11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 6674ae819SStefano Zampini 7669cc0f4SStefano Zampini #undef __FUNCT__ 8669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 98ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 10669cc0f4SStefano Zampini { 11a198735bSStefano Zampini Mat loc_divudotp; 12fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 138ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 14669cc0f4SStefano Zampini IS *faces,*edges; 15669cc0f4SStefano Zampini PetscScalar *vals; 16669cc0f4SStefano Zampini const PetscScalar *array; 17669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 181ae86dd6SStefano Zampini PetscMPIInt rank; 19a198735bSStefano Zampini PetscErrorCode ierr; 20669cc0f4SStefano Zampini 21669cc0f4SStefano Zampini PetscFunctionBegin; 22669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 23669cc0f4SStefano Zampini if (graph->twodim) { 24669cc0f4SStefano Zampini lmaxneighs = 2; 25669cc0f4SStefano Zampini } else { 26669cc0f4SStefano Zampini lmaxneighs = 1; 27669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 28669cc0f4SStefano Zampini const PetscInt *idxs; 29669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 30669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 31669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 32669cc0f4SStefano Zampini } 33669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 34669cc0f4SStefano Zampini } 35669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 36669cc0f4SStefano Zampini maxsize = 0; 37669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 38669cc0f4SStefano Zampini PetscInt nn; 39669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 40669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 41669cc0f4SStefano Zampini } 42669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 43669cc0f4SStefano Zampini PetscInt nn; 44669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 45669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 46669cc0f4SStefano Zampini } 47669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 48669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 49669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 508ae0ca82SStefano Zampini if (!transpose) { 518ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 528ae0ca82SStefano Zampini } else { 538ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 548ae0ca82SStefano Zampini } 55669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 561ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 57669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 58669cc0f4SStefano Zampini PetscInt first,last; 59669cc0f4SStefano Zampini 608ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 61669cc0f4SStefano Zampini /* the near-null space fo BDDC carries information on quadrature weights, 62669cc0f4SStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 63669cc0f4SStefano Zampini and create a suitable set of basis vectors first */ 64669cc0f4SStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 65669cc0f4SStefano Zampini if (i>=first && i < last) { 66669cc0f4SStefano Zampini PetscScalar *data; 67669cc0f4SStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 68669cc0f4SStefano Zampini data[i-first] = 1.; 69669cc0f4SStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 70669cc0f4SStefano Zampini } 71669cc0f4SStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 72669cc0f4SStefano Zampini } 73669cc0f4SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 741ae86dd6SStefano Zampini for (i=0;i<maxneighs;i++) { /* reset vectors */ 751ae86dd6SStefano Zampini PetscInt first,last; 761ae86dd6SStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 771ae86dd6SStefano Zampini if (i>=first && i < last) { 781ae86dd6SStefano Zampini PetscScalar *data; 791ae86dd6SStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 801ae86dd6SStefano Zampini data[i-first] = 0.; 811ae86dd6SStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 821ae86dd6SStefano Zampini } 831ae86dd6SStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 841ae86dd6SStefano Zampini } 85669cc0f4SStefano Zampini /* compute local quad vec */ 86a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 878ae0ca82SStefano Zampini if (!transpose) { 88a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 898ae0ca82SStefano Zampini } else { 908ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 918ae0ca82SStefano Zampini } 92669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 938ae0ca82SStefano Zampini if (!transpose) { 94a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 958ae0ca82SStefano Zampini } else { 968ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 978ae0ca82SStefano Zampini } 98fa23a32eSStefano Zampini if (vl2l) { 99fa23a32eSStefano Zampini ierr = VecGetSubVector(v,vl2l,&vins);CHKERRQ(ierr); 100fa23a32eSStefano Zampini } else { 101fa23a32eSStefano Zampini vins = v; 102fa23a32eSStefano Zampini } 103fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 104669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 1059a962809SStefano Zampini 1061ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 1071ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 108669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 109669cc0f4SStefano Zampini const PetscInt *idxs; 110669cc0f4SStefano Zampini PetscInt idx,nn,j; 111669cc0f4SStefano Zampini 112669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 113669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 114669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 1151ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 116669cc0f4SStefano Zampini idx = -(idx+1); 117669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 118669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 119669cc0f4SStefano Zampini ierr = ISDestroy(&faces[i]);CHKERRQ(ierr); 120669cc0f4SStefano Zampini } 121669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 122669cc0f4SStefano Zampini const PetscInt *idxs; 123669cc0f4SStefano Zampini PetscInt idx,nn,j; 124669cc0f4SStefano Zampini 125669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 126669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 127669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 1281ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 129669cc0f4SStefano Zampini idx = -(idx+1); 130669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 131669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 132669cc0f4SStefano Zampini ierr = ISDestroy(&edges[i]);CHKERRQ(ierr); 133669cc0f4SStefano Zampini } 134669cc0f4SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 135669cc0f4SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 136fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 137fa23a32eSStefano Zampini if (vl2l) { 138fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 139fa23a32eSStefano Zampini } 140669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 141669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 142669cc0f4SStefano Zampini 143669cc0f4SStefano Zampini /* assemble near null space */ 144669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 145669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 146669cc0f4SStefano Zampini } 147669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 148669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 149669cc0f4SStefano Zampini } 150669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 151669cc0f4SStefano Zampini PetscFunctionReturn(0); 152669cc0f4SStefano Zampini } 153669cc0f4SStefano Zampini 154669cc0f4SStefano Zampini 155a3df083aSStefano Zampini #undef __FUNCT__ 1561f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 1571f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 1581f4df5f7SStefano Zampini { 1591f4df5f7SStefano Zampini PetscErrorCode ierr; 1601f4df5f7SStefano Zampini Vec local,global; 1611f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1621f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1631f4df5f7SStefano Zampini 1641f4df5f7SStefano Zampini PetscFunctionBegin; 1651f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 1661f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 1671f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 1681f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 1691f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 1701f4df5f7SStefano Zampini PetscInt i; 1711f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1721f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 1731f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1741f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 1751f4df5f7SStefano Zampini } 1761f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 1771f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 1781f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 1791f4df5f7SStefano Zampini } 1801f4df5f7SStefano Zampini } else { 181986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 1821f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 183986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 184986cdee1SStefano Zampini if (i > 1) { 185986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 1861f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1871f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 1881f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1891f4df5f7SStefano Zampini } 1901f4df5f7SStefano Zampini } 1911f4df5f7SStefano Zampini } 192986cdee1SStefano Zampini } 1931f4df5f7SStefano Zampini 1941f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 1951f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 1961f4df5f7SStefano Zampini } 1971f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 1981f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1991f4df5f7SStefano Zampini } 2001f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 2011f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 2021f4df5f7SStefano Zampini } 2031f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 2041f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 2051f4df5f7SStefano Zampini PetscFunctionReturn(0); 2061f4df5f7SStefano Zampini } 2071f4df5f7SStefano Zampini 2081f4df5f7SStefano Zampini #undef __FUNCT__ 2093e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 2103e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 2113e589ea0SStefano Zampini { 2123e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 2133e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2143e589ea0SStefano Zampini PetscErrorCode ierr; 2153e589ea0SStefano Zampini 2163e589ea0SStefano Zampini PetscFunctionBegin; 2173e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 2183e589ea0SStefano Zampini PetscFunctionReturn(0); 2193e589ea0SStefano Zampini } 2203e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2213e589ea0SStefano Zampini Vec swap; 2223e589ea0SStefano Zampini 2233e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 2243e589ea0SStefano Zampini swap = pcbddc->work_change; 2253e589ea0SStefano Zampini pcbddc->work_change = r; 2263e589ea0SStefano Zampini r = swap; 2273e589ea0SStefano Zampini } 2283e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2293e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2303e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 2313e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 2323e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2333e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2343e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2353e589ea0SStefano Zampini Vec swap; 2363e589ea0SStefano Zampini 2373e589ea0SStefano Zampini swap = r; 2383e589ea0SStefano Zampini r = pcbddc->work_change; 2393e589ea0SStefano Zampini pcbddc->work_change = swap; 2403e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 2413e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 2423e589ea0SStefano Zampini } 2433e589ea0SStefano Zampini PetscFunctionReturn(0); 2443e589ea0SStefano Zampini } 2453e589ea0SStefano Zampini 2463e589ea0SStefano Zampini #undef __FUNCT__ 247a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 248a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 249a3df083aSStefano Zampini { 250a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 251a3df083aSStefano Zampini PetscErrorCode ierr; 252a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 253a3df083aSStefano Zampini 254a3df083aSStefano Zampini PetscFunctionBegin; 255a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 256a3df083aSStefano Zampini if (transpose) { 257a3df083aSStefano Zampini apply_right = ctx->apply_left; 258a3df083aSStefano Zampini apply_left = ctx->apply_right; 259a3df083aSStefano Zampini } else { 260a3df083aSStefano Zampini apply_right = ctx->apply_right; 261a3df083aSStefano Zampini apply_left = ctx->apply_left; 262a3df083aSStefano Zampini } 263a3df083aSStefano Zampini reset_x = PETSC_FALSE; 264a3df083aSStefano Zampini if (apply_right) { 265a3df083aSStefano Zampini const PetscScalar *ax; 266a3df083aSStefano Zampini PetscInt nl,i; 267a3df083aSStefano Zampini 268a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 269a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 270a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 271a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 272a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 273a3df083aSStefano Zampini PetscScalar sum,val; 274a3df083aSStefano Zampini const PetscInt *idxs; 275a3df083aSStefano Zampini PetscInt nz,j; 276a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 277a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 278a3df083aSStefano Zampini sum = 0.; 279a3df083aSStefano Zampini if (ctx->apply_p0) { 280a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 281a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 282a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 283a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 284a3df083aSStefano Zampini } 285a3df083aSStefano Zampini } else { 286a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 287a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 288a3df083aSStefano Zampini } 289a3df083aSStefano Zampini } 290a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 291a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 292a3df083aSStefano Zampini } 293a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 294a3df083aSStefano Zampini reset_x = PETSC_TRUE; 295a3df083aSStefano Zampini } 296a3df083aSStefano Zampini if (transpose) { 297a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 298a3df083aSStefano Zampini } else { 299a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 300a3df083aSStefano Zampini } 301a3df083aSStefano Zampini if (reset_x) { 302a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 303a3df083aSStefano Zampini } 304a3df083aSStefano Zampini if (apply_left) { 305a3df083aSStefano Zampini PetscScalar *ay; 306a3df083aSStefano Zampini PetscInt i; 307a3df083aSStefano Zampini 308a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 309a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 310a3df083aSStefano Zampini PetscScalar sum,val; 311a3df083aSStefano Zampini const PetscInt *idxs; 312a3df083aSStefano Zampini PetscInt nz,j; 313a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 314a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 315a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 316a3df083aSStefano Zampini if (ctx->apply_p0) { 317a3df083aSStefano Zampini sum = 0.; 318a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 319a3df083aSStefano Zampini sum += ay[idxs[j]]; 320a3df083aSStefano Zampini ay[idxs[j]] += val; 321a3df083aSStefano Zampini } 322a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 323a3df083aSStefano Zampini } else { 324a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 325a3df083aSStefano Zampini ay[idxs[j]] += val; 326a3df083aSStefano Zampini } 327a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 328a3df083aSStefano Zampini } 329a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 330a3df083aSStefano Zampini } 331a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 332a3df083aSStefano Zampini } 333a3df083aSStefano Zampini PetscFunctionReturn(0); 334a3df083aSStefano Zampini } 335a3df083aSStefano Zampini 336a3df083aSStefano Zampini #undef __FUNCT__ 337a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 338a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 339a3df083aSStefano Zampini { 340a3df083aSStefano Zampini PetscErrorCode ierr; 341a3df083aSStefano Zampini 342a3df083aSStefano Zampini PetscFunctionBegin; 343a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 344a3df083aSStefano Zampini PetscFunctionReturn(0); 345a3df083aSStefano Zampini } 346a3df083aSStefano Zampini 347a3df083aSStefano Zampini #undef __FUNCT__ 348a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 349a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 350a3df083aSStefano Zampini { 351a3df083aSStefano Zampini PetscErrorCode ierr; 352a3df083aSStefano Zampini 353a3df083aSStefano Zampini PetscFunctionBegin; 354a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 355a3df083aSStefano Zampini PetscFunctionReturn(0); 356a3df083aSStefano Zampini } 357a3df083aSStefano Zampini 358a3df083aSStefano Zampini #undef __FUNCT__ 359a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 360a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 361a3df083aSStefano Zampini { 362a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 363a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 364a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 365a3df083aSStefano Zampini PetscErrorCode ierr; 366a3df083aSStefano Zampini 367a3df083aSStefano Zampini PetscFunctionBegin; 368a3df083aSStefano Zampini if (!restore) { 3691dd7afcfSStefano Zampini Mat A_IB,A_BI; 370a3df083aSStefano Zampini PetscScalar *work; 371b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 372a3df083aSStefano Zampini 3739a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 3749a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 375a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 376a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 377a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 378a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 379a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 380a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 381a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 382a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 383a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 384a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 385a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 386a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 387059032f7SStefano Zampini if (reuse) { 388a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 3891dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 390059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 391059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 392059032f7SStefano Zampini PetscInt i; 393059032f7SStefano Zampini 394059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 395059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 396059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 397059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 398059032f7SStefano Zampini } 399059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 4001dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 401059032f7SStefano Zampini } 402a3df083aSStefano Zampini ctx->A = pcis->A_IB; 403a3df083aSStefano Zampini ctx->work = work; 404a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 405a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 406a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 407a3df083aSStefano Zampini pcis->A_IB = A_IB; 408a3df083aSStefano Zampini 409a3df083aSStefano Zampini /* A_BI as A_IB^T */ 410a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 411a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 412a3df083aSStefano Zampini pcis->A_BI = A_BI; 413a3df083aSStefano Zampini } else { 4141dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 4151dd7afcfSStefano Zampini PetscFunctionReturn(0); 4161dd7afcfSStefano Zampini } 417a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 418a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 419a3df083aSStefano Zampini pcis->A_IB = ctx->A; 4201dd7afcfSStefano Zampini ctx->A = NULL; 4211dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 4221dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 4231dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 4241dd7afcfSStefano Zampini if (ctx->free) { 425059032f7SStefano Zampini PetscInt i; 4261dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 427059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 428059032f7SStefano Zampini } 429059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 430059032f7SStefano Zampini } 431a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 432a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 433a3df083aSStefano Zampini } 434a3df083aSStefano Zampini PetscFunctionReturn(0); 435a3df083aSStefano Zampini } 436a3df083aSStefano Zampini 437a3df083aSStefano Zampini /* used just in bddc debug mode */ 438a3df083aSStefano Zampini #undef __FUNCT__ 439a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 440a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 441a3df083aSStefano Zampini { 442a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 443a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 444a3df083aSStefano Zampini Mat An; 445a3df083aSStefano Zampini PetscErrorCode ierr; 446a3df083aSStefano Zampini 447a3df083aSStefano Zampini PetscFunctionBegin; 448a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 449a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 450a3df083aSStefano Zampini if (is1) { 451a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 452a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 453a3df083aSStefano Zampini } else { 454a3df083aSStefano Zampini *B = An; 455a3df083aSStefano Zampini } 456a3df083aSStefano Zampini PetscFunctionReturn(0); 457a3df083aSStefano Zampini } 458a3df083aSStefano Zampini 4591cf9b237SStefano Zampini /* TODO: add reuse flag */ 4601cf9b237SStefano Zampini #undef __FUNCT__ 4611cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 4621cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 4631cf9b237SStefano Zampini { 4641cf9b237SStefano Zampini Mat Bt; 4651cf9b237SStefano Zampini PetscScalar *a,*bdata; 4661cf9b237SStefano Zampini const PetscInt *ii,*ij; 4671cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 4681cf9b237SStefano Zampini PetscBool flg_row; 4691cf9b237SStefano Zampini PetscErrorCode ierr; 4701cf9b237SStefano Zampini 4711cf9b237SStefano Zampini PetscFunctionBegin; 4721cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 4731cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4741cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 4751cf9b237SStefano Zampini nnz = n; 4761cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 4771cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 4781cf9b237SStefano Zampini } 4791cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 4801cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 4811cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 4821cf9b237SStefano Zampini nnz = 0; 4831cf9b237SStefano Zampini bii[0] = 0; 4841cf9b237SStefano Zampini for (i=0;i<n;i++) { 4851cf9b237SStefano Zampini PetscInt j; 4861cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 4871cf9b237SStefano Zampini PetscScalar entry = a[j]; 4881cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 4891cf9b237SStefano Zampini bij[nnz] = ij[j]; 4901cf9b237SStefano Zampini bdata[nnz] = entry; 4911cf9b237SStefano Zampini nnz++; 4921cf9b237SStefano Zampini } 4931cf9b237SStefano Zampini } 4941cf9b237SStefano Zampini bii[i+1] = nnz; 4951cf9b237SStefano Zampini } 4961cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 4971cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 4981cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4991cf9b237SStefano Zampini { 5001cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 5011cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 5021cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 5031cf9b237SStefano Zampini } 5041cf9b237SStefano Zampini *B = Bt; 5051cf9b237SStefano Zampini PetscFunctionReturn(0); 5061cf9b237SStefano Zampini } 5071cf9b237SStefano Zampini 508674ae819SStefano Zampini #undef __FUNCT__ 5094f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 5104f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 5114f1b2e48SStefano Zampini { 5124f1b2e48SStefano Zampini Mat B; 5134f1b2e48SStefano Zampini IS is_dummy,*cc_n; 5144f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 5154f1b2e48SStefano Zampini PCBDDCGraph graph; 5164f1b2e48SStefano Zampini PetscInt i,n; 5174f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 5184f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 5194f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 5204f1b2e48SStefano Zampini PetscErrorCode ierr; 5214f1b2e48SStefano Zampini 5224f1b2e48SStefano Zampini PetscFunctionBegin; 52363c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 52463c961adSStefano Zampini *ncc = 0; 52563c961adSStefano Zampini *cc = NULL; 52663c961adSStefano Zampini PetscFunctionReturn(0); 52763c961adSStefano Zampini } 5284f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 5294f1b2e48SStefano Zampini if (!isseqaij && filter) { 5301cf9b237SStefano Zampini PetscBool isseqdense; 5311cf9b237SStefano Zampini 5321cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 5331cf9b237SStefano Zampini if (!isseqdense) { 5344f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 5351cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 5361cf9b237SStefano Zampini PetscScalar *array; 5371cf9b237SStefano Zampini PetscReal chop=1.e-6; 5381cf9b237SStefano Zampini 5391cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 5401cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 5411cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 5421cf9b237SStefano Zampini for (i=0;i<n;i++) { 5431cf9b237SStefano Zampini PetscInt j; 5441cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 5451cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 5461cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 5471cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 5481cf9b237SStefano Zampini } 5491cf9b237SStefano Zampini } 5501cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 5519d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 5521cf9b237SStefano Zampini } 5534f1b2e48SStefano Zampini } else { 5544f1b2e48SStefano Zampini B = A; 5554f1b2e48SStefano Zampini } 5564f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5574f1b2e48SStefano Zampini 5584f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 5594f1b2e48SStefano Zampini if (filter) { 5604f1b2e48SStefano Zampini PetscScalar *data; 5614f1b2e48SStefano Zampini PetscInt j,cum; 5624f1b2e48SStefano Zampini 5634f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 5644f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 5654f1b2e48SStefano Zampini cum = 0; 5664f1b2e48SStefano Zampini for (i=0;i<n;i++) { 5674f1b2e48SStefano Zampini PetscInt t; 5684f1b2e48SStefano Zampini 5694f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 5704f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 5714f1b2e48SStefano Zampini continue; 5724f1b2e48SStefano Zampini } 5734f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 5744f1b2e48SStefano Zampini } 5754f1b2e48SStefano Zampini t = xadj_filtered[i]; 5764f1b2e48SStefano Zampini xadj_filtered[i] = cum; 5774f1b2e48SStefano Zampini cum += t; 5784f1b2e48SStefano Zampini } 5794f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 5804f1b2e48SStefano Zampini } else { 5814f1b2e48SStefano Zampini xadj_filtered = NULL; 5824f1b2e48SStefano Zampini adjncy_filtered = NULL; 5834f1b2e48SStefano Zampini } 5844f1b2e48SStefano Zampini 5854f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 5864f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 5874f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 5894f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 5904f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 5914f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 5924f1b2e48SStefano Zampini if (xadj_filtered) { 5934f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 5944f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 5954f1b2e48SStefano Zampini } else { 5964f1b2e48SStefano Zampini graph->xadj = xadj; 5974f1b2e48SStefano Zampini graph->adjncy = adjncy; 5984f1b2e48SStefano Zampini } 5994f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 6004f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6014f1b2e48SStefano Zampini /* partial clean up */ 6024f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 6041cf9b237SStefano Zampini if (A != B) { 6054f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini } 6074f1b2e48SStefano Zampini 6084f1b2e48SStefano Zampini /* get back data */ 6091cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 6101cf9b237SStefano Zampini if (cc) { 6114f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 6124f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 6134f1b2e48SStefano 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); 6144f1b2e48SStefano Zampini } 6154f1b2e48SStefano Zampini *cc = cc_n; 6161cf9b237SStefano Zampini } 6174f1b2e48SStefano Zampini /* clean up graph */ 6184f1b2e48SStefano Zampini graph->xadj = 0; 6194f1b2e48SStefano Zampini graph->adjncy = 0; 6204f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini PetscFunctionReturn(0); 6224f1b2e48SStefano Zampini } 6234f1b2e48SStefano Zampini 6244f1b2e48SStefano Zampini #undef __FUNCT__ 6255408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 6265408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 6275408967cSStefano Zampini { 6285408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6295408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 630dee84bffSStefano Zampini IS dirIS = NULL; 6314f1b2e48SStefano Zampini PetscInt i; 6325408967cSStefano Zampini PetscErrorCode ierr; 6335408967cSStefano Zampini 6345408967cSStefano Zampini PetscFunctionBegin; 635dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 6365408967cSStefano Zampini if (zerodiag) { 6375408967cSStefano Zampini Mat A; 6385408967cSStefano Zampini Vec vec3_N; 6395408967cSStefano Zampini PetscScalar *vals; 6405408967cSStefano Zampini const PetscInt *idxs; 641d12d3064SStefano Zampini PetscInt nz,*count; 6425408967cSStefano Zampini 6435408967cSStefano Zampini /* p0 */ 6445408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 6455408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 6465408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6475408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6484f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 6495408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6505408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6515408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6525408967cSStefano Zampini /* v_I */ 6535408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 6545408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 6555408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6565408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 6575408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6585408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 6595408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6605408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6615408967cSStefano Zampini if (dirIS) { 6625408967cSStefano Zampini PetscInt n; 6635408967cSStefano Zampini 6645408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 6655408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 6665408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 6675408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6685408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 6695408967cSStefano Zampini } 6705408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 6715408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 6725408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 6735408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 674669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 6755408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 6765408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 6779a962809SStefano 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])); 6785408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6795408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 680d12d3064SStefano Zampini 681d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 682d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 683d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 684d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 685d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 686d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6879a962809SStefano 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]); 688d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 689d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 6905408967cSStefano Zampini } 691dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 6925408967cSStefano Zampini 6935408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 6945408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 6954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 6965408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 6974f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 6985408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 699*f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 700*f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 701*f2a566d8SStefano 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); 702*f2a566d8SStefano Zampini } 7035408967cSStefano Zampini PetscFunctionReturn(0); 7045408967cSStefano Zampini } 7055408967cSStefano Zampini 7065408967cSStefano Zampini #undef __FUNCT__ 707339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 708339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 709339f8db1SStefano Zampini { 710339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7114f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 712b0f5fe93SStefano Zampini PetscInt nz,n; 7131f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 7144f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 715339f8db1SStefano Zampini PetscErrorCode ierr; 716339f8db1SStefano Zampini 717339f8db1SStefano Zampini PetscFunctionBegin; 7189f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 7199f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 720a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 721a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 722a3df083aSStefano Zampini } 723a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 724a3df083aSStefano Zampini pcbddc->benign_n = 0; 7251ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 7264f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 7274f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 7284f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 7291ae86dd6SStefano Zampini since the local Schur complements are already SPD 7304f1b2e48SStefano Zampini */ 7314f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 7324f1b2e48SStefano Zampini have_null = PETSC_TRUE; 73340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 7344f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 7354f1b2e48SStefano Zampini 7364f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 7374f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 7384f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 7394f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 740ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 74140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 74240fa8d13SStefano Zampini if (!sorted) { 74340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 74440fa8d13SStefano Zampini } 74540fa8d13SStefano Zampini } else { 74640fa8d13SStefano Zampini pressures = NULL; 74740fa8d13SStefano Zampini } 74897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 74997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 75027b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 75197d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 752339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 753339f8db1SStefano Zampini if (!sorted) { 754339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 755339f8db1SStefano Zampini } 756339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 7574f1b2e48SStefano Zampini if (!nz) { 7584f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 7594f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 76040fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 76140fa8d13SStefano Zampini } 7624f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 7634f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 7644f1b2e48SStefano Zampini zerodiag_subs = NULL; 7654f1b2e48SStefano Zampini pcbddc->benign_n = 0; 7661f4df5f7SStefano Zampini n_interior_dofs = 0; 7671f4df5f7SStefano Zampini interior_dofs = NULL; 7681f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 7691f4df5f7SStefano Zampini PetscInt n,i,j; 7701f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 7711f4df5f7SStefano Zampini PetscInt *iwork; 7721f4df5f7SStefano Zampini 7731f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 7741f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7751f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 7761f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 7771f4df5f7SStefano Zampini for (i=0;i<n_neigh;i++) 7781f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 7791f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 7801f4df5f7SStefano Zampini for (i=0;i<n;i++) 7811f4df5f7SStefano Zampini if (!iwork[i]) 7821f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 7831f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 7841f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7851f4df5f7SStefano Zampini } 7864f1b2e48SStefano Zampini if (has_null_pressures) { 7874f1b2e48SStefano Zampini IS *subs; 7881f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 7891f4df5f7SStefano Zampini const PetscInt *idxs; 7901f4df5f7SStefano Zampini PetscScalar *array; 7911f4df5f7SStefano Zampini Vec *work; 7921f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 7934f1b2e48SStefano Zampini 7944f1b2e48SStefano Zampini subs = pcbddc->local_subs; 7954f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 7961f4df5f7SStefano 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) */ 7971f4df5f7SStefano Zampini if (pcbddc->current_level) { 7981f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 7991f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 8001f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 8011f4df5f7SStefano Zampini /* work[0] = 1_p */ 8021f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 8031f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 8041f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8051f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 8061f4df5f7SStefano Zampini /* work[0] = 1_v */ 8071f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 8081f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 8091f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 8101f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 8111f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 8121f4df5f7SStefano Zampini } 8134f1b2e48SStefano Zampini if (nsubs > 1) { 8144f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 8154f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 8164f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 8174f1b2e48SStefano Zampini IS t_zerodiag_subs; 8184f1b2e48SStefano Zampini PetscInt nl; 8194f1b2e48SStefano Zampini 8204f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 8214f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 8224f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 8234f1b2e48SStefano Zampini if (nl) { 8244f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8254f1b2e48SStefano Zampini 8261f4df5f7SStefano Zampini if (pcbddc->current_level) { 8271f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 8281f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 8291f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 8301f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8311f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8321f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8331f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 8341f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 8351f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 8361f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 8371f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 8381f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8391f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8401f4df5f7SStefano Zampini valid = PETSC_FALSE; 8411f4df5f7SStefano Zampini break; 8421f4df5f7SStefano Zampini } 8431f4df5f7SStefano Zampini } 8441f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 8451f4df5f7SStefano Zampini } 8461f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 8471f4df5f7SStefano Zampini IS t_bc; 8481f4df5f7SStefano Zampini PetscInt nzb; 8491f4df5f7SStefano Zampini 8501f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 8511f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 8521f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 8531f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8541f4df5f7SStefano Zampini } 8551f4df5f7SStefano Zampini if (valid && pressures) { 8564f1b2e48SStefano Zampini IS t_pressure_subs; 8574f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 8584f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 8594f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 8604f1b2e48SStefano Zampini } 8614f1b2e48SStefano Zampini if (valid) { 8624f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 8634f1b2e48SStefano Zampini pcbddc->benign_n++; 8644f1b2e48SStefano Zampini } else { 8654f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 8664f1b2e48SStefano Zampini } 8674f1b2e48SStefano Zampini } 8684f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 8694f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 8704f1b2e48SStefano Zampini } 8714f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 8724f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8731f4df5f7SStefano Zampini 8741f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 8751f4df5f7SStefano Zampini PetscInt nzb; 8761f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 8771f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8781f4df5f7SStefano Zampini } 8791f4df5f7SStefano Zampini if (valid && pressures) { 8804f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 8814f1b2e48SStefano Zampini } 8821f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 8831f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 8841f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 8851f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8861f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8871f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8881f4df5f7SStefano Zampini valid = PETSC_FALSE; 8891f4df5f7SStefano Zampini break; 8901f4df5f7SStefano Zampini } 8911f4df5f7SStefano Zampini } 8921f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8931f4df5f7SStefano Zampini } 8944f1b2e48SStefano Zampini if (valid) { 8954f1b2e48SStefano Zampini pcbddc->benign_n = 1; 896ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 8974f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 8984f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 8994f1b2e48SStefano Zampini } 9004f1b2e48SStefano Zampini } 9011f4df5f7SStefano Zampini if (pcbddc->current_level) { 9021f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 9034f1b2e48SStefano Zampini } 9041f4df5f7SStefano Zampini } 9051f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 9064f1b2e48SStefano Zampini 9074f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 908b9b0e38cSStefano Zampini PetscInt n; 909b9b0e38cSStefano Zampini 9104f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9114f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 912b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 913b9b0e38cSStefano Zampini if (n) { 9144f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 9154f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9164f1b2e48SStefano Zampini } 917b9b0e38cSStefano Zampini } 9184f1b2e48SStefano Zampini 9194f1b2e48SStefano Zampini /* final check for null pressures */ 9204f1b2e48SStefano Zampini if (zerodiag && pressures) { 9214f1b2e48SStefano Zampini PetscInt nz,np; 9224f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 9234f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 9244f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 9254f1b2e48SStefano Zampini } 9264f1b2e48SStefano Zampini 9274f1b2e48SStefano Zampini if (recompute_zerodiag) { 9284f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9294f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 9304f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 9314f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 9324f1b2e48SStefano Zampini } else { 9334f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 9344f1b2e48SStefano Zampini 9354f1b2e48SStefano Zampini nzn = 0; 9364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9374f1b2e48SStefano Zampini PetscInt ns; 9384f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9394f1b2e48SStefano Zampini nzn += ns; 9404f1b2e48SStefano Zampini } 9414f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 9424f1b2e48SStefano Zampini nzn = 0; 9434f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9444f1b2e48SStefano Zampini PetscInt ns,*idxs; 9454f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9464f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9474f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 9484f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9494f1b2e48SStefano Zampini nzn += ns; 9504f1b2e48SStefano Zampini } 9514f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 9524f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 9534f1b2e48SStefano Zampini } 9544f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9554f1b2e48SStefano Zampini } 9564f1b2e48SStefano Zampini 957669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 958a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 959a198735bSStefano Zampini Mat A,loc_divudotp; 960a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 961a198735bSStefano Zampini IS row,col,isused = NULL; 962a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 963a198735bSStefano Zampini 9641f4df5f7SStefano Zampini if (pressures) { 9651f4df5f7SStefano Zampini isused = pressures; 9661f4df5f7SStefano Zampini } else { 9671f4df5f7SStefano Zampini isused = zerodiag; 9681f4df5f7SStefano Zampini } 969a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 970669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 9711ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 9721ae86dd6SStefano 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"); 973a198735bSStefano Zampini n_isused = 0; 974a198735bSStefano Zampini if (isused) { 975a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 976a198735bSStefano Zampini } 977a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 978a198735bSStefano Zampini st = st-n_isused; 9791ae86dd6SStefano Zampini if (n) { 980a198735bSStefano Zampini const PetscInt *gidxs; 981a198735bSStefano Zampini 982a198735bSStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 983a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 984a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 985a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 986a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 987a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 9881ae86dd6SStefano Zampini } else { 989a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 990a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 991a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 992a198735bSStefano Zampini } 993a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 994a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 995a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 996a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 997a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 998a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 999a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 1000a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 1001a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 1002a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 1003a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 1004a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 1005a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 1006a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 10071ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10081ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10091ae86dd6SStefano Zampini } 1010b3afcdbeSStefano Zampini 1011b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 10124f1b2e48SStefano Zampini if (has_null_pressures) { 10134f1b2e48SStefano Zampini IS zerodiagc; 10144f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 10154f1b2e48SStefano Zampini PetscInt i,s,*nnz; 10164f1b2e48SStefano Zampini 10174f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 1018339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 1019339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 1020339f8db1SStefano Zampini /* local change of basis for pressures */ 1021339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 102297d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 1023339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 1024339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1025339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 10264f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 10274f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 10284f1b2e48SStefano Zampini PetscInt nzs,j; 10294f1b2e48SStefano Zampini 10304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 10314f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 10324f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 10334f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 10344f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 10354f1b2e48SStefano Zampini } 1036339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 1037339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1038339f8db1SStefano Zampini /* set identity on velocities */ 1039339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 1040339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 1041339f8db1SStefano Zampini } 10424f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 10434f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 10449f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 10454f1b2e48SStefano 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); 1046339f8db1SStefano Zampini /* set change on pressures */ 10474f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 10484f1b2e48SStefano Zampini PetscScalar *array; 10494f1b2e48SStefano Zampini PetscInt nzs; 10504f1b2e48SStefano Zampini 10514f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 10524f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 10534f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 1054339f8db1SStefano Zampini PetscScalar vals[2]; 1055339f8db1SStefano Zampini PetscInt cols[2]; 1056339f8db1SStefano Zampini 1057339f8db1SStefano Zampini cols[0] = idxs[i]; 10584f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 1059339f8db1SStefano Zampini vals[0] = 1.; 1060b0f5fe93SStefano Zampini vals[1] = 1.; 10614f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1062339f8db1SStefano Zampini } 10634f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 10644f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 10654f1b2e48SStefano Zampini array[nzs-1] = 1.; 10664f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 10674f1b2e48SStefano Zampini /* store local idxs for p0 */ 10684f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 10694f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 1070339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 10714f1b2e48SStefano Zampini } 1072339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1073339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1074a3df083aSStefano Zampini /* project if needed */ 1075a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10761dd7afcfSStefano Zampini Mat M; 10771dd7afcfSStefano Zampini 10781dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 1079339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 10801dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 10811dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 1082a3df083aSStefano Zampini } 10834f1b2e48SStefano Zampini /* store global idxs for p0 */ 10844f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1085339f8db1SStefano Zampini } 1086ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 10874f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 1088b0f5fe93SStefano Zampini 1089b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 1090b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 109127b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 109227b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1093339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 1094339f8db1SStefano Zampini PetscFunctionReturn(0); 1095339f8db1SStefano Zampini } 1096339f8db1SStefano Zampini 1097339f8db1SStefano Zampini #undef __FUNCT__ 1098015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 1099015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 1100efc2fbd9SStefano Zampini { 1101efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1102de9d7bd0SStefano Zampini PetscScalar *array; 1103efc2fbd9SStefano Zampini PetscErrorCode ierr; 1104efc2fbd9SStefano Zampini 1105efc2fbd9SStefano Zampini PetscFunctionBegin; 1106efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 1107efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 11084f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1109efc2fbd9SStefano Zampini } 1110de9d7bd0SStefano Zampini if (get) { 1111efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 11124f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 11134f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 1114efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 1115de9d7bd0SStefano Zampini } else { 1116de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 1117de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1118de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1119de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 1120efc2fbd9SStefano Zampini } 1121efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1122efc2fbd9SStefano Zampini } 1123efc2fbd9SStefano Zampini 1124efc2fbd9SStefano Zampini #undef __FUNCT__ 1125c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 1126c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 1127c263805aSStefano Zampini { 1128c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1129c263805aSStefano Zampini PetscErrorCode ierr; 1130c263805aSStefano Zampini 1131c263805aSStefano Zampini PetscFunctionBegin; 1132c263805aSStefano Zampini /* TODO: add error checking 1133c263805aSStefano Zampini - avoid nested pop (or push) calls. 1134c263805aSStefano Zampini - cannot push before pop. 11351c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 1136c263805aSStefano Zampini */ 11374f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1138efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1139efc2fbd9SStefano Zampini } 1140c263805aSStefano Zampini if (pop) { 1141a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 11424f1b2e48SStefano Zampini IS is_p0; 11434f1b2e48SStefano Zampini MatReuse reuse; 1144c263805aSStefano Zampini 1145c263805aSStefano Zampini /* extract B_0 */ 11464f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 11474f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 11484f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 11494f1b2e48SStefano Zampini } 11504f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 11514f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 1152c263805aSStefano Zampini /* remove rows and cols from local problem */ 1153c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 115497d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 11554f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 11564f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 1157a3df083aSStefano Zampini } else { 1158a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1159a3df083aSStefano Zampini PetscScalar *vals; 1160a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 1161a3df083aSStefano Zampini 1162a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 1163a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 1164a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 11650b5adadeSStefano Zampini PetscInt *nnz; 1166a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 1167a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 1168a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1169331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 1170331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1171331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 1172331e053bSStefano Zampini nnz[i] = n - nnz[i]; 1173331e053bSStefano Zampini } 1174331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 1175331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1176331e053bSStefano Zampini } 1177a3df083aSStefano Zampini 1178a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1179a3df083aSStefano Zampini PetscScalar *array; 1180a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 1181a3df083aSStefano Zampini 1182a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 1183a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1184a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1185a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 1186a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1187a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 1188a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 1189a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 1190a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 1191a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 1192a3df083aSStefano Zampini cum = 0; 1193a3df083aSStefano Zampini for (j=0;j<n;j++) { 119422db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 1195a3df083aSStefano Zampini vals[cum] = array[j]; 1196a3df083aSStefano Zampini idxs_ins[cum] = j; 1197a3df083aSStefano Zampini cum++; 1198a3df083aSStefano Zampini } 1199a3df083aSStefano Zampini } 1200a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 1201a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 1202a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1203a3df083aSStefano Zampini } 1204a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1205a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1206a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 1207a3df083aSStefano Zampini } 1208c263805aSStefano Zampini } else { /* push */ 1209a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 12104f1b2e48SStefano Zampini PetscInt i; 12114f1b2e48SStefano Zampini 12124f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 12134f1b2e48SStefano Zampini PetscScalar *B0_vals; 12144f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 12154f1b2e48SStefano Zampini 12164f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 12174f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 12187b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 12194f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 12204f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 12214f1b2e48SStefano Zampini } 1222c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1223c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1224a3df083aSStefano Zampini } else { 1225a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 1226a3df083aSStefano Zampini } 1227c263805aSStefano Zampini } 1228c263805aSStefano Zampini PetscFunctionReturn(0); 1229c263805aSStefano Zampini } 1230c263805aSStefano Zampini 1231c263805aSStefano Zampini #undef __FUNCT__ 1232b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 123308122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 1234b1b3d7a2SStefano Zampini { 1235b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 123608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 123708122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 123808122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 123908122e43SStefano Zampini PetscScalar *work,lwork; 124008122e43SStefano Zampini PetscScalar *St,*S,*eigv; 124108122e43SStefano Zampini PetscScalar *Sarray,*Starray; 124208122e43SStefano Zampini PetscReal *eigs,thresh; 12431b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 1244f6f667cfSStefano Zampini PetscBool allocated_S_St; 124508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 124608122e43SStefano Zampini PetscReal *rwork; 124708122e43SStefano Zampini #endif 1248b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1249b1b3d7a2SStefano Zampini 1250b1b3d7a2SStefano Zampini PetscFunctionBegin; 1251b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 1252af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 1253af25d912SStefano 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); 125406a4e24aSStefano Zampini 1255fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1256fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1257fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1258fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 12591575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1260fd14bc51SStefano Zampini } 1261fd14bc51SStefano Zampini 1262e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 1263e496cd5dSStefano 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); 1264e496cd5dSStefano Zampini } 1265e496cd5dSStefano Zampini 126608122e43SStefano Zampini /* max size of subsets */ 126708122e43SStefano Zampini mss = 0; 126808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 126908122e43SStefano Zampini PetscInt subset_size; 1270862806e4SStefano Zampini 127108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 127208122e43SStefano Zampini mss = PetscMax(mss,subset_size); 127308122e43SStefano Zampini } 127408122e43SStefano Zampini 127508122e43SStefano Zampini /* min/max and threshold */ 127608122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 1277f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 127808122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 1279f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 1280f6f667cfSStefano Zampini if (nmin) { 1281f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 1282f6f667cfSStefano Zampini } 128308122e43SStefano Zampini 128408122e43SStefano Zampini /* allocate lapack workspace */ 128508122e43SStefano Zampini cum = cum2 = 0; 128608122e43SStefano Zampini maxneigs = 0; 128708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 128808122e43SStefano Zampini PetscInt n,subset_size; 1289f6f667cfSStefano Zampini 129008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 129108122e43SStefano Zampini n = PetscMin(subset_size,nmax); 12929162d606SStefano Zampini cum += subset_size; 12939162d606SStefano Zampini cum2 += subset_size*n; 129408122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 129508122e43SStefano Zampini } 129608122e43SStefano Zampini if (mss) { 12979ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 129808122e43SStefano Zampini PetscBLASInt B_itype = 1; 129908122e43SStefano Zampini PetscBLASInt B_N = mss; 13004c6709b3SStefano Zampini PetscReal zero = 0.0; 13014c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 130208122e43SStefano Zampini 130308122e43SStefano Zampini B_lwork = -1; 130408122e43SStefano Zampini S = NULL; 130508122e43SStefano Zampini St = NULL; 1306a58a30b4SStefano Zampini eigs = NULL; 1307a58a30b4SStefano Zampini eigv = NULL; 1308a58a30b4SStefano Zampini B_iwork = NULL; 1309a58a30b4SStefano Zampini B_ifail = NULL; 1310d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1311d1710679SStefano Zampini rwork = NULL; 1312d1710679SStefano Zampini #endif 13138bec7fa6SStefano Zampini thresh = 1.0; 131408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 131508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 131608122e43SStefano 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)); 131708122e43SStefano Zampini #else 131808122e43SStefano 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)); 131908122e43SStefano Zampini #endif 132008122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 132108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 132208122e43SStefano Zampini } else { 132308122e43SStefano Zampini /* TODO */ 132408122e43SStefano Zampini } 132508122e43SStefano Zampini } else { 132608122e43SStefano Zampini lwork = 0; 132708122e43SStefano Zampini } 132808122e43SStefano Zampini 132908122e43SStefano Zampini nv = 0; 1330d62866d3SStefano 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) */ 1331d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 133208122e43SStefano Zampini } 13334c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1334f6f667cfSStefano Zampini if (allocated_S_St) { 1335f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1336f6f667cfSStefano Zampini } 1337f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 133808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 133908122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 134008122e43SStefano Zampini #endif 13419162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 13429162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 13439162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 134408122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 13459162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 134608122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 134708122e43SStefano Zampini 134808122e43SStefano Zampini maxneigs = 0; 134972b8c272SStefano Zampini cum = cumarray = 0; 13509162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 13519162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1352d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 135308122e43SStefano Zampini const PetscInt *idxs; 135408122e43SStefano Zampini 1355d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 135608122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 135708122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 135808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 135908122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 13609162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 13619162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 136208122e43SStefano Zampini } 1363d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 136408122e43SStefano Zampini } 136508122e43SStefano Zampini 136608122e43SStefano Zampini if (mss) { /* multilevel */ 136708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 136808122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 136908122e43SStefano Zampini } 137008122e43SStefano Zampini 1371ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 137208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 137308122e43SStefano Zampini const PetscInt *idxs; 13749d54b7f4SStefano Zampini PetscReal upper,lower; 1375862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 137608122e43SStefano Zampini PetscBLASInt B_N; 1377aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 137808122e43SStefano Zampini 13799d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 13809d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 13819d54b7f4SStefano Zampini lower = thresh; 13829d54b7f4SStefano Zampini } else { 13839d54b7f4SStefano Zampini upper = 1./thresh; 13849d54b7f4SStefano Zampini lower = 0.; 13859d54b7f4SStefano Zampini } 1386862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1387ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1388f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1389f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 13909ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1391aff50787SStefano Zampini PetscInt j,k; 1392aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1393aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1394aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 139508122e43SStefano Zampini } 139608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1397aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1398aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1399aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1400aff50787SStefano Zampini } 140108122e43SStefano Zampini } 140208122e43SStefano Zampini } else { 140308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 140408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 140508122e43SStefano Zampini } 14068bec7fa6SStefano Zampini } else { 1407f6f667cfSStefano Zampini S = Sarray + cumarray; 1408f6f667cfSStefano Zampini St = Starray + cumarray; 14098bec7fa6SStefano Zampini } 1410aff50787SStefano Zampini /* see if we can save some work */ 1411b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1412aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1413aff50787SStefano Zampini } 1414aff50787SStefano Zampini 1415b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1416aff50787SStefano Zampini B_neigs = 0; 1417aff50787SStefano Zampini } else { 14189ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 141908122e43SStefano Zampini PetscBLASInt B_itype = 1; 1420f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 14214c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 14229552c7c7SStefano Zampini PetscInt nmin_s; 1423b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 142408122e43SStefano Zampini 1425fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 14268bec7fa6SStefano 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]]); 1427fd14bc51SStefano Zampini } 1428d16cbb6bSStefano Zampini 1429b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1430b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1431b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1432b7ab4a40SStefano Zampini } 1433b7ab4a40SStefano Zampini 143408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1435b7ab4a40SStefano Zampini if (compute_range) { 1436d16cbb6bSStefano Zampini 1437d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 143808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 14399d54b7f4SStefano 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)); 144008122e43SStefano Zampini #else 14419d54b7f4SStefano 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)); 144208122e43SStefano Zampini #endif 1443b7ab4a40SStefano Zampini } else if (!same_data) { 1444d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1445d16cbb6bSStefano Zampini B_IL = 1; 1446d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 14479d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1448d16cbb6bSStefano Zampini #else 14499d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1450d16cbb6bSStefano Zampini #endif 1451b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1452b7ab4a40SStefano Zampini PetscInt k; 1453b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1454b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1455b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1456b7ab4a40SStefano Zampini nmin = nmax; 1457b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1458b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1459b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1460b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1461b7ab4a40SStefano Zampini } 1462d16cbb6bSStefano Zampini } 146308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 146408122e43SStefano Zampini if (B_ierr) { 14656c4ed002SBarry 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); 14666c4ed002SBarry 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); 14676c4ed002SBarry 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); 146808122e43SStefano Zampini } 146908122e43SStefano Zampini 147008122e43SStefano Zampini if (B_neigs > nmax) { 1471fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1472fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1473fd14bc51SStefano Zampini } 14749d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 147508122e43SStefano Zampini B_neigs = nmax; 147608122e43SStefano Zampini } 147708122e43SStefano Zampini 14789552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 14799552c7c7SStefano Zampini if (B_neigs < nmin_s) { 148008122e43SStefano Zampini PetscBLASInt B_neigs2; 148108122e43SStefano Zampini 14829d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1483f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 14849d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 14859d54b7f4SStefano Zampini } else { 14869d54b7f4SStefano Zampini B_IL = B_neigs + 1; 14879d54b7f4SStefano Zampini B_IU = nmin_s; 14889d54b7f4SStefano Zampini } 1489fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1490fd14bc51SStefano 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); 1491fd14bc51SStefano Zampini } 14929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 14931ae86dd6SStefano Zampini PetscInt j,k; 149408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 14951ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 14961ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 14971ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 149808122e43SStefano Zampini } 149908122e43SStefano Zampini } 150008122e43SStefano Zampini } else { 150108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 150208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 150308122e43SStefano Zampini } 150408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 150508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 15069d54b7f4SStefano 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)); 150708122e43SStefano Zampini #else 15089d54b7f4SStefano 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)); 150908122e43SStefano Zampini #endif 151008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 151108122e43SStefano Zampini B_neigs += B_neigs2; 151208122e43SStefano Zampini } 151308122e43SStefano Zampini if (B_ierr) { 15146c4ed002SBarry 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); 15156c4ed002SBarry 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); 15166c4ed002SBarry 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); 151708122e43SStefano Zampini } 1518fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1519ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 152008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 152108122e43SStefano Zampini if (eigs[j] == 0.0) { 1522ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 152308122e43SStefano Zampini } else { 15249d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1525ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 15269d54b7f4SStefano Zampini } else { 15279d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 15289d54b7f4SStefano Zampini } 1529fd14bc51SStefano Zampini } 153008122e43SStefano Zampini } 153108122e43SStefano Zampini } 153208122e43SStefano Zampini } else { 153308122e43SStefano Zampini /* TODO */ 153408122e43SStefano Zampini } 1535aff50787SStefano Zampini } 15366c3e6151SStefano Zampini /* change the basis back to the original one */ 15376c3e6151SStefano Zampini if (sub_schurs->change) { 153872b8c272SStefano Zampini Mat change,phi,phit; 15396c3e6151SStefano Zampini 15406c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 15416c3e6151SStefano Zampini PetscInt ii; 15426c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 15436c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15446c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 15456c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 15466c3e6151SStefano Zampini } 15476c3e6151SStefano Zampini } 15486c3e6151SStefano Zampini } 154972b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 15506c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 155172b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 15526c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 15536c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 15546c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 15556c3e6151SStefano Zampini } 15568bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 15578bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 15589162d606SStefano Zampini if (B_neigs) { 15599162d606SStefano 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); 1560fd14bc51SStefano Zampini 1561fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 15629552c7c7SStefano Zampini PetscInt ii; 15639552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1564ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15659552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1566ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1567ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1568ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1569ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1570ac47001eSStefano Zampini #else 1571ac47001eSStefano 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); 1572ac47001eSStefano Zampini #endif 15739552c7c7SStefano Zampini } 15749552c7c7SStefano Zampini } 1575fd14bc51SStefano Zampini } 15769162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 15779162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 15789162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 15799162d606SStefano Zampini cum++; 158008122e43SStefano Zampini } 158108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 158208122e43SStefano Zampini /* shift for next computation */ 158308122e43SStefano Zampini cumarray += subset_size*subset_size; 158408122e43SStefano Zampini } 1585fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1586fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1587fd14bc51SStefano Zampini } 158808122e43SStefano Zampini 158908122e43SStefano Zampini if (mss) { 159008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 159108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1592f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1593f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1594f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 159508122e43SStefano Zampini } 1596f6f667cfSStefano Zampini if (allocated_S_St) { 1597f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1598f6f667cfSStefano Zampini } 1599f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 160008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 160108122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 160208122e43SStefano Zampini #endif 160308122e43SStefano Zampini if (pcbddc->dbg_flag) { 16041b968477SStefano Zampini PetscInt maxneigs_r; 1605b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 16069b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 160708122e43SStefano Zampini } 160808122e43SStefano Zampini PetscFunctionReturn(0); 160908122e43SStefano Zampini } 1610b1b3d7a2SStefano Zampini 1611674ae819SStefano Zampini #undef __FUNCT__ 1612c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1613c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1614c8587f34SStefano Zampini { 16158629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1616c8587f34SStefano Zampini PetscErrorCode ierr; 1617c8587f34SStefano Zampini 1618c8587f34SStefano Zampini PetscFunctionBegin; 1619f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 16205e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1621c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1622c8587f34SStefano Zampini 1623684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 16240fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1625684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1626c8587f34SStefano Zampini 16278629588bSStefano Zampini /* 16288629588bSStefano Zampini Setup local correction and local part of coarse basis. 16298629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 16308629588bSStefano Zampini */ 163147f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 16328629588bSStefano Zampini 16338629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 16348629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 16358629588bSStefano Zampini 16368629588bSStefano Zampini /* free */ 16378629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1638c8587f34SStefano Zampini PetscFunctionReturn(0); 1639c8587f34SStefano Zampini } 1640c8587f34SStefano Zampini 1641c8587f34SStefano Zampini #undef __FUNCT__ 1642674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1643674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1644674ae819SStefano Zampini { 1645674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1646674ae819SStefano Zampini PetscErrorCode ierr; 1647674ae819SStefano Zampini 1648674ae819SStefano Zampini PetscFunctionBegin; 1649674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1650674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 165130368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1652674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1653785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1654674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1655f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1656f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1657785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 165863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 165963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1660674ae819SStefano Zampini PetscFunctionReturn(0); 1661674ae819SStefano Zampini } 1662674ae819SStefano Zampini 1663674ae819SStefano Zampini #undef __FUNCT__ 1664674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1665674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1666674ae819SStefano Zampini { 1667674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16684f1b2e48SStefano Zampini PetscInt i; 1669674ae819SStefano Zampini PetscErrorCode ierr; 1670674ae819SStefano Zampini 1671674ae819SStefano Zampini PetscFunctionBegin; 1672b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1673674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 167416909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 16751dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1676674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1677669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 1678fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 1679674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 168071582508SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 16814f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 16824f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 16834f1b2e48SStefano Zampini } 16844f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1685b334f244SStefano Zampini if (pcbddc->sub_schurs) { 1686b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1687b334f244SStefano Zampini } 1688674ae819SStefano Zampini PetscFunctionReturn(0); 1689674ae819SStefano Zampini } 1690674ae819SStefano Zampini 1691674ae819SStefano Zampini #undef __FUNCT__ 1692674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1693674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1694674ae819SStefano Zampini { 1695674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1696674ae819SStefano Zampini PetscErrorCode ierr; 1697674ae819SStefano Zampini 1698674ae819SStefano Zampini PetscFunctionBegin; 1699674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 170058da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1701ca92afb2SStefano Zampini PetscScalar *array; 170206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 170306656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 170458da7f69SStefano Zampini } 1705674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1706674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 170715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 170815aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1709674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1710674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1711674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 171206656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1713674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1714674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17158ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1716674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1717674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1718674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1719f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1720f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1721f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1722f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1723727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 17240e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1725f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 172670cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 172781d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 17280369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 17291dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 17304f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 17318b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1732ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1733ca92afb2SStefano Zampini PetscInt i; 1734ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1735ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1736ca92afb2SStefano Zampini } 1737ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1738ca92afb2SStefano Zampini } 17394f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1740674ae819SStefano Zampini PetscFunctionReturn(0); 1741674ae819SStefano Zampini } 1742674ae819SStefano Zampini 1743674ae819SStefano Zampini #undef __FUNCT__ 1744f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1745f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 17466bfb1811SStefano Zampini { 17476bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 17486bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 17496bfb1811SStefano Zampini VecType impVecType; 17504f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 17516bfb1811SStefano Zampini PetscErrorCode ierr; 17526bfb1811SStefano Zampini 17536bfb1811SStefano Zampini PetscFunctionBegin; 17546c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1755e7b262bdSStefano Zampini /* get sizes */ 17564f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1757b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 17586bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1759e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1760e7b262bdSStefano Zampini /* R nodes */ 1761e7b262bdSStefano Zampini old_size = -1; 1762e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1763e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1764e7b262bdSStefano Zampini } 1765e7b262bdSStefano Zampini if (n_R != old_size) { 1766e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1767e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17686bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 17696bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 17706bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 17716bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1772e7b262bdSStefano Zampini } 1773e7b262bdSStefano Zampini /* local primal dofs */ 1774e7b262bdSStefano Zampini old_size = -1; 1775e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1776e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1777e7b262bdSStefano Zampini } 1778e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1779e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 178083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1781e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 17826bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1783e7b262bdSStefano Zampini } 1784e7b262bdSStefano Zampini /* local explicit constraints */ 1785e7b262bdSStefano Zampini old_size = -1; 1786e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1787e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1788e7b262bdSStefano Zampini } 1789e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1790e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 179183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 179283b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 179383b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 179483b7ccabSStefano Zampini } 17956bfb1811SStefano Zampini PetscFunctionReturn(0); 17966bfb1811SStefano Zampini } 17976bfb1811SStefano Zampini 17986bfb1811SStefano Zampini #undef __FUNCT__ 179947f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 180047f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 180188ebb749SStefano Zampini { 180225084f0cSStefano Zampini PetscErrorCode ierr; 180325084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 180488ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 180588ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1806d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 180725084f0cSStefano Zampini /* submatrices of local problem */ 180880677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 180906656605SStefano Zampini /* submatrices of local coarse problem */ 181006656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 181125084f0cSStefano Zampini /* working matrices */ 181206656605SStefano Zampini Mat C_CR; 181325084f0cSStefano Zampini /* additional working stuff */ 181406656605SStefano Zampini PC pc_R; 18154f1b2e48SStefano Zampini Mat F; 18165cbda25cSStefano Zampini Vec dummy_vec; 1817a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 181825084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 181906656605SStefano Zampini PetscScalar *work; 182006656605SStefano Zampini PetscInt *idx_V_B; 1821ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 182206656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1823ffd830a3SStefano Zampini 182425084f0cSStefano Zampini /* some shortcuts to scalars */ 182506656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 182688ebb749SStefano Zampini 182788ebb749SStefano Zampini PetscFunctionBegin; 18289a962809SStefano 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"); 1829ffd830a3SStefano Zampini 1830ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1831b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 18324f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1833b371cd4fSStefano Zampini n_B = pcis->n_B; 1834b371cd4fSStefano Zampini n_D = pcis->n - n_B; 183588ebb749SStefano Zampini n_R = pcis->n - n_vertices; 183688ebb749SStefano Zampini 183788ebb749SStefano Zampini /* vertices in boundary numbering */ 1838785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 18390e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 18406c4ed002SBarry 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); 184188ebb749SStefano Zampini 184206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1843019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 184406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 184506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 184606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 184706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 184806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 184906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 185006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 185106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 185206656605SStefano Zampini 185306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 185406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 185506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 185606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 185706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1858ffd830a3SStefano Zampini lda_rhs = n_R; 1859a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 186006656605SStefano Zampini if (isLU || isILU || isCHOL) { 186106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1862b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 1863df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1864d62866d3SStefano Zampini MatFactorType type; 1865d62866d3SStefano Zampini 1866df4d28bfSStefano Zampini F = reuse_solver->F; 18676816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1868d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1869ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 187022db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 187106656605SStefano Zampini } else { 187206656605SStefano Zampini F = NULL; 187306656605SStefano Zampini } 187406656605SStefano Zampini 1875ffd830a3SStefano Zampini /* allocate workspace */ 1876ffd830a3SStefano Zampini n = 0; 1877ffd830a3SStefano Zampini if (n_constraints) { 1878ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1879ffd830a3SStefano Zampini } 1880ffd830a3SStefano Zampini if (n_vertices) { 1881ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1882ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1883ffd830a3SStefano Zampini } 1884ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1885ffd830a3SStefano Zampini 18865cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 18875cbda25cSStefano Zampini dummy_vec = NULL; 18885cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 18895cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 18905cbda25cSStefano Zampini } 18915cbda25cSStefano Zampini 189288ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 189388ebb749SStefano Zampini if (n_constraints) { 189472b8c272SStefano Zampini Mat M1,M2,M3,C_B; 189506656605SStefano Zampini IS is_aux; 189680677318SStefano Zampini PetscScalar *array,*array2; 189706656605SStefano Zampini 1898f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 189980677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 190088ebb749SStefano Zampini 190125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 190225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 19038ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 190472b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 190588ebb749SStefano Zampini 190680677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 190780677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1908ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 190988ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 191006656605SStefano Zampini const PetscScalar *row_cmat_values; 191106656605SStefano Zampini const PetscInt *row_cmat_indices; 191206656605SStefano Zampini PetscInt size_of_constraint,j; 191388ebb749SStefano Zampini 191406656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 191506656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1916ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 191706656605SStefano Zampini } 191806656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 191906656605SStefano Zampini } 1920ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 192106656605SStefano Zampini if (F) { 192206656605SStefano Zampini Mat B; 192306656605SStefano Zampini 1924ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1925a3df083aSStefano Zampini if (need_benign_correction) { 1926df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1927a3df083aSStefano Zampini 192872b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 192972b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1930a3df083aSStefano Zampini } 193180677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1932a3df083aSStefano Zampini if (need_benign_correction) { 1933a3df083aSStefano Zampini PetscScalar *marr; 1934df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1935a3df083aSStefano Zampini 1936a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 19375cbda25cSStefano Zampini if (lda_rhs != n_R) { 19385cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 19395cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19405cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19415cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19425cbda25cSStefano Zampini } 19435cbda25cSStefano Zampini } else { 1944a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1945a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19465cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1947a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1948a3df083aSStefano Zampini } 19495cbda25cSStefano Zampini } 1950a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1951a3df083aSStefano Zampini } 195206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195306656605SStefano Zampini } else { 195480677318SStefano Zampini PetscScalar *marr; 195580677318SStefano Zampini 195680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 195706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1958ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1959ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 196006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 196106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 196206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 196306656605SStefano Zampini } 196480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 196506656605SStefano Zampini } 196680677318SStefano Zampini if (!pcbddc->switch_static) { 196780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 196880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 196980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 197080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1971ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 197280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 197380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 197680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 197780677318SStefano Zampini } 197880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 197980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 198072b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 198180677318SStefano Zampini } else { 1982ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1983ffd830a3SStefano Zampini IS dummy; 1984ffd830a3SStefano Zampini 1985ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 198672b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1987ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1988ffd830a3SStefano Zampini } else { 198980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 199080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1991ffd830a3SStefano Zampini } 199225084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 199380677318SStefano Zampini } 199480677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 199580677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 199680677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 199706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 199806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 199980677318SStefano Zampini if (isCHOL) { 200080677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 200180677318SStefano Zampini } else { 200225084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 200380677318SStefano Zampini } 200480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 200506656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 200625084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 200725084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 200825084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 200980677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 201072b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 201172b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 201206656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 201306656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 2014f4ddd8eeSStefano Zampini } 2015fc227af8SStefano Zampini 2016fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 201788ebb749SStefano Zampini if (n_vertices) { 201806656605SStefano Zampini IS is_aux; 20193a50541eSStefano Zampini 2020b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 20216816873aSStefano Zampini IS tis; 20226816873aSStefano Zampini 20236816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 20246816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 20256816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 20266816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 20276816873aSStefano Zampini } else { 20283a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 20296816873aSStefano Zampini } 20309577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 20319577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 203204708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 203325084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 203488ebb749SStefano Zampini } 203588ebb749SStefano Zampini 203688ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 2037f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 203806656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 203906656605SStefano Zampini if (pcbddc->coarse_phi_D) { 204006656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 204106656605SStefano Zampini } 2042f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 204306656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 204406656605SStefano Zampini PetscScalar *marray; 204506656605SStefano Zampini 204606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 204706656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 2048f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2049f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 2050f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 2051f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2052f4ddd8eeSStefano Zampini } 2053f4ddd8eeSStefano Zampini } 205406656605SStefano Zampini 2055f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 205606656605SStefano Zampini PetscScalar *marray; 205788ebb749SStefano Zampini 205806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20598eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 206006656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 206188ebb749SStefano Zampini } 20623301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 206306656605SStefano Zampini n *= 2; 206488ebb749SStefano Zampini } 206506656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 206606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 206706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20688eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 206906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 207006656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 207188ebb749SStefano Zampini } 20723301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 207306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 20748eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 207506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 207606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 207788ebb749SStefano Zampini } 207888ebb749SStefano Zampini } else { 2079c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 2080c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 20811b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2082c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 2083c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 2084c0553b1fSStefano Zampini } 208588ebb749SStefano Zampini } 208606656605SStefano Zampini } 2087019a44ceSStefano Zampini 208806656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 20894f1b2e48SStefano Zampini p0_lidx_I = NULL; 20904f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 2091d12edf2fSStefano Zampini const PetscInt *idxs; 2092d12edf2fSStefano Zampini 2093d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 20944f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 20954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20964f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 20974f1b2e48SStefano Zampini } 2098d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 2099d12edf2fSStefano Zampini } 2100d16cbb6bSStefano Zampini 210106656605SStefano Zampini /* vertices */ 210206656605SStefano Zampini if (n_vertices) { 210316f15bc4SStefano Zampini 2104af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 210504708bb6SStefano Zampini 210616f15bc4SStefano Zampini if (n_R) { 210714393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 210806656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 210916f15bc4SStefano Zampini PetscScalar *x,*y; 211004708bb6SStefano Zampini PetscBool isseqaij; 211106656605SStefano Zampini 211221eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 211314393ed6SStefano Zampini if (need_benign_correction) { 211414393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 211514393ed6SStefano Zampini IS is_p0; 211614393ed6SStefano Zampini PetscInt *idxs_p0,n; 211714393ed6SStefano Zampini 211814393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 211914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 212014393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2121af25d912SStefano 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); 212214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 212314393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 212414393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 212514393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 212614393ed6SStefano Zampini } 212714393ed6SStefano Zampini 2128ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2129af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2130ffd830a3SStefano Zampini } else { 2131ca92afb2SStefano Zampini PetscScalar *av,*array; 2132ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2133ca92afb2SStefano Zampini PetscInt n; 2134ca92afb2SStefano Zampini PetscBool flg_row; 2135ffd830a3SStefano Zampini 2136ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2137ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 21389d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2139ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2140ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2141ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2142ca92afb2SStefano Zampini PetscInt j; 2143ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2144ffd830a3SStefano Zampini } 2145ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2146ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2147ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2148ffd830a3SStefano Zampini } 2149ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2150a3df083aSStefano Zampini if (need_benign_correction) { 2151df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2152a3df083aSStefano Zampini PetscScalar *marr; 2153a3df083aSStefano Zampini 2154a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 215514393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 215614393ed6SStefano Zampini 215714393ed6SStefano Zampini | 0 0 0 | (V) 215814393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 215914393ed6SStefano Zampini | 0 0 -1 | (p0) 216014393ed6SStefano Zampini 216114393ed6SStefano Zampini */ 2162df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 216314393ed6SStefano Zampini const PetscScalar *vals; 216414393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 216514393ed6SStefano Zampini PetscInt n,j,nz; 216614393ed6SStefano Zampini 2167df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2168df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 216914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 217014393ed6SStefano Zampini for (j=0;j<n;j++) { 217114393ed6SStefano Zampini PetscScalar val = vals[j]; 217214393ed6SStefano Zampini PetscInt k,col = idxs[j]; 217314393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 217414393ed6SStefano Zampini } 217514393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2176df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 217714393ed6SStefano Zampini } 217872b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 217972b8c272SStefano Zampini } 218072b8c272SStefano Zampini if (F) { 218114393ed6SStefano Zampini /* need to correct the rhs */ 218272b8c272SStefano Zampini if (need_benign_correction) { 218372b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 218472b8c272SStefano Zampini PetscScalar *marr; 218572b8c272SStefano Zampini 218672b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 21875cbda25cSStefano Zampini if (lda_rhs != n_R) { 21885cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21895cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21905cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 21915cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21925cbda25cSStefano Zampini } 21935cbda25cSStefano Zampini } else { 2194a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2195a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21965cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2197a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2198a3df083aSStefano Zampini } 21995cbda25cSStefano Zampini } 2200a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2201a3df083aSStefano Zampini } 220206656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 220314393ed6SStefano Zampini /* need to correct the solution */ 2204a3df083aSStefano Zampini if (need_benign_correction) { 2205df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2206a3df083aSStefano Zampini PetscScalar *marr; 2207a3df083aSStefano Zampini 2208a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 22095cbda25cSStefano Zampini if (lda_rhs != n_R) { 22105cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 22115cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 22125cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 22135cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 22145cbda25cSStefano Zampini } 22155cbda25cSStefano Zampini } else { 2216a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2217a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 22185cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2219a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2220a3df083aSStefano Zampini } 22215cbda25cSStefano Zampini } 2222a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2223a3df083aSStefano Zampini } 222406656605SStefano Zampini } else { 222506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 222606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2227ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2228ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 222906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 223006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 223106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 223206656605SStefano Zampini } 223306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 223406656605SStefano Zampini } 223580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2236ffd830a3SStefano Zampini /* S_VV and S_CV */ 223706656605SStefano Zampini if (n_constraints) { 223806656605SStefano Zampini Mat B; 223980677318SStefano Zampini 2240ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 224180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2242ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2243ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 224480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 224580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 224680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 224780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 224880677318SStefano Zampini } 2249ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 225080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 225180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2252ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 225380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 225406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2255ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2256ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 225706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 225806656605SStefano Zampini } 225904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 226004708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2261511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 226204708bb6SStefano Zampini } 2263ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2264ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2265ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2266ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2267ffd830a3SStefano Zampini } 226806656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 226914393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 227014393ed6SStefano Zampini if (need_benign_correction) { 2271df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 227214393ed6SStefano Zampini PetscScalar *marr,*sums; 227314393ed6SStefano Zampini 227414393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 227514393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2276df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 227714393ed6SStefano Zampini const PetscScalar *vals; 227814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 227914393ed6SStefano Zampini PetscInt n,j,nz; 228014393ed6SStefano Zampini 2281df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2282df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 228314393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 228414393ed6SStefano Zampini PetscInt k; 228514393ed6SStefano Zampini sums[j] = 0.; 228614393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 228714393ed6SStefano Zampini } 228814393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 228914393ed6SStefano Zampini for (j=0;j<n;j++) { 229014393ed6SStefano Zampini PetscScalar val = vals[j]; 229114393ed6SStefano Zampini PetscInt k; 229214393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 229314393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 229414393ed6SStefano Zampini } 229514393ed6SStefano Zampini } 229614393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2297df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 229814393ed6SStefano Zampini } 229914393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 230014393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 230114393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 230214393ed6SStefano Zampini } 230380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 230406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 230506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 230606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 230706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 230806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 230906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 231006656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2311d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2312019a44ceSStefano Zampini } else { 2313d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2314d16cbb6bSStefano Zampini } 231521eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2316d16cbb6bSStefano Zampini 231706656605SStefano Zampini /* coarse basis functions */ 231806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 231916f15bc4SStefano Zampini PetscScalar *y; 232016f15bc4SStefano Zampini 2321ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 232206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 232306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 232406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 232706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 232806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 232906656605SStefano Zampini 233006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23314f1b2e48SStefano Zampini PetscInt j; 23324f1b2e48SStefano Zampini 233306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 233406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 233506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23384f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 233906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 234006656605SStefano Zampini } 234106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 234206656605SStefano Zampini } 234304708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 234404708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 234506656605SStefano Zampini } 23465cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 234706656605SStefano Zampini 234806656605SStefano Zampini if (n_constraints) { 234906656605SStefano Zampini Mat B; 235006656605SStefano Zampini 2351ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 235206656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 235380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 235406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 235506656605SStefano Zampini if (n_vertices) { 235680677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 235780677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 235880677318SStefano Zampini } else { 235980677318SStefano Zampini Mat S_VCt; 236080677318SStefano Zampini 2361ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2362ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 236372b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2364ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2365ffd830a3SStefano Zampini } 236680677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 236780677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 236880677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 236980677318SStefano Zampini } 237006656605SStefano Zampini } 237106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 237206656605SStefano Zampini /* coarse basis functions */ 237306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 237406656605SStefano Zampini PetscScalar *y; 237506656605SStefano Zampini 2376ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 237706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 237806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 237906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 238206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 238306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23844f1b2e48SStefano Zampini PetscInt j; 23854f1b2e48SStefano Zampini 238606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 238706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 238806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23914f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 239206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 239306656605SStefano Zampini } 239406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 239506656605SStefano Zampini } 239606656605SStefano Zampini } 239780677318SStefano Zampini if (n_constraints) { 239880677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 239980677318SStefano Zampini } 24004f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 240172b8c272SStefano Zampini 240272b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 240372b8c272SStefano Zampini if (pcbddc->benign_n) { 240472b8c272SStefano Zampini Mat B0_B,B0_BPHI; 240572b8c272SStefano Zampini IS is_dummy; 240672b8c272SStefano Zampini PetscScalar *data; 240772b8c272SStefano Zampini PetscInt j; 240872b8c272SStefano Zampini 240972b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 241072b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 241172b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 241272b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 241386c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 241472b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 241572b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 241672b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 241772b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 241872b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 241972b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 242072b8c272SStefano Zampini } 242172b8c272SStefano Zampini } 242272b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 242372b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 242472b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 242572b8c272SStefano Zampini } 2426019a44ceSStefano Zampini 242706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 24283301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2429ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2430ffd830a3SStefano Zampini PetscScalar *marray; 243106656605SStefano Zampini 243206656605SStefano Zampini if (n_constraints) { 2433ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 243406656605SStefano Zampini 2435af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 243606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2437ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 243816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 243906656605SStefano Zampini if (n_vertices) { 2440ffd830a3SStefano Zampini Mat S_VCT; 244106656605SStefano Zampini 244206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2443ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 244416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 244506656605SStefano Zampini } 2446ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 24475b782168SStefano Zampini } else { 24485b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 244906656605SStefano Zampini } 245016f15bc4SStefano Zampini if (n_vertices && n_R) { 2451ffd830a3SStefano Zampini PetscScalar *av,*marray; 2452ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2453ffd830a3SStefano Zampini PetscInt n; 2454ffd830a3SStefano Zampini PetscBool flg_row; 245506656605SStefano Zampini 2456ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2457af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2458ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2459ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2460ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2461ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2462ffd830a3SStefano Zampini PetscInt j; 2463ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2464ffd830a3SStefano Zampini } 2465ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2466ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2467ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 246806656605SStefano Zampini } 246906656605SStefano Zampini 2470ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2471ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2472ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2473ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2474ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 247506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 247606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 247706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 247806656605SStefano Zampini } 2479ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 24805b782168SStefano Zampini if (B_C) { 2481ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2482ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2483ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2484ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2485ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2486ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2487ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 248806656605SStefano Zampini } 2489ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 24905b782168SStefano Zampini } 249106656605SStefano Zampini /* coarse basis functions */ 249206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 249306656605SStefano Zampini PetscScalar *y; 249406656605SStefano Zampini 2495ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 249606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 249706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 249806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 249906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250006656605SStefano Zampini if (i<n_vertices) { 250106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 250206656605SStefano Zampini } 250306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 250406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 250506656605SStefano Zampini 250606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 250706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 250806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 250906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 251206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 251306656605SStefano Zampini } 251406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 251506656605SStefano Zampini } 2516ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2517ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 251806656605SStefano Zampini } 2519d62866d3SStefano Zampini /* free memory */ 252088ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 252106656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 252206656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 252306656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 252406656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2525d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2526d62866d3SStefano Zampini if (n_vertices) { 2527d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2528d62866d3SStefano Zampini } 2529d62866d3SStefano Zampini if (n_constraints) { 2530d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2531d62866d3SStefano Zampini } 253288ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 253388ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 253488ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2535d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 253688ebb749SStefano Zampini Mat coarse_sub_mat; 253725084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 253888ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 253988ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 254088ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 25418bec7fa6SStefano Zampini Mat C_B,CPHI; 25428bec7fa6SStefano Zampini IS is_dummy; 25438bec7fa6SStefano Zampini Vec mones; 254488ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 254588ebb749SStefano Zampini PetscReal real_value; 254688ebb749SStefano Zampini 2547a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2548a3df083aSStefano Zampini Mat A; 2549a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2550a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2551a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2552a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2553a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2554a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2555a3df083aSStefano Zampini } else { 255688ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 255788ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 255888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 255988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2560a3df083aSStefano Zampini } 256188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 256288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2563ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 256488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 256588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 256688ebb749SStefano Zampini } 256788ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 256888ebb749SStefano Zampini 256925084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 25703301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 257125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2572ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 257388ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 257488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 257588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 257688ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 257788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 257888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 257988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 258188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 258488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258588ebb749SStefano Zampini } else { 258688ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 258788ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 258888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 259088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 259188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 259288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 259388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 259488ebb749SStefano Zampini } 259588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 259688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 259788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2598511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 25994f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2600fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2601d12edf2fSStefano Zampini PetscScalar *data,*data2; 26024f1b2e48SStefano Zampini PetscInt j; 2603d12edf2fSStefano Zampini 26044f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2605fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2606d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 260786c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2608d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2609d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 26104f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 26114f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2612d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 26134f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 26144f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 26154f1b2e48SStefano Zampini } 2616d12edf2fSStefano Zampini } 2617d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2618d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2619d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2620d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2621d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2622d12edf2fSStefano Zampini } 2623d12edf2fSStefano Zampini #if 0 2624d12edf2fSStefano Zampini { 2625d12edf2fSStefano Zampini PetscViewer viewer; 2626d12edf2fSStefano Zampini char filename[256]; 2627ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2628d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2629d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2630ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2631ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2632ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2633d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 263472b8c272SStefano Zampini if (save_change) { 263572b8c272SStefano Zampini Mat phi_B; 263672b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 263772b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 263872b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 263972b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 264072b8c272SStefano Zampini } else { 2641ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2642ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 264372b8c272SStefano Zampini } 2644ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2645ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2646ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2647ffd830a3SStefano Zampini } 2648ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2649ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2650ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2651ffd830a3SStefano Zampini } 265272b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2653ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2654ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2655ffd830a3SStefano Zampini } 2656d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2657d12edf2fSStefano Zampini } 2658d12edf2fSStefano Zampini #endif 265981d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 26608bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 26611575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 266206656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 26638bec7fa6SStefano Zampini 26648bec7fa6SStefano Zampini /* check constraints */ 2665a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2666a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 26674f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 26688bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2669a00504b5SStefano Zampini } else { 2670a00504b5SStefano Zampini PetscScalar *data; 2671a00504b5SStefano Zampini Mat tmat; 2672a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2673a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2674a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2675a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2676a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2677a00504b5SStefano Zampini } 26788bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 26798bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 26808bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 26818bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2682bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2683ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2684bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2685bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2686bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2687bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2688bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 268988ebb749SStefano Zampini } 26908bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 26918bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 26928bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 26938bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 269425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 269588ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 269688ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 269788ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 269888ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 269988ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 270088ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 270188ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 270288ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 270388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 270488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2705ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 270688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 270788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 270888ebb749SStefano Zampini } 270988ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 271088ebb749SStefano Zampini } 27118629588bSStefano Zampini /* get back data */ 27128629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 271388ebb749SStefano Zampini PetscFunctionReturn(0); 271488ebb749SStefano Zampini } 271588ebb749SStefano Zampini 271688ebb749SStefano Zampini #undef __FUNCT__ 2717d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2718d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2719aa0d41d4SStefano Zampini { 2720d65f70fdSStefano Zampini Mat *work_mat; 2721d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2722d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2723c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2724aa0d41d4SStefano Zampini PetscErrorCode ierr; 2725aa0d41d4SStefano Zampini 2726aa0d41d4SStefano Zampini PetscFunctionBegin; 2727d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2728d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2729d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2730d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2731aa0d41d4SStefano Zampini 2732d65f70fdSStefano Zampini if (!rsorted) { 2733906d46d4SStefano Zampini const PetscInt *idxs; 2734906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2735aa0d41d4SStefano Zampini 2736d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2737d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2738d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2739d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2740aa0d41d4SStefano Zampini } 2741d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2742d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2743d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2744d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2745aa0d41d4SStefano Zampini } 2746d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2747d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2748d65f70fdSStefano Zampini } else { 2749d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2750d65f70fdSStefano Zampini isrow_s = isrow; 2751aa0d41d4SStefano Zampini } 2752906d46d4SStefano Zampini 2753d65f70fdSStefano Zampini if (!csorted) { 2754d65f70fdSStefano Zampini if (isrow == iscol) { 2755d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2756d65f70fdSStefano Zampini iscol_s = isrow_s; 2757d65f70fdSStefano Zampini } else { 2758d65f70fdSStefano Zampini const PetscInt *idxs; 2759d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2760906d46d4SStefano Zampini 2761d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2762d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2763d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2764d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2765d65f70fdSStefano Zampini } 2766d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2767d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2768d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2769d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2770d65f70fdSStefano Zampini } 2771d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2772d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2773d65f70fdSStefano Zampini } 2774d65f70fdSStefano Zampini } else { 2775d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2776d65f70fdSStefano Zampini iscol_s = iscol; 2777d65f70fdSStefano Zampini } 2778d65f70fdSStefano Zampini 2779d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2780d65f70fdSStefano Zampini 2781d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2782906d46d4SStefano Zampini Mat new_mat; 2783d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2784906d46d4SStefano Zampini 2785d65f70fdSStefano Zampini if (!rsorted) { 2786d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2787d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2788d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2789d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2790906d46d4SStefano Zampini } 2791d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2792d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2793d65f70fdSStefano Zampini } else { 2794d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2795906d46d4SStefano Zampini } 2796d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2797d65f70fdSStefano Zampini 2798d65f70fdSStefano Zampini if (!csorted) { 2799d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2800d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2801d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2802d65f70fdSStefano Zampini } else { 2803d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2804d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2805d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2806d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2807d65f70fdSStefano Zampini } 2808d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2809d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2810d65f70fdSStefano Zampini } 2811d65f70fdSStefano Zampini } else { 2812d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2813d65f70fdSStefano Zampini } 2814d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2815d65f70fdSStefano Zampini 2816d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2817d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2818d65f70fdSStefano Zampini work_mat[0] = new_mat; 2819d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2820d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2821d65f70fdSStefano Zampini } 2822d65f70fdSStefano Zampini 2823d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2824d65f70fdSStefano Zampini *B = work_mat[0]; 2825d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2826d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2827d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2828d65f70fdSStefano Zampini PetscFunctionReturn(0); 2829d65f70fdSStefano Zampini } 2830d65f70fdSStefano Zampini 2831d65f70fdSStefano Zampini #undef __FUNCT__ 28325e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 28335e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2834aa0d41d4SStefano Zampini { 2835aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 28365e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2837d65f70fdSStefano Zampini Mat new_mat; 28385e8657edSStefano Zampini IS is_local,is_global; 2839d65f70fdSStefano Zampini PetscInt local_size; 2840d65f70fdSStefano Zampini PetscBool isseqaij; 2841aa0d41d4SStefano Zampini PetscErrorCode ierr; 2842aa0d41d4SStefano Zampini 2843aa0d41d4SStefano Zampini PetscFunctionBegin; 2844aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 28455e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 28465e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2847b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2848aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2849d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2850aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2851906d46d4SStefano Zampini 2852906d46d4SStefano Zampini /* check */ 2853906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2854906d46d4SStefano Zampini Vec x,x_change; 2855906d46d4SStefano Zampini PetscReal error; 2856906d46d4SStefano Zampini 28575e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2858906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 28595e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2860e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2861e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2862d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2863e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2864e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2865906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2866906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2867906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2868906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2869906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2870906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2871906d46d4SStefano Zampini } 2872906d46d4SStefano Zampini 287322d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 28749b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 287522d5777bSStefano Zampini if (isseqaij) { 2876a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2877a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2878aa0d41d4SStefano Zampini } else { 2879a00504b5SStefano Zampini Mat work_mat; 28801cf9b237SStefano Zampini 2881a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2882aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2883a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 28841d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2885aa0d41d4SStefano Zampini } 28863301b35fSStefano Zampini if (matis->A->symmetric_set) { 28873301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2888e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 28893301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2890e496cd5dSStefano Zampini #endif 28913301b35fSStefano Zampini } 2892d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2893aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2894aa0d41d4SStefano Zampini } 2895aa0d41d4SStefano Zampini 2896aa0d41d4SStefano Zampini #undef __FUNCT__ 2897a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 28988ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2899a64d13efSStefano Zampini { 2900a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2901a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2902d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 290353892102SStefano Zampini PetscInt *idx_R_local=NULL; 29043a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 29053a50541eSStefano Zampini PetscInt vbs,bs; 29066816873aSStefano Zampini PetscBT bitmask=NULL; 2907a64d13efSStefano Zampini PetscErrorCode ierr; 2908a64d13efSStefano Zampini 2909a64d13efSStefano Zampini PetscFunctionBegin; 2910b23d619eSStefano Zampini /* 2911b23d619eSStefano Zampini No need to setup local scatters if 2912b23d619eSStefano Zampini - primal space is unchanged 2913b23d619eSStefano Zampini AND 2914b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2915b23d619eSStefano Zampini AND 2916b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2917b23d619eSStefano Zampini */ 2918b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2919f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2920f4ddd8eeSStefano Zampini } 2921f4ddd8eeSStefano Zampini /* destroy old objects */ 2922f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2923f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2924f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2925a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2926b371cd4fSStefano Zampini n_B = pcis->n_B; 2927b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2928b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 29293a50541eSStefano Zampini 2930a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 29316816873aSStefano Zampini 293253892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2933b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2934854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2935a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2936a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 29370e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2938a64d13efSStefano Zampini } 2939a64d13efSStefano Zampini 2940a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 29414641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 29426816873aSStefano Zampini idx_R_local[n_R++] = i; 2943a64d13efSStefano Zampini } 2944a64d13efSStefano Zampini } 2945df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2946df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 29476816873aSStefano Zampini 2948df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2949df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 29506816873aSStefano Zampini } 29513a50541eSStefano Zampini 29523a50541eSStefano Zampini /* Block code */ 29533a50541eSStefano Zampini vbs = 1; 29543a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 29553a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 29563a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 29573a50541eSStefano Zampini PetscInt *vary; 2958b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2959785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 29603a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2961d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2962d3df7717SStefano 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 */ 29630e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2964d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 29653a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 29663a50541eSStefano Zampini is_blocked = PETSC_FALSE; 29673a50541eSStefano Zampini break; 29683a50541eSStefano Zampini } 29693a50541eSStefano Zampini } 2970d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2971d3df7717SStefano Zampini } else { 2972d3df7717SStefano Zampini /* Verify directly the R set */ 2973d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2974d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2975d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2976d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2977d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2978d3df7717SStefano Zampini break; 2979d3df7717SStefano Zampini } 2980d3df7717SStefano Zampini } 2981d3df7717SStefano Zampini } 2982d3df7717SStefano Zampini } 29833a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 29843a50541eSStefano Zampini vbs = bs; 29853a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 29863a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 29873a50541eSStefano Zampini } 29883a50541eSStefano Zampini } 29893a50541eSStefano Zampini } 29903a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2991b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 2992df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 299353892102SStefano Zampini 2994df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2995df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 299653892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2997df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 299853892102SStefano Zampini } else { 29993a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 300053892102SStefano Zampini } 3001a64d13efSStefano Zampini 3002a64d13efSStefano Zampini /* print some info if requested */ 3003a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 3004a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3005a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 30061575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3007a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 3008a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 30094f1b2e48SStefano 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); 3010a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3011a64d13efSStefano Zampini } 3012a64d13efSStefano Zampini 3013a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 3014b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 30156816873aSStefano Zampini IS is_aux1,is_aux2; 30166816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 30176816873aSStefano Zampini 30183a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3019854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 3020854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 3021a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 30224641a718SStefano Zampini for (i=0; i<n_D; i++) { 30234641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 30244641a718SStefano Zampini } 3025a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3026a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30274641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 30284641a718SStefano Zampini aux_array1[j++] = i; 3029a64d13efSStefano Zampini } 3030a64d13efSStefano Zampini } 3031a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3032a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3033a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 30344641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 30354641a718SStefano Zampini aux_array2[j++] = i; 3036a64d13efSStefano Zampini } 3037a64d13efSStefano Zampini } 3038a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3039a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 3040a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 3041a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3042a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 3043a64d13efSStefano Zampini 30448eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3045785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 3046a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30474641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 30484641a718SStefano Zampini aux_array1[j++] = i; 3049a64d13efSStefano Zampini } 3050a64d13efSStefano Zampini } 3051a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3052a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 3053a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3054a64d13efSStefano Zampini } 30554641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 30563a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3057d62866d3SStefano Zampini } else { 3058df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 30596816873aSStefano Zampini IS tis; 30606816873aSStefano Zampini PetscInt schur_size; 30616816873aSStefano Zampini 3062df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 30636816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 3064df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 30656816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 30666816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 30676816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 30686816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 30696816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 3070d62866d3SStefano Zampini } 3071d62866d3SStefano Zampini } 3072a64d13efSStefano Zampini PetscFunctionReturn(0); 3073a64d13efSStefano Zampini } 3074a64d13efSStefano Zampini 3075304d26faSStefano Zampini 3076304d26faSStefano Zampini #undef __FUNCT__ 3077304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 3078684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 3079304d26faSStefano Zampini { 3080304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3081304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3082304d26faSStefano Zampini PC pc_temp; 3083304d26faSStefano Zampini Mat A_RR; 3084f4ddd8eeSStefano Zampini MatReuse reuse; 3085304d26faSStefano Zampini PetscScalar m_one = -1.0; 3086304d26faSStefano Zampini PetscReal value; 308704708bb6SStefano Zampini PetscInt n_D,n_R; 3088c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 3089304d26faSStefano Zampini PetscErrorCode ierr; 3090e604994aSStefano Zampini /* prefixes stuff */ 3091312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 3092e604994aSStefano Zampini size_t len; 3093304d26faSStefano Zampini 3094304d26faSStefano Zampini PetscFunctionBegin; 3095304d26faSStefano Zampini 3096e604994aSStefano Zampini /* compute prefixes */ 3097e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 3098e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 3099e604994aSStefano Zampini if (!pcbddc->current_level) { 3100e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3101e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3102e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3103e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3104e604994aSStefano Zampini } else { 3105e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3106312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3107e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3108e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 3109312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3110312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 311134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 311234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3113e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3114e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3115e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 3116e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 3117e604994aSStefano Zampini } 3118e604994aSStefano Zampini 3119304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3120684f6988SStefano Zampini if (dirichlet) { 3121d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3122450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 31239a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3124450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3125a3df083aSStefano Zampini Mat A_IIn; 3126a3df083aSStefano Zampini 3127a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3128a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3129a3df083aSStefano Zampini pcis->A_II = A_IIn; 3130a3df083aSStefano Zampini } 3131450f8f5eSStefano Zampini } 31323301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31333301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3134964fefecSStefano Zampini } 3135ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3136964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3137304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3138304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3139304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3140304d26faSStefano Zampini /* default */ 3141304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3142e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 31439577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3144304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 31459577ea80SStefano Zampini if (issbaij) { 31469577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 31479577ea80SStefano Zampini } else { 3148304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31499577ea80SStefano Zampini } 3150304d26faSStefano Zampini /* Allow user's customization */ 3151304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3152304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3153304d26faSStefano Zampini } 3154d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3155b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3156df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3157d62866d3SStefano Zampini 3158df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3159d5574798SStefano Zampini } 3160304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3161304d26faSStefano Zampini if (!n_D) { 3162304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3163304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3164304d26faSStefano Zampini } 3165304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3166304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3167304d26faSStefano Zampini /* set ksp_D into pcis data */ 3168304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3169304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3170304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3171684f6988SStefano Zampini } 3172304d26faSStefano Zampini 3173304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3174684f6988SStefano Zampini A_RR = 0; 3175684f6988SStefano Zampini if (neumann) { 3176d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 317704708bb6SStefano Zampini PetscInt ibs,mbs; 317804708bb6SStefano Zampini PetscBool issbaij; 317904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3180f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 31818ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3182f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3183f4ddd8eeSStefano Zampini PetscInt nn_R; 318481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3185f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3186f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3187f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3188f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3189f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3190f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3191f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3192727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3193f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3194f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3195f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3196f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3197f4ddd8eeSStefano Zampini } 3198f4ddd8eeSStefano Zampini } 3199f4ddd8eeSStefano Zampini /* last check */ 3200d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3201f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3202f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3203f4ddd8eeSStefano Zampini } 3204f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3205f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3206f4ddd8eeSStefano Zampini } 3207a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3208af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3209af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 321004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 321104708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 321204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 321304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 321404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3215af732b37SStefano Zampini } else { 3216511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 32176816873aSStefano Zampini } 321804708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 321904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 322004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 322104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 322204708bb6SStefano Zampini } else { 3223511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 322404708bb6SStefano Zampini } 322504708bb6SStefano Zampini } 3226a00504b5SStefano Zampini /* extract A_RR */ 3227b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3228a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3229a00504b5SStefano Zampini 3230a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 323116e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3232a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 323316e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 323416e386b8SStefano Zampini } else { 3235a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3236a00504b5SStefano Zampini } 3237a00504b5SStefano Zampini } else { 3238a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3239a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3240a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3241a00504b5SStefano Zampini } 3242a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3243f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 324416e386b8SStefano Zampini } 32453301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 32463301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 32476816873aSStefano Zampini } 3248f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3249304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3250304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3251304d26faSStefano Zampini /* default */ 3252304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3253e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3254304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 32559577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 32569577ea80SStefano Zampini if (issbaij) { 32579577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 32589577ea80SStefano Zampini } else { 3259304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 32609577ea80SStefano Zampini } 3261304d26faSStefano Zampini /* Allow user's customization */ 3262304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3263304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3264304d26faSStefano Zampini } 3265304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3266304d26faSStefano Zampini if (!n_R) { 3267304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3268304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3269304d26faSStefano Zampini } 32705cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3271df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3272b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3273df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3274d62866d3SStefano Zampini 3275df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3276d62866d3SStefano Zampini } 3277304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3278304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3279684f6988SStefano Zampini } 3280304d26faSStefano Zampini 3281684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3282684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32831575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3284684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3285684f6988SStefano Zampini } 3286c7017625SStefano Zampini 3287c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3288c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3289c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3290c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3291c7017625SStefano Zampini } 3292c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3293c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3294c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3295c7017625SStefano Zampini } 3296c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3297c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3298c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3299c7017625SStefano Zampini } 3300c7017625SStefano Zampini 3301c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3302c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3303684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 33040fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 33050fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 33060fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 33070fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 33080fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3309e604994aSStefano 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); 3310c7017625SStefano Zampini if (check_corr[0]) { 3311c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3312c7017625SStefano Zampini } 3313304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3314304d26faSStefano Zampini } 3315684f6988SStefano Zampini if (neumann) { /* Neumann */ 33160fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 33170fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 33180fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 33190fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 33200fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3321e604994aSStefano 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); 3322c7017625SStefano Zampini if (check_corr[1]) { 3323c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3324c7017625SStefano Zampini } 3325304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3326304d26faSStefano Zampini } 3327684f6988SStefano Zampini } 33285cbda25cSStefano Zampini /* free Neumann problem's matrix */ 33295cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3330304d26faSStefano Zampini PetscFunctionReturn(0); 3331304d26faSStefano Zampini } 3332304d26faSStefano Zampini 3333304d26faSStefano Zampini #undef __FUNCT__ 3334ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 333580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3336674ae819SStefano Zampini { 3337674ae819SStefano Zampini PetscErrorCode ierr; 3338674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3339be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3340b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 3341674ae819SStefano Zampini 3342674ae819SStefano Zampini PetscFunctionBegin; 3343b334f244SStefano Zampini if (!reuse_solver) { 334480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 334520c7b377SStefano Zampini } 334680677318SStefano Zampini if (!pcbddc->switch_static) { 334780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 334880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 334980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 335020c7b377SStefano Zampini } 3351b334f244SStefano Zampini if (!reuse_solver) { 335280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335420c7b377SStefano Zampini } else { 3355df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3356be83ff47SStefano Zampini 3357df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3358df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335920c7b377SStefano Zampini } 3360be83ff47SStefano Zampini } else { 336180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 336680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 336780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 336880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3370674ae819SStefano Zampini } 3371674ae819SStefano Zampini } 3372b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 337380677318SStefano Zampini if (applytranspose) { 337480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 337580677318SStefano Zampini } else { 337680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 337780677318SStefano Zampini } 3378be83ff47SStefano Zampini } else { 3379df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3380be83ff47SStefano Zampini 3381be83ff47SStefano Zampini if (applytranspose) { 3382df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3383be83ff47SStefano Zampini } else { 3384df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3385be83ff47SStefano Zampini } 3386be83ff47SStefano Zampini } 338780677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 338880677318SStefano Zampini if (!pcbddc->switch_static) { 3389b334f244SStefano Zampini if (!reuse_solver) { 339080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 339180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3392be83ff47SStefano Zampini } else { 3393df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3394be83ff47SStefano Zampini 3395df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3396df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3397be83ff47SStefano Zampini } 339880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 339980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 340080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 340180677318SStefano Zampini } 340280677318SStefano Zampini } else { 340380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 340880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 340980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 341080677318SStefano Zampini } 341180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3415674ae819SStefano Zampini } 3416674ae819SStefano Zampini PetscFunctionReturn(0); 3417674ae819SStefano Zampini } 3418674ae819SStefano Zampini 3419dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3420674ae819SStefano Zampini #undef __FUNCT__ 3421674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3422dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3423674ae819SStefano Zampini { 3424674ae819SStefano Zampini PetscErrorCode ierr; 3425674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3426674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3427674ae819SStefano Zampini const PetscScalar zero = 0.0; 3428674ae819SStefano Zampini 3429674ae819SStefano Zampini PetscFunctionBegin; 3430dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 34314fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3432dc359a40SStefano Zampini if (applytranspose) { 3433674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 34348eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3435dc359a40SStefano Zampini } else { 3436674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3437674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 343815aaf578SStefano Zampini } 34394fee134fSStefano Zampini } else { 34404fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 34414fee134fSStefano Zampini } 3442efc2fbd9SStefano Zampini 3443efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 34444f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3445efc2fbd9SStefano Zampini PetscScalar *array; 34464f1b2e48SStefano Zampini PetscInt j; 3447efc2fbd9SStefano Zampini 3448efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34494f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3450efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3451efc2fbd9SStefano Zampini } 3452efc2fbd9SStefano Zampini 345312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 345412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 345512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 345612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 345712edc857SStefano Zampini 34589f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 345912edc857SStefano Zampini if (pcbddc->coarse_ksp) { 346051694757SStefano Zampini Mat coarse_mat; 3461964fefecSStefano Zampini Vec rhs,sol; 346251694757SStefano Zampini MatNullSpace nullsp; 346327b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3464964fefecSStefano Zampini 346527b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 346627b6a85dSStefano Zampini PC coarse_pc; 346727b6a85dSStefano Zampini 346827b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 346927b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 347027b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 347127b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 347227b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 347327b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 34743bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 347527b6a85dSStefano Zampini } 347627b6a85dSStefano Zampini } 3477964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3478964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 347951694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 348051694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 348151694757SStefano Zampini if (nullsp) { 348251694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 348351694757SStefano Zampini } 348412edc857SStefano Zampini if (applytranspose) { 34859a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 3486964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 34872701bc32SStefano Zampini } else { 34881f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 34892701bc32SStefano Zampini PC coarse_pc; 34902701bc32SStefano Zampini 34912701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 34922701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 34933e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 34942701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 349512edc857SStefano Zampini } else { 3496964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 349712edc857SStefano Zampini } 34982701bc32SStefano Zampini } 34991d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 350027b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 350127b6a85dSStefano Zampini PC coarse_pc; 350227b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 350327b6a85dSStefano Zampini 350427b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 350527b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 350627b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 35073bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 350827b6a85dSStefano Zampini } 350951694757SStefano Zampini if (nullsp) { 351051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 351151694757SStefano Zampini } 351212edc857SStefano Zampini } 3513674ae819SStefano Zampini 3514674ae819SStefano Zampini /* Local solution on R nodes */ 35154fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 351680677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 35179f00e9b4SStefano Zampini } 35189f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 35199f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 352012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3521674ae819SStefano Zampini 35224fee134fSStefano Zampini /* Sum contributions from the two levels */ 35234fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3524dc359a40SStefano Zampini if (applytranspose) { 3525dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3526dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3527dc359a40SStefano Zampini } else { 3528674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 35298eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3530dc359a40SStefano Zampini } 3531efc2fbd9SStefano Zampini /* store p0 */ 35324f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3533efc2fbd9SStefano Zampini PetscScalar *array; 35344f1b2e48SStefano Zampini PetscInt j; 3535efc2fbd9SStefano Zampini 3536efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 35374f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3538efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3539efc2fbd9SStefano Zampini } 35404fee134fSStefano Zampini } else { /* expand the coarse solution */ 35414fee134fSStefano Zampini if (applytranspose) { 35424fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35434fee134fSStefano Zampini } else { 35444fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35454fee134fSStefano Zampini } 35464fee134fSStefano Zampini } 3547674ae819SStefano Zampini PetscFunctionReturn(0); 3548674ae819SStefano Zampini } 3549674ae819SStefano Zampini 3550674ae819SStefano Zampini #undef __FUNCT__ 3551674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 355212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3553674ae819SStefano Zampini { 3554674ae819SStefano Zampini PetscErrorCode ierr; 3555674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 355658da7f69SStefano Zampini PetscScalar *array; 355712edc857SStefano Zampini Vec from,to; 3558674ae819SStefano Zampini 3559674ae819SStefano Zampini PetscFunctionBegin; 356012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 356112edc857SStefano Zampini from = pcbddc->coarse_vec; 356212edc857SStefano Zampini to = pcbddc->vec1_P; 356312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 356412edc857SStefano Zampini Vec tvec; 356558da7f69SStefano Zampini 356658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 356758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 356812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 356958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 357058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 357158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 357212edc857SStefano Zampini } 357312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 357412edc857SStefano Zampini from = pcbddc->vec1_P; 357512edc857SStefano Zampini to = pcbddc->coarse_vec; 357612edc857SStefano Zampini } 357712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3578674ae819SStefano Zampini PetscFunctionReturn(0); 3579674ae819SStefano Zampini } 3580674ae819SStefano Zampini 3581674ae819SStefano Zampini #undef __FUNCT__ 3582674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 358312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3584674ae819SStefano Zampini { 3585674ae819SStefano Zampini PetscErrorCode ierr; 3586674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 358758da7f69SStefano Zampini PetscScalar *array; 358812edc857SStefano Zampini Vec from,to; 3589674ae819SStefano Zampini 3590674ae819SStefano Zampini PetscFunctionBegin; 359112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 359212edc857SStefano Zampini from = pcbddc->coarse_vec; 359312edc857SStefano Zampini to = pcbddc->vec1_P; 359412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 359512edc857SStefano Zampini from = pcbddc->vec1_P; 359612edc857SStefano Zampini to = pcbddc->coarse_vec; 359712edc857SStefano Zampini } 359812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 359912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 360012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 360112edc857SStefano Zampini Vec tvec; 360258da7f69SStefano Zampini 360312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 360458da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 360558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 360658da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 360758da7f69SStefano Zampini } 360858da7f69SStefano Zampini } else { 360958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 361058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 361112edc857SStefano Zampini } 361212edc857SStefano Zampini } 3613674ae819SStefano Zampini PetscFunctionReturn(0); 3614674ae819SStefano Zampini } 3615674ae819SStefano Zampini 3616984c4197SStefano Zampini /* uncomment for testing purposes */ 3617984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3618674ae819SStefano Zampini #undef __FUNCT__ 3619674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3620674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3621674ae819SStefano Zampini { 3622674ae819SStefano Zampini PetscErrorCode ierr; 3623674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3624674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3625674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3626984c4197SStefano Zampini /* one and zero */ 3627984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3628984c4197SStefano Zampini /* space to store constraints and their local indices */ 36299162d606SStefano Zampini PetscScalar *constraints_data; 36309162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 36319162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 36329162d606SStefano Zampini PetscInt *constraints_n; 3633984c4197SStefano Zampini /* iterators */ 3634b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3635984c4197SStefano Zampini /* BLAS integers */ 3636e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3637e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3638c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3639727cdba6SStefano Zampini /* reuse */ 36400e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 36410e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3642984c4197SStefano Zampini /* change of basis */ 3643b3d85658SStefano Zampini PetscBool qr_needed; 36449162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3645984c4197SStefano Zampini /* auxiliary stuff */ 364664efe560SStefano Zampini PetscInt *nnz,*is_indices; 36478a0068c3SStefano Zampini PetscInt ncc; 3648984c4197SStefano Zampini /* some quantities */ 364945a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3650a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3651984c4197SStefano Zampini 3652674ae819SStefano Zampini PetscFunctionBegin; 36538e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 36548e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 36558e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 365616909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3657088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3658088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 36590e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 36600e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 36610e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36620e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36630e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3664088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3665cf5a6209SStefano Zampini 3666cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 36679162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3668cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3669cf5a6209SStefano Zampini const Vec *nearnullvecs; 3670cf5a6209SStefano Zampini Vec *localnearnullsp; 3671cf5a6209SStefano Zampini PetscScalar *array; 3672cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3673cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3674674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3675b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3676674ae819SStefano Zampini PetscScalar *work; 3677674ae819SStefano Zampini PetscReal *singular_vals; 3678674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3679674ae819SStefano Zampini PetscReal *rwork; 3680674ae819SStefano Zampini #endif 3681674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3682674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3683674ae819SStefano Zampini #else 3684964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3685964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3686674ae819SStefano Zampini #endif 3687674ae819SStefano Zampini 3688674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3689d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3690e4d548c7SStefano Zampini /* print some info */ 36911f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3692e4d548c7SStefano Zampini PetscInt nv; 3693e4d548c7SStefano Zampini 36948ae0ca82SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag-1,pcbddc->dbg_viewer);CHKERRQ(ierr); 3695e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3696e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3697e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3698e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3699e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3700e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3701e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3702e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3703e4d548c7SStefano Zampini } 3704e4d548c7SStefano Zampini 3705d06fc5fdSStefano Zampini /* free unneeded index sets */ 3706d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3707d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3708674ae819SStefano Zampini } 3709d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3710d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3711d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3712d06fc5fdSStefano Zampini } 3713d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3714d06fc5fdSStefano Zampini n_ISForEdges = 0; 3715d06fc5fdSStefano Zampini } 3716d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3717d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3718d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3719d06fc5fdSStefano Zampini } 3720d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3721d06fc5fdSStefano Zampini n_ISForFaces = 0; 3722d06fc5fdSStefano Zampini } 372370022509SStefano Zampini 3724674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3725674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3726674ae819SStefano Zampini if (nearnullsp) { 3727674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3728f4ddd8eeSStefano Zampini /* remove any stored info */ 3729f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3730f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3731f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3732f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3733f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3734473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3735f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3736f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3737f4ddd8eeSStefano Zampini } 3738984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3739984c4197SStefano Zampini nnsp_size = 0; 3740674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3741674ae819SStefano Zampini } 3742984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3743984c4197SStefano Zampini max_constraints = nnsp_size; 3744984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3745984c4197SStefano Zampini 3746674ae819SStefano Zampini /* 3747674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 37489162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 37499162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 37509162d606SStefano Zampini There can be multiple constraints per connected component 3751674ae819SStefano Zampini */ 3752674ae819SStefano Zampini n_vertices = 0; 3753674ae819SStefano Zampini if (ISForVertices) { 3754674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3755674ae819SStefano Zampini } 37569162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 37579162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 37589162d606SStefano Zampini 37599162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 37609162d606SStefano Zampini total_counts *= max_constraints; 3761674ae819SStefano Zampini total_counts += n_vertices; 37624641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 37639162d606SStefano Zampini 3764674ae819SStefano Zampini total_counts = 0; 3765674ae819SStefano Zampini max_size_of_constraint = 0; 3766674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 37679162d606SStefano Zampini IS used_is; 3768674ae819SStefano Zampini if (i<n_ISForEdges) { 37699162d606SStefano Zampini used_is = ISForEdges[i]; 3770674ae819SStefano Zampini } else { 37719162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3772674ae819SStefano Zampini } 37739162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3774674ae819SStefano Zampini total_counts += j; 3775674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3776674ae819SStefano Zampini } 37779162d606SStefano 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); 37789162d606SStefano Zampini 3779984c4197SStefano Zampini /* get local part of global near null space vectors */ 3780785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3781984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3782984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3783e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3784e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3785984c4197SStefano Zampini } 3786674ae819SStefano Zampini 3787242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3788242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3789a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3790242a89d7SStefano Zampini 3791984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3792a773dcb8SStefano Zampini if (!skip_lapack) { 3793674ae819SStefano Zampini PetscScalar temp_work; 3794911cabfeSStefano Zampini 3795674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3796984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3797785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3798785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3799785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3800674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3801785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3802674ae819SStefano Zampini #endif 3803674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3804c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3805c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3806674ae819SStefano Zampini lwork = -1; 3807674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3808674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3809c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3810674ae819SStefano Zampini #else 3811c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3812674ae819SStefano Zampini #endif 3813674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3814984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3815674ae819SStefano Zampini #else /* on missing GESVD */ 3816674ae819SStefano Zampini /* SVD */ 3817674ae819SStefano Zampini PetscInt max_n,min_n; 3818674ae819SStefano Zampini max_n = max_size_of_constraint; 3819984c4197SStefano Zampini min_n = max_constraints; 3820984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3821674ae819SStefano Zampini min_n = max_size_of_constraint; 3822984c4197SStefano Zampini max_n = max_constraints; 3823674ae819SStefano Zampini } 3824785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3825674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3826785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3827674ae819SStefano Zampini #endif 3828674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3829674ae819SStefano Zampini lwork = -1; 3830e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3831e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3832b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3833674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3834674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 38359162d606SStefano 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)); 3836674ae819SStefano Zampini #else 38379162d606SStefano 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)); 3838674ae819SStefano Zampini #endif 3839674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3840984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3841984c4197SStefano Zampini #endif /* on missing GESVD */ 3842674ae819SStefano Zampini /* Allocate optimal workspace */ 3843674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3844854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3845674ae819SStefano Zampini } 3846674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3847674ae819SStefano Zampini total_counts = 0; 38489162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 38499162d606SStefano Zampini constraints_data_ptr[0] = 0; 3850674ae819SStefano Zampini /* vertices */ 38519162d606SStefano Zampini if (n_vertices) { 3852674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 38539162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3854674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 38559162d606SStefano Zampini constraints_n[total_counts] = 1; 38569162d606SStefano Zampini constraints_data[total_counts] = 1.0; 38579162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 38589162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3859674ae819SStefano Zampini total_counts++; 3860674ae819SStefano Zampini } 3861674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3862674ae819SStefano Zampini n_vertices = total_counts; 3863674ae819SStefano Zampini } 3864984c4197SStefano Zampini 3865674ae819SStefano Zampini /* edges and faces */ 38669162d606SStefano Zampini total_counts_cc = total_counts; 3867911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 38689162d606SStefano Zampini IS used_is; 38699162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 38709162d606SStefano Zampini 3871911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 38729162d606SStefano Zampini used_is = ISForEdges[ncc]; 3873984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3874674ae819SStefano Zampini } else { 38759162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3876984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3877674ae819SStefano Zampini } 3878674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 38799162d606SStefano Zampini 38809162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 38819162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3882984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3883984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3884674ae819SStefano Zampini if (nnsp_has_cnst) { 38855b08dc53SStefano Zampini PetscScalar quad_value; 38869162d606SStefano Zampini 38879162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38889162d606SStefano Zampini idxs_copied = PETSC_TRUE; 38899162d606SStefano Zampini 3890a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3891674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3892a773dcb8SStefano Zampini } else { 3893a773dcb8SStefano Zampini quad_value = 1.0; 3894a773dcb8SStefano Zampini } 3895674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38969162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3897674ae819SStefano Zampini } 38989162d606SStefano Zampini temp_constraints++; 3899674ae819SStefano Zampini total_counts++; 3900674ae819SStefano Zampini } 3901674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3902984c4197SStefano Zampini PetscReal real_value; 39039162d606SStefano Zampini PetscScalar *ptr_to_data; 39049162d606SStefano Zampini 3905984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 39069162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3907674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 39089162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3909674ae819SStefano Zampini } 3910984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3911984c4197SStefano Zampini /* check if array is null on the connected component */ 3912e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39139162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 39145b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3915674ae819SStefano Zampini temp_constraints++; 3916674ae819SStefano Zampini total_counts++; 39179162d606SStefano Zampini if (!idxs_copied) { 39189162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 39199162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3920674ae819SStefano Zampini } 3921674ae819SStefano Zampini } 39229162d606SStefano Zampini } 39239162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 392445a1bb75SStefano Zampini valid_constraints = temp_constraints; 3925eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3926a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 39279162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 39289162d606SStefano Zampini 39299162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3930a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39319162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3932a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 39339162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3934a773dcb8SStefano Zampini } else { /* perform SVD */ 3935984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 39369162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3937674ae819SStefano Zampini 3938674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3939984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3940984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3941984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3942984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3943984c4197SStefano Zampini from that computed using LAPACKgesvd 3944984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3945984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3946984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3947674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3948e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3949984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3950674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3951674ae819SStefano Zampini for (k=0;k<j+1;k++) { 39529162d606SStefano 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)); 3953674ae819SStefano Zampini } 3954674ae819SStefano Zampini } 3955e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3956e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3957e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3958674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3959c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3960674ae819SStefano Zampini #else 3961c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3962674ae819SStefano Zampini #endif 3963674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3964984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3965984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3966674ae819SStefano Zampini j = 0; 3967984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3968674ae819SStefano Zampini total_counts = total_counts-j; 396945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3970e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3971c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3972c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3973c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3974c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3975c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3976c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3977674ae819SStefano Zampini if (j<temp_constraints) { 3978984c4197SStefano Zampini PetscInt ii; 3979984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3980674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39819162d606SStefano 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)); 3982674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3983984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3984674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 39859162d606SStefano 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]; 3986674ae819SStefano Zampini } 3987674ae819SStefano Zampini } 3988674ae819SStefano Zampini } 3989674ae819SStefano Zampini #else /* on missing GESVD */ 3990e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3991e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3992b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3993674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3994674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 39959162d606SStefano 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)); 3996674ae819SStefano Zampini #else 39979162d606SStefano 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)); 3998674ae819SStefano Zampini #endif 3999984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 4000674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4001984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 4002e310c8b4SStefano Zampini k = temp_constraints; 4003e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 4004674ae819SStefano Zampini j = 0; 4005e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 400645a1bb75SStefano Zampini valid_constraints = k-j; 4007911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 4008984c4197SStefano Zampini #endif /* on missing GESVD */ 4009674ae819SStefano Zampini } 4010a773dcb8SStefano Zampini } 40119162d606SStefano Zampini /* update pointers information */ 40129162d606SStefano Zampini if (valid_constraints) { 40139162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 40149162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 40159162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 40169162d606SStefano Zampini /* set change_of_basis flag */ 401745a1bb75SStefano Zampini if (boolforchange) { 4018b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 40199162d606SStefano Zampini } 4020b3d85658SStefano Zampini total_counts_cc++; 402145a1bb75SStefano Zampini } 402245a1bb75SStefano Zampini } 4023984c4197SStefano Zampini /* free workspace */ 40248f1c130eSStefano Zampini if (!skip_lapack) { 4025984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4026984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4027984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 4028984c4197SStefano Zampini #endif 4029984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 4030984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4031984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 4032984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 4033984c4197SStefano Zampini #endif 4034984c4197SStefano Zampini } 4035984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4036984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 4037984c4197SStefano Zampini } 4038984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 4039cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 4040cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4041cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4042cf5a6209SStefano Zampini } 4043cf5a6209SStefano Zampini if (n_ISForFaces) { 4044cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4045cf5a6209SStefano Zampini } 4046cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4047cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4048cf5a6209SStefano Zampini } 4049cf5a6209SStefano Zampini if (n_ISForEdges) { 4050cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4051cf5a6209SStefano Zampini } 4052cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 405308122e43SStefano Zampini } else { 405408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4055984c4197SStefano Zampini 405608122e43SStefano Zampini total_counts = 0; 405708122e43SStefano Zampini n_vertices = 0; 4058d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 4059d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 406008122e43SStefano Zampini } 406108122e43SStefano Zampini max_constraints = 0; 40629162d606SStefano Zampini total_counts_cc = 0; 406308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 406408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 40659162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 406608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 406708122e43SStefano Zampini } 40689162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 40699162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 40709162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 40719162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 407274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 40739162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 40749162d606SStefano Zampini total_counts_cc = 0; 40759162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 40769162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 40779162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 407808122e43SStefano Zampini } 407908122e43SStefano Zampini } 40809162d606SStefano Zampini #if 0 40819162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 40829162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40839162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 40849162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 40859162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 40869162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 40879162d606SStefano Zampini } 40889162d606SStefano Zampini printf("\n"); 40899162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 40909162d606SStefano Zampini } 40911b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 40928bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 40931b968477SStefano Zampini } 40941b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 40958bec7fa6SStefano 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]); 40961b968477SStefano Zampini } 409708122e43SStefano Zampini #endif 409808122e43SStefano Zampini 40998bec7fa6SStefano Zampini max_size_of_constraint = 0; 41009162d606SStefano 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]); 41019162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 410208122e43SStefano Zampini /* Change of basis */ 4103b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 410408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 410508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 410608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 4107b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 410808122e43SStefano Zampini } 410908122e43SStefano Zampini } 411008122e43SStefano Zampini } 411108122e43SStefano Zampini } 4112984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 41134f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 411408122e43SStefano Zampini 41159162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 41169162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 41176c4ed002SBarry 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); 4118674ae819SStefano Zampini 4119674ae819SStefano Zampini /* Create constraint matrix */ 4120674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 412116f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4122984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4123984c4197SStefano Zampini 4124984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4125a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4126a717540cSStefano Zampini qr_needed = PETSC_FALSE; 412774d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4128984c4197SStefano Zampini total_primal_vertices=0; 4129b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 41309162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 41319162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 413272b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 41339162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4134b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 413564efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 41369162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41379162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4138a717540cSStefano Zampini } 4139b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 414091af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4141a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4142a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4143a717540cSStefano Zampini } 4144fa434743SStefano Zampini } else { 4145b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4146fa434743SStefano Zampini } 4147a717540cSStefano Zampini } 4148b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4149b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4150674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 415170022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4152b3d85658SStefano Zampini 41534f1b2e48SStefano 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); 41540e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 41550e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4156984c4197SStefano Zampini 4157984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 415874d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4159785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4160984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 416174d5cdf7SStefano Zampini 4162984c4197SStefano Zampini j = total_primal_vertices; 416374d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4164b3d85658SStefano Zampini cum = total_primal_vertices; 41659162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41664641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4167b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4168b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4169b3d85658SStefano Zampini cum++; 41709162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 417174d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 417274d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 417374d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 417474d5cdf7SStefano Zampini } 41759162d606SStefano Zampini j += constraints_n[i]; 4176674ae819SStefano Zampini } 4177674ae819SStefano Zampini } 4178674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4179674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4180088faed8SStefano Zampini 4181674ae819SStefano Zampini /* set values in constraint matrix */ 4182984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 41830e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4184674ae819SStefano Zampini } 4185984c4197SStefano Zampini total_counts = total_primal_vertices; 41869162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41874641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 41889162d606SStefano Zampini PetscInt *cols; 41899162d606SStefano Zampini 41909162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 41919162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 41929162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41939162d606SStefano Zampini PetscInt row = total_counts+k; 41949162d606SStefano Zampini PetscScalar *vals; 41959162d606SStefano Zampini 41969162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 41979162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 41989162d606SStefano Zampini } 41999162d606SStefano Zampini total_counts += constraints_n[i]; 4200674ae819SStefano Zampini } 4201674ae819SStefano Zampini } 4202674ae819SStefano Zampini /* assembling */ 4203674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4204674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4205088faed8SStefano Zampini 4206984c4197SStefano Zampini /* 42076a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4208984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4209f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4210984c4197SStefano Zampini */ 4211674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4212674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4213026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4214984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4215984c4197SStefano Zampini /* working stuff for GEQRF */ 421681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4217984c4197SStefano Zampini PetscBLASInt lqr_work; 4218984c4197SStefano Zampini /* working stuff for UNGQR */ 4219984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4220984c4197SStefano Zampini PetscBLASInt lgqr_work; 4221984c4197SStefano Zampini /* working stuff for TRTRS */ 4222984c4197SStefano Zampini PetscScalar *trs_rhs; 42233f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4224984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4225984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4226984c4197SStefano Zampini PetscScalar *start_vals; 4227984c4197SStefano Zampini /* working stuff for values insertion */ 42284641a718SStefano Zampini PetscBT is_primal; 422964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4230906d46d4SStefano Zampini /* matrix sizes */ 4231906d46d4SStefano Zampini PetscInt global_size,local_size; 4232906d46d4SStefano Zampini /* temporary change of basis */ 4233906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4234cf5a6209SStefano Zampini /* extra space for debugging */ 4235cf5a6209SStefano Zampini PetscScalar *dbg_work; 4236984c4197SStefano Zampini 4237906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4238906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 423916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4240bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4241906d46d4SStefano Zampini /* nonzeros for local mat */ 4242bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 42431dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4244bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 42451dd7afcfSStefano Zampini } else { 42461dd7afcfSStefano Zampini const PetscInt *ii; 42471dd7afcfSStefano Zampini PetscInt n; 42481dd7afcfSStefano Zampini PetscBool flg_row; 42491dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42501dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 42511dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42521dd7afcfSStefano Zampini } 42539162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4254a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 42559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4256a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 42579162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4258a717540cSStefano Zampini } else { 42599162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 42609162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4261a717540cSStefano Zampini } 4262a717540cSStefano Zampini } 4263a717540cSStefano Zampini } 4264906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4265bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 42661dd7afcfSStefano Zampini /* Set interior change in the matrix */ 42671dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4268bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4269906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4270a717540cSStefano Zampini } 42711dd7afcfSStefano Zampini } else { 42721dd7afcfSStefano Zampini const PetscInt *ii,*jj; 42731dd7afcfSStefano Zampini PetscScalar *aa; 42741dd7afcfSStefano Zampini PetscInt n; 42751dd7afcfSStefano Zampini PetscBool flg_row; 42761dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42771dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42781dd7afcfSStefano Zampini for (i=0;i<n;i++) { 42791dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 42801dd7afcfSStefano Zampini } 42811dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42821dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42831dd7afcfSStefano Zampini } 4284a717540cSStefano Zampini 4285a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4286a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4287a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4288a717540cSStefano Zampini } 4289a717540cSStefano Zampini 4290a717540cSStefano Zampini 4291a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4292a717540cSStefano Zampini /* 4293a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4294a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4295a717540cSStefano Zampini 4296a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4297a717540cSStefano Zampini 4298a6b551f4SStefano 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) 4299a6b551f4SStefano Zampini 4300a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4301a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4302a717540cSStefano Zampini | ... | 4303a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4304a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4305a717540cSStefano Zampini 4306a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4307a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4308a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4309a6b551f4SStefano Zampini 4310a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4311a717540cSStefano Zampini */ 4312a717540cSStefano Zampini if (qr_needed) { 4313984c4197SStefano Zampini /* space to store Q */ 4314854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4315984c4197SStefano Zampini /* first we issue queries for optimal work */ 43163f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43173f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 43183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4319984c4197SStefano Zampini lqr_work = -1; 43203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4321984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4322984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4323785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4324984c4197SStefano Zampini lgqr_work = -1; 43253f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43263f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 43273f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 43283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43293f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 43303f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4331984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4332984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4333785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4334984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4335785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4336984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4337785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4338a717540cSStefano Zampini /* allocating workspace for check */ 4339a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4340cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4341a717540cSStefano Zampini } 4342a717540cSStefano Zampini } 4343984c4197SStefano Zampini /* array to store whether a node is primal or not */ 43444641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4345473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 43460e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 43476c4ed002SBarry 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); 434839e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 434939e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 435039e2fb2aSStefano Zampini } 435139e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4352984c4197SStefano Zampini 4353a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 43549162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 43559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 43564641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4357984c4197SStefano Zampini /* get constraint info */ 43589162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4359984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4360984c4197SStefano Zampini 4361984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43629162d606SStefano 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); 4363674ae819SStefano Zampini } 4364984c4197SStefano Zampini 4365fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4366a717540cSStefano Zampini 4367a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4368a717540cSStefano Zampini if (pcbddc->dbg_flag) { 43699162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4370a717540cSStefano Zampini } 4371984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 43729162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4373984c4197SStefano Zampini 4374984c4197SStefano Zampini /* compute QR decomposition of constraints */ 43753f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43763f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43773f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4378674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43793f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4380984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4381674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4382984c4197SStefano Zampini 4383984c4197SStefano Zampini /* explictly compute R^-T */ 4384984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4385984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 43863f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43873f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 43883f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43893f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4390984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43913f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4392984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4393984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4394984c4197SStefano Zampini 4395a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 43963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43973f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 43983f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 43993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4400984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44013f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4402984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4403984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4404984c4197SStefano Zampini 4405984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4406984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4407984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 44083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44093f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 44103f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 44113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 44123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 44133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4414984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44159162d606SStefano 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)); 4416984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 44179162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4418984c4197SStefano Zampini 4419984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 44209162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4421984c4197SStefano Zampini /* insert cols for primal dofs */ 4422984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4423984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 44249162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4425906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4426984c4197SStefano Zampini } 4427984c4197SStefano Zampini /* insert cols for dual dofs */ 4428984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 44299162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4430984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 44319162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4432906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4433984c4197SStefano Zampini j++; 4434674ae819SStefano Zampini } 4435674ae819SStefano Zampini } 4436984c4197SStefano Zampini 4437984c4197SStefano Zampini /* check change of basis */ 4438984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4439984c4197SStefano Zampini PetscInt ii,jj; 4440984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4441c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4442c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4443c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4444c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4445c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4446c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4447984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4448cf5a6209SStefano 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)); 4449984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4450984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4451984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4452cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4453cf5a6209SStefano 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; 4454674ae819SStefano Zampini } 4455674ae819SStefano Zampini } 4456984c4197SStefano Zampini if (!valid_qr) { 445722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4458984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4459984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4460cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4461cf5a6209SStefano 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])); 4462674ae819SStefano Zampini } 4463cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4464cf5a6209SStefano 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])); 4465984c4197SStefano Zampini } 4466984c4197SStefano Zampini } 4467984c4197SStefano Zampini } 4468674ae819SStefano Zampini } else { 446922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4470674ae819SStefano Zampini } 4471674ae819SStefano Zampini } 4472a717540cSStefano Zampini } else { /* simple transformation block */ 4473a717540cSStefano Zampini PetscInt row,col; 4474a6b551f4SStefano Zampini PetscScalar val,norm; 4475a6b551f4SStefano Zampini 4476a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44779162d606SStefano 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)); 4478a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 44799162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 44809162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4481bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 44829162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4483906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 44849162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4485a717540cSStefano Zampini } else { 4486a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 44879162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4488a717540cSStefano Zampini if (row != col) { 44899162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4490a717540cSStefano Zampini } else { 44919162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4492a717540cSStefano Zampini } 4493906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4494a717540cSStefano Zampini } 4495a717540cSStefano Zampini } 4496a717540cSStefano Zampini } 449798a51de6SStefano Zampini if (pcbddc->dbg_flag) { 449822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4499a717540cSStefano Zampini } 4500674ae819SStefano Zampini } 4501984c4197SStefano Zampini } else { 4502984c4197SStefano Zampini if (pcbddc->dbg_flag) { 45039162d606SStefano 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); 4504674ae819SStefano Zampini } 4505674ae819SStefano Zampini } 4506674ae819SStefano Zampini } 4507a717540cSStefano Zampini 4508a717540cSStefano Zampini /* free workspace */ 4509a717540cSStefano Zampini if (qr_needed) { 4510984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4511cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4512984c4197SStefano Zampini } 4513984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4514984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4515984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4516984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4517984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4518674ae819SStefano Zampini } 4519a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4520906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4521906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4522906d46d4SStefano Zampini 4523906d46d4SStefano Zampini /* assembling of global change of variable */ 452488c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4525bbb9e6c6SStefano Zampini Mat tmat; 452616f15bc4SStefano Zampini PetscInt bs; 452716f15bc4SStefano Zampini 4528906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4529906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4530bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4531bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4532bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4533bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 453416f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 453516f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4536906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4537bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4538bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4539bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4540bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4541bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4542e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4543e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4544bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4545bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 454688c03ad3SStefano Zampini 4547906d46d4SStefano Zampini /* check */ 4548906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4549906d46d4SStefano Zampini PetscReal error; 4550906d46d4SStefano Zampini Vec x,x_change; 4551906d46d4SStefano Zampini 4552906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4553906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4554906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4555906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4556e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4557e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4558bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4559e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4560e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4561906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4562906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4563906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4564906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4565bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4566906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4567906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4568906d46d4SStefano Zampini } 4569b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4570b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4571b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4572bf3a8328SStefano Zampini 45739a962809SStefano 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); 4574b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 4575ac632422SStefano Zampini Mat S_new,tmat; 4576bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4577bbb9e6c6SStefano Zampini 4578bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 45796816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4580bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4581bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4582bf3a8328SStefano Zampini IS is_V; 4583b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4584b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4585b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4586b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4587b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4588bf3a8328SStefano Zampini } 4589bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4590ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4591b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4592ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4593bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4594bf3a8328SStefano Zampini const PetscScalar *array; 4595bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4596bf3a8328SStefano Zampini PetscInt i,n_V; 4597bf3a8328SStefano Zampini 4598b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4599b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4600b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4601b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4602b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4603b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4604b087196eSStefano Zampini PetscScalar val; 4605b087196eSStefano Zampini PetscInt idx; 4606b087196eSStefano Zampini 4607b087196eSStefano Zampini idx = idxs_V[i]; 4608b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4609b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4610b087196eSStefano Zampini } 4611b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4612b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4613bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4614bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4615bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4616bf3a8328SStefano Zampini } 4617ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4618ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4619ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4620ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4621b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4622ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4623bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4624b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4625bf3a8328SStefano Zampini } 4626ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4627ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4628ac632422SStefano Zampini } 4629b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 463088c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4631b96c3477SStefano Zampini } 4632c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4633b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 4634c9db6a07SStefano Zampini PetscInt i; 4635c9db6a07SStefano Zampini 4636c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4637c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4638c9db6a07SStefano Zampini } 4639c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4640c9db6a07SStefano Zampini } 4641b96c3477SStefano Zampini } 464216909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 464316909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 464416909a7fSStefano Zampini } else { 4645906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 464616909a7fSStefano Zampini } 46471dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 464827b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 464972b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 465072b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 465172b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 465272b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 465372b8c272SStefano Zampini } 46541dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 465527b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4656b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4657b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4658906d46d4SStefano Zampini } else { 46591dd7afcfSStefano Zampini Mat benign_global = NULL; 466027b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 46611dd7afcfSStefano Zampini Mat tmat; 46621dd7afcfSStefano Zampini 46631dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 46641dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 46651dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 46661dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46671dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46681dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 46691dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46701dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46711dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 46721dd7afcfSStefano Zampini if (pcbddc->benign_change) { 46731dd7afcfSStefano Zampini Mat M; 46741dd7afcfSStefano Zampini 46751dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 46761dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 46771dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 46781dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4679906d46d4SStefano Zampini } else { 46801dd7afcfSStefano Zampini Mat eye; 46811dd7afcfSStefano Zampini PetscScalar *array; 46821dd7afcfSStefano Zampini 46831dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46841dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 46851dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 46861dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4687906d46d4SStefano Zampini } 46881dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46891dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46901dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46911dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 46921dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 46931dd7afcfSStefano Zampini } 46941dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 46951dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 46961dd7afcfSStefano Zampini } 46971dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 46981dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 46991dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 470027b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 47011dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 47021dd7afcfSStefano Zampini } 47031dd7afcfSStefano Zampini } 470416909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 470516909a7fSStefano Zampini IS is_global; 470616909a7fSStefano Zampini const PetscInt *gidxs; 470716909a7fSStefano Zampini 470816909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 470916909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 471016909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 471116909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 471216909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 471316909a7fSStefano Zampini } 47141dd7afcfSStefano Zampini } 47151dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 47161dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4717b9b85e73SStefano Zampini } 4718a717540cSStefano Zampini 471972b8c272SStefano Zampini if (!pcbddc->fake_change) { 47204f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 47214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 47224f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 47234f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4724019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4725019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4726019a44ceSStefano Zampini pcbddc->local_primal_size++; 4727019a44ceSStefano Zampini } 4728019a44ceSStefano Zampini 4729019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4730727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4731727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 47329f47a83aSStefano 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); 4733c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 47340e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 47359f47a83aSStefano 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); 4736727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4737727cdba6SStefano Zampini } 47380e6343abSStefano Zampini } 4739727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4740b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 474172b8c272SStefano Zampini } 474272b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4743727cdba6SStefano Zampini 4744a717540cSStefano Zampini /* flush dbg viewer */ 4745b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4746b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4747b8ffe317SStefano Zampini } 4748a717540cSStefano Zampini 4749e310c8b4SStefano Zampini /* free workspace */ 4750a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 47514641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 475208122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 47539162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 47549162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 475508122e43SStefano Zampini } else { 47569162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 47579162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 47589162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 475908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 476008122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 47619162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 47629162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 476308122e43SStefano Zampini } 4764674ae819SStefano Zampini PetscFunctionReturn(0); 4765674ae819SStefano Zampini } 4766674ae819SStefano Zampini 4767674ae819SStefano Zampini #undef __FUNCT__ 4768674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4769674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4770674ae819SStefano Zampini { 477171582508SStefano Zampini ISLocalToGlobalMapping map; 4772674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4773674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 477414f95afaSStefano Zampini PetscInt ierr,i,N; 4775674ae819SStefano Zampini 4776674ae819SStefano Zampini PetscFunctionBegin; 477771582508SStefano Zampini if (pcbddc->graphanalyzed) PetscFunctionReturn(0); 47788e61c736SStefano Zampini /* Reset previously computed graph */ 47798e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4780674ae819SStefano Zampini /* Init local Graph struct */ 47817fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 478271582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 478371582508SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N);CHKERRQ(ierr); 4784674ae819SStefano Zampini 4785575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 47869a962809SStefano 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); 47879577ea80SStefano Zampini 4788674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4789d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 47904d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 47914d379d7bSStefano Zampini PetscInt nvtxs; 4792e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4793674ae819SStefano Zampini 47942fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 47952fffb893SStefano Zampini if (flg_row) { 47964d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4797b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 47982fffb893SStefano Zampini } 47992fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4800674ae819SStefano Zampini } 48019b28b941SStefano Zampini if (pcbddc->dbg_flag) { 48029b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4803674ae819SStefano Zampini } 4804674ae819SStefano Zampini 4805674ae819SStefano Zampini /* Setup of Graph */ 48064b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 480714f95afaSStefano 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); 4808674ae819SStefano Zampini 48094f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 48104f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 48114f1b2e48SStefano Zampini PetscInt *local_subs; 48124f1b2e48SStefano Zampini 48134f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 48144f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 48154f1b2e48SStefano Zampini const PetscInt *idxs; 48164f1b2e48SStefano Zampini PetscInt nl,j; 48174f1b2e48SStefano Zampini 48184f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 48194f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 482071582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 48214f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 48224f1b2e48SStefano Zampini } 48234f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 48244f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 48254f1b2e48SStefano Zampini } 48264f1b2e48SStefano Zampini 4827674ae819SStefano Zampini /* Graph's connected components analysis */ 4828674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 482971582508SStefano Zampini 483071582508SStefano Zampini /* set flag indicating analysis has been done */ 483171582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 4832674ae819SStefano Zampini PetscFunctionReturn(0); 4833674ae819SStefano Zampini } 4834674ae819SStefano Zampini 48359a7d3425SStefano Zampini #undef __FUNCT__ 48369a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 48379a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 48389a7d3425SStefano Zampini { 48399a7d3425SStefano Zampini PetscInt i,j; 48409a7d3425SStefano Zampini PetscScalar *alphas; 48419a7d3425SStefano Zampini PetscErrorCode ierr; 48429a7d3425SStefano Zampini 48439a7d3425SStefano Zampini PetscFunctionBegin; 4844785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 48459a7d3425SStefano Zampini for (i=0;i<n;i++) { 48469a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 4847669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 4848669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 4849669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 48509a7d3425SStefano Zampini } 48519a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 48529a7d3425SStefano Zampini PetscFunctionReturn(0); 48539a7d3425SStefano Zampini } 48549a7d3425SStefano Zampini 4855e7931f94SStefano Zampini #undef __FUNCT__ 485670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 485757de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4858e7931f94SStefano Zampini { 485957de7509SStefano Zampini Mat A; 4860e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4861e7931f94SStefano Zampini PetscMPIInt size,rank,color; 486252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 486352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 486427b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 486557de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 486627b6a85dSStefano Zampini PetscInt xadj_count, *count; 486727b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 486827b6a85dSStefano Zampini PetscSubcomm psubcomm; 486927b6a85dSStefano Zampini MPI_Comm subcomm; 487052e5ac9dSStefano Zampini PetscErrorCode ierr; 4871a57a6d2fSStefano Zampini 4872e7931f94SStefano Zampini PetscFunctionBegin; 487357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 487457de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 487557de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 487657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 487757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 487857de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 487957de7509SStefano Zampini 488057de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 488157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 488257de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 488357de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 488457de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 488557de7509SStefano Zampini im_active = !!(n); 488657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 488757de7509SStefano Zampini void_procs = size - active_procs; 488857de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 488957de7509SStefano Zampini if (void_procs) { 489057de7509SStefano Zampini PetscInt ncand; 489157de7509SStefano Zampini 489257de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 489357de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 489457de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 489557de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 489657de7509SStefano Zampini if (!procs_candidates[i]) { 489757de7509SStefano Zampini procs_candidates[ncand++] = i; 489857de7509SStefano Zampini } 489957de7509SStefano Zampini } 490057de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 490157de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 490257de7509SStefano Zampini } 490357de7509SStefano Zampini 490457de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 490557de7509SStefano Zampini if (active_procs < *n_subdomains) { 490657de7509SStefano Zampini PetscInt issize,isidx; 490757de7509SStefano Zampini if (im_active) { 490857de7509SStefano Zampini issize = 1; 490957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 491057de7509SStefano Zampini isidx = procs_candidates[rank]; 491157de7509SStefano Zampini } else { 491257de7509SStefano Zampini isidx = rank; 491357de7509SStefano Zampini } 491457de7509SStefano Zampini } else { 491557de7509SStefano Zampini issize = 0; 491657de7509SStefano Zampini isidx = -1; 491757de7509SStefano Zampini } 491857de7509SStefano Zampini *n_subdomains = active_procs; 491957de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 4920daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 492157de7509SStefano Zampini PetscFunctionReturn(0); 492257de7509SStefano Zampini } 4923c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4924c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 492527b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 4926e7931f94SStefano Zampini 4927e7931f94SStefano Zampini /* Get info on mapping */ 49283bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 49293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4930e7931f94SStefano Zampini 4931e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4932785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4933e7931f94SStefano Zampini xadj[0] = 0; 4934e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4935785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4936785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 493727b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 493827b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 493927b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 494027b6a85dSStefano Zampini count[shared[i][j]] += 1; 4941e7931f94SStefano Zampini 494227b6a85dSStefano Zampini xadj_count = 0; 49432b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 494427b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 494527b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 4946d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4947d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4948d023bfaeSStefano Zampini xadj_count++; 494927b6a85dSStefano Zampini break; 495027b6a85dSStefano Zampini } 4951e7931f94SStefano Zampini } 4952e7931f94SStefano Zampini } 4953d023bfaeSStefano Zampini xadj[1] = xadj_count; 495427b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 49553bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4956e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4957e7931f94SStefano Zampini 49583837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4959e7931f94SStefano Zampini 496027b6a85dSStefano Zampini /* Restrict work on active processes only */ 496127b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 496227b6a85dSStefano Zampini if (void_procs) { 496327b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 496427b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 496527b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 496627b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 496727b6a85dSStefano Zampini } else { 496827b6a85dSStefano Zampini psubcomm = NULL; 496927b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 497027b6a85dSStefano Zampini } 497127b6a85dSStefano Zampini 497227b6a85dSStefano Zampini v_wgt = NULL; 497327b6a85dSStefano Zampini if (!color) { 4974e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4975e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4976e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4977c8587f34SStefano Zampini } else { 497852e5ac9dSStefano Zampini Mat subdomain_adj; 497952e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 498052e5ac9dSStefano Zampini MatPartitioning partitioner; 498127b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 498252e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 498357de7509SStefano Zampini PetscMPIInt size; 4984b0c7d250SStefano Zampini PetscBool aggregate; 4985b0c7d250SStefano Zampini 498627b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 498727b6a85dSStefano Zampini if (void_procs) { 498827b6a85dSStefano Zampini PetscInt prank = rank; 4989785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 499027b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 4991e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4992e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4993c8587f34SStefano Zampini } 4994e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 499527b6a85dSStefano Zampini } else { 499627b6a85dSStefano Zampini oldranks = NULL; 499727b6a85dSStefano Zampini } 4998b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 499927b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 5000b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 5001b0c7d250SStefano Zampini PetscMPIInt nrank; 5002b0c7d250SStefano Zampini PetscScalar *vals; 5003b0c7d250SStefano Zampini 500427b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5005b0c7d250SStefano Zampini lrows = 0; 5006b0c7d250SStefano Zampini if (nrank<redprocs) { 5007b0c7d250SStefano Zampini lrows = size/redprocs; 5008b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5009b0c7d250SStefano Zampini } 501027b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5011b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5012b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5013b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5014b0c7d250SStefano Zampini row = nrank; 5015b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5016b0c7d250SStefano Zampini cols = adjncy; 5017b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5018b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5019b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5020b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5021b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 502252e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 502352e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 502452e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5025b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 502627b6a85dSStefano Zampini if (use_vwgt) { 502727b6a85dSStefano Zampini Vec v; 502827b6a85dSStefano Zampini const PetscScalar *array; 502927b6a85dSStefano Zampini PetscInt nl; 503027b6a85dSStefano Zampini 503127b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 503227b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 503327b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 503427b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 503527b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 503627b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 503727b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 503822db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 503927b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 504027b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 504127b6a85dSStefano Zampini } 5042b0c7d250SStefano Zampini } else { 504327b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 504427b6a85dSStefano Zampini if (use_vwgt) { 504527b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 504627b6a85dSStefano Zampini v_wgt[0] = local_size; 504727b6a85dSStefano Zampini } 5048b0c7d250SStefano Zampini } 504922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5050e7931f94SStefano Zampini 5051e7931f94SStefano Zampini /* Partition */ 505227b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5053e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 505427b6a85dSStefano Zampini if (v_wgt) { 5055e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5056c8587f34SStefano Zampini } 505757de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 505857de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5059e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5060e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 506122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5062e7931f94SStefano Zampini 506352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 50646583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 506552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 506652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 506757de7509SStefano Zampini if (!aggregate) { 506857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 506927b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 507027b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 507127b6a85dSStefano Zampini #endif 507257de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 507327b6a85dSStefano Zampini } else if (oldranks) { 5074b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 507527b6a85dSStefano Zampini } else { 507627b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 507757de7509SStefano Zampini } 507828143c3dSStefano Zampini } else { 5079b0c7d250SStefano Zampini PetscInt idxs[1]; 5080b0c7d250SStefano Zampini PetscMPIInt tag; 5081b0c7d250SStefano Zampini MPI_Request *reqs; 5082b0c7d250SStefano Zampini 5083b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5084b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5085b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 508627b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 508728143c3dSStefano Zampini } 508827b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5089b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5090b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 509157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 509227b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 509327b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 509427b6a85dSStefano Zampini #endif 509557de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 509627b6a85dSStefano Zampini } else if (oldranks) { 5097b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 509827b6a85dSStefano Zampini } else { 509927b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5100e7931f94SStefano Zampini } 510157de7509SStefano Zampini } 510252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5103e7931f94SStefano Zampini /* clean up */ 5104e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 510552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5106e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5107e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5108e7931f94SStefano Zampini } 510927b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 511057de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5111e7931f94SStefano Zampini 5112e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5113e7931f94SStefano Zampini i = 1; 511427b6a85dSStefano Zampini if (!color) i=0; 511557de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5116e7931f94SStefano Zampini PetscFunctionReturn(0); 5117e7931f94SStefano Zampini } 5118e7931f94SStefano Zampini 5119e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5120e7931f94SStefano Zampini 5121e7931f94SStefano Zampini #undef __FUNCT__ 5122e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 51231ae86dd6SStefano 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[]) 5124e7931f94SStefano Zampini { 512570cf5478SStefano Zampini Mat local_mat; 5126e7931f94SStefano Zampini IS is_sends_internal; 51279d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 51281ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 51299d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5130e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5131e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5132e7931f94SStefano Zampini const PetscInt* is_indices; 5133e7931f94SStefano Zampini MatType new_local_type; 5134e7931f94SStefano Zampini /* buffers */ 5135e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 513628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 51379d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5138e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 51391ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 5140e7931f94SStefano Zampini /* MPI */ 514128143c3dSStefano Zampini MPI_Comm comm,comm_n; 514228143c3dSStefano Zampini PetscSubcomm subcomm; 5143e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 514428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 514528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 51461ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 51471ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 51481ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 5149e7931f94SStefano Zampini PetscErrorCode ierr; 5150e7931f94SStefano Zampini 5151e7931f94SStefano Zampini PetscFunctionBegin; 515257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5153e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 515428143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 515557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 515657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 515757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 515857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 515957de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 51601ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 51611ae86dd6SStefano Zampini if (nvecs) { 51621ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 51631ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 51641ae86dd6SStefano Zampini } 516557de7509SStefano Zampini /* further checks */ 5166e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5167e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5168e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5169e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5170e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 517157de7509SStefano Zampini if (reuse && *mat_n) { 517270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 517357de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 517470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 517528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 517670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 517770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 517870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 517970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 518070cf5478SStefano Zampini } 5181e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5182e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 518357de7509SStefano Zampini 5184e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5185e7931f94SStefano Zampini if (!is_sends) { 518628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 518757de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5188c8587f34SStefano Zampini } else { 5189e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5190e7931f94SStefano Zampini is_sends_internal = is_sends; 5191c8587f34SStefano Zampini } 5192e7931f94SStefano Zampini 5193e7931f94SStefano Zampini /* get comm */ 5194a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5195e7931f94SStefano Zampini 5196e7931f94SStefano Zampini /* compute number of sends */ 5197e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5198e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5199e7931f94SStefano Zampini 5200e7931f94SStefano Zampini /* compute number of receives */ 5201e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5202785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5203e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5204e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5205e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5206e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5207e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5208e7931f94SStefano Zampini 520928143c3dSStefano Zampini /* restrict comm if requested */ 521028143c3dSStefano Zampini subcomm = 0; 521128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 521228143c3dSStefano Zampini if (restrict_comm) { 5213779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5214779c1cceSStefano Zampini 521528143c3dSStefano Zampini color = 0; 521653a05cb3SStefano Zampini if (restrict_full) { 521753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 521853a05cb3SStefano Zampini } else { 521953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 522053a05cb3SStefano Zampini } 5221b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 522228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 522328143c3dSStefano Zampini /* check if reuse has been requested */ 522457de7509SStefano Zampini if (reuse) { 522528143c3dSStefano Zampini if (*mat_n) { 522628143c3dSStefano Zampini PetscMPIInt subcommsize2; 522728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 522828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 522928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 523028143c3dSStefano Zampini } else { 523128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 523228143c3dSStefano Zampini } 523328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5234779c1cceSStefano Zampini PetscMPIInt rank; 5235779c1cceSStefano Zampini 5236779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 523728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 523828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 523928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5240306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 524128143c3dSStefano Zampini } 524228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 524328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 524428143c3dSStefano Zampini } else { 524528143c3dSStefano Zampini comm_n = comm; 524628143c3dSStefano Zampini } 524728143c3dSStefano Zampini 5248e7931f94SStefano Zampini /* prepare send/receive buffers */ 5249785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5250e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5251785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5252e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 525328143c3dSStefano Zampini if (nis) { 5254854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 525528143c3dSStefano Zampini } 5256e7931f94SStefano Zampini 525728143c3dSStefano Zampini /* Get data from local matrices */ 52586c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5259e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5260e7931f94SStefano Zampini /* 5261e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5262e7931f94SStefano Zampini send_buffer_idxs should contain: 5263e7931f94SStefano Zampini - MatType_PRIVATE type 5264e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5265e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5266e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5267e7931f94SStefano Zampini */ 52686c4ed002SBarry Smith else { 5269e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 52703bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5271854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5272e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5273e7931f94SStefano Zampini send_buffer_idxs[1] = i; 52743bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5275e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 52763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5277e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5278e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5279e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5280e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5281c8587f34SStefano Zampini } 5282c8587f34SStefano Zampini } 5283e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 528428143c3dSStefano Zampini /* additional is (if any) */ 528528143c3dSStefano Zampini if (nis) { 528628143c3dSStefano Zampini PetscMPIInt psum; 528728143c3dSStefano Zampini PetscInt j; 528828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 528928143c3dSStefano Zampini PetscInt plen; 529028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 529128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 529228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 529328143c3dSStefano Zampini } 5294854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 529528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 529628143c3dSStefano Zampini PetscInt plen; 529728143c3dSStefano Zampini const PetscInt *is_array_idxs; 529828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 529928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 530028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 530128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 530228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 530328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 530428143c3dSStefano Zampini } 530528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 530628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 530728143c3dSStefano Zampini } 530828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 530928143c3dSStefano Zampini } 531028143c3dSStefano Zampini 5311e7931f94SStefano Zampini buf_size_idxs = 0; 5312e7931f94SStefano Zampini buf_size_vals = 0; 531328143c3dSStefano Zampini buf_size_idxs_is = 0; 53141ae86dd6SStefano Zampini buf_size_vecs = 0; 5315e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5316e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5317e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 531828143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 53191ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 5320e7931f94SStefano Zampini } 5321785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5322785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 532395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 53241ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 5325e7931f94SStefano Zampini 5326e7931f94SStefano Zampini /* get new tags for clean communications */ 5327e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5328e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 532928143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 53301ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 5331e7931f94SStefano Zampini 5332e7931f94SStefano Zampini /* allocate for requests */ 5333785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5334785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 533595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 53361ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 5337785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5338785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 533995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 53401ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 5341e7931f94SStefano Zampini 5342e7931f94SStefano Zampini /* communications */ 5343e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5344e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 534528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 53461ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 5347e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5348e7931f94SStefano Zampini source_dest = onodes[i]; 5349e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5350e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5351e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5352e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 535328143c3dSStefano Zampini if (nis) { 535457de7509SStefano Zampini source_dest = onodes_is[i]; 535528143c3dSStefano 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); 535628143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 535728143c3dSStefano Zampini } 53581ae86dd6SStefano Zampini if (nvecs) { 53591ae86dd6SStefano Zampini source_dest = onodes[i]; 53601ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 53611ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 53621ae86dd6SStefano Zampini } 5363e7931f94SStefano Zampini } 5364e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5365e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5366e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5367e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 536828143c3dSStefano Zampini if (nis) { 536928143c3dSStefano 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); 537028143c3dSStefano Zampini } 53711ae86dd6SStefano Zampini if (nvecs) { 53721ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 53731ae86dd6SStefano 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); 53741ae86dd6SStefano Zampini } 5375e7931f94SStefano Zampini } 5376e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5377e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5378e7931f94SStefano Zampini 5379e7931f94SStefano Zampini /* assemble new l2g map */ 5380e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5381e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53829d30be91SStefano Zampini new_local_rows = 0; 5383e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53849d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5385e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5386e7931f94SStefano Zampini } 53879d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5388e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53899d30be91SStefano Zampini new_local_rows = 0; 5390e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53919d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 53929d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5393e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5394e7931f94SStefano Zampini } 53959d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 53969d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5397e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5398e7931f94SStefano Zampini 5399e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5400e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5401e7931f94SStefano 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) */ 5402e7931f94SStefano Zampini if (n_recvs) { 540328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5404e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5405e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5406e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5407e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5408e7931f94SStefano Zampini break; 5409e7931f94SStefano Zampini } 5410e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5411e7931f94SStefano Zampini } 5412e7931f94SStefano Zampini switch (new_local_type_private) { 541328143c3dSStefano Zampini case MATDENSE_PRIVATE: 541428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5415e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5416e7931f94SStefano Zampini bs = 1; 541728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 541828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 541928143c3dSStefano Zampini bs = 1; 542028143c3dSStefano Zampini } 5421e7931f94SStefano Zampini break; 5422e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5423e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5424e7931f94SStefano Zampini bs = 1; 5425e7931f94SStefano Zampini break; 5426e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5427e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5428e7931f94SStefano Zampini break; 5429e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5430e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5431e7931f94SStefano Zampini break; 5432e7931f94SStefano Zampini default: 54339d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5434e7931f94SStefano Zampini break; 5435e7931f94SStefano Zampini } 543628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 543728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 543828143c3dSStefano Zampini bs = 1; 5439e7931f94SStefano Zampini } 5440e7931f94SStefano Zampini 544170cf5478SStefano Zampini /* create MATIS object if needed */ 544257de7509SStefano Zampini if (!reuse) { 5443e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5444e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 544570cf5478SStefano Zampini } else { 544670cf5478SStefano Zampini /* it also destroys the local matrices */ 544757de7509SStefano Zampini if (*mat_n) { 544870cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 544957de7509SStefano Zampini } else { /* this is a fake object */ 545057de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 545157de7509SStefano Zampini } 545270cf5478SStefano Zampini } 545370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5454e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 54559d30be91SStefano Zampini 54569d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 54579d30be91SStefano Zampini 54589d30be91SStefano Zampini /* Global to local map of received indices */ 54599d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 54609d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 54619d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 54629d30be91SStefano Zampini 54639d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 54649d30be91SStefano Zampini buf_size_idxs = 0; 54659d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54669d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 54679d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 54689d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 54699d30be91SStefano Zampini } 54709d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 54719d30be91SStefano Zampini 54729d30be91SStefano Zampini /* set preallocation */ 54739d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 54749d30be91SStefano Zampini if (!newisdense) { 54759d30be91SStefano Zampini PetscInt *new_local_nnz=0; 54769d30be91SStefano Zampini 54779d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 54789d30be91SStefano Zampini if (n_recvs) { 54799d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 54809d30be91SStefano Zampini } 54819d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54829d30be91SStefano Zampini PetscInt j; 54839d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 54849d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 54859d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 54869d30be91SStefano Zampini } 54879d30be91SStefano Zampini } else { 54889d30be91SStefano Zampini /* TODO */ 54899d30be91SStefano Zampini } 54909d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 54919d30be91SStefano Zampini } 54929d30be91SStefano Zampini if (new_local_nnz) { 54939d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 54949d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 54959d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 54969d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 54979d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 54989d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 54999d30be91SStefano Zampini } else { 55009d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55019d30be91SStefano Zampini } 55029d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 55039d30be91SStefano Zampini } else { 55049d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55059d30be91SStefano Zampini } 5506e7931f94SStefano Zampini 5507e7931f94SStefano Zampini /* set values */ 5508e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 55099d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5510e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5511e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5512e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 55139d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5514e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5515e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5516e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 551728143c3dSStefano Zampini } else { 551828143c3dSStefano Zampini /* TODO */ 5519e7931f94SStefano Zampini } 5520e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5521e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5522e7931f94SStefano Zampini } 5523e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5524e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 552570cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 552670cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55279d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 5528e7931f94SStefano Zampini 5529dfd14d43SStefano Zampini #if 0 553028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5531e7931f94SStefano Zampini Vec lvec,rvec; 5532e7931f94SStefano Zampini PetscReal infty_error; 5533e7931f94SStefano Zampini 55342a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5535e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5536e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5537e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 553870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5539e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5540e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5541e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5542e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5543e7931f94SStefano Zampini } 554428143c3dSStefano Zampini #endif 5545e7931f94SStefano Zampini 554628143c3dSStefano Zampini /* assemble new additional is (if any) */ 554728143c3dSStefano Zampini if (nis) { 554828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 554928143c3dSStefano Zampini 555028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5551854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 555228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 555328143c3dSStefano Zampini psum = 0; 555428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 555528143c3dSStefano Zampini for (j=0;j<nis;j++) { 555628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 555728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 555828143c3dSStefano Zampini psum += plen; 555928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 556028143c3dSStefano Zampini } 556128143c3dSStefano Zampini } 5562854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5563854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 556428143c3dSStefano Zampini for (i=1;i<nis;i++) { 556528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 556628143c3dSStefano Zampini } 556728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 556828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 556928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 557028143c3dSStefano Zampini for (j=0;j<nis;j++) { 557128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 557228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 557328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 557428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 557528143c3dSStefano Zampini } 557628143c3dSStefano Zampini } 557728143c3dSStefano Zampini for (i=0;i<nis;i++) { 557828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 557928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 558028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 558128143c3dSStefano Zampini } 558228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 558328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 558428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 558528143c3dSStefano Zampini } 5586e7931f94SStefano Zampini /* free workspace */ 558728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5588e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5589e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5590e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5591e7931f94SStefano Zampini if (isdense) { 5592e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5593e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5594e7931f94SStefano Zampini } else { 5595e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5596e7931f94SStefano Zampini } 559728143c3dSStefano Zampini if (nis) { 559828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 559928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 560028143c3dSStefano Zampini } 56011ae86dd6SStefano Zampini 56021ae86dd6SStefano Zampini if (nvecs) { 56031ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56041ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56051ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56061ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56071ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 56081ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 56091ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 56101ae86dd6SStefano Zampini /* set values */ 56111ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 56121ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 56131ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56141ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 56151ae86dd6SStefano Zampini PetscInt j; 56161ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 56171ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 56181ae86dd6SStefano Zampini } 56191ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 56201ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 56211ae86dd6SStefano Zampini } 56221ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56231ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 56241ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 56251ae86dd6SStefano Zampini } 56261ae86dd6SStefano Zampini 56271ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 56281ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5629e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5630e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 56311ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 563228143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5633e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5634e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 56351ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 563628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5637e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5638e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5639e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5640e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5641e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 564228143c3dSStefano Zampini if (nis) { 564328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 564428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 564528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 564628143c3dSStefano Zampini } 564728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 564828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 564928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 565028143c3dSStefano Zampini for (i=0;i<nis;i++) { 565128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 565228143c3dSStefano Zampini } 56531ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 56541ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56551ae86dd6SStefano Zampini } 565653a05cb3SStefano Zampini *mat_n = NULL; 565728143c3dSStefano Zampini } 5658e7931f94SStefano Zampini PetscFunctionReturn(0); 5659e7931f94SStefano Zampini } 5660a57a6d2fSStefano Zampini 566112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5662af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 566312edc857SStefano Zampini 5664c8587f34SStefano Zampini #undef __FUNCT__ 5665c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5666c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5667c8587f34SStefano Zampini { 5668c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5669c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 567020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 56711ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 56729881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 567320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 56746e683305SStefano Zampini IS coarse_is,*isarray; 56756e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 567630368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5677f9eb5b7dSStefano Zampini PC pc_temp; 5678c8587f34SStefano Zampini PCType coarse_pc_type; 5679c8587f34SStefano Zampini KSPType coarse_ksp_type; 5680f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 56814f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 56826e683305SStefano Zampini Mat t_coarse_mat_is; 568357de7509SStefano Zampini PetscInt ncoarse; 568468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 568522bc73bbSStefano Zampini PetscScalar *array; 568657de7509SStefano Zampini MatReuse coarse_mat_reuse; 568757de7509SStefano Zampini PetscBool restr, full_restr, have_void; 56889881197aSStefano Zampini PetscErrorCode ierr; 5689fdc09c96SStefano Zampini 5690c8587f34SStefano Zampini PetscFunctionBegin; 5691c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 569268457ee5SStefano 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 */ 5693fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 56945a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5695fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5696f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5697f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5698f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5699fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 570051bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 570151bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5702dc4bcba2SStefano Zampini PC pc; 5703dc4bcba2SStefano Zampini PetscBool isbddc; 5704dc4bcba2SStefano Zampini 5705dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5706dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5707dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5708dc4bcba2SStefano Zampini if (isbddc) { 570963c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 571063c961adSStefano Zampini } else { 5711727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 571263c961adSStefano Zampini } 5713fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5714fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5715fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5716f4ddd8eeSStefano Zampini } 5717fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5718fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5719f4ddd8eeSStefano Zampini } 572070cf5478SStefano Zampini /* reset any subassembling information */ 572157de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 572270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 572357de7509SStefano Zampini } 57246e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5725fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5726f4ddd8eeSStefano Zampini } 572757de7509SStefano Zampini /* assemble coarse matrix */ 572857de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 572957de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 573057de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 573157de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 573218a45a71SStefano Zampini } else { 573357de7509SStefano Zampini coarse_mat = NULL; 573457de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 57356e683305SStefano Zampini } 5736e7931f94SStefano Zampini 5737abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5738abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5739abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5740abbbba34SStefano Zampini 5741abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 574222bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 574322bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 574422bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 574522bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5746e176bc59SStefano 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); 57476e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 57486e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57496e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5750abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5751abbbba34SStefano Zampini 575257de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 575357de7509SStefano Zampini im_active = !!(pcis->n); 575457de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 575557de7509SStefano Zampini 575657de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 575757de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 575857de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 575957de7509SStefano Zampini coarse_mat_is = NULL; 576057de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 576157de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 576257de7509SStefano Zampini full_restr = PETSC_TRUE; 57631ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 576457de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 576557de7509SStefano Zampini if (multilevel_requested) { 576657de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 576757de7509SStefano Zampini restr = PETSC_FALSE; 576857de7509SStefano Zampini full_restr = PETSC_FALSE; 576957de7509SStefano Zampini } else { 577057de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 577157de7509SStefano Zampini restr = PETSC_TRUE; 577257de7509SStefano Zampini full_restr = PETSC_TRUE; 577357de7509SStefano Zampini } 57744b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 577557de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 577657de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 5777a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 577857de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 5779a198735bSStefano Zampini } else { 5780a198735bSStefano Zampini PetscMPIInt size,rank; 5781a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 5782a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 5783a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 5784a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5785a198735bSStefano Zampini } 578657de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 578757de7509SStefano Zampini PetscInt psum; 578857de7509SStefano Zampini PetscMPIInt size; 578957de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 579057de7509SStefano Zampini else psum = 0; 579157de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 579257de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 579357de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 579457de7509SStefano Zampini } 579557de7509SStefano Zampini /* determine if we can go multilevel */ 579657de7509SStefano Zampini if (multilevel_requested) { 579757de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 579857de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 579957de7509SStefano Zampini } 580057de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 580157de7509SStefano Zampini 5802e4d548c7SStefano Zampini /* dump subassembling pattern */ 5803e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5804e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5805e4d548c7SStefano Zampini } 5806e4d548c7SStefano Zampini 58076e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 580827b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 58096e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 58106e683305SStefano Zampini const PetscInt *idxs; 58116e683305SStefano Zampini ISLocalToGlobalMapping tmap; 58126e683305SStefano Zampini 58136e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 58140be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 58156e683305SStefano Zampini /* allocate space for temporary storage */ 5816854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5817854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 58186e683305SStefano Zampini /* allocate for IS array */ 58196e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 58206e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 582127b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 582230368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5823854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 58246e683305SStefano Zampini /* dofs splitting */ 58256e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 58266e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 58276e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 58286e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58296e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58306e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58316e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 583230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 58336e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 58346e683305SStefano Zampini } 58356e683305SStefano Zampini /* neumann boundaries */ 58366e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 58376e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 58386e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 58396e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58406e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58416e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58426e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 584330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 58446e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 58456e683305SStefano Zampini } 58466e683305SStefano Zampini /* free memory */ 58476e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 58486e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 58496e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 58506e683305SStefano Zampini } else { 58516e683305SStefano Zampini nis = 0; 58526e683305SStefano Zampini nisdofs = 0; 58536e683305SStefano Zampini nisneu = 0; 585430368db7SStefano Zampini nisvert = 0; 58556e683305SStefano Zampini isarray = NULL; 58566e683305SStefano Zampini } 58576e683305SStefano Zampini /* destroy no longer needed map */ 58586e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 58596e683305SStefano Zampini 586057de7509SStefano Zampini /* subassemble */ 586157de7509SStefano Zampini if (multilevel_allowed) { 58621ae86dd6SStefano Zampini Vec vp[1]; 58631ae86dd6SStefano Zampini PetscInt nvecs = 0; 586457de7509SStefano Zampini PetscBool reuse,reuser; 58651ae86dd6SStefano Zampini 586657de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 586757de7509SStefano Zampini else reuse = PETSC_FALSE; 586857de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 58691ae86dd6SStefano Zampini vp[0] = NULL; 58701ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 58711ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 58721ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 58731ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 58741ae86dd6SStefano Zampini nvecs = 1; 58751ae86dd6SStefano Zampini 58761ae86dd6SStefano Zampini if (pcbddc->divudotp) { 5877a198735bSStefano Zampini Mat B,loc_divudotp; 58781ae86dd6SStefano Zampini Vec v,p; 58791ae86dd6SStefano Zampini IS dummy; 58801ae86dd6SStefano Zampini PetscInt np; 58811ae86dd6SStefano Zampini 5882a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 5883a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 58841ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 5885a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 58861ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 58871ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 58881ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 58891ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 58901ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 58911ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 58921ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 58931ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 58941ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 58951ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 58961ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 58971ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 589874e2c79eSStefano Zampini } 58991ae86dd6SStefano Zampini } 59001ae86dd6SStefano Zampini if (reuser) { 59011ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 590274e2c79eSStefano Zampini } else { 59031ae86dd6SStefano 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); 59041ae86dd6SStefano Zampini } 59051ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 59061ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 59071ae86dd6SStefano Zampini PetscInt nl; 59081ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 59091ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 59101ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59111ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 59121ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 59131ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 59141ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59151ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 5916a198735bSStefano Zampini } else { 5917a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 59181ae86dd6SStefano Zampini } 59191ae86dd6SStefano Zampini } else { 59201ae86dd6SStefano 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); 59216e683305SStefano Zampini } 592257de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 592357de7509SStefano Zampini PetscMPIInt size; 592457de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 592557de7509SStefano Zampini if (!multilevel_allowed) { 592657de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 59276e683305SStefano Zampini } else { 592857de7509SStefano Zampini Mat A; 5929779c1cceSStefano Zampini 593057de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 593157de7509SStefano Zampini if (coarse_mat_is) { 593257de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 593357de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 593457de7509SStefano Zampini coarse_mat = coarse_mat_is; 593557de7509SStefano Zampini } 593657de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 593757de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 593857de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5939779c1cceSStefano Zampini } 5940779c1cceSStefano Zampini } 594157de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 594257de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 59436e683305SStefano Zampini 59446e683305SStefano Zampini /* create local to global scatters for coarse problem */ 594568457ee5SStefano Zampini if (compute_vecs) { 59466e683305SStefano Zampini PetscInt lrows; 59476e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 594857de7509SStefano Zampini if (coarse_mat) { 594957de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 59506e683305SStefano Zampini } else { 59516e683305SStefano Zampini lrows = 0; 59526e683305SStefano Zampini } 59536e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 59546e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 59556e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 59566e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59576e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59586e683305SStefano Zampini } 59596e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5960c8587f34SStefano Zampini 5961f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5962f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5963f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5964f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5965f9eb5b7dSStefano Zampini } else { 5966f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5967f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5968c8587f34SStefano Zampini } 5969c8587f34SStefano Zampini 59706e683305SStefano Zampini /* print some info if requested */ 59716e683305SStefano Zampini if (pcbddc->dbg_flag) { 59726e683305SStefano Zampini if (!multilevel_allowed) { 59736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 59746e683305SStefano Zampini if (multilevel_requested) { 59756e683305SStefano 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); 59766e683305SStefano Zampini } else if (pcbddc->max_levels) { 59776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 59786e683305SStefano Zampini } 59796e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59806e683305SStefano Zampini } 59816e683305SStefano Zampini } 59826e683305SStefano Zampini 5983f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 598457de7509SStefano Zampini if (coarse_mat) { 59856a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 59866e683305SStefano Zampini if (pcbddc->dbg_flag) { 598757de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 59886e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 59896e683305SStefano Zampini } 5990f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5991312be037SStefano Zampini char prefix[256],str_level[16]; 5992e604994aSStefano Zampini size_t len; 599357de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5994422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5995c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5996f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 599757de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5998c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 59996e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 6000c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 6001c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6002e604994aSStefano Zampini /* prefix */ 6003e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 6004e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 6005e604994aSStefano Zampini if (!pcbddc->current_level) { 6006e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 6007e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 6008c8587f34SStefano Zampini } else { 6009e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 6010312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 6011312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 601234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 6013312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 6014e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 6015e604994aSStefano Zampini } 6016e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 60173e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 60183e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 60193e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 60203e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 6021f9eb5b7dSStefano Zampini /* allow user customization */ 6022f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 60233e3c6dadSStefano Zampini } 60243e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 602551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 60263e3c6dadSStefano Zampini if (nisdofs) { 60273e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 60283e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 60293e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 60303e3c6dadSStefano Zampini } 60313e3c6dadSStefano Zampini } 60323e3c6dadSStefano Zampini if (nisneu) { 60333e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 60343e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 6035312be037SStefano Zampini } 603630368db7SStefano Zampini if (nisvert) { 603730368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 603830368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 603930368db7SStefano Zampini } 6040f9eb5b7dSStefano Zampini 6041f9eb5b7dSStefano Zampini /* get some info after set from options */ 6042f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 6043f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 60444f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 60456e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 6046f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6047f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 6048f9eb5b7dSStefano Zampini } 604939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 60504f3a063dSStefano Zampini if (isredundant) { 60514f3a063dSStefano Zampini KSP inner_ksp; 60524f3a063dSStefano Zampini PC inner_pc; 60534f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 60544f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 60554f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 60564f3a063dSStefano Zampini } 6057f9eb5b7dSStefano Zampini 605857de7509SStefano Zampini /* parameters which miss an API */ 605957de7509SStefano Zampini if (isbddc) { 6060720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 6061720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 606257de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 606327b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 606427b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 6065a198735bSStefano Zampini Mat coarsedivudotp_is; 6066a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 6067a198735bSStefano Zampini IS row,col; 6068a198735bSStefano Zampini const PetscInt *gidxs; 6069a198735bSStefano Zampini PetscInt n,st,M,N; 6070a198735bSStefano Zampini 6071a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 6072a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 6073a198735bSStefano Zampini st = st-n; 6074a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 6075a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 6076a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 6077a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6078a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 6079a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6080a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 6081a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 6082a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 6083a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 6084a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 6085a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 6086a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 6087a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 6088a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 6089a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 6090a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 6091a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 6092a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 6093a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 60948ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 6095a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 6096720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 609759e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 6098720d30f9SStefano Zampini } 6099d4d8cf7bSStefano Zampini } 61009881197aSStefano Zampini 61013301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 61025a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 61033301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 61043301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 61053301b35fSStefano Zampini } 61063301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 61073301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 61083301b35fSStefano Zampini } 61093301b35fSStefano Zampini if (pc->pmat->spd_set) { 61103301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 61113301b35fSStefano Zampini } 611227b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 611327b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 611427b6a85dSStefano Zampini } 61156e683305SStefano Zampini /* set operators */ 61165f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 61176e683305SStefano Zampini if (pcbddc->dbg_flag) { 61186e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 61196e683305SStefano Zampini } 61206e683305SStefano Zampini } 61216e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6122b1ecc7b1SStefano Zampini #if 0 6123b9b85e73SStefano Zampini { 6124b9b85e73SStefano Zampini PetscViewer viewer; 6125b9b85e73SStefano Zampini char filename[256]; 6126b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6127b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 61286a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6129b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6130f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6131b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6132b9b85e73SStefano Zampini } 6133b9b85e73SStefano Zampini #endif 6134f9eb5b7dSStefano Zampini 613598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 613698a51de6SStefano Zampini Vec crhs,csol; 613704708bb6SStefano Zampini 6138f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6139f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6140f347579bSStefano Zampini if (!csol) { 61412a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6142f9eb5b7dSStefano Zampini } 6143f347579bSStefano Zampini if (!crhs) { 61442a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6145f347579bSStefano Zampini } 6146b0f5fe93SStefano Zampini } 61471ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 6148b0f5fe93SStefano Zampini 6149b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6150b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6151b0f5fe93SStefano Zampini 6152b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 61534f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 61544f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 61554f1b2e48SStefano Zampini } 6156b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6157b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6158b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6159b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6160b0f5fe93SStefano Zampini if (coarse_mat) { 6161b0f5fe93SStefano Zampini Vec nullv; 6162b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6163b0f5fe93SStefano Zampini PetscInt nl; 6164b0f5fe93SStefano Zampini 6165b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6166b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6167b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6168b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6169b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6170b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6171b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6172b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6173b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6174b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6175b0f5fe93SStefano Zampini } 6176b0f5fe93SStefano Zampini } 6177b0f5fe93SStefano Zampini 6178b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6179b0f5fe93SStefano Zampini PetscBool ispreonly; 6180b0f5fe93SStefano Zampini 6181b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6182b0f5fe93SStefano Zampini PetscBool isnull; 6183b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6184bef83e63SStefano Zampini if (isnull) { 6185b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6186b0f5fe93SStefano Zampini } 6187bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6188b0f5fe93SStefano Zampini } 6189b0f5fe93SStefano Zampini /* setup coarse ksp */ 6190b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6191cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6192cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 61936e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6194c8587f34SStefano Zampini KSP check_ksp; 61952b510759SStefano Zampini KSPType check_ksp_type; 6196c8587f34SStefano Zampini PC check_pc; 61976e683305SStefano Zampini Vec check_vec,coarse_vec; 61986a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 61992b510759SStefano Zampini PetscInt its; 62006e683305SStefano Zampini PetscBool compute_eigs; 62016e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 62026e683305SStefano Zampini PetscInt neigs; 62038e185a42SStefano Zampini const char *prefix; 6204c8587f34SStefano Zampini 62052b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 62066e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6207422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 620823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6209f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6210e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6211e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6212e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 62132b510759SStefano Zampini if (ispreonly) { 62142b510759SStefano Zampini check_ksp_type = KSPPREONLY; 62156e683305SStefano Zampini compute_eigs = PETSC_FALSE; 62162b510759SStefano Zampini } else { 6217cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 62186e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6219c8587f34SStefano Zampini } 6220c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 62216e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 62226e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 62236e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6224a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6225a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6226a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6227a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6228c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6229c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6230c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6231c8587f34SStefano Zampini /* create random vec */ 62322701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6233c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 62346e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6235c8587f34SStefano Zampini /* solve coarse problem */ 62366e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6237cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 62386e683305SStefano Zampini if (compute_eigs) { 6239854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6240854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 62416e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 62421ae86dd6SStefano Zampini if (neigs) { 62436e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 62446e683305SStefano Zampini lambda_min = eigs_r[0]; 62456e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 62462701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 62472701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6248cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6249cbcc2c2aSStefano Zampini } 6250c8587f34SStefano Zampini } 6251c8587f34SStefano Zampini } 62521ae86dd6SStefano Zampini } 6253cbcc2c2aSStefano Zampini 6254c8587f34SStefano Zampini /* check coarse problem residual error */ 62556e683305SStefano Zampini if (pcbddc->dbg_flag) { 62566e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 62576e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62586e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6259c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 62606e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 62616e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6262779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 62636e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 62646e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 62656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 62666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6267b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6268b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6269b0f5fe93SStefano Zampini } 62706e683305SStefano Zampini if (compute_eigs) { 62716e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6272deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6273c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 62746e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 62756e683305SStefano 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); 62766e683305SStefano Zampini for (i=0;i<neigs;i++) { 62776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6278c8587f34SStefano Zampini } 62796e683305SStefano Zampini } 62806e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 62816e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62826e683305SStefano Zampini } 6283e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 62842701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6285c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 62866e683305SStefano Zampini if (compute_eigs) { 62876e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 62886e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6289c8587f34SStefano Zampini } 62906e683305SStefano Zampini } 62916e683305SStefano Zampini } 6292bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6293cbcc2c2aSStefano Zampini /* print additional info */ 6294cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 62956e683305SStefano Zampini /* waits until all processes reaches this point */ 62966e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6297cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6298cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6299cbcc2c2aSStefano Zampini } 6300cbcc2c2aSStefano Zampini 63012b510759SStefano Zampini /* free memory */ 6302fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6303c8587f34SStefano Zampini PetscFunctionReturn(0); 6304c8587f34SStefano Zampini } 6305674ae819SStefano Zampini 6306f34684f1SStefano Zampini #undef __FUNCT__ 6307f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6308f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6309f34684f1SStefano Zampini { 6310f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6311f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6312f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6313dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6314dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 631573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6316dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6317f34684f1SStefano Zampini PetscErrorCode ierr; 6318f34684f1SStefano Zampini 6319f34684f1SStefano Zampini PetscFunctionBegin; 6320f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 63216c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6322dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 63233bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6324dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6325dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 63266583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6327dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6328dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6329dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 63306c4ed002SBarry 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); 6331dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6332dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6333dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6334dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6335dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6336f34684f1SStefano Zampini 6337f34684f1SStefano Zampini /* check numbering */ 6338f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6339019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6340dc456d91SStefano Zampini PetscInt i; 6341b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6342f34684f1SStefano Zampini 6343f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6344f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6345f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 63461575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6347019a44ceSStefano Zampini /* counter */ 6348019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6349019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6350019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6351019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6352019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6353019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6354f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6355f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6356727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6357f34684f1SStefano Zampini } 6358f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6359f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6360f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6361e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6362e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6363e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6364e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6365f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6366019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6367f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6368019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 63692c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 637075c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6371b9b85e73SStefano Zampini set_error = PETSC_TRUE; 63722c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 63732c66d082SStefano 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); 6374f34684f1SStefano Zampini } 6375f34684f1SStefano Zampini } 6376019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6377b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6378f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6379f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6380f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6381f34684f1SStefano Zampini } 6382f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6383f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6384e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6385e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6386f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6387f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6388b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6389ca8b9ea9SStefano Zampini PetscInt *gidxs; 6390ca8b9ea9SStefano Zampini 6391ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 63923bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6393f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6394f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6395f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6396f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 63974bc2dc4bSStefano 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); 6398f34684f1SStefano Zampini } 6399f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6400ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6401f34684f1SStefano Zampini } 6402f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 64031575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6404302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6405f34684f1SStefano Zampini } 64068bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6407f34684f1SStefano Zampini /* get back data */ 6408f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6409f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6410674ae819SStefano Zampini PetscFunctionReturn(0); 6411674ae819SStefano Zampini } 6412674ae819SStefano Zampini 6413e456f2a8SStefano Zampini #undef __FUNCT__ 6414e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6415a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6416e456f2a8SStefano Zampini { 6417e456f2a8SStefano Zampini IS localis_t; 6418a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6419e456f2a8SStefano Zampini PetscScalar *vals; 6420e456f2a8SStefano Zampini PetscErrorCode ierr; 6421e456f2a8SStefano Zampini 6422e456f2a8SStefano Zampini PetscFunctionBegin; 6423a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6424e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6425854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6426e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6427e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6428a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6429a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 64301035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6431a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 64321035eff8SStefano Zampini } 6433a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6434e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6435e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6436a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6437a7dc3881SStefano Zampini /* now compute set in local ordering */ 6438a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6439a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6440a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6441a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6442a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6443ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6444e456f2a8SStefano Zampini lsize++; 6445e456f2a8SStefano Zampini } 6446e456f2a8SStefano Zampini } 6447854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6448a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6449ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6450e456f2a8SStefano Zampini idxs[lsize++] = i; 6451e456f2a8SStefano Zampini } 6452e456f2a8SStefano Zampini } 6453a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6454a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6455e456f2a8SStefano Zampini *localis = localis_t; 6456e456f2a8SStefano Zampini PetscFunctionReturn(0); 6457e456f2a8SStefano Zampini } 6458906d46d4SStefano Zampini 6459b96c3477SStefano Zampini #undef __FUNCT__ 6460b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 646108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6462b96c3477SStefano Zampini { 6463a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6464b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6465b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6466a64f4aa4SStefano Zampini Mat S_j; 6467b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6468b96c3477SStefano Zampini PetscBool free_used_adj; 6469b96c3477SStefano Zampini PetscErrorCode ierr; 6470b96c3477SStefano Zampini 6471b96c3477SStefano Zampini PetscFunctionBegin; 6472b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6473b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 647408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6475b96c3477SStefano Zampini used_xadj = NULL; 6476b96c3477SStefano Zampini used_adjncy = NULL; 6477b96c3477SStefano Zampini } else { 647808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 647908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 648008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 648108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6482b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6483b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6484b96c3477SStefano Zampini } else { 64852fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6486b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6487b96c3477SStefano Zampini PetscInt nvtxs; 6488b96c3477SStefano Zampini 64892fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 64902fffb893SStefano Zampini if (flg_row) { 6491b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6492b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6493b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6494b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 64952fffb893SStefano Zampini } else { 64962fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 64972fffb893SStefano Zampini used_xadj = NULL; 64982fffb893SStefano Zampini used_adjncy = NULL; 64992fffb893SStefano Zampini } 65002fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6501b96c3477SStefano Zampini } 6502b96c3477SStefano Zampini } 6503d5574798SStefano Zampini 6504d5574798SStefano Zampini /* setup sub_schurs data */ 6505a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6506df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6507df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6508a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 650991af6908SStefano 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); 6510a64f4aa4SStefano Zampini } else { 65116816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 65124d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 6513a3df083aSStefano Zampini PetscInt benign_n; 651472b8c272SStefano Zampini Mat change = NULL; 65159d54b7f4SStefano Zampini Vec scaling = NULL; 651672b8c272SStefano Zampini IS change_primal = NULL; 6517a3df083aSStefano Zampini 65185feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 65195feab87aSStefano Zampini PetscInt n_vertices; 65205feab87aSStefano Zampini 65215feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 65222034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 65235feab87aSStefano Zampini } 652404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 652504708bb6SStefano Zampini if (!isseqaij) { 652604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 652704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 652804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 652904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 653004708bb6SStefano Zampini } else { 6531511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 653204708bb6SStefano Zampini } 653304708bb6SStefano Zampini } 6534a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6535a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6536ca92afb2SStefano Zampini } else { 6537a3df083aSStefano Zampini benign_n = 0; 6538ca92afb2SStefano Zampini } 6539b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6540b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6541b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 654272b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 654322db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6544b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 654522db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6546b7ab4a40SStefano Zampini } 6547b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6548b7ab4a40SStefano 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 */ 6549b7ab4a40SStefano Zampini if (need_change) { 655088c03ad3SStefano Zampini PC_IS *pcisf; 655188c03ad3SStefano Zampini PC_BDDC *pcbddcf; 655288c03ad3SStefano Zampini PC pcf; 655388c03ad3SStefano Zampini 6554e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 655588c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 655688c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 655788c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 655888c03ad3SStefano Zampini /* hacks */ 655988c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 656072b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 656172b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 656272b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 656372b8c272SStefano Zampini pcisf->n = pcis->n; 656472b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 656588c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 656688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 656788c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 656888c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 656988c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 657088c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 657172b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 657288c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 657388c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 657472b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 657572b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 657672b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 657772b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 657872b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 657988c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 658072b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 658188c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 658288c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 658388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 658488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 658588c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 658688c03ad3SStefano Zampini pcf->ops->destroy = NULL; 658788c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 658888c03ad3SStefano Zampini } 65899d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 659091af6908SStefano 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); 659172b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 659272b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6593ca92afb2SStefano Zampini } 6594d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6595b96c3477SStefano Zampini 6596b96c3477SStefano Zampini /* free adjacency */ 6597b96c3477SStefano Zampini if (free_used_adj) { 6598b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6599b96c3477SStefano Zampini } 6600b96c3477SStefano Zampini PetscFunctionReturn(0); 6601b96c3477SStefano Zampini } 6602b96c3477SStefano Zampini 6603b96c3477SStefano Zampini #undef __FUNCT__ 6604b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 660508122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6606b96c3477SStefano Zampini { 6607b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6608b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6609b96c3477SStefano Zampini PCBDDCGraph graph; 6610b96c3477SStefano Zampini PetscErrorCode ierr; 6611b96c3477SStefano Zampini 6612b96c3477SStefano Zampini PetscFunctionBegin; 6613b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 661408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 66153301b35fSStefano Zampini IS verticesIS,verticescomm; 66163301b35fSStefano Zampini PetscInt vsize,*idxs; 6617b96c3477SStefano Zampini 6618b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 66193301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 66203301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66213301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 66223301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66233301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6624b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 66257fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6626441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 66273301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6628b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6629b96c3477SStefano Zampini } else { 6630b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6631b96c3477SStefano Zampini } 6632e4d548c7SStefano Zampini /* print some info */ 6633e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6634e4d548c7SStefano Zampini IS vertices; 6635e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 66368ae0ca82SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag-1,pcbddc->dbg_viewer);CHKERRQ(ierr); 6637e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6638e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6639e4d548c7SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 6640e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6641e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6642e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6643e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6644e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6645e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6646e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6647e4d548c7SStefano Zampini } 6648b96c3477SStefano Zampini 6649b96c3477SStefano Zampini /* sub_schurs init */ 6650b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 6651b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 6652b334f244SStefano Zampini } 6653b334f244SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6654a64f4aa4SStefano Zampini 6655b96c3477SStefano Zampini /* free graph struct */ 665608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6657b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6658b96c3477SStefano Zampini } 6659b96c3477SStefano Zampini PetscFunctionReturn(0); 6660b96c3477SStefano Zampini } 6661fa34dd3eSStefano Zampini 6662fa34dd3eSStefano Zampini #undef __FUNCT__ 6663fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6664fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6665fa34dd3eSStefano Zampini { 6666fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6667fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6668fa34dd3eSStefano Zampini PetscErrorCode ierr; 6669fa34dd3eSStefano Zampini 6670fa34dd3eSStefano Zampini PetscFunctionBegin; 6671fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6672fa34dd3eSStefano Zampini IS zerodiag = NULL; 66734f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6674fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 66754f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 667675c01103SStefano Zampini PetscReal norm; 6677fa34dd3eSStefano Zampini PetscInt i; 6678fa34dd3eSStefano Zampini 6679fa34dd3eSStefano Zampini /* B0 and B0_B */ 6680fa34dd3eSStefano Zampini if (zerodiag) { 6681fa34dd3eSStefano Zampini IS dummy; 6682fa34dd3eSStefano Zampini 66834f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 66844f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6685fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6686fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6687fa34dd3eSStefano Zampini } 6688fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6689fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6690fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6691fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6692fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6693fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6694fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6695fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6696fa34dd3eSStefano Zampini /* S_j */ 6697fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6698fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6699fa34dd3eSStefano Zampini 6700fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6701fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6702fa34dd3eSStefano Zampini /* continuous in primal space */ 6703fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6704fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6705fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6706fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 67074f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 67084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6709fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6710fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6711fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6712fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6713fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6714fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6715fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6716fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6717fa34dd3eSStefano Zampini 6718fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6719fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6720fa34dd3eSStefano Zampini /* local with Schur */ 6721fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6722fa34dd3eSStefano Zampini if (zerodiag) { 6723fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6725fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6726fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6727fa34dd3eSStefano Zampini } 6728fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6729fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6730fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6731fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6732fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6733fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6734fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6735fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6736fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6737fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6738fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6739fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6740fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6741fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6742fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6743fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6744fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6745fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6746fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6747fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6748fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6749fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6750fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6751fa34dd3eSStefano Zampini if (zerodiag) { 6752fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6753fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67544f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6755fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6756fa34dd3eSStefano Zampini } 6757fa34dd3eSStefano Zampini /* BDDC */ 6758fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6759fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6760fa34dd3eSStefano Zampini 6761fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6762fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6763fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6764fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 67654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 67664f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6767fa34dd3eSStefano Zampini } 67684f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6769fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6770fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6771fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6772fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6773fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6774fa34dd3eSStefano Zampini } 6775fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6776fa34dd3eSStefano Zampini } 6777