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 } 120669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 121669cc0f4SStefano Zampini const PetscInt *idxs; 122669cc0f4SStefano Zampini PetscInt idx,nn,j; 123669cc0f4SStefano Zampini 124669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 125669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 126669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 1271ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 128669cc0f4SStefano Zampini idx = -(idx+1); 129669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 130669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 131669cc0f4SStefano Zampini } 132*c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 133fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 134fa23a32eSStefano Zampini if (vl2l) { 135fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 136fa23a32eSStefano Zampini } 137669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 138669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 139669cc0f4SStefano Zampini 140669cc0f4SStefano Zampini /* assemble near null space */ 141669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 142669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 143669cc0f4SStefano Zampini } 144669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 145669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 146669cc0f4SStefano Zampini } 147669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 148669cc0f4SStefano Zampini PetscFunctionReturn(0); 149669cc0f4SStefano Zampini } 150669cc0f4SStefano Zampini 151669cc0f4SStefano Zampini 152a3df083aSStefano Zampini #undef __FUNCT__ 1531f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 1541f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 1551f4df5f7SStefano Zampini { 1561f4df5f7SStefano Zampini PetscErrorCode ierr; 1571f4df5f7SStefano Zampini Vec local,global; 1581f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1591f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1601f4df5f7SStefano Zampini 1611f4df5f7SStefano Zampini PetscFunctionBegin; 1621f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 1631f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 1641f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 1651f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 1661f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 1671f4df5f7SStefano Zampini PetscInt i; 1681f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1691f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 1701f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1711f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 1721f4df5f7SStefano Zampini } 1731f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 1741f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 1751f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 1761f4df5f7SStefano Zampini } 1771f4df5f7SStefano Zampini } else { 178986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 1791f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 180986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 181986cdee1SStefano Zampini if (i > 1) { 182986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 1831f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 1841f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 1851f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 1861f4df5f7SStefano Zampini } 1871f4df5f7SStefano Zampini } 1881f4df5f7SStefano Zampini } 189986cdee1SStefano Zampini } 1901f4df5f7SStefano Zampini 1911f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 1921f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 1931f4df5f7SStefano Zampini } 1941f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 1951f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1961f4df5f7SStefano Zampini } 1971f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 1981f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1991f4df5f7SStefano Zampini } 2001f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 2011f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 2021f4df5f7SStefano Zampini PetscFunctionReturn(0); 2031f4df5f7SStefano Zampini } 2041f4df5f7SStefano Zampini 2051f4df5f7SStefano Zampini #undef __FUNCT__ 2063e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 2073e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 2083e589ea0SStefano Zampini { 2093e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 2103e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 2113e589ea0SStefano Zampini PetscErrorCode ierr; 2123e589ea0SStefano Zampini 2133e589ea0SStefano Zampini PetscFunctionBegin; 2143e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 2153e589ea0SStefano Zampini PetscFunctionReturn(0); 2163e589ea0SStefano Zampini } 2173e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2183e589ea0SStefano Zampini Vec swap; 2193e589ea0SStefano Zampini 2203e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 2213e589ea0SStefano Zampini swap = pcbddc->work_change; 2223e589ea0SStefano Zampini pcbddc->work_change = r; 2233e589ea0SStefano Zampini r = swap; 2243e589ea0SStefano Zampini } 2253e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2263e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2273e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 2283e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 2293e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2303e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2313e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2323e589ea0SStefano Zampini Vec swap; 2333e589ea0SStefano Zampini 2343e589ea0SStefano Zampini swap = r; 2353e589ea0SStefano Zampini r = pcbddc->work_change; 2363e589ea0SStefano Zampini pcbddc->work_change = swap; 2373e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 2383e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 2393e589ea0SStefano Zampini } 2403e589ea0SStefano Zampini PetscFunctionReturn(0); 2413e589ea0SStefano Zampini } 2423e589ea0SStefano Zampini 2433e589ea0SStefano Zampini #undef __FUNCT__ 244a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 245a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 246a3df083aSStefano Zampini { 247a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 248a3df083aSStefano Zampini PetscErrorCode ierr; 249a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 250a3df083aSStefano Zampini 251a3df083aSStefano Zampini PetscFunctionBegin; 252a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 253a3df083aSStefano Zampini if (transpose) { 254a3df083aSStefano Zampini apply_right = ctx->apply_left; 255a3df083aSStefano Zampini apply_left = ctx->apply_right; 256a3df083aSStefano Zampini } else { 257a3df083aSStefano Zampini apply_right = ctx->apply_right; 258a3df083aSStefano Zampini apply_left = ctx->apply_left; 259a3df083aSStefano Zampini } 260a3df083aSStefano Zampini reset_x = PETSC_FALSE; 261a3df083aSStefano Zampini if (apply_right) { 262a3df083aSStefano Zampini const PetscScalar *ax; 263a3df083aSStefano Zampini PetscInt nl,i; 264a3df083aSStefano Zampini 265a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 266a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 267a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 268a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 269a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 270a3df083aSStefano Zampini PetscScalar sum,val; 271a3df083aSStefano Zampini const PetscInt *idxs; 272a3df083aSStefano Zampini PetscInt nz,j; 273a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 274a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 275a3df083aSStefano Zampini sum = 0.; 276a3df083aSStefano Zampini if (ctx->apply_p0) { 277a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 278a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 279a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 280a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 281a3df083aSStefano Zampini } 282a3df083aSStefano Zampini } else { 283a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 284a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 285a3df083aSStefano Zampini } 286a3df083aSStefano Zampini } 287a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 288a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 289a3df083aSStefano Zampini } 290a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 291a3df083aSStefano Zampini reset_x = PETSC_TRUE; 292a3df083aSStefano Zampini } 293a3df083aSStefano Zampini if (transpose) { 294a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 295a3df083aSStefano Zampini } else { 296a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 297a3df083aSStefano Zampini } 298a3df083aSStefano Zampini if (reset_x) { 299a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 300a3df083aSStefano Zampini } 301a3df083aSStefano Zampini if (apply_left) { 302a3df083aSStefano Zampini PetscScalar *ay; 303a3df083aSStefano Zampini PetscInt i; 304a3df083aSStefano Zampini 305a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 306a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 307a3df083aSStefano Zampini PetscScalar sum,val; 308a3df083aSStefano Zampini const PetscInt *idxs; 309a3df083aSStefano Zampini PetscInt nz,j; 310a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 311a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 312a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 313a3df083aSStefano Zampini if (ctx->apply_p0) { 314a3df083aSStefano Zampini sum = 0.; 315a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 316a3df083aSStefano Zampini sum += ay[idxs[j]]; 317a3df083aSStefano Zampini ay[idxs[j]] += val; 318a3df083aSStefano Zampini } 319a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 320a3df083aSStefano Zampini } else { 321a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 322a3df083aSStefano Zampini ay[idxs[j]] += val; 323a3df083aSStefano Zampini } 324a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 325a3df083aSStefano Zampini } 326a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 327a3df083aSStefano Zampini } 328a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 329a3df083aSStefano Zampini } 330a3df083aSStefano Zampini PetscFunctionReturn(0); 331a3df083aSStefano Zampini } 332a3df083aSStefano Zampini 333a3df083aSStefano Zampini #undef __FUNCT__ 334a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 335a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 336a3df083aSStefano Zampini { 337a3df083aSStefano Zampini PetscErrorCode ierr; 338a3df083aSStefano Zampini 339a3df083aSStefano Zampini PetscFunctionBegin; 340a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 341a3df083aSStefano Zampini PetscFunctionReturn(0); 342a3df083aSStefano Zampini } 343a3df083aSStefano Zampini 344a3df083aSStefano Zampini #undef __FUNCT__ 345a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 346a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 347a3df083aSStefano Zampini { 348a3df083aSStefano Zampini PetscErrorCode ierr; 349a3df083aSStefano Zampini 350a3df083aSStefano Zampini PetscFunctionBegin; 351a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 352a3df083aSStefano Zampini PetscFunctionReturn(0); 353a3df083aSStefano Zampini } 354a3df083aSStefano Zampini 355a3df083aSStefano Zampini #undef __FUNCT__ 356a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 357a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 358a3df083aSStefano Zampini { 359a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 360a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 361a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 362a3df083aSStefano Zampini PetscErrorCode ierr; 363a3df083aSStefano Zampini 364a3df083aSStefano Zampini PetscFunctionBegin; 365a3df083aSStefano Zampini if (!restore) { 3661dd7afcfSStefano Zampini Mat A_IB,A_BI; 367a3df083aSStefano Zampini PetscScalar *work; 368b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 369a3df083aSStefano Zampini 3709a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 3719a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 372a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 373a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 374a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 375a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 376a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 377a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 378a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 379a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 380a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 381a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 382a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 383a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 384059032f7SStefano Zampini if (reuse) { 385a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 3861dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 387059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 388059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 389059032f7SStefano Zampini PetscInt i; 390059032f7SStefano Zampini 391059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 392059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 393059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 394059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 395059032f7SStefano Zampini } 396059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 3971dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 398059032f7SStefano Zampini } 399a3df083aSStefano Zampini ctx->A = pcis->A_IB; 400a3df083aSStefano Zampini ctx->work = work; 401a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 402a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 403a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 404a3df083aSStefano Zampini pcis->A_IB = A_IB; 405a3df083aSStefano Zampini 406a3df083aSStefano Zampini /* A_BI as A_IB^T */ 407a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 408a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 409a3df083aSStefano Zampini pcis->A_BI = A_BI; 410a3df083aSStefano Zampini } else { 4111dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 4121dd7afcfSStefano Zampini PetscFunctionReturn(0); 4131dd7afcfSStefano Zampini } 414a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 415a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 416a3df083aSStefano Zampini pcis->A_IB = ctx->A; 4171dd7afcfSStefano Zampini ctx->A = NULL; 4181dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 4191dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 4201dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 4211dd7afcfSStefano Zampini if (ctx->free) { 422059032f7SStefano Zampini PetscInt i; 4231dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 424059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 425059032f7SStefano Zampini } 426059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 427059032f7SStefano Zampini } 428a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 429a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 430a3df083aSStefano Zampini } 431a3df083aSStefano Zampini PetscFunctionReturn(0); 432a3df083aSStefano Zampini } 433a3df083aSStefano Zampini 434a3df083aSStefano Zampini /* used just in bddc debug mode */ 435a3df083aSStefano Zampini #undef __FUNCT__ 436a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 437a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 438a3df083aSStefano Zampini { 439a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 440a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 441a3df083aSStefano Zampini Mat An; 442a3df083aSStefano Zampini PetscErrorCode ierr; 443a3df083aSStefano Zampini 444a3df083aSStefano Zampini PetscFunctionBegin; 445a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 446a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 447a3df083aSStefano Zampini if (is1) { 448a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 449a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 450a3df083aSStefano Zampini } else { 451a3df083aSStefano Zampini *B = An; 452a3df083aSStefano Zampini } 453a3df083aSStefano Zampini PetscFunctionReturn(0); 454a3df083aSStefano Zampini } 455a3df083aSStefano Zampini 4561cf9b237SStefano Zampini /* TODO: add reuse flag */ 4571cf9b237SStefano Zampini #undef __FUNCT__ 4581cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 4591cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 4601cf9b237SStefano Zampini { 4611cf9b237SStefano Zampini Mat Bt; 4621cf9b237SStefano Zampini PetscScalar *a,*bdata; 4631cf9b237SStefano Zampini const PetscInt *ii,*ij; 4641cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 4651cf9b237SStefano Zampini PetscBool flg_row; 4661cf9b237SStefano Zampini PetscErrorCode ierr; 4671cf9b237SStefano Zampini 4681cf9b237SStefano Zampini PetscFunctionBegin; 4691cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 4701cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4711cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 4721cf9b237SStefano Zampini nnz = n; 4731cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 4741cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 4751cf9b237SStefano Zampini } 4761cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 4771cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 4781cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 4791cf9b237SStefano Zampini nnz = 0; 4801cf9b237SStefano Zampini bii[0] = 0; 4811cf9b237SStefano Zampini for (i=0;i<n;i++) { 4821cf9b237SStefano Zampini PetscInt j; 4831cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 4841cf9b237SStefano Zampini PetscScalar entry = a[j]; 4851cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 4861cf9b237SStefano Zampini bij[nnz] = ij[j]; 4871cf9b237SStefano Zampini bdata[nnz] = entry; 4881cf9b237SStefano Zampini nnz++; 4891cf9b237SStefano Zampini } 4901cf9b237SStefano Zampini } 4911cf9b237SStefano Zampini bii[i+1] = nnz; 4921cf9b237SStefano Zampini } 4931cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 4941cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 4951cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 4961cf9b237SStefano Zampini { 4971cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 4981cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 4991cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 5001cf9b237SStefano Zampini } 5011cf9b237SStefano Zampini *B = Bt; 5021cf9b237SStefano Zampini PetscFunctionReturn(0); 5031cf9b237SStefano Zampini } 5041cf9b237SStefano Zampini 505674ae819SStefano Zampini #undef __FUNCT__ 5064f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 5074f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 5084f1b2e48SStefano Zampini { 5094f1b2e48SStefano Zampini Mat B; 5104f1b2e48SStefano Zampini IS is_dummy,*cc_n; 5114f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 5124f1b2e48SStefano Zampini PCBDDCGraph graph; 5134f1b2e48SStefano Zampini PetscInt i,n; 5144f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 5154f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 5164f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 5174f1b2e48SStefano Zampini PetscErrorCode ierr; 5184f1b2e48SStefano Zampini 5194f1b2e48SStefano Zampini PetscFunctionBegin; 52063c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 52163c961adSStefano Zampini *ncc = 0; 52263c961adSStefano Zampini *cc = NULL; 52363c961adSStefano Zampini PetscFunctionReturn(0); 52463c961adSStefano Zampini } 5254f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 5264f1b2e48SStefano Zampini if (!isseqaij && filter) { 5271cf9b237SStefano Zampini PetscBool isseqdense; 5281cf9b237SStefano Zampini 5291cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 5301cf9b237SStefano Zampini if (!isseqdense) { 5314f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 5321cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 5331cf9b237SStefano Zampini PetscScalar *array; 5341cf9b237SStefano Zampini PetscReal chop=1.e-6; 5351cf9b237SStefano Zampini 5361cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 5371cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 5381cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 5391cf9b237SStefano Zampini for (i=0;i<n;i++) { 5401cf9b237SStefano Zampini PetscInt j; 5411cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 5421cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 5431cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 5441cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 5451cf9b237SStefano Zampini } 5461cf9b237SStefano Zampini } 5471cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 5489d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 5491cf9b237SStefano Zampini } 5504f1b2e48SStefano Zampini } else { 5514f1b2e48SStefano Zampini B = A; 5524f1b2e48SStefano Zampini } 5534f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5544f1b2e48SStefano Zampini 5554f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 5564f1b2e48SStefano Zampini if (filter) { 5574f1b2e48SStefano Zampini PetscScalar *data; 5584f1b2e48SStefano Zampini PetscInt j,cum; 5594f1b2e48SStefano Zampini 5604f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 5614f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini cum = 0; 5634f1b2e48SStefano Zampini for (i=0;i<n;i++) { 5644f1b2e48SStefano Zampini PetscInt t; 5654f1b2e48SStefano Zampini 5664f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 5674f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 5684f1b2e48SStefano Zampini continue; 5694f1b2e48SStefano Zampini } 5704f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 5714f1b2e48SStefano Zampini } 5724f1b2e48SStefano Zampini t = xadj_filtered[i]; 5734f1b2e48SStefano Zampini xadj_filtered[i] = cum; 5744f1b2e48SStefano Zampini cum += t; 5754f1b2e48SStefano Zampini } 5764f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 5774f1b2e48SStefano Zampini } else { 5784f1b2e48SStefano Zampini xadj_filtered = NULL; 5794f1b2e48SStefano Zampini adjncy_filtered = NULL; 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini 5824f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 5834f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 5844f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 5854f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 5864f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 5874f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 5894f1b2e48SStefano Zampini if (xadj_filtered) { 5904f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 5914f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 5924f1b2e48SStefano Zampini } else { 5934f1b2e48SStefano Zampini graph->xadj = xadj; 5944f1b2e48SStefano Zampini graph->adjncy = adjncy; 5954f1b2e48SStefano Zampini } 5964f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 5974f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5984f1b2e48SStefano Zampini /* partial clean up */ 5994f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 6004f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 6011cf9b237SStefano Zampini if (A != B) { 6024f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini } 6044f1b2e48SStefano Zampini 6054f1b2e48SStefano Zampini /* get back data */ 6061cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 6071cf9b237SStefano Zampini if (cc) { 6084f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 6094f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 6104f1b2e48SStefano 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); 6114f1b2e48SStefano Zampini } 6124f1b2e48SStefano Zampini *cc = cc_n; 6131cf9b237SStefano Zampini } 6144f1b2e48SStefano Zampini /* clean up graph */ 6154f1b2e48SStefano Zampini graph->xadj = 0; 6164f1b2e48SStefano Zampini graph->adjncy = 0; 6174f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6184f1b2e48SStefano Zampini PetscFunctionReturn(0); 6194f1b2e48SStefano Zampini } 6204f1b2e48SStefano Zampini 6214f1b2e48SStefano Zampini #undef __FUNCT__ 6225408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 6235408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 6245408967cSStefano Zampini { 6255408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6265408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 627dee84bffSStefano Zampini IS dirIS = NULL; 6284f1b2e48SStefano Zampini PetscInt i; 6295408967cSStefano Zampini PetscErrorCode ierr; 6305408967cSStefano Zampini 6315408967cSStefano Zampini PetscFunctionBegin; 632dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 6335408967cSStefano Zampini if (zerodiag) { 6345408967cSStefano Zampini Mat A; 6355408967cSStefano Zampini Vec vec3_N; 6365408967cSStefano Zampini PetscScalar *vals; 6375408967cSStefano Zampini const PetscInt *idxs; 638d12d3064SStefano Zampini PetscInt nz,*count; 6395408967cSStefano Zampini 6405408967cSStefano Zampini /* p0 */ 6415408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 6425408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 6435408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6445408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6454f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 6465408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6475408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6485408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6495408967cSStefano Zampini /* v_I */ 6505408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 6515408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 6525408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6535408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 6545408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6555408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 6565408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6575408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 6585408967cSStefano Zampini if (dirIS) { 6595408967cSStefano Zampini PetscInt n; 6605408967cSStefano Zampini 6615408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 6625408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 6635408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 6645408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 6655408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 6665408967cSStefano Zampini } 6675408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 6685408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 6695408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 6705408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 671669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 6725408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 6735408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 6749a962809SStefano 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])); 6755408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6765408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 677d12d3064SStefano Zampini 678d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 679d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 680d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 681d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 682d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 683d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 6849a962809SStefano 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]); 685d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 686d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 6875408967cSStefano Zampini } 688dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 6895408967cSStefano Zampini 6905408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 6915408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 6924f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 6935408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 6944f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 6955408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 696f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 697f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 698f2a566d8SStefano Zampini if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %d instead of %g\n",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i);CHKERRQ(ierr); 699f2a566d8SStefano Zampini } 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++) { 1542684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1543684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 1544684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 1545684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1546684229deSStefano Zampini #else 15476c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 1548684229deSStefano Zampini #endif 15496c3e6151SStefano Zampini } 15506c3e6151SStefano Zampini } 15516c3e6151SStefano Zampini } 155272b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 15536c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 155472b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 15556c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 15566c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 15576c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 15586c3e6151SStefano Zampini } 15598bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 15608bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 15619162d606SStefano Zampini if (B_neigs) { 15629162d606SStefano 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); 1563fd14bc51SStefano Zampini 1564fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 15659552c7c7SStefano Zampini PetscInt ii; 15669552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1567ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 15689552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1569ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1570ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1571ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1572ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1573ac47001eSStefano Zampini #else 1574ac47001eSStefano 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); 1575ac47001eSStefano Zampini #endif 15769552c7c7SStefano Zampini } 15779552c7c7SStefano Zampini } 1578fd14bc51SStefano Zampini } 15799162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 15809162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 15819162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 15829162d606SStefano Zampini cum++; 158308122e43SStefano Zampini } 158408122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 158508122e43SStefano Zampini /* shift for next computation */ 158608122e43SStefano Zampini cumarray += subset_size*subset_size; 158708122e43SStefano Zampini } 1588fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1589fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1590fd14bc51SStefano Zampini } 159108122e43SStefano Zampini 159208122e43SStefano Zampini if (mss) { 159308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 159408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1595f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1596f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1597f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 159808122e43SStefano Zampini } 1599f6f667cfSStefano Zampini if (allocated_S_St) { 1600f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1601f6f667cfSStefano Zampini } 1602f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 160308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 160408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 160508122e43SStefano Zampini #endif 160608122e43SStefano Zampini if (pcbddc->dbg_flag) { 16071b968477SStefano Zampini PetscInt maxneigs_r; 1608b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 16099b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 161008122e43SStefano Zampini } 161108122e43SStefano Zampini PetscFunctionReturn(0); 161208122e43SStefano Zampini } 1613b1b3d7a2SStefano Zampini 1614674ae819SStefano Zampini #undef __FUNCT__ 1615c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1616c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1617c8587f34SStefano Zampini { 16188629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1619c8587f34SStefano Zampini PetscErrorCode ierr; 1620c8587f34SStefano Zampini 1621c8587f34SStefano Zampini PetscFunctionBegin; 1622f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 16235e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1624c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1625c8587f34SStefano Zampini 1626684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 16270fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1628684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1629c8587f34SStefano Zampini 16308629588bSStefano Zampini /* 16318629588bSStefano Zampini Setup local correction and local part of coarse basis. 16328629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 16338629588bSStefano Zampini */ 163447f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 16358629588bSStefano Zampini 16368629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 16378629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 16388629588bSStefano Zampini 16398629588bSStefano Zampini /* free */ 16408629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1641c8587f34SStefano Zampini PetscFunctionReturn(0); 1642c8587f34SStefano Zampini } 1643c8587f34SStefano Zampini 1644c8587f34SStefano Zampini #undef __FUNCT__ 1645674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1646674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1647674ae819SStefano Zampini { 1648674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1649674ae819SStefano Zampini PetscErrorCode ierr; 1650674ae819SStefano Zampini 1651674ae819SStefano Zampini PetscFunctionBegin; 1652674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1653674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 165430368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1655674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1656785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1657674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1658f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1659f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1660785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 166163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 166263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1663674ae819SStefano Zampini PetscFunctionReturn(0); 1664674ae819SStefano Zampini } 1665674ae819SStefano Zampini 1666674ae819SStefano Zampini #undef __FUNCT__ 1667674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1668674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1669674ae819SStefano Zampini { 1670674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16714f1b2e48SStefano Zampini PetscInt i; 1672674ae819SStefano Zampini PetscErrorCode ierr; 1673674ae819SStefano Zampini 1674674ae819SStefano Zampini PetscFunctionBegin; 1675b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1676674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 167716909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 16781dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1679674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1680669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 1681fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 1682674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 168371582508SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 16844f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 16854f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 16864f1b2e48SStefano Zampini } 16874f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1688b334f244SStefano Zampini if (pcbddc->sub_schurs) { 1689b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1690b334f244SStefano Zampini } 1691674ae819SStefano Zampini PetscFunctionReturn(0); 1692674ae819SStefano Zampini } 1693674ae819SStefano Zampini 1694674ae819SStefano Zampini #undef __FUNCT__ 1695674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1696674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1697674ae819SStefano Zampini { 1698674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1699674ae819SStefano Zampini PetscErrorCode ierr; 1700674ae819SStefano Zampini 1701674ae819SStefano Zampini PetscFunctionBegin; 1702674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 170358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1704ca92afb2SStefano Zampini PetscScalar *array; 170506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 170606656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 170758da7f69SStefano Zampini } 1708674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1709674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 171015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 171115aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1712674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1713674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1714674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 171506656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1716674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1717674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17188ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1719674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1720674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1721674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1722f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1723f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1724f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1725f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1726727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 17270e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1728f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 172970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 173081d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 17310369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 17321dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 17334f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 17348b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1735ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1736ca92afb2SStefano Zampini PetscInt i; 1737ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1738ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1739ca92afb2SStefano Zampini } 1740ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1741ca92afb2SStefano Zampini } 17424f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1743674ae819SStefano Zampini PetscFunctionReturn(0); 1744674ae819SStefano Zampini } 1745674ae819SStefano Zampini 1746674ae819SStefano Zampini #undef __FUNCT__ 1747f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1748f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 17496bfb1811SStefano Zampini { 17506bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 17516bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 17526bfb1811SStefano Zampini VecType impVecType; 17534f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 17546bfb1811SStefano Zampini PetscErrorCode ierr; 17556bfb1811SStefano Zampini 17566bfb1811SStefano Zampini PetscFunctionBegin; 17576c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1758e7b262bdSStefano Zampini /* get sizes */ 17594f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1760b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 17616bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1762e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1763e7b262bdSStefano Zampini /* R nodes */ 1764e7b262bdSStefano Zampini old_size = -1; 1765e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1766e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1767e7b262bdSStefano Zampini } 1768e7b262bdSStefano Zampini if (n_R != old_size) { 1769e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1770e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 17716bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 17726bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 17736bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 17746bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1775e7b262bdSStefano Zampini } 1776e7b262bdSStefano Zampini /* local primal dofs */ 1777e7b262bdSStefano Zampini old_size = -1; 1778e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1779e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1780e7b262bdSStefano Zampini } 1781e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1782e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 178383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1784e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 17856bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1786e7b262bdSStefano Zampini } 1787e7b262bdSStefano Zampini /* local explicit constraints */ 1788e7b262bdSStefano Zampini old_size = -1; 1789e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1790e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1791e7b262bdSStefano Zampini } 1792e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1793e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 179483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 179583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 179683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 179783b7ccabSStefano Zampini } 17986bfb1811SStefano Zampini PetscFunctionReturn(0); 17996bfb1811SStefano Zampini } 18006bfb1811SStefano Zampini 18016bfb1811SStefano Zampini #undef __FUNCT__ 180247f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 180347f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 180488ebb749SStefano Zampini { 180525084f0cSStefano Zampini PetscErrorCode ierr; 180625084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 180788ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 180888ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1809d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 181025084f0cSStefano Zampini /* submatrices of local problem */ 181180677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 181206656605SStefano Zampini /* submatrices of local coarse problem */ 181306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 181425084f0cSStefano Zampini /* working matrices */ 181506656605SStefano Zampini Mat C_CR; 181625084f0cSStefano Zampini /* additional working stuff */ 181706656605SStefano Zampini PC pc_R; 18184f1b2e48SStefano Zampini Mat F; 18195cbda25cSStefano Zampini Vec dummy_vec; 1820a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 182125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 182206656605SStefano Zampini PetscScalar *work; 182306656605SStefano Zampini PetscInt *idx_V_B; 1824ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 182506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1826ffd830a3SStefano Zampini 182725084f0cSStefano Zampini /* some shortcuts to scalars */ 182806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 182988ebb749SStefano Zampini 183088ebb749SStefano Zampini PetscFunctionBegin; 18319a962809SStefano 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"); 1832ffd830a3SStefano Zampini 1833ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1834b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 18354f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1836b371cd4fSStefano Zampini n_B = pcis->n_B; 1837b371cd4fSStefano Zampini n_D = pcis->n - n_B; 183888ebb749SStefano Zampini n_R = pcis->n - n_vertices; 183988ebb749SStefano Zampini 184088ebb749SStefano Zampini /* vertices in boundary numbering */ 1841785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 18420e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 18436c4ed002SBarry 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); 184488ebb749SStefano Zampini 184506656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1846019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 184706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 184806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 184906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 185006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 185106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 185206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 185306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 185406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 185506656605SStefano Zampini 185606656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 185706656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 185806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 185906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 186006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1861ffd830a3SStefano Zampini lda_rhs = n_R; 1862a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 186306656605SStefano Zampini if (isLU || isILU || isCHOL) { 186406656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1865b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 1866df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1867d62866d3SStefano Zampini MatFactorType type; 1868d62866d3SStefano Zampini 1869df4d28bfSStefano Zampini F = reuse_solver->F; 18706816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1871d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1872ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 187322db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 187406656605SStefano Zampini } else { 187506656605SStefano Zampini F = NULL; 187606656605SStefano Zampini } 187706656605SStefano Zampini 1878ffd830a3SStefano Zampini /* allocate workspace */ 1879ffd830a3SStefano Zampini n = 0; 1880ffd830a3SStefano Zampini if (n_constraints) { 1881ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1882ffd830a3SStefano Zampini } 1883ffd830a3SStefano Zampini if (n_vertices) { 1884ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1885ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1886ffd830a3SStefano Zampini } 1887ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1888ffd830a3SStefano Zampini 18895cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 18905cbda25cSStefano Zampini dummy_vec = NULL; 18915cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 18925cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 18935cbda25cSStefano Zampini } 18945cbda25cSStefano Zampini 189588ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 189688ebb749SStefano Zampini if (n_constraints) { 189772b8c272SStefano Zampini Mat M1,M2,M3,C_B; 189806656605SStefano Zampini IS is_aux; 189980677318SStefano Zampini PetscScalar *array,*array2; 190006656605SStefano Zampini 1901f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 190280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 190388ebb749SStefano Zampini 190425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 190525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 19068ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 190772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 190888ebb749SStefano Zampini 190980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 191080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1911ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 191288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 191306656605SStefano Zampini const PetscScalar *row_cmat_values; 191406656605SStefano Zampini const PetscInt *row_cmat_indices; 191506656605SStefano Zampini PetscInt size_of_constraint,j; 191688ebb749SStefano Zampini 191706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 191806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1919ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 192006656605SStefano Zampini } 192106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 192206656605SStefano Zampini } 1923ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 192406656605SStefano Zampini if (F) { 192506656605SStefano Zampini Mat B; 192606656605SStefano Zampini 1927ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1928a3df083aSStefano Zampini if (need_benign_correction) { 1929df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1930a3df083aSStefano Zampini 193172b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 193272b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1933a3df083aSStefano Zampini } 193480677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1935a3df083aSStefano Zampini if (need_benign_correction) { 1936a3df083aSStefano Zampini PetscScalar *marr; 1937df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1938a3df083aSStefano Zampini 1939a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 19405cbda25cSStefano Zampini if (lda_rhs != n_R) { 19415cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 19425cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19435cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19445cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19455cbda25cSStefano Zampini } 19465cbda25cSStefano Zampini } else { 1947a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1948a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19495cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1950a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1951a3df083aSStefano Zampini } 19525cbda25cSStefano Zampini } 1953a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1954a3df083aSStefano Zampini } 195506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195606656605SStefano Zampini } else { 195780677318SStefano Zampini PetscScalar *marr; 195880677318SStefano Zampini 195980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 196006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1961ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1962ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 196306656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 196406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 196506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 196606656605SStefano Zampini } 196780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 196806656605SStefano Zampini } 196980677318SStefano Zampini if (!pcbddc->switch_static) { 197080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 197180677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 197280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 197380677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1974ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 197580677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 197680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 197980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 198080677318SStefano Zampini } 198180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 198280677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 198372b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 198480677318SStefano Zampini } else { 1985ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1986ffd830a3SStefano Zampini IS dummy; 1987ffd830a3SStefano Zampini 1988ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 198972b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1990ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1991ffd830a3SStefano Zampini } else { 199280677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 199380677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1994ffd830a3SStefano Zampini } 199525084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 199680677318SStefano Zampini } 199780677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 199880677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 199980677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 200006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 200280677318SStefano Zampini if (isCHOL) { 200380677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 200480677318SStefano Zampini } else { 200525084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 200680677318SStefano Zampini } 200780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 200806656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 200925084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 201025084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 201125084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 201280677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 201372b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 201472b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 201506656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 201606656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 2017f4ddd8eeSStefano Zampini } 2018fc227af8SStefano Zampini 2019fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 202088ebb749SStefano Zampini if (n_vertices) { 202106656605SStefano Zampini IS is_aux; 20223a50541eSStefano Zampini 2023b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 20246816873aSStefano Zampini IS tis; 20256816873aSStefano Zampini 20266816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 20276816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 20286816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 20296816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 20306816873aSStefano Zampini } else { 20313a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 20326816873aSStefano Zampini } 20339577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 20349577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 203504708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 203625084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 203788ebb749SStefano Zampini } 203888ebb749SStefano Zampini 203988ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 2040f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 204106656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 204206656605SStefano Zampini if (pcbddc->coarse_phi_D) { 204306656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 204406656605SStefano Zampini } 2045f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 204606656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 204706656605SStefano Zampini PetscScalar *marray; 204806656605SStefano Zampini 204906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 205006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 2051f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2052f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 2053f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 2054f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2055f4ddd8eeSStefano Zampini } 2056f4ddd8eeSStefano Zampini } 205706656605SStefano Zampini 2058f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 205906656605SStefano Zampini PetscScalar *marray; 206088ebb749SStefano Zampini 206106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 206306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 206488ebb749SStefano Zampini } 20653301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 206606656605SStefano Zampini n *= 2; 206788ebb749SStefano Zampini } 206806656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 206906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 207006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 20718eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 207206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 207306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 207488ebb749SStefano Zampini } 20753301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 207606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 20778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 207806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 207906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 208088ebb749SStefano Zampini } 208188ebb749SStefano Zampini } else { 2082c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 2083c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 20841b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2085c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 2086c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 2087c0553b1fSStefano Zampini } 208888ebb749SStefano Zampini } 208906656605SStefano Zampini } 2090019a44ceSStefano Zampini 209106656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 20924f1b2e48SStefano Zampini p0_lidx_I = NULL; 20934f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 2094d12edf2fSStefano Zampini const PetscInt *idxs; 2095d12edf2fSStefano Zampini 2096d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 20974f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 20984f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20994f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 21004f1b2e48SStefano Zampini } 2101d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 2102d12edf2fSStefano Zampini } 2103d16cbb6bSStefano Zampini 210406656605SStefano Zampini /* vertices */ 210506656605SStefano Zampini if (n_vertices) { 210616f15bc4SStefano Zampini 2107af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 210804708bb6SStefano Zampini 210916f15bc4SStefano Zampini if (n_R) { 211014393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 211106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 211216f15bc4SStefano Zampini PetscScalar *x,*y; 211304708bb6SStefano Zampini PetscBool isseqaij; 211406656605SStefano Zampini 211521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 211614393ed6SStefano Zampini if (need_benign_correction) { 211714393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 211814393ed6SStefano Zampini IS is_p0; 211914393ed6SStefano Zampini PetscInt *idxs_p0,n; 212014393ed6SStefano Zampini 212114393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 212214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 212314393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 2124af25d912SStefano 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); 212514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 212614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 212714393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 212814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 212914393ed6SStefano Zampini } 213014393ed6SStefano Zampini 2131ffd830a3SStefano Zampini if (lda_rhs == n_R) { 2132af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2133ffd830a3SStefano Zampini } else { 2134ca92afb2SStefano Zampini PetscScalar *av,*array; 2135ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 2136ca92afb2SStefano Zampini PetscInt n; 2137ca92afb2SStefano Zampini PetscBool flg_row; 2138ffd830a3SStefano Zampini 2139ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 2140ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 21419d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 2142ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2143ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 2144ca92afb2SStefano Zampini for (i=0;i<n;i++) { 2145ca92afb2SStefano Zampini PetscInt j; 2146ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 2147ffd830a3SStefano Zampini } 2148ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2149ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2150ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 2151ffd830a3SStefano Zampini } 2152ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2153a3df083aSStefano Zampini if (need_benign_correction) { 2154df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2155a3df083aSStefano Zampini PetscScalar *marr; 2156a3df083aSStefano Zampini 2157a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 215814393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 215914393ed6SStefano Zampini 216014393ed6SStefano Zampini | 0 0 0 | (V) 216114393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 216214393ed6SStefano Zampini | 0 0 -1 | (p0) 216314393ed6SStefano Zampini 216414393ed6SStefano Zampini */ 2165df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 216614393ed6SStefano Zampini const PetscScalar *vals; 216714393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 216814393ed6SStefano Zampini PetscInt n,j,nz; 216914393ed6SStefano Zampini 2170df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2171df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 217214393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 217314393ed6SStefano Zampini for (j=0;j<n;j++) { 217414393ed6SStefano Zampini PetscScalar val = vals[j]; 217514393ed6SStefano Zampini PetscInt k,col = idxs[j]; 217614393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 217714393ed6SStefano Zampini } 217814393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2179df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 218014393ed6SStefano Zampini } 218172b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 218272b8c272SStefano Zampini } 218372b8c272SStefano Zampini if (F) { 218414393ed6SStefano Zampini /* need to correct the rhs */ 218572b8c272SStefano Zampini if (need_benign_correction) { 218672b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 218772b8c272SStefano Zampini PetscScalar *marr; 218872b8c272SStefano Zampini 218972b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 21905cbda25cSStefano Zampini if (lda_rhs != n_R) { 21915cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 21925cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 21935cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 21945cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 21955cbda25cSStefano Zampini } 21965cbda25cSStefano Zampini } else { 2197a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2198a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 21995cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2200a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2201a3df083aSStefano Zampini } 22025cbda25cSStefano Zampini } 2203a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 2204a3df083aSStefano Zampini } 220506656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 220614393ed6SStefano Zampini /* need to correct the solution */ 2207a3df083aSStefano Zampini if (need_benign_correction) { 2208df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2209a3df083aSStefano Zampini PetscScalar *marr; 2210a3df083aSStefano Zampini 2211a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 22125cbda25cSStefano Zampini if (lda_rhs != n_R) { 22135cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 22145cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 22155cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 22165cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 22175cbda25cSStefano Zampini } 22185cbda25cSStefano Zampini } else { 2219a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 2220a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 22215cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 2222a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2223a3df083aSStefano Zampini } 22245cbda25cSStefano Zampini } 2225a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 2226a3df083aSStefano Zampini } 222706656605SStefano Zampini } else { 222806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 222906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 2230ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 2231ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 223206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 223306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 223406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 223506656605SStefano Zampini } 223606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 223706656605SStefano Zampini } 223880677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 2239ffd830a3SStefano Zampini /* S_VV and S_CV */ 224006656605SStefano Zampini if (n_constraints) { 224106656605SStefano Zampini Mat B; 224280677318SStefano Zampini 2243ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 224480677318SStefano Zampini for (i=0;i<n_vertices;i++) { 2245ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 2246ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 224780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 224880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 224980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 225080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 225180677318SStefano Zampini } 2252ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 225380677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 225480677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 2255ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 225680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 225706656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 2258ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 2259ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 226006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 226106656605SStefano Zampini } 226204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 226304708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 2264511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 226504708bb6SStefano Zampini } 2266ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2267ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 2268ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 2269ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 2270ffd830a3SStefano Zampini } 227106656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 227214393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 227314393ed6SStefano Zampini if (need_benign_correction) { 2274df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 227514393ed6SStefano Zampini PetscScalar *marr,*sums; 227614393ed6SStefano Zampini 227714393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 227814393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 2279df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 228014393ed6SStefano Zampini const PetscScalar *vals; 228114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 228214393ed6SStefano Zampini PetscInt n,j,nz; 228314393ed6SStefano Zampini 2284df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2285df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 228614393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 228714393ed6SStefano Zampini PetscInt k; 228814393ed6SStefano Zampini sums[j] = 0.; 228914393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 229014393ed6SStefano Zampini } 229114393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 229214393ed6SStefano Zampini for (j=0;j<n;j++) { 229314393ed6SStefano Zampini PetscScalar val = vals[j]; 229414393ed6SStefano Zampini PetscInt k; 229514393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 229614393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 229714393ed6SStefano Zampini } 229814393ed6SStefano Zampini } 229914393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2300df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 230114393ed6SStefano Zampini } 230214393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 230314393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 230414393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 230514393ed6SStefano Zampini } 230680677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 230706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 230806656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 230906656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 231006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 231106656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 231206656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 231306656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2314d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2315019a44ceSStefano Zampini } else { 2316d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2317d16cbb6bSStefano Zampini } 231821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2319d16cbb6bSStefano Zampini 232006656605SStefano Zampini /* coarse basis functions */ 232106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 232216f15bc4SStefano Zampini PetscScalar *y; 232316f15bc4SStefano Zampini 2324ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 232506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 232606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 232706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 232906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 233006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 233106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 233206656605SStefano Zampini 233306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23344f1b2e48SStefano Zampini PetscInt j; 23354f1b2e48SStefano Zampini 233606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 233706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 233806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 233906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 234006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23414f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 234206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 234306656605SStefano Zampini } 234406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 234506656605SStefano Zampini } 234604708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 234704708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 234806656605SStefano Zampini } 23495cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 235006656605SStefano Zampini 235106656605SStefano Zampini if (n_constraints) { 235206656605SStefano Zampini Mat B; 235306656605SStefano Zampini 2354ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 235506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 235680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 235706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 235806656605SStefano Zampini if (n_vertices) { 235980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 236080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 236180677318SStefano Zampini } else { 236280677318SStefano Zampini Mat S_VCt; 236380677318SStefano Zampini 2364ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2365ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 236672b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2367ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2368ffd830a3SStefano Zampini } 236980677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 237080677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 237180677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 237280677318SStefano Zampini } 237306656605SStefano Zampini } 237406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 237506656605SStefano Zampini /* coarse basis functions */ 237606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 237706656605SStefano Zampini PetscScalar *y; 237806656605SStefano Zampini 2379ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 238006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 238106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 238206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 238506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 238606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 23874f1b2e48SStefano Zampini PetscInt j; 23884f1b2e48SStefano Zampini 238906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 239006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 239106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 239306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 23944f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 239506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 239606656605SStefano Zampini } 239706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 239806656605SStefano Zampini } 239906656605SStefano Zampini } 240080677318SStefano Zampini if (n_constraints) { 240180677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 240280677318SStefano Zampini } 24034f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 240472b8c272SStefano Zampini 240572b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 240672b8c272SStefano Zampini if (pcbddc->benign_n) { 240772b8c272SStefano Zampini Mat B0_B,B0_BPHI; 240872b8c272SStefano Zampini IS is_dummy; 240972b8c272SStefano Zampini PetscScalar *data; 241072b8c272SStefano Zampini PetscInt j; 241172b8c272SStefano Zampini 241272b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 241372b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 241472b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 241572b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 241686c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 241772b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 241872b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 241972b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 242072b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 242172b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 242272b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 242372b8c272SStefano Zampini } 242472b8c272SStefano Zampini } 242572b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 242672b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 242772b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 242872b8c272SStefano Zampini } 2429019a44ceSStefano Zampini 243006656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 24313301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2432ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2433ffd830a3SStefano Zampini PetscScalar *marray; 243406656605SStefano Zampini 243506656605SStefano Zampini if (n_constraints) { 2436ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 243706656605SStefano Zampini 2438af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 243906656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2440ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 244116f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 244206656605SStefano Zampini if (n_vertices) { 2443ffd830a3SStefano Zampini Mat S_VCT; 244406656605SStefano Zampini 244506656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2446ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 244716f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 244806656605SStefano Zampini } 2449ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 24505b782168SStefano Zampini } else { 24515b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 245206656605SStefano Zampini } 245316f15bc4SStefano Zampini if (n_vertices && n_R) { 2454ffd830a3SStefano Zampini PetscScalar *av,*marray; 2455ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2456ffd830a3SStefano Zampini PetscInt n; 2457ffd830a3SStefano Zampini PetscBool flg_row; 245806656605SStefano Zampini 2459ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2460af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2461ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2462ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2463ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2464ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2465ffd830a3SStefano Zampini PetscInt j; 2466ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2467ffd830a3SStefano Zampini } 2468ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2469ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2470ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 247106656605SStefano Zampini } 247206656605SStefano Zampini 2473ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2474ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2475ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2476ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2477ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 247806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 247906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 248006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 248106656605SStefano Zampini } 2482ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 24835b782168SStefano Zampini if (B_C) { 2484ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2485ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2486ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2487ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2488ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2489ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2490ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 249106656605SStefano Zampini } 2492ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 24935b782168SStefano Zampini } 249406656605SStefano Zampini /* coarse basis functions */ 249506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 249606656605SStefano Zampini PetscScalar *y; 249706656605SStefano Zampini 2498ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 249906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 250006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 250106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 250306656605SStefano Zampini if (i<n_vertices) { 250406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 250506656605SStefano Zampini } 250606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 250706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 250806656605SStefano Zampini 250906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 251006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 251106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 251206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 251406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 251506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 251606656605SStefano Zampini } 251706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 251806656605SStefano Zampini } 2519ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2520ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 252106656605SStefano Zampini } 2522d62866d3SStefano Zampini /* free memory */ 252388ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 252406656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 252506656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 252606656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 252706656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2528d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2529d62866d3SStefano Zampini if (n_vertices) { 2530d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2531d62866d3SStefano Zampini } 2532d62866d3SStefano Zampini if (n_constraints) { 2533d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2534d62866d3SStefano Zampini } 253588ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 253688ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 253788ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2538d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 253988ebb749SStefano Zampini Mat coarse_sub_mat; 254025084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 254188ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 254288ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 254388ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 25448bec7fa6SStefano Zampini Mat C_B,CPHI; 25458bec7fa6SStefano Zampini IS is_dummy; 25468bec7fa6SStefano Zampini Vec mones; 254788ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 254888ebb749SStefano Zampini PetscReal real_value; 254988ebb749SStefano Zampini 2550a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2551a3df083aSStefano Zampini Mat A; 2552a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2553a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2554a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2555a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2556a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2557a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2558a3df083aSStefano Zampini } else { 255988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 256088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 256188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 256288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2563a3df083aSStefano Zampini } 256488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 256588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2566ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 256788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 256888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 256988ebb749SStefano Zampini } 257088ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 257188ebb749SStefano Zampini 257225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 25733301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 257425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2575ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 257688ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 257788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 257888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 257988ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 258188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 258488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 258688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 258788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 258888ebb749SStefano Zampini } else { 258988ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 259088ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 259188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 259288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 259388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 259488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 259588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 259688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 259788ebb749SStefano Zampini } 259888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 259988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 260088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2601511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 26024f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2603fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2604d12edf2fSStefano Zampini PetscScalar *data,*data2; 26054f1b2e48SStefano Zampini PetscInt j; 2606d12edf2fSStefano Zampini 26074f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2608fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2609d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 261086c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2611d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2612d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 26134f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 26144f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2615d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 26164f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 26174f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 26184f1b2e48SStefano Zampini } 2619d12edf2fSStefano Zampini } 2620d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2621d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2622d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2623d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2624d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2625d12edf2fSStefano Zampini } 2626d12edf2fSStefano Zampini #if 0 2627d12edf2fSStefano Zampini { 2628d12edf2fSStefano Zampini PetscViewer viewer; 2629d12edf2fSStefano Zampini char filename[256]; 2630ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2631d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2632d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2633ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2634ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2635ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2636d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 263772b8c272SStefano Zampini if (save_change) { 263872b8c272SStefano Zampini Mat phi_B; 263972b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 264072b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 264172b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 264272b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 264372b8c272SStefano Zampini } else { 2644ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2645ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 264672b8c272SStefano Zampini } 2647ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2648ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2649ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2650ffd830a3SStefano Zampini } 2651ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2652ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2653ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2654ffd830a3SStefano Zampini } 265572b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2656ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2657ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2658ffd830a3SStefano Zampini } 2659d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2660d12edf2fSStefano Zampini } 2661d12edf2fSStefano Zampini #endif 266281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 26638bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 26641575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 266506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 26668bec7fa6SStefano Zampini 26678bec7fa6SStefano Zampini /* check constraints */ 2668a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2669a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 26704f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 26718bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2672a00504b5SStefano Zampini } else { 2673a00504b5SStefano Zampini PetscScalar *data; 2674a00504b5SStefano Zampini Mat tmat; 2675a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2676a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2677a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2678a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2679a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2680a00504b5SStefano Zampini } 26818bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 26828bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 26838bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 26848bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2685bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2686ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2687bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2688bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2689bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2690bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2691bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 269288ebb749SStefano Zampini } 26938bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 26948bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 26958bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 26968bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 269725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 269888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 269988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 270088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 270188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 270288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 270388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 270488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 270588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 270688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 270788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2708ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 270988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 271088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 271188ebb749SStefano Zampini } 271288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 271388ebb749SStefano Zampini } 27148629588bSStefano Zampini /* get back data */ 27158629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 271688ebb749SStefano Zampini PetscFunctionReturn(0); 271788ebb749SStefano Zampini } 271888ebb749SStefano Zampini 271988ebb749SStefano Zampini #undef __FUNCT__ 2720d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2721d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2722aa0d41d4SStefano Zampini { 2723d65f70fdSStefano Zampini Mat *work_mat; 2724d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2725d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2726c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 2727aa0d41d4SStefano Zampini PetscErrorCode ierr; 2728aa0d41d4SStefano Zampini 2729aa0d41d4SStefano Zampini PetscFunctionBegin; 2730d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2731d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2732d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2733d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2734aa0d41d4SStefano Zampini 2735d65f70fdSStefano Zampini if (!rsorted) { 2736906d46d4SStefano Zampini const PetscInt *idxs; 2737906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2738aa0d41d4SStefano Zampini 2739d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2740d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2741d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2742d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2743aa0d41d4SStefano Zampini } 2744d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2745d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2746d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2747d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2748aa0d41d4SStefano Zampini } 2749d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2750d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2751d65f70fdSStefano Zampini } else { 2752d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2753d65f70fdSStefano Zampini isrow_s = isrow; 2754aa0d41d4SStefano Zampini } 2755906d46d4SStefano Zampini 2756d65f70fdSStefano Zampini if (!csorted) { 2757d65f70fdSStefano Zampini if (isrow == iscol) { 2758d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2759d65f70fdSStefano Zampini iscol_s = isrow_s; 2760d65f70fdSStefano Zampini } else { 2761d65f70fdSStefano Zampini const PetscInt *idxs; 2762d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2763906d46d4SStefano Zampini 2764d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2765d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2766d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2767d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2768d65f70fdSStefano Zampini } 2769d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2770d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2771d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2772d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2773d65f70fdSStefano Zampini } 2774d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2775d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2776d65f70fdSStefano Zampini } 2777d65f70fdSStefano Zampini } else { 2778d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2779d65f70fdSStefano Zampini iscol_s = iscol; 2780d65f70fdSStefano Zampini } 2781d65f70fdSStefano Zampini 2782d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2783d65f70fdSStefano Zampini 2784d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2785906d46d4SStefano Zampini Mat new_mat; 2786d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2787906d46d4SStefano Zampini 2788d65f70fdSStefano Zampini if (!rsorted) { 2789d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2790d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2791d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2792d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2793906d46d4SStefano Zampini } 2794d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2795d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2796d65f70fdSStefano Zampini } else { 2797d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2798906d46d4SStefano Zampini } 2799d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2800d65f70fdSStefano Zampini 2801d65f70fdSStefano Zampini if (!csorted) { 2802d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2803d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2804d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2805d65f70fdSStefano Zampini } else { 2806d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2807d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2808d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2809d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2810d65f70fdSStefano Zampini } 2811d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2812d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2813d65f70fdSStefano Zampini } 2814d65f70fdSStefano Zampini } else { 2815d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2816d65f70fdSStefano Zampini } 2817d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2818d65f70fdSStefano Zampini 2819d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2820d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2821d65f70fdSStefano Zampini work_mat[0] = new_mat; 2822d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2823d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2824d65f70fdSStefano Zampini } 2825d65f70fdSStefano Zampini 2826d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2827d65f70fdSStefano Zampini *B = work_mat[0]; 2828d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2829d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2830d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2831d65f70fdSStefano Zampini PetscFunctionReturn(0); 2832d65f70fdSStefano Zampini } 2833d65f70fdSStefano Zampini 2834d65f70fdSStefano Zampini #undef __FUNCT__ 28355e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 28365e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2837aa0d41d4SStefano Zampini { 2838aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 28395e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2840d65f70fdSStefano Zampini Mat new_mat; 28415e8657edSStefano Zampini IS is_local,is_global; 2842d65f70fdSStefano Zampini PetscInt local_size; 2843d65f70fdSStefano Zampini PetscBool isseqaij; 2844aa0d41d4SStefano Zampini PetscErrorCode ierr; 2845aa0d41d4SStefano Zampini 2846aa0d41d4SStefano Zampini PetscFunctionBegin; 2847aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 28485e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 28495e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2850b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2851aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2852d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2853aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2854906d46d4SStefano Zampini 2855906d46d4SStefano Zampini /* check */ 2856906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2857906d46d4SStefano Zampini Vec x,x_change; 2858906d46d4SStefano Zampini PetscReal error; 2859906d46d4SStefano Zampini 28605e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2861906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 28625e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2863e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2864e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2865d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2866e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2867e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2868906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2869906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2870906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2871906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2872906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2873906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2874906d46d4SStefano Zampini } 2875906d46d4SStefano Zampini 287622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 28779b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 287822d5777bSStefano Zampini if (isseqaij) { 2879a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2880a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2881aa0d41d4SStefano Zampini } else { 2882a00504b5SStefano Zampini Mat work_mat; 28831cf9b237SStefano Zampini 2884a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2885aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2886a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 28871d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2888aa0d41d4SStefano Zampini } 28893301b35fSStefano Zampini if (matis->A->symmetric_set) { 28903301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2891e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 28923301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2893e496cd5dSStefano Zampini #endif 28943301b35fSStefano Zampini } 2895d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2896aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2897aa0d41d4SStefano Zampini } 2898aa0d41d4SStefano Zampini 2899aa0d41d4SStefano Zampini #undef __FUNCT__ 2900a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 29018ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2902a64d13efSStefano Zampini { 2903a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2904a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2905d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 290653892102SStefano Zampini PetscInt *idx_R_local=NULL; 29073a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 29083a50541eSStefano Zampini PetscInt vbs,bs; 29096816873aSStefano Zampini PetscBT bitmask=NULL; 2910a64d13efSStefano Zampini PetscErrorCode ierr; 2911a64d13efSStefano Zampini 2912a64d13efSStefano Zampini PetscFunctionBegin; 2913b23d619eSStefano Zampini /* 2914b23d619eSStefano Zampini No need to setup local scatters if 2915b23d619eSStefano Zampini - primal space is unchanged 2916b23d619eSStefano Zampini AND 2917b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2918b23d619eSStefano Zampini AND 2919b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2920b23d619eSStefano Zampini */ 2921b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2922f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2923f4ddd8eeSStefano Zampini } 2924f4ddd8eeSStefano Zampini /* destroy old objects */ 2925f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2926f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2927f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2928a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2929b371cd4fSStefano Zampini n_B = pcis->n_B; 2930b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2931b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 29323a50541eSStefano Zampini 2933a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 29346816873aSStefano Zampini 293553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2936b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2937854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2938a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2939a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 29400e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2941a64d13efSStefano Zampini } 2942a64d13efSStefano Zampini 2943a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 29444641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 29456816873aSStefano Zampini idx_R_local[n_R++] = i; 2946a64d13efSStefano Zampini } 2947a64d13efSStefano Zampini } 2948df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2949df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 29506816873aSStefano Zampini 2951df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2952df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 29536816873aSStefano Zampini } 29543a50541eSStefano Zampini 29553a50541eSStefano Zampini /* Block code */ 29563a50541eSStefano Zampini vbs = 1; 29573a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 29583a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 29593a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 29603a50541eSStefano Zampini PetscInt *vary; 2961b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 2962785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 29633a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2964d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2965d3df7717SStefano 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 */ 29660e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2967d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 29683a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 29693a50541eSStefano Zampini is_blocked = PETSC_FALSE; 29703a50541eSStefano Zampini break; 29713a50541eSStefano Zampini } 29723a50541eSStefano Zampini } 2973d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2974d3df7717SStefano Zampini } else { 2975d3df7717SStefano Zampini /* Verify directly the R set */ 2976d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2977d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2978d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2979d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2980d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2981d3df7717SStefano Zampini break; 2982d3df7717SStefano Zampini } 2983d3df7717SStefano Zampini } 2984d3df7717SStefano Zampini } 2985d3df7717SStefano Zampini } 29863a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 29873a50541eSStefano Zampini vbs = bs; 29883a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 29893a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 29903a50541eSStefano Zampini } 29913a50541eSStefano Zampini } 29923a50541eSStefano Zampini } 29933a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2994b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 2995df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 299653892102SStefano Zampini 2997df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2998df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 299953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 3000df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 300153892102SStefano Zampini } else { 30023a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 300353892102SStefano Zampini } 3004a64d13efSStefano Zampini 3005a64d13efSStefano Zampini /* print some info if requested */ 3006a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 3007a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3008a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 30091575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3010a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 3011a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 30124f1b2e48SStefano 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); 3013a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3014a64d13efSStefano Zampini } 3015a64d13efSStefano Zampini 3016a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 3017b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 30186816873aSStefano Zampini IS is_aux1,is_aux2; 30196816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 30206816873aSStefano Zampini 30213a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3022854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 3023854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 3024a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 30254641a718SStefano Zampini for (i=0; i<n_D; i++) { 30264641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 30274641a718SStefano Zampini } 3028a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3029a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30304641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 30314641a718SStefano Zampini aux_array1[j++] = i; 3032a64d13efSStefano Zampini } 3033a64d13efSStefano Zampini } 3034a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3035a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3036a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 30374641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 30384641a718SStefano Zampini aux_array2[j++] = i; 3039a64d13efSStefano Zampini } 3040a64d13efSStefano Zampini } 3041a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3042a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 3043a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 3044a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3045a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 3046a64d13efSStefano Zampini 30478eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3048785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 3049a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 30504641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 30514641a718SStefano Zampini aux_array1[j++] = i; 3052a64d13efSStefano Zampini } 3053a64d13efSStefano Zampini } 3054a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 3055a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 3056a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 3057a64d13efSStefano Zampini } 30584641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 30593a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 3060d62866d3SStefano Zampini } else { 3061df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 30626816873aSStefano Zampini IS tis; 30636816873aSStefano Zampini PetscInt schur_size; 30646816873aSStefano Zampini 3065df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 30666816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 3067df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 30686816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 30696816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 30706816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 30716816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 30726816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 3073d62866d3SStefano Zampini } 3074d62866d3SStefano Zampini } 3075a64d13efSStefano Zampini PetscFunctionReturn(0); 3076a64d13efSStefano Zampini } 3077a64d13efSStefano Zampini 3078304d26faSStefano Zampini 3079304d26faSStefano Zampini #undef __FUNCT__ 3080304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 3081684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 3082304d26faSStefano Zampini { 3083304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3084304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3085304d26faSStefano Zampini PC pc_temp; 3086304d26faSStefano Zampini Mat A_RR; 3087f4ddd8eeSStefano Zampini MatReuse reuse; 3088304d26faSStefano Zampini PetscScalar m_one = -1.0; 3089304d26faSStefano Zampini PetscReal value; 309004708bb6SStefano Zampini PetscInt n_D,n_R; 3091c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 3092304d26faSStefano Zampini PetscErrorCode ierr; 3093e604994aSStefano Zampini /* prefixes stuff */ 3094312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 3095e604994aSStefano Zampini size_t len; 3096304d26faSStefano Zampini 3097304d26faSStefano Zampini PetscFunctionBegin; 3098304d26faSStefano Zampini 3099e604994aSStefano Zampini /* compute prefixes */ 3100e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 3101e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 3102e604994aSStefano Zampini if (!pcbddc->current_level) { 3103e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3104e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 3105e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3106e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3107e604994aSStefano Zampini } else { 3108e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 3109312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 3110e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 3111e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 3112312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 3113312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 311434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 311534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 3116e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 3117e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 3118e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 3119e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 3120e604994aSStefano Zampini } 3121e604994aSStefano Zampini 3122304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 3123684f6988SStefano Zampini if (dirichlet) { 3124d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3125450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 31269a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 3127450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 3128a3df083aSStefano Zampini Mat A_IIn; 3129a3df083aSStefano Zampini 3130a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 3131a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 3132a3df083aSStefano Zampini pcis->A_II = A_IIn; 3133a3df083aSStefano Zampini } 3134450f8f5eSStefano Zampini } 31353301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 31363301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 3137964fefecSStefano Zampini } 3138ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 3139964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 3140304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 3141304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 3142304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 3143304d26faSStefano Zampini /* default */ 3144304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 3145e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 31469577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 3147304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 31489577ea80SStefano Zampini if (issbaij) { 31499577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 31509577ea80SStefano Zampini } else { 3151304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 31529577ea80SStefano Zampini } 3153304d26faSStefano Zampini /* Allow user's customization */ 3154304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 3155304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3156304d26faSStefano Zampini } 3157d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 3158b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3159df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3160d62866d3SStefano Zampini 3161df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 3162d5574798SStefano Zampini } 3163304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3164304d26faSStefano Zampini if (!n_D) { 3165304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 3166304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3167304d26faSStefano Zampini } 3168304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 3169304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 3170304d26faSStefano Zampini /* set ksp_D into pcis data */ 3171304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 3172304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 3173304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 3174684f6988SStefano Zampini } 3175304d26faSStefano Zampini 3176304d26faSStefano Zampini /* NEUMANN PROBLEM */ 3177684f6988SStefano Zampini A_RR = 0; 3178684f6988SStefano Zampini if (neumann) { 3179d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 318004708bb6SStefano Zampini PetscInt ibs,mbs; 318104708bb6SStefano Zampini PetscBool issbaij; 318204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3183f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 31848ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 3185f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 3186f4ddd8eeSStefano Zampini PetscInt nn_R; 318781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 3188f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3189f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 3190f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 3191f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 3192f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3193f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3194f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 3195727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 3196f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3197f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3198f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 3199f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 3200f4ddd8eeSStefano Zampini } 3201f4ddd8eeSStefano Zampini } 3202f4ddd8eeSStefano Zampini /* last check */ 3203d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 3204f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3205f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3206f4ddd8eeSStefano Zampini } 3207f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 3208f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 3209f4ddd8eeSStefano Zampini } 3210a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 3211af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 3212af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 321304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 321404708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 321504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 321604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 321704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 3218af732b37SStefano Zampini } else { 3219511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 32206816873aSStefano Zampini } 322104708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 322204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 322304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 322404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 322504708bb6SStefano Zampini } else { 3226511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 322704708bb6SStefano Zampini } 322804708bb6SStefano Zampini } 3229a00504b5SStefano Zampini /* extract A_RR */ 3230b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3231a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3232a00504b5SStefano Zampini 3233a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 323416e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3235a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 323616e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 323716e386b8SStefano Zampini } else { 3238a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 3239a00504b5SStefano Zampini } 3240a00504b5SStefano Zampini } else { 3241a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3242a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 3243a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 3244a00504b5SStefano Zampini } 3245a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 3246f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 324716e386b8SStefano Zampini } 32483301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 32493301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 32506816873aSStefano Zampini } 3251f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 3252304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 3253304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 3254304d26faSStefano Zampini /* default */ 3255304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 3256e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 3257304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 32589577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 32599577ea80SStefano Zampini if (issbaij) { 32609577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 32619577ea80SStefano Zampini } else { 3262304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 32639577ea80SStefano Zampini } 3264304d26faSStefano Zampini /* Allow user's customization */ 3265304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 3266304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 3267304d26faSStefano Zampini } 3268304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 3269304d26faSStefano Zampini if (!n_R) { 3270304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 3271304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 3272304d26faSStefano Zampini } 32735cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 3274df4d28bfSStefano Zampini /* Reuse solver if it is present */ 3275b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 3276df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3277d62866d3SStefano Zampini 3278df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 3279d62866d3SStefano Zampini } 3280304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 3281304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 3282684f6988SStefano Zampini } 3283304d26faSStefano Zampini 3284684f6988SStefano Zampini if (pcbddc->dbg_flag) { 3285684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32861575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3287684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3288684f6988SStefano Zampini } 3289c7017625SStefano Zampini 3290c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 3291c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 3292c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 3293c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 3294c7017625SStefano Zampini } 3295c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 3296c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 3297c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 3298c7017625SStefano Zampini } 3299c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 3300c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 3301c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 3302c7017625SStefano Zampini } 3303c7017625SStefano Zampini 3304c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 3305c7017625SStefano Zampini if (pcbddc->dbg_flag) { 3306684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 33070fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 33080fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 33090fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 33100fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 33110fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 3312e604994aSStefano 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); 3313c7017625SStefano Zampini if (check_corr[0]) { 3314c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 3315c7017625SStefano Zampini } 3316304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3317304d26faSStefano Zampini } 3318684f6988SStefano Zampini if (neumann) { /* Neumann */ 33190fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 33200fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 33210fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 33220fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 33230fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3324e604994aSStefano 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); 3325c7017625SStefano Zampini if (check_corr[1]) { 3326c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 3327c7017625SStefano Zampini } 3328304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3329304d26faSStefano Zampini } 3330684f6988SStefano Zampini } 33315cbda25cSStefano Zampini /* free Neumann problem's matrix */ 33325cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3333304d26faSStefano Zampini PetscFunctionReturn(0); 3334304d26faSStefano Zampini } 3335304d26faSStefano Zampini 3336304d26faSStefano Zampini #undef __FUNCT__ 3337ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 333880677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3339674ae819SStefano Zampini { 3340674ae819SStefano Zampini PetscErrorCode ierr; 3341674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3342be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3343b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 3344674ae819SStefano Zampini 3345674ae819SStefano Zampini PetscFunctionBegin; 3346b334f244SStefano Zampini if (!reuse_solver) { 334780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 334820c7b377SStefano Zampini } 334980677318SStefano Zampini if (!pcbddc->switch_static) { 335080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 335180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 335280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 335320c7b377SStefano Zampini } 3354b334f244SStefano Zampini if (!reuse_solver) { 335580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 335720c7b377SStefano Zampini } else { 3358df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3359be83ff47SStefano Zampini 3360df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3361df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 336220c7b377SStefano Zampini } 3363be83ff47SStefano Zampini } else { 336480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 336880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 336980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 337080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 337180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 337280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3373674ae819SStefano Zampini } 3374674ae819SStefano Zampini } 3375b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 337680677318SStefano Zampini if (applytranspose) { 337780677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 337880677318SStefano Zampini } else { 337980677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 338080677318SStefano Zampini } 3381be83ff47SStefano Zampini } else { 3382df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3383be83ff47SStefano Zampini 3384be83ff47SStefano Zampini if (applytranspose) { 3385df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3386be83ff47SStefano Zampini } else { 3387df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3388be83ff47SStefano Zampini } 3389be83ff47SStefano Zampini } 339080677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 339180677318SStefano Zampini if (!pcbddc->switch_static) { 3392b334f244SStefano Zampini if (!reuse_solver) { 339380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 339480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3395be83ff47SStefano Zampini } else { 3396df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3397be83ff47SStefano Zampini 3398df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3399df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3400be83ff47SStefano Zampini } 340180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 340280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 340380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 340480677318SStefano Zampini } 340580677318SStefano Zampini } else { 340680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 340980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 341180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 341280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 341380677318SStefano Zampini } 341480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3418674ae819SStefano Zampini } 3419674ae819SStefano Zampini PetscFunctionReturn(0); 3420674ae819SStefano Zampini } 3421674ae819SStefano Zampini 3422dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3423674ae819SStefano Zampini #undef __FUNCT__ 3424674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3425dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3426674ae819SStefano Zampini { 3427674ae819SStefano Zampini PetscErrorCode ierr; 3428674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3429674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3430674ae819SStefano Zampini const PetscScalar zero = 0.0; 3431674ae819SStefano Zampini 3432674ae819SStefano Zampini PetscFunctionBegin; 3433dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 34344fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3435dc359a40SStefano Zampini if (applytranspose) { 3436674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 34378eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3438dc359a40SStefano Zampini } else { 3439674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3440674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 344115aaf578SStefano Zampini } 34424fee134fSStefano Zampini } else { 34434fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 34444fee134fSStefano Zampini } 3445efc2fbd9SStefano Zampini 3446efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 34474f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3448efc2fbd9SStefano Zampini PetscScalar *array; 34494f1b2e48SStefano Zampini PetscInt j; 3450efc2fbd9SStefano Zampini 3451efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 34524f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3453efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3454efc2fbd9SStefano Zampini } 3455efc2fbd9SStefano Zampini 345612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 345712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 345812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 345912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 346012edc857SStefano Zampini 34619f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 346212edc857SStefano Zampini if (pcbddc->coarse_ksp) { 346351694757SStefano Zampini Mat coarse_mat; 3464964fefecSStefano Zampini Vec rhs,sol; 346551694757SStefano Zampini MatNullSpace nullsp; 346627b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3467964fefecSStefano Zampini 346827b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 346927b6a85dSStefano Zampini PC coarse_pc; 347027b6a85dSStefano Zampini 347127b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 347227b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 347327b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 347427b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 347527b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 347627b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 34773bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 347827b6a85dSStefano Zampini } 347927b6a85dSStefano Zampini } 3480964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3481964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 348251694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 348351694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 348451694757SStefano Zampini if (nullsp) { 348551694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 348651694757SStefano Zampini } 348712edc857SStefano Zampini if (applytranspose) { 34889a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 3489964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 34902701bc32SStefano Zampini } else { 34911f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 34922701bc32SStefano Zampini PC coarse_pc; 34932701bc32SStefano Zampini 34942701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 34952701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 34963e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 34972701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 349812edc857SStefano Zampini } else { 3499964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 350012edc857SStefano Zampini } 35012701bc32SStefano Zampini } 35021d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 350327b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 350427b6a85dSStefano Zampini PC coarse_pc; 350527b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 350627b6a85dSStefano Zampini 350727b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 350827b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 350927b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 35103bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 351127b6a85dSStefano Zampini } 351251694757SStefano Zampini if (nullsp) { 351351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 351451694757SStefano Zampini } 351512edc857SStefano Zampini } 3516674ae819SStefano Zampini 3517674ae819SStefano Zampini /* Local solution on R nodes */ 35184fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 351980677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 35209f00e9b4SStefano Zampini } 35219f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 35229f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 352312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3524674ae819SStefano Zampini 35254fee134fSStefano Zampini /* Sum contributions from the two levels */ 35264fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3527dc359a40SStefano Zampini if (applytranspose) { 3528dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3529dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3530dc359a40SStefano Zampini } else { 3531674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 35328eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3533dc359a40SStefano Zampini } 3534efc2fbd9SStefano Zampini /* store p0 */ 35354f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3536efc2fbd9SStefano Zampini PetscScalar *array; 35374f1b2e48SStefano Zampini PetscInt j; 3538efc2fbd9SStefano Zampini 3539efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 35404f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3541efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3542efc2fbd9SStefano Zampini } 35434fee134fSStefano Zampini } else { /* expand the coarse solution */ 35444fee134fSStefano Zampini if (applytranspose) { 35454fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35464fee134fSStefano Zampini } else { 35474fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 35484fee134fSStefano Zampini } 35494fee134fSStefano Zampini } 3550674ae819SStefano Zampini PetscFunctionReturn(0); 3551674ae819SStefano Zampini } 3552674ae819SStefano Zampini 3553674ae819SStefano Zampini #undef __FUNCT__ 3554674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 355512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3556674ae819SStefano Zampini { 3557674ae819SStefano Zampini PetscErrorCode ierr; 3558674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 355958da7f69SStefano Zampini PetscScalar *array; 356012edc857SStefano Zampini Vec from,to; 3561674ae819SStefano Zampini 3562674ae819SStefano Zampini PetscFunctionBegin; 356312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 356412edc857SStefano Zampini from = pcbddc->coarse_vec; 356512edc857SStefano Zampini to = pcbddc->vec1_P; 356612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 356712edc857SStefano Zampini Vec tvec; 356858da7f69SStefano Zampini 356958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 357058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 357112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 357258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 357358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 357458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 357512edc857SStefano Zampini } 357612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 357712edc857SStefano Zampini from = pcbddc->vec1_P; 357812edc857SStefano Zampini to = pcbddc->coarse_vec; 357912edc857SStefano Zampini } 358012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3581674ae819SStefano Zampini PetscFunctionReturn(0); 3582674ae819SStefano Zampini } 3583674ae819SStefano Zampini 3584674ae819SStefano Zampini #undef __FUNCT__ 3585674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 358612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3587674ae819SStefano Zampini { 3588674ae819SStefano Zampini PetscErrorCode ierr; 3589674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 359058da7f69SStefano Zampini PetscScalar *array; 359112edc857SStefano Zampini Vec from,to; 3592674ae819SStefano Zampini 3593674ae819SStefano Zampini PetscFunctionBegin; 359412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 359512edc857SStefano Zampini from = pcbddc->coarse_vec; 359612edc857SStefano Zampini to = pcbddc->vec1_P; 359712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 359812edc857SStefano Zampini from = pcbddc->vec1_P; 359912edc857SStefano Zampini to = pcbddc->coarse_vec; 360012edc857SStefano Zampini } 360112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 360212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 360312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 360412edc857SStefano Zampini Vec tvec; 360558da7f69SStefano Zampini 360612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 360758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 360858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 360958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 361058da7f69SStefano Zampini } 361158da7f69SStefano Zampini } else { 361258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 361358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 361412edc857SStefano Zampini } 361512edc857SStefano Zampini } 3616674ae819SStefano Zampini PetscFunctionReturn(0); 3617674ae819SStefano Zampini } 3618674ae819SStefano Zampini 3619984c4197SStefano Zampini /* uncomment for testing purposes */ 3620984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3621674ae819SStefano Zampini #undef __FUNCT__ 3622674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3623674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3624674ae819SStefano Zampini { 3625674ae819SStefano Zampini PetscErrorCode ierr; 3626674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3627674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3628674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3629984c4197SStefano Zampini /* one and zero */ 3630984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3631984c4197SStefano Zampini /* space to store constraints and their local indices */ 36329162d606SStefano Zampini PetscScalar *constraints_data; 36339162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 36349162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 36359162d606SStefano Zampini PetscInt *constraints_n; 3636984c4197SStefano Zampini /* iterators */ 3637b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3638984c4197SStefano Zampini /* BLAS integers */ 3639e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3640e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3641c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3642727cdba6SStefano Zampini /* reuse */ 36430e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 36440e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3645984c4197SStefano Zampini /* change of basis */ 3646b3d85658SStefano Zampini PetscBool qr_needed; 36479162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3648984c4197SStefano Zampini /* auxiliary stuff */ 364964efe560SStefano Zampini PetscInt *nnz,*is_indices; 36508a0068c3SStefano Zampini PetscInt ncc; 3651984c4197SStefano Zampini /* some quantities */ 365245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3653a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3654984c4197SStefano Zampini 3655674ae819SStefano Zampini PetscFunctionBegin; 36568e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 36578e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 36588e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 365916909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3660088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3661088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 36620e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 36630e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 36640e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36650e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 36660e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3667088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3668cf5a6209SStefano Zampini 3669cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 36709162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3671cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3672cf5a6209SStefano Zampini const Vec *nearnullvecs; 3673cf5a6209SStefano Zampini Vec *localnearnullsp; 3674cf5a6209SStefano Zampini PetscScalar *array; 3675cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3676cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3677674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3678b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3679674ae819SStefano Zampini PetscScalar *work; 3680674ae819SStefano Zampini PetscReal *singular_vals; 3681674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3682674ae819SStefano Zampini PetscReal *rwork; 3683674ae819SStefano Zampini #endif 3684674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3685674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3686674ae819SStefano Zampini #else 3687964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3688964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3689674ae819SStefano Zampini #endif 3690674ae819SStefano Zampini 3691674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3692d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3693e4d548c7SStefano Zampini /* print some info */ 36941f4df5f7SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs) { 3695e4d548c7SStefano Zampini PetscInt nv; 3696e4d548c7SStefano Zampini 3697*c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 3698e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3699e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3700e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3701e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3702e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3703e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3704e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3705e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3706e4d548c7SStefano Zampini } 3707e4d548c7SStefano Zampini 3708d06fc5fdSStefano Zampini /* free unneeded index sets */ 3709d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3710d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3711674ae819SStefano Zampini } 3712d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3713d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3714d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3715d06fc5fdSStefano Zampini } 3716d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3717d06fc5fdSStefano Zampini n_ISForEdges = 0; 3718d06fc5fdSStefano Zampini } 3719d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3720d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3721d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3722d06fc5fdSStefano Zampini } 3723d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3724d06fc5fdSStefano Zampini n_ISForFaces = 0; 3725d06fc5fdSStefano Zampini } 372670022509SStefano Zampini 3727674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3728674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3729674ae819SStefano Zampini if (nearnullsp) { 3730674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3731f4ddd8eeSStefano Zampini /* remove any stored info */ 3732f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3733f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3734f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3735f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3736f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3737473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3738f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3739f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3740f4ddd8eeSStefano Zampini } 3741984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3742984c4197SStefano Zampini nnsp_size = 0; 3743674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3744674ae819SStefano Zampini } 3745984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3746984c4197SStefano Zampini max_constraints = nnsp_size; 3747984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3748984c4197SStefano Zampini 3749674ae819SStefano Zampini /* 3750674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 37519162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 37529162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 37539162d606SStefano Zampini There can be multiple constraints per connected component 3754674ae819SStefano Zampini */ 3755674ae819SStefano Zampini n_vertices = 0; 3756674ae819SStefano Zampini if (ISForVertices) { 3757674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3758674ae819SStefano Zampini } 37599162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 37609162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 37619162d606SStefano Zampini 37629162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 37639162d606SStefano Zampini total_counts *= max_constraints; 3764674ae819SStefano Zampini total_counts += n_vertices; 37654641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 37669162d606SStefano Zampini 3767674ae819SStefano Zampini total_counts = 0; 3768674ae819SStefano Zampini max_size_of_constraint = 0; 3769674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 37709162d606SStefano Zampini IS used_is; 3771674ae819SStefano Zampini if (i<n_ISForEdges) { 37729162d606SStefano Zampini used_is = ISForEdges[i]; 3773674ae819SStefano Zampini } else { 37749162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3775674ae819SStefano Zampini } 37769162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3777674ae819SStefano Zampini total_counts += j; 3778674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3779674ae819SStefano Zampini } 37809162d606SStefano 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); 37819162d606SStefano Zampini 3782984c4197SStefano Zampini /* get local part of global near null space vectors */ 3783785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3784984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3785984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3786e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3787e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3788984c4197SStefano Zampini } 3789674ae819SStefano Zampini 3790242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3791242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3792a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3793242a89d7SStefano Zampini 3794984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3795a773dcb8SStefano Zampini if (!skip_lapack) { 3796674ae819SStefano Zampini PetscScalar temp_work; 3797911cabfeSStefano Zampini 3798674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3799984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3800785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3801785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3802785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3803674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3804785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3805674ae819SStefano Zampini #endif 3806674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3807c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3808c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3809674ae819SStefano Zampini lwork = -1; 3810674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3811674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3812c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3813674ae819SStefano Zampini #else 3814c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3815674ae819SStefano Zampini #endif 3816674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3817984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3818674ae819SStefano Zampini #else /* on missing GESVD */ 3819674ae819SStefano Zampini /* SVD */ 3820674ae819SStefano Zampini PetscInt max_n,min_n; 3821674ae819SStefano Zampini max_n = max_size_of_constraint; 3822984c4197SStefano Zampini min_n = max_constraints; 3823984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3824674ae819SStefano Zampini min_n = max_size_of_constraint; 3825984c4197SStefano Zampini max_n = max_constraints; 3826674ae819SStefano Zampini } 3827785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3828674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3829785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3830674ae819SStefano Zampini #endif 3831674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3832674ae819SStefano Zampini lwork = -1; 3833e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3834e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3835b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3836674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3837674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 38389162d606SStefano 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)); 3839674ae819SStefano Zampini #else 38409162d606SStefano 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)); 3841674ae819SStefano Zampini #endif 3842674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3843984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3844984c4197SStefano Zampini #endif /* on missing GESVD */ 3845674ae819SStefano Zampini /* Allocate optimal workspace */ 3846674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3847854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3848674ae819SStefano Zampini } 3849674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3850674ae819SStefano Zampini total_counts = 0; 38519162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 38529162d606SStefano Zampini constraints_data_ptr[0] = 0; 3853674ae819SStefano Zampini /* vertices */ 38549162d606SStefano Zampini if (n_vertices) { 3855674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 38569162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3857674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 38589162d606SStefano Zampini constraints_n[total_counts] = 1; 38599162d606SStefano Zampini constraints_data[total_counts] = 1.0; 38609162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 38619162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3862674ae819SStefano Zampini total_counts++; 3863674ae819SStefano Zampini } 3864674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3865674ae819SStefano Zampini n_vertices = total_counts; 3866674ae819SStefano Zampini } 3867984c4197SStefano Zampini 3868674ae819SStefano Zampini /* edges and faces */ 38699162d606SStefano Zampini total_counts_cc = total_counts; 3870911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 38719162d606SStefano Zampini IS used_is; 38729162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 38739162d606SStefano Zampini 3874911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 38759162d606SStefano Zampini used_is = ISForEdges[ncc]; 3876984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3877674ae819SStefano Zampini } else { 38789162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3879984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3880674ae819SStefano Zampini } 3881674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 38829162d606SStefano Zampini 38839162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 38849162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3885984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3886984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3887674ae819SStefano Zampini if (nnsp_has_cnst) { 38885b08dc53SStefano Zampini PetscScalar quad_value; 38899162d606SStefano Zampini 38909162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 38919162d606SStefano Zampini idxs_copied = PETSC_TRUE; 38929162d606SStefano Zampini 3893a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3894674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3895a773dcb8SStefano Zampini } else { 3896a773dcb8SStefano Zampini quad_value = 1.0; 3897a773dcb8SStefano Zampini } 3898674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 38999162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3900674ae819SStefano Zampini } 39019162d606SStefano Zampini temp_constraints++; 3902674ae819SStefano Zampini total_counts++; 3903674ae819SStefano Zampini } 3904674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3905984c4197SStefano Zampini PetscReal real_value; 39069162d606SStefano Zampini PetscScalar *ptr_to_data; 39079162d606SStefano Zampini 3908984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 39099162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3910674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 39119162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3912674ae819SStefano Zampini } 3913984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3914984c4197SStefano Zampini /* check if array is null on the connected component */ 3915e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39169162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 39175b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3918674ae819SStefano Zampini temp_constraints++; 3919674ae819SStefano Zampini total_counts++; 39209162d606SStefano Zampini if (!idxs_copied) { 39219162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 39229162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3923674ae819SStefano Zampini } 3924674ae819SStefano Zampini } 39259162d606SStefano Zampini } 39269162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 392745a1bb75SStefano Zampini valid_constraints = temp_constraints; 3928eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3929a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 39309162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 39319162d606SStefano Zampini 39329162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3933a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39349162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3935a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 39369162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3937a773dcb8SStefano Zampini } else { /* perform SVD */ 3938984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 39399162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3940674ae819SStefano Zampini 3941674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3942984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3943984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3944984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3945984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3946984c4197SStefano Zampini from that computed using LAPACKgesvd 3947984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3948984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3949984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3950674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3951e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3952984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3953674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3954674ae819SStefano Zampini for (k=0;k<j+1;k++) { 39559162d606SStefano 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)); 3956674ae819SStefano Zampini } 3957674ae819SStefano Zampini } 3958e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3959e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3960e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3961674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3962c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3963674ae819SStefano Zampini #else 3964c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3965674ae819SStefano Zampini #endif 3966674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3967984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3968984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3969674ae819SStefano Zampini j = 0; 3970984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3971674ae819SStefano Zampini total_counts = total_counts-j; 397245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3973e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3974c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3975c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3976c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3977c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3978c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3979c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3980674ae819SStefano Zampini if (j<temp_constraints) { 3981984c4197SStefano Zampini PetscInt ii; 3982984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3983674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39849162d606SStefano 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)); 3985674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3986984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3987674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 39889162d606SStefano 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]; 3989674ae819SStefano Zampini } 3990674ae819SStefano Zampini } 3991674ae819SStefano Zampini } 3992674ae819SStefano Zampini #else /* on missing GESVD */ 3993e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3994e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3995b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3996674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3997674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 39989162d606SStefano 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)); 3999674ae819SStefano Zampini #else 40009162d606SStefano 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)); 4001674ae819SStefano Zampini #endif 4002984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 4003674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4004984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 4005e310c8b4SStefano Zampini k = temp_constraints; 4006e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 4007674ae819SStefano Zampini j = 0; 4008e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 400945a1bb75SStefano Zampini valid_constraints = k-j; 4010911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 4011984c4197SStefano Zampini #endif /* on missing GESVD */ 4012674ae819SStefano Zampini } 4013a773dcb8SStefano Zampini } 40149162d606SStefano Zampini /* update pointers information */ 40159162d606SStefano Zampini if (valid_constraints) { 40169162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 40179162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 40189162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 40199162d606SStefano Zampini /* set change_of_basis flag */ 402045a1bb75SStefano Zampini if (boolforchange) { 4021b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 40229162d606SStefano Zampini } 4023b3d85658SStefano Zampini total_counts_cc++; 402445a1bb75SStefano Zampini } 402545a1bb75SStefano Zampini } 4026984c4197SStefano Zampini /* free workspace */ 40278f1c130eSStefano Zampini if (!skip_lapack) { 4028984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4029984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4030984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 4031984c4197SStefano Zampini #endif 4032984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 4033984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4034984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 4035984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 4036984c4197SStefano Zampini #endif 4037984c4197SStefano Zampini } 4038984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4039984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 4040984c4197SStefano Zampini } 4041984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 4042cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 4043cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4044cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4045cf5a6209SStefano Zampini } 4046cf5a6209SStefano Zampini if (n_ISForFaces) { 4047cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4048cf5a6209SStefano Zampini } 4049cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4050cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4051cf5a6209SStefano Zampini } 4052cf5a6209SStefano Zampini if (n_ISForEdges) { 4053cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4054cf5a6209SStefano Zampini } 4055cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 405608122e43SStefano Zampini } else { 405708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4058984c4197SStefano Zampini 405908122e43SStefano Zampini total_counts = 0; 406008122e43SStefano Zampini n_vertices = 0; 4061d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 4062d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 406308122e43SStefano Zampini } 406408122e43SStefano Zampini max_constraints = 0; 40659162d606SStefano Zampini total_counts_cc = 0; 406608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 406708122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 40689162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 406908122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 407008122e43SStefano Zampini } 40719162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 40729162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 40739162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 40749162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 407574d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 40769162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 40779162d606SStefano Zampini total_counts_cc = 0; 40789162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 40799162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 40809162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 408108122e43SStefano Zampini } 408208122e43SStefano Zampini } 40839162d606SStefano Zampini #if 0 40849162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 40859162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 40869162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 40879162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 40889162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 40899162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 40909162d606SStefano Zampini } 40919162d606SStefano Zampini printf("\n"); 40929162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 40939162d606SStefano Zampini } 40941b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 40958bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 40961b968477SStefano Zampini } 40971b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 40988bec7fa6SStefano 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]); 40991b968477SStefano Zampini } 410008122e43SStefano Zampini #endif 410108122e43SStefano Zampini 41028bec7fa6SStefano Zampini max_size_of_constraint = 0; 41039162d606SStefano 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]); 41049162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 410508122e43SStefano Zampini /* Change of basis */ 4106b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 410708122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 410808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 410908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 4110b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 411108122e43SStefano Zampini } 411208122e43SStefano Zampini } 411308122e43SStefano Zampini } 411408122e43SStefano Zampini } 4115984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 41164f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 411708122e43SStefano Zampini 41189162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 41199162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 41206c4ed002SBarry 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); 4121674ae819SStefano Zampini 4122674ae819SStefano Zampini /* Create constraint matrix */ 4123674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 412416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 4125984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 4126984c4197SStefano Zampini 4127984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 4128a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 4129a717540cSStefano Zampini qr_needed = PETSC_FALSE; 413074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 4131984c4197SStefano Zampini total_primal_vertices=0; 4132b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 41339162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 41349162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 413572b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 41369162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 4137b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 413864efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 41399162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41409162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 4141a717540cSStefano Zampini } 4142b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 414391af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 4144a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 4145a717540cSStefano Zampini qr_needed = PETSC_TRUE; 4146a717540cSStefano Zampini } 4147fa434743SStefano Zampini } else { 4148b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 4149fa434743SStefano Zampini } 4150a717540cSStefano Zampini } 4151b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 4152b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 4153674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 415470022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4155b3d85658SStefano Zampini 41564f1b2e48SStefano 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); 41570e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 41580e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 4159984c4197SStefano Zampini 4160984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 416174d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 4162785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 4163984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 416474d5cdf7SStefano Zampini 4165984c4197SStefano Zampini j = total_primal_vertices; 416674d5cdf7SStefano Zampini total_counts = total_primal_vertices; 4167b3d85658SStefano Zampini cum = total_primal_vertices; 41689162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41694641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 4170b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 4171b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 4172b3d85658SStefano Zampini cum++; 41739162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 417474d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 417574d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 417674d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 417774d5cdf7SStefano Zampini } 41789162d606SStefano Zampini j += constraints_n[i]; 4179674ae819SStefano Zampini } 4180674ae819SStefano Zampini } 4181674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 4182674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 4183088faed8SStefano Zampini 4184674ae819SStefano Zampini /* set values in constraint matrix */ 4185984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 41860e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4187674ae819SStefano Zampini } 4188984c4197SStefano Zampini total_counts = total_primal_vertices; 41899162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 41904641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 41919162d606SStefano Zampini PetscInt *cols; 41929162d606SStefano Zampini 41939162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 41949162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 41959162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 41969162d606SStefano Zampini PetscInt row = total_counts+k; 41979162d606SStefano Zampini PetscScalar *vals; 41989162d606SStefano Zampini 41999162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 42009162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 42019162d606SStefano Zampini } 42029162d606SStefano Zampini total_counts += constraints_n[i]; 4203674ae819SStefano Zampini } 4204674ae819SStefano Zampini } 4205674ae819SStefano Zampini /* assembling */ 4206674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4207674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4208088faed8SStefano Zampini 4209984c4197SStefano Zampini /* 42106a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4211984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 4212f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 4213984c4197SStefano Zampini */ 4214674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 4215674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 4216026de310SStefano Zampini /* dual and primal dofs on a single cc */ 4217984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 4218984c4197SStefano Zampini /* working stuff for GEQRF */ 421981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 4220984c4197SStefano Zampini PetscBLASInt lqr_work; 4221984c4197SStefano Zampini /* working stuff for UNGQR */ 4222984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 4223984c4197SStefano Zampini PetscBLASInt lgqr_work; 4224984c4197SStefano Zampini /* working stuff for TRTRS */ 4225984c4197SStefano Zampini PetscScalar *trs_rhs; 42263f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 4227984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 4228984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 4229984c4197SStefano Zampini PetscScalar *start_vals; 4230984c4197SStefano Zampini /* working stuff for values insertion */ 42314641a718SStefano Zampini PetscBT is_primal; 423264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 4233906d46d4SStefano Zampini /* matrix sizes */ 4234906d46d4SStefano Zampini PetscInt global_size,local_size; 4235906d46d4SStefano Zampini /* temporary change of basis */ 4236906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 4237cf5a6209SStefano Zampini /* extra space for debugging */ 4238cf5a6209SStefano Zampini PetscScalar *dbg_work; 4239984c4197SStefano Zampini 4240906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 4241906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 424216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 4243bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 4244906d46d4SStefano Zampini /* nonzeros for local mat */ 4245bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 42461dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4247bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 42481dd7afcfSStefano Zampini } else { 42491dd7afcfSStefano Zampini const PetscInt *ii; 42501dd7afcfSStefano Zampini PetscInt n; 42511dd7afcfSStefano Zampini PetscBool flg_row; 42521dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42531dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 42541dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 42551dd7afcfSStefano Zampini } 42569162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 4257a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 42589162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 4259a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 42609162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 4261a717540cSStefano Zampini } else { 42629162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 42639162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 4264a717540cSStefano Zampini } 4265a717540cSStefano Zampini } 4266a717540cSStefano Zampini } 4267906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 4268bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 42691dd7afcfSStefano Zampini /* Set interior change in the matrix */ 42701dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 4271bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 4272906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 4273a717540cSStefano Zampini } 42741dd7afcfSStefano Zampini } else { 42751dd7afcfSStefano Zampini const PetscInt *ii,*jj; 42761dd7afcfSStefano Zampini PetscScalar *aa; 42771dd7afcfSStefano Zampini PetscInt n; 42781dd7afcfSStefano Zampini PetscBool flg_row; 42791dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42801dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42811dd7afcfSStefano Zampini for (i=0;i<n;i++) { 42821dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 42831dd7afcfSStefano Zampini } 42841dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 42851dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 42861dd7afcfSStefano Zampini } 4287a717540cSStefano Zampini 4288a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4289a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4290a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4291a717540cSStefano Zampini } 4292a717540cSStefano Zampini 4293a717540cSStefano Zampini 4294a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4295a717540cSStefano Zampini /* 4296a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4297a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4298a717540cSStefano Zampini 4299a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4300a717540cSStefano Zampini 4301a6b551f4SStefano 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) 4302a6b551f4SStefano Zampini 4303a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4304a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4305a717540cSStefano Zampini | ... | 4306a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4307a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4308a717540cSStefano Zampini 4309a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4310a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4311a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4312a6b551f4SStefano Zampini 4313a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4314a717540cSStefano Zampini */ 4315a717540cSStefano Zampini if (qr_needed) { 4316984c4197SStefano Zampini /* space to store Q */ 4317854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4318984c4197SStefano Zampini /* first we issue queries for optimal work */ 43193f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43203f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 43213f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4322984c4197SStefano Zampini lqr_work = -1; 43233f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4324984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4325984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4326785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4327984c4197SStefano Zampini lgqr_work = -1; 43283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 43293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 43303f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 43313f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43323f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 43333f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4334984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4335984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4336785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4337984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4338785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4339984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4340785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4341a717540cSStefano Zampini /* allocating workspace for check */ 4342a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4343cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4344a717540cSStefano Zampini } 4345a717540cSStefano Zampini } 4346984c4197SStefano Zampini /* array to store whether a node is primal or not */ 43474641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4348473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 43490e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 43506c4ed002SBarry 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); 435139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 435239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 435339e2fb2aSStefano Zampini } 435439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4355984c4197SStefano Zampini 4356a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 43579162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 43589162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 43594641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4360984c4197SStefano Zampini /* get constraint info */ 43619162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4362984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4363984c4197SStefano Zampini 4364984c4197SStefano Zampini if (pcbddc->dbg_flag) { 43659162d606SStefano 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); 4366674ae819SStefano Zampini } 4367984c4197SStefano Zampini 4368fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4369a717540cSStefano Zampini 4370a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4371a717540cSStefano Zampini if (pcbddc->dbg_flag) { 43729162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4373a717540cSStefano Zampini } 4374984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 43759162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4376984c4197SStefano Zampini 4377984c4197SStefano Zampini /* compute QR decomposition of constraints */ 43783f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 43793f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4381674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43823f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4383984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4384674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4385984c4197SStefano Zampini 4386984c4197SStefano Zampini /* explictly compute R^-T */ 4387984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4388984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 43893f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 43903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 43913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 43923f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4393984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 43943f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4395984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4396984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4397984c4197SStefano Zampini 4398a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 43993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44013f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 44023f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4403984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44043f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4405984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4406984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4407984c4197SStefano Zampini 4408984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4409984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4410984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 44113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 44123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 44133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 44143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 44153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 44163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4417984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 44189162d606SStefano 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)); 4419984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 44209162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4421984c4197SStefano Zampini 4422984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 44239162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4424984c4197SStefano Zampini /* insert cols for primal dofs */ 4425984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4426984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 44279162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4428906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4429984c4197SStefano Zampini } 4430984c4197SStefano Zampini /* insert cols for dual dofs */ 4431984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 44329162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4433984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 44349162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4435906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4436984c4197SStefano Zampini j++; 4437674ae819SStefano Zampini } 4438674ae819SStefano Zampini } 4439984c4197SStefano Zampini 4440984c4197SStefano Zampini /* check change of basis */ 4441984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4442984c4197SStefano Zampini PetscInt ii,jj; 4443984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4444c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4445c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4446c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4447c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4448c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4449c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4450984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4451cf5a6209SStefano 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)); 4452984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4453984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4454984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4455cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4456cf5a6209SStefano 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; 4457674ae819SStefano Zampini } 4458674ae819SStefano Zampini } 4459984c4197SStefano Zampini if (!valid_qr) { 446022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4461984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4462984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4463cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4464cf5a6209SStefano 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])); 4465674ae819SStefano Zampini } 4466cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4467cf5a6209SStefano 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])); 4468984c4197SStefano Zampini } 4469984c4197SStefano Zampini } 4470984c4197SStefano Zampini } 4471674ae819SStefano Zampini } else { 447222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4473674ae819SStefano Zampini } 4474674ae819SStefano Zampini } 4475a717540cSStefano Zampini } else { /* simple transformation block */ 4476a717540cSStefano Zampini PetscInt row,col; 4477a6b551f4SStefano Zampini PetscScalar val,norm; 4478a6b551f4SStefano Zampini 4479a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 44809162d606SStefano 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)); 4481a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 44829162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 44839162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4484bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 44859162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4486906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 44879162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4488a717540cSStefano Zampini } else { 4489a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 44909162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4491a717540cSStefano Zampini if (row != col) { 44929162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4493a717540cSStefano Zampini } else { 44949162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4495a717540cSStefano Zampini } 4496906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4497a717540cSStefano Zampini } 4498a717540cSStefano Zampini } 4499a717540cSStefano Zampini } 450098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 450122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4502a717540cSStefano Zampini } 4503674ae819SStefano Zampini } 4504984c4197SStefano Zampini } else { 4505984c4197SStefano Zampini if (pcbddc->dbg_flag) { 45069162d606SStefano 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); 4507674ae819SStefano Zampini } 4508674ae819SStefano Zampini } 4509674ae819SStefano Zampini } 4510a717540cSStefano Zampini 4511a717540cSStefano Zampini /* free workspace */ 4512a717540cSStefano Zampini if (qr_needed) { 4513984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4514cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4515984c4197SStefano Zampini } 4516984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4517984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4518984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4519984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4520984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4521674ae819SStefano Zampini } 4522a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4523906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4524906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4525906d46d4SStefano Zampini 4526906d46d4SStefano Zampini /* assembling of global change of variable */ 452788c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4528bbb9e6c6SStefano Zampini Mat tmat; 452916f15bc4SStefano Zampini PetscInt bs; 453016f15bc4SStefano Zampini 4531906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4532906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4533bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4534bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4535bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4536bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 453716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 453816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4539906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4540bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4541bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4542bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4543bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4544bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4545e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4546e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4547bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4548bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 454988c03ad3SStefano Zampini 4550906d46d4SStefano Zampini /* check */ 4551906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4552906d46d4SStefano Zampini PetscReal error; 4553906d46d4SStefano Zampini Vec x,x_change; 4554906d46d4SStefano Zampini 4555906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4556906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4557906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4558906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4559e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4560e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4561bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4562e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4563e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4564906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4565906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4566906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4567906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4568bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4569906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4570906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4571906d46d4SStefano Zampini } 4572b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4573b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4574b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4575bf3a8328SStefano Zampini 45769a962809SStefano 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); 4577b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 4578ac632422SStefano Zampini Mat S_new,tmat; 4579bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4580bbb9e6c6SStefano Zampini 4581bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 45826816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4583bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4584bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4585bf3a8328SStefano Zampini IS is_V; 4586b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4587b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4588b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4589b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4590b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4591bf3a8328SStefano Zampini } 4592bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4593ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4594b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4595ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4596bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4597bf3a8328SStefano Zampini const PetscScalar *array; 4598bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4599bf3a8328SStefano Zampini PetscInt i,n_V; 4600bf3a8328SStefano Zampini 4601b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4602b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4603b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4604b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4605b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4606b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4607b087196eSStefano Zampini PetscScalar val; 4608b087196eSStefano Zampini PetscInt idx; 4609b087196eSStefano Zampini 4610b087196eSStefano Zampini idx = idxs_V[i]; 4611b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4612b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4613b087196eSStefano Zampini } 4614b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4615b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4616bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4617bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4618bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4619bf3a8328SStefano Zampini } 4620ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4621ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4622ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4623ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4624b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4625ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4626bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4627b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4628bf3a8328SStefano Zampini } 4629ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4630ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4631ac632422SStefano Zampini } 4632b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 463388c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4634b96c3477SStefano Zampini } 4635c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4636b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 4637c9db6a07SStefano Zampini PetscInt i; 4638c9db6a07SStefano Zampini 4639c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4640c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4641c9db6a07SStefano Zampini } 4642c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4643c9db6a07SStefano Zampini } 4644b96c3477SStefano Zampini } 464516909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 464616909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 464716909a7fSStefano Zampini } else { 4648906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 464916909a7fSStefano Zampini } 46501dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 465127b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 465272b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 465372b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 465472b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 465572b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 465672b8c272SStefano Zampini } 46571dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 465827b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4659b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4660b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4661906d46d4SStefano Zampini } else { 46621dd7afcfSStefano Zampini Mat benign_global = NULL; 466327b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 46641dd7afcfSStefano Zampini Mat tmat; 46651dd7afcfSStefano Zampini 46661dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 46671dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 46681dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 46691dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46701dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 46711dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 46721dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46731dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 46741dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 46751dd7afcfSStefano Zampini if (pcbddc->benign_change) { 46761dd7afcfSStefano Zampini Mat M; 46771dd7afcfSStefano Zampini 46781dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 46791dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 46801dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 46811dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4682906d46d4SStefano Zampini } else { 46831dd7afcfSStefano Zampini Mat eye; 46841dd7afcfSStefano Zampini PetscScalar *array; 46851dd7afcfSStefano Zampini 46861dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46871dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 46881dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 46891dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4690906d46d4SStefano Zampini } 46911dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 46921dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46931dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 46941dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 46951dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 46961dd7afcfSStefano Zampini } 46971dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 46981dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 46991dd7afcfSStefano Zampini } 47001dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 47011dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 47021dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 470327b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 47041dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 47051dd7afcfSStefano Zampini } 47061dd7afcfSStefano Zampini } 470716909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 470816909a7fSStefano Zampini IS is_global; 470916909a7fSStefano Zampini const PetscInt *gidxs; 471016909a7fSStefano Zampini 471116909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 471216909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 471316909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 471416909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 471516909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 471616909a7fSStefano Zampini } 47171dd7afcfSStefano Zampini } 47181dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 47191dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4720b9b85e73SStefano Zampini } 4721a717540cSStefano Zampini 472272b8c272SStefano Zampini if (!pcbddc->fake_change) { 47234f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 47244f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 47254f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 47264f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4727019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4728019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4729019a44ceSStefano Zampini pcbddc->local_primal_size++; 4730019a44ceSStefano Zampini } 4731019a44ceSStefano Zampini 4732019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4733727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4734727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 47359f47a83aSStefano 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); 4736c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 47370e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 47389f47a83aSStefano 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); 4739727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4740727cdba6SStefano Zampini } 47410e6343abSStefano Zampini } 4742727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4743b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 474472b8c272SStefano Zampini } 474572b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4746727cdba6SStefano Zampini 4747a717540cSStefano Zampini /* flush dbg viewer */ 4748b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4749b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4750b8ffe317SStefano Zampini } 4751a717540cSStefano Zampini 4752e310c8b4SStefano Zampini /* free workspace */ 4753a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 47544641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 475508122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 47569162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 47579162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 475808122e43SStefano Zampini } else { 47599162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 47609162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 47619162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 476208122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 476308122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 47649162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 47659162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 476608122e43SStefano Zampini } 4767674ae819SStefano Zampini PetscFunctionReturn(0); 4768674ae819SStefano Zampini } 4769674ae819SStefano Zampini 4770674ae819SStefano Zampini #undef __FUNCT__ 4771674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4772674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4773674ae819SStefano Zampini { 477471582508SStefano Zampini ISLocalToGlobalMapping map; 4775674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4776674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 477714f95afaSStefano Zampini PetscInt ierr,i,N; 4778674ae819SStefano Zampini 4779674ae819SStefano Zampini PetscFunctionBegin; 4780*c8272957SStefano Zampini if (pcbddc->graphanalyzed && !pcbddc->recompute_topography) PetscFunctionReturn(0); 47818e61c736SStefano Zampini /* Reset previously computed graph */ 47828e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4783674ae819SStefano Zampini /* Init local Graph struct */ 47847fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 478571582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 478671582508SStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N);CHKERRQ(ierr); 4787674ae819SStefano Zampini 4788575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 47899a962809SStefano 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); 47909577ea80SStefano Zampini 4791674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4792d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 47934d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 47944d379d7bSStefano Zampini PetscInt nvtxs; 4795e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4796674ae819SStefano Zampini 47972fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 47982fffb893SStefano Zampini if (flg_row) { 47994d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4800b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 48012fffb893SStefano Zampini } 48022fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4803674ae819SStefano Zampini } 48049b28b941SStefano Zampini if (pcbddc->dbg_flag) { 48059b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4806674ae819SStefano Zampini } 4807674ae819SStefano Zampini 4808674ae819SStefano Zampini /* Setup of Graph */ 48094b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 481014f95afaSStefano 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); 4811674ae819SStefano Zampini 48124f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 48134f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 48144f1b2e48SStefano Zampini PetscInt *local_subs; 48154f1b2e48SStefano Zampini 48164f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 48174f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 48184f1b2e48SStefano Zampini const PetscInt *idxs; 48194f1b2e48SStefano Zampini PetscInt nl,j; 48204f1b2e48SStefano Zampini 48214f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 48224f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 482371582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 48244f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 48254f1b2e48SStefano Zampini } 48264f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 48274f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 48284f1b2e48SStefano Zampini } 48294f1b2e48SStefano Zampini 4830674ae819SStefano Zampini /* Graph's connected components analysis */ 4831674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 483271582508SStefano Zampini 483371582508SStefano Zampini /* set flag indicating analysis has been done */ 483471582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 4835674ae819SStefano Zampini PetscFunctionReturn(0); 4836674ae819SStefano Zampini } 4837674ae819SStefano Zampini 48389a7d3425SStefano Zampini #undef __FUNCT__ 48399a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 48409a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 48419a7d3425SStefano Zampini { 48429a7d3425SStefano Zampini PetscInt i,j; 48439a7d3425SStefano Zampini PetscScalar *alphas; 48449a7d3425SStefano Zampini PetscErrorCode ierr; 48459a7d3425SStefano Zampini 48469a7d3425SStefano Zampini PetscFunctionBegin; 4847785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 48489a7d3425SStefano Zampini for (i=0;i<n;i++) { 48499a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 4850669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 4851669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 4852669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 48539a7d3425SStefano Zampini } 48549a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 48559a7d3425SStefano Zampini PetscFunctionReturn(0); 48569a7d3425SStefano Zampini } 48579a7d3425SStefano Zampini 4858e7931f94SStefano Zampini #undef __FUNCT__ 485970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 486057de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4861e7931f94SStefano Zampini { 486257de7509SStefano Zampini Mat A; 4863e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4864e7931f94SStefano Zampini PetscMPIInt size,rank,color; 486552e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 486652e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 486727b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 486857de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 486927b6a85dSStefano Zampini PetscInt xadj_count, *count; 487027b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 487127b6a85dSStefano Zampini PetscSubcomm psubcomm; 487227b6a85dSStefano Zampini MPI_Comm subcomm; 487352e5ac9dSStefano Zampini PetscErrorCode ierr; 4874a57a6d2fSStefano Zampini 4875e7931f94SStefano Zampini PetscFunctionBegin; 487657de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 487757de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 487857de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 487957de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 488057de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 488157de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 488257de7509SStefano Zampini 488357de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 488457de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 488557de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 488657de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 488757de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 488857de7509SStefano Zampini im_active = !!(n); 488957de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 489057de7509SStefano Zampini void_procs = size - active_procs; 489157de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 489257de7509SStefano Zampini if (void_procs) { 489357de7509SStefano Zampini PetscInt ncand; 489457de7509SStefano Zampini 489557de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 489657de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 489757de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 489857de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 489957de7509SStefano Zampini if (!procs_candidates[i]) { 490057de7509SStefano Zampini procs_candidates[ncand++] = i; 490157de7509SStefano Zampini } 490257de7509SStefano Zampini } 490357de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 490457de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 490557de7509SStefano Zampini } 490657de7509SStefano Zampini 490714f0bfb9SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix 490814f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 490914f0bfb9SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1) { 491014f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 491114f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 491214f0bfb9SStefano Zampini else dest = rank; 491357de7509SStefano Zampini if (im_active) { 491457de7509SStefano Zampini issize = 1; 491557de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 491614f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 491757de7509SStefano Zampini } else { 491814f0bfb9SStefano Zampini isidx = dest; 491957de7509SStefano Zampini } 492057de7509SStefano Zampini } else { 492157de7509SStefano Zampini issize = 0; 492257de7509SStefano Zampini isidx = -1; 492357de7509SStefano Zampini } 492457de7509SStefano Zampini *n_subdomains = active_procs; 492557de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 4926daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 492757de7509SStefano Zampini PetscFunctionReturn(0); 492857de7509SStefano Zampini } 4929c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4930c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 493127b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 4932e7931f94SStefano Zampini 4933e7931f94SStefano Zampini /* Get info on mapping */ 49343bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 49353bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4936e7931f94SStefano Zampini 4937e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4938785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4939e7931f94SStefano Zampini xadj[0] = 0; 4940e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4941785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4942785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 494327b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 494427b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 494527b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 494627b6a85dSStefano Zampini count[shared[i][j]] += 1; 4947e7931f94SStefano Zampini 494827b6a85dSStefano Zampini xadj_count = 0; 49492b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 495027b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 495127b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 4952d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4953d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4954d023bfaeSStefano Zampini xadj_count++; 495527b6a85dSStefano Zampini break; 495627b6a85dSStefano Zampini } 4957e7931f94SStefano Zampini } 4958e7931f94SStefano Zampini } 4959d023bfaeSStefano Zampini xadj[1] = xadj_count; 496027b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 49613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4962e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4963e7931f94SStefano Zampini 49643837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4965e7931f94SStefano Zampini 496627b6a85dSStefano Zampini /* Restrict work on active processes only */ 496727b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 496827b6a85dSStefano Zampini if (void_procs) { 496927b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 497027b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 497127b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 497227b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 497327b6a85dSStefano Zampini } else { 497427b6a85dSStefano Zampini psubcomm = NULL; 497527b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 497627b6a85dSStefano Zampini } 497727b6a85dSStefano Zampini 497827b6a85dSStefano Zampini v_wgt = NULL; 497927b6a85dSStefano Zampini if (!color) { 4980e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4981e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4982e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4983c8587f34SStefano Zampini } else { 498452e5ac9dSStefano Zampini Mat subdomain_adj; 498552e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 498652e5ac9dSStefano Zampini MatPartitioning partitioner; 498727b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 498852e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 498957de7509SStefano Zampini PetscMPIInt size; 4990b0c7d250SStefano Zampini PetscBool aggregate; 4991b0c7d250SStefano Zampini 499227b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 499327b6a85dSStefano Zampini if (void_procs) { 499427b6a85dSStefano Zampini PetscInt prank = rank; 4995785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 499627b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 4997e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4998e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4999c8587f34SStefano Zampini } 5000e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 500127b6a85dSStefano Zampini } else { 500227b6a85dSStefano Zampini oldranks = NULL; 500327b6a85dSStefano Zampini } 5004b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 500527b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 5006b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 5007b0c7d250SStefano Zampini PetscMPIInt nrank; 5008b0c7d250SStefano Zampini PetscScalar *vals; 5009b0c7d250SStefano Zampini 501027b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 5011b0c7d250SStefano Zampini lrows = 0; 5012b0c7d250SStefano Zampini if (nrank<redprocs) { 5013b0c7d250SStefano Zampini lrows = size/redprocs; 5014b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 5015b0c7d250SStefano Zampini } 501627b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 5017b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 5018b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5019b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 5020b0c7d250SStefano Zampini row = nrank; 5021b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 5022b0c7d250SStefano Zampini cols = adjncy; 5023b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 5024b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 5025b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 5026b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5027b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 502852e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 502952e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 503052e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 5031b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 503227b6a85dSStefano Zampini if (use_vwgt) { 503327b6a85dSStefano Zampini Vec v; 503427b6a85dSStefano Zampini const PetscScalar *array; 503527b6a85dSStefano Zampini PetscInt nl; 503627b6a85dSStefano Zampini 503727b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 503827b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 503927b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 504027b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 504127b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 504227b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 504327b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 504422db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 504527b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 504627b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 504727b6a85dSStefano Zampini } 5048b0c7d250SStefano Zampini } else { 504927b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 505027b6a85dSStefano Zampini if (use_vwgt) { 505127b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 505227b6a85dSStefano Zampini v_wgt[0] = local_size; 505327b6a85dSStefano Zampini } 5054b0c7d250SStefano Zampini } 505522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 5056e7931f94SStefano Zampini 5057e7931f94SStefano Zampini /* Partition */ 505827b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 5059e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 506027b6a85dSStefano Zampini if (v_wgt) { 5061e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 5062c8587f34SStefano Zampini } 506357de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 506457de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 5065e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 5066e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 506722b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 5068e7931f94SStefano Zampini 506952e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 50706583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 507152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 507252e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 507357de7509SStefano Zampini if (!aggregate) { 507457de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 507527b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 507627b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 507727b6a85dSStefano Zampini #endif 507857de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 507927b6a85dSStefano Zampini } else if (oldranks) { 5080b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 508127b6a85dSStefano Zampini } else { 508227b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 508357de7509SStefano Zampini } 508428143c3dSStefano Zampini } else { 5085b0c7d250SStefano Zampini PetscInt idxs[1]; 5086b0c7d250SStefano Zampini PetscMPIInt tag; 5087b0c7d250SStefano Zampini MPI_Request *reqs; 5088b0c7d250SStefano Zampini 5089b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 5090b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 5091b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 509227b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 509328143c3dSStefano Zampini } 509427b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 5095b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5096b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 509757de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 509827b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 509927b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 510027b6a85dSStefano Zampini #endif 510157de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 510227b6a85dSStefano Zampini } else if (oldranks) { 5103b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 510427b6a85dSStefano Zampini } else { 510527b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 5106e7931f94SStefano Zampini } 510757de7509SStefano Zampini } 510852e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5109e7931f94SStefano Zampini /* clean up */ 5110e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 511152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 5112e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 5113e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5114e7931f94SStefano Zampini } 511527b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 511657de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5117e7931f94SStefano Zampini 5118e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5119e7931f94SStefano Zampini i = 1; 512027b6a85dSStefano Zampini if (!color) i=0; 512157de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5122e7931f94SStefano Zampini PetscFunctionReturn(0); 5123e7931f94SStefano Zampini } 5124e7931f94SStefano Zampini 5125e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5126e7931f94SStefano Zampini 5127e7931f94SStefano Zampini #undef __FUNCT__ 5128e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 51291ae86dd6SStefano 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[]) 5130e7931f94SStefano Zampini { 513170cf5478SStefano Zampini Mat local_mat; 5132e7931f94SStefano Zampini IS is_sends_internal; 51339d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 51341ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 51359d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5136e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5137e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5138e7931f94SStefano Zampini const PetscInt* is_indices; 5139e7931f94SStefano Zampini MatType new_local_type; 5140e7931f94SStefano Zampini /* buffers */ 5141e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 514228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 51439d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5144e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 51451ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 5146e7931f94SStefano Zampini /* MPI */ 514728143c3dSStefano Zampini MPI_Comm comm,comm_n; 514828143c3dSStefano Zampini PetscSubcomm subcomm; 5149e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 515028143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 515128143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 51521ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 51531ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 51541ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 5155e7931f94SStefano Zampini PetscErrorCode ierr; 5156e7931f94SStefano Zampini 5157e7931f94SStefano Zampini PetscFunctionBegin; 515857de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5159e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 516028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 516157de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 516257de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 516357de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 516457de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 516557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 51661ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 51671ae86dd6SStefano Zampini if (nvecs) { 51681ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 51691ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 51701ae86dd6SStefano Zampini } 517157de7509SStefano Zampini /* further checks */ 5172e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5173e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5174e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5175e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5176e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 517757de7509SStefano Zampini if (reuse && *mat_n) { 517870cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 517957de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 518070cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 518128143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 518270cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 518370cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 518470cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 518570cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 518670cf5478SStefano Zampini } 5187e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5188e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 518957de7509SStefano Zampini 5190e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5191e7931f94SStefano Zampini if (!is_sends) { 519228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 519357de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5194c8587f34SStefano Zampini } else { 5195e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5196e7931f94SStefano Zampini is_sends_internal = is_sends; 5197c8587f34SStefano Zampini } 5198e7931f94SStefano Zampini 5199e7931f94SStefano Zampini /* get comm */ 5200a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5201e7931f94SStefano Zampini 5202e7931f94SStefano Zampini /* compute number of sends */ 5203e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5204e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5205e7931f94SStefano Zampini 5206e7931f94SStefano Zampini /* compute number of receives */ 5207e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5208785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5209e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5210e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5211e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5212e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5213e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5214e7931f94SStefano Zampini 521528143c3dSStefano Zampini /* restrict comm if requested */ 521628143c3dSStefano Zampini subcomm = 0; 521728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 521828143c3dSStefano Zampini if (restrict_comm) { 5219779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5220779c1cceSStefano Zampini 522128143c3dSStefano Zampini color = 0; 522253a05cb3SStefano Zampini if (restrict_full) { 522353a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 522453a05cb3SStefano Zampini } else { 522553a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 522653a05cb3SStefano Zampini } 5227b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 522828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 522928143c3dSStefano Zampini /* check if reuse has been requested */ 523057de7509SStefano Zampini if (reuse) { 523128143c3dSStefano Zampini if (*mat_n) { 523228143c3dSStefano Zampini PetscMPIInt subcommsize2; 523328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 523428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 523528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 523628143c3dSStefano Zampini } else { 523728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 523828143c3dSStefano Zampini } 523928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5240779c1cceSStefano Zampini PetscMPIInt rank; 5241779c1cceSStefano Zampini 5242779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 524328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 524428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 524528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5246306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 524728143c3dSStefano Zampini } 524828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 524928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 525028143c3dSStefano Zampini } else { 525128143c3dSStefano Zampini comm_n = comm; 525228143c3dSStefano Zampini } 525328143c3dSStefano Zampini 5254e7931f94SStefano Zampini /* prepare send/receive buffers */ 5255785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5256e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5257785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5258e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 525928143c3dSStefano Zampini if (nis) { 5260854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 526128143c3dSStefano Zampini } 5262e7931f94SStefano Zampini 526328143c3dSStefano Zampini /* Get data from local matrices */ 52646c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5265e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5266e7931f94SStefano Zampini /* 5267e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5268e7931f94SStefano Zampini send_buffer_idxs should contain: 5269e7931f94SStefano Zampini - MatType_PRIVATE type 5270e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5271e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5272e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5273e7931f94SStefano Zampini */ 52746c4ed002SBarry Smith else { 5275e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 52763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5277854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5278e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5279e7931f94SStefano Zampini send_buffer_idxs[1] = i; 52803bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5281e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 52823bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5283e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5284e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5285e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5286e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5287c8587f34SStefano Zampini } 5288c8587f34SStefano Zampini } 5289e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 529028143c3dSStefano Zampini /* additional is (if any) */ 529128143c3dSStefano Zampini if (nis) { 529228143c3dSStefano Zampini PetscMPIInt psum; 529328143c3dSStefano Zampini PetscInt j; 529428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 529528143c3dSStefano Zampini PetscInt plen; 529628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 529728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 529828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 529928143c3dSStefano Zampini } 5300854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 530128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 530228143c3dSStefano Zampini PetscInt plen; 530328143c3dSStefano Zampini const PetscInt *is_array_idxs; 530428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 530528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 530628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 530728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 530828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 530928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 531028143c3dSStefano Zampini } 531128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 531228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 531328143c3dSStefano Zampini } 531428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 531528143c3dSStefano Zampini } 531628143c3dSStefano Zampini 5317e7931f94SStefano Zampini buf_size_idxs = 0; 5318e7931f94SStefano Zampini buf_size_vals = 0; 531928143c3dSStefano Zampini buf_size_idxs_is = 0; 53201ae86dd6SStefano Zampini buf_size_vecs = 0; 5321e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5322e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5323e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 532428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 53251ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 5326e7931f94SStefano Zampini } 5327785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5328785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 532995ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 53301ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 5331e7931f94SStefano Zampini 5332e7931f94SStefano Zampini /* get new tags for clean communications */ 5333e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5334e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 533528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 53361ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 5337e7931f94SStefano Zampini 5338e7931f94SStefano Zampini /* allocate for requests */ 5339785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5340785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 534195ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 53421ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 5343785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5344785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 534595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 53461ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 5347e7931f94SStefano Zampini 5348e7931f94SStefano Zampini /* communications */ 5349e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5350e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 535128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 53521ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 5353e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5354e7931f94SStefano Zampini source_dest = onodes[i]; 5355e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5356e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5357e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5358e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 535928143c3dSStefano Zampini if (nis) { 536057de7509SStefano Zampini source_dest = onodes_is[i]; 536128143c3dSStefano 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); 536228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 536328143c3dSStefano Zampini } 53641ae86dd6SStefano Zampini if (nvecs) { 53651ae86dd6SStefano Zampini source_dest = onodes[i]; 53661ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 53671ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 53681ae86dd6SStefano Zampini } 5369e7931f94SStefano Zampini } 5370e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5371e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5372e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5373e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 537428143c3dSStefano Zampini if (nis) { 537528143c3dSStefano 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); 537628143c3dSStefano Zampini } 53771ae86dd6SStefano Zampini if (nvecs) { 53781ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 53791ae86dd6SStefano 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); 53801ae86dd6SStefano Zampini } 5381e7931f94SStefano Zampini } 5382e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5383e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5384e7931f94SStefano Zampini 5385e7931f94SStefano Zampini /* assemble new l2g map */ 5386e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5387e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53889d30be91SStefano Zampini new_local_rows = 0; 5389e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53909d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5391e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5392e7931f94SStefano Zampini } 53939d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5394e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 53959d30be91SStefano Zampini new_local_rows = 0; 5396e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 53979d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 53989d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5399e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5400e7931f94SStefano Zampini } 54019d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 54029d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5403e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5404e7931f94SStefano Zampini 5405e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5406e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5407e7931f94SStefano 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) */ 5408e7931f94SStefano Zampini if (n_recvs) { 540928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5410e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5411e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5412e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5413e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5414e7931f94SStefano Zampini break; 5415e7931f94SStefano Zampini } 5416e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5417e7931f94SStefano Zampini } 5418e7931f94SStefano Zampini switch (new_local_type_private) { 541928143c3dSStefano Zampini case MATDENSE_PRIVATE: 542028143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5421e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5422e7931f94SStefano Zampini bs = 1; 542328143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 542428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 542528143c3dSStefano Zampini bs = 1; 542628143c3dSStefano Zampini } 5427e7931f94SStefano Zampini break; 5428e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5429e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5430e7931f94SStefano Zampini bs = 1; 5431e7931f94SStefano Zampini break; 5432e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5433e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5434e7931f94SStefano Zampini break; 5435e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5436e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5437e7931f94SStefano Zampini break; 5438e7931f94SStefano Zampini default: 54399d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5440e7931f94SStefano Zampini break; 5441e7931f94SStefano Zampini } 544228143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 544328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 544428143c3dSStefano Zampini bs = 1; 5445e7931f94SStefano Zampini } 5446e7931f94SStefano Zampini 544770cf5478SStefano Zampini /* create MATIS object if needed */ 544857de7509SStefano Zampini if (!reuse) { 5449e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5450e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 545170cf5478SStefano Zampini } else { 545270cf5478SStefano Zampini /* it also destroys the local matrices */ 545357de7509SStefano Zampini if (*mat_n) { 545470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 545557de7509SStefano Zampini } else { /* this is a fake object */ 545657de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 545757de7509SStefano Zampini } 545870cf5478SStefano Zampini } 545970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5460e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 54619d30be91SStefano Zampini 54629d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 54639d30be91SStefano Zampini 54649d30be91SStefano Zampini /* Global to local map of received indices */ 54659d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 54669d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 54679d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 54689d30be91SStefano Zampini 54699d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 54709d30be91SStefano Zampini buf_size_idxs = 0; 54719d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54729d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 54739d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 54749d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 54759d30be91SStefano Zampini } 54769d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 54779d30be91SStefano Zampini 54789d30be91SStefano Zampini /* set preallocation */ 54799d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 54809d30be91SStefano Zampini if (!newisdense) { 54819d30be91SStefano Zampini PetscInt *new_local_nnz=0; 54829d30be91SStefano Zampini 54839d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 54849d30be91SStefano Zampini if (n_recvs) { 54859d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 54869d30be91SStefano Zampini } 54879d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 54889d30be91SStefano Zampini PetscInt j; 54899d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 54909d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 54919d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 54929d30be91SStefano Zampini } 54939d30be91SStefano Zampini } else { 54949d30be91SStefano Zampini /* TODO */ 54959d30be91SStefano Zampini } 54969d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 54979d30be91SStefano Zampini } 54989d30be91SStefano Zampini if (new_local_nnz) { 54999d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 55009d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 55019d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 55029d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 55039d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 55049d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 55059d30be91SStefano Zampini } else { 55069d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55079d30be91SStefano Zampini } 55089d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 55099d30be91SStefano Zampini } else { 55109d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 55119d30be91SStefano Zampini } 5512e7931f94SStefano Zampini 5513e7931f94SStefano Zampini /* set values */ 5514e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 55159d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5516e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5517e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5518e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 55199d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5520e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5521e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5522e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 552328143c3dSStefano Zampini } else { 552428143c3dSStefano Zampini /* TODO */ 5525e7931f94SStefano Zampini } 5526e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5527e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5528e7931f94SStefano Zampini } 5529e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5530e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 553170cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 553270cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55339d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 5534e7931f94SStefano Zampini 5535dfd14d43SStefano Zampini #if 0 553628143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5537e7931f94SStefano Zampini Vec lvec,rvec; 5538e7931f94SStefano Zampini PetscReal infty_error; 5539e7931f94SStefano Zampini 55402a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5541e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5542e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5543e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 554470cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5545e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5546e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5547e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5548e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5549e7931f94SStefano Zampini } 555028143c3dSStefano Zampini #endif 5551e7931f94SStefano Zampini 555228143c3dSStefano Zampini /* assemble new additional is (if any) */ 555328143c3dSStefano Zampini if (nis) { 555428143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 555528143c3dSStefano Zampini 555628143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5557854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 555828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 555928143c3dSStefano Zampini psum = 0; 556028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 556128143c3dSStefano Zampini for (j=0;j<nis;j++) { 556228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 556328143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 556428143c3dSStefano Zampini psum += plen; 556528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 556628143c3dSStefano Zampini } 556728143c3dSStefano Zampini } 5568854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5569854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 557028143c3dSStefano Zampini for (i=1;i<nis;i++) { 557128143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 557228143c3dSStefano Zampini } 557328143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 557428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 557528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 557628143c3dSStefano Zampini for (j=0;j<nis;j++) { 557728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 557828143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 557928143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 558028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 558128143c3dSStefano Zampini } 558228143c3dSStefano Zampini } 558328143c3dSStefano Zampini for (i=0;i<nis;i++) { 558428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 558528143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 558628143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 558728143c3dSStefano Zampini } 558828143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 558928143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 559028143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 559128143c3dSStefano Zampini } 5592e7931f94SStefano Zampini /* free workspace */ 559328143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5594e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5595e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5596e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5597e7931f94SStefano Zampini if (isdense) { 5598e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5599e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5600e7931f94SStefano Zampini } else { 5601e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5602e7931f94SStefano Zampini } 560328143c3dSStefano Zampini if (nis) { 560428143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 560528143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 560628143c3dSStefano Zampini } 56071ae86dd6SStefano Zampini 56081ae86dd6SStefano Zampini if (nvecs) { 56091ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56101ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 56111ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56121ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56131ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 56141ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 56151ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 56161ae86dd6SStefano Zampini /* set values */ 56171ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 56181ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 56191ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56201ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 56211ae86dd6SStefano Zampini PetscInt j; 56221ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 56231ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 56241ae86dd6SStefano Zampini } 56251ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 56261ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 56271ae86dd6SStefano Zampini } 56281ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 56291ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 56301ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 56311ae86dd6SStefano Zampini } 56321ae86dd6SStefano Zampini 56331ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 56341ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5635e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5636e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 56371ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 563828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5639e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5640e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 56411ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 564228143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5643e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5644e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5645e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5646e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5647e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 564828143c3dSStefano Zampini if (nis) { 564928143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 565028143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 565128143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 565228143c3dSStefano Zampini } 565328143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 565428143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 565528143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 565628143c3dSStefano Zampini for (i=0;i<nis;i++) { 565728143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 565828143c3dSStefano Zampini } 56591ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 56601ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 56611ae86dd6SStefano Zampini } 566253a05cb3SStefano Zampini *mat_n = NULL; 566328143c3dSStefano Zampini } 5664e7931f94SStefano Zampini PetscFunctionReturn(0); 5665e7931f94SStefano Zampini } 5666a57a6d2fSStefano Zampini 566712edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5668af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 566912edc857SStefano Zampini 5670c8587f34SStefano Zampini #undef __FUNCT__ 5671c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5672c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5673c8587f34SStefano Zampini { 5674c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5675c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 567620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 56771ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 56789881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 567920a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 56806e683305SStefano Zampini IS coarse_is,*isarray; 56816e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 568230368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5683f9eb5b7dSStefano Zampini PC pc_temp; 5684c8587f34SStefano Zampini PCType coarse_pc_type; 5685c8587f34SStefano Zampini KSPType coarse_ksp_type; 5686f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 56874f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 56886e683305SStefano Zampini Mat t_coarse_mat_is; 568957de7509SStefano Zampini PetscInt ncoarse; 569068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 569122bc73bbSStefano Zampini PetscScalar *array; 569257de7509SStefano Zampini MatReuse coarse_mat_reuse; 569357de7509SStefano Zampini PetscBool restr, full_restr, have_void; 56949881197aSStefano Zampini PetscErrorCode ierr; 5695fdc09c96SStefano Zampini 5696c8587f34SStefano Zampini PetscFunctionBegin; 5697c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 569868457ee5SStefano 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 */ 5699fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 57005a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5701fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5702f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5703f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5704f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5705fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 570651bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 570751bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5708dc4bcba2SStefano Zampini PC pc; 5709dc4bcba2SStefano Zampini PetscBool isbddc; 5710dc4bcba2SStefano Zampini 5711dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5712dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5713dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5714dc4bcba2SStefano Zampini if (isbddc) { 571563c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 571663c961adSStefano Zampini } else { 5717727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 571863c961adSStefano Zampini } 5719fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5720fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5721fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5722f4ddd8eeSStefano Zampini } 5723fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5724fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5725f4ddd8eeSStefano Zampini } 572670cf5478SStefano Zampini /* reset any subassembling information */ 572757de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 572870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 572957de7509SStefano Zampini } 57306e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5731fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5732f4ddd8eeSStefano Zampini } 573357de7509SStefano Zampini /* assemble coarse matrix */ 573457de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 573557de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 573657de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 573757de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 573818a45a71SStefano Zampini } else { 573957de7509SStefano Zampini coarse_mat = NULL; 574057de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 57416e683305SStefano Zampini } 5742e7931f94SStefano Zampini 5743abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5744abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5745abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5746abbbba34SStefano Zampini 5747abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 574822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 574922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 575022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 575122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5752e176bc59SStefano 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); 57536e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 57546e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57556e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5756abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5757abbbba34SStefano Zampini 575857de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 575957de7509SStefano Zampini im_active = !!(pcis->n); 576057de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 576157de7509SStefano Zampini 576214f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 576357de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 576457de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 576557de7509SStefano Zampini coarse_mat_is = NULL; 576657de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 576757de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 576857de7509SStefano Zampini full_restr = PETSC_TRUE; 57691ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 577057de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 577157de7509SStefano Zampini if (multilevel_requested) { 577257de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 577357de7509SStefano Zampini restr = PETSC_FALSE; 577457de7509SStefano Zampini full_restr = PETSC_FALSE; 577557de7509SStefano Zampini } else { 577657de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 577757de7509SStefano Zampini restr = PETSC_TRUE; 577857de7509SStefano Zampini full_restr = PETSC_TRUE; 577957de7509SStefano Zampini } 57804b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 578157de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 578257de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 5783a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 578457de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 5785a198735bSStefano Zampini } else { 5786a198735bSStefano Zampini PetscMPIInt size,rank; 5787a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 5788a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 5789a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 5790a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5791a198735bSStefano Zampini } 579257de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 579357de7509SStefano Zampini PetscInt psum; 579457de7509SStefano Zampini PetscMPIInt size; 579557de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 579657de7509SStefano Zampini else psum = 0; 579757de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 579857de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 579957de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 580057de7509SStefano Zampini } 580157de7509SStefano Zampini /* determine if we can go multilevel */ 580257de7509SStefano Zampini if (multilevel_requested) { 580357de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 580457de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 580557de7509SStefano Zampini } 580657de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 580757de7509SStefano Zampini 5808e4d548c7SStefano Zampini /* dump subassembling pattern */ 5809e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5810e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5811e4d548c7SStefano Zampini } 5812e4d548c7SStefano Zampini 58136e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 581427b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 58156e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 58166e683305SStefano Zampini const PetscInt *idxs; 58176e683305SStefano Zampini ISLocalToGlobalMapping tmap; 58186e683305SStefano Zampini 58196e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 58200be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 58216e683305SStefano Zampini /* allocate space for temporary storage */ 5822854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5823854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 58246e683305SStefano Zampini /* allocate for IS array */ 58256e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 58266e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 582727b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 582830368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5829854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 58306e683305SStefano Zampini /* dofs splitting */ 58316e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 58326e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 58336e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 58346e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58356e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58366e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 58376e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 583830368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 58396e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 58406e683305SStefano Zampini } 58416e683305SStefano Zampini /* neumann boundaries */ 58426e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 58436e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 58446e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 58456e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58466e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 58476e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 58486e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 584930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 58506e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 58516e683305SStefano Zampini } 58526e683305SStefano Zampini /* free memory */ 58536e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 58546e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 58556e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 58566e683305SStefano Zampini } else { 58576e683305SStefano Zampini nis = 0; 58586e683305SStefano Zampini nisdofs = 0; 58596e683305SStefano Zampini nisneu = 0; 586030368db7SStefano Zampini nisvert = 0; 58616e683305SStefano Zampini isarray = NULL; 58626e683305SStefano Zampini } 58636e683305SStefano Zampini /* destroy no longer needed map */ 58646e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 58656e683305SStefano Zampini 586657de7509SStefano Zampini /* subassemble */ 586757de7509SStefano Zampini if (multilevel_allowed) { 58681ae86dd6SStefano Zampini Vec vp[1]; 58691ae86dd6SStefano Zampini PetscInt nvecs = 0; 587057de7509SStefano Zampini PetscBool reuse,reuser; 58711ae86dd6SStefano Zampini 587257de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 587357de7509SStefano Zampini else reuse = PETSC_FALSE; 587457de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 58751ae86dd6SStefano Zampini vp[0] = NULL; 58761ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 58771ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 58781ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 58791ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 58801ae86dd6SStefano Zampini nvecs = 1; 58811ae86dd6SStefano Zampini 58821ae86dd6SStefano Zampini if (pcbddc->divudotp) { 5883a198735bSStefano Zampini Mat B,loc_divudotp; 58841ae86dd6SStefano Zampini Vec v,p; 58851ae86dd6SStefano Zampini IS dummy; 58861ae86dd6SStefano Zampini PetscInt np; 58871ae86dd6SStefano Zampini 5888a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 5889a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 58901ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 5891a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 58921ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 58931ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 58941ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 58951ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 58961ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 58971ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 58981ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 58991ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 59001ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 59011ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 59021ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 59031ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 590474e2c79eSStefano Zampini } 59051ae86dd6SStefano Zampini } 59061ae86dd6SStefano Zampini if (reuser) { 59071ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 590874e2c79eSStefano Zampini } else { 59091ae86dd6SStefano 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); 59101ae86dd6SStefano Zampini } 59111ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 59121ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 59131ae86dd6SStefano Zampini PetscInt nl; 59141ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 59151ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 59161ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59171ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 59181ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 59191ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 59201ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 59211ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 5922a198735bSStefano Zampini } else { 5923a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 59241ae86dd6SStefano Zampini } 59251ae86dd6SStefano Zampini } else { 59261ae86dd6SStefano 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); 59276e683305SStefano Zampini } 592857de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 592957de7509SStefano Zampini PetscMPIInt size; 593057de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 593157de7509SStefano Zampini if (!multilevel_allowed) { 593257de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 59336e683305SStefano Zampini } else { 593457de7509SStefano Zampini Mat A; 5935779c1cceSStefano Zampini 593657de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 593757de7509SStefano Zampini if (coarse_mat_is) { 593857de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 593957de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 594057de7509SStefano Zampini coarse_mat = coarse_mat_is; 594157de7509SStefano Zampini } 594257de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 594357de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 594457de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5945779c1cceSStefano Zampini } 5946779c1cceSStefano Zampini } 594757de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 594857de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 59496e683305SStefano Zampini 59506e683305SStefano Zampini /* create local to global scatters for coarse problem */ 595168457ee5SStefano Zampini if (compute_vecs) { 59526e683305SStefano Zampini PetscInt lrows; 59536e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 595457de7509SStefano Zampini if (coarse_mat) { 595557de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 59566e683305SStefano Zampini } else { 59576e683305SStefano Zampini lrows = 0; 59586e683305SStefano Zampini } 59596e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 59606e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 59616e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 59626e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59636e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 59646e683305SStefano Zampini } 59656e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5966c8587f34SStefano Zampini 5967f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5968f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5969f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5970f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5971f9eb5b7dSStefano Zampini } else { 5972f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5973f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5974c8587f34SStefano Zampini } 5975c8587f34SStefano Zampini 59766e683305SStefano Zampini /* print some info if requested */ 59776e683305SStefano Zampini if (pcbddc->dbg_flag) { 59786e683305SStefano Zampini if (!multilevel_allowed) { 59796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 59806e683305SStefano Zampini if (multilevel_requested) { 59816e683305SStefano 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); 59826e683305SStefano Zampini } else if (pcbddc->max_levels) { 59836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 59846e683305SStefano Zampini } 59856e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59866e683305SStefano Zampini } 59876e683305SStefano Zampini } 59886e683305SStefano Zampini 5989f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 599057de7509SStefano Zampini if (coarse_mat) { 59916a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 59926e683305SStefano Zampini if (pcbddc->dbg_flag) { 599357de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 59946e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 59956e683305SStefano Zampini } 5996f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5997312be037SStefano Zampini char prefix[256],str_level[16]; 5998e604994aSStefano Zampini size_t len; 599957de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 6000422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 6001c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 6002f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 600357de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6004c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 60056e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 6006c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 6007c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6008e604994aSStefano Zampini /* prefix */ 6009e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 6010e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 6011e604994aSStefano Zampini if (!pcbddc->current_level) { 6012e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 6013e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 6014c8587f34SStefano Zampini } else { 6015e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 6016312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 6017312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 601834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 6019312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 6020e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 6021e604994aSStefano Zampini } 6022e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 60233e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 60243e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 60253e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 60263e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 6027f9eb5b7dSStefano Zampini /* allow user customization */ 6028f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 60293e3c6dadSStefano Zampini } 60303e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 603151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 60323e3c6dadSStefano Zampini if (nisdofs) { 60333e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 60343e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 60353e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 60363e3c6dadSStefano Zampini } 60373e3c6dadSStefano Zampini } 60383e3c6dadSStefano Zampini if (nisneu) { 60393e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 60403e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 6041312be037SStefano Zampini } 604230368db7SStefano Zampini if (nisvert) { 604330368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 604430368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 604530368db7SStefano Zampini } 6046f9eb5b7dSStefano Zampini 6047f9eb5b7dSStefano Zampini /* get some info after set from options */ 6048f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 6049f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 60504f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 60516e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 6052f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 6053f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 6054f9eb5b7dSStefano Zampini } 605539f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 60564f3a063dSStefano Zampini if (isredundant) { 60574f3a063dSStefano Zampini KSP inner_ksp; 60584f3a063dSStefano Zampini PC inner_pc; 60594f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 60604f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 60614f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 60624f3a063dSStefano Zampini } 6063f9eb5b7dSStefano Zampini 606457de7509SStefano Zampini /* parameters which miss an API */ 606557de7509SStefano Zampini if (isbddc) { 6066720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 6067720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 606857de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 606927b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 607027b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 6071a198735bSStefano Zampini Mat coarsedivudotp_is; 6072a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 6073a198735bSStefano Zampini IS row,col; 6074a198735bSStefano Zampini const PetscInt *gidxs; 6075a198735bSStefano Zampini PetscInt n,st,M,N; 6076a198735bSStefano Zampini 6077a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 6078a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 6079a198735bSStefano Zampini st = st-n; 6080a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 6081a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 6082a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 6083a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6084a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 6085a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 6086a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 6087a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 6088a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 6089a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 6090a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 6091a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 6092a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 6093a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 6094a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 6095a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 6096a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 6097a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 6098a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 6099a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 61008ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 6101a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 6102720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 610359e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 6104720d30f9SStefano Zampini } 6105d4d8cf7bSStefano Zampini } 61069881197aSStefano Zampini 61073301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 61085a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 61093301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 61103301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 61113301b35fSStefano Zampini } 61123301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 61133301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 61143301b35fSStefano Zampini } 61153301b35fSStefano Zampini if (pc->pmat->spd_set) { 61163301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 61173301b35fSStefano Zampini } 611827b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 611927b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 612027b6a85dSStefano Zampini } 61216e683305SStefano Zampini /* set operators */ 61225f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 61236e683305SStefano Zampini if (pcbddc->dbg_flag) { 61246e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 61256e683305SStefano Zampini } 61266e683305SStefano Zampini } 61276e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 6128b1ecc7b1SStefano Zampini #if 0 6129b9b85e73SStefano Zampini { 6130b9b85e73SStefano Zampini PetscViewer viewer; 6131b9b85e73SStefano Zampini char filename[256]; 6132b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 6133b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 61346a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6135b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 6136f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 6137b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 6138b9b85e73SStefano Zampini } 6139b9b85e73SStefano Zampini #endif 6140f9eb5b7dSStefano Zampini 614198a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 614298a51de6SStefano Zampini Vec crhs,csol; 614304708bb6SStefano Zampini 6144f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 6145f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 6146f347579bSStefano Zampini if (!csol) { 61472a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 6148f9eb5b7dSStefano Zampini } 6149f347579bSStefano Zampini if (!crhs) { 61502a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 6151f347579bSStefano Zampini } 6152b0f5fe93SStefano Zampini } 61531ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 6154b0f5fe93SStefano Zampini 6155b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 6156b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 6157b0f5fe93SStefano Zampini 6158b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 61594f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 61604f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 61614f1b2e48SStefano Zampini } 6162b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 6163b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 6164b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6165b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6166b0f5fe93SStefano Zampini if (coarse_mat) { 6167b0f5fe93SStefano Zampini Vec nullv; 6168b0f5fe93SStefano Zampini PetscScalar *array,*array2; 6169b0f5fe93SStefano Zampini PetscInt nl; 6170b0f5fe93SStefano Zampini 6171b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 6172b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 6173b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6174b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 6175b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 6176b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 6177b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 6178b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 6179b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 6180b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 6181b0f5fe93SStefano Zampini } 6182b0f5fe93SStefano Zampini } 6183b0f5fe93SStefano Zampini 6184b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 6185b0f5fe93SStefano Zampini PetscBool ispreonly; 6186b0f5fe93SStefano Zampini 6187b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6188b0f5fe93SStefano Zampini PetscBool isnull; 6189b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 6190bef83e63SStefano Zampini if (isnull) { 6191b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 6192b0f5fe93SStefano Zampini } 6193bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 6194b0f5fe93SStefano Zampini } 6195b0f5fe93SStefano Zampini /* setup coarse ksp */ 6196b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 6197cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 6198cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 61996e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 6200c8587f34SStefano Zampini KSP check_ksp; 62012b510759SStefano Zampini KSPType check_ksp_type; 6202c8587f34SStefano Zampini PC check_pc; 62036e683305SStefano Zampini Vec check_vec,coarse_vec; 62046a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 62052b510759SStefano Zampini PetscInt its; 62066e683305SStefano Zampini PetscBool compute_eigs; 62076e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 62086e683305SStefano Zampini PetscInt neigs; 62098e185a42SStefano Zampini const char *prefix; 6210c8587f34SStefano Zampini 62112b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 62126e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 6213422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 621423ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 6215f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 6216e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 6217e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 6218e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 62192b510759SStefano Zampini if (ispreonly) { 62202b510759SStefano Zampini check_ksp_type = KSPPREONLY; 62216e683305SStefano Zampini compute_eigs = PETSC_FALSE; 62222b510759SStefano Zampini } else { 6223cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 62246e683305SStefano Zampini compute_eigs = PETSC_TRUE; 6225c8587f34SStefano Zampini } 6226c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 62276e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 62286e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 62296e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 6230a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 6231a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 6232a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 6233a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 6234c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 6235c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 6236c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 6237c8587f34SStefano Zampini /* create random vec */ 62382701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 6239c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 62406e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6241c8587f34SStefano Zampini /* solve coarse problem */ 62426e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6243cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 62446e683305SStefano Zampini if (compute_eigs) { 6245854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6246854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 62476e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 62481ae86dd6SStefano Zampini if (neigs) { 62496e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 62506e683305SStefano Zampini lambda_min = eigs_r[0]; 62516e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 62522701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 62532701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6254cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6255cbcc2c2aSStefano Zampini } 6256c8587f34SStefano Zampini } 6257c8587f34SStefano Zampini } 62581ae86dd6SStefano Zampini } 6259cbcc2c2aSStefano Zampini 6260c8587f34SStefano Zampini /* check coarse problem residual error */ 62616e683305SStefano Zampini if (pcbddc->dbg_flag) { 62626e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 62636e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62646e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6265c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 62666e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 62676e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6268779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 62696e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 62706e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 62716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 62726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6273b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6274b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6275b0f5fe93SStefano Zampini } 62766e683305SStefano Zampini if (compute_eigs) { 62776e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6278deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6279c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 62806e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 62816e683305SStefano 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); 62826e683305SStefano Zampini for (i=0;i<neigs;i++) { 62836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6284c8587f34SStefano Zampini } 62856e683305SStefano Zampini } 62866e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 62876e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 62886e683305SStefano Zampini } 6289e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 62902701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6291c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 62926e683305SStefano Zampini if (compute_eigs) { 62936e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 62946e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6295c8587f34SStefano Zampini } 62966e683305SStefano Zampini } 62976e683305SStefano Zampini } 6298bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6299cbcc2c2aSStefano Zampini /* print additional info */ 6300cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 63016e683305SStefano Zampini /* waits until all processes reaches this point */ 63026e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6303cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6304cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6305cbcc2c2aSStefano Zampini } 6306cbcc2c2aSStefano Zampini 63072b510759SStefano Zampini /* free memory */ 6308fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6309c8587f34SStefano Zampini PetscFunctionReturn(0); 6310c8587f34SStefano Zampini } 6311674ae819SStefano Zampini 6312f34684f1SStefano Zampini #undef __FUNCT__ 6313f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6314f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6315f34684f1SStefano Zampini { 6316f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6317f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6318f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6319dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6320dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 632173be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6322dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6323f34684f1SStefano Zampini PetscErrorCode ierr; 6324f34684f1SStefano Zampini 6325f34684f1SStefano Zampini PetscFunctionBegin; 6326f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 63276c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6328dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 63293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6330dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6331dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 63326583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6333dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6334dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6335dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 63366c4ed002SBarry 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); 6337dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6338dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6339dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6340dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6341dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6342f34684f1SStefano Zampini 6343f34684f1SStefano Zampini /* check numbering */ 6344f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6345019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6346dc456d91SStefano Zampini PetscInt i; 6347b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6348f34684f1SStefano Zampini 6349f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6350f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6351f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 63521575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6353019a44ceSStefano Zampini /* counter */ 6354019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6355019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6356019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6357019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6358019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6359019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6360f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6361f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6362727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6363f34684f1SStefano Zampini } 6364f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6365f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6366f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6367e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6368e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6369e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6370e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6371f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6372019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6373f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6374019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 63752c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 637675c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6377b9b85e73SStefano Zampini set_error = PETSC_TRUE; 63782c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 63792c66d082SStefano 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); 6380f34684f1SStefano Zampini } 6381f34684f1SStefano Zampini } 6382019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6383b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6384f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6385f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6386f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6387f34684f1SStefano Zampini } 6388f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6389f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6390e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6391e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6392f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6393f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6394b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6395ca8b9ea9SStefano Zampini PetscInt *gidxs; 6396ca8b9ea9SStefano Zampini 6397ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 63983bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6399f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6400f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6401f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6402f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 64034bc2dc4bSStefano 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); 6404f34684f1SStefano Zampini } 6405f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6406ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6407f34684f1SStefano Zampini } 6408f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 64091575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6410302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6411f34684f1SStefano Zampini } 64128bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6413f34684f1SStefano Zampini /* get back data */ 6414f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6415f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6416674ae819SStefano Zampini PetscFunctionReturn(0); 6417674ae819SStefano Zampini } 6418674ae819SStefano Zampini 6419e456f2a8SStefano Zampini #undef __FUNCT__ 6420e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6421a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6422e456f2a8SStefano Zampini { 6423e456f2a8SStefano Zampini IS localis_t; 6424a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6425e456f2a8SStefano Zampini PetscScalar *vals; 6426e456f2a8SStefano Zampini PetscErrorCode ierr; 6427e456f2a8SStefano Zampini 6428e456f2a8SStefano Zampini PetscFunctionBegin; 6429a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6430e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6431854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6432e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6433e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6434a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6435a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 64361035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6437a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 64381035eff8SStefano Zampini } 6439a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6440e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6441e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6442a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6443a7dc3881SStefano Zampini /* now compute set in local ordering */ 6444a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6445a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6446a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6447a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6448a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6449ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6450e456f2a8SStefano Zampini lsize++; 6451e456f2a8SStefano Zampini } 6452e456f2a8SStefano Zampini } 6453854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6454a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6455ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6456e456f2a8SStefano Zampini idxs[lsize++] = i; 6457e456f2a8SStefano Zampini } 6458e456f2a8SStefano Zampini } 6459a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6460a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6461e456f2a8SStefano Zampini *localis = localis_t; 6462e456f2a8SStefano Zampini PetscFunctionReturn(0); 6463e456f2a8SStefano Zampini } 6464906d46d4SStefano Zampini 6465b96c3477SStefano Zampini #undef __FUNCT__ 6466b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 646708122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6468b96c3477SStefano Zampini { 6469a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6470b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6471b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6472a64f4aa4SStefano Zampini Mat S_j; 6473b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6474b96c3477SStefano Zampini PetscBool free_used_adj; 6475b96c3477SStefano Zampini PetscErrorCode ierr; 6476b96c3477SStefano Zampini 6477b96c3477SStefano Zampini PetscFunctionBegin; 6478b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6479b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 648008122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6481b96c3477SStefano Zampini used_xadj = NULL; 6482b96c3477SStefano Zampini used_adjncy = NULL; 6483b96c3477SStefano Zampini } else { 648408122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 648508122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 648608122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 648708122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6488b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6489b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6490b96c3477SStefano Zampini } else { 64912fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6492b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6493b96c3477SStefano Zampini PetscInt nvtxs; 6494b96c3477SStefano Zampini 64952fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 64962fffb893SStefano Zampini if (flg_row) { 6497b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6498b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6499b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6500b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 65012fffb893SStefano Zampini } else { 65022fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 65032fffb893SStefano Zampini used_xadj = NULL; 65042fffb893SStefano Zampini used_adjncy = NULL; 65052fffb893SStefano Zampini } 65062fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6507b96c3477SStefano Zampini } 6508b96c3477SStefano Zampini } 6509d5574798SStefano Zampini 6510d5574798SStefano Zampini /* setup sub_schurs data */ 6511a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6512df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6513df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6514a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 651591af6908SStefano 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); 6516a64f4aa4SStefano Zampini } else { 65176816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 65184d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 6519a3df083aSStefano Zampini PetscInt benign_n; 652072b8c272SStefano Zampini Mat change = NULL; 65219d54b7f4SStefano Zampini Vec scaling = NULL; 652272b8c272SStefano Zampini IS change_primal = NULL; 6523a3df083aSStefano Zampini 65245feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 65255feab87aSStefano Zampini PetscInt n_vertices; 65265feab87aSStefano Zampini 65275feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 65282034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 65295feab87aSStefano Zampini } 653004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 653104708bb6SStefano Zampini if (!isseqaij) { 653204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 653304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 653404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 653504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 653604708bb6SStefano Zampini } else { 6537511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 653804708bb6SStefano Zampini } 653904708bb6SStefano Zampini } 6540a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6541a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6542ca92afb2SStefano Zampini } else { 6543a3df083aSStefano Zampini benign_n = 0; 6544ca92afb2SStefano Zampini } 6545b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6546b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6547b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 654872b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 654922db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6550b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 655122db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6552b7ab4a40SStefano Zampini } 6553b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6554b7ab4a40SStefano 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 */ 6555b7ab4a40SStefano Zampini if (need_change) { 655688c03ad3SStefano Zampini PC_IS *pcisf; 655788c03ad3SStefano Zampini PC_BDDC *pcbddcf; 655888c03ad3SStefano Zampini PC pcf; 655988c03ad3SStefano Zampini 6560e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 656188c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 656288c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 656388c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 656488c03ad3SStefano Zampini /* hacks */ 656588c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 656672b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 656772b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 656872b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 656972b8c272SStefano Zampini pcisf->n = pcis->n; 657072b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 657188c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 657288c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 657388c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 657488c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 657588c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 657688c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 657772b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 657888c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 657988c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 658072b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 658172b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 658272b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 658372b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 658472b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 658588c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 658672b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 658788c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 658888c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 658988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 659088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 659188c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 659288c03ad3SStefano Zampini pcf->ops->destroy = NULL; 659388c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 659488c03ad3SStefano Zampini } 65959d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 659691af6908SStefano 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); 659772b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 659872b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6599ca92afb2SStefano Zampini } 6600d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6601b96c3477SStefano Zampini 6602b96c3477SStefano Zampini /* free adjacency */ 6603b96c3477SStefano Zampini if (free_used_adj) { 6604b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6605b96c3477SStefano Zampini } 6606b96c3477SStefano Zampini PetscFunctionReturn(0); 6607b96c3477SStefano Zampini } 6608b96c3477SStefano Zampini 6609b96c3477SStefano Zampini #undef __FUNCT__ 6610b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 661108122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6612b96c3477SStefano Zampini { 6613b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6614b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6615b96c3477SStefano Zampini PCBDDCGraph graph; 6616b96c3477SStefano Zampini PetscErrorCode ierr; 6617b96c3477SStefano Zampini 6618b96c3477SStefano Zampini PetscFunctionBegin; 6619b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 662008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 66213301b35fSStefano Zampini IS verticesIS,verticescomm; 66223301b35fSStefano Zampini PetscInt vsize,*idxs; 6623b96c3477SStefano Zampini 6624b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 66253301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 66263301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 66273301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 66283301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 6629*c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 6630b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 66317fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6632441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 66333301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6634b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6635b96c3477SStefano Zampini } else { 6636b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6637b96c3477SStefano Zampini } 6638e4d548c7SStefano Zampini /* print some info */ 6639e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6640e4d548c7SStefano Zampini IS vertices; 6641e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 6642*c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6643e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6644e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6645e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6646e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6647e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6648e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6649e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6650e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6651e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6652*c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6653e4d548c7SStefano Zampini } 6654b96c3477SStefano Zampini 6655b96c3477SStefano Zampini /* sub_schurs init */ 6656b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 6657b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 6658b334f244SStefano Zampini } 6659b334f244SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6660a64f4aa4SStefano Zampini 6661b96c3477SStefano Zampini /* free graph struct */ 666208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6663b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6664b96c3477SStefano Zampini } 6665b96c3477SStefano Zampini PetscFunctionReturn(0); 6666b96c3477SStefano Zampini } 6667fa34dd3eSStefano Zampini 6668fa34dd3eSStefano Zampini #undef __FUNCT__ 6669fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6670fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6671fa34dd3eSStefano Zampini { 6672fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6673fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6674fa34dd3eSStefano Zampini PetscErrorCode ierr; 6675fa34dd3eSStefano Zampini 6676fa34dd3eSStefano Zampini PetscFunctionBegin; 6677fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6678fa34dd3eSStefano Zampini IS zerodiag = NULL; 66794f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6680fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 66814f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 668275c01103SStefano Zampini PetscReal norm; 6683fa34dd3eSStefano Zampini PetscInt i; 6684fa34dd3eSStefano Zampini 6685fa34dd3eSStefano Zampini /* B0 and B0_B */ 6686fa34dd3eSStefano Zampini if (zerodiag) { 6687fa34dd3eSStefano Zampini IS dummy; 6688fa34dd3eSStefano Zampini 66894f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 66904f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6691fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6692fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6693fa34dd3eSStefano Zampini } 6694fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6695fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6696fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6697fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6698fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6699fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6700fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6701fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6702fa34dd3eSStefano Zampini /* S_j */ 6703fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6704fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6705fa34dd3eSStefano Zampini 6706fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6707fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6708fa34dd3eSStefano Zampini /* continuous in primal space */ 6709fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6710fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6711fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6712fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 67134f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 67144f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6715fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6716fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6717fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6718fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6719fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6720fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6721fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6722fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6723fa34dd3eSStefano Zampini 6724fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6725fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6726fa34dd3eSStefano Zampini /* local with Schur */ 6727fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6728fa34dd3eSStefano Zampini if (zerodiag) { 6729fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67304f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6731fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6732fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6733fa34dd3eSStefano Zampini } 6734fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6735fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6736fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6737fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6738fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6739fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6740fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6741fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6742fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6743fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6744fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6745fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6746fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6747fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6748fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6749fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6750fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6751fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6752fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6753fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6754fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6755fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6756fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6757fa34dd3eSStefano Zampini if (zerodiag) { 6758fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6759fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 67604f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6761fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6762fa34dd3eSStefano Zampini } 6763fa34dd3eSStefano Zampini /* BDDC */ 6764fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6765fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6766fa34dd3eSStefano Zampini 6767fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6768fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6769fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6770fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 67714f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 67724f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6773fa34dd3eSStefano Zampini } 67744f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6775fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6776fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6777fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6778fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6779fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6780fa34dd3eSStefano Zampini } 6781fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6782fa34dd3eSStefano Zampini } 6783