687d12edf2fSStefano Zampini #endif 268881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 26898bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 26901575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 269106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 26928bec7fa6SStefano Zampini 26938bec7fa6SStefano Zampini /* check constraints */ 2694a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2695a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 26964f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 26978bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2698a00504b5SStefano Zampini } else { 2699a00504b5SStefano Zampini PetscScalar *data; 2700a00504b5SStefano Zampini Mat tmat; 2701a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2702a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2703a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2704a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2705a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2706a00504b5SStefano Zampini } 27078bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 27088bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 27098bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 27108bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2711bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2712ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2713bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2714bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2715bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2716bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2717bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 271888ebb749SStefano Zampini } 27198bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 27208bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 27218bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 27228bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 272325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 272488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 272588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 272688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 272788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 272888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 272988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 273088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 273188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 273288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 273388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2734ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 273588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 273688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 273788ebb749SStefano Zampini } 273888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 273988ebb749SStefano Zampini } 27408629588bSStefano Zampini /* get back data */ 27418629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 274288ebb749SStefano Zampini PetscFunctionReturn(0); 274388ebb749SStefano Zampini } 274488ebb749SStefano Zampini 274588ebb749SStefano Zampini #undef __FUNCT__ 2746d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2747d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2748aa0d41d4SStefano Zampini { 2749d65f70fdSStefano Zampini Mat *work_mat; 2750d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2751d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2752c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2753aa0d41d4SStefano Zampini PetscErrorCode ierr; 2754aa0d41d4SStefano Zampini 2755aa0d41d4SStefano Zampini PetscFunctionBegin; 2756d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2757d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2758d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2759d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2760aa0d41d4SStefano Zampini 2761d65f70fdSStefano Zampini if (!rsorted) { 2762906d46d4SStefano Zampini const PetscInt *idxs; 2763906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2764aa0d41d4SStefano Zampini 2765d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2766d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2767d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2768d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2769aa0d41d4SStefano Zampini } 2770d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2771d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2772d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2773d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2774aa0d41d4SStefano Zampini } 2775d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2776d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2777d65f70fdSStefano Zampini } else { 2778d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2779d65f70fdSStefano Zampini isrow_s = isrow; 2780aa0d41d4SStefano Zampini } 2781906d46d4SStefano Zampini 2782d65f70fdSStefano Zampini if (!csorted) { 2783d65f70fdSStefano Zampini if (isrow == iscol) { 2784d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2785d65f70fdSStefano Zampini iscol_s = isrow_s; 2786d65f70fdSStefano Zampini } else { 2787d65f70fdSStefano Zampini const PetscInt *idxs; 2788d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2789906d46d4SStefano Zampini 2790d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2791d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2792d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2793d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2794d65f70fdSStefano Zampini } 2795d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2796d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2797d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2798d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2799d65f70fdSStefano Zampini } 2800d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2801d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2802d65f70fdSStefano Zampini } 2803d65f70fdSStefano Zampini } else { 2804d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2805d65f70fdSStefano Zampini iscol_s = iscol; 2806d65f70fdSStefano Zampini } 2807d65f70fdSStefano Zampini 2808d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2809d65f70fdSStefano Zampini 2810d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2811906d46d4SStefano Zampini Mat new_mat; 2812d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2813906d46d4SStefano Zampini 2814d65f70fdSStefano Zampini if (!rsorted) { 2815d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2816d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2817d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2818d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2819906d46d4SStefano Zampini } 2820d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2821d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2822d65f70fdSStefano Zampini } else { 2823d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2824906d46d4SStefano Zampini } 2825d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2826d65f70fdSStefano Zampini 2827d65f70fdSStefano Zampini if (!csorted) { 2828d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2829d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2830d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2831d65f70fdSStefano Zampini } else { 2832d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2833d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2834d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2835d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2836d65f70fdSStefano Zampini } 2837d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2838d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2839d65f70fdSStefano Zampini } 2840d65f70fdSStefano Zampini } else { 2841d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2842d65f70fdSStefano Zampini } 2843d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2844d65f70fdSStefano Zampini 2845d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2846d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2847d65f70fdSStefano Zampini work_mat[0] = new_mat; 2848d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2849d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2850d65f70fdSStefano Zampini } 2851d65f70fdSStefano Zampini 2852d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2853d65f70fdSStefano Zampini *B = work_mat[0]; 2854d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2855d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2856d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2857d65f70fdSStefano Zampini PetscFunctionReturn(0); 2858d65f70fdSStefano Zampini } 2859d65f70fdSStefano Zampini 2860d65f70fdSStefano Zampini #undef __FUNCT__ 28615e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 28625e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2863aa0d41d4SStefano Zampini { 2864aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 28655e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2866d65f70fdSStefano Zampini Mat new_mat; 28675e8657edSStefano Zampini IS is_local,is_global; 2868d65f70fdSStefano Zampini PetscInt local_size; 2869d65f70fdSStefano Zampini PetscBool isseqaij; 2870aa0d41d4SStefano Zampini PetscErrorCode ierr; 2871aa0d41d4SStefano Zampini 2872aa0d41d4SStefano Zampini PetscFunctionBegin; 2873aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 28745e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 28755e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2876b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2877aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2878d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2879aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2880906d46d4SStefano Zampini 2881906d46d4SStefano Zampini /* check */ 2882906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2883906d46d4SStefano Zampini Vec x,x_change; 2884906d46d4SStefano Zampini PetscReal error; 2885906d46d4SStefano Zampini 28865e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2887906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 28885e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2889e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2890e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2891d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2892e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2893e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2894906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2895906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2896906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2897906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2898906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2899906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2900906d46d4SStefano Zampini } 2901906d46d4SStefano Zampini 290222d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 29039b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 290422d5777bSStefano Zampini if (isseqaij) { 2905a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2906a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2907aa0d41d4SStefano Zampini } else { 2908a00504b5SStefano Zampini Mat work_mat; 29091cf9b237SStefano Zampini 2910a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2911aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2912a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 29131d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2914aa0d41d4SStefano Zampini } 29153301b35fSStefano Zampini if (matis->A->symmetric_set) { 29163301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2917e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 29183301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2919e496cd5dSStefano Zampini #endif 29203301b35fSStefano Zampini } 2921d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2922aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2923aa0d41d4SStefano Zampini } 2924aa0d41d4SStefano Zampini 2925aa0d41d4SStefano Zampini #undef __FUNCT__ 2926a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 29278ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2928a64d13efSStefano Zampini { 2929a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2930a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2931d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 293253892102SStefano Zampini PetscInt *idx_R_local=NULL; 29333a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 29343a50541eSStefano Zampini PetscInt vbs,bs; 29356816873aSStefano Zampini PetscBT bitmask=NULL; 2936a64d13efSStefano Zampini PetscErrorCode ierr; 2937a64d13efSStefano Zampini 2938a64d13efSStefano Zampini PetscFunctionBegin; 2939b23d619eSStefano Zampini /* 2940b23d619eSStefano Zampini No need to setup local scatters if 2941b23d619eSStefano Zampini - primal space is unchanged 2942b23d619eSStefano Zampini AND 2943b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2944b23d619eSStefano Zampini AND 2945b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2946b23d619eSStefano Zampini */ 2947b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2948f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2949f4ddd8eeSStefano Zampini } 2950f4ddd8eeSStefano Zampini /* destroy old objects */ 2951f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2952f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2953f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2954a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2955b371cd4fSStefano Zampini n_B = pcis->n_B; 2956b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2957b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 29583a50541eSStefano Zampini 2959a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 29606816873aSStefano Zampini 296153892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2962b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2963854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2964a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2965a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 29660e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2967a64d13efSStefano Zampini } 2968a64d13efSStefano Zampini 2969a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 29704641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 29716816873aSStefano Zampini idx_R_local[n_R++] = i; 2972a64d13efSStefano Zampini } 2973a64d13efSStefano Zampini } 2974df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2975df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 29766816873aSStefano Zampini 2977df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2978df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 29796816873aSStefano Zampini } 29803a50541eSStefano Zampini 29813a50541eSStefano Zampini /* Block code */ 29823a50541eSStefano Zampini vbs = 1; 29833a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 29843a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 29853a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 29863a50541eSStefano Zampini PetscInt *vary; 2987b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2988785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 29893a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2990d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2991d3df7717SStefano 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 */ 29920e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2993d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 29943a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 29953a50541eSStefano Zampini is_blocked = PETSC_FALSE; 29963a50541eSStefano Zampini break; 29973a50541eSStefano Zampini } 29983a50541eSStefano Zampini } 2999d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 3000d3df7717SStefano Zampini } else { 3001d3df7717SStefano Zampini /* Verify directly the R set */ 3002d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 3003d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 3004d3df7717SStefano Zampini for (j=1; j<bs; j++) { 3005d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 3006d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 3007d3df7717SStefano Zampini break; 3008d3df7717SStefano Zampini } 3009d3df7717SStefano Zampini } 3010d3df7717SStefano Zampini } 3011d3df7717SStefano Zampini } 30123a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 30133a50541eSStefano Zampini vbs = bs; 30143a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 30153a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 30163a50541eSStefano Zampini } 30173a50541eSStefano Zampini } 30183a50541eSStefano Zampini } 30193a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 3020b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3021df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 302253892102SStefano Zampini 3023df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3024df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 302553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 3026df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 302753892102SStefano Zampini } else { 30283a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 302953892102SStefano Zampini } 3030a64d13efSStefano Zampini 3031a64d13efSStefano Zampini /* print some info if requested */ 3032a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 3033a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3034a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 30351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3036a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 3037a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 30384f1b2e48SStefano 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); 3039a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3040a64d13efSStefano Zampini } 3041a64d13efSStefano Zampini 3042a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 3043b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 30446816873aSStefano Zampini IS is_aux1,is_aux2; 30456816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 30466816873aSStefano Zampini 30473a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3048854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 3049854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 3050a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 30514641a718SStefano Zampini for (i=0; i<n_D; i++) { 30524641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 30534641a718SStefano Zampini } 3054a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3055a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30564641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 30574641a718SStefano Zampini aux_array1[j++] = i; 3058a64d13efSStefano Zampini } 3059a64d13efSStefano Zampini } 3060a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3061a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3062a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 30634641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 30644641a718SStefano Zampini aux_array2[j++] = i; 3065a64d13efSStefano Zampini } 3066a64d13efSStefano Zampini } 3067a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3068a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 3069a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 3070a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3071a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 3072a64d13efSStefano Zampini 30738eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3074785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 3075a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30764641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 30774641a718SStefano Zampini aux_array1[j++] = i; 3078a64d13efSStefano Zampini } 3079a64d13efSStefano Zampini } 3080a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3081a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 3082a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3083a64d13efSStefano Zampini } 30844641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 30853a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3086d62866d3SStefano Zampini } else { 3087df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 30886816873aSStefano Zampini IS tis; 30896816873aSStefano Zampini PetscInt schur_size; 30906816873aSStefano Zampini 3091df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 30926816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 3093df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 30946816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 30956816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 30966816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 30976816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 30986816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 3099d62866d3SStefano Zampini } 3100d62866d3SStefano Zampini } 3101a64d13efSStefano Zampini PetscFunctionReturn(0); 3102a64d13efSStefano Zampini } 3103a64d13efSStefano Zampini 3104304d26faSStefano Zampini 3105304d26faSStefano Zampini #undef __FUNCT__ 3106304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 3107684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 3108304d26faSStefano Zampini { 3109304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3110304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3111304d26faSStefano Zampini PC pc_temp; 3112304d26faSStefano Zampini Mat A_RR; 3113f4ddd8eeSStefano Zampini MatReuse reuse; 3114304d26faSStefano Zampini PetscScalar m_one = -1.0; 3115304d26faSStefano Zampini PetscReal value; 311604708bb6SStefano Zampini PetscInt n_D,n_R; 3117c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 3118304d26faSStefano Zampini PetscErrorCode ierr; 3119e604994aSStefano Zampini /* prefixes stuff */ 3120312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 3121e604994aSStefano Zampini size_t len; 3122304d26faSStefano Zampini 3123304d26faSStefano Zampini PetscFunctionBegin; 3124304d26faSStefano Zampini 3125e604994aSStefano Zampini /* compute prefixes */ 3126e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 3127e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 3128e604994aSStefano Zampini if (!pcbddc->current_level) { 3129e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3130e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3131e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3132e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3133e604994aSStefano Zampini } else { 3134e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3135312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3136e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3137e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 3138312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3139312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 314034d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 314134d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3142e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3143e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3144e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 3145e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 3146e604994aSStefano Zampini } 3147e604994aSStefano Zampini 3148304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3149684f6988SStefano Zampini if (dirichlet) { 3150d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3151450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 31529a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3153450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3154a3df083aSStefano Zampini Mat A_IIn; 3155a3df083aSStefano Zampini 3156a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3157a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3158a3df083aSStefano Zampini pcis->A_II = A_IIn; 3159a3df083aSStefano Zampini } 3160450f8f5eSStefano Zampini } 31613301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31623301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3163964fefecSStefano Zampini } 3164ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3165964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3166304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3167304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3168304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3169304d26faSStefano Zampini /* default */ 3170304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3171e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 31729577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3173304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 31749577ea80SStefano Zampini if (issbaij) { 31759577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 31769577ea80SStefano Zampini } else { 3177304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31789577ea80SStefano Zampini } 3179304d26faSStefano Zampini /* Allow user's customization */ 3180304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3181304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3182304d26faSStefano Zampini } 3183d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3184b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3185df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3186d62866d3SStefano Zampini 3187df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3188d5574798SStefano Zampini } 3189304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3190304d26faSStefano Zampini if (!n_D) { 3191304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3192304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3193304d26faSStefano Zampini } 3194304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3195304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3196304d26faSStefano Zampini /* set ksp_D into pcis data */ 3197304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3198304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3199304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3200684f6988SStefano Zampini } 3201304d26faSStefano Zampini 3202304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3203684f6988SStefano Zampini A_RR = 0; 3204684f6988SStefano Zampini if (neumann) { 3205d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 320604708bb6SStefano Zampini PetscInt ibs,mbs; 320704708bb6SStefano Zampini PetscBool issbaij; 320804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3209f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 32108ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3211f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3212f4ddd8eeSStefano Zampini PetscInt nn_R; 321381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3214f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3215f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3216f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3217f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3218f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3219f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3220f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3221727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3222f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3223f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3224f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3225f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3226f4ddd8eeSStefano Zampini } 3227f4ddd8eeSStefano Zampini } 3228f4ddd8eeSStefano Zampini /* last check */ 3229d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3230f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3231f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3232f4ddd8eeSStefano Zampini } 3233f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3234f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3235f4ddd8eeSStefano Zampini } 3236a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3237af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3238af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 323904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 324004708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 324104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 324204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 324304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3244af732b37SStefano Zampini } else { 3245511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 32466816873aSStefano Zampini } 324704708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 324804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 324904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 325004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 325104708bb6SStefano Zampini } else { 3252511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 325304708bb6SStefano Zampini } 325404708bb6SStefano Zampini } 3255a00504b5SStefano Zampini /* extract A_RR */ 3256b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3257a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3258a00504b5SStefano Zampini 3259a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 326016e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3261a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 326216e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 326316e386b8SStefano Zampini } else { 3264a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3265a00504b5SStefano Zampini } 3266a00504b5SStefano Zampini } else { 3267a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3268a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3269a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3270a00504b5SStefano Zampini } 3271a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3272f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 327316e386b8SStefano Zampini } 32743301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 32753301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 32766816873aSStefano Zampini } 3277f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3278304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3279304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3280304d26faSStefano Zampini /* default */ 3281304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3282e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3283304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 32849577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 32859577ea80SStefano Zampini if (issbaij) { 32869577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 32879577ea80SStefano Zampini } else { 3288304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 32899577ea80SStefano Zampini } 3290304d26faSStefano Zampini /* Allow user's customization */ 3291304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3292304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3293304d26faSStefano Zampini } 3294304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3295304d26faSStefano Zampini if (!n_R) { 3296304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3297304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3298304d26faSStefano Zampini } 32995cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3300df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3301b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3302df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3303d62866d3SStefano Zampini 3304df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3305d62866d3SStefano Zampini } 3306304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3307304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3308684f6988SStefano Zampini } 3309304d26faSStefano Zampini 3310684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3311684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33121575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3313684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3314684f6988SStefano Zampini } 3315c7017625SStefano Zampini 3316c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3317c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3318c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3319c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3320c7017625SStefano Zampini } 3321c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3322c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3323c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3324c7017625SStefano Zampini } 3325c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3326c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3327c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3328c7017625SStefano Zampini } 3329c7017625SStefano Zampini 3330c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3331c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3332684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 33330fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 33340fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 33350fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 33360fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 33370fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3338e604994aSStefano 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); 3339c7017625SStefano Zampini if (check_corr[0]) { 3340c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3341c7017625SStefano Zampini } 3342304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3343304d26faSStefano Zampini } 3344684f6988SStefano Zampini if (neumann) { /* Neumann */ 33450fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 33460fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 33470fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 33480fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 33490fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3350e604994aSStefano 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); 3351c7017625SStefano Zampini if (check_corr[1]) { 3352c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3353c7017625SStefano Zampini } 3354304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3355304d26faSStefano Zampini } 3356684f6988SStefano Zampini } 33575cbda25cSStefano Zampini /* free Neumann problem's matrix */ 33585cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3359304d26faSStefano Zampini PetscFunctionReturn(0); 3360304d26faSStefano Zampini } 3361304d26faSStefano Zampini 3362304d26faSStefano Zampini #undef __FUNCT__ 3363ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 336480677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3365674ae819SStefano Zampini { 3366674ae819SStefano Zampini PetscErrorCode ierr; 3367674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3368be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3369b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 3370674ae819SStefano Zampini 3371674ae819SStefano Zampini PetscFunctionBegin; 3372b334f244SStefano Zampini if (!reuse_solver) { 337380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 337420c7b377SStefano Zampini } 337580677318SStefano Zampini if (!pcbddc->switch_static) { 337680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 337780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 337880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 337920c7b377SStefano Zampini } 3380b334f244SStefano Zampini if (!reuse_solver) { 338180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 338280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 338320c7b377SStefano Zampini } else { 3384df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3385be83ff47SStefano Zampini 3386df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3387df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 338820c7b377SStefano Zampini } 3389be83ff47SStefano Zampini } else { 339080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 339180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 339280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 339380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 339480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 339580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 339680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 339780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 339880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3399674ae819SStefano Zampini } 3400674ae819SStefano Zampini } 3401b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 340280677318SStefano Zampini if (applytranspose) { 340380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 340480677318SStefano Zampini } else { 340580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 340680677318SStefano Zampini } 3407be83ff47SStefano Zampini } else { 3408df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3409be83ff47SStefano Zampini 3410be83ff47SStefano Zampini if (applytranspose) { 3411df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3412be83ff47SStefano Zampini } else { 3413df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3414be83ff47SStefano Zampini } 3415be83ff47SStefano Zampini } 341680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 341780677318SStefano Zampini if (!pcbddc->switch_static) { 3418b334f244SStefano Zampini if (!reuse_solver) { 341980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3421be83ff47SStefano Zampini } else { 3422df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3423be83ff47SStefano Zampini 3424df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3425df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3426be83ff47SStefano Zampini } 342780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 342880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 342980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 343080677318SStefano Zampini } 343180677318SStefano Zampini } else { 343280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 343380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 343480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 343580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 343680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 343780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 343880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 343980677318SStefano Zampini } 344080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 344180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 344280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 344380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3444674ae819SStefano Zampini } 3445674ae819SStefano Zampini PetscFunctionReturn(0); 3446674ae819SStefano Zampini } 3447674ae819SStefano Zampini 3448dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3449674ae819SStefano Zampini #undef __FUNCT__ 3450674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3451dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3452674ae819SStefano Zampini { 3453674ae819SStefano Zampini PetscErrorCode ierr; 3454674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3455674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3456674ae819SStefano Zampini const PetscScalar zero = 0.0; 3457674ae819SStefano Zampini 3458674ae819SStefano Zampini PetscFunctionBegin; 3459dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 34604fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3461dc359a40SStefano Zampini if (applytranspose) { 3462674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 34638eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3464dc359a40SStefano Zampini } else { 3465674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3466674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 346715aaf578SStefano Zampini } 34684fee134fSStefano Zampini } else { 34694fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 34704fee134fSStefano Zampini } 3471efc2fbd9SStefano Zampini 3472efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 34734f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3474efc2fbd9SStefano Zampini PetscScalar *array; 34754f1b2e48SStefano Zampini PetscInt j; 3476efc2fbd9SStefano Zampini 3477efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34784f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3479efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3480efc2fbd9SStefano Zampini } 3481efc2fbd9SStefano Zampini 348212edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 348312edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 348412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 348512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 348612edc857SStefano Zampini 34879f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 348812edc857SStefano Zampini if (pcbddc->coarse_ksp) { 348951694757SStefano Zampini Mat coarse_mat; 3490964fefecSStefano Zampini Vec rhs,sol; 349151694757SStefano Zampini MatNullSpace nullsp; 349227b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3493964fefecSStefano Zampini 349427b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 349527b6a85dSStefano Zampini PC coarse_pc; 349627b6a85dSStefano Zampini 349727b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 349827b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 349927b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 350027b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 350127b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 350227b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 35033bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 350427b6a85dSStefano Zampini } 350527b6a85dSStefano Zampini } 3506964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3507964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 350851694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 350951694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 351051694757SStefano Zampini if (nullsp) { 351151694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 351251694757SStefano Zampini } 351312edc857SStefano Zampini if (applytranspose) { 35149a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 3515964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 35162701bc32SStefano Zampini } else { 35171f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 35182701bc32SStefano Zampini PC coarse_pc; 35192701bc32SStefano Zampini 35202701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 35212701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 35223e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 35232701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 352412edc857SStefano Zampini } else { 3525964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 352612edc857SStefano Zampini } 35272701bc32SStefano Zampini } 35281d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 352927b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 353027b6a85dSStefano Zampini PC coarse_pc; 353127b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 353227b6a85dSStefano Zampini 353327b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 353427b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 353527b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 35363bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 353727b6a85dSStefano Zampini } 353851694757SStefano Zampini if (nullsp) { 353951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 354051694757SStefano Zampini } 354112edc857SStefano Zampini } 3542674ae819SStefano Zampini 3543674ae819SStefano Zampini /* Local solution on R nodes */ 35444fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 354580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 35469f00e9b4SStefano Zampini } 35479f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 35489f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 354912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3550674ae819SStefano Zampini 35514fee134fSStefano Zampini /* Sum contributions from the two levels */ 35524fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3553dc359a40SStefano Zampini if (applytranspose) { 3554dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3555dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3556dc359a40SStefano Zampini } else { 3557674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 35588eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3559dc359a40SStefano Zampini } 3560efc2fbd9SStefano Zampini /* store p0 */ 35614f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3562efc2fbd9SStefano Zampini PetscScalar *array; 35634f1b2e48SStefano Zampini PetscInt j; 3564efc2fbd9SStefano Zampini 3565efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 35664f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3567efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3568efc2fbd9SStefano Zampini } 35694fee134fSStefano Zampini } else { /* expand the coarse solution */ 35704fee134fSStefano Zampini if (applytranspose) { 35714fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35724fee134fSStefano Zampini } else { 35734fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35744fee134fSStefano Zampini } 35754fee134fSStefano Zampini } 3576674ae819SStefano Zampini PetscFunctionReturn(0); 3577674ae819SStefano Zampini } 3578674ae819SStefano Zampini 3579674ae819SStefano Zampini #undef __FUNCT__ 3580674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 358112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3582674ae819SStefano Zampini { 3583674ae819SStefano Zampini PetscErrorCode ierr; 3584674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 358558da7f69SStefano Zampini PetscScalar *array; 358612edc857SStefano Zampini Vec from,to; 3587674ae819SStefano Zampini 3588674ae819SStefano Zampini PetscFunctionBegin; 358912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 359012edc857SStefano Zampini from = pcbddc->coarse_vec; 359112edc857SStefano Zampini to = pcbddc->vec1_P; 359212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 359312edc857SStefano Zampini Vec tvec; 359458da7f69SStefano Zampini 359558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 359658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 359712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 359858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 359958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 360058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 360112edc857SStefano Zampini } 360212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 360312edc857SStefano Zampini from = pcbddc->vec1_P; 360412edc857SStefano Zampini to = pcbddc->coarse_vec; 360512edc857SStefano Zampini } 360612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3607674ae819SStefano Zampini PetscFunctionReturn(0); 3608674ae819SStefano Zampini } 3609674ae819SStefano Zampini 3610674ae819SStefano Zampini #undef __FUNCT__ 3611674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 361212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3613674ae819SStefano Zampini { 3614674ae819SStefano Zampini PetscErrorCode ierr; 3615674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 361658da7f69SStefano Zampini PetscScalar *array; 361712edc857SStefano Zampini Vec from,to; 3618674ae819SStefano Zampini 3619674ae819SStefano Zampini PetscFunctionBegin; 362012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 362112edc857SStefano Zampini from = pcbddc->coarse_vec; 362212edc857SStefano Zampini to = pcbddc->vec1_P; 362312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 362412edc857SStefano Zampini from = pcbddc->vec1_P; 362512edc857SStefano Zampini to = pcbddc->coarse_vec; 362612edc857SStefano Zampini } 362712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 362812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 362912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 363012edc857SStefano Zampini Vec tvec; 363158da7f69SStefano Zampini 363212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 363358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 363458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 363558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 363658da7f69SStefano Zampini } 363758da7f69SStefano Zampini } else { 363858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 363958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 364012edc857SStefano Zampini } 364112edc857SStefano Zampini } 3642674ae819SStefano Zampini PetscFunctionReturn(0); 3643674ae819SStefano Zampini } 3644674ae819SStefano Zampini 3645984c4197SStefano Zampini /* uncomment for testing purposes */ 3646984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3647674ae819SStefano Zampini #undef __FUNCT__ 3648674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3649674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3650674ae819SStefano Zampini { 3651674ae819SStefano Zampini PetscErrorCode ierr; 3652674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3653674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3654674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3655984c4197SStefano Zampini /* one and zero */ 3656984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3657984c4197SStefano Zampini /* space to store constraints and their local indices */ 36589162d606SStefano Zampini PetscScalar *constraints_data; 36599162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 36609162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 36619162d606SStefano Zampini PetscInt *constraints_n; 3662984c4197SStefano Zampini /* iterators */ 3663b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3664984c4197SStefano Zampini /* BLAS integers */ 3665e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3666e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3667c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3668727cdba6SStefano Zampini /* reuse */ 36690e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 36700e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3671984c4197SStefano Zampini /* change of basis */ 3672b3d85658SStefano Zampini PetscBool qr_needed; 36739162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3674984c4197SStefano Zampini /* auxiliary stuff */ 367564efe560SStefano Zampini PetscInt *nnz,*is_indices; 36768a0068c3SStefano Zampini PetscInt ncc; 3677984c4197SStefano Zampini /* some quantities */ 367845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3679a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3680984c4197SStefano Zampini 3681674ae819SStefano Zampini PetscFunctionBegin; 36828e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 36838e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 36848e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 368516909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3686088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3687088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 36880e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 36890e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 36900e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36910e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36920e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3693088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3694cf5a6209SStefano Zampini 3695cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 36969162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3697cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3698cf5a6209SStefano Zampini const Vec *nearnullvecs; 3699cf5a6209SStefano Zampini Vec *localnearnullsp; 3700cf5a6209SStefano Zampini PetscScalar *array; 3701cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3702cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3703674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3704b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3705674ae819SStefano Zampini PetscScalar *work; 3706674ae819SStefano Zampini PetscReal *singular_vals; 3707674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3708674ae819SStefano Zampini PetscReal *rwork; 3709674ae819SStefano Zampini #endif 3710674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3711674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3712674ae819SStefano Zampini #else 3713964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3714964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3715674ae819SStefano Zampini #endif 3716674ae819SStefano Zampini 3717674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3718d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3719e4d548c7SStefano Zampini /* print some info */ 37201f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3721e4d548c7SStefano Zampini PetscInt nv; 3722e4d548c7SStefano Zampini 3723c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 3724e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3725e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3726e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3727e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3728e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3729e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3730e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3731e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3732e4d548c7SStefano Zampini } 3733e4d548c7SStefano Zampini 3734d06fc5fdSStefano Zampini /* free unneeded index sets */ 3735d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3736d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3737674ae819SStefano Zampini } 3738d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3739d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3740d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3741d06fc5fdSStefano Zampini } 3742d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3743d06fc5fdSStefano Zampini n_ISForEdges = 0; 3744d06fc5fdSStefano Zampini } 3745d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3746d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3747d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3748d06fc5fdSStefano Zampini } 3749d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3750d06fc5fdSStefano Zampini n_ISForFaces = 0; 3751d06fc5fdSStefano Zampini } 375270022509SStefano Zampini 3753674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3754674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3755674ae819SStefano Zampini if (nearnullsp) { 3756674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3757f4ddd8eeSStefano Zampini /* remove any stored info */ 3758f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3759f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3760f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3761f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3762f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3763473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3764f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3765f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3766f4ddd8eeSStefano Zampini } 3767984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3768984c4197SStefano Zampini nnsp_size = 0; 3769674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3770674ae819SStefano Zampini } 3771984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3772984c4197SStefano Zampini max_constraints = nnsp_size; 3773984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3774984c4197SStefano Zampini 3775674ae819SStefano Zampini /* 3776674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 37779162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 37789162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 37799162d606SStefano Zampini There can be multiple constraints per connected component 3780674ae819SStefano Zampini */ 3781674ae819SStefano Zampini n_vertices = 0; 3782674ae819SStefano Zampini if (ISForVertices) { 3783674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3784674ae819SStefano Zampini } 37859162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 37869162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 37879162d606SStefano Zampini 37889162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 37899162d606SStefano Zampini total_counts *= max_constraints; 3790674ae819SStefano Zampini total_counts += n_vertices; 37914641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 37929162d606SStefano Zampini 3793674ae819SStefano Zampini total_counts = 0; 3794674ae819SStefano Zampini max_size_of_constraint = 0; 3795674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 37969162d606SStefano Zampini IS used_is; 3797674ae819SStefano Zampini if (i<n_ISForEdges) { 37989162d606SStefano Zampini used_is = ISForEdges[i]; 3799674ae819SStefano Zampini } else { 38009162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3801674ae819SStefano Zampini } 38029162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3803674ae819SStefano Zampini total_counts += j; 3804674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3805674ae819SStefano Zampini } 38069162d606SStefano 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); 38079162d606SStefano Zampini 3808984c4197SStefano Zampini /* get local part of global near null space vectors */ 3809785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3810984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3811984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3812e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3813e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3814984c4197SStefano Zampini } 3815674ae819SStefano Zampini 3816242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3817242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3818a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3819242a89d7SStefano Zampini 3820984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3821a773dcb8SStefano Zampini if (!skip_lapack) { 3822674ae819SStefano Zampini PetscScalar temp_work; 3823911cabfeSStefano Zampini 3824674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3825984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3826785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3827785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3828785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3829674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3830785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3831674ae819SStefano Zampini #endif 3832674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3833c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3834c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3835674ae819SStefano Zampini lwork = -1; 3836674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3837674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3838c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3839674ae819SStefano Zampini #else 3840c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3841674ae819SStefano Zampini #endif 3842674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3843984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3844674ae819SStefano Zampini #else /* on missing GESVD */ 3845674ae819SStefano Zampini /* SVD */ 3846674ae819SStefano Zampini PetscInt max_n,min_n; 3847674ae819SStefano Zampini max_n = max_size_of_constraint; 3848984c4197SStefano Zampini min_n = max_constraints; 3849984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3850674ae819SStefano Zampini min_n = max_size_of_constraint; 3851984c4197SStefano Zampini max_n = max_constraints; 3852674ae819SStefano Zampini } 3853785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3854674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3855785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3856674ae819SStefano Zampini #endif 3857674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3858674ae819SStefano Zampini lwork = -1; 3859e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3860e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3861b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3862674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3863674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 38649162d606SStefano 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)); 3865674ae819SStefano Zampini #else 38669162d606SStefano 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)); 3867674ae819SStefano Zampini #endif 3868674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3869984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3870984c4197SStefano Zampini #endif /* on missing GESVD */ 3871674ae819SStefano Zampini /* Allocate optimal workspace */ 3872674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3873854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3874674ae819SStefano Zampini } 3875674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3876674ae819SStefano Zampini total_counts = 0; 38779162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 38789162d606SStefano Zampini constraints_data_ptr[0] = 0; 3879674ae819SStefano Zampini /* vertices */ 38809162d606SStefano Zampini if (n_vertices) { 3881674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 38829162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3883674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 38849162d606SStefano Zampini constraints_n[total_counts] = 1; 38859162d606SStefano Zampini constraints_data[total_counts] = 1.0; 38869162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 38879162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3888674ae819SStefano Zampini total_counts++; 3889674ae819SStefano Zampini } 3890674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3891674ae819SStefano Zampini n_vertices = total_counts; 3892674ae819SStefano Zampini } 3893984c4197SStefano Zampini 3894674ae819SStefano Zampini /* edges and faces */ 38959162d606SStefano Zampini total_counts_cc = total_counts; 3896911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 38979162d606SStefano Zampini IS used_is; 38989162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 38999162d606SStefano Zampini 3900911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 39019162d606SStefano Zampini used_is = ISForEdges[ncc]; 3902984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3903674ae819SStefano Zampini } else { 39049162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3905984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3906674ae819SStefano Zampini } 3907674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 39089162d606SStefano Zampini 39099162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 39109162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3911984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3912984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3913674ae819SStefano Zampini if (nnsp_has_cnst) { 39145b08dc53SStefano Zampini PetscScalar quad_value; 39159162d606SStefano Zampini 39169162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 39179162d606SStefano Zampini idxs_copied = PETSC_TRUE; 39189162d606SStefano Zampini 3919a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3920674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3921a773dcb8SStefano Zampini } else { 3922a773dcb8SStefano Zampini quad_value = 1.0; 3923a773dcb8SStefano Zampini } 3924674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 39259162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3926674ae819SStefano Zampini } 39279162d606SStefano Zampini temp_constraints++; 3928674ae819SStefano Zampini total_counts++; 3929674ae819SStefano Zampini } 3930674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3931984c4197SStefano Zampini PetscReal real_value; 39329162d606SStefano Zampini PetscScalar *ptr_to_data; 39339162d606SStefano Zampini 3934984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 39359162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3936674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 39379162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3938674ae819SStefano Zampini } 3939984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3940984c4197SStefano Zampini /* check if array is null on the connected component */ 3941e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39429162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 39435b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3944674ae819SStefano Zampini temp_constraints++; 3945674ae819SStefano Zampini total_counts++; 39469162d606SStefano Zampini if (!idxs_copied) { 39479162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 39489162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3949674ae819SStefano Zampini } 3950674ae819SStefano Zampini } 39519162d606SStefano Zampini } 39529162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 395345a1bb75SStefano Zampini valid_constraints = temp_constraints; 3954eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3955a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 39569162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 39579162d606SStefano Zampini 39589162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3959a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39609162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3961a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 39629162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3963a773dcb8SStefano Zampini } else { /* perform SVD */ 3964984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 39659162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3966674ae819SStefano Zampini 3967674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3968984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3969984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3970984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3971984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3972984c4197SStefano Zampini from that computed using LAPACKgesvd 3973984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3974984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3975984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3976674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3977e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3978984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3979674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3980674ae819SStefano Zampini for (k=0;k<j+1;k++) { 39819162d606SStefano 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)); 3982674ae819SStefano Zampini } 3983674ae819SStefano Zampini } 3984e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3985e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3986e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3987674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3988c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3989674ae819SStefano Zampini #else 3990c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3991674ae819SStefano Zampini #endif 3992674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3993984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3994984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3995674ae819SStefano Zampini j = 0; 3996984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3997674ae819SStefano Zampini total_counts = total_counts-j; 399845a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3999e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 4000c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 4001c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 4002c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 4003c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4004c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 4005c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4006674ae819SStefano Zampini if (j<temp_constraints) { 4007984c4197SStefano Zampini PetscInt ii; 4008984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 4009674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40109162d606SStefano 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)); 4011674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4012984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 4013674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 40149162d606SStefano 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]; 4015674ae819SStefano Zampini } 4016674ae819SStefano Zampini } 4017674ae819SStefano Zampini } 4018674ae819SStefano Zampini #else /* on missing GESVD */ 4019e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 4020e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 4021b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4022674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4023674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 40249162d606SStefano 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)); 4025674ae819SStefano Zampini #else 40269162d606SStefano 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)); 4027674ae819SStefano Zampini #endif 4028984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 4029674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4030984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 4031e310c8b4SStefano Zampini k = temp_constraints; 4032e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 4033674ae819SStefano Zampini j = 0; 4034e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 403545a1bb75SStefano Zampini valid_constraints = k-j; 4036911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 4037984c4197SStefano Zampini #endif /* on missing GESVD */ 4038674ae819SStefano Zampini } 4039a773dcb8SStefano Zampini } 40409162d606SStefano Zampini /* update pointers information */ 40419162d606SStefano Zampini if (valid_constraints) { 40429162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 40439162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 40449162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 40459162d606SStefano Zampini /* set change_of_basis flag */ 404645a1bb75SStefano Zampini if (boolforchange) { 4047b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 40489162d606SStefano Zampini } 4049b3d85658SStefano Zampini total_counts_cc++; 405045a1bb75SStefano Zampini } 405145a1bb75SStefano Zampini } 4052984c4197SStefano Zampini /* free workspace */ 40538f1c130eSStefano Zampini if (!skip_lapack) { 4054984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4055984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4056984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 4057984c4197SStefano Zampini #endif 4058984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 4059984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4060984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 4061984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 4062984c4197SStefano Zampini #endif 4063984c4197SStefano Zampini } 4064984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4065984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 4066984c4197SStefano Zampini } 4067984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 4068cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 4069cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4070cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4071cf5a6209SStefano Zampini } 4072cf5a6209SStefano Zampini if (n_ISForFaces) { 4073cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4074cf5a6209SStefano Zampini } 4075cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4076cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4077cf5a6209SStefano Zampini } 4078cf5a6209SStefano Zampini if (n_ISForEdges) { 4079cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4080cf5a6209SStefano Zampini } 4081cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 408208122e43SStefano Zampini } else { 408308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4084984c4197SStefano Zampini 408508122e43SStefano Zampini total_counts = 0; 408608122e43SStefano Zampini n_vertices = 0; 4087d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 4088d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 408908122e43SStefano Zampini } 409008122e43SStefano Zampini max_constraints = 0; 40919162d606SStefano Zampini total_counts_cc = 0; 409208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 409308122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 40949162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 409508122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 409608122e43SStefano Zampini } 40979162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 40989162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 40999162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 41009162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 410174d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 41029162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 41039162d606SStefano Zampini total_counts_cc = 0; 41049162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 41059162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 41069162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 410708122e43SStefano Zampini } 410808122e43SStefano Zampini } 41099162d606SStefano Zampini #if 0 41109162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 41119162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 41129162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 41139162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 41149162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 41159162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 41169162d606SStefano Zampini } 41179162d606SStefano Zampini printf("\n"); 41189162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 41199162d606SStefano Zampini } 41201b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 41218bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 41221b968477SStefano Zampini } 41231b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 41248bec7fa6SStefano 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]); 41251b968477SStefano Zampini } 412608122e43SStefano Zampini #endif 412708122e43SStefano Zampini 41288bec7fa6SStefano Zampini max_size_of_constraint = 0; 41299162d606SStefano 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]); 41309162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 413108122e43SStefano Zampini /* Change of basis */ 4132b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 413308122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 413408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 413508122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 4136b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 413708122e43SStefano Zampini } 413808122e43SStefano Zampini } 413908122e43SStefano Zampini } 414008122e43SStefano Zampini } 4141984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 41424f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 414308122e43SStefano Zampini 41449162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 41459162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 41466c4ed002SBarry 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); 4147674ae819SStefano Zampini 4148674ae819SStefano Zampini /* Create constraint matrix */ 4149674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 415016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4151984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4152984c4197SStefano Zampini 4153984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4154a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4155a717540cSStefano Zampini qr_needed = PETSC_FALSE; 415674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4157984c4197SStefano Zampini total_primal_vertices=0; 4158b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 41599162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 41609162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 416172b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 41629162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4163b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 416464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 41659162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41669162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4167a717540cSStefano Zampini } 4168b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 416991af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4170a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4171a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4172a717540cSStefano Zampini } 4173fa434743SStefano Zampini } else { 4174b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4175fa434743SStefano Zampini } 4176a717540cSStefano Zampini } 4177b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4178b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4179674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 418070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4181b3d85658SStefano Zampini 41824f1b2e48SStefano 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); 41830e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 41840e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4185984c4197SStefano Zampini 4186984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 418774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4188785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4189984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 419074d5cdf7SStefano Zampini 4191984c4197SStefano Zampini j = total_primal_vertices; 419274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4193b3d85658SStefano Zampini cum = total_primal_vertices; 41949162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41954641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4196b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4197b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4198b3d85658SStefano Zampini cum++; 41999162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 420074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 420174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 420274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 420374d5cdf7SStefano Zampini } 42049162d606SStefano Zampini j += constraints_n[i]; 4205674ae819SStefano Zampini } 4206674ae819SStefano Zampini } 4207674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4208674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4209088faed8SStefano Zampini 4210674ae819SStefano Zampini /* set values in constraint matrix */ 4211984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 42120e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4213674ae819SStefano Zampini } 4214984c4197SStefano Zampini total_counts = total_primal_vertices; 42159162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 42164641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 42179162d606SStefano Zampini PetscInt *cols; 42189162d606SStefano Zampini 42199162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 42209162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 42219162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 42229162d606SStefano Zampini PetscInt row = total_counts+k; 42239162d606SStefano Zampini PetscScalar *vals; 42249162d606SStefano Zampini 42259162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 42269162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 42279162d606SStefano Zampini } 42289162d606SStefano Zampini total_counts += constraints_n[i]; 4229674ae819SStefano Zampini } 4230674ae819SStefano Zampini } 4231674ae819SStefano Zampini /* assembling */ 4232674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4233674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4234088faed8SStefano Zampini 4235984c4197SStefano Zampini /* 42366a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4237984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4238f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4239984c4197SStefano Zampini */ 4240674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4241674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4242026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4243984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4244984c4197SStefano Zampini /* working stuff for GEQRF */ 424581d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4246984c4197SStefano Zampini PetscBLASInt lqr_work; 4247984c4197SStefano Zampini /* working stuff for UNGQR */ 4248984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4249984c4197SStefano Zampini PetscBLASInt lgqr_work; 4250984c4197SStefano Zampini /* working stuff for TRTRS */ 4251984c4197SStefano Zampini PetscScalar *trs_rhs; 42523f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4253984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4254984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4255984c4197SStefano Zampini PetscScalar *start_vals; 4256984c4197SStefano Zampini /* working stuff for values insertion */ 42574641a718SStefano Zampini PetscBT is_primal; 425864efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4259906d46d4SStefano Zampini /* matrix sizes */ 4260906d46d4SStefano Zampini PetscInt global_size,local_size; 4261906d46d4SStefano Zampini /* temporary change of basis */ 4262906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4263cf5a6209SStefano Zampini /* extra space for debugging */ 4264cf5a6209SStefano Zampini PetscScalar *dbg_work; 4265984c4197SStefano Zampini 4266906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4267906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 426816f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4269bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4270906d46d4SStefano Zampini /* nonzeros for local mat */ 4271bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 42721dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4273bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 42741dd7afcfSStefano Zampini } else { 42751dd7afcfSStefano Zampini const PetscInt *ii; 42761dd7afcfSStefano Zampini PetscInt n; 42771dd7afcfSStefano Zampini PetscBool flg_row; 42781dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42791dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 42801dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42811dd7afcfSStefano Zampini } 42829162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4283a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 42849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4285a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 42869162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4287a717540cSStefano Zampini } else { 42889162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 42899162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4290a717540cSStefano Zampini } 4291a717540cSStefano Zampini } 4292a717540cSStefano Zampini } 4293906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4294bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 42951dd7afcfSStefano Zampini /* Set interior change in the matrix */ 42961dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4297bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4298906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4299a717540cSStefano Zampini } 43001dd7afcfSStefano Zampini } else { 43011dd7afcfSStefano Zampini const PetscInt *ii,*jj; 43021dd7afcfSStefano Zampini PetscScalar *aa; 43031dd7afcfSStefano Zampini PetscInt n; 43041dd7afcfSStefano Zampini PetscBool flg_row; 43051dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 43061dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 43071dd7afcfSStefano Zampini for (i=0;i<n;i++) { 43081dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 43091dd7afcfSStefano Zampini } 43101dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 43111dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 43121dd7afcfSStefano Zampini } 4313a717540cSStefano Zampini 4314a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4315a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4316a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4317a717540cSStefano Zampini } 4318a717540cSStefano Zampini 4319a717540cSStefano Zampini 4320a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4321a717540cSStefano Zampini /* 4322a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4323a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4324a717540cSStefano Zampini 4325a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4326a717540cSStefano Zampini 4327a6b551f4SStefano 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) 4328a6b551f4SStefano Zampini 4329a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4330a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4331a717540cSStefano Zampini | ... | 4332a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4333a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4334a717540cSStefano Zampini 4335a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4336a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4337a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4338a6b551f4SStefano Zampini 4339a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4340a717540cSStefano Zampini */ 4341a717540cSStefano Zampini if (qr_needed) { 4342984c4197SStefano Zampini /* space to store Q */ 4343854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4344984c4197SStefano Zampini /* first we issue queries for optimal work */ 43453f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43463f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 43473f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4348984c4197SStefano Zampini lqr_work = -1; 43493f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4350984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4351984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4352785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4353984c4197SStefano Zampini lgqr_work = -1; 43543f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43553f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 43563f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 43573f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43583f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 43593f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4360984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4361984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4362785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4363984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4364785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4365984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4366785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4367a717540cSStefano Zampini /* allocating workspace for check */ 4368a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4369cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4370a717540cSStefano Zampini } 4371a717540cSStefano Zampini } 4372984c4197SStefano Zampini /* array to store whether a node is primal or not */ 43734641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4374473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 43750e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 43766c4ed002SBarry 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); 437739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 437839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 437939e2fb2aSStefano Zampini } 438039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4381984c4197SStefano Zampini 4382a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 43839162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 43849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 43854641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4386984c4197SStefano Zampini /* get constraint info */ 43879162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4388984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4389984c4197SStefano Zampini 4390984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43919162d606SStefano 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); 4392674ae819SStefano Zampini } 4393984c4197SStefano Zampini 4394fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4395a717540cSStefano Zampini 4396a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4397a717540cSStefano Zampini if (pcbddc->dbg_flag) { 43989162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4399a717540cSStefano Zampini } 4400984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 44019162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4402984c4197SStefano Zampini 4403984c4197SStefano Zampini /* compute QR decomposition of constraints */ 44043f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44053f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 44063f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4407674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44083f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4409984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4410674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4411984c4197SStefano Zampini 4412984c4197SStefano Zampini /* explictly compute R^-T */ 4413984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4414984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 44153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 44163f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 44173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 44183f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4419984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4421984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4422984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4423984c4197SStefano Zampini 4424a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 44253f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44263f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44273f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 44283f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4429984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44303f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4431984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4432984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4433984c4197SStefano Zampini 4434984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4435984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4436984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 44373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44383f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 44393f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 44403f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 44413f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 44423f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4443984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44449162d606SStefano 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)); 4445984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 44469162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4447984c4197SStefano Zampini 4448984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 44499162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4450984c4197SStefano Zampini /* insert cols for primal dofs */ 4451984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4452984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 44539162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4454906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4455984c4197SStefano Zampini } 4456984c4197SStefano Zampini /* insert cols for dual dofs */ 4457984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 44589162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4459984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 44609162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4461906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4462984c4197SStefano Zampini j++; 4463674ae819SStefano Zampini } 4464674ae819SStefano Zampini } 4465984c4197SStefano Zampini 4466984c4197SStefano Zampini /* check change of basis */ 4467984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4468984c4197SStefano Zampini PetscInt ii,jj; 4469984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4470c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4471c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4472c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4473c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4474c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4475c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4476984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4477cf5a6209SStefano 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)); 4478984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4479984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4480984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4481cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4482cf5a6209SStefano 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; 4483674ae819SStefano Zampini } 4484674ae819SStefano Zampini } 4485984c4197SStefano Zampini if (!valid_qr) { 448622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4487984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4488984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4489cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4490cf5a6209SStefano 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])); 4491674ae819SStefano Zampini } 4492cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4493cf5a6209SStefano 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])); 4494984c4197SStefano Zampini } 4495984c4197SStefano Zampini } 4496984c4197SStefano Zampini } 4497674ae819SStefano Zampini } else { 449822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4499674ae819SStefano Zampini } 4500674ae819SStefano Zampini } 4501a717540cSStefano Zampini } else { /* simple transformation block */ 4502a717540cSStefano Zampini PetscInt row,col; 4503a6b551f4SStefano Zampini PetscScalar val,norm; 4504a6b551f4SStefano Zampini 4505a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 45069162d606SStefano 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)); 4507a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 45089162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 45099162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4510bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 45119162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4512906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 45139162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4514a717540cSStefano Zampini } else { 4515a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 45169162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4517a717540cSStefano Zampini if (row != col) { 45189162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4519a717540cSStefano Zampini } else { 45209162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4521a717540cSStefano Zampini } 4522906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4523a717540cSStefano Zampini } 4524a717540cSStefano Zampini } 4525a717540cSStefano Zampini } 452698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 452722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4528a717540cSStefano Zampini } 4529674ae819SStefano Zampini } 4530984c4197SStefano Zampini } else { 4531984c4197SStefano Zampini if (pcbddc->dbg_flag) { 45329162d606SStefano 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); 4533674ae819SStefano Zampini } 4534674ae819SStefano Zampini } 4535674ae819SStefano Zampini } 4536a717540cSStefano Zampini 4537a717540cSStefano Zampini /* free workspace */ 4538a717540cSStefano Zampini if (qr_needed) { 4539984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4540cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4541984c4197SStefano Zampini } 4542984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4543984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4544984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4545984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4546984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4547674ae819SStefano Zampini } 4548a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4549906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4550906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4551906d46d4SStefano Zampini 4552906d46d4SStefano Zampini /* assembling of global change of variable */ 455388c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4554bbb9e6c6SStefano Zampini Mat tmat; 455516f15bc4SStefano Zampini PetscInt bs; 455616f15bc4SStefano Zampini 4557906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4558906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4559bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4560bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4561bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4562bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 456316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 456416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4565906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4566bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4567bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4568bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4569bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4570bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4571e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4572e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4573bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4574bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 457588c03ad3SStefano Zampini 4576906d46d4SStefano Zampini /* check */ 4577906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4578906d46d4SStefano Zampini PetscReal error; 4579906d46d4SStefano Zampini Vec x,x_change; 4580906d46d4SStefano Zampini 4581906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4582906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4583906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4584906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4585e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4586e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4587bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4588e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4589e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4590906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4591906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4592906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4593906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4594bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4595906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4596906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4597906d46d4SStefano Zampini } 4598b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4599b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4600b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4601bf3a8328SStefano Zampini 46029a962809SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4603b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 4604ac632422SStefano Zampini Mat S_new,tmat; 4605bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4606bbb9e6c6SStefano Zampini 4607bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 46086816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4609bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4610bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4611bf3a8328SStefano Zampini IS is_V; 4612b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4613b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4614b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4615b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4616b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4617bf3a8328SStefano Zampini } 4618bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4619ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4620b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4621ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4622bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4623bf3a8328SStefano Zampini const PetscScalar *array; 4624bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4625bf3a8328SStefano Zampini PetscInt i,n_V; 4626bf3a8328SStefano Zampini 4627b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4628b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4629b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4630b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4631b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4632b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4633b087196eSStefano Zampini PetscScalar val; 4634b087196eSStefano Zampini PetscInt idx; 4635b087196eSStefano Zampini 4636b087196eSStefano Zampini idx = idxs_V[i]; 4637b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4638b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4639b087196eSStefano Zampini } 4640b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4641b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4642bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4643bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4644bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4645bf3a8328SStefano Zampini } 4646ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4647ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4648ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4649ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4650b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4651ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4652bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4653b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4654bf3a8328SStefano Zampini } 4655ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4656ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4657ac632422SStefano Zampini } 4658b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 465988c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4660b96c3477SStefano Zampini } 4661c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4662b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 4663c9db6a07SStefano Zampini PetscInt i; 4664c9db6a07SStefano Zampini 4665c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4666c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4667c9db6a07SStefano Zampini } 4668c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4669c9db6a07SStefano Zampini } 4670b96c3477SStefano Zampini } 467116909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 467216909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 467316909a7fSStefano Zampini } else { 4674906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 467516909a7fSStefano Zampini } 46761dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 467727b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 467872b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 467972b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 468072b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 468172b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 468272b8c272SStefano Zampini } 46831dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 468427b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4685b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4686b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4687906d46d4SStefano Zampini } else { 46881dd7afcfSStefano Zampini Mat benign_global = NULL; 468927b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 46901dd7afcfSStefano Zampini Mat tmat; 46911dd7afcfSStefano Zampini 46921dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 46931dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 46941dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 46951dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46961dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46971dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 46981dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46991dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 47001dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 47011dd7afcfSStefano Zampini if (pcbddc->benign_change) { 47021dd7afcfSStefano Zampini Mat M; 47031dd7afcfSStefano Zampini 47041dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 47051dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 47061dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 47071dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4708906d46d4SStefano Zampini } else { 47091dd7afcfSStefano Zampini Mat eye; 47101dd7afcfSStefano Zampini PetscScalar *array; 47111dd7afcfSStefano Zampini 47121dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 47131dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 47141dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 47151dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4716906d46d4SStefano Zampini } 47171dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 47181dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 47191dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 47201dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 47211dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 47221dd7afcfSStefano Zampini } 47231dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 47241dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 47251dd7afcfSStefano Zampini } 47261dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 47271dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 47281dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 472927b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 47301dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 47311dd7afcfSStefano Zampini } 47321dd7afcfSStefano Zampini } 473316909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 473416909a7fSStefano Zampini IS is_global; 473516909a7fSStefano Zampini const PetscInt *gidxs; 473616909a7fSStefano Zampini 473716909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 473816909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 473916909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 474016909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 474116909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 474216909a7fSStefano Zampini } 47431dd7afcfSStefano Zampini } 47441dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 47451dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4746b9b85e73SStefano Zampini } 4747a717540cSStefano Zampini 474872b8c272SStefano Zampini if (!pcbddc->fake_change) { 47494f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 47504f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 47514f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 47524f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4753019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4754019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4755019a44ceSStefano Zampini pcbddc->local_primal_size++; 4756019a44ceSStefano Zampini } 4757019a44ceSStefano Zampini 4758019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4759727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4760727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 47619f47a83aSStefano 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); 4762c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 47630e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 47649f47a83aSStefano 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); 4765727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4766727cdba6SStefano Zampini } 47670e6343abSStefano Zampini } 4768727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4769b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 477072b8c272SStefano Zampini } 477172b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4772727cdba6SStefano Zampini 4773a717540cSStefano Zampini /* flush dbg viewer */ 4774b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4775b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4776b8ffe317SStefano Zampini } 4777a717540cSStefano Zampini 4778e310c8b4SStefano Zampini /* free workspace */ 4779a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 47804641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 478108122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 47829162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 47839162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 478408122e43SStefano Zampini } else { 47859162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 47869162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 47879162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 478808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 478908122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 47909162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 47919162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 479208122e43SStefano Zampini } 4793674ae819SStefano Zampini PetscFunctionReturn(0); 4794674ae819SStefano Zampini } 4795674ae819SStefano Zampini 4796674ae819SStefano Zampini #undef __FUNCT__ 4797674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4798674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4799674ae819SStefano Zampini { 480071582508SStefano Zampini ISLocalToGlobalMapping map; 4801674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4802674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 480314f95afaSStefano Zampini PetscInt ierr,i,N; 4804674ae819SStefano Zampini 4805674ae819SStefano Zampini PetscFunctionBegin; 4806c8272957SStefano Zampini if (pcbddc->graphanalyzed && !pcbddc->recompute_topography) PetscFunctionReturn(0); 48078e61c736SStefano Zampini /* Reset previously computed graph */ 48088e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4809674ae819SStefano Zampini /* Init local Graph struct */ 48107fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 481171582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 481271582508SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N);CHKERRQ(ierr); 4813674ae819SStefano Zampini 4814575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 48159a962809SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) 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); 48169577ea80SStefano Zampini 4817674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4818d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 48194d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 48204d379d7bSStefano Zampini PetscInt nvtxs; 4821e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4822674ae819SStefano Zampini 48232fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 48242fffb893SStefano Zampini if (flg_row) { 48254d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4826b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 48272fffb893SStefano Zampini } 48282fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4829674ae819SStefano Zampini } 48309b28b941SStefano Zampini if (pcbddc->dbg_flag) { 48319b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4832674ae819SStefano Zampini } 4833674ae819SStefano Zampini 4834674ae819SStefano Zampini /* Setup of Graph */ 48354b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 483614f95afaSStefano 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); 4837674ae819SStefano Zampini 48384f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 48394f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 48404f1b2e48SStefano Zampini PetscInt *local_subs; 48414f1b2e48SStefano Zampini 48424f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 48434f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 48444f1b2e48SStefano Zampini const PetscInt *idxs; 48454f1b2e48SStefano Zampini PetscInt nl,j; 48464f1b2e48SStefano Zampini 48474f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 48484f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 484971582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 48504f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 48514f1b2e48SStefano Zampini } 48524f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 48534f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 48544f1b2e48SStefano Zampini } 48554f1b2e48SStefano Zampini 4856674ae819SStefano Zampini /* Graph's connected components analysis */ 4857674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 485871582508SStefano Zampini 485971582508SStefano Zampini /* set flag indicating analysis has been done */ 486071582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 4861674ae819SStefano Zampini PetscFunctionReturn(0); 4862674ae819SStefano Zampini } 4863674ae819SStefano Zampini 48649a7d3425SStefano Zampini #undef __FUNCT__ 48659a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 48669a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 48679a7d3425SStefano Zampini { 48689a7d3425SStefano Zampini PetscInt i,j; 48699a7d3425SStefano Zampini PetscScalar *alphas; 48709a7d3425SStefano Zampini PetscErrorCode ierr; 48719a7d3425SStefano Zampini 48729a7d3425SStefano Zampini PetscFunctionBegin; 4873785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 48749a7d3425SStefano Zampini for (i=0;i<n;i++) { 48759a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 4876669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 4877669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 4878669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 48799a7d3425SStefano Zampini } 48809a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 48819a7d3425SStefano Zampini PetscFunctionReturn(0); 48829a7d3425SStefano Zampini } 48839a7d3425SStefano Zampini 4884e7931f94SStefano Zampini #undef __FUNCT__ 488570cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 488657de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4887e7931f94SStefano Zampini { 488857de7509SStefano Zampini Mat A; 4889e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4890e7931f94SStefano Zampini PetscMPIInt size,rank,color; 489152e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 489252e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 489327b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 489457de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 489527b6a85dSStefano Zampini PetscInt xadj_count, *count; 489627b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 489727b6a85dSStefano Zampini PetscSubcomm psubcomm; 489827b6a85dSStefano Zampini MPI_Comm subcomm; 489952e5ac9dSStefano Zampini PetscErrorCode ierr; 4900a57a6d2fSStefano Zampini 4901e7931f94SStefano Zampini PetscFunctionBegin; 490257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 490357de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 490457de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 490557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 490657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 490757de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 490857de7509SStefano Zampini 490957de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 491057de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 491157de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 491257de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 491357de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 491457de7509SStefano Zampini im_active = !!(n); 491557de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 491657de7509SStefano Zampini void_procs = size - active_procs; 491757de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 491857de7509SStefano Zampini if (void_procs) { 491957de7509SStefano Zampini PetscInt ncand; 492057de7509SStefano Zampini 492157de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 492257de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 492357de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 492457de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 492557de7509SStefano Zampini if (!procs_candidates[i]) { 492657de7509SStefano Zampini procs_candidates[ncand++] = i; 492757de7509SStefano Zampini } 492857de7509SStefano Zampini } 492957de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 493057de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 493157de7509SStefano Zampini } 493257de7509SStefano Zampini 493314f0bfb9SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix 493414f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 493514f0bfb9SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1) { 493614f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 493714f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 493814f0bfb9SStefano Zampini else dest = rank; 493957de7509SStefano Zampini if (im_active) { 494057de7509SStefano Zampini issize = 1; 494157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 494214f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 494357de7509SStefano Zampini } else { 494414f0bfb9SStefano Zampini isidx = dest; 494557de7509SStefano Zampini } 494657de7509SStefano Zampini } else { 494757de7509SStefano Zampini issize = 0; 494857de7509SStefano Zampini isidx = -1; 494957de7509SStefano Zampini } 495057de7509SStefano Zampini *n_subdomains = active_procs; 495157de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 4952daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 495357de7509SStefano Zampini PetscFunctionReturn(0); 495457de7509SStefano Zampini } 4955c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4956c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 495727b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 4958e7931f94SStefano Zampini 4959e7931f94SStefano Zampini /* Get info on mapping */ 49603bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 49613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4962e7931f94SStefano Zampini 4963e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4964785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4965e7931f94SStefano Zampini xadj[0] = 0; 4966e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4967785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4968785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 496927b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 497027b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 497127b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 497227b6a85dSStefano Zampini count[shared[i][j]] += 1; 4973e7931f94SStefano Zampini 497427b6a85dSStefano Zampini xadj_count = 0; 49752b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 497627b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 497727b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 4978d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4979d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4980d023bfaeSStefano Zampini xadj_count++; 498127b6a85dSStefano Zampini break; 498227b6a85dSStefano Zampini } 4983e7931f94SStefano Zampini } 4984e7931f94SStefano Zampini } 4985d023bfaeSStefano Zampini xadj[1] = xadj_count; 498627b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 49873bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4988e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4989e7931f94SStefano Zampini 49903837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4991e7931f94SStefano Zampini 499227b6a85dSStefano Zampini /* Restrict work on active processes only */ 499327b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 499427b6a85dSStefano Zampini if (void_procs) { 499527b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 499627b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 499727b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 499827b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 499927b6a85dSStefano Zampini } else { 500027b6a85dSStefano Zampini psubcomm = NULL; 500127b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 500227b6a85dSStefano Zampini } 500327b6a85dSStefano Zampini 500427b6a85dSStefano Zampini v_wgt = NULL; 500527b6a85dSStefano Zampini if (!color) { 5006e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 5007e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 5008e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5009c8587f34SStefano Zampini } else { 501052e5ac9dSStefano Zampini Mat subdomain_adj; 501152e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 501252e5ac9dSStefano Zampini MatPartitioning partitioner; 501327b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 501452e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 501557de7509SStefano Zampini PetscMPIInt size; 5016b0c7d250SStefano Zampini PetscBool aggregate; 5017b0c7d250SStefano Zampini 501827b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 501927b6a85dSStefano Zampini if (void_procs) { 502027b6a85dSStefano Zampini PetscInt prank = rank; 5021785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 502227b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 5023e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 5024e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 5025c8587f34SStefano Zampini } 5026e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 502727b6a85dSStefano Zampini } else { 502827b6a85dSStefano Zampini oldranks = NULL; 502927b6a85dSStefano Zampini } 5030b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 503127b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 5032b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 5033b0c7d250SStefano Zampini PetscMPIInt nrank; 5034b0c7d250SStefano Zampini PetscScalar *vals; 5035b0c7d250SStefano Zampini 503627b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5037b0c7d250SStefano Zampini lrows = 0; 5038b0c7d250SStefano Zampini if (nrank<redprocs) { 5039b0c7d250SStefano Zampini lrows = size/redprocs; 5040b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5041b0c7d250SStefano Zampini } 504227b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5043b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5044b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5045b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5046b0c7d250SStefano Zampini row = nrank; 5047b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5048b0c7d250SStefano Zampini cols = adjncy; 5049b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5050b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5051b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5052b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5053b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 505452e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 505552e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 505652e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5057b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 505827b6a85dSStefano Zampini if (use_vwgt) { 505927b6a85dSStefano Zampini Vec v; 506027b6a85dSStefano Zampini const PetscScalar *array; 506127b6a85dSStefano Zampini PetscInt nl; 506227b6a85dSStefano Zampini 506327b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 506427b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 506527b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 506627b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 506727b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 506827b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 506927b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 507022db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 507127b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 507227b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 507327b6a85dSStefano Zampini } 5074b0c7d250SStefano Zampini } else { 507527b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 507627b6a85dSStefano Zampini if (use_vwgt) { 507727b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 507827b6a85dSStefano Zampini v_wgt[0] = local_size; 507927b6a85dSStefano Zampini } 5080b0c7d250SStefano Zampini } 508122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5082e7931f94SStefano Zampini 5083e7931f94SStefano Zampini /* Partition */ 508427b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5085e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 508627b6a85dSStefano Zampini if (v_wgt) { 5087e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5088c8587f34SStefano Zampini } 508957de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 509057de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5091e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5092e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 509322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5094e7931f94SStefano Zampini 509552e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 50966583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 509752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 509852e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 509957de7509SStefano Zampini if (!aggregate) { 510057de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 510127b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 510227b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 510327b6a85dSStefano Zampini #endif 510457de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 510527b6a85dSStefano Zampini } else if (oldranks) { 5106b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 510727b6a85dSStefano Zampini } else { 510827b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 510957de7509SStefano Zampini } 511028143c3dSStefano Zampini } else { 5111b0c7d250SStefano Zampini PetscInt idxs[1]; 5112b0c7d250SStefano Zampini PetscMPIInt tag; 5113b0c7d250SStefano Zampini MPI_Request *reqs; 5114b0c7d250SStefano Zampini 5115b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5116b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5117b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 511827b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 511928143c3dSStefano Zampini } 512027b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5121b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5122b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 512357de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 512427b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 512527b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 512627b6a85dSStefano Zampini #endif 512757de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 512827b6a85dSStefano Zampini } else if (oldranks) { 5129b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 513027b6a85dSStefano Zampini } else { 513127b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5132e7931f94SStefano Zampini } 513357de7509SStefano Zampini } 513452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5135e7931f94SStefano Zampini /* clean up */ 5136e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 513752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5138e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5139e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5140e7931f94SStefano Zampini } 514127b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 514257de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5143e7931f94SStefano Zampini 5144e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5145e7931f94SStefano Zampini i = 1; 514627b6a85dSStefano Zampini if (!color) i=0; 514757de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5148e7931f94SStefano Zampini PetscFunctionReturn(0); 5149e7931f94SStefano Zampini } 5150e7931f94SStefano Zampini 5151e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5152e7931f94SStefano Zampini 5153e7931f94SStefano Zampini #undef __FUNCT__ 5154e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 51551ae86dd6SStefano 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[]) 5156e7931f94SStefano Zampini { 515770cf5478SStefano Zampini Mat local_mat; 5158e7931f94SStefano Zampini IS is_sends_internal; 51599d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 51601ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 51619d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5162e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5163e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5164e7931f94SStefano Zampini const PetscInt* is_indices; 5165e7931f94SStefano Zampini MatType new_local_type; 5166e7931f94SStefano Zampini /* buffers */ 5167e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 516828143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 51699d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5170e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 51711ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 5172e7931f94SStefano Zampini /* MPI */ 517328143c3dSStefano Zampini MPI_Comm comm,comm_n; 517428143c3dSStefano Zampini PetscSubcomm subcomm; 5175e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 517628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 517728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 51781ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 51791ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 51801ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 5181e7931f94SStefano Zampini PetscErrorCode ierr; 5182e7931f94SStefano Zampini 5183e7931f94SStefano Zampini PetscFunctionBegin; 518457de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5185e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 518628143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 518757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 518857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 518957de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 519057de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 519157de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 51921ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 51931ae86dd6SStefano Zampini if (nvecs) { 51941ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 51951ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 51961ae86dd6SStefano Zampini } 519757de7509SStefano Zampini /* further checks */ 5198e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5199e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5200e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5201e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5202e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 520357de7509SStefano Zampini if (reuse && *mat_n) { 520470cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 520557de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 520670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 520728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 520870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 520970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 521070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 521170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 521270cf5478SStefano Zampini } 5213e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5214e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 521557de7509SStefano Zampini 5216e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5217e7931f94SStefano Zampini if (!is_sends) { 521828143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 521957de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5220c8587f34SStefano Zampini } else { 5221e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5222e7931f94SStefano Zampini is_sends_internal = is_sends; 5223c8587f34SStefano Zampini } 5224e7931f94SStefano Zampini 5225e7931f94SStefano Zampini /* get comm */ 5226a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5227e7931f94SStefano Zampini 5228e7931f94SStefano Zampini /* compute number of sends */ 5229e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5230e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5231e7931f94SStefano Zampini 5232e7931f94SStefano Zampini /* compute number of receives */ 5233e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5234785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5235e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5236e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5237e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5238e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5239e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5240e7931f94SStefano Zampini 524128143c3dSStefano Zampini /* restrict comm if requested */ 524228143c3dSStefano Zampini subcomm = 0; 524328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 524428143c3dSStefano Zampini if (restrict_comm) { 5245779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5246779c1cceSStefano Zampini 524728143c3dSStefano Zampini color = 0; 524853a05cb3SStefano Zampini if (restrict_full) { 524953a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 525053a05cb3SStefano Zampini } else { 525153a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 525253a05cb3SStefano Zampini } 5253b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 525428143c3dSStefano Zampini subcommsize = commsize - subcommsize; 525528143c3dSStefano Zampini /* check if reuse has been requested */ 525657de7509SStefano Zampini if (reuse) { 525728143c3dSStefano Zampini if (*mat_n) { 525828143c3dSStefano Zampini PetscMPIInt subcommsize2; 525928143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 526028143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 526128143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 526228143c3dSStefano Zampini } else { 526328143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 526428143c3dSStefano Zampini } 526528143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5266779c1cceSStefano Zampini PetscMPIInt rank; 5267779c1cceSStefano Zampini 5268779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 526928143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 527028143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 527128143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5272306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 527328143c3dSStefano Zampini } 527428143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 527528143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 527628143c3dSStefano Zampini } else { 527728143c3dSStefano Zampini comm_n = comm; 527828143c3dSStefano Zampini } 527928143c3dSStefano Zampini 5280e7931f94SStefano Zampini /* prepare send/receive buffers */ 5281785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5282e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5283785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5284e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 528528143c3dSStefano Zampini if (nis) { 5286854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 528728143c3dSStefano Zampini } 5288e7931f94SStefano Zampini 528928143c3dSStefano Zampini /* Get data from local matrices */ 52906c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5291e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5292e7931f94SStefano Zampini /* 5293e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5294e7931f94SStefano Zampini send_buffer_idxs should contain: 5295e7931f94SStefano Zampini - MatType_PRIVATE type 5296e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5297e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5298e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5299e7931f94SStefano Zampini */ 53006c4ed002SBarry Smith else { 5301e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 53023bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5303854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5304e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5305e7931f94SStefano Zampini send_buffer_idxs[1] = i; 53063bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5307e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 53083bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5309e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5310e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5311e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5312e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5313c8587f34SStefano Zampini } 5314c8587f34SStefano Zampini } 5315e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 531628143c3dSStefano Zampini /* additional is (if any) */ 531728143c3dSStefano Zampini if (nis) { 531828143c3dSStefano Zampini PetscMPIInt psum; 531928143c3dSStefano Zampini PetscInt j; 532028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 532128143c3dSStefano Zampini PetscInt plen; 532228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 532328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 532428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 532528143c3dSStefano Zampini } 5326854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 532728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 532828143c3dSStefano Zampini PetscInt plen; 532928143c3dSStefano Zampini const PetscInt *is_array_idxs; 533028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 533128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 533228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 533328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 533428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 533528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 533628143c3dSStefano Zampini } 533728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 533828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 533928143c3dSStefano Zampini } 534028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 534128143c3dSStefano Zampini } 534228143c3dSStefano Zampini 5343e7931f94SStefano Zampini buf_size_idxs = 0; 5344e7931f94SStefano Zampini buf_size_vals = 0; 534528143c3dSStefano Zampini buf_size_idxs_is = 0; 53461ae86dd6SStefano Zampini buf_size_vecs = 0; 5347e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5348e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5349e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 535028143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 53511ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 5352e7931f94SStefano Zampini } 5353785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5354785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 535595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 53561ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 5357e7931f94SStefano Zampini 5358e7931f94SStefano Zampini /* get new tags for clean communications */ 5359e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5360e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 536128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 53621ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 5363e7931f94SStefano Zampini 5364e7931f94SStefano Zampini /* allocate for requests */ 5365785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5366785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 536795ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 53681ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 5369785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5370785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 537195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 53721ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 5373e7931f94SStefano Zampini 5374e7931f94SStefano Zampini /* communications */ 5375e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5376e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 537728143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 53781ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 5379e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5380e7931f94SStefano Zampini source_dest = onodes[i]; 5381e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5382e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5383e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5384e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 538528143c3dSStefano Zampini if (nis) { 538657de7509SStefano Zampini source_dest = onodes_is[i]; 538728143c3dSStefano 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); 538828143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 538928143c3dSStefano Zampini } 53901ae86dd6SStefano Zampini if (nvecs) { 53911ae86dd6SStefano Zampini source_dest = onodes[i]; 53921ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 53931ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 53941ae86dd6SStefano Zampini } 5395e7931f94SStefano Zampini } 5396e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5397e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5398e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5399e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 540028143c3dSStefano Zampini if (nis) { 540128143c3dSStefano 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); 540228143c3dSStefano Zampini } 54031ae86dd6SStefano Zampini if (nvecs) { 54041ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 54051ae86dd6SStefano 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); 54061ae86dd6SStefano Zampini } 5407e7931f94SStefano Zampini } 5408e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5409e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5410e7931f94SStefano Zampini 5411e7931f94SStefano Zampini /* assemble new l2g map */ 5412e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5413e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 54149d30be91SStefano Zampini new_local_rows = 0; 5415e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 54169d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5417e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5418e7931f94SStefano Zampini } 54199d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5420e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 54219d30be91SStefano Zampini new_local_rows = 0; 5422e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 54239d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 54249d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5425e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5426e7931f94SStefano Zampini } 54279d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 54289d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5429e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5430e7931f94SStefano Zampini 5431e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5432e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5433e7931f94SStefano 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) */ 5434e7931f94SStefano Zampini if (n_recvs) { 543528143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5436e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5437e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5438e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5439e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5440e7931f94SStefano Zampini break; 5441e7931f94SStefano Zampini } 5442e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5443e7931f94SStefano Zampini } 5444e7931f94SStefano Zampini switch (new_local_type_private) { 544528143c3dSStefano Zampini case MATDENSE_PRIVATE: 544628143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5447e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5448e7931f94SStefano Zampini bs = 1; 544928143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 545028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 545128143c3dSStefano Zampini bs = 1; 545228143c3dSStefano Zampini } 5453e7931f94SStefano Zampini break; 5454e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5455e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5456e7931f94SStefano Zampini bs = 1; 5457e7931f94SStefano Zampini break; 5458e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5459e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5460e7931f94SStefano Zampini break; 5461e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5462e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5463e7931f94SStefano Zampini break; 5464e7931f94SStefano Zampini default: 54659d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5466e7931f94SStefano Zampini break; 5467e7931f94SStefano Zampini } 546828143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 546928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 547028143c3dSStefano Zampini bs = 1; 5471e7931f94SStefano Zampini } 5472e7931f94SStefano Zampini 547370cf5478SStefano Zampini /* create MATIS object if needed */ 547457de7509SStefano Zampini if (!reuse) { 5475e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5476e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 547770cf5478SStefano Zampini } else { 547870cf5478SStefano Zampini /* it also destroys the local matrices */ 547957de7509SStefano Zampini if (*mat_n) { 548070cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 548157de7509SStefano Zampini } else { /* this is a fake object */ 548257de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 548357de7509SStefano Zampini } 548470cf5478SStefano Zampini } 548570cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5486e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 54879d30be91SStefano Zampini 54889d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 54899d30be91SStefano Zampini 54909d30be91SStefano Zampini /* Global to local map of received indices */ 54919d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 54929d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 54939d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 54949d30be91SStefano Zampini 54959d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 54969d30be91SStefano Zampini buf_size_idxs = 0; 54979d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54989d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 54999d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 55009d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 55019d30be91SStefano Zampini } 55029d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 55039d30be91SStefano Zampini 55049d30be91SStefano Zampini /* set preallocation */ 55059d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 55069d30be91SStefano Zampini if (!newisdense) { 55079d30be91SStefano Zampini PetscInt *new_local_nnz=0; 55089d30be91SStefano Zampini 55099d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 55109d30be91SStefano Zampini if (n_recvs) { 55119d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 55129d30be91SStefano Zampini } 55139d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 55149d30be91SStefano Zampini PetscInt j; 55159d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 55169d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 55179d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 55189d30be91SStefano Zampini } 55199d30be91SStefano Zampini } else { 55209d30be91SStefano Zampini /* TODO */ 55219d30be91SStefano Zampini } 55229d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 55239d30be91SStefano Zampini } 55249d30be91SStefano Zampini if (new_local_nnz) { 55259d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 55269d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 55279d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 55289d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 55299d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 55309d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 55319d30be91SStefano Zampini } else { 55329d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55339d30be91SStefano Zampini } 55349d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 55359d30be91SStefano Zampini } else { 55369d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55379d30be91SStefano Zampini } 5538e7931f94SStefano Zampini 5539e7931f94SStefano Zampini /* set values */ 5540e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 55419d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5542e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5543e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5544e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 55459d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5546e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5547e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5548e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 554928143c3dSStefano Zampini } else { 555028143c3dSStefano Zampini /* TODO */ 5551e7931f94SStefano Zampini } 5552e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5553e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5554e7931f94SStefano Zampini } 5555e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5556e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 555770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 555870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55599d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 5560e7931f94SStefano Zampini 5561dfd14d43SStefano Zampini #if 0 556228143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5563e7931f94SStefano Zampini Vec lvec,rvec; 5564e7931f94SStefano Zampini PetscReal infty_error; 5565e7931f94SStefano Zampini 55662a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5567e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5568e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5569e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 557070cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5571e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5572e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5573e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5574e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5575e7931f94SStefano Zampini } 557628143c3dSStefano Zampini #endif 5577e7931f94SStefano Zampini 557828143c3dSStefano Zampini /* assemble new additional is (if any) */ 557928143c3dSStefano Zampini if (nis) { 558028143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 558128143c3dSStefano Zampini 558228143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5583854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 558428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 558528143c3dSStefano Zampini psum = 0; 558628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 558728143c3dSStefano Zampini for (j=0;j<nis;j++) { 558828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 558928143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 559028143c3dSStefano Zampini psum += plen; 559128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 559228143c3dSStefano Zampini } 559328143c3dSStefano Zampini } 5594854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5595854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 559628143c3dSStefano Zampini for (i=1;i<nis;i++) { 559728143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 559828143c3dSStefano Zampini } 559928143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 560028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 560128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 560228143c3dSStefano Zampini for (j=0;j<nis;j++) { 560328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 560428143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 560528143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 560628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 560728143c3dSStefano Zampini } 560828143c3dSStefano Zampini } 560928143c3dSStefano Zampini for (i=0;i<nis;i++) { 561028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 561128143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 561228143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 561328143c3dSStefano Zampini } 561428143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 561528143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 561628143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 561728143c3dSStefano Zampini } 5618e7931f94SStefano Zampini /* free workspace */ 561928143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5620e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5621e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5622e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5623e7931f94SStefano Zampini if (isdense) { 5624e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5625e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5626e7931f94SStefano Zampini } else { 5627e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5628e7931f94SStefano Zampini } 562928143c3dSStefano Zampini if (nis) { 563028143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 563128143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 563228143c3dSStefano Zampini } 56331ae86dd6SStefano Zampini 56341ae86dd6SStefano Zampini if (nvecs) { 56351ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56361ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56371ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56381ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56391ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 56401ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 56411ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 56421ae86dd6SStefano Zampini /* set values */ 56431ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 56441ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 56451ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56461ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 56471ae86dd6SStefano Zampini PetscInt j; 56481ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 56491ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 56501ae86dd6SStefano Zampini } 56511ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 56521ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 56531ae86dd6SStefano Zampini } 56541ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56551ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 56561ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 56571ae86dd6SStefano Zampini } 56581ae86dd6SStefano Zampini 56591ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 56601ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5661e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5662e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 56631ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 566428143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5665e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5666e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 56671ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 566828143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5669e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5670e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5671e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5672e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5673e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 567428143c3dSStefano Zampini if (nis) { 567528143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 567628143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 567728143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 567828143c3dSStefano Zampini } 567928143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 568028143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 568128143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 568228143c3dSStefano Zampini for (i=0;i<nis;i++) { 568328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 568428143c3dSStefano Zampini } 56851ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 56861ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56871ae86dd6SStefano Zampini } 568853a05cb3SStefano Zampini *mat_n = NULL; 568928143c3dSStefano Zampini } 5690e7931f94SStefano Zampini PetscFunctionReturn(0); 5691e7931f94SStefano Zampini } 5692a57a6d2fSStefano Zampini 569312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5694af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 569512edc857SStefano Zampini 5696c8587f34SStefano Zampini #undef __FUNCT__ 5697c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5698c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5699c8587f34SStefano Zampini { 5700c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5701c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 570220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 57031ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 57049881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 570520a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 57066e683305SStefano Zampini IS coarse_is,*isarray; 57076e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 570830368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5709f9eb5b7dSStefano Zampini PC pc_temp; 5710c8587f34SStefano Zampini PCType coarse_pc_type; 5711c8587f34SStefano Zampini KSPType coarse_ksp_type; 5712f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 57134f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 57146e683305SStefano Zampini Mat t_coarse_mat_is; 571557de7509SStefano Zampini PetscInt ncoarse; 571668457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 571722bc73bbSStefano Zampini PetscScalar *array; 571857de7509SStefano Zampini MatReuse coarse_mat_reuse; 571957de7509SStefano Zampini PetscBool restr, full_restr, have_void; 57209881197aSStefano Zampini PetscErrorCode ierr; 5721fdc09c96SStefano Zampini 5722c8587f34SStefano Zampini PetscFunctionBegin; 5723c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 572468457ee5SStefano 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 */ 5725fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 57265a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5727fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5728f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5729f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5730f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5731fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 573251bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 573351bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5734dc4bcba2SStefano Zampini PC pc; 5735dc4bcba2SStefano Zampini PetscBool isbddc; 5736dc4bcba2SStefano Zampini 5737dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5738dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5739dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5740dc4bcba2SStefano Zampini if (isbddc) { 574163c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 574263c961adSStefano Zampini } else { 5743727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 574463c961adSStefano Zampini } 5745fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5746fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5747fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5748f4ddd8eeSStefano Zampini } 5749fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5750fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5751f4ddd8eeSStefano Zampini } 575270cf5478SStefano Zampini /* reset any subassembling information */ 575357de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 575470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 575557de7509SStefano Zampini } 57566e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5757fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5758f4ddd8eeSStefano Zampini } 575957de7509SStefano Zampini /* assemble coarse matrix */ 576057de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 576157de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 576257de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 576357de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 576418a45a71SStefano Zampini } else { 576557de7509SStefano Zampini coarse_mat = NULL; 576657de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 57676e683305SStefano Zampini } 5768e7931f94SStefano Zampini 5769abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5770abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5771abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5772abbbba34SStefano Zampini 5773abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 577422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 577522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 577622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 577722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5778e176bc59SStefano 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); 57796e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 57806e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57816e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5782abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5783abbbba34SStefano Zampini 578457de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 578557de7509SStefano Zampini im_active = !!(pcis->n); 578657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 578757de7509SStefano Zampini 578814f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 578957de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 579057de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 579157de7509SStefano Zampini coarse_mat_is = NULL; 579257de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 579357de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 579457de7509SStefano Zampini full_restr = PETSC_TRUE; 57951ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 579657de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 579757de7509SStefano Zampini if (multilevel_requested) { 579857de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 579957de7509SStefano Zampini restr = PETSC_FALSE; 580057de7509SStefano Zampini full_restr = PETSC_FALSE; 580157de7509SStefano Zampini } else { 580257de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 580357de7509SStefano Zampini restr = PETSC_TRUE; 580457de7509SStefano Zampini full_restr = PETSC_TRUE; 580557de7509SStefano Zampini } 58064b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 580757de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 580857de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 5809a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 581057de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 5811a198735bSStefano Zampini } else { 5812a198735bSStefano Zampini PetscMPIInt size,rank; 5813a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 5814a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 5815a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 5816a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5817a198735bSStefano Zampini } 581857de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 581957de7509SStefano Zampini PetscInt psum; 582057de7509SStefano Zampini PetscMPIInt size; 582157de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 582257de7509SStefano Zampini else psum = 0; 582357de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 582457de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 582557de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 582657de7509SStefano Zampini } 582757de7509SStefano Zampini /* determine if we can go multilevel */ 582857de7509SStefano Zampini if (multilevel_requested) { 582957de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 583057de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 583157de7509SStefano Zampini } 583257de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 583357de7509SStefano Zampini 5834e4d548c7SStefano Zampini /* dump subassembling pattern */ 5835e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5836e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5837e4d548c7SStefano Zampini } 5838e4d548c7SStefano Zampini 58396e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 584027b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 58416e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 58426e683305SStefano Zampini const PetscInt *idxs; 58436e683305SStefano Zampini ISLocalToGlobalMapping tmap; 58446e683305SStefano Zampini 58456e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 58460be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 58476e683305SStefano Zampini /* allocate space for temporary storage */ 5848854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5849854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 58506e683305SStefano Zampini /* allocate for IS array */ 58516e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 58526e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 585327b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 585430368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5855854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 58566e683305SStefano Zampini /* dofs splitting */ 58576e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 58586e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 58596e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 58606e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58616e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58626e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58636e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 586430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 58656e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 58666e683305SStefano Zampini } 58676e683305SStefano Zampini /* neumann boundaries */ 58686e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 58696e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 58706e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 58716e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58726e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58736e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58746e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 587530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 58766e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 58776e683305SStefano Zampini } 58786e683305SStefano Zampini /* free memory */ 58796e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 58806e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 58816e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 58826e683305SStefano Zampini } else { 58836e683305SStefano Zampini nis = 0; 58846e683305SStefano Zampini nisdofs = 0; 58856e683305SStefano Zampini nisneu = 0; 588630368db7SStefano Zampini nisvert = 0; 58876e683305SStefano Zampini isarray = NULL; 58886e683305SStefano Zampini } 58896e683305SStefano Zampini /* destroy no longer needed map */ 58906e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 58916e683305SStefano Zampini 589257de7509SStefano Zampini /* subassemble */ 589357de7509SStefano Zampini if (multilevel_allowed) { 58941ae86dd6SStefano Zampini Vec vp[1]; 58951ae86dd6SStefano Zampini PetscInt nvecs = 0; 589657de7509SStefano Zampini PetscBool reuse,reuser; 58971ae86dd6SStefano Zampini 589857de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 589957de7509SStefano Zampini else reuse = PETSC_FALSE; 590057de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 59011ae86dd6SStefano Zampini vp[0] = NULL; 59021ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 59031ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 59041ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 59051ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 59061ae86dd6SStefano Zampini nvecs = 1; 59071ae86dd6SStefano Zampini 59081ae86dd6SStefano Zampini if (pcbddc->divudotp) { 5909a198735bSStefano Zampini Mat B,loc_divudotp; 59101ae86dd6SStefano Zampini Vec v,p; 59111ae86dd6SStefano Zampini IS dummy; 59121ae86dd6SStefano Zampini PetscInt np; 59131ae86dd6SStefano Zampini 5914a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 5915a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 59161ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 5917a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 59181ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 59191ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 59201ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 59211ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 59221ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 59231ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 59241ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 59251ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 59261ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 59271ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 59281ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 59291ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 593074e2c79eSStefano Zampini } 59311ae86dd6SStefano Zampini } 59321ae86dd6SStefano Zampini if (reuser) { 59331ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 593474e2c79eSStefano Zampini } else { 59351ae86dd6SStefano 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); 59361ae86dd6SStefano Zampini } 59371ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 59381ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 59391ae86dd6SStefano Zampini PetscInt nl; 59401ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 59411ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 59421ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59431ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 59441ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 59451ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 59461ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59471ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 5948a198735bSStefano Zampini } else { 5949a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 59501ae86dd6SStefano Zampini } 59511ae86dd6SStefano Zampini } else { 59521ae86dd6SStefano 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); 59536e683305SStefano Zampini } 595457de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 595557de7509SStefano Zampini PetscMPIInt size; 595657de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 595757de7509SStefano Zampini if (!multilevel_allowed) { 595857de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 59596e683305SStefano Zampini } else { 596057de7509SStefano Zampini Mat A; 5961779c1cceSStefano Zampini 596257de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 596357de7509SStefano Zampini if (coarse_mat_is) { 596457de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 596557de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 596657de7509SStefano Zampini coarse_mat = coarse_mat_is; 596757de7509SStefano Zampini } 596857de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 596957de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 597057de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5971779c1cceSStefano Zampini } 5972779c1cceSStefano Zampini } 597357de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 597457de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 59756e683305SStefano Zampini 59766e683305SStefano Zampini /* create local to global scatters for coarse problem */ 597768457ee5SStefano Zampini if (compute_vecs) { 59786e683305SStefano Zampini PetscInt lrows; 59796e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 598057de7509SStefano Zampini if (coarse_mat) { 598157de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 59826e683305SStefano Zampini } else { 59836e683305SStefano Zampini lrows = 0; 59846e683305SStefano Zampini } 59856e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 59866e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 59876e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 59886e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59896e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59906e683305SStefano Zampini } 59916e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5992c8587f34SStefano Zampini 5993f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5994f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5995f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5996f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5997f9eb5b7dSStefano Zampini } else { 5998f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5999f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 6000c8587f34SStefano Zampini } 6001c8587f34SStefano Zampini 60026e683305SStefano Zampini /* print some info if requested */ 60036e683305SStefano Zampini if (pcbddc->dbg_flag) { 60046e683305SStefano Zampini if (!multilevel_allowed) { 60056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 60066e683305SStefano Zampini if (multilevel_requested) { 60076e683305SStefano 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); 60086e683305SStefano Zampini } else if (pcbddc->max_levels) { 60096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 60106e683305SStefano Zampini } 60116e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 60126e683305SStefano Zampini } 60136e683305SStefano Zampini } 60146e683305SStefano Zampini 6015f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 601657de7509SStefano Zampini if (coarse_mat) { 60176a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 60186e683305SStefano Zampini if (pcbddc->dbg_flag) { 601957de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 60206e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 60216e683305SStefano Zampini } 6022f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 6023312be037SStefano Zampini char prefix[256],str_level[16]; 6024e604994aSStefano Zampini size_t len; 602557de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 6026422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 6027c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 6028f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 602957de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6030c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 60316e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 6032c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 6033c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6034e604994aSStefano Zampini /* prefix */ 6035e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 6036e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 6037e604994aSStefano Zampini if (!pcbddc->current_level) { 6038e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 6039e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 6040c8587f34SStefano Zampini } else { 6041e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 6042312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 6043312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 604434d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 6045312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 6046e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 6047e604994aSStefano Zampini } 6048e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 60493e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 60503e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 60513e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 60523e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 6053f9eb5b7dSStefano Zampini /* allow user customization */ 6054f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 60553e3c6dadSStefano Zampini } 60563e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 605751bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 60583e3c6dadSStefano Zampini if (nisdofs) { 60593e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 60603e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 60613e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 60623e3c6dadSStefano Zampini } 60633e3c6dadSStefano Zampini } 60643e3c6dadSStefano Zampini if (nisneu) { 60653e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 60663e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 6067312be037SStefano Zampini } 606830368db7SStefano Zampini if (nisvert) { 606930368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 607030368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 607130368db7SStefano Zampini } 6072f9eb5b7dSStefano Zampini 6073f9eb5b7dSStefano Zampini /* get some info after set from options */ 6074f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 6075f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 60764f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 60776e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 6078f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6079f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 6080f9eb5b7dSStefano Zampini } 608139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 60824f3a063dSStefano Zampini if (isredundant) { 60834f3a063dSStefano Zampini KSP inner_ksp; 60844f3a063dSStefano Zampini PC inner_pc; 60854f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 60864f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 60874f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 60884f3a063dSStefano Zampini } 6089f9eb5b7dSStefano Zampini 609057de7509SStefano Zampini /* parameters which miss an API */ 609157de7509SStefano Zampini if (isbddc) { 6092720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 6093720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 609457de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 609527b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 609627b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 6097a198735bSStefano Zampini Mat coarsedivudotp_is; 6098a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 6099a198735bSStefano Zampini IS row,col; 6100a198735bSStefano Zampini const PetscInt *gidxs; 6101a198735bSStefano Zampini PetscInt n,st,M,N; 6102a198735bSStefano Zampini 6103a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 6104a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 6105a198735bSStefano Zampini st = st-n; 6106a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 6107a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 6108a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 6109a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6110a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 6111a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6112a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 6113a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 6114a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 6115a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 6116a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 6117a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 6118a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 6119a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 6120a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 6121a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 6122a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 6123a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 6124a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 6125a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 61268ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 6127a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 6128720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 612959e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 6130720d30f9SStefano Zampini } 6131d4d8cf7bSStefano Zampini } 61329881197aSStefano Zampini 61333301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 61345a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 61353301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 61363301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 61373301b35fSStefano Zampini } 61383301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 61393301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 61403301b35fSStefano Zampini } 61413301b35fSStefano Zampini if (pc->pmat->spd_set) { 61423301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 61433301b35fSStefano Zampini } 614427b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 614527b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 614627b6a85dSStefano Zampini } 61476e683305SStefano Zampini /* set operators */ 61485f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 61496e683305SStefano Zampini if (pcbddc->dbg_flag) { 61506e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 61516e683305SStefano Zampini } 61526e683305SStefano Zampini } 61536e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6154b1ecc7b1SStefano Zampini #if 0 6155b9b85e73SStefano Zampini { 6156b9b85e73SStefano Zampini PetscViewer viewer; 6157b9b85e73SStefano Zampini char filename[256]; 6158b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6159b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 61606a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6161b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6162f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6163b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6164b9b85e73SStefano Zampini } 6165b9b85e73SStefano Zampini #endif 6166f9eb5b7dSStefano Zampini 616798a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 616898a51de6SStefano Zampini Vec crhs,csol; 616904708bb6SStefano Zampini 6170f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6171f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6172f347579bSStefano Zampini if (!csol) { 61732a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6174f9eb5b7dSStefano Zampini } 6175f347579bSStefano Zampini if (!crhs) { 61762a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6177f347579bSStefano Zampini } 6178b0f5fe93SStefano Zampini } 61791ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 6180b0f5fe93SStefano Zampini 6181b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6182b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6183b0f5fe93SStefano Zampini 6184b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 61854f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 61864f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 61874f1b2e48SStefano Zampini } 6188b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6189b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6190b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6191b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6192b0f5fe93SStefano Zampini if (coarse_mat) { 6193b0f5fe93SStefano Zampini Vec nullv; 6194b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6195b0f5fe93SStefano Zampini PetscInt nl; 6196b0f5fe93SStefano Zampini 6197b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6198b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6199b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6200b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6201b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6202b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6203b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6204b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6205b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6206b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6207b0f5fe93SStefano Zampini } 6208b0f5fe93SStefano Zampini } 6209b0f5fe93SStefano Zampini 6210b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6211b0f5fe93SStefano Zampini PetscBool ispreonly; 6212b0f5fe93SStefano Zampini 6213b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6214b0f5fe93SStefano Zampini PetscBool isnull; 6215b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6216bef83e63SStefano Zampini if (isnull) { 6217b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6218b0f5fe93SStefano Zampini } 6219bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6220b0f5fe93SStefano Zampini } 6221b0f5fe93SStefano Zampini /* setup coarse ksp */ 6222b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6223cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6224cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 62256e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6226c8587f34SStefano Zampini KSP check_ksp; 62272b510759SStefano Zampini KSPType check_ksp_type; 6228c8587f34SStefano Zampini PC check_pc; 62296e683305SStefano Zampini Vec check_vec,coarse_vec; 62306a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 62312b510759SStefano Zampini PetscInt its; 62326e683305SStefano Zampini PetscBool compute_eigs; 62336e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 62346e683305SStefano Zampini PetscInt neigs; 62358e185a42SStefano Zampini const char *prefix; 6236c8587f34SStefano Zampini 62372b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 62386e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6239422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 624023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6241f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6242e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6243e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6244e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 62452b510759SStefano Zampini if (ispreonly) { 62462b510759SStefano Zampini check_ksp_type = KSPPREONLY; 62476e683305SStefano Zampini compute_eigs = PETSC_FALSE; 62482b510759SStefano Zampini } else { 6249cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 62506e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6251c8587f34SStefano Zampini } 6252c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 62536e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 62546e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 62556e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6256a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6257a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6258a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6259a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6260c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6261c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6262c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6263c8587f34SStefano Zampini /* create random vec */ 62642701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6265c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 62666e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6267c8587f34SStefano Zampini /* solve coarse problem */ 62686e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6269cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 62706e683305SStefano Zampini if (compute_eigs) { 6271854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6272854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 62736e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 62741ae86dd6SStefano Zampini if (neigs) { 62756e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 62766e683305SStefano Zampini lambda_min = eigs_r[0]; 62776e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 62782701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 62792701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6280cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6281cbcc2c2aSStefano Zampini } 6282c8587f34SStefano Zampini } 6283c8587f34SStefano Zampini } 62841ae86dd6SStefano Zampini } 6285cbcc2c2aSStefano Zampini 6286c8587f34SStefano Zampini /* check coarse problem residual error */ 62876e683305SStefano Zampini if (pcbddc->dbg_flag) { 62886e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 62896e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62906e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6291c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 62926e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 62936e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6294779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 62956e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 62966e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 62976e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 62986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6299b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6300b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6301b0f5fe93SStefano Zampini } 63026e683305SStefano Zampini if (compute_eigs) { 63036e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6304deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6305c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 63066e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 63076e683305SStefano 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); 63086e683305SStefano Zampini for (i=0;i<neigs;i++) { 63096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6310c8587f34SStefano Zampini } 63116e683305SStefano Zampini } 63126e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 63136e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 63146e683305SStefano Zampini } 6315e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 63162701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6317c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 63186e683305SStefano Zampini if (compute_eigs) { 63196e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 63206e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6321c8587f34SStefano Zampini } 63226e683305SStefano Zampini } 63236e683305SStefano Zampini } 6324bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6325cbcc2c2aSStefano Zampini /* print additional info */ 6326cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 63276e683305SStefano Zampini /* waits until all processes reaches this point */ 63286e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6329cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6330cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6331cbcc2c2aSStefano Zampini } 6332cbcc2c2aSStefano Zampini 63332b510759SStefano Zampini /* free memory */ 6334fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6335c8587f34SStefano Zampini PetscFunctionReturn(0); 6336c8587f34SStefano Zampini } 6337674ae819SStefano Zampini 6338f34684f1SStefano Zampini #undef __FUNCT__ 6339f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6340f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6341f34684f1SStefano Zampini { 6342f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6343f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6344f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6345dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6346dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 634773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6348dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6349f34684f1SStefano Zampini PetscErrorCode ierr; 6350f34684f1SStefano Zampini 6351f34684f1SStefano Zampini PetscFunctionBegin; 6352f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 63536c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6354dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 63553bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6356dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6357dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 63586583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6359dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6360dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6361dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 63626c4ed002SBarry 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); 6363dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6364dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6365dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6366dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6367dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6368f34684f1SStefano Zampini 6369f34684f1SStefano Zampini /* check numbering */ 6370f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6371019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6372dc456d91SStefano Zampini PetscInt i; 6373b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6374f34684f1SStefano Zampini 6375f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6376f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6377f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 63781575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6379019a44ceSStefano Zampini /* counter */ 6380019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6381019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6382019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6383019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6384019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6385019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6386f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6387f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6388727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6389f34684f1SStefano Zampini } 6390f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6391f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6392f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6393e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6394e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6395e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6396e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6397f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6398019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6399f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6400019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 64012c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 640275c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6403b9b85e73SStefano Zampini set_error = PETSC_TRUE; 64042c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 64052c66d082SStefano 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); 6406f34684f1SStefano Zampini } 6407f34684f1SStefano Zampini } 6408019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6409b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6410f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6411f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6412f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6413f34684f1SStefano Zampini } 6414f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6415f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6416e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6417e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6418f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6419f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6420b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6421ca8b9ea9SStefano Zampini PetscInt *gidxs; 6422ca8b9ea9SStefano Zampini 6423ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 64243bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6425f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6426f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6427f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6428f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 64294bc2dc4bSStefano 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); 6430f34684f1SStefano Zampini } 6431f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6432ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6433f34684f1SStefano Zampini } 6434f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 64351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6436302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6437f34684f1SStefano Zampini } 64388bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6439f34684f1SStefano Zampini /* get back data */ 6440f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6441f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6442674ae819SStefano Zampini PetscFunctionReturn(0); 6443674ae819SStefano Zampini } 6444674ae819SStefano Zampini 6445e456f2a8SStefano Zampini #undef __FUNCT__ 6446e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6447a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6448e456f2a8SStefano Zampini { 6449e456f2a8SStefano Zampini IS localis_t; 6450a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6451e456f2a8SStefano Zampini PetscScalar *vals; 6452e456f2a8SStefano Zampini PetscErrorCode ierr; 6453e456f2a8SStefano Zampini 6454e456f2a8SStefano Zampini PetscFunctionBegin; 6455a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6456e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6457854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6458e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6459e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6460a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6461a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 64621035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6463a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 64641035eff8SStefano Zampini } 6465a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6466e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6467e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6468a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6469a7dc3881SStefano Zampini /* now compute set in local ordering */ 6470a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6471a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6472a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6473a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6474a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6475ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6476e456f2a8SStefano Zampini lsize++; 6477e456f2a8SStefano Zampini } 6478e456f2a8SStefano Zampini } 6479854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6480a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6481ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6482e456f2a8SStefano Zampini idxs[lsize++] = i; 6483e456f2a8SStefano Zampini } 6484e456f2a8SStefano Zampini } 6485a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6486a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6487e456f2a8SStefano Zampini *localis = localis_t; 6488e456f2a8SStefano Zampini PetscFunctionReturn(0); 6489e456f2a8SStefano Zampini } 6490906d46d4SStefano Zampini 6491b96c3477SStefano Zampini #undef __FUNCT__ 6492b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 649308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6494b96c3477SStefano Zampini { 6495a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6496b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6497b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6498a64f4aa4SStefano Zampini Mat S_j; 6499b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6500b96c3477SStefano Zampini PetscBool free_used_adj; 6501b96c3477SStefano Zampini PetscErrorCode ierr; 6502b96c3477SStefano Zampini 6503b96c3477SStefano Zampini PetscFunctionBegin; 6504b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6505b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 650608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6507b96c3477SStefano Zampini used_xadj = NULL; 6508b96c3477SStefano Zampini used_adjncy = NULL; 6509b96c3477SStefano Zampini } else { 651008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 651108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 651208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 651308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6514b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6515b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6516b96c3477SStefano Zampini } else { 65172fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6518b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6519b96c3477SStefano Zampini PetscInt nvtxs; 6520b96c3477SStefano Zampini 65212fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 65222fffb893SStefano Zampini if (flg_row) { 6523b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6524b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6525b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6526b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 65272fffb893SStefano Zampini } else { 65282fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 65292fffb893SStefano Zampini used_xadj = NULL; 65302fffb893SStefano Zampini used_adjncy = NULL; 65312fffb893SStefano Zampini } 65322fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6533b96c3477SStefano Zampini } 6534b96c3477SStefano Zampini } 6535d5574798SStefano Zampini 6536d5574798SStefano Zampini /* setup sub_schurs data */ 6537a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6538df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6539df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6540a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 654191af6908SStefano 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); 6542a64f4aa4SStefano Zampini } else { 65436816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 65444d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 6545a3df083aSStefano Zampini PetscInt benign_n; 654672b8c272SStefano Zampini Mat change = NULL; 65479d54b7f4SStefano Zampini Vec scaling = NULL; 654872b8c272SStefano Zampini IS change_primal = NULL; 6549a3df083aSStefano Zampini 65505feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 65515feab87aSStefano Zampini PetscInt n_vertices; 65525feab87aSStefano Zampini 65535feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 65542034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 65555feab87aSStefano Zampini } 655604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 655704708bb6SStefano Zampini if (!isseqaij) { 655804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 655904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 656004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 656104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 656204708bb6SStefano Zampini } else { 6563511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 656404708bb6SStefano Zampini } 656504708bb6SStefano Zampini } 6566a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6567a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6568ca92afb2SStefano Zampini } else { 6569a3df083aSStefano Zampini benign_n = 0; 6570ca92afb2SStefano Zampini } 6571b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6572b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6573b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 657472b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 657522db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6576b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 657722db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6578b7ab4a40SStefano Zampini } 6579b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6580b7ab4a40SStefano 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 */ 6581b7ab4a40SStefano Zampini if (need_change) { 658288c03ad3SStefano Zampini PC_IS *pcisf; 658388c03ad3SStefano Zampini PC_BDDC *pcbddcf; 658488c03ad3SStefano Zampini PC pcf; 658588c03ad3SStefano Zampini 6586e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 658788c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 658888c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 658988c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 659088c03ad3SStefano Zampini /* hacks */ 659188c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 659272b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 659372b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 659472b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 659572b8c272SStefano Zampini pcisf->n = pcis->n; 659672b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 659788c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 659888c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 659988c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 660088c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 660188c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 660288c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 660372b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 660488c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 660588c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 660672b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 660772b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 660872b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 660972b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 661072b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 661188c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 661272b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 661388c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 661488c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 661588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 661688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 661788c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 661888c03ad3SStefano Zampini pcf->ops->destroy = NULL; 661988c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 662088c03ad3SStefano Zampini } 66219d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 662291af6908SStefano 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); 662372b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 662472b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6625ca92afb2SStefano Zampini } 6626d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6627b96c3477SStefano Zampini 6628b96c3477SStefano Zampini /* free adjacency */ 6629b96c3477SStefano Zampini if (free_used_adj) { 6630b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6631b96c3477SStefano Zampini } 6632b96c3477SStefano Zampini PetscFunctionReturn(0); 6633b96c3477SStefano Zampini } 6634b96c3477SStefano Zampini 6635b96c3477SStefano Zampini #undef __FUNCT__ 6636b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 663708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6638b96c3477SStefano Zampini { 6639b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6640b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6641b96c3477SStefano Zampini PCBDDCGraph graph; 6642b96c3477SStefano Zampini PetscErrorCode ierr; 6643b96c3477SStefano Zampini 6644b96c3477SStefano Zampini PetscFunctionBegin; 6645b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 664608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 66473301b35fSStefano Zampini IS verticesIS,verticescomm; 66483301b35fSStefano Zampini PetscInt vsize,*idxs; 6649b96c3477SStefano Zampini 6650b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 66513301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 66523301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66533301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 66543301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 6655c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 6656b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 66577fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6658441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 66593301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6660b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6661b96c3477SStefano Zampini } else { 6662b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6663b96c3477SStefano Zampini } 6664e4d548c7SStefano Zampini /* print some info */ 6665e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6666e4d548c7SStefano Zampini IS vertices; 6667e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 6668c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6669e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6670e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6671e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6672e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6673e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6674e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6675e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6676e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6677e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6678c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6679e4d548c7SStefano Zampini } 6680b96c3477SStefano Zampini 6681b96c3477SStefano Zampini /* sub_schurs init */ 6682b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 6683b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 6684b334f244SStefano Zampini } 6685b334f244SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6686a64f4aa4SStefano Zampini 6687b96c3477SStefano Zampini /* free graph struct */ 668808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6689b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6690b96c3477SStefano Zampini } 6691b96c3477SStefano Zampini PetscFunctionReturn(0); 6692b96c3477SStefano Zampini } 6693fa34dd3eSStefano Zampini 6694fa34dd3eSStefano Zampini #undef __FUNCT__ 6695fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6696fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6697fa34dd3eSStefano Zampini { 6698fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6699fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6700fa34dd3eSStefano Zampini PetscErrorCode ierr; 6701fa34dd3eSStefano Zampini 6702fa34dd3eSStefano Zampini PetscFunctionBegin; 6703fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6704fa34dd3eSStefano Zampini IS zerodiag = NULL; 67054f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6706fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 67074f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 670875c01103SStefano Zampini PetscReal norm; 6709fa34dd3eSStefano Zampini PetscInt i; 6710fa34dd3eSStefano Zampini 6711fa34dd3eSStefano Zampini /* B0 and B0_B */ 6712fa34dd3eSStefano Zampini if (zerodiag) { 6713fa34dd3eSStefano Zampini IS dummy; 6714fa34dd3eSStefano Zampini 67154f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 67164f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6717fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6718fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6719fa34dd3eSStefano Zampini } 6720fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6721fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6722fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6723fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6724fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6725fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6726fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6727fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6728fa34dd3eSStefano Zampini /* S_j */ 6729fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6730fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6731fa34dd3eSStefano Zampini 6732fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6733fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6734fa34dd3eSStefano Zampini /* continuous in primal space */ 6735fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6736fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6737fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6738fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 67394f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 67404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6741fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6742fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6743fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6744fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6745fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6746fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6747fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6748fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6749fa34dd3eSStefano Zampini 6750fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6751fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6752fa34dd3eSStefano Zampini /* local with Schur */ 6753fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6754fa34dd3eSStefano Zampini if (zerodiag) { 6755fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67564f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6757fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6758fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6759fa34dd3eSStefano Zampini } 6760fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6761fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6762fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6763fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6764fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6765fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6766fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6767fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6768fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6769fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6770fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6771fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6772fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6773fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6774fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6775fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6776fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6777fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6778fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6779fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6780fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6781fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6782fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6783fa34dd3eSStefano Zampini if (zerodiag) { 6784fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6785fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67864f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6787fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6788fa34dd3eSStefano Zampini } 6789fa34dd3eSStefano Zampini /* BDDC */ 6790fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6791fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6792fa34dd3eSStefano Zampini 6793fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6794fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6795fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6796fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 67974f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 67984f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6799fa34dd3eSStefano Zampini } 68004f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6801fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6802fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6803fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6804fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6805fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6806fa34dd3eSStefano Zampini } 6807fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6808fa34dd3eSStefano Zampini } 6809