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); 6994d7f8f00SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) if ((PetscInt)(PetscRealPart(pcbddc->benign_p0[i])) != -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); 7005408967cSStefano Zampini PetscFunctionReturn(0); 7015408967cSStefano Zampini } 7025408967cSStefano Zampini 7035408967cSStefano Zampini #undef __FUNCT__ 704339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 705339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 706339f8db1SStefano Zampini { 707339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7084f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 709b0f5fe93SStefano Zampini PetscInt nz,n; 7101f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 7114f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 712339f8db1SStefano Zampini PetscErrorCode ierr; 713339f8db1SStefano Zampini 714339f8db1SStefano Zampini PetscFunctionBegin; 7159f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 7169f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 717a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 718a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 719a3df083aSStefano Zampini } 720a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 721a3df083aSStefano Zampini pcbddc->benign_n = 0; 7221ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 7234f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 7244f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 7254f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 7261ae86dd6SStefano Zampini since the local Schur complements are already SPD 7274f1b2e48SStefano Zampini */ 7284f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 7294f1b2e48SStefano Zampini have_null = PETSC_TRUE; 73040fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 7314f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 7324f1b2e48SStefano Zampini 7334f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 7344f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 7354f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 7364f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 737ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 73840fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 73940fa8d13SStefano Zampini if (!sorted) { 74040fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 74140fa8d13SStefano Zampini } 74240fa8d13SStefano Zampini } else { 74340fa8d13SStefano Zampini pressures = NULL; 74440fa8d13SStefano Zampini } 74597d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 74697d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 74727b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 74897d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 749339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 750339f8db1SStefano Zampini if (!sorted) { 751339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 752339f8db1SStefano Zampini } 753339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 7544f1b2e48SStefano Zampini if (!nz) { 7554f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 7564f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 75740fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 75840fa8d13SStefano Zampini } 7594f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 7604f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 7614f1b2e48SStefano Zampini zerodiag_subs = NULL; 7624f1b2e48SStefano Zampini pcbddc->benign_n = 0; 7631f4df5f7SStefano Zampini n_interior_dofs = 0; 7641f4df5f7SStefano Zampini interior_dofs = NULL; 7651f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 7661f4df5f7SStefano Zampini PetscInt n,i,j; 7671f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 7681f4df5f7SStefano Zampini PetscInt *iwork; 7691f4df5f7SStefano Zampini 7701f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 7711f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7721f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 7731f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 7741f4df5f7SStefano Zampini for (i=0;i<n_neigh;i++) 7751f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 7761f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 7771f4df5f7SStefano Zampini for (i=0;i<n;i++) 7781f4df5f7SStefano Zampini if (!iwork[i]) 7791f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 7801f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 7811f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 7821f4df5f7SStefano Zampini } 7834f1b2e48SStefano Zampini if (has_null_pressures) { 7844f1b2e48SStefano Zampini IS *subs; 7851f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 7861f4df5f7SStefano Zampini const PetscInt *idxs; 7871f4df5f7SStefano Zampini PetscScalar *array; 7881f4df5f7SStefano Zampini Vec *work; 7891f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 7904f1b2e48SStefano Zampini 7914f1b2e48SStefano Zampini subs = pcbddc->local_subs; 7924f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 7931f4df5f7SStefano 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) */ 7941f4df5f7SStefano Zampini if (pcbddc->current_level) { 7951f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 7961f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 7971f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 7981f4df5f7SStefano Zampini /* work[0] = 1_p */ 7991f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 8001f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 8011f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8021f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 8031f4df5f7SStefano Zampini /* work[0] = 1_v */ 8041f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 8051f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 8061f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 8071f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 8081f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 8091f4df5f7SStefano Zampini } 8104f1b2e48SStefano Zampini if (nsubs > 1) { 8114f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 8124f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 8134f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 8144f1b2e48SStefano Zampini IS t_zerodiag_subs; 8154f1b2e48SStefano Zampini PetscInt nl; 8164f1b2e48SStefano Zampini 8174f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 8184f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 8194f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 8204f1b2e48SStefano Zampini if (nl) { 8214f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8224f1b2e48SStefano Zampini 8231f4df5f7SStefano Zampini if (pcbddc->current_level) { 8241f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 8251f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 8261f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 8271f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8281f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 8291f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8301f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 8311f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 8321f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 8331f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 8341f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 8351f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8361f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8371f4df5f7SStefano Zampini valid = PETSC_FALSE; 8381f4df5f7SStefano Zampini break; 8391f4df5f7SStefano Zampini } 8401f4df5f7SStefano Zampini } 8411f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 8421f4df5f7SStefano Zampini } 8431f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 8441f4df5f7SStefano Zampini IS t_bc; 8451f4df5f7SStefano Zampini PetscInt nzb; 8461f4df5f7SStefano Zampini 8471f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 8481f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 8491f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 8501f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8511f4df5f7SStefano Zampini } 8521f4df5f7SStefano Zampini if (valid && pressures) { 8534f1b2e48SStefano Zampini IS t_pressure_subs; 8544f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 8554f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 8564f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 8574f1b2e48SStefano Zampini } 8584f1b2e48SStefano Zampini if (valid) { 8594f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 8604f1b2e48SStefano Zampini pcbddc->benign_n++; 8614f1b2e48SStefano Zampini } else { 8624f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 8634f1b2e48SStefano Zampini } 8644f1b2e48SStefano Zampini } 8654f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 8664f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 8674f1b2e48SStefano Zampini } 8684f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 8694f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 8701f4df5f7SStefano Zampini 8711f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 8721f4df5f7SStefano Zampini PetscInt nzb; 8731f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 8741f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 8751f4df5f7SStefano Zampini } 8761f4df5f7SStefano Zampini if (valid && pressures) { 8774f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 8784f1b2e48SStefano Zampini } 8791f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 8801f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 8811f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 8821f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 8831f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 8841f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 8851f4df5f7SStefano Zampini valid = PETSC_FALSE; 8861f4df5f7SStefano Zampini break; 8871f4df5f7SStefano Zampini } 8881f4df5f7SStefano Zampini } 8891f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 8901f4df5f7SStefano Zampini } 8914f1b2e48SStefano Zampini if (valid) { 8924f1b2e48SStefano Zampini pcbddc->benign_n = 1; 893ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 8944f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 8954f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 8964f1b2e48SStefano Zampini } 8974f1b2e48SStefano Zampini } 8981f4df5f7SStefano Zampini if (pcbddc->current_level) { 8991f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 9004f1b2e48SStefano Zampini } 9011f4df5f7SStefano Zampini } 9021f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 9034f1b2e48SStefano Zampini 9044f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 905b9b0e38cSStefano Zampini PetscInt n; 906b9b0e38cSStefano Zampini 9074f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9084f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 909b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 910b9b0e38cSStefano Zampini if (n) { 9114f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 9124f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9134f1b2e48SStefano Zampini } 914b9b0e38cSStefano Zampini } 9154f1b2e48SStefano Zampini 9164f1b2e48SStefano Zampini /* final check for null pressures */ 9174f1b2e48SStefano Zampini if (zerodiag && pressures) { 9184f1b2e48SStefano Zampini PetscInt nz,np; 9194f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 9204f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 9214f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 9224f1b2e48SStefano Zampini } 9234f1b2e48SStefano Zampini 9244f1b2e48SStefano Zampini if (recompute_zerodiag) { 9254f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 9264f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 9274f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 9284f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 9294f1b2e48SStefano Zampini } else { 9304f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 9314f1b2e48SStefano Zampini 9324f1b2e48SStefano Zampini nzn = 0; 9334f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9344f1b2e48SStefano Zampini PetscInt ns; 9354f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9364f1b2e48SStefano Zampini nzn += ns; 9374f1b2e48SStefano Zampini } 9384f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 9394f1b2e48SStefano Zampini nzn = 0; 9404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9414f1b2e48SStefano Zampini PetscInt ns,*idxs; 9424f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 9434f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9444f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 9454f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 9464f1b2e48SStefano Zampini nzn += ns; 9474f1b2e48SStefano Zampini } 9484f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 9494f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 9504f1b2e48SStefano Zampini } 9514f1b2e48SStefano Zampini have_null = PETSC_FALSE; 9524f1b2e48SStefano Zampini } 9534f1b2e48SStefano Zampini 954669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 955a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 956a198735bSStefano Zampini Mat A,loc_divudotp; 957a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 958a198735bSStefano Zampini IS row,col,isused = NULL; 959a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 960a198735bSStefano Zampini 9611f4df5f7SStefano Zampini if (pressures) { 9621f4df5f7SStefano Zampini isused = pressures; 9631f4df5f7SStefano Zampini } else { 9641f4df5f7SStefano Zampini isused = zerodiag; 9651f4df5f7SStefano Zampini } 966a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 967669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 9681ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 9691ae86dd6SStefano 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"); 970a198735bSStefano Zampini n_isused = 0; 971a198735bSStefano Zampini if (isused) { 972a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 973a198735bSStefano Zampini } 974a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 975a198735bSStefano Zampini st = st-n_isused; 9761ae86dd6SStefano Zampini if (n) { 977a198735bSStefano Zampini const PetscInt *gidxs; 978a198735bSStefano Zampini 979a198735bSStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 980a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 981a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 982a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 983a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 984a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 9851ae86dd6SStefano Zampini } else { 986a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 987a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 988a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 989a198735bSStefano Zampini } 990a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 991a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 992a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 993a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 994a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 995a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 996a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 997a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 998a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 999a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 1000a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 1001a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 1002a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 1003a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 10041ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10051ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 10061ae86dd6SStefano Zampini } 1007b3afcdbeSStefano Zampini 1008b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 10094f1b2e48SStefano Zampini if (has_null_pressures) { 10104f1b2e48SStefano Zampini IS zerodiagc; 10114f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 10124f1b2e48SStefano Zampini PetscInt i,s,*nnz; 10134f1b2e48SStefano Zampini 10144f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 1015339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 1016339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 1017339f8db1SStefano Zampini /* local change of basis for pressures */ 1018339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 101997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 1020339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 1021339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1022339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 10234f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 10244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 10254f1b2e48SStefano Zampini PetscInt nzs,j; 10264f1b2e48SStefano Zampini 10274f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 10284f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 10294f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 10304f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 10314f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 10324f1b2e48SStefano Zampini } 1033339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 1034339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1035339f8db1SStefano Zampini /* set identity on velocities */ 1036339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 1037339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 1038339f8db1SStefano Zampini } 10394f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 10404f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 10419f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 10424f1b2e48SStefano 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); 1043339f8db1SStefano Zampini /* set change on pressures */ 10444f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 10454f1b2e48SStefano Zampini PetscScalar *array; 10464f1b2e48SStefano Zampini PetscInt nzs; 10474f1b2e48SStefano Zampini 10484f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 10494f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 10504f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 1051339f8db1SStefano Zampini PetscScalar vals[2]; 1052339f8db1SStefano Zampini PetscInt cols[2]; 1053339f8db1SStefano Zampini 1054339f8db1SStefano Zampini cols[0] = idxs[i]; 10554f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 1056339f8db1SStefano Zampini vals[0] = 1.; 1057b0f5fe93SStefano Zampini vals[1] = 1.; 10584f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1059339f8db1SStefano Zampini } 10604f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 10614f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 10624f1b2e48SStefano Zampini array[nzs-1] = 1.; 10634f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 10644f1b2e48SStefano Zampini /* store local idxs for p0 */ 10654f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 10664f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 1067339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 10684f1b2e48SStefano Zampini } 1069339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1070339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1071a3df083aSStefano Zampini /* project if needed */ 1072a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 10731dd7afcfSStefano Zampini Mat M; 10741dd7afcfSStefano Zampini 10751dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 1076339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 10771dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 10781dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 1079a3df083aSStefano Zampini } 10804f1b2e48SStefano Zampini /* store global idxs for p0 */ 10814f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1082339f8db1SStefano Zampini } 1083ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 10844f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 1085b0f5fe93SStefano Zampini 1086b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 1087b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 108827b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 108927b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1090339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 1091339f8db1SStefano Zampini PetscFunctionReturn(0); 1092339f8db1SStefano Zampini } 1093339f8db1SStefano Zampini 1094339f8db1SStefano Zampini #undef __FUNCT__ 1095015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 1096015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 1097efc2fbd9SStefano Zampini { 1098efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1099de9d7bd0SStefano Zampini PetscScalar *array; 1100efc2fbd9SStefano Zampini PetscErrorCode ierr; 1101efc2fbd9SStefano Zampini 1102efc2fbd9SStefano Zampini PetscFunctionBegin; 1103efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 1104efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 11054f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 1106efc2fbd9SStefano Zampini } 1107de9d7bd0SStefano Zampini if (get) { 1108efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 11094f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 11104f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 1111efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 1112de9d7bd0SStefano Zampini } else { 1113de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 1114de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1115de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 1116de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 1117efc2fbd9SStefano Zampini } 1118efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1119efc2fbd9SStefano Zampini } 1120efc2fbd9SStefano Zampini 1121efc2fbd9SStefano Zampini #undef __FUNCT__ 1122c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 1123c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 1124c263805aSStefano Zampini { 1125c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1126c263805aSStefano Zampini PetscErrorCode ierr; 1127c263805aSStefano Zampini 1128c263805aSStefano Zampini PetscFunctionBegin; 1129c263805aSStefano Zampini /* TODO: add error checking 1130c263805aSStefano Zampini - avoid nested pop (or push) calls. 1131c263805aSStefano Zampini - cannot push before pop. 11321c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 1133c263805aSStefano Zampini */ 11344f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1135efc2fbd9SStefano Zampini PetscFunctionReturn(0); 1136efc2fbd9SStefano Zampini } 1137c263805aSStefano Zampini if (pop) { 1138a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 11394f1b2e48SStefano Zampini IS is_p0; 11404f1b2e48SStefano Zampini MatReuse reuse; 1141c263805aSStefano Zampini 1142c263805aSStefano Zampini /* extract B_0 */ 11434f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 11444f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 11454f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 11464f1b2e48SStefano Zampini } 11474f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 11484f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 1149c263805aSStefano Zampini /* remove rows and cols from local problem */ 1150c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 115197d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 11524f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 11534f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 1154a3df083aSStefano Zampini } else { 1155a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1156a3df083aSStefano Zampini PetscScalar *vals; 1157a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 1158a3df083aSStefano Zampini 1159a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 1160a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 1161a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 11620b5adadeSStefano Zampini PetscInt *nnz; 1163a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 1164a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 1165a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1166331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 1167331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1168331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 1169331e053bSStefano Zampini nnz[i] = n - nnz[i]; 1170331e053bSStefano Zampini } 1171331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 1172331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1173331e053bSStefano Zampini } 1174a3df083aSStefano Zampini 1175a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1176a3df083aSStefano Zampini PetscScalar *array; 1177a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 1178a3df083aSStefano Zampini 1179a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 1180a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1181a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1182a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 1183a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1184a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 1185a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 1186a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 1187a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 1188a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 1189a3df083aSStefano Zampini cum = 0; 1190a3df083aSStefano Zampini for (j=0;j<n;j++) { 119122db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 1192a3df083aSStefano Zampini vals[cum] = array[j]; 1193a3df083aSStefano Zampini idxs_ins[cum] = j; 1194a3df083aSStefano Zampini cum++; 1195a3df083aSStefano Zampini } 1196a3df083aSStefano Zampini } 1197a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 1198a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 1199a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 1200a3df083aSStefano Zampini } 1201a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1202a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1203a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 1204a3df083aSStefano Zampini } 1205c263805aSStefano Zampini } else { /* push */ 1206a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 12074f1b2e48SStefano Zampini PetscInt i; 12084f1b2e48SStefano Zampini 12094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 12104f1b2e48SStefano Zampini PetscScalar *B0_vals; 12114f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 12124f1b2e48SStefano Zampini 12134f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 12144f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 12157b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 12164f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 12174f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 12184f1b2e48SStefano Zampini } 1219c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1220c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1221a3df083aSStefano Zampini } else { 1222a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 1223a3df083aSStefano Zampini } 1224c263805aSStefano Zampini } 1225c263805aSStefano Zampini PetscFunctionReturn(0); 1226c263805aSStefano Zampini } 1227c263805aSStefano Zampini 1228c263805aSStefano Zampini #undef __FUNCT__ 1229b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 123008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 1231b1b3d7a2SStefano Zampini { 1232b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 123308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 123408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 123508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 123608122e43SStefano Zampini PetscScalar *work,lwork; 123708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 123808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 123908122e43SStefano Zampini PetscReal *eigs,thresh; 12401b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 1241f6f667cfSStefano Zampini PetscBool allocated_S_St; 124208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 124308122e43SStefano Zampini PetscReal *rwork; 124408122e43SStefano Zampini #endif 1245b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1246b1b3d7a2SStefano Zampini 1247b1b3d7a2SStefano Zampini PetscFunctionBegin; 1248b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 1249af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 1250af25d912SStefano 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); 125106a4e24aSStefano Zampini 1252fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1253fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1254fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1255fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 12561575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 1257fd14bc51SStefano Zampini } 1258fd14bc51SStefano Zampini 1259e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 1260e496cd5dSStefano 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); 1261e496cd5dSStefano Zampini } 1262e496cd5dSStefano Zampini 126308122e43SStefano Zampini /* max size of subsets */ 126408122e43SStefano Zampini mss = 0; 126508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 126608122e43SStefano Zampini PetscInt subset_size; 1267862806e4SStefano Zampini 126808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 126908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 127008122e43SStefano Zampini } 127108122e43SStefano Zampini 127208122e43SStefano Zampini /* min/max and threshold */ 127308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 1274f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 127508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 1276f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 1277f6f667cfSStefano Zampini if (nmin) { 1278f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 1279f6f667cfSStefano Zampini } 128008122e43SStefano Zampini 128108122e43SStefano Zampini /* allocate lapack workspace */ 128208122e43SStefano Zampini cum = cum2 = 0; 128308122e43SStefano Zampini maxneigs = 0; 128408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 128508122e43SStefano Zampini PetscInt n,subset_size; 1286f6f667cfSStefano Zampini 128708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 128808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 12899162d606SStefano Zampini cum += subset_size; 12909162d606SStefano Zampini cum2 += subset_size*n; 129108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 129208122e43SStefano Zampini } 129308122e43SStefano Zampini if (mss) { 12949ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 129508122e43SStefano Zampini PetscBLASInt B_itype = 1; 129608122e43SStefano Zampini PetscBLASInt B_N = mss; 12974c6709b3SStefano Zampini PetscReal zero = 0.0; 12984c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 129908122e43SStefano Zampini 130008122e43SStefano Zampini B_lwork = -1; 130108122e43SStefano Zampini S = NULL; 130208122e43SStefano Zampini St = NULL; 1303a58a30b4SStefano Zampini eigs = NULL; 1304a58a30b4SStefano Zampini eigv = NULL; 1305a58a30b4SStefano Zampini B_iwork = NULL; 1306a58a30b4SStefano Zampini B_ifail = NULL; 1307d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1308d1710679SStefano Zampini rwork = NULL; 1309d1710679SStefano Zampini #endif 13108bec7fa6SStefano Zampini thresh = 1.0; 131108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 131208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 131308122e43SStefano 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)); 131408122e43SStefano Zampini #else 131508122e43SStefano 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)); 131608122e43SStefano Zampini #endif 131708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 131808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 131908122e43SStefano Zampini } else { 132008122e43SStefano Zampini /* TODO */ 132108122e43SStefano Zampini } 132208122e43SStefano Zampini } else { 132308122e43SStefano Zampini lwork = 0; 132408122e43SStefano Zampini } 132508122e43SStefano Zampini 132608122e43SStefano Zampini nv = 0; 1327d62866d3SStefano 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) */ 1328d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 132908122e43SStefano Zampini } 13304c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1331f6f667cfSStefano Zampini if (allocated_S_St) { 1332f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1333f6f667cfSStefano Zampini } 1334f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 133508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 133608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 133708122e43SStefano Zampini #endif 13389162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 13399162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 13409162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 134108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 13429162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 134308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 134408122e43SStefano Zampini 134508122e43SStefano Zampini maxneigs = 0; 134672b8c272SStefano Zampini cum = cumarray = 0; 13479162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 13489162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1349d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 135008122e43SStefano Zampini const PetscInt *idxs; 135108122e43SStefano Zampini 1352d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 135308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 135408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 135508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 135608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 13579162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 13589162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 135908122e43SStefano Zampini } 1360d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 136108122e43SStefano Zampini } 136208122e43SStefano Zampini 136308122e43SStefano Zampini if (mss) { /* multilevel */ 136408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 136508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 136608122e43SStefano Zampini } 136708122e43SStefano Zampini 1368ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 136908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 137008122e43SStefano Zampini const PetscInt *idxs; 13719d54b7f4SStefano Zampini PetscReal upper,lower; 1372862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 137308122e43SStefano Zampini PetscBLASInt B_N; 1374aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 137508122e43SStefano Zampini 13769d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 13779d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 13789d54b7f4SStefano Zampini lower = thresh; 13799d54b7f4SStefano Zampini } else { 13809d54b7f4SStefano Zampini upper = 1./thresh; 13819d54b7f4SStefano Zampini lower = 0.; 13829d54b7f4SStefano Zampini } 1383862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1384ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1385f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1386f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 13879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1388aff50787SStefano Zampini PetscInt j,k; 1389aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1390aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1391aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 139208122e43SStefano Zampini } 139308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1394aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1395aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1396aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1397aff50787SStefano Zampini } 139808122e43SStefano Zampini } 139908122e43SStefano Zampini } else { 140008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 140108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 140208122e43SStefano Zampini } 14038bec7fa6SStefano Zampini } else { 1404f6f667cfSStefano Zampini S = Sarray + cumarray; 1405f6f667cfSStefano Zampini St = Starray + cumarray; 14068bec7fa6SStefano Zampini } 1407aff50787SStefano Zampini /* see if we can save some work */ 1408b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1409aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1410aff50787SStefano Zampini } 1411aff50787SStefano Zampini 1412b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1413aff50787SStefano Zampini B_neigs = 0; 1414aff50787SStefano Zampini } else { 14159ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 141608122e43SStefano Zampini PetscBLASInt B_itype = 1; 1417f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 14184c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 14199552c7c7SStefano Zampini PetscInt nmin_s; 1420b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 142108122e43SStefano Zampini 1422fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 14238bec7fa6SStefano 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]]); 1424fd14bc51SStefano Zampini } 1425d16cbb6bSStefano Zampini 1426b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1427b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1428b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1429b7ab4a40SStefano Zampini } 1430b7ab4a40SStefano Zampini 143108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1432b7ab4a40SStefano Zampini if (compute_range) { 1433d16cbb6bSStefano Zampini 1434d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 143508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 14369d54b7f4SStefano 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)); 143708122e43SStefano Zampini #else 14389d54b7f4SStefano 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)); 143908122e43SStefano Zampini #endif 1440b7ab4a40SStefano Zampini } else if (!same_data) { 1441d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1442d16cbb6bSStefano Zampini B_IL = 1; 1443d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 14449d54b7f4SStefano 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)); 1445d16cbb6bSStefano Zampini #else 14469d54b7f4SStefano 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)); 1447d16cbb6bSStefano Zampini #endif 1448b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1449b7ab4a40SStefano Zampini PetscInt k; 1450b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1451b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1452b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1453b7ab4a40SStefano Zampini nmin = nmax; 1454b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1455b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1456b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1457b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1458b7ab4a40SStefano Zampini } 1459d16cbb6bSStefano Zampini } 146008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 146108122e43SStefano Zampini if (B_ierr) { 14626c4ed002SBarry 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); 14636c4ed002SBarry 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); 14646c4ed002SBarry 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); 146508122e43SStefano Zampini } 146608122e43SStefano Zampini 146708122e43SStefano Zampini if (B_neigs > nmax) { 1468fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1469fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1470fd14bc51SStefano Zampini } 14719d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 147208122e43SStefano Zampini B_neigs = nmax; 147308122e43SStefano Zampini } 147408122e43SStefano Zampini 14759552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 14769552c7c7SStefano Zampini if (B_neigs < nmin_s) { 147708122e43SStefano Zampini PetscBLASInt B_neigs2; 147808122e43SStefano Zampini 14799d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1480f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 14819d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 14829d54b7f4SStefano Zampini } else { 14839d54b7f4SStefano Zampini B_IL = B_neigs + 1; 14849d54b7f4SStefano Zampini B_IU = nmin_s; 14859d54b7f4SStefano Zampini } 1486fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1487fd14bc51SStefano 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); 1488fd14bc51SStefano Zampini } 14899ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 14901ae86dd6SStefano Zampini PetscInt j,k; 149108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 14921ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 14931ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 14941ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 149508122e43SStefano Zampini } 149608122e43SStefano Zampini } 149708122e43SStefano Zampini } else { 149808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 149908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 150008122e43SStefano Zampini } 150108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 150208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 15039d54b7f4SStefano 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)); 150408122e43SStefano Zampini #else 15059d54b7f4SStefano 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)); 150608122e43SStefano Zampini #endif 150708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 150808122e43SStefano Zampini B_neigs += B_neigs2; 150908122e43SStefano Zampini } 151008122e43SStefano Zampini if (B_ierr) { 15116c4ed002SBarry 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); 15126c4ed002SBarry 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); 15136c4ed002SBarry 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); 151408122e43SStefano Zampini } 1515fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1516ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 151708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 151808122e43SStefano Zampini if (eigs[j] == 0.0) { 1519ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 152008122e43SStefano Zampini } else { 15219d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1522ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 15239d54b7f4SStefano Zampini } else { 15249d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 15259d54b7f4SStefano Zampini } 1526fd14bc51SStefano Zampini } 152708122e43SStefano Zampini } 152808122e43SStefano Zampini } 152908122e43SStefano Zampini } else { 153008122e43SStefano Zampini /* TODO */ 153108122e43SStefano Zampini } 1532aff50787SStefano Zampini } 15336c3e6151SStefano Zampini /* change the basis back to the original one */ 15346c3e6151SStefano Zampini if (sub_schurs->change) { 153572b8c272SStefano Zampini Mat change,phi,phit; 15366c3e6151SStefano Zampini 15376c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 15386c3e6151SStefano Zampini PetscInt ii; 15396c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 15406c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15416c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 15426c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 15436c3e6151SStefano Zampini } 15446c3e6151SStefano Zampini } 15456c3e6151SStefano Zampini } 154672b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 15476c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 154872b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 15496c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 15506c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 15516c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 15526c3e6151SStefano Zampini } 15538bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 15548bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 15559162d606SStefano Zampini if (B_neigs) { 15569162d606SStefano 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); 1557fd14bc51SStefano Zampini 1558fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 15599552c7c7SStefano Zampini PetscInt ii; 15609552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1561ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15629552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1563ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1564ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1565ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1566ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1567ac47001eSStefano Zampini #else 1568ac47001eSStefano 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); 1569ac47001eSStefano Zampini #endif 15709552c7c7SStefano Zampini } 15719552c7c7SStefano Zampini } 1572fd14bc51SStefano Zampini } 15739162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 15749162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 15759162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 15769162d606SStefano Zampini cum++; 157708122e43SStefano Zampini } 157808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 157908122e43SStefano Zampini /* shift for next computation */ 158008122e43SStefano Zampini cumarray += subset_size*subset_size; 158108122e43SStefano Zampini } 1582fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1583fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1584fd14bc51SStefano Zampini } 158508122e43SStefano Zampini 158608122e43SStefano Zampini if (mss) { 158708122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 158808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1589f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1590f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1591f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 159208122e43SStefano Zampini } 1593f6f667cfSStefano Zampini if (allocated_S_St) { 1594f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1595f6f667cfSStefano Zampini } 1596f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 159708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 159808122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 159908122e43SStefano Zampini #endif 160008122e43SStefano Zampini if (pcbddc->dbg_flag) { 16011b968477SStefano Zampini PetscInt maxneigs_r; 1602b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 16039b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 160408122e43SStefano Zampini } 160508122e43SStefano Zampini PetscFunctionReturn(0); 160608122e43SStefano Zampini } 1607b1b3d7a2SStefano Zampini 1608674ae819SStefano Zampini #undef __FUNCT__ 1609c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1610c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1611c8587f34SStefano Zampini { 16128629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1613c8587f34SStefano Zampini PetscErrorCode ierr; 1614c8587f34SStefano Zampini 1615c8587f34SStefano Zampini PetscFunctionBegin; 1616f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 16175e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1618c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1619c8587f34SStefano Zampini 1620684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 16210fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1622684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1623c8587f34SStefano Zampini 16248629588bSStefano Zampini /* 16258629588bSStefano Zampini Setup local correction and local part of coarse basis. 16268629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 16278629588bSStefano Zampini */ 162847f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 16298629588bSStefano Zampini 16308629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 16318629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 16328629588bSStefano Zampini 16338629588bSStefano Zampini /* free */ 16348629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1635c8587f34SStefano Zampini PetscFunctionReturn(0); 1636c8587f34SStefano Zampini } 1637c8587f34SStefano Zampini 1638c8587f34SStefano Zampini #undef __FUNCT__ 1639674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1640674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1641674ae819SStefano Zampini { 1642674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1643674ae819SStefano Zampini PetscErrorCode ierr; 1644674ae819SStefano Zampini 1645674ae819SStefano Zampini PetscFunctionBegin; 1646674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1647674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 164830368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1649674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1650785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1651674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1652f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1653f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1654785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 165563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 165663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1657674ae819SStefano Zampini PetscFunctionReturn(0); 1658674ae819SStefano Zampini } 1659674ae819SStefano Zampini 1660674ae819SStefano Zampini #undef __FUNCT__ 1661674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1662674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1663674ae819SStefano Zampini { 1664674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16654f1b2e48SStefano Zampini PetscInt i; 1666674ae819SStefano Zampini PetscErrorCode ierr; 1667674ae819SStefano Zampini 1668674ae819SStefano Zampini PetscFunctionBegin; 1669b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1670674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 167116909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 16721dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1673674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1674669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 1675fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 1676674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 1677*71582508SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 16784f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 16794f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 16804f1b2e48SStefano Zampini } 16814f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1682b334f244SStefano Zampini if (pcbddc->sub_schurs) { 1683b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1684b334f244SStefano Zampini } 1685674ae819SStefano Zampini PetscFunctionReturn(0); 1686674ae819SStefano Zampini } 1687674ae819SStefano Zampini 1688674ae819SStefano Zampini #undef __FUNCT__ 1689674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1690674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1691674ae819SStefano Zampini { 1692674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1693674ae819SStefano Zampini PetscErrorCode ierr; 1694674ae819SStefano Zampini 1695674ae819SStefano Zampini PetscFunctionBegin; 1696674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 169758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1698ca92afb2SStefano Zampini PetscScalar *array; 169906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 170006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 170158da7f69SStefano Zampini } 1702674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1703674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 170415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 170515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1706674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1707674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1708674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 170906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1710674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1711674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17128ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1713674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1714674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1715674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1716f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1717f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1718f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1719f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1720727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 17210e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1722f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 172370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 172481d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 17250369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 17261dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 17274f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 17288b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1729ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1730ca92afb2SStefano Zampini PetscInt i; 1731ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1732ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1733ca92afb2SStefano Zampini } 1734ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1735ca92afb2SStefano Zampini } 17364f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1737674ae819SStefano Zampini PetscFunctionReturn(0); 1738674ae819SStefano Zampini } 1739674ae819SStefano Zampini 1740674ae819SStefano Zampini #undef __FUNCT__ 1741f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1742f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 17436bfb1811SStefano Zampini { 17446bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 17456bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 17466bfb1811SStefano Zampini VecType impVecType; 17474f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 17486bfb1811SStefano Zampini PetscErrorCode ierr; 17496bfb1811SStefano Zampini 17506bfb1811SStefano Zampini PetscFunctionBegin; 17516c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1752e7b262bdSStefano Zampini /* get sizes */ 17534f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1754b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 17556bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1756e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1757e7b262bdSStefano Zampini /* R nodes */ 1758e7b262bdSStefano Zampini old_size = -1; 1759e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1760e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1761e7b262bdSStefano Zampini } 1762e7b262bdSStefano Zampini if (n_R != old_size) { 1763e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1764e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17656bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 17666bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 17676bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 17686bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1769e7b262bdSStefano Zampini } 1770e7b262bdSStefano Zampini /* local primal dofs */ 1771e7b262bdSStefano Zampini old_size = -1; 1772e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1773e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1774e7b262bdSStefano Zampini } 1775e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1776e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 177783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1778e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 17796bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1780e7b262bdSStefano Zampini } 1781e7b262bdSStefano Zampini /* local explicit constraints */ 1782e7b262bdSStefano Zampini old_size = -1; 1783e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1784e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1785e7b262bdSStefano Zampini } 1786e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1787e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 178883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 178983b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 179083b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 179183b7ccabSStefano Zampini } 17926bfb1811SStefano Zampini PetscFunctionReturn(0); 17936bfb1811SStefano Zampini } 17946bfb1811SStefano Zampini 17956bfb1811SStefano Zampini #undef __FUNCT__ 179647f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 179747f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 179888ebb749SStefano Zampini { 179925084f0cSStefano Zampini PetscErrorCode ierr; 180025084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 180188ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 180288ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1803d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 180425084f0cSStefano Zampini /* submatrices of local problem */ 180580677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 180606656605SStefano Zampini /* submatrices of local coarse problem */ 180706656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 180825084f0cSStefano Zampini /* working matrices */ 180906656605SStefano Zampini Mat C_CR; 181025084f0cSStefano Zampini /* additional working stuff */ 181106656605SStefano Zampini PC pc_R; 18124f1b2e48SStefano Zampini Mat F; 18135cbda25cSStefano Zampini Vec dummy_vec; 1814a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 181525084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 181606656605SStefano Zampini PetscScalar *work; 181706656605SStefano Zampini PetscInt *idx_V_B; 1818ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 181906656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1820ffd830a3SStefano Zampini 182125084f0cSStefano Zampini /* some shortcuts to scalars */ 182206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 182388ebb749SStefano Zampini 182488ebb749SStefano Zampini PetscFunctionBegin; 18259a962809SStefano 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"); 1826ffd830a3SStefano Zampini 1827ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1828b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 18294f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1830b371cd4fSStefano Zampini n_B = pcis->n_B; 1831b371cd4fSStefano Zampini n_D = pcis->n - n_B; 183288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 183388ebb749SStefano Zampini 183488ebb749SStefano Zampini /* vertices in boundary numbering */ 1835785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 18360e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 18376c4ed002SBarry 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); 183888ebb749SStefano Zampini 183906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1840019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 184106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 184206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 184306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 184406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 184506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 184606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 184706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 184806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 184906656605SStefano Zampini 185006656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 185106656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 185206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 185306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 185406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1855ffd830a3SStefano Zampini lda_rhs = n_R; 1856a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 185706656605SStefano Zampini if (isLU || isILU || isCHOL) { 185806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1859b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 1860df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1861d62866d3SStefano Zampini MatFactorType type; 1862d62866d3SStefano Zampini 1863df4d28bfSStefano Zampini F = reuse_solver->F; 18646816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1865d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1866ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 186722db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 186806656605SStefano Zampini } else { 186906656605SStefano Zampini F = NULL; 187006656605SStefano Zampini } 187106656605SStefano Zampini 1872ffd830a3SStefano Zampini /* allocate workspace */ 1873ffd830a3SStefano Zampini n = 0; 1874ffd830a3SStefano Zampini if (n_constraints) { 1875ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1876ffd830a3SStefano Zampini } 1877ffd830a3SStefano Zampini if (n_vertices) { 1878ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1879ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1880ffd830a3SStefano Zampini } 1881ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1882ffd830a3SStefano Zampini 18835cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 18845cbda25cSStefano Zampini dummy_vec = NULL; 18855cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 18865cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 18875cbda25cSStefano Zampini } 18885cbda25cSStefano Zampini 188988ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 189088ebb749SStefano Zampini if (n_constraints) { 189172b8c272SStefano Zampini Mat M1,M2,M3,C_B; 189206656605SStefano Zampini IS is_aux; 189380677318SStefano Zampini PetscScalar *array,*array2; 189406656605SStefano Zampini 1895f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 189680677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 189788ebb749SStefano Zampini 189825084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 189925084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 19008ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 190172b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 190288ebb749SStefano Zampini 190380677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 190480677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1905ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 190688ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 190706656605SStefano Zampini const PetscScalar *row_cmat_values; 190806656605SStefano Zampini const PetscInt *row_cmat_indices; 190906656605SStefano Zampini PetscInt size_of_constraint,j; 191088ebb749SStefano Zampini 191106656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 191206656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1913ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 191406656605SStefano Zampini } 191506656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 191606656605SStefano Zampini } 1917ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 191806656605SStefano Zampini if (F) { 191906656605SStefano Zampini Mat B; 192006656605SStefano Zampini 1921ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1922a3df083aSStefano Zampini if (need_benign_correction) { 1923df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1924a3df083aSStefano Zampini 192572b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 192672b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1927a3df083aSStefano Zampini } 192880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1929a3df083aSStefano Zampini if (need_benign_correction) { 1930a3df083aSStefano Zampini PetscScalar *marr; 1931df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1932a3df083aSStefano Zampini 1933a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 19345cbda25cSStefano Zampini if (lda_rhs != n_R) { 19355cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 19365cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19375cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19385cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19395cbda25cSStefano Zampini } 19405cbda25cSStefano Zampini } else { 1941a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1942a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19435cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1944a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1945a3df083aSStefano Zampini } 19465cbda25cSStefano Zampini } 1947a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1948a3df083aSStefano Zampini } 194906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195006656605SStefano Zampini } else { 195180677318SStefano Zampini PetscScalar *marr; 195280677318SStefano Zampini 195380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 195406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1955ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1956ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 195706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 195806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 195906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 196006656605SStefano Zampini } 196180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 196206656605SStefano Zampini } 196380677318SStefano Zampini if (!pcbddc->switch_static) { 196480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 196580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 196680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 196780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1968ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 196980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 197080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 197380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 197480677318SStefano Zampini } 197580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 197680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 197772b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 197880677318SStefano Zampini } else { 1979ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1980ffd830a3SStefano Zampini IS dummy; 1981ffd830a3SStefano Zampini 1982ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 198372b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1984ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1985ffd830a3SStefano Zampini } else { 198680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 198780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1988ffd830a3SStefano Zampini } 198925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 199080677318SStefano Zampini } 199180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 199280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 199380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 199406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 199506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 199680677318SStefano Zampini if (isCHOL) { 199780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 199880677318SStefano Zampini } else { 199925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 200080677318SStefano Zampini } 200180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 200206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 200325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 200425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 200525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 200680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 200772b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 200872b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 200906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 201006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 2011f4ddd8eeSStefano Zampini } 2012fc227af8SStefano Zampini 2013fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 201488ebb749SStefano Zampini if (n_vertices) { 201506656605SStefano Zampini IS is_aux; 20163a50541eSStefano Zampini 2017b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 20186816873aSStefano Zampini IS tis; 20196816873aSStefano Zampini 20206816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 20216816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 20226816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 20236816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 20246816873aSStefano Zampini } else { 20253a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 20266816873aSStefano Zampini } 20279577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 20289577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 202904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 203025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 203188ebb749SStefano Zampini } 203288ebb749SStefano Zampini 203388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 2034f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 203506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 203606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 203706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 203806656605SStefano Zampini } 2039f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 204006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 204106656605SStefano Zampini PetscScalar *marray; 204206656605SStefano Zampini 204306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 204406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 2045f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2046f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 2047f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 2048f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2049f4ddd8eeSStefano Zampini } 2050f4ddd8eeSStefano Zampini } 205106656605SStefano Zampini 2052f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 205306656605SStefano Zampini PetscScalar *marray; 205488ebb749SStefano Zampini 205506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20568eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 205706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 205888ebb749SStefano Zampini } 20593301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 206006656605SStefano Zampini n *= 2; 206188ebb749SStefano Zampini } 206206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 206306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 206406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20658eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 206606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 206706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 206888ebb749SStefano Zampini } 20693301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 207006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 20718eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 207206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 207306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 207488ebb749SStefano Zampini } 207588ebb749SStefano Zampini } else { 2076c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 2077c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 20781b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2079c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 2080c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 2081c0553b1fSStefano Zampini } 208288ebb749SStefano Zampini } 208306656605SStefano Zampini } 2084019a44ceSStefano Zampini 208506656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 20864f1b2e48SStefano Zampini p0_lidx_I = NULL; 20874f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 2088d12edf2fSStefano Zampini const PetscInt *idxs; 2089d12edf2fSStefano Zampini 2090d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 20914f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 20924f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20934f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 20944f1b2e48SStefano Zampini } 2095d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 2096d12edf2fSStefano Zampini } 2097d16cbb6bSStefano Zampini 209806656605SStefano Zampini /* vertices */ 209906656605SStefano Zampini if (n_vertices) { 210016f15bc4SStefano Zampini 2101af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 210204708bb6SStefano Zampini 210316f15bc4SStefano Zampini if (n_R) { 210414393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 210506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 210616f15bc4SStefano Zampini PetscScalar *x,*y; 210704708bb6SStefano Zampini PetscBool isseqaij; 210806656605SStefano Zampini 210921eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 211014393ed6SStefano Zampini if (need_benign_correction) { 211114393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 211214393ed6SStefano Zampini IS is_p0; 211314393ed6SStefano Zampini PetscInt *idxs_p0,n; 211414393ed6SStefano Zampini 211514393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 211614393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 211714393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2118af25d912SStefano 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); 211914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 212014393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 212114393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 212214393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 212314393ed6SStefano Zampini } 212414393ed6SStefano Zampini 2125ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2126af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2127ffd830a3SStefano Zampini } else { 2128ca92afb2SStefano Zampini PetscScalar *av,*array; 2129ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2130ca92afb2SStefano Zampini PetscInt n; 2131ca92afb2SStefano Zampini PetscBool flg_row; 2132ffd830a3SStefano Zampini 2133ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2134ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 21359d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2136ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2137ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2138ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2139ca92afb2SStefano Zampini PetscInt j; 2140ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2141ffd830a3SStefano Zampini } 2142ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2143ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2144ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2145ffd830a3SStefano Zampini } 2146ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2147a3df083aSStefano Zampini if (need_benign_correction) { 2148df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2149a3df083aSStefano Zampini PetscScalar *marr; 2150a3df083aSStefano Zampini 2151a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 215214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 215314393ed6SStefano Zampini 215414393ed6SStefano Zampini | 0 0 0 | (V) 215514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 215614393ed6SStefano Zampini | 0 0 -1 | (p0) 215714393ed6SStefano Zampini 215814393ed6SStefano Zampini */ 2159df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 216014393ed6SStefano Zampini const PetscScalar *vals; 216114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 216214393ed6SStefano Zampini PetscInt n,j,nz; 216314393ed6SStefano Zampini 2164df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2165df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 216614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 216714393ed6SStefano Zampini for (j=0;j<n;j++) { 216814393ed6SStefano Zampini PetscScalar val = vals[j]; 216914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 217014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 217114393ed6SStefano Zampini } 217214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2173df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 217414393ed6SStefano Zampini } 217572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 217672b8c272SStefano Zampini } 217772b8c272SStefano Zampini if (F) { 217814393ed6SStefano Zampini /* need to correct the rhs */ 217972b8c272SStefano Zampini if (need_benign_correction) { 218072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 218172b8c272SStefano Zampini PetscScalar *marr; 218272b8c272SStefano Zampini 218372b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 21845cbda25cSStefano Zampini if (lda_rhs != n_R) { 21855cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21865cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21875cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 21885cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21895cbda25cSStefano Zampini } 21905cbda25cSStefano Zampini } else { 2191a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2192a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21935cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2194a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2195a3df083aSStefano Zampini } 21965cbda25cSStefano Zampini } 2197a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2198a3df083aSStefano Zampini } 219906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 220014393ed6SStefano Zampini /* need to correct the solution */ 2201a3df083aSStefano Zampini if (need_benign_correction) { 2202df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2203a3df083aSStefano Zampini PetscScalar *marr; 2204a3df083aSStefano Zampini 2205a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 22065cbda25cSStefano Zampini if (lda_rhs != n_R) { 22075cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 22085cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 22095cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 22105cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 22115cbda25cSStefano Zampini } 22125cbda25cSStefano Zampini } else { 2213a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2214a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 22155cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2216a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2217a3df083aSStefano Zampini } 22185cbda25cSStefano Zampini } 2219a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2220a3df083aSStefano Zampini } 222106656605SStefano Zampini } else { 222206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 222306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2224ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2225ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 222606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 222706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 222806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 222906656605SStefano Zampini } 223006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 223106656605SStefano Zampini } 223280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2233ffd830a3SStefano Zampini /* S_VV and S_CV */ 223406656605SStefano Zampini if (n_constraints) { 223506656605SStefano Zampini Mat B; 223680677318SStefano Zampini 2237ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 223880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2239ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2240ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 224180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 224280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 224380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 224480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 224580677318SStefano Zampini } 2246ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 224780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 224880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2249ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 225080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 225106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2252ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2253ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 225406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 225506656605SStefano Zampini } 225604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 225704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2258511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 225904708bb6SStefano Zampini } 2260ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2261ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2262ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2263ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2264ffd830a3SStefano Zampini } 226506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 226614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 226714393ed6SStefano Zampini if (need_benign_correction) { 2268df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 226914393ed6SStefano Zampini PetscScalar *marr,*sums; 227014393ed6SStefano Zampini 227114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 227214393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2273df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 227414393ed6SStefano Zampini const PetscScalar *vals; 227514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 227614393ed6SStefano Zampini PetscInt n,j,nz; 227714393ed6SStefano Zampini 2278df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2279df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 228014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 228114393ed6SStefano Zampini PetscInt k; 228214393ed6SStefano Zampini sums[j] = 0.; 228314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 228414393ed6SStefano Zampini } 228514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 228614393ed6SStefano Zampini for (j=0;j<n;j++) { 228714393ed6SStefano Zampini PetscScalar val = vals[j]; 228814393ed6SStefano Zampini PetscInt k; 228914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 229014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 229114393ed6SStefano Zampini } 229214393ed6SStefano Zampini } 229314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2294df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 229514393ed6SStefano Zampini } 229614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 229714393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 229814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 229914393ed6SStefano Zampini } 230080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 230106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 230206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 230306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 230406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 230506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 230606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 230706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2308d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2309019a44ceSStefano Zampini } else { 2310d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2311d16cbb6bSStefano Zampini } 231221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2313d16cbb6bSStefano Zampini 231406656605SStefano Zampini /* coarse basis functions */ 231506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 231616f15bc4SStefano Zampini PetscScalar *y; 231716f15bc4SStefano Zampini 2318ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 231906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 232006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 232106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 232406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 232506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 232606656605SStefano Zampini 232706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23284f1b2e48SStefano Zampini PetscInt j; 23294f1b2e48SStefano Zampini 233006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 233106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 233206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23354f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 233606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 233706656605SStefano Zampini } 233806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 233906656605SStefano Zampini } 234004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 234104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 234206656605SStefano Zampini } 23435cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 234406656605SStefano Zampini 234506656605SStefano Zampini if (n_constraints) { 234606656605SStefano Zampini Mat B; 234706656605SStefano Zampini 2348ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 234906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 235080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 235106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 235206656605SStefano Zampini if (n_vertices) { 235380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 235480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 235580677318SStefano Zampini } else { 235680677318SStefano Zampini Mat S_VCt; 235780677318SStefano Zampini 2358ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2359ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 236072b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2361ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2362ffd830a3SStefano Zampini } 236380677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 236480677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 236580677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 236680677318SStefano Zampini } 236706656605SStefano Zampini } 236806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 236906656605SStefano Zampini /* coarse basis functions */ 237006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 237106656605SStefano Zampini PetscScalar *y; 237206656605SStefano Zampini 2373ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 237406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 237506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 237606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 237706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 237806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 237906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 238006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23814f1b2e48SStefano Zampini PetscInt j; 23824f1b2e48SStefano Zampini 238306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 238406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 238506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23884f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 238906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 239006656605SStefano Zampini } 239106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 239206656605SStefano Zampini } 239306656605SStefano Zampini } 239480677318SStefano Zampini if (n_constraints) { 239580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 239680677318SStefano Zampini } 23974f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 239872b8c272SStefano Zampini 239972b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 240072b8c272SStefano Zampini if (pcbddc->benign_n) { 240172b8c272SStefano Zampini Mat B0_B,B0_BPHI; 240272b8c272SStefano Zampini IS is_dummy; 240372b8c272SStefano Zampini PetscScalar *data; 240472b8c272SStefano Zampini PetscInt j; 240572b8c272SStefano Zampini 240672b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 240772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 240872b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 240972b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 241086c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 241172b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 241272b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 241372b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 241472b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 241572b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 241672b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 241772b8c272SStefano Zampini } 241872b8c272SStefano Zampini } 241972b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 242072b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 242172b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 242272b8c272SStefano Zampini } 2423019a44ceSStefano Zampini 242406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 24253301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2426ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2427ffd830a3SStefano Zampini PetscScalar *marray; 242806656605SStefano Zampini 242906656605SStefano Zampini if (n_constraints) { 2430ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 243106656605SStefano Zampini 2432af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 243306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2434ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 243516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 243606656605SStefano Zampini if (n_vertices) { 2437ffd830a3SStefano Zampini Mat S_VCT; 243806656605SStefano Zampini 243906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2440ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 244116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 244206656605SStefano Zampini } 2443ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 24445b782168SStefano Zampini } else { 24455b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 244606656605SStefano Zampini } 244716f15bc4SStefano Zampini if (n_vertices && n_R) { 2448ffd830a3SStefano Zampini PetscScalar *av,*marray; 2449ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2450ffd830a3SStefano Zampini PetscInt n; 2451ffd830a3SStefano Zampini PetscBool flg_row; 245206656605SStefano Zampini 2453ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2454af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2455ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2456ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2457ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2458ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2459ffd830a3SStefano Zampini PetscInt j; 2460ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2461ffd830a3SStefano Zampini } 2462ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2463ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2464ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 246506656605SStefano Zampini } 246606656605SStefano Zampini 2467ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2468ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2469ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2470ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2471ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 247206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 247306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 247406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 247506656605SStefano Zampini } 2476ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 24775b782168SStefano Zampini if (B_C) { 2478ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2479ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2480ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2481ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2482ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2483ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2484ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 248506656605SStefano Zampini } 2486ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 24875b782168SStefano Zampini } 248806656605SStefano Zampini /* coarse basis functions */ 248906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 249006656605SStefano Zampini PetscScalar *y; 249106656605SStefano Zampini 2492ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 249306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 249406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 249506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 249606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 249706656605SStefano Zampini if (i<n_vertices) { 249806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 249906656605SStefano Zampini } 250006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 250106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 250206656605SStefano Zampini 250306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 250406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 250506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 250606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 250906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 251006656605SStefano Zampini } 251106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 251206656605SStefano Zampini } 2513ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2514ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 251506656605SStefano Zampini } 2516d62866d3SStefano Zampini /* free memory */ 251788ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 251806656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 251906656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 252006656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 252106656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2522d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2523d62866d3SStefano Zampini if (n_vertices) { 2524d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2525d62866d3SStefano Zampini } 2526d62866d3SStefano Zampini if (n_constraints) { 2527d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2528d62866d3SStefano Zampini } 252988ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 253088ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 253188ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2532d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 253388ebb749SStefano Zampini Mat coarse_sub_mat; 253425084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 253588ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 253688ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 253788ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 25388bec7fa6SStefano Zampini Mat C_B,CPHI; 25398bec7fa6SStefano Zampini IS is_dummy; 25408bec7fa6SStefano Zampini Vec mones; 254188ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 254288ebb749SStefano Zampini PetscReal real_value; 254388ebb749SStefano Zampini 2544a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2545a3df083aSStefano Zampini Mat A; 2546a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2547a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2548a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2549a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2550a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2551a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2552a3df083aSStefano Zampini } else { 255388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 255488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 255588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 255688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2557a3df083aSStefano Zampini } 255888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 255988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2560ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 256188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 256288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 256388ebb749SStefano Zampini } 256488ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 256588ebb749SStefano Zampini 256625084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 25673301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 256825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2569ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 257088ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 257188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 257288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 257388ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 257488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 257588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 257688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 257788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 257888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 257988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 258188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258288ebb749SStefano Zampini } else { 258388ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 258488ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 258588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 258788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 259088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 259188ebb749SStefano Zampini } 259288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 259388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 259488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2595511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 25964f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2597fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2598d12edf2fSStefano Zampini PetscScalar *data,*data2; 25994f1b2e48SStefano Zampini PetscInt j; 2600d12edf2fSStefano Zampini 26014f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2602fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2603d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 260486c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2605d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2606d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 26074f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 26084f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2609d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 26104f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 26114f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 26124f1b2e48SStefano Zampini } 2613d12edf2fSStefano Zampini } 2614d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2615d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2616d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2617d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2618d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2619d12edf2fSStefano Zampini } 2620d12edf2fSStefano Zampini #if 0 2621d12edf2fSStefano Zampini { 2622d12edf2fSStefano Zampini PetscViewer viewer; 2623d12edf2fSStefano Zampini char filename[256]; 2624ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2625d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2626d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2627ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2628ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2629ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2630d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 263172b8c272SStefano Zampini if (save_change) { 263272b8c272SStefano Zampini Mat phi_B; 263372b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 263472b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 263572b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 263672b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 263772b8c272SStefano Zampini } else { 2638ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2639ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 264072b8c272SStefano Zampini } 2641ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2642ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2643ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2644ffd830a3SStefano Zampini } 2645ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2646ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2647ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2648ffd830a3SStefano Zampini } 264972b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2650ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2651ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2652ffd830a3SStefano Zampini } 2653d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2654d12edf2fSStefano Zampini } 2655d12edf2fSStefano Zampini #endif 265681d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 26578bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 26581575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 265906656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 26608bec7fa6SStefano Zampini 26618bec7fa6SStefano Zampini /* check constraints */ 2662a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2663a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 26644f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 26658bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2666a00504b5SStefano Zampini } else { 2667a00504b5SStefano Zampini PetscScalar *data; 2668a00504b5SStefano Zampini Mat tmat; 2669a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2670a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2671a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2672a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2673a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2674a00504b5SStefano Zampini } 26758bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 26768bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 26778bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 26788bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2679bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2680ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2681bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2682bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2683bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2684bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2685bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 268688ebb749SStefano Zampini } 26878bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 26888bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 26898bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 26908bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 269125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 269288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 269388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 269488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 269588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 269688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 269788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 269888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 269988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 270088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 270188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2702ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 270388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 270488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 270588ebb749SStefano Zampini } 270688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 270788ebb749SStefano Zampini } 27088629588bSStefano Zampini /* get back data */ 27098629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 271088ebb749SStefano Zampini PetscFunctionReturn(0); 271188ebb749SStefano Zampini } 271288ebb749SStefano Zampini 271388ebb749SStefano Zampini #undef __FUNCT__ 2714d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2715d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2716aa0d41d4SStefano Zampini { 2717d65f70fdSStefano Zampini Mat *work_mat; 2718d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2719d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2720c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2721aa0d41d4SStefano Zampini PetscErrorCode ierr; 2722aa0d41d4SStefano Zampini 2723aa0d41d4SStefano Zampini PetscFunctionBegin; 2724d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2725d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2726d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2727d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2728aa0d41d4SStefano Zampini 2729d65f70fdSStefano Zampini if (!rsorted) { 2730906d46d4SStefano Zampini const PetscInt *idxs; 2731906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2732aa0d41d4SStefano Zampini 2733d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2734d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2735d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2736d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2737aa0d41d4SStefano Zampini } 2738d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2739d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2740d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2741d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2742aa0d41d4SStefano Zampini } 2743d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2744d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2745d65f70fdSStefano Zampini } else { 2746d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2747d65f70fdSStefano Zampini isrow_s = isrow; 2748aa0d41d4SStefano Zampini } 2749906d46d4SStefano Zampini 2750d65f70fdSStefano Zampini if (!csorted) { 2751d65f70fdSStefano Zampini if (isrow == iscol) { 2752d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2753d65f70fdSStefano Zampini iscol_s = isrow_s; 2754d65f70fdSStefano Zampini } else { 2755d65f70fdSStefano Zampini const PetscInt *idxs; 2756d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2757906d46d4SStefano Zampini 2758d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2759d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2760d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2761d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2762d65f70fdSStefano Zampini } 2763d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2764d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2765d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2766d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2767d65f70fdSStefano Zampini } 2768d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2769d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2770d65f70fdSStefano Zampini } 2771d65f70fdSStefano Zampini } else { 2772d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2773d65f70fdSStefano Zampini iscol_s = iscol; 2774d65f70fdSStefano Zampini } 2775d65f70fdSStefano Zampini 2776d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2777d65f70fdSStefano Zampini 2778d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2779906d46d4SStefano Zampini Mat new_mat; 2780d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2781906d46d4SStefano Zampini 2782d65f70fdSStefano Zampini if (!rsorted) { 2783d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2784d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2785d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2786d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2787906d46d4SStefano Zampini } 2788d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2789d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2790d65f70fdSStefano Zampini } else { 2791d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2792906d46d4SStefano Zampini } 2793d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2794d65f70fdSStefano Zampini 2795d65f70fdSStefano Zampini if (!csorted) { 2796d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2797d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2798d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2799d65f70fdSStefano Zampini } else { 2800d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2801d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2802d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2803d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2804d65f70fdSStefano Zampini } 2805d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2806d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2807d65f70fdSStefano Zampini } 2808d65f70fdSStefano Zampini } else { 2809d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2810d65f70fdSStefano Zampini } 2811d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2812d65f70fdSStefano Zampini 2813d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2814d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2815d65f70fdSStefano Zampini work_mat[0] = new_mat; 2816d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2817d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2818d65f70fdSStefano Zampini } 2819d65f70fdSStefano Zampini 2820d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2821d65f70fdSStefano Zampini *B = work_mat[0]; 2822d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2823d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2824d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2825d65f70fdSStefano Zampini PetscFunctionReturn(0); 2826d65f70fdSStefano Zampini } 2827d65f70fdSStefano Zampini 2828d65f70fdSStefano Zampini #undef __FUNCT__ 28295e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 28305e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2831aa0d41d4SStefano Zampini { 2832aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 28335e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2834d65f70fdSStefano Zampini Mat new_mat; 28355e8657edSStefano Zampini IS is_local,is_global; 2836d65f70fdSStefano Zampini PetscInt local_size; 2837d65f70fdSStefano Zampini PetscBool isseqaij; 2838aa0d41d4SStefano Zampini PetscErrorCode ierr; 2839aa0d41d4SStefano Zampini 2840aa0d41d4SStefano Zampini PetscFunctionBegin; 2841aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 28425e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 28435e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2844b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2845aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2846d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2847aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2848906d46d4SStefano Zampini 2849906d46d4SStefano Zampini /* check */ 2850906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2851906d46d4SStefano Zampini Vec x,x_change; 2852906d46d4SStefano Zampini PetscReal error; 2853906d46d4SStefano Zampini 28545e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2855906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 28565e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2857e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2858e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2859d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2860e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2861e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2862906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2863906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2864906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2865906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2866906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2867906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2868906d46d4SStefano Zampini } 2869906d46d4SStefano Zampini 287022d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 28719b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 287222d5777bSStefano Zampini if (isseqaij) { 2873a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2874a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2875aa0d41d4SStefano Zampini } else { 2876a00504b5SStefano Zampini Mat work_mat; 28771cf9b237SStefano Zampini 2878a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2879aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2880a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 28811d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2882aa0d41d4SStefano Zampini } 28833301b35fSStefano Zampini if (matis->A->symmetric_set) { 28843301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2885e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 28863301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2887e496cd5dSStefano Zampini #endif 28883301b35fSStefano Zampini } 2889d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2890aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2891aa0d41d4SStefano Zampini } 2892aa0d41d4SStefano Zampini 2893aa0d41d4SStefano Zampini #undef __FUNCT__ 2894a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 28958ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2896a64d13efSStefano Zampini { 2897a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2898a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2899d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 290053892102SStefano Zampini PetscInt *idx_R_local=NULL; 29013a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 29023a50541eSStefano Zampini PetscInt vbs,bs; 29036816873aSStefano Zampini PetscBT bitmask=NULL; 2904a64d13efSStefano Zampini PetscErrorCode ierr; 2905a64d13efSStefano Zampini 2906a64d13efSStefano Zampini PetscFunctionBegin; 2907b23d619eSStefano Zampini /* 2908b23d619eSStefano Zampini No need to setup local scatters if 2909b23d619eSStefano Zampini - primal space is unchanged 2910b23d619eSStefano Zampini AND 2911b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2912b23d619eSStefano Zampini AND 2913b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2914b23d619eSStefano Zampini */ 2915b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2916f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2917f4ddd8eeSStefano Zampini } 2918f4ddd8eeSStefano Zampini /* destroy old objects */ 2919f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2920f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2921f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2922a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2923b371cd4fSStefano Zampini n_B = pcis->n_B; 2924b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2925b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 29263a50541eSStefano Zampini 2927a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 29286816873aSStefano Zampini 292953892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2930b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2931854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2932a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2933a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 29340e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2935a64d13efSStefano Zampini } 2936a64d13efSStefano Zampini 2937a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 29384641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 29396816873aSStefano Zampini idx_R_local[n_R++] = i; 2940a64d13efSStefano Zampini } 2941a64d13efSStefano Zampini } 2942df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2943df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 29446816873aSStefano Zampini 2945df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2946df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 29476816873aSStefano Zampini } 29483a50541eSStefano Zampini 29493a50541eSStefano Zampini /* Block code */ 29503a50541eSStefano Zampini vbs = 1; 29513a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 29523a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 29533a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 29543a50541eSStefano Zampini PetscInt *vary; 2955b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2956785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 29573a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2958d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2959d3df7717SStefano 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 */ 29600e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2961d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 29623a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 29633a50541eSStefano Zampini is_blocked = PETSC_FALSE; 29643a50541eSStefano Zampini break; 29653a50541eSStefano Zampini } 29663a50541eSStefano Zampini } 2967d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2968d3df7717SStefano Zampini } else { 2969d3df7717SStefano Zampini /* Verify directly the R set */ 2970d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2971d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2972d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2973d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2974d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2975d3df7717SStefano Zampini break; 2976d3df7717SStefano Zampini } 2977d3df7717SStefano Zampini } 2978d3df7717SStefano Zampini } 2979d3df7717SStefano Zampini } 29803a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 29813a50541eSStefano Zampini vbs = bs; 29823a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 29833a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 29843a50541eSStefano Zampini } 29853a50541eSStefano Zampini } 29863a50541eSStefano Zampini } 29873a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2988b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 2989df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 299053892102SStefano Zampini 2991df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2992df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 299353892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2994df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 299553892102SStefano Zampini } else { 29963a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 299753892102SStefano Zampini } 2998a64d13efSStefano Zampini 2999a64d13efSStefano Zampini /* print some info if requested */ 3000a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 3001a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3002a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 30031575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3004a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 3005a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 30064f1b2e48SStefano 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); 3007a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3008a64d13efSStefano Zampini } 3009a64d13efSStefano Zampini 3010a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 3011b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 30126816873aSStefano Zampini IS is_aux1,is_aux2; 30136816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 30146816873aSStefano Zampini 30153a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3016854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 3017854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 3018a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 30194641a718SStefano Zampini for (i=0; i<n_D; i++) { 30204641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 30214641a718SStefano Zampini } 3022a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3023a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30244641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 30254641a718SStefano Zampini aux_array1[j++] = i; 3026a64d13efSStefano Zampini } 3027a64d13efSStefano Zampini } 3028a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3029a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3030a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 30314641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 30324641a718SStefano Zampini aux_array2[j++] = i; 3033a64d13efSStefano Zampini } 3034a64d13efSStefano Zampini } 3035a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3036a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 3037a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 3038a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3039a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 3040a64d13efSStefano Zampini 30418eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3042785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 3043a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30444641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 30454641a718SStefano Zampini aux_array1[j++] = i; 3046a64d13efSStefano Zampini } 3047a64d13efSStefano Zampini } 3048a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3049a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 3050a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3051a64d13efSStefano Zampini } 30524641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 30533a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3054d62866d3SStefano Zampini } else { 3055df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 30566816873aSStefano Zampini IS tis; 30576816873aSStefano Zampini PetscInt schur_size; 30586816873aSStefano Zampini 3059df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 30606816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 3061df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 30626816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 30636816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 30646816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 30656816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 30666816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 3067d62866d3SStefano Zampini } 3068d62866d3SStefano Zampini } 3069a64d13efSStefano Zampini PetscFunctionReturn(0); 3070a64d13efSStefano Zampini } 3071a64d13efSStefano Zampini 3072304d26faSStefano Zampini 3073304d26faSStefano Zampini #undef __FUNCT__ 3074304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 3075684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 3076304d26faSStefano Zampini { 3077304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3078304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3079304d26faSStefano Zampini PC pc_temp; 3080304d26faSStefano Zampini Mat A_RR; 3081f4ddd8eeSStefano Zampini MatReuse reuse; 3082304d26faSStefano Zampini PetscScalar m_one = -1.0; 3083304d26faSStefano Zampini PetscReal value; 308404708bb6SStefano Zampini PetscInt n_D,n_R; 3085c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 3086304d26faSStefano Zampini PetscErrorCode ierr; 3087e604994aSStefano Zampini /* prefixes stuff */ 3088312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 3089e604994aSStefano Zampini size_t len; 3090304d26faSStefano Zampini 3091304d26faSStefano Zampini PetscFunctionBegin; 3092304d26faSStefano Zampini 3093e604994aSStefano Zampini /* compute prefixes */ 3094e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 3095e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 3096e604994aSStefano Zampini if (!pcbddc->current_level) { 3097e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3098e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3099e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3100e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3101e604994aSStefano Zampini } else { 3102e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3103312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3104e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3105e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 3106312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3107312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 310834d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 310934d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3110e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3111e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3112e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 3113e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 3114e604994aSStefano Zampini } 3115e604994aSStefano Zampini 3116304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3117684f6988SStefano Zampini if (dirichlet) { 3118d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3119450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 31209a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3121450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3122a3df083aSStefano Zampini Mat A_IIn; 3123a3df083aSStefano Zampini 3124a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3125a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3126a3df083aSStefano Zampini pcis->A_II = A_IIn; 3127a3df083aSStefano Zampini } 3128450f8f5eSStefano Zampini } 31293301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31303301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3131964fefecSStefano Zampini } 3132ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3133964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3134304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3135304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3136304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3137304d26faSStefano Zampini /* default */ 3138304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3139e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 31409577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3141304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 31429577ea80SStefano Zampini if (issbaij) { 31439577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 31449577ea80SStefano Zampini } else { 3145304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31469577ea80SStefano Zampini } 3147304d26faSStefano Zampini /* Allow user's customization */ 3148304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3149304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3150304d26faSStefano Zampini } 3151d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3152b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3153df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3154d62866d3SStefano Zampini 3155df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3156d5574798SStefano Zampini } 3157304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3158304d26faSStefano Zampini if (!n_D) { 3159304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3160304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3161304d26faSStefano Zampini } 3162304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3163304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3164304d26faSStefano Zampini /* set ksp_D into pcis data */ 3165304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3166304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3167304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3168684f6988SStefano Zampini } 3169304d26faSStefano Zampini 3170304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3171684f6988SStefano Zampini A_RR = 0; 3172684f6988SStefano Zampini if (neumann) { 3173d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 317404708bb6SStefano Zampini PetscInt ibs,mbs; 317504708bb6SStefano Zampini PetscBool issbaij; 317604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3177f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 31788ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3179f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3180f4ddd8eeSStefano Zampini PetscInt nn_R; 318181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3182f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3183f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3184f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3185f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3186f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3187f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3188f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3189727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3190f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3191f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3192f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3193f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3194f4ddd8eeSStefano Zampini } 3195f4ddd8eeSStefano Zampini } 3196f4ddd8eeSStefano Zampini /* last check */ 3197d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3198f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3199f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3200f4ddd8eeSStefano Zampini } 3201f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3202f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3203f4ddd8eeSStefano Zampini } 3204a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3205af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3206af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 320704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 320804708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 320904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 321004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 321104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3212af732b37SStefano Zampini } else { 3213511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 32146816873aSStefano Zampini } 321504708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 321604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 321704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 321804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 321904708bb6SStefano Zampini } else { 3220511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 322104708bb6SStefano Zampini } 322204708bb6SStefano Zampini } 3223a00504b5SStefano Zampini /* extract A_RR */ 3224b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3225a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3226a00504b5SStefano Zampini 3227a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 322816e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3229a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 323016e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 323116e386b8SStefano Zampini } else { 3232a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3233a00504b5SStefano Zampini } 3234a00504b5SStefano Zampini } else { 3235a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3236a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3237a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3238a00504b5SStefano Zampini } 3239a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3240f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 324116e386b8SStefano Zampini } 32423301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 32433301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 32446816873aSStefano Zampini } 3245f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3246304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3247304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3248304d26faSStefano Zampini /* default */ 3249304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3250e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3251304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 32529577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 32539577ea80SStefano Zampini if (issbaij) { 32549577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 32559577ea80SStefano Zampini } else { 3256304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 32579577ea80SStefano Zampini } 3258304d26faSStefano Zampini /* Allow user's customization */ 3259304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3260304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3261304d26faSStefano Zampini } 3262304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3263304d26faSStefano Zampini if (!n_R) { 3264304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3265304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3266304d26faSStefano Zampini } 32675cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3268df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3269b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3270df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3271d62866d3SStefano Zampini 3272df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3273d62866d3SStefano Zampini } 3274304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3275304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3276684f6988SStefano Zampini } 3277304d26faSStefano Zampini 3278684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3279684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32801575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3281684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3282684f6988SStefano Zampini } 3283c7017625SStefano Zampini 3284c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3285c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3286c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3287c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3288c7017625SStefano Zampini } 3289c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3290c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3291c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3292c7017625SStefano Zampini } 3293c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3294c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3295c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3296c7017625SStefano Zampini } 3297c7017625SStefano Zampini 3298c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3299c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3300684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 33010fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 33020fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 33030fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 33040fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 33050fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3306e604994aSStefano 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); 3307c7017625SStefano Zampini if (check_corr[0]) { 3308c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3309c7017625SStefano Zampini } 3310304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3311304d26faSStefano Zampini } 3312684f6988SStefano Zampini if (neumann) { /* Neumann */ 33130fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 33140fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 33150fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 33160fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 33170fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3318e604994aSStefano 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); 3319c7017625SStefano Zampini if (check_corr[1]) { 3320c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3321c7017625SStefano Zampini } 3322304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3323304d26faSStefano Zampini } 3324684f6988SStefano Zampini } 33255cbda25cSStefano Zampini /* free Neumann problem's matrix */ 33265cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3327304d26faSStefano Zampini PetscFunctionReturn(0); 3328304d26faSStefano Zampini } 3329304d26faSStefano Zampini 3330304d26faSStefano Zampini #undef __FUNCT__ 3331ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 333280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3333674ae819SStefano Zampini { 3334674ae819SStefano Zampini PetscErrorCode ierr; 3335674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3336be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3337b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 3338674ae819SStefano Zampini 3339674ae819SStefano Zampini PetscFunctionBegin; 3340b334f244SStefano Zampini if (!reuse_solver) { 334180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 334220c7b377SStefano Zampini } 334380677318SStefano Zampini if (!pcbddc->switch_static) { 334480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 334580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 334680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 334720c7b377SStefano Zampini } 3348b334f244SStefano Zampini if (!reuse_solver) { 334980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335120c7b377SStefano Zampini } else { 3352df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3353be83ff47SStefano Zampini 3354df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3355df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 335620c7b377SStefano Zampini } 3357be83ff47SStefano Zampini } else { 335880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 336380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 336480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 336580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3367674ae819SStefano Zampini } 3368674ae819SStefano Zampini } 3369b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 337080677318SStefano Zampini if (applytranspose) { 337180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 337280677318SStefano Zampini } else { 337380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 337480677318SStefano Zampini } 3375be83ff47SStefano Zampini } else { 3376df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3377be83ff47SStefano Zampini 3378be83ff47SStefano Zampini if (applytranspose) { 3379df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3380be83ff47SStefano Zampini } else { 3381df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3382be83ff47SStefano Zampini } 3383be83ff47SStefano Zampini } 338480677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 338580677318SStefano Zampini if (!pcbddc->switch_static) { 3386b334f244SStefano Zampini if (!reuse_solver) { 338780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 338880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3389be83ff47SStefano Zampini } else { 3390df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3391be83ff47SStefano Zampini 3392df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3393df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3394be83ff47SStefano Zampini } 339580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 339680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 339780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 339880677318SStefano Zampini } 339980677318SStefano Zampini } else { 340080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 340580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 340680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 340780677318SStefano Zampini } 340880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3412674ae819SStefano Zampini } 3413674ae819SStefano Zampini PetscFunctionReturn(0); 3414674ae819SStefano Zampini } 3415674ae819SStefano Zampini 3416dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3417674ae819SStefano Zampini #undef __FUNCT__ 3418674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3419dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3420674ae819SStefano Zampini { 3421674ae819SStefano Zampini PetscErrorCode ierr; 3422674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3423674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3424674ae819SStefano Zampini const PetscScalar zero = 0.0; 3425674ae819SStefano Zampini 3426674ae819SStefano Zampini PetscFunctionBegin; 3427dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 34284fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3429dc359a40SStefano Zampini if (applytranspose) { 3430674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 34318eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3432dc359a40SStefano Zampini } else { 3433674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3434674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 343515aaf578SStefano Zampini } 34364fee134fSStefano Zampini } else { 34374fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 34384fee134fSStefano Zampini } 3439efc2fbd9SStefano Zampini 3440efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 34414f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3442efc2fbd9SStefano Zampini PetscScalar *array; 34434f1b2e48SStefano Zampini PetscInt j; 3444efc2fbd9SStefano Zampini 3445efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34464f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3447efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3448efc2fbd9SStefano Zampini } 3449efc2fbd9SStefano Zampini 345012edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 345112edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 345212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 345312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 345412edc857SStefano Zampini 34559f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 345612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 345751694757SStefano Zampini Mat coarse_mat; 3458964fefecSStefano Zampini Vec rhs,sol; 345951694757SStefano Zampini MatNullSpace nullsp; 346027b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3461964fefecSStefano Zampini 346227b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 346327b6a85dSStefano Zampini PC coarse_pc; 346427b6a85dSStefano Zampini 346527b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 346627b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 346727b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 346827b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 346927b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 347027b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 34713bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 347227b6a85dSStefano Zampini } 347327b6a85dSStefano Zampini } 3474964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3475964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 347651694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 347751694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 347851694757SStefano Zampini if (nullsp) { 347951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 348051694757SStefano Zampini } 348112edc857SStefano Zampini if (applytranspose) { 34829a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 3483964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 34842701bc32SStefano Zampini } else { 34851f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 34862701bc32SStefano Zampini PC coarse_pc; 34872701bc32SStefano Zampini 34882701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 34892701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 34903e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 34912701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 349212edc857SStefano Zampini } else { 3493964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 349412edc857SStefano Zampini } 34952701bc32SStefano Zampini } 34961d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 349727b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 349827b6a85dSStefano Zampini PC coarse_pc; 349927b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 350027b6a85dSStefano Zampini 350127b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 350227b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 350327b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 35043bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 350527b6a85dSStefano Zampini } 350651694757SStefano Zampini if (nullsp) { 350751694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 350851694757SStefano Zampini } 350912edc857SStefano Zampini } 3510674ae819SStefano Zampini 3511674ae819SStefano Zampini /* Local solution on R nodes */ 35124fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 351380677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 35149f00e9b4SStefano Zampini } 35159f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 35169f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 351712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3518674ae819SStefano Zampini 35194fee134fSStefano Zampini /* Sum contributions from the two levels */ 35204fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3521dc359a40SStefano Zampini if (applytranspose) { 3522dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3523dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3524dc359a40SStefano Zampini } else { 3525674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 35268eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3527dc359a40SStefano Zampini } 3528efc2fbd9SStefano Zampini /* store p0 */ 35294f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3530efc2fbd9SStefano Zampini PetscScalar *array; 35314f1b2e48SStefano Zampini PetscInt j; 3532efc2fbd9SStefano Zampini 3533efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 35344f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3535efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3536efc2fbd9SStefano Zampini } 35374fee134fSStefano Zampini } else { /* expand the coarse solution */ 35384fee134fSStefano Zampini if (applytranspose) { 35394fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35404fee134fSStefano Zampini } else { 35414fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35424fee134fSStefano Zampini } 35434fee134fSStefano Zampini } 3544674ae819SStefano Zampini PetscFunctionReturn(0); 3545674ae819SStefano Zampini } 3546674ae819SStefano Zampini 3547674ae819SStefano Zampini #undef __FUNCT__ 3548674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 354912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3550674ae819SStefano Zampini { 3551674ae819SStefano Zampini PetscErrorCode ierr; 3552674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 355358da7f69SStefano Zampini PetscScalar *array; 355412edc857SStefano Zampini Vec from,to; 3555674ae819SStefano Zampini 3556674ae819SStefano Zampini PetscFunctionBegin; 355712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 355812edc857SStefano Zampini from = pcbddc->coarse_vec; 355912edc857SStefano Zampini to = pcbddc->vec1_P; 356012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 356112edc857SStefano Zampini Vec tvec; 356258da7f69SStefano Zampini 356358da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 356458da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 356512edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 356658da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 356758da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 356858da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 356912edc857SStefano Zampini } 357012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 357112edc857SStefano Zampini from = pcbddc->vec1_P; 357212edc857SStefano Zampini to = pcbddc->coarse_vec; 357312edc857SStefano Zampini } 357412edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3575674ae819SStefano Zampini PetscFunctionReturn(0); 3576674ae819SStefano Zampini } 3577674ae819SStefano Zampini 3578674ae819SStefano Zampini #undef __FUNCT__ 3579674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 358012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3581674ae819SStefano Zampini { 3582674ae819SStefano Zampini PetscErrorCode ierr; 3583674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 358458da7f69SStefano Zampini PetscScalar *array; 358512edc857SStefano Zampini Vec from,to; 3586674ae819SStefano Zampini 3587674ae819SStefano Zampini PetscFunctionBegin; 358812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 358912edc857SStefano Zampini from = pcbddc->coarse_vec; 359012edc857SStefano Zampini to = pcbddc->vec1_P; 359112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 359212edc857SStefano Zampini from = pcbddc->vec1_P; 359312edc857SStefano Zampini to = pcbddc->coarse_vec; 359412edc857SStefano Zampini } 359512edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 359612edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 359712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 359812edc857SStefano Zampini Vec tvec; 359958da7f69SStefano Zampini 360012edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 360158da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 360258da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 360358da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 360458da7f69SStefano Zampini } 360558da7f69SStefano Zampini } else { 360658da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 360758da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 360812edc857SStefano Zampini } 360912edc857SStefano Zampini } 3610674ae819SStefano Zampini PetscFunctionReturn(0); 3611674ae819SStefano Zampini } 3612674ae819SStefano Zampini 3613984c4197SStefano Zampini /* uncomment for testing purposes */ 3614984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3615674ae819SStefano Zampini #undef __FUNCT__ 3616674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3617674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3618674ae819SStefano Zampini { 3619674ae819SStefano Zampini PetscErrorCode ierr; 3620674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3621674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3622674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3623984c4197SStefano Zampini /* one and zero */ 3624984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3625984c4197SStefano Zampini /* space to store constraints and their local indices */ 36269162d606SStefano Zampini PetscScalar *constraints_data; 36279162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 36289162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 36299162d606SStefano Zampini PetscInt *constraints_n; 3630984c4197SStefano Zampini /* iterators */ 3631b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3632984c4197SStefano Zampini /* BLAS integers */ 3633e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3634e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3635c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3636727cdba6SStefano Zampini /* reuse */ 36370e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 36380e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3639984c4197SStefano Zampini /* change of basis */ 3640b3d85658SStefano Zampini PetscBool qr_needed; 36419162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3642984c4197SStefano Zampini /* auxiliary stuff */ 364364efe560SStefano Zampini PetscInt *nnz,*is_indices; 36448a0068c3SStefano Zampini PetscInt ncc; 3645984c4197SStefano Zampini /* some quantities */ 364645a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3647a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3648984c4197SStefano Zampini 3649674ae819SStefano Zampini PetscFunctionBegin; 36508e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 36518e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 36528e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 365316909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3654088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3655088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 36560e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 36570e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 36580e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36590e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36600e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3661088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3662cf5a6209SStefano Zampini 3663cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 36649162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3665cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3666cf5a6209SStefano Zampini const Vec *nearnullvecs; 3667cf5a6209SStefano Zampini Vec *localnearnullsp; 3668cf5a6209SStefano Zampini PetscScalar *array; 3669cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3670cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3671674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3672b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3673674ae819SStefano Zampini PetscScalar *work; 3674674ae819SStefano Zampini PetscReal *singular_vals; 3675674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3676674ae819SStefano Zampini PetscReal *rwork; 3677674ae819SStefano Zampini #endif 3678674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3679674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3680674ae819SStefano Zampini #else 3681964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3682964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3683674ae819SStefano Zampini #endif 3684674ae819SStefano Zampini 3685674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3686d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3687e4d548c7SStefano Zampini /* print some info */ 36881f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3689e4d548c7SStefano Zampini PetscInt nv; 3690e4d548c7SStefano Zampini 36918ae0ca82SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag-1,pcbddc->dbg_viewer);CHKERRQ(ierr); 3692e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3693e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3694e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3695e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3696e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3697e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3698e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3699e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3700e4d548c7SStefano Zampini } 3701e4d548c7SStefano Zampini 3702d06fc5fdSStefano Zampini /* free unneeded index sets */ 3703d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3704d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3705674ae819SStefano Zampini } 3706d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3707d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3708d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3709d06fc5fdSStefano Zampini } 3710d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3711d06fc5fdSStefano Zampini n_ISForEdges = 0; 3712d06fc5fdSStefano Zampini } 3713d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3714d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3715d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3716d06fc5fdSStefano Zampini } 3717d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3718d06fc5fdSStefano Zampini n_ISForFaces = 0; 3719d06fc5fdSStefano Zampini } 372070022509SStefano Zampini 3721674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3722674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3723674ae819SStefano Zampini if (nearnullsp) { 3724674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3725f4ddd8eeSStefano Zampini /* remove any stored info */ 3726f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3727f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3728f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3729f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3730f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3731473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3732f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3733f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3734f4ddd8eeSStefano Zampini } 3735984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3736984c4197SStefano Zampini nnsp_size = 0; 3737674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3738674ae819SStefano Zampini } 3739984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3740984c4197SStefano Zampini max_constraints = nnsp_size; 3741984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3742984c4197SStefano Zampini 3743674ae819SStefano Zampini /* 3744674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 37459162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 37469162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 37479162d606SStefano Zampini There can be multiple constraints per connected component 3748674ae819SStefano Zampini */ 3749674ae819SStefano Zampini n_vertices = 0; 3750674ae819SStefano Zampini if (ISForVertices) { 3751674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3752674ae819SStefano Zampini } 37539162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 37549162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 37559162d606SStefano Zampini 37569162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 37579162d606SStefano Zampini total_counts *= max_constraints; 3758674ae819SStefano Zampini total_counts += n_vertices; 37594641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 37609162d606SStefano Zampini 3761674ae819SStefano Zampini total_counts = 0; 3762674ae819SStefano Zampini max_size_of_constraint = 0; 3763674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 37649162d606SStefano Zampini IS used_is; 3765674ae819SStefano Zampini if (i<n_ISForEdges) { 37669162d606SStefano Zampini used_is = ISForEdges[i]; 3767674ae819SStefano Zampini } else { 37689162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3769674ae819SStefano Zampini } 37709162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3771674ae819SStefano Zampini total_counts += j; 3772674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3773674ae819SStefano Zampini } 37749162d606SStefano 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); 37759162d606SStefano Zampini 3776984c4197SStefano Zampini /* get local part of global near null space vectors */ 3777785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3778984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3779984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3780e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3781e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3782984c4197SStefano Zampini } 3783674ae819SStefano Zampini 3784242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3785242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3786a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3787242a89d7SStefano Zampini 3788984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3789a773dcb8SStefano Zampini if (!skip_lapack) { 3790674ae819SStefano Zampini PetscScalar temp_work; 3791911cabfeSStefano Zampini 3792674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3793984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3794785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3795785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3796785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3797674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3798785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3799674ae819SStefano Zampini #endif 3800674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3801c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3802c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3803674ae819SStefano Zampini lwork = -1; 3804674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3805674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3806c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3807674ae819SStefano Zampini #else 3808c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3809674ae819SStefano Zampini #endif 3810674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3811984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3812674ae819SStefano Zampini #else /* on missing GESVD */ 3813674ae819SStefano Zampini /* SVD */ 3814674ae819SStefano Zampini PetscInt max_n,min_n; 3815674ae819SStefano Zampini max_n = max_size_of_constraint; 3816984c4197SStefano Zampini min_n = max_constraints; 3817984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3818674ae819SStefano Zampini min_n = max_size_of_constraint; 3819984c4197SStefano Zampini max_n = max_constraints; 3820674ae819SStefano Zampini } 3821785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3822674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3823785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3824674ae819SStefano Zampini #endif 3825674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3826674ae819SStefano Zampini lwork = -1; 3827e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3828e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3829b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3830674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3831674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 38329162d606SStefano 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)); 3833674ae819SStefano Zampini #else 38349162d606SStefano 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)); 3835674ae819SStefano Zampini #endif 3836674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3837984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3838984c4197SStefano Zampini #endif /* on missing GESVD */ 3839674ae819SStefano Zampini /* Allocate optimal workspace */ 3840674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3841854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3842674ae819SStefano Zampini } 3843674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3844674ae819SStefano Zampini total_counts = 0; 38459162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 38469162d606SStefano Zampini constraints_data_ptr[0] = 0; 3847674ae819SStefano Zampini /* vertices */ 38489162d606SStefano Zampini if (n_vertices) { 3849674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 38509162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3851674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 38529162d606SStefano Zampini constraints_n[total_counts] = 1; 38539162d606SStefano Zampini constraints_data[total_counts] = 1.0; 38549162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 38559162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3856674ae819SStefano Zampini total_counts++; 3857674ae819SStefano Zampini } 3858674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3859674ae819SStefano Zampini n_vertices = total_counts; 3860674ae819SStefano Zampini } 3861984c4197SStefano Zampini 3862674ae819SStefano Zampini /* edges and faces */ 38639162d606SStefano Zampini total_counts_cc = total_counts; 3864911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 38659162d606SStefano Zampini IS used_is; 38669162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 38679162d606SStefano Zampini 3868911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 38699162d606SStefano Zampini used_is = ISForEdges[ncc]; 3870984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3871674ae819SStefano Zampini } else { 38729162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3873984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3874674ae819SStefano Zampini } 3875674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 38769162d606SStefano Zampini 38779162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 38789162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3879984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3880984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3881674ae819SStefano Zampini if (nnsp_has_cnst) { 38825b08dc53SStefano Zampini PetscScalar quad_value; 38839162d606SStefano Zampini 38849162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38859162d606SStefano Zampini idxs_copied = PETSC_TRUE; 38869162d606SStefano Zampini 3887a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3888674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3889a773dcb8SStefano Zampini } else { 3890a773dcb8SStefano Zampini quad_value = 1.0; 3891a773dcb8SStefano Zampini } 3892674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38939162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3894674ae819SStefano Zampini } 38959162d606SStefano Zampini temp_constraints++; 3896674ae819SStefano Zampini total_counts++; 3897674ae819SStefano Zampini } 3898674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3899984c4197SStefano Zampini PetscReal real_value; 39009162d606SStefano Zampini PetscScalar *ptr_to_data; 39019162d606SStefano Zampini 3902984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 39039162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3904674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 39059162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3906674ae819SStefano Zampini } 3907984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3908984c4197SStefano Zampini /* check if array is null on the connected component */ 3909e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39109162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 39115b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3912674ae819SStefano Zampini temp_constraints++; 3913674ae819SStefano Zampini total_counts++; 39149162d606SStefano Zampini if (!idxs_copied) { 39159162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 39169162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3917674ae819SStefano Zampini } 3918674ae819SStefano Zampini } 39199162d606SStefano Zampini } 39209162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 392145a1bb75SStefano Zampini valid_constraints = temp_constraints; 3922eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3923a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 39249162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 39259162d606SStefano Zampini 39269162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3927a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39289162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3929a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 39309162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3931a773dcb8SStefano Zampini } else { /* perform SVD */ 3932984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 39339162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3934674ae819SStefano Zampini 3935674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3936984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3937984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3938984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3939984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3940984c4197SStefano Zampini from that computed using LAPACKgesvd 3941984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3942984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3943984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3944674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3945e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3946984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3947674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3948674ae819SStefano Zampini for (k=0;k<j+1;k++) { 39499162d606SStefano 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)); 3950674ae819SStefano Zampini } 3951674ae819SStefano Zampini } 3952e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3953e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3954e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3955674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3956c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3957674ae819SStefano Zampini #else 3958c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3959674ae819SStefano Zampini #endif 3960674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3961984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3962984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3963674ae819SStefano Zampini j = 0; 3964984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3965674ae819SStefano Zampini total_counts = total_counts-j; 396645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3967e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3968c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3969c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3970c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3971c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3972c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3973c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3974674ae819SStefano Zampini if (j<temp_constraints) { 3975984c4197SStefano Zampini PetscInt ii; 3976984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3977674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39789162d606SStefano 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)); 3979674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3980984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3981674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 39829162d606SStefano 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]; 3983674ae819SStefano Zampini } 3984674ae819SStefano Zampini } 3985674ae819SStefano Zampini } 3986674ae819SStefano Zampini #else /* on missing GESVD */ 3987e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3988e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3989b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3990674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3991674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 39929162d606SStefano 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)); 3993674ae819SStefano Zampini #else 39949162d606SStefano 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)); 3995674ae819SStefano Zampini #endif 3996984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3997674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3998984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3999e310c8b4SStefano Zampini k = temp_constraints; 4000e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 4001674ae819SStefano Zampini j = 0; 4002e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 400345a1bb75SStefano Zampini valid_constraints = k-j; 4004911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 4005984c4197SStefano Zampini #endif /* on missing GESVD */ 4006674ae819SStefano Zampini } 4007a773dcb8SStefano Zampini } 40089162d606SStefano Zampini /* update pointers information */ 40099162d606SStefano Zampini if (valid_constraints) { 40109162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 40119162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 40129162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 40139162d606SStefano Zampini /* set change_of_basis flag */ 401445a1bb75SStefano Zampini if (boolforchange) { 4015b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 40169162d606SStefano Zampini } 4017b3d85658SStefano Zampini total_counts_cc++; 401845a1bb75SStefano Zampini } 401945a1bb75SStefano Zampini } 4020984c4197SStefano Zampini /* free workspace */ 40218f1c130eSStefano Zampini if (!skip_lapack) { 4022984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4023984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4024984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 4025984c4197SStefano Zampini #endif 4026984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 4027984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4028984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 4029984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 4030984c4197SStefano Zampini #endif 4031984c4197SStefano Zampini } 4032984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4033984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 4034984c4197SStefano Zampini } 4035984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 4036cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 4037cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4038cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4039cf5a6209SStefano Zampini } 4040cf5a6209SStefano Zampini if (n_ISForFaces) { 4041cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4042cf5a6209SStefano Zampini } 4043cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4044cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4045cf5a6209SStefano Zampini } 4046cf5a6209SStefano Zampini if (n_ISForEdges) { 4047cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4048cf5a6209SStefano Zampini } 4049cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 405008122e43SStefano Zampini } else { 405108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4052984c4197SStefano Zampini 405308122e43SStefano Zampini total_counts = 0; 405408122e43SStefano Zampini n_vertices = 0; 4055d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 4056d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 405708122e43SStefano Zampini } 405808122e43SStefano Zampini max_constraints = 0; 40599162d606SStefano Zampini total_counts_cc = 0; 406008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 406108122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 40629162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 406308122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 406408122e43SStefano Zampini } 40659162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 40669162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 40679162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 40689162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 406974d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 40709162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 40719162d606SStefano Zampini total_counts_cc = 0; 40729162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 40739162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 40749162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 407508122e43SStefano Zampini } 407608122e43SStefano Zampini } 40779162d606SStefano Zampini #if 0 40789162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 40799162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40809162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 40819162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 40829162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 40839162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 40849162d606SStefano Zampini } 40859162d606SStefano Zampini printf("\n"); 40869162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 40879162d606SStefano Zampini } 40881b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 40898bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 40901b968477SStefano Zampini } 40911b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 40928bec7fa6SStefano 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]); 40931b968477SStefano Zampini } 409408122e43SStefano Zampini #endif 409508122e43SStefano Zampini 40968bec7fa6SStefano Zampini max_size_of_constraint = 0; 40979162d606SStefano 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]); 40989162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 409908122e43SStefano Zampini /* Change of basis */ 4100b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 410108122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 410208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 410308122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 4104b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 410508122e43SStefano Zampini } 410608122e43SStefano Zampini } 410708122e43SStefano Zampini } 410808122e43SStefano Zampini } 4109984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 41104f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 411108122e43SStefano Zampini 41129162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 41139162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 41146c4ed002SBarry 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); 4115674ae819SStefano Zampini 4116674ae819SStefano Zampini /* Create constraint matrix */ 4117674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 411816f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4119984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4120984c4197SStefano Zampini 4121984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4122a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4123a717540cSStefano Zampini qr_needed = PETSC_FALSE; 412474d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4125984c4197SStefano Zampini total_primal_vertices=0; 4126b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 41279162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 41289162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 412972b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 41309162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4131b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 413264efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 41339162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41349162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4135a717540cSStefano Zampini } 4136b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 413791af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4138a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4139a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4140a717540cSStefano Zampini } 4141fa434743SStefano Zampini } else { 4142b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4143fa434743SStefano Zampini } 4144a717540cSStefano Zampini } 4145b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4146b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4147674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 414870022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4149b3d85658SStefano Zampini 41504f1b2e48SStefano 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); 41510e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 41520e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4153984c4197SStefano Zampini 4154984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 415574d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4156785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4157984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 415874d5cdf7SStefano Zampini 4159984c4197SStefano Zampini j = total_primal_vertices; 416074d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4161b3d85658SStefano Zampini cum = total_primal_vertices; 41629162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41634641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4164b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4165b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4166b3d85658SStefano Zampini cum++; 41679162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 416874d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 416974d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 417074d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 417174d5cdf7SStefano Zampini } 41729162d606SStefano Zampini j += constraints_n[i]; 4173674ae819SStefano Zampini } 4174674ae819SStefano Zampini } 4175674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4176674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4177088faed8SStefano Zampini 4178674ae819SStefano Zampini /* set values in constraint matrix */ 4179984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 41800e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4181674ae819SStefano Zampini } 4182984c4197SStefano Zampini total_counts = total_primal_vertices; 41839162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41844641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 41859162d606SStefano Zampini PetscInt *cols; 41869162d606SStefano Zampini 41879162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 41889162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 41899162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41909162d606SStefano Zampini PetscInt row = total_counts+k; 41919162d606SStefano Zampini PetscScalar *vals; 41929162d606SStefano Zampini 41939162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 41949162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 41959162d606SStefano Zampini } 41969162d606SStefano Zampini total_counts += constraints_n[i]; 4197674ae819SStefano Zampini } 4198674ae819SStefano Zampini } 4199674ae819SStefano Zampini /* assembling */ 4200674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4201674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4202088faed8SStefano Zampini 4203984c4197SStefano Zampini /* 42046a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4205984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4206f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4207984c4197SStefano Zampini */ 4208674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4209674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4210026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4211984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4212984c4197SStefano Zampini /* working stuff for GEQRF */ 421381d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4214984c4197SStefano Zampini PetscBLASInt lqr_work; 4215984c4197SStefano Zampini /* working stuff for UNGQR */ 4216984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4217984c4197SStefano Zampini PetscBLASInt lgqr_work; 4218984c4197SStefano Zampini /* working stuff for TRTRS */ 4219984c4197SStefano Zampini PetscScalar *trs_rhs; 42203f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4221984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4222984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4223984c4197SStefano Zampini PetscScalar *start_vals; 4224984c4197SStefano Zampini /* working stuff for values insertion */ 42254641a718SStefano Zampini PetscBT is_primal; 422664efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4227906d46d4SStefano Zampini /* matrix sizes */ 4228906d46d4SStefano Zampini PetscInt global_size,local_size; 4229906d46d4SStefano Zampini /* temporary change of basis */ 4230906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4231cf5a6209SStefano Zampini /* extra space for debugging */ 4232cf5a6209SStefano Zampini PetscScalar *dbg_work; 4233984c4197SStefano Zampini 4234906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4235906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 423616f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4237bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4238906d46d4SStefano Zampini /* nonzeros for local mat */ 4239bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 42401dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4241bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 42421dd7afcfSStefano Zampini } else { 42431dd7afcfSStefano Zampini const PetscInt *ii; 42441dd7afcfSStefano Zampini PetscInt n; 42451dd7afcfSStefano Zampini PetscBool flg_row; 42461dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42471dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 42481dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42491dd7afcfSStefano Zampini } 42509162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4251a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 42529162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4253a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 42549162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4255a717540cSStefano Zampini } else { 42569162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 42579162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4258a717540cSStefano Zampini } 4259a717540cSStefano Zampini } 4260a717540cSStefano Zampini } 4261906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4262bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 42631dd7afcfSStefano Zampini /* Set interior change in the matrix */ 42641dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4265bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4266906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4267a717540cSStefano Zampini } 42681dd7afcfSStefano Zampini } else { 42691dd7afcfSStefano Zampini const PetscInt *ii,*jj; 42701dd7afcfSStefano Zampini PetscScalar *aa; 42711dd7afcfSStefano Zampini PetscInt n; 42721dd7afcfSStefano Zampini PetscBool flg_row; 42731dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42741dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42751dd7afcfSStefano Zampini for (i=0;i<n;i++) { 42761dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 42771dd7afcfSStefano Zampini } 42781dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42791dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42801dd7afcfSStefano Zampini } 4281a717540cSStefano Zampini 4282a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4283a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4284a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4285a717540cSStefano Zampini } 4286a717540cSStefano Zampini 4287a717540cSStefano Zampini 4288a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4289a717540cSStefano Zampini /* 4290a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4291a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4292a717540cSStefano Zampini 4293a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4294a717540cSStefano Zampini 4295a6b551f4SStefano 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) 4296a6b551f4SStefano Zampini 4297a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4298a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4299a717540cSStefano Zampini | ... | 4300a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4301a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4302a717540cSStefano Zampini 4303a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4304a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4305a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4306a6b551f4SStefano Zampini 4307a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4308a717540cSStefano Zampini */ 4309a717540cSStefano Zampini if (qr_needed) { 4310984c4197SStefano Zampini /* space to store Q */ 4311854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4312984c4197SStefano Zampini /* first we issue queries for optimal work */ 43133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43143f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 43153f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4316984c4197SStefano Zampini lqr_work = -1; 43173f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4318984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4319984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4320785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4321984c4197SStefano Zampini lgqr_work = -1; 43223f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43233f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 43243f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 43253f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43263f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 43273f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4328984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4329984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4330785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4331984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4332785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4333984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4334785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4335a717540cSStefano Zampini /* allocating workspace for check */ 4336a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4337cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4338a717540cSStefano Zampini } 4339a717540cSStefano Zampini } 4340984c4197SStefano Zampini /* array to store whether a node is primal or not */ 43414641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4342473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 43430e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 43446c4ed002SBarry 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); 434539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 434639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 434739e2fb2aSStefano Zampini } 434839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4349984c4197SStefano Zampini 4350a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 43519162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 43529162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 43534641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4354984c4197SStefano Zampini /* get constraint info */ 43559162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4356984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4357984c4197SStefano Zampini 4358984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43599162d606SStefano 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); 4360674ae819SStefano Zampini } 4361984c4197SStefano Zampini 4362fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4363a717540cSStefano Zampini 4364a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4365a717540cSStefano Zampini if (pcbddc->dbg_flag) { 43669162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4367a717540cSStefano Zampini } 4368984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 43699162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4370984c4197SStefano Zampini 4371984c4197SStefano Zampini /* compute QR decomposition of constraints */ 43723f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4375674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43763f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4377984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4378674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4379984c4197SStefano Zampini 4380984c4197SStefano Zampini /* explictly compute R^-T */ 4381984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4382984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 43833f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43843f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 43853f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43863f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4387984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43883f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4389984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4390984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4391984c4197SStefano Zampini 4392a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 43933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 43953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 43963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4397984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43983f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4399984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4400984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4401984c4197SStefano Zampini 4402984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4403984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4404984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 44053f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44063f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 44073f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 44083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 44093f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 44103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4411984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44129162d606SStefano 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)); 4413984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 44149162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4415984c4197SStefano Zampini 4416984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 44179162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4418984c4197SStefano Zampini /* insert cols for primal dofs */ 4419984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4420984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 44219162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4422906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4423984c4197SStefano Zampini } 4424984c4197SStefano Zampini /* insert cols for dual dofs */ 4425984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 44269162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4427984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 44289162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4429906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4430984c4197SStefano Zampini j++; 4431674ae819SStefano Zampini } 4432674ae819SStefano Zampini } 4433984c4197SStefano Zampini 4434984c4197SStefano Zampini /* check change of basis */ 4435984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4436984c4197SStefano Zampini PetscInt ii,jj; 4437984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4438c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4439c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4440c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4441c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4442c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4443c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4444984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4445cf5a6209SStefano 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)); 4446984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4447984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4448984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4449cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4450cf5a6209SStefano 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; 4451674ae819SStefano Zampini } 4452674ae819SStefano Zampini } 4453984c4197SStefano Zampini if (!valid_qr) { 445422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4455984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4456984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4457cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4458cf5a6209SStefano 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])); 4459674ae819SStefano Zampini } 4460cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4461cf5a6209SStefano 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])); 4462984c4197SStefano Zampini } 4463984c4197SStefano Zampini } 4464984c4197SStefano Zampini } 4465674ae819SStefano Zampini } else { 446622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4467674ae819SStefano Zampini } 4468674ae819SStefano Zampini } 4469a717540cSStefano Zampini } else { /* simple transformation block */ 4470a717540cSStefano Zampini PetscInt row,col; 4471a6b551f4SStefano Zampini PetscScalar val,norm; 4472a6b551f4SStefano Zampini 4473a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44749162d606SStefano 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)); 4475a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 44769162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 44779162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4478bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 44799162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4480906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 44819162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4482a717540cSStefano Zampini } else { 4483a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 44849162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4485a717540cSStefano Zampini if (row != col) { 44869162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4487a717540cSStefano Zampini } else { 44889162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4489a717540cSStefano Zampini } 4490906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4491a717540cSStefano Zampini } 4492a717540cSStefano Zampini } 4493a717540cSStefano Zampini } 449498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 449522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4496a717540cSStefano Zampini } 4497674ae819SStefano Zampini } 4498984c4197SStefano Zampini } else { 4499984c4197SStefano Zampini if (pcbddc->dbg_flag) { 45009162d606SStefano 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); 4501674ae819SStefano Zampini } 4502674ae819SStefano Zampini } 4503674ae819SStefano Zampini } 4504a717540cSStefano Zampini 4505a717540cSStefano Zampini /* free workspace */ 4506a717540cSStefano Zampini if (qr_needed) { 4507984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4508cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4509984c4197SStefano Zampini } 4510984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4511984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4512984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4513984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4514984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4515674ae819SStefano Zampini } 4516a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4517906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4518906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4519906d46d4SStefano Zampini 4520906d46d4SStefano Zampini /* assembling of global change of variable */ 452188c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4522bbb9e6c6SStefano Zampini Mat tmat; 452316f15bc4SStefano Zampini PetscInt bs; 452416f15bc4SStefano Zampini 4525906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4526906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4527bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4528bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4529bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4530bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 453116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 453216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4533906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4534bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4535bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4536bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4537bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4538bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4539e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4540e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4541bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4542bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 454388c03ad3SStefano Zampini 4544906d46d4SStefano Zampini /* check */ 4545906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4546906d46d4SStefano Zampini PetscReal error; 4547906d46d4SStefano Zampini Vec x,x_change; 4548906d46d4SStefano Zampini 4549906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4550906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4551906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4552906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4553e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4554e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4555bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4556e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4557e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4558906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4559906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4560906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4561906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4562bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4563906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4564906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4565906d46d4SStefano Zampini } 4566b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4567b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4568b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4569bf3a8328SStefano Zampini 45709a962809SStefano 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); 4571b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 4572ac632422SStefano Zampini Mat S_new,tmat; 4573bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4574bbb9e6c6SStefano Zampini 4575bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 45766816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4577bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4578bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4579bf3a8328SStefano Zampini IS is_V; 4580b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4581b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4582b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4583b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4584b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4585bf3a8328SStefano Zampini } 4586bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4587ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4588b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4589ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4590bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4591bf3a8328SStefano Zampini const PetscScalar *array; 4592bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4593bf3a8328SStefano Zampini PetscInt i,n_V; 4594bf3a8328SStefano Zampini 4595b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4596b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4597b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4598b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4599b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4600b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4601b087196eSStefano Zampini PetscScalar val; 4602b087196eSStefano Zampini PetscInt idx; 4603b087196eSStefano Zampini 4604b087196eSStefano Zampini idx = idxs_V[i]; 4605b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4606b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4607b087196eSStefano Zampini } 4608b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4609b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4610bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4611bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4612bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4613bf3a8328SStefano Zampini } 4614ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4615ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4616ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4617ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4618b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4619ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4620bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4621b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4622bf3a8328SStefano Zampini } 4623ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4624ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4625ac632422SStefano Zampini } 4626b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 462788c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4628b96c3477SStefano Zampini } 4629c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4630b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 4631c9db6a07SStefano Zampini PetscInt i; 4632c9db6a07SStefano Zampini 4633c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4634c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4635c9db6a07SStefano Zampini } 4636c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4637c9db6a07SStefano Zampini } 4638b96c3477SStefano Zampini } 463916909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 464016909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 464116909a7fSStefano Zampini } else { 4642906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 464316909a7fSStefano Zampini } 46441dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 464527b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 464672b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 464772b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 464872b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 464972b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 465072b8c272SStefano Zampini } 46511dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 465227b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4653b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4654b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4655906d46d4SStefano Zampini } else { 46561dd7afcfSStefano Zampini Mat benign_global = NULL; 465727b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 46581dd7afcfSStefano Zampini Mat tmat; 46591dd7afcfSStefano Zampini 46601dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 46611dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 46621dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 46631dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46641dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46651dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 46661dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46671dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46681dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 46691dd7afcfSStefano Zampini if (pcbddc->benign_change) { 46701dd7afcfSStefano Zampini Mat M; 46711dd7afcfSStefano Zampini 46721dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 46731dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 46741dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 46751dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4676906d46d4SStefano Zampini } else { 46771dd7afcfSStefano Zampini Mat eye; 46781dd7afcfSStefano Zampini PetscScalar *array; 46791dd7afcfSStefano Zampini 46801dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46811dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 46821dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 46831dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4684906d46d4SStefano Zampini } 46851dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46861dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46871dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46881dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 46891dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 46901dd7afcfSStefano Zampini } 46911dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 46921dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 46931dd7afcfSStefano Zampini } 46941dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 46951dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 46961dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 469727b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 46981dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 46991dd7afcfSStefano Zampini } 47001dd7afcfSStefano Zampini } 470116909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 470216909a7fSStefano Zampini IS is_global; 470316909a7fSStefano Zampini const PetscInt *gidxs; 470416909a7fSStefano Zampini 470516909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 470616909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 470716909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 470816909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 470916909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 471016909a7fSStefano Zampini } 47111dd7afcfSStefano Zampini } 47121dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 47131dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4714b9b85e73SStefano Zampini } 4715a717540cSStefano Zampini 471672b8c272SStefano Zampini if (!pcbddc->fake_change) { 47174f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 47184f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 47194f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 47204f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4721019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4722019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4723019a44ceSStefano Zampini pcbddc->local_primal_size++; 4724019a44ceSStefano Zampini } 4725019a44ceSStefano Zampini 4726019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4727727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4728727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 47299f47a83aSStefano 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); 4730c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 47310e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 47329f47a83aSStefano 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); 4733727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4734727cdba6SStefano Zampini } 47350e6343abSStefano Zampini } 4736727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4737b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 473872b8c272SStefano Zampini } 473972b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4740727cdba6SStefano Zampini 4741a717540cSStefano Zampini /* flush dbg viewer */ 4742b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4743b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4744b8ffe317SStefano Zampini } 4745a717540cSStefano Zampini 4746e310c8b4SStefano Zampini /* free workspace */ 4747a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 47484641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 474908122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 47509162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 47519162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 475208122e43SStefano Zampini } else { 47539162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 47549162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 47559162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 475608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 475708122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 47589162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 47599162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 476008122e43SStefano Zampini } 4761674ae819SStefano Zampini PetscFunctionReturn(0); 4762674ae819SStefano Zampini } 4763674ae819SStefano Zampini 4764674ae819SStefano Zampini #undef __FUNCT__ 4765674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4766674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4767674ae819SStefano Zampini { 4768*71582508SStefano Zampini ISLocalToGlobalMapping map; 4769674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4770674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 477114f95afaSStefano Zampini PetscInt ierr,i,N; 4772674ae819SStefano Zampini 4773674ae819SStefano Zampini PetscFunctionBegin; 4774*71582508SStefano Zampini if (pcbddc->graphanalyzed) PetscFunctionReturn(0); 47758e61c736SStefano Zampini /* Reset previously computed graph */ 47768e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4777674ae819SStefano Zampini /* Init local Graph struct */ 47787fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 4779*71582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 4780*71582508SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N);CHKERRQ(ierr); 4781674ae819SStefano Zampini 4782575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 47839a962809SStefano 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); 47849577ea80SStefano Zampini 4785674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4786d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 47874d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 47884d379d7bSStefano Zampini PetscInt nvtxs; 4789e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4790674ae819SStefano Zampini 47912fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 47922fffb893SStefano Zampini if (flg_row) { 47934d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4794b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 47952fffb893SStefano Zampini } 47962fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4797674ae819SStefano Zampini } 47989b28b941SStefano Zampini if (pcbddc->dbg_flag) { 47999b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4800674ae819SStefano Zampini } 4801674ae819SStefano Zampini 4802674ae819SStefano Zampini /* Setup of Graph */ 48034b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 480414f95afaSStefano 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); 4805674ae819SStefano Zampini 48064f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 48074f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 48084f1b2e48SStefano Zampini PetscInt *local_subs; 48094f1b2e48SStefano Zampini 48104f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 48114f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 48124f1b2e48SStefano Zampini const PetscInt *idxs; 48134f1b2e48SStefano Zampini PetscInt nl,j; 48144f1b2e48SStefano Zampini 48154f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 48164f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 4817*71582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 48184f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 48194f1b2e48SStefano Zampini } 48204f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 48214f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 48224f1b2e48SStefano Zampini } 48234f1b2e48SStefano Zampini 4824674ae819SStefano Zampini /* Graph's connected components analysis */ 4825674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4826*71582508SStefano Zampini 4827*71582508SStefano Zampini /* set flag indicating analysis has been done */ 4828*71582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 4829674ae819SStefano Zampini PetscFunctionReturn(0); 4830674ae819SStefano Zampini } 4831674ae819SStefano Zampini 48329a7d3425SStefano Zampini #undef __FUNCT__ 48339a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 48349a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 48359a7d3425SStefano Zampini { 48369a7d3425SStefano Zampini PetscInt i,j; 48379a7d3425SStefano Zampini PetscScalar *alphas; 48389a7d3425SStefano Zampini PetscErrorCode ierr; 48399a7d3425SStefano Zampini 48409a7d3425SStefano Zampini PetscFunctionBegin; 4841785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 48429a7d3425SStefano Zampini for (i=0;i<n;i++) { 48439a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 4844669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 4845669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 4846669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 48479a7d3425SStefano Zampini } 48489a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 48499a7d3425SStefano Zampini PetscFunctionReturn(0); 48509a7d3425SStefano Zampini } 48519a7d3425SStefano Zampini 4852e7931f94SStefano Zampini #undef __FUNCT__ 485370cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 485457de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4855e7931f94SStefano Zampini { 485657de7509SStefano Zampini Mat A; 4857e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4858e7931f94SStefano Zampini PetscMPIInt size,rank,color; 485952e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 486052e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 486127b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 486257de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 486327b6a85dSStefano Zampini PetscInt xadj_count, *count; 486427b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 486527b6a85dSStefano Zampini PetscSubcomm psubcomm; 486627b6a85dSStefano Zampini MPI_Comm subcomm; 486752e5ac9dSStefano Zampini PetscErrorCode ierr; 4868a57a6d2fSStefano Zampini 4869e7931f94SStefano Zampini PetscFunctionBegin; 487057de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 487157de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 487257de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 487357de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 487457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 487557de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 487657de7509SStefano Zampini 487757de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 487857de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 487957de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 488057de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 488157de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 488257de7509SStefano Zampini im_active = !!(n); 488357de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 488457de7509SStefano Zampini void_procs = size - active_procs; 488557de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 488657de7509SStefano Zampini if (void_procs) { 488757de7509SStefano Zampini PetscInt ncand; 488857de7509SStefano Zampini 488957de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 489057de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 489157de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 489257de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 489357de7509SStefano Zampini if (!procs_candidates[i]) { 489457de7509SStefano Zampini procs_candidates[ncand++] = i; 489557de7509SStefano Zampini } 489657de7509SStefano Zampini } 489757de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 489857de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 489957de7509SStefano Zampini } 490057de7509SStefano Zampini 490157de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 490257de7509SStefano Zampini if (active_procs < *n_subdomains) { 490357de7509SStefano Zampini PetscInt issize,isidx; 490457de7509SStefano Zampini if (im_active) { 490557de7509SStefano Zampini issize = 1; 490657de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 490757de7509SStefano Zampini isidx = procs_candidates[rank]; 490857de7509SStefano Zampini } else { 490957de7509SStefano Zampini isidx = rank; 491057de7509SStefano Zampini } 491157de7509SStefano Zampini } else { 491257de7509SStefano Zampini issize = 0; 491357de7509SStefano Zampini isidx = -1; 491457de7509SStefano Zampini } 491557de7509SStefano Zampini *n_subdomains = active_procs; 491657de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 4917daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 491857de7509SStefano Zampini PetscFunctionReturn(0); 491957de7509SStefano Zampini } 4920c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4921c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 492227b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 4923e7931f94SStefano Zampini 4924e7931f94SStefano Zampini /* Get info on mapping */ 49253bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 49263bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4927e7931f94SStefano Zampini 4928e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4929785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4930e7931f94SStefano Zampini xadj[0] = 0; 4931e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4932785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4933785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 493427b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 493527b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 493627b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 493727b6a85dSStefano Zampini count[shared[i][j]] += 1; 4938e7931f94SStefano Zampini 493927b6a85dSStefano Zampini xadj_count = 0; 49402b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 494127b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 494227b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 4943d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4944d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4945d023bfaeSStefano Zampini xadj_count++; 494627b6a85dSStefano Zampini break; 494727b6a85dSStefano Zampini } 4948e7931f94SStefano Zampini } 4949e7931f94SStefano Zampini } 4950d023bfaeSStefano Zampini xadj[1] = xadj_count; 495127b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 49523bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4953e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4954e7931f94SStefano Zampini 49553837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4956e7931f94SStefano Zampini 495727b6a85dSStefano Zampini /* Restrict work on active processes only */ 495827b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 495927b6a85dSStefano Zampini if (void_procs) { 496027b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 496127b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 496227b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 496327b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 496427b6a85dSStefano Zampini } else { 496527b6a85dSStefano Zampini psubcomm = NULL; 496627b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 496727b6a85dSStefano Zampini } 496827b6a85dSStefano Zampini 496927b6a85dSStefano Zampini v_wgt = NULL; 497027b6a85dSStefano Zampini if (!color) { 4971e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4972e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4973e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4974c8587f34SStefano Zampini } else { 497552e5ac9dSStefano Zampini Mat subdomain_adj; 497652e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 497752e5ac9dSStefano Zampini MatPartitioning partitioner; 497827b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 497952e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 498057de7509SStefano Zampini PetscMPIInt size; 4981b0c7d250SStefano Zampini PetscBool aggregate; 4982b0c7d250SStefano Zampini 498327b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 498427b6a85dSStefano Zampini if (void_procs) { 498527b6a85dSStefano Zampini PetscInt prank = rank; 4986785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 498727b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 4988e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4989e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4990c8587f34SStefano Zampini } 4991e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 499227b6a85dSStefano Zampini } else { 499327b6a85dSStefano Zampini oldranks = NULL; 499427b6a85dSStefano Zampini } 4995b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 499627b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 4997b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4998b0c7d250SStefano Zampini PetscMPIInt nrank; 4999b0c7d250SStefano Zampini PetscScalar *vals; 5000b0c7d250SStefano Zampini 500127b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5002b0c7d250SStefano Zampini lrows = 0; 5003b0c7d250SStefano Zampini if (nrank<redprocs) { 5004b0c7d250SStefano Zampini lrows = size/redprocs; 5005b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5006b0c7d250SStefano Zampini } 500727b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5008b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5009b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5010b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5011b0c7d250SStefano Zampini row = nrank; 5012b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5013b0c7d250SStefano Zampini cols = adjncy; 5014b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5015b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5016b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5017b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5018b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 501952e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 502052e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 502152e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5022b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 502327b6a85dSStefano Zampini if (use_vwgt) { 502427b6a85dSStefano Zampini Vec v; 502527b6a85dSStefano Zampini const PetscScalar *array; 502627b6a85dSStefano Zampini PetscInt nl; 502727b6a85dSStefano Zampini 502827b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 502927b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 503027b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 503127b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 503227b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 503327b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 503427b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 503522db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 503627b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 503727b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 503827b6a85dSStefano Zampini } 5039b0c7d250SStefano Zampini } else { 504027b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 504127b6a85dSStefano Zampini if (use_vwgt) { 504227b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 504327b6a85dSStefano Zampini v_wgt[0] = local_size; 504427b6a85dSStefano Zampini } 5045b0c7d250SStefano Zampini } 504622b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5047e7931f94SStefano Zampini 5048e7931f94SStefano Zampini /* Partition */ 504927b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5050e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 505127b6a85dSStefano Zampini if (v_wgt) { 5052e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5053c8587f34SStefano Zampini } 505457de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 505557de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5056e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5057e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 505822b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5059e7931f94SStefano Zampini 506052e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 50616583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 506252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 506352e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 506457de7509SStefano Zampini if (!aggregate) { 506557de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 506627b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 506727b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 506827b6a85dSStefano Zampini #endif 506957de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 507027b6a85dSStefano Zampini } else if (oldranks) { 5071b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 507227b6a85dSStefano Zampini } else { 507327b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 507457de7509SStefano Zampini } 507528143c3dSStefano Zampini } else { 5076b0c7d250SStefano Zampini PetscInt idxs[1]; 5077b0c7d250SStefano Zampini PetscMPIInt tag; 5078b0c7d250SStefano Zampini MPI_Request *reqs; 5079b0c7d250SStefano Zampini 5080b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5081b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5082b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 508327b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 508428143c3dSStefano Zampini } 508527b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5086b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5087b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 508857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 508927b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 509027b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 509127b6a85dSStefano Zampini #endif 509257de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 509327b6a85dSStefano Zampini } else if (oldranks) { 5094b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 509527b6a85dSStefano Zampini } else { 509627b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5097e7931f94SStefano Zampini } 509857de7509SStefano Zampini } 509952e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5100e7931f94SStefano Zampini /* clean up */ 5101e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 510252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5103e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5104e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5105e7931f94SStefano Zampini } 510627b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 510757de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5108e7931f94SStefano Zampini 5109e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5110e7931f94SStefano Zampini i = 1; 511127b6a85dSStefano Zampini if (!color) i=0; 511257de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5113e7931f94SStefano Zampini PetscFunctionReturn(0); 5114e7931f94SStefano Zampini } 5115e7931f94SStefano Zampini 5116e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5117e7931f94SStefano Zampini 5118e7931f94SStefano Zampini #undef __FUNCT__ 5119e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 51201ae86dd6SStefano 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[]) 5121e7931f94SStefano Zampini { 512270cf5478SStefano Zampini Mat local_mat; 5123e7931f94SStefano Zampini IS is_sends_internal; 51249d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 51251ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 51269d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5127e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5128e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5129e7931f94SStefano Zampini const PetscInt* is_indices; 5130e7931f94SStefano Zampini MatType new_local_type; 5131e7931f94SStefano Zampini /* buffers */ 5132e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 513328143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 51349d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5135e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 51361ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 5137e7931f94SStefano Zampini /* MPI */ 513828143c3dSStefano Zampini MPI_Comm comm,comm_n; 513928143c3dSStefano Zampini PetscSubcomm subcomm; 5140e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 514128143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 514228143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 51431ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 51441ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 51451ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 5146e7931f94SStefano Zampini PetscErrorCode ierr; 5147e7931f94SStefano Zampini 5148e7931f94SStefano Zampini PetscFunctionBegin; 514957de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5150e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 515128143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 515257de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 515357de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 515457de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 515557de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 515657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 51571ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 51581ae86dd6SStefano Zampini if (nvecs) { 51591ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 51601ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 51611ae86dd6SStefano Zampini } 516257de7509SStefano Zampini /* further checks */ 5163e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5164e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5165e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5166e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5167e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 516857de7509SStefano Zampini if (reuse && *mat_n) { 516970cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 517057de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 517170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 517228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 517370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 517470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 517570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 517670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 517770cf5478SStefano Zampini } 5178e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5179e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 518057de7509SStefano Zampini 5181e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5182e7931f94SStefano Zampini if (!is_sends) { 518328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 518457de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5185c8587f34SStefano Zampini } else { 5186e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5187e7931f94SStefano Zampini is_sends_internal = is_sends; 5188c8587f34SStefano Zampini } 5189e7931f94SStefano Zampini 5190e7931f94SStefano Zampini /* get comm */ 5191a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5192e7931f94SStefano Zampini 5193e7931f94SStefano Zampini /* compute number of sends */ 5194e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5195e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5196e7931f94SStefano Zampini 5197e7931f94SStefano Zampini /* compute number of receives */ 5198e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5199785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5200e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5201e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5202e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5203e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5204e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5205e7931f94SStefano Zampini 520628143c3dSStefano Zampini /* restrict comm if requested */ 520728143c3dSStefano Zampini subcomm = 0; 520828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 520928143c3dSStefano Zampini if (restrict_comm) { 5210779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5211779c1cceSStefano Zampini 521228143c3dSStefano Zampini color = 0; 521353a05cb3SStefano Zampini if (restrict_full) { 521453a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 521553a05cb3SStefano Zampini } else { 521653a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 521753a05cb3SStefano Zampini } 5218b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 521928143c3dSStefano Zampini subcommsize = commsize - subcommsize; 522028143c3dSStefano Zampini /* check if reuse has been requested */ 522157de7509SStefano Zampini if (reuse) { 522228143c3dSStefano Zampini if (*mat_n) { 522328143c3dSStefano Zampini PetscMPIInt subcommsize2; 522428143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 522528143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 522628143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 522728143c3dSStefano Zampini } else { 522828143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 522928143c3dSStefano Zampini } 523028143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5231779c1cceSStefano Zampini PetscMPIInt rank; 5232779c1cceSStefano Zampini 5233779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 523428143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 523528143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 523628143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5237306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 523828143c3dSStefano Zampini } 523928143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 524028143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 524128143c3dSStefano Zampini } else { 524228143c3dSStefano Zampini comm_n = comm; 524328143c3dSStefano Zampini } 524428143c3dSStefano Zampini 5245e7931f94SStefano Zampini /* prepare send/receive buffers */ 5246785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5247e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5248785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5249e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 525028143c3dSStefano Zampini if (nis) { 5251854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 525228143c3dSStefano Zampini } 5253e7931f94SStefano Zampini 525428143c3dSStefano Zampini /* Get data from local matrices */ 52556c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5256e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5257e7931f94SStefano Zampini /* 5258e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5259e7931f94SStefano Zampini send_buffer_idxs should contain: 5260e7931f94SStefano Zampini - MatType_PRIVATE type 5261e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5262e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5263e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5264e7931f94SStefano Zampini */ 52656c4ed002SBarry Smith else { 5266e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 52673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5268854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5269e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5270e7931f94SStefano Zampini send_buffer_idxs[1] = i; 52713bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5272e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 52733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5274e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5275e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5276e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5277e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5278c8587f34SStefano Zampini } 5279c8587f34SStefano Zampini } 5280e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 528128143c3dSStefano Zampini /* additional is (if any) */ 528228143c3dSStefano Zampini if (nis) { 528328143c3dSStefano Zampini PetscMPIInt psum; 528428143c3dSStefano Zampini PetscInt j; 528528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 528628143c3dSStefano Zampini PetscInt plen; 528728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 528828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 528928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 529028143c3dSStefano Zampini } 5291854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 529228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 529328143c3dSStefano Zampini PetscInt plen; 529428143c3dSStefano Zampini const PetscInt *is_array_idxs; 529528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 529628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 529728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 529828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 529928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 530028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 530128143c3dSStefano Zampini } 530228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 530328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 530428143c3dSStefano Zampini } 530528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 530628143c3dSStefano Zampini } 530728143c3dSStefano Zampini 5308e7931f94SStefano Zampini buf_size_idxs = 0; 5309e7931f94SStefano Zampini buf_size_vals = 0; 531028143c3dSStefano Zampini buf_size_idxs_is = 0; 53111ae86dd6SStefano Zampini buf_size_vecs = 0; 5312e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5313e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5314e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 531528143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 53161ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 5317e7931f94SStefano Zampini } 5318785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5319785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 532095ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 53211ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 5322e7931f94SStefano Zampini 5323e7931f94SStefano Zampini /* get new tags for clean communications */ 5324e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5325e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 532628143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 53271ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 5328e7931f94SStefano Zampini 5329e7931f94SStefano Zampini /* allocate for requests */ 5330785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5331785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 533295ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 53331ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 5334785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5335785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 533695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 53371ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 5338e7931f94SStefano Zampini 5339e7931f94SStefano Zampini /* communications */ 5340e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5341e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 534228143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 53431ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 5344e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5345e7931f94SStefano Zampini source_dest = onodes[i]; 5346e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5347e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5348e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5349e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 535028143c3dSStefano Zampini if (nis) { 535157de7509SStefano Zampini source_dest = onodes_is[i]; 535228143c3dSStefano 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); 535328143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 535428143c3dSStefano Zampini } 53551ae86dd6SStefano Zampini if (nvecs) { 53561ae86dd6SStefano Zampini source_dest = onodes[i]; 53571ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 53581ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 53591ae86dd6SStefano Zampini } 5360e7931f94SStefano Zampini } 5361e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5362e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5363e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5364e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 536528143c3dSStefano Zampini if (nis) { 536628143c3dSStefano 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); 536728143c3dSStefano Zampini } 53681ae86dd6SStefano Zampini if (nvecs) { 53691ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 53701ae86dd6SStefano 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); 53711ae86dd6SStefano Zampini } 5372e7931f94SStefano Zampini } 5373e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5374e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5375e7931f94SStefano Zampini 5376e7931f94SStefano Zampini /* assemble new l2g map */ 5377e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5378e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53799d30be91SStefano Zampini new_local_rows = 0; 5380e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53819d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5382e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5383e7931f94SStefano Zampini } 53849d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5385e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53869d30be91SStefano Zampini new_local_rows = 0; 5387e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53889d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 53899d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5390e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5391e7931f94SStefano Zampini } 53929d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 53939d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5394e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5395e7931f94SStefano Zampini 5396e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5397e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5398e7931f94SStefano 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) */ 5399e7931f94SStefano Zampini if (n_recvs) { 540028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5401e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5402e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5403e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5404e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5405e7931f94SStefano Zampini break; 5406e7931f94SStefano Zampini } 5407e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5408e7931f94SStefano Zampini } 5409e7931f94SStefano Zampini switch (new_local_type_private) { 541028143c3dSStefano Zampini case MATDENSE_PRIVATE: 541128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5412e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5413e7931f94SStefano Zampini bs = 1; 541428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 541528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 541628143c3dSStefano Zampini bs = 1; 541728143c3dSStefano Zampini } 5418e7931f94SStefano Zampini break; 5419e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5420e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5421e7931f94SStefano Zampini bs = 1; 5422e7931f94SStefano Zampini break; 5423e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5424e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5425e7931f94SStefano Zampini break; 5426e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5427e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5428e7931f94SStefano Zampini break; 5429e7931f94SStefano Zampini default: 54309d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5431e7931f94SStefano Zampini break; 5432e7931f94SStefano Zampini } 543328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 543428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 543528143c3dSStefano Zampini bs = 1; 5436e7931f94SStefano Zampini } 5437e7931f94SStefano Zampini 543870cf5478SStefano Zampini /* create MATIS object if needed */ 543957de7509SStefano Zampini if (!reuse) { 5440e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5441e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 544270cf5478SStefano Zampini } else { 544370cf5478SStefano Zampini /* it also destroys the local matrices */ 544457de7509SStefano Zampini if (*mat_n) { 544570cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 544657de7509SStefano Zampini } else { /* this is a fake object */ 544757de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 544857de7509SStefano Zampini } 544970cf5478SStefano Zampini } 545070cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5451e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 54529d30be91SStefano Zampini 54539d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 54549d30be91SStefano Zampini 54559d30be91SStefano Zampini /* Global to local map of received indices */ 54569d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 54579d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 54589d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 54599d30be91SStefano Zampini 54609d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 54619d30be91SStefano Zampini buf_size_idxs = 0; 54629d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54639d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 54649d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 54659d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 54669d30be91SStefano Zampini } 54679d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 54689d30be91SStefano Zampini 54699d30be91SStefano Zampini /* set preallocation */ 54709d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 54719d30be91SStefano Zampini if (!newisdense) { 54729d30be91SStefano Zampini PetscInt *new_local_nnz=0; 54739d30be91SStefano Zampini 54749d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 54759d30be91SStefano Zampini if (n_recvs) { 54769d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 54779d30be91SStefano Zampini } 54789d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54799d30be91SStefano Zampini PetscInt j; 54809d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 54819d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 54829d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 54839d30be91SStefano Zampini } 54849d30be91SStefano Zampini } else { 54859d30be91SStefano Zampini /* TODO */ 54869d30be91SStefano Zampini } 54879d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 54889d30be91SStefano Zampini } 54899d30be91SStefano Zampini if (new_local_nnz) { 54909d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 54919d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 54929d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 54939d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 54949d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 54959d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 54969d30be91SStefano Zampini } else { 54979d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 54989d30be91SStefano Zampini } 54999d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 55009d30be91SStefano Zampini } else { 55019d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55029d30be91SStefano Zampini } 5503e7931f94SStefano Zampini 5504e7931f94SStefano Zampini /* set values */ 5505e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 55069d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5507e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5508e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5509e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 55109d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5511e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5512e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5513e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 551428143c3dSStefano Zampini } else { 551528143c3dSStefano Zampini /* TODO */ 5516e7931f94SStefano Zampini } 5517e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5518e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5519e7931f94SStefano Zampini } 5520e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5521e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 552270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 552370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55249d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 5525e7931f94SStefano Zampini 5526dfd14d43SStefano Zampini #if 0 552728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5528e7931f94SStefano Zampini Vec lvec,rvec; 5529e7931f94SStefano Zampini PetscReal infty_error; 5530e7931f94SStefano Zampini 55312a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5532e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5533e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5534e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 553570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5536e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5537e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5538e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5539e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5540e7931f94SStefano Zampini } 554128143c3dSStefano Zampini #endif 5542e7931f94SStefano Zampini 554328143c3dSStefano Zampini /* assemble new additional is (if any) */ 554428143c3dSStefano Zampini if (nis) { 554528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 554628143c3dSStefano Zampini 554728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5548854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 554928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 555028143c3dSStefano Zampini psum = 0; 555128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 555228143c3dSStefano Zampini for (j=0;j<nis;j++) { 555328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 555428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 555528143c3dSStefano Zampini psum += plen; 555628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 555728143c3dSStefano Zampini } 555828143c3dSStefano Zampini } 5559854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5560854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 556128143c3dSStefano Zampini for (i=1;i<nis;i++) { 556228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 556328143c3dSStefano Zampini } 556428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 556528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 556628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 556728143c3dSStefano Zampini for (j=0;j<nis;j++) { 556828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 556928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 557028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 557128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 557228143c3dSStefano Zampini } 557328143c3dSStefano Zampini } 557428143c3dSStefano Zampini for (i=0;i<nis;i++) { 557528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 557628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 557728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 557828143c3dSStefano Zampini } 557928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 558028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 558128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 558228143c3dSStefano Zampini } 5583e7931f94SStefano Zampini /* free workspace */ 558428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5585e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5586e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5587e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5588e7931f94SStefano Zampini if (isdense) { 5589e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5590e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5591e7931f94SStefano Zampini } else { 5592e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5593e7931f94SStefano Zampini } 559428143c3dSStefano Zampini if (nis) { 559528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 559628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 559728143c3dSStefano Zampini } 55981ae86dd6SStefano Zampini 55991ae86dd6SStefano Zampini if (nvecs) { 56001ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56011ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56021ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56031ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56041ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 56051ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 56061ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 56071ae86dd6SStefano Zampini /* set values */ 56081ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 56091ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 56101ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56111ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 56121ae86dd6SStefano Zampini PetscInt j; 56131ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 56141ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 56151ae86dd6SStefano Zampini } 56161ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 56171ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 56181ae86dd6SStefano Zampini } 56191ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56201ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 56211ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 56221ae86dd6SStefano Zampini } 56231ae86dd6SStefano Zampini 56241ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 56251ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5626e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5627e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 56281ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 562928143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5630e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5631e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 56321ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 563328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5634e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5635e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5636e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5637e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5638e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 563928143c3dSStefano Zampini if (nis) { 564028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 564128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 564228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 564328143c3dSStefano Zampini } 564428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 564528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 564628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 564728143c3dSStefano Zampini for (i=0;i<nis;i++) { 564828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 564928143c3dSStefano Zampini } 56501ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 56511ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56521ae86dd6SStefano Zampini } 565353a05cb3SStefano Zampini *mat_n = NULL; 565428143c3dSStefano Zampini } 5655e7931f94SStefano Zampini PetscFunctionReturn(0); 5656e7931f94SStefano Zampini } 5657a57a6d2fSStefano Zampini 565812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5659af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 566012edc857SStefano Zampini 5661c8587f34SStefano Zampini #undef __FUNCT__ 5662c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5663c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5664c8587f34SStefano Zampini { 5665c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5666c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 566720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 56681ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 56699881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 567020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 56716e683305SStefano Zampini IS coarse_is,*isarray; 56726e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 567330368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5674f9eb5b7dSStefano Zampini PC pc_temp; 5675c8587f34SStefano Zampini PCType coarse_pc_type; 5676c8587f34SStefano Zampini KSPType coarse_ksp_type; 5677f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 56784f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 56796e683305SStefano Zampini Mat t_coarse_mat_is; 568057de7509SStefano Zampini PetscInt ncoarse; 568168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 568222bc73bbSStefano Zampini PetscScalar *array; 568357de7509SStefano Zampini MatReuse coarse_mat_reuse; 568457de7509SStefano Zampini PetscBool restr, full_restr, have_void; 56859881197aSStefano Zampini PetscErrorCode ierr; 5686fdc09c96SStefano Zampini 5687c8587f34SStefano Zampini PetscFunctionBegin; 5688c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 568968457ee5SStefano 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 */ 5690fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 56915a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5692fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5693f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5694f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5695f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5696fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 569751bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 569851bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5699dc4bcba2SStefano Zampini PC pc; 5700dc4bcba2SStefano Zampini PetscBool isbddc; 5701dc4bcba2SStefano Zampini 5702dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5703dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5704dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5705dc4bcba2SStefano Zampini if (isbddc) { 570663c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 570763c961adSStefano Zampini } else { 5708727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 570963c961adSStefano Zampini } 5710fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5711fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5712fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5713f4ddd8eeSStefano Zampini } 5714fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5715fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5716f4ddd8eeSStefano Zampini } 571770cf5478SStefano Zampini /* reset any subassembling information */ 571857de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 571970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 572057de7509SStefano Zampini } 57216e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5722fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5723f4ddd8eeSStefano Zampini } 572457de7509SStefano Zampini /* assemble coarse matrix */ 572557de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 572657de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 572757de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 572857de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 572918a45a71SStefano Zampini } else { 573057de7509SStefano Zampini coarse_mat = NULL; 573157de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 57326e683305SStefano Zampini } 5733e7931f94SStefano Zampini 5734abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5735abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5736abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5737abbbba34SStefano Zampini 5738abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 573922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 574022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 574122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 574222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5743e176bc59SStefano 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); 57446e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 57456e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57466e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5747abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5748abbbba34SStefano Zampini 574957de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 575057de7509SStefano Zampini im_active = !!(pcis->n); 575157de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 575257de7509SStefano Zampini 575357de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 575457de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 575557de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 575657de7509SStefano Zampini coarse_mat_is = NULL; 575757de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 575857de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 575957de7509SStefano Zampini full_restr = PETSC_TRUE; 57601ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 576157de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 576257de7509SStefano Zampini if (multilevel_requested) { 576357de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 576457de7509SStefano Zampini restr = PETSC_FALSE; 576557de7509SStefano Zampini full_restr = PETSC_FALSE; 576657de7509SStefano Zampini } else { 576757de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 576857de7509SStefano Zampini restr = PETSC_TRUE; 576957de7509SStefano Zampini full_restr = PETSC_TRUE; 577057de7509SStefano Zampini } 57714b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 577257de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 577357de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 5774a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 577557de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 5776a198735bSStefano Zampini } else { 5777a198735bSStefano Zampini PetscMPIInt size,rank; 5778a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 5779a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 5780a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 5781a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5782a198735bSStefano Zampini } 578357de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 578457de7509SStefano Zampini PetscInt psum; 578557de7509SStefano Zampini PetscMPIInt size; 578657de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 578757de7509SStefano Zampini else psum = 0; 578857de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 578957de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 579057de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 579157de7509SStefano Zampini } 579257de7509SStefano Zampini /* determine if we can go multilevel */ 579357de7509SStefano Zampini if (multilevel_requested) { 579457de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 579557de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 579657de7509SStefano Zampini } 579757de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 579857de7509SStefano Zampini 5799e4d548c7SStefano Zampini /* dump subassembling pattern */ 5800e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5801e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5802e4d548c7SStefano Zampini } 5803e4d548c7SStefano Zampini 58046e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 580527b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 58066e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 58076e683305SStefano Zampini const PetscInt *idxs; 58086e683305SStefano Zampini ISLocalToGlobalMapping tmap; 58096e683305SStefano Zampini 58106e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 58110be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 58126e683305SStefano Zampini /* allocate space for temporary storage */ 5813854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5814854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 58156e683305SStefano Zampini /* allocate for IS array */ 58166e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 58176e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 581827b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 581930368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5820854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 58216e683305SStefano Zampini /* dofs splitting */ 58226e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 58236e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 58246e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 58256e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58266e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58276e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58286e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 582930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 58306e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 58316e683305SStefano Zampini } 58326e683305SStefano Zampini /* neumann boundaries */ 58336e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 58346e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 58356e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 58366e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58376e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58386e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58396e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 584030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 58416e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 58426e683305SStefano Zampini } 58436e683305SStefano Zampini /* free memory */ 58446e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 58456e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 58466e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 58476e683305SStefano Zampini } else { 58486e683305SStefano Zampini nis = 0; 58496e683305SStefano Zampini nisdofs = 0; 58506e683305SStefano Zampini nisneu = 0; 585130368db7SStefano Zampini nisvert = 0; 58526e683305SStefano Zampini isarray = NULL; 58536e683305SStefano Zampini } 58546e683305SStefano Zampini /* destroy no longer needed map */ 58556e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 58566e683305SStefano Zampini 585757de7509SStefano Zampini /* subassemble */ 585857de7509SStefano Zampini if (multilevel_allowed) { 58591ae86dd6SStefano Zampini Vec vp[1]; 58601ae86dd6SStefano Zampini PetscInt nvecs = 0; 586157de7509SStefano Zampini PetscBool reuse,reuser; 58621ae86dd6SStefano Zampini 586357de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 586457de7509SStefano Zampini else reuse = PETSC_FALSE; 586557de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 58661ae86dd6SStefano Zampini vp[0] = NULL; 58671ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 58681ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 58691ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 58701ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 58711ae86dd6SStefano Zampini nvecs = 1; 58721ae86dd6SStefano Zampini 58731ae86dd6SStefano Zampini if (pcbddc->divudotp) { 5874a198735bSStefano Zampini Mat B,loc_divudotp; 58751ae86dd6SStefano Zampini Vec v,p; 58761ae86dd6SStefano Zampini IS dummy; 58771ae86dd6SStefano Zampini PetscInt np; 58781ae86dd6SStefano Zampini 5879a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 5880a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 58811ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 5882a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 58831ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 58841ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 58851ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 58861ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 58871ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 58881ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 58891ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 58901ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 58911ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 58921ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 58931ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 58941ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 589574e2c79eSStefano Zampini } 58961ae86dd6SStefano Zampini } 58971ae86dd6SStefano Zampini if (reuser) { 58981ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 589974e2c79eSStefano Zampini } else { 59001ae86dd6SStefano 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); 59011ae86dd6SStefano Zampini } 59021ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 59031ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 59041ae86dd6SStefano Zampini PetscInt nl; 59051ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 59061ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 59071ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59081ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 59091ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 59101ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 59111ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59121ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 5913a198735bSStefano Zampini } else { 5914a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 59151ae86dd6SStefano Zampini } 59161ae86dd6SStefano Zampini } else { 59171ae86dd6SStefano 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); 59186e683305SStefano Zampini } 591957de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 592057de7509SStefano Zampini PetscMPIInt size; 592157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 592257de7509SStefano Zampini if (!multilevel_allowed) { 592357de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 59246e683305SStefano Zampini } else { 592557de7509SStefano Zampini Mat A; 5926779c1cceSStefano Zampini 592757de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 592857de7509SStefano Zampini if (coarse_mat_is) { 592957de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 593057de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 593157de7509SStefano Zampini coarse_mat = coarse_mat_is; 593257de7509SStefano Zampini } 593357de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 593457de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 593557de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5936779c1cceSStefano Zampini } 5937779c1cceSStefano Zampini } 593857de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 593957de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 59406e683305SStefano Zampini 59416e683305SStefano Zampini /* create local to global scatters for coarse problem */ 594268457ee5SStefano Zampini if (compute_vecs) { 59436e683305SStefano Zampini PetscInt lrows; 59446e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 594557de7509SStefano Zampini if (coarse_mat) { 594657de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 59476e683305SStefano Zampini } else { 59486e683305SStefano Zampini lrows = 0; 59496e683305SStefano Zampini } 59506e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 59516e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 59526e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 59536e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59546e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59556e683305SStefano Zampini } 59566e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5957c8587f34SStefano Zampini 5958f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5959f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5960f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5961f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5962f9eb5b7dSStefano Zampini } else { 5963f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5964f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5965c8587f34SStefano Zampini } 5966c8587f34SStefano Zampini 59676e683305SStefano Zampini /* print some info if requested */ 59686e683305SStefano Zampini if (pcbddc->dbg_flag) { 59696e683305SStefano Zampini if (!multilevel_allowed) { 59706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 59716e683305SStefano Zampini if (multilevel_requested) { 59726e683305SStefano 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); 59736e683305SStefano Zampini } else if (pcbddc->max_levels) { 59746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 59756e683305SStefano Zampini } 59766e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59776e683305SStefano Zampini } 59786e683305SStefano Zampini } 59796e683305SStefano Zampini 5980f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 598157de7509SStefano Zampini if (coarse_mat) { 59826a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 59836e683305SStefano Zampini if (pcbddc->dbg_flag) { 598457de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 59856e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 59866e683305SStefano Zampini } 5987f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5988312be037SStefano Zampini char prefix[256],str_level[16]; 5989e604994aSStefano Zampini size_t len; 599057de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5991422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5992c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5993f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 599457de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5995c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 59966e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5997c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5998c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5999e604994aSStefano Zampini /* prefix */ 6000e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 6001e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 6002e604994aSStefano Zampini if (!pcbddc->current_level) { 6003e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 6004e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 6005c8587f34SStefano Zampini } else { 6006e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 6007312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 6008312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 600934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 6010312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 6011e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 6012e604994aSStefano Zampini } 6013e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 60143e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 60153e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 60163e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 60173e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 6018f9eb5b7dSStefano Zampini /* allow user customization */ 6019f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 60203e3c6dadSStefano Zampini } 60213e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 602251bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 60233e3c6dadSStefano Zampini if (nisdofs) { 60243e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 60253e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 60263e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 60273e3c6dadSStefano Zampini } 60283e3c6dadSStefano Zampini } 60293e3c6dadSStefano Zampini if (nisneu) { 60303e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 60313e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 6032312be037SStefano Zampini } 603330368db7SStefano Zampini if (nisvert) { 603430368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 603530368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 603630368db7SStefano Zampini } 6037f9eb5b7dSStefano Zampini 6038f9eb5b7dSStefano Zampini /* get some info after set from options */ 6039f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 6040f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 60414f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 60426e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 6043f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6044f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 6045f9eb5b7dSStefano Zampini } 604639f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 60474f3a063dSStefano Zampini if (isredundant) { 60484f3a063dSStefano Zampini KSP inner_ksp; 60494f3a063dSStefano Zampini PC inner_pc; 60504f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 60514f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 60524f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 60534f3a063dSStefano Zampini } 6054f9eb5b7dSStefano Zampini 605557de7509SStefano Zampini /* parameters which miss an API */ 605657de7509SStefano Zampini if (isbddc) { 6057720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 6058720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 605957de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 606027b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 606127b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 6062a198735bSStefano Zampini Mat coarsedivudotp_is; 6063a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 6064a198735bSStefano Zampini IS row,col; 6065a198735bSStefano Zampini const PetscInt *gidxs; 6066a198735bSStefano Zampini PetscInt n,st,M,N; 6067a198735bSStefano Zampini 6068a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 6069a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 6070a198735bSStefano Zampini st = st-n; 6071a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 6072a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 6073a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 6074a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6075a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 6076a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6077a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 6078a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 6079a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 6080a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 6081a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 6082a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 6083a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 6084a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 6085a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 6086a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 6087a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 6088a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 6089a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 6090a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 60918ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 6092a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 6093720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 609459e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 6095720d30f9SStefano Zampini } 6096d4d8cf7bSStefano Zampini } 60979881197aSStefano Zampini 60983301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 60995a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 61003301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 61013301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 61023301b35fSStefano Zampini } 61033301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 61043301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 61053301b35fSStefano Zampini } 61063301b35fSStefano Zampini if (pc->pmat->spd_set) { 61073301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 61083301b35fSStefano Zampini } 610927b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 611027b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 611127b6a85dSStefano Zampini } 61126e683305SStefano Zampini /* set operators */ 61135f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 61146e683305SStefano Zampini if (pcbddc->dbg_flag) { 61156e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 61166e683305SStefano Zampini } 61176e683305SStefano Zampini } 61186e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6119b1ecc7b1SStefano Zampini #if 0 6120b9b85e73SStefano Zampini { 6121b9b85e73SStefano Zampini PetscViewer viewer; 6122b9b85e73SStefano Zampini char filename[256]; 6123b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6124b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 61256a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6126b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6127f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6128b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6129b9b85e73SStefano Zampini } 6130b9b85e73SStefano Zampini #endif 6131f9eb5b7dSStefano Zampini 613298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 613398a51de6SStefano Zampini Vec crhs,csol; 613404708bb6SStefano Zampini 6135f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6136f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6137f347579bSStefano Zampini if (!csol) { 61382a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6139f9eb5b7dSStefano Zampini } 6140f347579bSStefano Zampini if (!crhs) { 61412a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6142f347579bSStefano Zampini } 6143b0f5fe93SStefano Zampini } 61441ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 6145b0f5fe93SStefano Zampini 6146b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6147b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6148b0f5fe93SStefano Zampini 6149b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 61504f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 61514f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 61524f1b2e48SStefano Zampini } 6153b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6154b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6155b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6156b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6157b0f5fe93SStefano Zampini if (coarse_mat) { 6158b0f5fe93SStefano Zampini Vec nullv; 6159b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6160b0f5fe93SStefano Zampini PetscInt nl; 6161b0f5fe93SStefano Zampini 6162b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6163b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6164b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6165b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6166b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6167b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6168b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6169b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6170b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6171b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6172b0f5fe93SStefano Zampini } 6173b0f5fe93SStefano Zampini } 6174b0f5fe93SStefano Zampini 6175b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6176b0f5fe93SStefano Zampini PetscBool ispreonly; 6177b0f5fe93SStefano Zampini 6178b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6179b0f5fe93SStefano Zampini PetscBool isnull; 6180b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6181bef83e63SStefano Zampini if (isnull) { 6182b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6183b0f5fe93SStefano Zampini } 6184bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6185b0f5fe93SStefano Zampini } 6186b0f5fe93SStefano Zampini /* setup coarse ksp */ 6187b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6188cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6189cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 61906e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6191c8587f34SStefano Zampini KSP check_ksp; 61922b510759SStefano Zampini KSPType check_ksp_type; 6193c8587f34SStefano Zampini PC check_pc; 61946e683305SStefano Zampini Vec check_vec,coarse_vec; 61956a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 61962b510759SStefano Zampini PetscInt its; 61976e683305SStefano Zampini PetscBool compute_eigs; 61986e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 61996e683305SStefano Zampini PetscInt neigs; 62008e185a42SStefano Zampini const char *prefix; 6201c8587f34SStefano Zampini 62022b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 62036e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6204422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 620523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6206f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6207e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6208e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6209e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 62102b510759SStefano Zampini if (ispreonly) { 62112b510759SStefano Zampini check_ksp_type = KSPPREONLY; 62126e683305SStefano Zampini compute_eigs = PETSC_FALSE; 62132b510759SStefano Zampini } else { 6214cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 62156e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6216c8587f34SStefano Zampini } 6217c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 62186e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 62196e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 62206e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6221a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6222a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6223a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6224a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6225c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6226c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6227c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6228c8587f34SStefano Zampini /* create random vec */ 62292701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6230c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 62316e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6232c8587f34SStefano Zampini /* solve coarse problem */ 62336e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6234cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 62356e683305SStefano Zampini if (compute_eigs) { 6236854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6237854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 62386e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 62391ae86dd6SStefano Zampini if (neigs) { 62406e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 62416e683305SStefano Zampini lambda_min = eigs_r[0]; 62426e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 62432701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 62442701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6245cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6246cbcc2c2aSStefano Zampini } 6247c8587f34SStefano Zampini } 6248c8587f34SStefano Zampini } 62491ae86dd6SStefano Zampini } 6250cbcc2c2aSStefano Zampini 6251c8587f34SStefano Zampini /* check coarse problem residual error */ 62526e683305SStefano Zampini if (pcbddc->dbg_flag) { 62536e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 62546e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62556e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6256c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 62576e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 62586e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6259779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 62606e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 62616e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 62626e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 62636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6264b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6265b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6266b0f5fe93SStefano Zampini } 62676e683305SStefano Zampini if (compute_eigs) { 62686e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6269deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6270c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 62716e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 62726e683305SStefano 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); 62736e683305SStefano Zampini for (i=0;i<neigs;i++) { 62746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6275c8587f34SStefano Zampini } 62766e683305SStefano Zampini } 62776e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 62786e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62796e683305SStefano Zampini } 6280e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 62812701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6282c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 62836e683305SStefano Zampini if (compute_eigs) { 62846e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 62856e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6286c8587f34SStefano Zampini } 62876e683305SStefano Zampini } 62886e683305SStefano Zampini } 6289bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6290cbcc2c2aSStefano Zampini /* print additional info */ 6291cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 62926e683305SStefano Zampini /* waits until all processes reaches this point */ 62936e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6294cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6295cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6296cbcc2c2aSStefano Zampini } 6297cbcc2c2aSStefano Zampini 62982b510759SStefano Zampini /* free memory */ 6299fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6300c8587f34SStefano Zampini PetscFunctionReturn(0); 6301c8587f34SStefano Zampini } 6302674ae819SStefano Zampini 6303f34684f1SStefano Zampini #undef __FUNCT__ 6304f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6305f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6306f34684f1SStefano Zampini { 6307f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6308f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6309f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6310dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6311dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 631273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6313dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6314f34684f1SStefano Zampini PetscErrorCode ierr; 6315f34684f1SStefano Zampini 6316f34684f1SStefano Zampini PetscFunctionBegin; 6317f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 63186c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6319dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 63203bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6321dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6322dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 63236583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6324dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6325dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6326dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 63276c4ed002SBarry 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); 6328dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6329dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6330dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6331dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6332dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6333f34684f1SStefano Zampini 6334f34684f1SStefano Zampini /* check numbering */ 6335f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6336019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6337dc456d91SStefano Zampini PetscInt i; 6338b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6339f34684f1SStefano Zampini 6340f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6341f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6342f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 63431575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6344019a44ceSStefano Zampini /* counter */ 6345019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6346019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6347019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6348019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6349019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6350019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6351f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6352f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6353727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6354f34684f1SStefano Zampini } 6355f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6356f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6357f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6358e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6359e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6360e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6361e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6362f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6363019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6364f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6365019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 63662c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 636775c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6368b9b85e73SStefano Zampini set_error = PETSC_TRUE; 63692c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 63702c66d082SStefano 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); 6371f34684f1SStefano Zampini } 6372f34684f1SStefano Zampini } 6373019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6374b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6375f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6376f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6377f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6378f34684f1SStefano Zampini } 6379f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6380f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6381e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6382e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6383f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6384f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6385b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6386ca8b9ea9SStefano Zampini PetscInt *gidxs; 6387ca8b9ea9SStefano Zampini 6388ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 63893bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6390f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6391f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6392f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6393f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 63944bc2dc4bSStefano 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); 6395f34684f1SStefano Zampini } 6396f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6397ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6398f34684f1SStefano Zampini } 6399f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 64001575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6401302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6402f34684f1SStefano Zampini } 64038bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6404f34684f1SStefano Zampini /* get back data */ 6405f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6406f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6407674ae819SStefano Zampini PetscFunctionReturn(0); 6408674ae819SStefano Zampini } 6409674ae819SStefano Zampini 6410e456f2a8SStefano Zampini #undef __FUNCT__ 6411e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6412a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6413e456f2a8SStefano Zampini { 6414e456f2a8SStefano Zampini IS localis_t; 6415a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6416e456f2a8SStefano Zampini PetscScalar *vals; 6417e456f2a8SStefano Zampini PetscErrorCode ierr; 6418e456f2a8SStefano Zampini 6419e456f2a8SStefano Zampini PetscFunctionBegin; 6420a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6421e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6422854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6423e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6424e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6425a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6426a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 64271035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6428a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 64291035eff8SStefano Zampini } 6430a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6431e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6432e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6433a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6434a7dc3881SStefano Zampini /* now compute set in local ordering */ 6435a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6436a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6437a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6438a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6439a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6440ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6441e456f2a8SStefano Zampini lsize++; 6442e456f2a8SStefano Zampini } 6443e456f2a8SStefano Zampini } 6444854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6445a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6446ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6447e456f2a8SStefano Zampini idxs[lsize++] = i; 6448e456f2a8SStefano Zampini } 6449e456f2a8SStefano Zampini } 6450a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6451a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6452e456f2a8SStefano Zampini *localis = localis_t; 6453e456f2a8SStefano Zampini PetscFunctionReturn(0); 6454e456f2a8SStefano Zampini } 6455906d46d4SStefano Zampini 6456b96c3477SStefano Zampini #undef __FUNCT__ 6457b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 645808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6459b96c3477SStefano Zampini { 6460a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6461b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6462b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6463a64f4aa4SStefano Zampini Mat S_j; 6464b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6465b96c3477SStefano Zampini PetscBool free_used_adj; 6466b96c3477SStefano Zampini PetscErrorCode ierr; 6467b96c3477SStefano Zampini 6468b96c3477SStefano Zampini PetscFunctionBegin; 6469b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6470b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 647108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6472b96c3477SStefano Zampini used_xadj = NULL; 6473b96c3477SStefano Zampini used_adjncy = NULL; 6474b96c3477SStefano Zampini } else { 647508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 647608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 647708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 647808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6479b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6480b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6481b96c3477SStefano Zampini } else { 64822fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6483b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6484b96c3477SStefano Zampini PetscInt nvtxs; 6485b96c3477SStefano Zampini 64862fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 64872fffb893SStefano Zampini if (flg_row) { 6488b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6489b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6490b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6491b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 64922fffb893SStefano Zampini } else { 64932fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 64942fffb893SStefano Zampini used_xadj = NULL; 64952fffb893SStefano Zampini used_adjncy = NULL; 64962fffb893SStefano Zampini } 64972fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6498b96c3477SStefano Zampini } 6499b96c3477SStefano Zampini } 6500d5574798SStefano Zampini 6501d5574798SStefano Zampini /* setup sub_schurs data */ 6502a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6503df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6504df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6505a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 650691af6908SStefano 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); 6507a64f4aa4SStefano Zampini } else { 65086816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 65094d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 6510a3df083aSStefano Zampini PetscInt benign_n; 651172b8c272SStefano Zampini Mat change = NULL; 65129d54b7f4SStefano Zampini Vec scaling = NULL; 651372b8c272SStefano Zampini IS change_primal = NULL; 6514a3df083aSStefano Zampini 65155feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 65165feab87aSStefano Zampini PetscInt n_vertices; 65175feab87aSStefano Zampini 65185feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 65192034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 65205feab87aSStefano Zampini } 652104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 652204708bb6SStefano Zampini if (!isseqaij) { 652304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 652404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 652504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 652604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 652704708bb6SStefano Zampini } else { 6528511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 652904708bb6SStefano Zampini } 653004708bb6SStefano Zampini } 6531a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6532a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6533ca92afb2SStefano Zampini } else { 6534a3df083aSStefano Zampini benign_n = 0; 6535ca92afb2SStefano Zampini } 6536b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6537b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6538b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 653972b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 654022db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6541b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 654222db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6543b7ab4a40SStefano Zampini } 6544b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6545b7ab4a40SStefano 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 */ 6546b7ab4a40SStefano Zampini if (need_change) { 654788c03ad3SStefano Zampini PC_IS *pcisf; 654888c03ad3SStefano Zampini PC_BDDC *pcbddcf; 654988c03ad3SStefano Zampini PC pcf; 655088c03ad3SStefano Zampini 6551e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 655288c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 655388c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 655488c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 655588c03ad3SStefano Zampini /* hacks */ 655688c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 655772b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 655872b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 655972b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 656072b8c272SStefano Zampini pcisf->n = pcis->n; 656172b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 656288c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 656388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 656488c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 656588c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 656688c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 656788c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 656872b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 656988c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 657088c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 657172b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 657272b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 657372b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 657472b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 657572b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 657688c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 657772b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 657888c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 657988c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 658088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 658188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 658288c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 658388c03ad3SStefano Zampini pcf->ops->destroy = NULL; 658488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 658588c03ad3SStefano Zampini } 65869d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 658791af6908SStefano 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); 658872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 658972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6590ca92afb2SStefano Zampini } 6591d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6592b96c3477SStefano Zampini 6593b96c3477SStefano Zampini /* free adjacency */ 6594b96c3477SStefano Zampini if (free_used_adj) { 6595b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6596b96c3477SStefano Zampini } 6597b96c3477SStefano Zampini PetscFunctionReturn(0); 6598b96c3477SStefano Zampini } 6599b96c3477SStefano Zampini 6600b96c3477SStefano Zampini #undef __FUNCT__ 6601b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 660208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6603b96c3477SStefano Zampini { 6604b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6605b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6606b96c3477SStefano Zampini PCBDDCGraph graph; 6607b96c3477SStefano Zampini PetscErrorCode ierr; 6608b96c3477SStefano Zampini 6609b96c3477SStefano Zampini PetscFunctionBegin; 6610b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 661108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 66123301b35fSStefano Zampini IS verticesIS,verticescomm; 66133301b35fSStefano Zampini PetscInt vsize,*idxs; 6614b96c3477SStefano Zampini 6615b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 66163301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 66173301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66183301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 66193301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66203301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6621b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 66227fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6623441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 66243301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6625b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6626b96c3477SStefano Zampini } else { 6627b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6628b96c3477SStefano Zampini } 6629e4d548c7SStefano Zampini /* print some info */ 6630e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6631e4d548c7SStefano Zampini IS vertices; 6632e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 66338ae0ca82SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag-1,pcbddc->dbg_viewer);CHKERRQ(ierr); 6634e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6635e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6636e4d548c7SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 6637e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6638e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6639e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6640e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6641e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6642e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6643e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6644e4d548c7SStefano Zampini } 6645b96c3477SStefano Zampini 6646b96c3477SStefano Zampini /* sub_schurs init */ 6647b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 6648b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 6649b334f244SStefano Zampini } 6650b334f244SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6651a64f4aa4SStefano Zampini 6652b96c3477SStefano Zampini /* free graph struct */ 665308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6654b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6655b96c3477SStefano Zampini } 6656b96c3477SStefano Zampini PetscFunctionReturn(0); 6657b96c3477SStefano Zampini } 6658fa34dd3eSStefano Zampini 6659fa34dd3eSStefano Zampini #undef __FUNCT__ 6660fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6661fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6662fa34dd3eSStefano Zampini { 6663fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6664fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6665fa34dd3eSStefano Zampini PetscErrorCode ierr; 6666fa34dd3eSStefano Zampini 6667fa34dd3eSStefano Zampini PetscFunctionBegin; 6668fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6669fa34dd3eSStefano Zampini IS zerodiag = NULL; 66704f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6671fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 66724f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 667375c01103SStefano Zampini PetscReal norm; 6674fa34dd3eSStefano Zampini PetscInt i; 6675fa34dd3eSStefano Zampini 6676fa34dd3eSStefano Zampini /* B0 and B0_B */ 6677fa34dd3eSStefano Zampini if (zerodiag) { 6678fa34dd3eSStefano Zampini IS dummy; 6679fa34dd3eSStefano Zampini 66804f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 66814f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6682fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6683fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6684fa34dd3eSStefano Zampini } 6685fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6686fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6687fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6688fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6689fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6690fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6691fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6692fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6693fa34dd3eSStefano Zampini /* S_j */ 6694fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6695fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6696fa34dd3eSStefano Zampini 6697fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6698fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6699fa34dd3eSStefano Zampini /* continuous in primal space */ 6700fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6701fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6702fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6703fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 67044f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 67054f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6706fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6707fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6708fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6709fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6710fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6711fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6712fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6713fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6714fa34dd3eSStefano Zampini 6715fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6716fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6717fa34dd3eSStefano Zampini /* local with Schur */ 6718fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6719fa34dd3eSStefano Zampini if (zerodiag) { 6720fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6722fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6723fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6724fa34dd3eSStefano Zampini } 6725fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6726fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6727fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6728fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6729fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6730fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6731fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6732fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6733fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6734fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6735fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6736fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6737fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6738fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6739fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6740fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6741fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6742fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6743fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6744fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6745fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6746fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6747fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6748fa34dd3eSStefano Zampini if (zerodiag) { 6749fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6750fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67514f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6752fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6753fa34dd3eSStefano Zampini } 6754fa34dd3eSStefano Zampini /* BDDC */ 6755fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6756fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6757fa34dd3eSStefano Zampini 6758fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6759fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6760fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6761fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 67624f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 67634f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6764fa34dd3eSStefano Zampini } 67654f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6766fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6767fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6768fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6769fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6770fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6771fa34dd3eSStefano Zampini } 6772fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6773fa34dd3eSStefano Zampini } 6774